commit 7e689d5775159d358987be4663c83ae11c7a651a
Author: aska-0096 <haocwang@amd.com>
Date:   Thu Jul 18 08:37:13 2024 +0000

    initial commit

diff --git a/.clang-format b/.clang-format
new file mode 100644
index 0000000000..22f2674966
--- /dev/null
+++ b/.clang-format
@@ -0,0 +1,90 @@
+---
+Language:        Cpp
+AccessModifierOffset: 0
+AlignAfterOpenBracket: Align
+AlignConsecutiveAssignments: true
+AlignConsecutiveDeclarations: false
+AlignEscapedNewlinesLeft: true
+AlignOperands:   true
+AlignTrailingComments: true
+AllowAllParametersOfDeclarationOnNextLine: true
+AllowShortBlocksOnASingleLine: true
+AllowShortCaseLabelsOnASingleLine: true
+AllowShortFunctionsOnASingleLine: All
+AllowShortIfStatementsOnASingleLine: false
+AllowShortLoopsOnASingleLine: false
+AlwaysBreakAfterDefinitionReturnType: None
+AlwaysBreakAfterReturnType: None
+AlwaysBreakBeforeMultilineStrings: false
+AlwaysBreakTemplateDeclarations: true
+BinPackArguments: false
+BinPackParameters: false
+BraceWrapping:   
+  AfterClass:      true
+  AfterControlStatement: true
+  AfterEnum:       true
+  AfterFunction:   true
+  AfterNamespace:  false
+  AfterObjCDeclaration: true
+  AfterStruct:     true
+  AfterUnion:      true
+  BeforeCatch:     true
+  BeforeElse:      true
+  IndentBraces:    false
+BreakBeforeBinaryOperators: None
+BreakBeforeBraces: Custom
+BreakBeforeTernaryOperators: true
+BreakConstructorInitializersBeforeComma: false
+ColumnLimit:     100
+CommentPragmas:  '^ IWYU pragma:'
+ConstructorInitializerAllOnOneLineOrOnePerLine: true
+ConstructorInitializerIndentWidth: 4
+ContinuationIndentWidth: 4
+Cpp11BracedListStyle: true
+DerivePointerAlignment: false
+DisableFormat:   false
+ExperimentalAutoDetectBinPacking: false
+ForEachMacros:   [ foreach, Q_FOREACH, BOOST_FOREACH ]
+IncludeCategories: 
+  - Regex:           '^"(llvm|llvm-c|clang|clang-c)/'
+    Priority:        2
+  - Regex:           '^(<|"(gtest|isl|json)/)'
+    Priority:        3
+  - Regex:           '.*'
+    Priority:        1
+IndentCaseLabels: false
+IndentWidth:     4
+IndentWrappedFunctionNames: false
+KeepEmptyLinesAtTheStartOfBlocks: true
+MacroBlockBegin: ''
+MacroBlockEnd:   ''
+MaxEmptyLinesToKeep: 1
+NamespaceIndentation: None
+ObjCBlockIndentWidth: 2
+ObjCSpaceAfterProperty: false
+ObjCSpaceBeforeProtocolList: true
+PenaltyBreakBeforeFirstCallParameter: 19
+PenaltyBreakComment: 300
+PenaltyBreakFirstLessLess: 120
+PenaltyBreakString: 1000
+PenaltyExcessCharacter: 1000000
+PenaltyReturnTypeOnItsOwnLine: 60
+PointerAlignment: Left
+ReflowComments:  true
+SortIncludes:    false
+SpaceAfterCStyleCast: false
+# SpaceAfterTemplateKeyword: true
+SpaceBeforeAssignmentOperators: true
+SpaceBeforeParens: Never
+SpaceInEmptyParentheses: false
+SpacesBeforeTrailingComments: 1
+SpacesInAngles:  false
+SpacesInContainerLiterals: true
+SpacesInCStyleCastParentheses: false
+SpacesInParentheses: false
+SpacesInSquareBrackets: false
+Standard:        Cpp11
+TabWidth:        8
+UseTab:          Never
+...
+
diff --git a/.clang-tidy b/.clang-tidy
new file mode 100644
index 0000000000..5c2b781687
--- /dev/null
+++ b/.clang-tidy
@@ -0,0 +1,3 @@
+CheckOptions:
+  - key: bugprone-reserved-identifier.AllowedIdentifiers
+    value: '__HIP_PLATFORM_HCC__;__HIP_ROCclr__'
diff --git a/.github/dependabot.yml b/.github/dependabot.yml
new file mode 100644
index 0000000000..276690bd4f
--- /dev/null
+++ b/.github/dependabot.yml
@@ -0,0 +1,12 @@
+# To get started with Dependabot version updates, you'll need to specify which
+# package ecosystems to update and where the package manifests are located.
+# Please see the documentation for all configuration options:
+# https://docs.github.com/github/administering-a-repository/configuration-options-for-dependency-updates
+
+version: 2
+updates:
+  - package-ecosystem: "pip" # See documentation for possible values
+    directory: "/docs/sphinx" # Location of package manifests
+    open-pull-requests-limit: 10
+    schedule:
+      interval: "daily"
diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000000..7af066c82d
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,58 @@
+# Compiled Object files
+*.slo
+*.lo
+*.o
+*.obj
+
+# Precompiled Headers
+*.gch
+*.pch
+*.ipch
+
+# Compiled Dynamic libraries
+*.so
+*.dylib
+*.dll
+
+# Fortran module files
+*.mod
+
+# Compiled Static libraries
+*.lai
+*.la
+*.a
+*.lib
+
+# Executables
+*.exe
+*.out
+*.app
+
+# vim tags
+tags
+.tags
+.*.swp
+
+# Editors
+.vscode
+
+# build-in-source directory
+build*
+
+# emacs temporary/backup files
+.\#*
+\#*\#
+*~
+
+# GDB temporary files
+.gdb_history
+install.dir*
+
+# documentation artifacts
+_build/
+_images/
+_static/
+_templates/
+_toc.yml
+docBin/
+_doxygen/
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
new file mode 100644
index 0000000000..d6700ae05b
--- /dev/null
+++ b/.pre-commit-config.yaml
@@ -0,0 +1,14 @@
+repos:
+-   repo: local
+    hooks:
+    -   id: clang-format
+        name: clang-format
+        entry: clang-format-12 -i --style=file
+        language: system
+        types_or: [c++, inc]
+    -   id: copyright-year-checker
+        name: copyright-year-checker
+        entry: script/check_copyright_year.sh
+        verbose: false
+        language: script
+        types: [c++]
diff --git a/.readthedocs.yaml b/.readthedocs.yaml
new file mode 100644
index 0000000000..5f50df2525
--- /dev/null
+++ b/.readthedocs.yaml
@@ -0,0 +1,18 @@
+# Read the Docs configuration file
+# See https://docs.readthedocs.io/en/stable/config-file/v2.html for details
+
+version: 2
+
+build:
+   os: ubuntu-22.04
+   tools:
+      python: "3.8"
+
+sphinx:
+   configuration: docs/conf.py
+
+formats: [htmlzip, pdf, epub]
+
+python:
+   install:
+   - requirements: docs/sphinx/requirements.txt
diff --git a/CHANGELOG.md b/CHANGELOG.md
new file mode 100644
index 0000000000..31f129b581
--- /dev/null
+++ b/CHANGELOG.md
@@ -0,0 +1,32 @@
+# Change Log for Composable Kernel
+
+Full documentation for Composable Kernel is not yet available.
+
+## CK 0.2.0 for ROCm 5.5.0
+
+### Fixed
+- Fixed a bug in 6-dimensional kernels (#555).
+- Fixed grouped ConvBwdWeight test case failure (#524).
+
+### Optimizations
+- Improve proformance of normalization kernel
+
+### Added
+- Added new cmake flag "DL_KERNELS" must be set to "ON" in order to build the gemm_dl and batched_gemm_multi_d_dl instances.
+- Added new cmake flag "DTYPES" which could be set to any subset of "fp64;fp32;fp16;fp8;bf16;int8" to build instance of select data types.
+- Added new cmake flag "INSTANCES_ONLY" which will only build CK library and instances without the tests, examples, or profiler.
+- Added new feature: if GPU_TARGETS is not set on cmake command line, CK will be built for all targets supported by compiler.
+- Added support on MI300A/MI300X.
+- Added support on NAVI3x.
+- Added user tutorial (#563).
+- Added more instances for irregular GEMM sizes (#560).
+- Added inter-wave consumer-producer programming model for GEMM kernels (#310).
+- Added multi-D GEMM client APIs (#534).
+- Added multi-embeddings support (#542).
+- Added Navi3x blockwise GEMM and real GEMM support (#541).
+- Added Navi grouped ConvBwdWeight support (#505).
+- Added MaxPool, AvgPool forward (#815).
+- Added MaxPool backward (#750).
+
+### Changed
+- Changed ...
diff --git a/CITATION.cff b/CITATION.cff
new file mode 100644
index 0000000000..d35fe9e587
--- /dev/null
+++ b/CITATION.cff
@@ -0,0 +1,67 @@
+cff-version: 1.2.0
+title: Composable Kernel
+message: If you use this software, please cite using the following metadata.
+type: software
+authors:
+  - given-names: Chao
+    family-names: Liu
+    email: chao.liu2@amd.com
+    affiliation: AMD
+  - given-names: Jing
+    family-names: Zhang
+    email: jing.zhang3@amd.com
+    affiliation: AMD
+  - given-names: Letao
+    family-names: Qin
+    email: letao.qin@amd.com
+    affiliation: AMD
+  - given-names: Qianfeng
+    family-names: Zhang
+    email: qianfeng.zhang@amd.com
+    affiliation: AMD
+  - given-names: Liang
+    family-names: Huang
+    email: carlus.huang@amd.com
+    affiliation: AMD
+  - given-names: Shaojie
+    family-names: Wang
+    email: shaojie.wang@amd.com
+    affiliation: AMD
+  - given-names: Anthony
+    family-names: Chang
+    email: antc@amd.com
+    affiliation: AMD
+  - given-names: Chunyu
+    family-names: Lai
+    email: chunyu.lai@amd.com
+    affiliation: AMD
+  - given-names: Illia
+    family-names: Silin
+    email: illia.silin@amd.com
+    affiliation: AMD
+  - given-names: Adam
+    family-names: Osewski
+    email: adam.osewski@amd.com
+    affiliation: AMD
+  - given-names: Poyen
+    family-names: Chen
+    email: poyen.chen@amd.com
+    affiliation: AMD
+  - given-names: Rosty
+    family-names: Geyyer
+    email: rosty.geyyer@amd.com
+    affiliation: AMD
+  - given-names: Hanwen
+    family-names: Chen
+  - given-names: Tejash
+    family-names: Shah
+  - given-names: Xiaoyan
+    family-names: Zhou
+  - given-names: Jianfeng
+    family-names: Yan
+repository-code: 'https://github.com/ROCmSoftwarePlatform/composable_kernel'
+abstract: Composable Kernel (CK) library aims to provide a programming model for writing performance critical kernels for Machine Learning workloads across multiple architectures including GPUs, CPUs, etc, through general purpose kernel progarmming languages, like HIP C++.
+keywords:
+  - 'CK, Composable Kernel, Tensor Coordinate Transformation'
+license: MIT
+license-url: https://github.com/ROCmSoftwarePlatform/composable_kernel/blob/7fc3ed761aa35709d87c8fbbe41dd368648b3541/LICENSE
diff --git a/CMakeLists.txt b/CMakeLists.txt
new file mode 100644
index 0000000000..4ca54d847d
--- /dev/null
+++ b/CMakeLists.txt
@@ -0,0 +1,467 @@
+cmake_minimum_required(VERSION 3.14)
+
+set(version 1.1.0)
+# Check support for CUDA/HIP in Cmake
+project(composable_kernel VERSION ${version})
+
+list(APPEND CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}/cmake")
+
+if (DTYPES)
+    add_definitions(-DDTYPES)
+    if (DTYPES MATCHES "int8")
+        add_definitions(-DCK_ENABLE_INT8)
+        set(CK_ENABLE_INT8 "ON")
+    endif()
+    if (DTYPES MATCHES "fp8")
+        add_definitions(-DCK_ENABLE_FP8)
+        set(CK_ENABLE_FP8 "ON")
+    endif()
+    if (DTYPES MATCHES "fp16")
+        add_definitions(-DCK_ENABLE_FP16)
+        set(CK_ENABLE_FP16 "ON")
+    endif()
+    if (DTYPES MATCHES "fp32")
+        add_definitions(-DCK_ENABLE_FP32)
+        set(CK_ENABLE_FP32 "ON")
+    endif()
+    if (DTYPES MATCHES "fp64")
+        add_definitions(-DCK_ENABLE_FP64)
+        set(CK_ENABLE_FP64 "ON")
+    endif()
+    if (DTYPES MATCHES "bf16")
+        add_definitions(-DCK_ENABLE_BF16)
+        set(CK_ENABLE_BF16 "ON")
+    endif()
+    message("DTYPES macro set to ${DTYPES}")
+else()
+    add_definitions(-DCK_ENABLE_INT8 -DCK_ENABLE_FP8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_BF16)
+    set(CK_ENABLE_ALL_DTYPES "ON")
+endif()
+
+if(DL_KERNELS)
+    add_definitions(-DDL_KERNELS)
+    set(CK_ENABLE_DL_KERNELS "ON")
+endif()
+
+if(INSTANCES_ONLY)
+    add_definitions(-DINSTANCES_ONLY)
+    set(CK_ENABLE_INSTANCES_ONLY "ON")
+endif()
+
+# CK config file to record supported datatypes, etc.
+configure_file("${PROJECT_SOURCE_DIR}/include/ck/config.h.in" "${PROJECT_BINARY_DIR}/include/ck/config.h")
+
+# CK version file to record release version as well as git commit hash
+find_package(Git REQUIRED)
+execute_process(COMMAND "${GIT_EXECUTABLE}" rev-parse HEAD OUTPUT_VARIABLE COMMIT_ID OUTPUT_STRIP_TRAILING_WHITESPACE)
+configure_file("${PROJECT_SOURCE_DIR}/include/ck/version.h.in" "${PROJECT_BINARY_DIR}/include/ck/version.h")
+
+enable_testing()
+
+set(ROCM_SYMLINK_LIBS OFF)
+find_package(ROCM REQUIRED PATHS /opt/rocm)
+
+include(ROCMInstallTargets)
+include(ROCMPackageConfigHelpers)
+include(ROCMSetupVersion)
+include(ROCMInstallSymlinks)
+include(ROCMCreatePackage)
+include(CheckCXXCompilerFlag)
+include(ROCMCheckTargetIds)
+include(TargetFlags)
+
+rocm_setup_version(VERSION ${version})
+
+list(APPEND CMAKE_PREFIX_PATH ${CMAKE_INSTALL_PREFIX} ${CMAKE_INSTALL_PREFIX}/llvm ${CMAKE_INSTALL_PREFIX}/hip /opt/rocm /opt/rocm/llvm /opt/rocm/hip)
+
+message("GPU_TARGETS= ${GPU_TARGETS}")
+
+message("checking which targets are supported")
+#This is the list of targets to be used in case GPU_TARGETS is not set on command line
+#These targets will be filtered and only supported ones will be used
+#Setting GPU_TARGETS on command line will override this list
+if(NOT PROFILER_ONLY)
+    rocm_check_target_ids(DEFAULT_GPU_TARGETS
+        TARGETS "gfx900;gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102")
+else()
+    add_definitions(-DPROFILER_ONLY)
+    if(GPU_TARGETS)
+        message(FATAL_ERROR "For PROFILE_ONLY build, please do not set GPU_TARGETS, use GPU_ARCH = gfx9, gfx10, or gfx11")
+    endif()
+    if(GPU_ARCH MATCHES "gfx9")
+        rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx900;gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942")
+    elseif(GPU_ARCH MATCHES "gfx10")
+        rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1030")
+    elseif(GPU_ARCH MATCHES "gfx11")
+        rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1100;gfx1101;gfx1102")
+    else()
+        message(FATAL_ERROR "For PROFILE_ONLY build, please specify GPU_ARCH as gfx9, gfx10, or gfx11")
+    endif()
+endif()
+
+message("Supported GPU_TARGETS= ${DEFAULT_GPU_TARGETS}")
+
+set(AMDGPU_TARGETS "${DEFAULT_GPU_TARGETS}" CACHE STRING " ")
+
+if(GPU_TARGETS)
+    message("Building CK for the following targets: ${GPU_TARGETS}")
+else()
+    message("Building CK for the following targets: ${AMDGPU_TARGETS}")
+endif()
+find_package(hip)
+# No assumption that HIP kernels are launched with uniform block size for backward compatibility
+# SWDEV-413293 and https://reviews.llvm.org/D155213
+math(EXPR hip_VERSION_FLAT "(${hip_VERSION_MAJOR} * 1000 + ${hip_VERSION_MINOR}) * 100000 + ${hip_VERSION_PATCH}")
+message("hip_version_flat=${hip_VERSION_FLAT}")
+if(${hip_VERSION_FLAT} GREATER 500723302)
+   message("Adding the fno-offload-uniform-block compiler flag")
+   add_compile_options(-fno-offload-uniform-block)
+endif()
+
+option(USE_BITINT_EXTENSION_INT4, "Whether to enable clang's BitInt extension to provide int4 data type." OFF)
+option(USE_OPT_NAVI3X, "Whether to enable LDS cumode and Wavefront32 mode for NAVI3X silicons." OFF)
+
+if(USE_BITINT_EXTENSION_INT4)
+    add_compile_definitions(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    add_compile_options(-Wno-bit-int-extension)
+    message("CK compiled with USE_BITINT_EXTENSION_INT4 set to ${USE_BITINT_EXTENSION_INT4}")
+endif()
+
+if(USE_OPT_NAVI3X)
+    add_compile_options(-mcumode)
+    add_compile_options(-mno-wavefrontsize64)
+    message("CK compiled with USE_OPT_NAVI3X set to ${USE_OPT_NAVI3X}")
+endif()
+
+## Threads
+set(THREADS_PREFER_PTHREAD_FLAG ON)
+find_package(Threads REQUIRED)
+link_libraries(Threads::Threads)
+
+## C++
+enable_language(CXX)
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_CXX_EXTENSIONS OFF)
+message("CMAKE_CXX_COMPILER_ID: ${CMAKE_CXX_COMPILER_ID}")
+
+## OpenMP
+if(CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+	# workaround issue hipcc in rocm3.5 cannot find openmp
+	set(OpenMP_CXX "${CMAKE_CXX_COMPILER}")
+	set(OpenMP_CXX_FLAGS "-fopenmp=libomp -Wno-unused-command-line-argument")
+	set(OpenMP_CXX_LIB_NAMES "libomp" "libgomp" "libiomp5")
+	set(OpenMP_libomp_LIBRARY ${OpenMP_CXX_LIB_NAMES})
+	set(OpenMP_libgomp_LIBRARY ${OpenMP_CXX_LIB_NAMES})
+	set(OpenMP_libiomp5_LIBRARY ${OpenMP_CXX_LIB_NAMES})
+else()
+	find_package(OpenMP REQUIRED)
+endif()
+
+message("OpenMP_CXX_LIB_NAMES: ${OpenMP_CXX_LIB_NAMES}")
+message("OpenMP_gomp_LIBRARY: ${OpenMP_gomp_LIBRARY}")
+message("OpenMP_pthread_LIBRARY: ${OpenMP_pthread_LIBRARY}")
+message("OpenMP_CXX_FLAGS: ${OpenMP_CXX_FLAGS}")
+
+link_libraries(${OpenMP_gomp_LIBRARY})
+link_libraries(${OpenMP_pthread_LIBRARY})
+
+## HIP
+find_package(HIP REQUIRED)
+# Override HIP version in config.h, if necessary.
+# The variables set by find_package() can't be overwritten,
+# therefore let's use intermediate variables.
+set(CK_HIP_VERSION_MAJOR "${HIP_VERSION_MAJOR}")
+set(CK_HIP_VERSION_MINOR "${HIP_VERSION_MINOR}")
+set(CK_HIP_VERSION_PATCH "${HIP_VERSION_PATCH}")
+if( DEFINED CK_OVERRIDE_HIP_VERSION_MAJOR )
+    set(CK_HIP_VERSION_MAJOR "${CK_OVERRIDE_HIP_VERSION_MAJOR}")
+    message(STATUS "CK_HIP_VERSION_MAJOR overriden with ${CK_OVERRIDE_HIP_VERSION_MAJOR}")
+endif()
+if( DEFINED CK_OVERRIDE_HIP_VERSION_MINOR )
+    set(CK_HIP_VERSION_MINOR "${CK_OVERRIDE_HIP_VERSION_MINOR}")
+    message(STATUS "CK_HIP_VERSION_MINOR overriden with ${CK_OVERRIDE_HIP_VERSION_MINOR}")
+endif()
+if( DEFINED CK_OVERRIDE_HIP_VERSION_PATCH )
+    set(CK_HIP_VERSION_PATCH "${CK_OVERRIDE_HIP_VERSION_PATCH}")
+    message(STATUS "CK_HIP_VERSION_PATCH overriden with ${CK_OVERRIDE_HIP_VERSION_PATCH}")
+endif()
+message(STATUS "Build with HIP ${HIP_VERSION}")
+link_libraries(hip::device)
+add_compile_definitions(__HIP_PLATFORM_HCC__=1)
+
+## tidy
+include(EnableCompilerWarnings)
+set(CK_TIDY_ERRORS ERRORS * -readability-inconsistent-declaration-parameter-name)
+if(CMAKE_CXX_COMPILER MATCHES ".*hcc" OR CMAKE_CXX_COMPILER MATCHES ".*clang\\+\\+")
+    set(CK_TIDY_CHECKS -modernize-use-override -readability-non-const-parameter)
+# Enable tidy on hip
+elseif(CK_BACKEND STREQUAL "HIP" OR CK_BACKEND STREQUAL "HIPNOGPU")
+    set(CK_TIDY_ERRORS ALL)
+endif()
+
+
+include(ClangTidy)
+enable_clang_tidy(
+    CHECKS
+        *
+        -abseil-*
+        -android-cloexec-fopen
+        # Yea we shouldn't be using rand()
+        -cert-msc30-c
+        -bugprone-exception-escape
+        -bugprone-macro-parentheses
+        -cert-env33-c
+        -cert-msc32-c
+        -cert-msc50-cpp
+        -cert-msc51-cpp
+        -cert-dcl37-c
+        -cert-dcl51-cpp
+        -clang-analyzer-alpha.core.CastToStruct
+        -clang-analyzer-optin.performance.Padding
+        -clang-diagnostic-deprecated-declarations
+        -clang-diagnostic-extern-c-compat
+        -clang-diagnostic-unused-command-line-argument
+        -cppcoreguidelines-avoid-c-arrays
+        -cppcoreguidelines-avoid-magic-numbers
+        -cppcoreguidelines-explicit-virtual-functions
+        -cppcoreguidelines-init-variables
+        -cppcoreguidelines-macro-usage
+        -cppcoreguidelines-non-private-member-variables-in-classes
+        -cppcoreguidelines-pro-bounds-array-to-pointer-decay
+        -cppcoreguidelines-pro-bounds-constant-array-index
+        -cppcoreguidelines-pro-bounds-pointer-arithmetic
+        -cppcoreguidelines-pro-type-member-init
+        -cppcoreguidelines-pro-type-reinterpret-cast
+        -cppcoreguidelines-pro-type-union-access
+        -cppcoreguidelines-pro-type-vararg
+        -cppcoreguidelines-special-member-functions
+        -fuchsia-*
+        -google-explicit-constructor
+        -google-readability-braces-around-statements
+        -google-readability-todo
+        -google-runtime-int
+        -google-runtime-references
+        -hicpp-vararg
+        -hicpp-braces-around-statements
+        -hicpp-explicit-conversions
+        -hicpp-named-parameter
+        -hicpp-no-array-decay
+        # We really shouldn't use bitwise operators with signed integers, but
+        # opencl leaves us no choice
+        -hicpp-avoid-c-arrays
+        -hicpp-signed-bitwise
+        -hicpp-special-member-functions
+        -hicpp-uppercase-literal-suffix
+        -hicpp-use-auto
+        -hicpp-use-equals-default
+        -hicpp-use-override
+        -llvm-header-guard
+        -llvm-include-order
+        #-llvmlibc-*
+        -llvmlibc-restrict-system-libc-headers
+        -llvmlibc-callee-namespace
+        -llvmlibc-implementation-in-namespace
+        -llvm-else-after-return
+        -llvm-qualified-auto
+        -misc-misplaced-const
+        -misc-non-private-member-variables-in-classes
+        -misc-no-recursion
+        -modernize-avoid-bind
+        -modernize-avoid-c-arrays
+        -modernize-pass-by-value
+        -modernize-use-auto
+        -modernize-use-default-member-init
+        -modernize-use-equals-default
+        -modernize-use-trailing-return-type
+        -modernize-use-transparent-functors
+        -performance-unnecessary-value-param
+        -readability-braces-around-statements
+        -readability-else-after-return
+        # we are not ready to use it, but very useful
+        -readability-function-cognitive-complexity
+        -readability-isolate-declaration
+        -readability-magic-numbers
+        -readability-named-parameter
+        -readability-uppercase-literal-suffix
+        -readability-convert-member-functions-to-static
+        -readability-qualified-auto
+        -readability-redundant-string-init
+        # too many narrowing conversions in our code
+        -bugprone-narrowing-conversions
+        -cppcoreguidelines-narrowing-conversions
+        -altera-struct-pack-align
+        -cppcoreguidelines-prefer-member-initializer
+        ${CK_TIDY_CHECKS}
+        ${CK_TIDY_ERRORS}
+    HEADER_FILTER
+        "\.hpp$"
+    EXTRA_ARGS
+        -DCK_USE_CLANG_TIDY
+)
+
+include(CppCheck)
+enable_cppcheck(
+    CHECKS
+        warning
+        style
+        performance
+        portability
+    SUPPRESS
+        ConfigurationNotChecked
+        constStatement
+        duplicateCondition
+        noExplicitConstructor
+        passedByValue
+        preprocessorErrorDirective
+        shadowVariable
+        unusedFunction
+        unusedPrivateFunction
+        unusedStructMember
+        unmatchedSuppression
+    FORCE
+    SOURCES
+        library/src
+    INCLUDE
+        ${CMAKE_CURRENT_SOURCE_DIR}/include
+        ${CMAKE_CURRENT_BINARY_DIR}/include
+        ${CMAKE_CURRENT_SOURCE_DIR}/library/include
+    DEFINE
+        CPPCHECK=1
+        __linux__=1
+)
+
+set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/lib)
+set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/lib)
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/bin)
+
+# set CK project include directories
+include_directories(BEFORE
+    ${PROJECT_BINARY_DIR}/include
+    ${PROJECT_SOURCE_DIR}/include
+    ${PROJECT_SOURCE_DIR}/library/include
+    ${HIP_INCLUDE_DIRS}
+)
+
+SET(BUILD_DEV ON CACHE BOOL "BUILD_DEV")
+if(BUILD_DEV)
+    add_compile_options(-Werror)
+    add_compile_options(-Weverything)
+endif()
+message("CMAKE_CXX_FLAGS: ${CMAKE_CXX_FLAGS}")
+
+add_custom_target(check COMMAND ${CMAKE_CTEST_COMMAND} --output-on-failure -C ${CMAKE_CFG_INTDIR})
+
+file(GLOB_RECURSE INSTANCE_FILES "${PROJECT_SOURCE_DIR}/*/device_*_instance.cpp")
+file(GLOB dir_list RELATIVE ${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu ${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu/*)
+set(CK_DEVICE_INSTANCES)
+FOREACH(subdir_path ${dir_list})
+set(target_dir)
+IF(IS_DIRECTORY "${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu/${subdir_path}")
+    set(cmake_instance)
+    file(READ "${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu/${subdir_path}/CMakeLists.txt" cmake_instance)
+    set(add_inst 0)
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp8\" " AND DTYPES MATCHES "fp8")
+        #message("fp8 instance found!")
+        set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp16\"" AND DTYPES MATCHES "fp16")
+        #message("fp16 instance found!")
+        set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp32\"" AND DTYPES MATCHES "fp32")
+        #message("fp32 instance found!")
+        set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp64\"" AND DTYPES MATCHES "fp64")
+        #message("fp64 instance found!")
+        set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"bf16\"" AND DTYPES MATCHES "bf16")
+        #message("bf16 instance found!")
+        set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"int8\"" AND DTYPES MATCHES "int8")
+        #message("int8 instance found!")
+        set(add_inst 1)
+    endif()
+    if(NOT "${cmake_instance}" MATCHES "DTYPES")
+        #message("instance should be built for all types!")
+        set(add_inst 1)
+    endif()
+    if(add_inst EQUAL 1 OR NOT DEFINED DTYPES)
+        list(APPEND CK_DEVICE_INSTANCES device_${subdir_path}_instance)
+    endif()
+ENDIF()
+ENDFOREACH()
+
+add_custom_target(instances DEPENDS utility;${CK_DEVICE_INSTANCES}  SOURCES ${INSTANCE_FILES})
+add_subdirectory(library)
+
+if(NOT DEFINED INSTANCES_ONLY)
+ if(NOT DEFINED PROFILER_ONLY)
+   rocm_package_setup_component(tests
+        LIBRARY_NAME composablekernel
+        PACKAGE_NAME tests # Prevent -static suffix on package name
+   )
+
+   rocm_package_setup_component(examples
+        LIBRARY_NAME composablekernel
+        PACKAGE_NAME examples
+   )
+   add_subdirectory(example)
+   add_subdirectory(test)
+
+   rocm_package_setup_component(profiler
+        LIBRARY_NAME composablekernel
+        PACKAGE_NAME ckProfiler
+   )
+   add_subdirectory(profiler)
+  else()
+    #When building PROFILER_ONLY, label the package with GPU_ARCH
+    rocm_package_setup_component(profiler
+       LIBRARY_NAME composablekernel
+       PACKAGE_NAME ckProfiler_${GPU_ARCH}
+    )
+    add_subdirectory(profiler)
+  endif()
+endif()
+
+#Create an interface target for the include only files and call it "composablekernels"
+include(CMakePackageConfigHelpers)
+
+write_basic_package_version_file(
+    "${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfigVersion.cmake"
+    VERSION "${version}"
+    COMPATIBILITY AnyNewerVersion
+)
+
+configure_package_config_file(${CMAKE_CURRENT_SOURCE_DIR}/Config.cmake.in
+    "${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfig.cmake"
+    INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
+    NO_CHECK_REQUIRED_COMPONENTS_MACRO
+)
+
+rocm_install(FILES
+    "${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfig.cmake"
+    "${CMAKE_CURRENT_BINARY_DIR}/composable_kernelConfigVersion.cmake"
+    DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
+)
+
+# Install CK version and configuration files
+rocm_install(FILES
+    ${PROJECT_BINARY_DIR}/include/ck/version.h
+    ${PROJECT_BINARY_DIR}/include/ck/config.h
+    DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/ck/
+)
+
+set(CPACK_RESOURCE_FILE_LICENSE "${CMAKE_CURRENT_SOURCE_DIR}/LICENSE")
+set(CPACK_RPM_PACKAGE_LICENSE "MIT")
+
+rocm_create_package(
+    NAME composablekernel
+    DESCRIPTION "High Performance Composable Kernel for AMD GPUs"
+    MAINTAINER "MIOpen Kernels Dev Team <dl.MIOpen@amd.com>"
+    LDCONFIG
+    HEADER_ONLY
+)
diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.md
new file mode 100644
index 0000000000..cdce5a4630
--- /dev/null
+++ b/CONTRIBUTORS.md
@@ -0,0 +1,33 @@
+# Composable Kernel Developers and Contributors
+
+This is the list of developers and contributors to Composable Kernel library
+
+
+## Developers
+[Chao Liu](https://github.com/asroy), [Jing Zhang](https://github.com/zjing14), 2018-2023
+
+[Letao Qin](https://github.com/ltqin), [Qianfeng Zhang](https://github.com/qianfengz), [Liang Huang](https://github.com/carlushuang), [Shaojie Wang](https://github.com/shaojiewang), 2019-2023
+
+[Anthony Chang](https://github.com/rosenrodt), [Chunyu Lai](https://github.com/rocking5566), [Illia Silin](https://github.com/illsilin), [Adam Osewski](https://github.com/aosewski), [Poyen Chen](https://github.com/poyenc), [Rosty Geyyer](https://github.com/geyyer), [Astha Rai](https://github.com/arai713), [Shi YanXing](https://github.com/Yanxing-Shi), 2022-2023
+
+[Hari Sadasivan](https://github.com/hsadasiv), [Bartlomiej Kocot](https://github.com/bartekxk), [Bartlomiej Wroblewski](https://github.com/bwroblew), 2023
+
+Hanwen Chang, 2019-2021,
+
+Tejash Shah, 2019-2020
+
+Xiaoyan Zhou, 2020
+
+[Jianfeng Yan](https://github.com/j4yan), 2021-2022
+
+
+## Product Manager
+[Jun Liu](https://github.com/junliume)
+
+
+## Contributors
+[Dan Yao](https://github.com/danyao12), [Guangzhao Lu](https://github.com/guangzlu), [Raman Jana](https://github.com/ramjana), [Jehandad Khan](https://github.com/JehandadKhan), [Wen-Heng (Jack) Chung](https://github.com/whchung)
+
+
+## Acknowledgement
+CK team works closely with Meta [AITemplate](https://github.com/facebookincubator/AITemplate) team ([Bing Xu](https://github.com/antinucleon), [Hao Lu](https://github.com/hlu1), [Ying Zhang](https://github.com/ipiszy), etc). Most of the lucrative graph optimization opportunities in ML models were identified by AITemplate team, and we also co-designed many high performance fused kernels for AMD GPUs. Without this collaboration, CK would not reach its current potential.
diff --git a/Config.cmake.in b/Config.cmake.in
new file mode 100644
index 0000000000..02978cd4dd
--- /dev/null
+++ b/Config.cmake.in
@@ -0,0 +1,11 @@
+@PACKAGE_INIT@
+
+set(_composable_kernel_supported_components device_operations utility)
+
+foreach(_comp ${composable_kernel_FIND_COMPONENTS})
+	if(NOT _comp IN_LIST _composable_kernel_supported_components)
+		set(composable_kernel_FOUND False)
+		set(composable_kernel_NOT_FOUND_MESSAGE "Unsupported component: ${_comp}")
+	endif()
+	include("${CMAKE_CURRENT_LIST_DIR}/composable_kernel${_comp}Targets.cmake")
+endforeach()
diff --git a/Dockerfile b/Dockerfile
new file mode 100644
index 0000000000..e479268f48
--- /dev/null
+++ b/Dockerfile
@@ -0,0 +1,139 @@
+FROM ubuntu:20.04
+ARG DEBIAN_FRONTEND=noninteractive
+ARG ROCMVERSION=5.6
+ARG compiler_version=""
+ARG compiler_commit=""
+
+RUN set -xe
+
+ARG DEB_ROCM_REPO=http://repo.radeon.com/rocm/apt/.apt_$ROCMVERSION/
+RUN useradd -rm -d /home/jenkins -s /bin/bash -u 1004 jenkins
+# Add rocm repository
+RUN chmod 1777 /tmp
+RUN apt-get update
+RUN apt-get install -y --allow-unauthenticated apt-utils wget gnupg2 curl
+
+ENV APT_KEY_DONT_WARN_ON_DANGEROUS_USAGE=DontWarn
+RUN curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg
+
+RUN wget https://repo.radeon.com/amdgpu-install/5.6/ubuntu/focal/amdgpu-install_5.6.50600-1_all.deb  --no-check-certificate
+RUN apt-get update && \
+DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
+    ./amdgpu-install_5.6.50600-1_all.deb
+
+RUN if [ "$ROCMVERSION" != "5.7" ]; then \
+        wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \
+        sh -c "echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] $DEB_ROCM_REPO focal main > /etc/apt/sources.list.d/rocm.list" && \
+        sh -c 'echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] https://repo.radeon.com/amdgpu/$ROCMVERSION/ubuntu focal main > /etc/apt/sources.list.d/amdgpu.list'; \
+    elif [ "$ROCMVERSION" = "5.7" ] && [ "$compiler_version" = "" ] || [ "$compiler_version" = "amd-stg-open" ]; then \
+        sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_5.7-20.04-1_all.deb" && \
+        apt update && apt-get install -y ./amdgpu-install-internal_5.7-20.04-1_all.deb && \
+        amdgpu-repo --amdgpu-build=1609671 --rocm-build=compute-rocm-npi-mi300/1354; \
+    elif [ "$ROCMVERSION" = "5.7" ] && [ "$compiler_version" = "rc1" ]; then \
+        sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_5.7-20.04-1_all.deb" && \
+        apt update && apt-get install -y ./amdgpu-install-internal_5.7-20.04-1_all.deb && \
+        sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 5.7 rel-19 > /etc/apt/sources.list.d/rocm-build.list' && \
+        amdgpu-repo --amdgpu-build=1637781; \
+    fi
+
+RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list"
+RUN amdgpu-install -y --usecase=rocm --no-dkms
+
+# Install dependencies
+RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated \
+    build-essential \
+    ccache \
+    cmake \
+    git \
+    hip-rocclr \
+    jq \
+    libelf-dev \
+    libncurses5-dev \
+    libnuma-dev \
+    libpthread-stubs0-dev \
+    llvm-amdgpu \
+    pkg-config \
+    python \
+    python3 \
+    python3-dev \
+    python3-pip \
+    sshpass \
+    software-properties-common \
+    vim \
+    nano \
+    zlib1g-dev \
+    openssh-server \
+    clang-format-12 \
+    kmod && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+#Install latest version of cmake
+RUN wget -qO /usr/local/bin/ninja.gz https://github.com/ninja-build/ninja/releases/latest/download/ninja-linux.zip
+RUN gunzip /usr/local/bin/ninja.gz
+RUN chmod a+x /usr/local/bin/ninja
+RUN git clone https://github.com/nico/ninjatracing.git
+RUN apt purge --auto-remove -y cmake
+RUN apt update
+RUN apt install -y software-properties-common lsb-release
+RUN apt clean all
+RUN wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | gpg --dearmor - | tee /etc/apt/trusted.gpg.d/kitware.gpg >/dev/null
+RUN apt-add-repository "deb https://apt.kitware.com/ubuntu/ $(lsb_release -cs) main"
+RUN apt install -y kitware-archive-keyring
+RUN rm /etc/apt/trusted.gpg.d/kitware.gpg
+RUN apt install -y cmake
+
+# Setup ubsan environment to printstacktrace
+RUN ln -s /usr/bin/llvm-symbolizer-3.8 /usr/local/bin/llvm-symbolizer
+ENV UBSAN_OPTIONS=print_stacktrace=1
+
+# Install an init system
+RUN wget https://github.com/Yelp/dumb-init/releases/download/v1.2.0/dumb-init_1.2.0_amd64.deb
+RUN dpkg -i dumb-init_*.deb && rm dumb-init_*.deb
+
+ARG PREFIX=/opt/rocm
+# Install packages for processing the performance results
+RUN pip3 install --upgrade pip
+RUN pip3 install sqlalchemy==1.4.46
+RUN pip3 install pymysql
+RUN pip3 install pandas
+RUN pip3 install setuptools-rust
+RUN pip3 install sshtunnel
+# Setup ubsan environment to printstacktrace
+ENV UBSAN_OPTIONS=print_stacktrace=1
+
+ENV LC_ALL=C.UTF-8
+ENV LANG=C.UTF-8
+RUN groupadd -f render
+
+# Install the new rocm-cmake version
+RUN git clone -b master https://github.com/RadeonOpenCompute/rocm-cmake.git  && \
+  cd rocm-cmake && mkdir build && cd build && \
+  cmake  .. && cmake --build . && cmake --build . --target install
+
+WORKDIR /
+
+ENV compiler_version=$compiler_version
+ENV compiler_commit=$compiler_commit
+RUN sh -c "echo compiler version = '$compiler_version'"
+RUN sh -c "echo compiler commit = '$compiler_commit'"
+
+RUN if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" = "" ]; then \
+        git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \
+        cd llvm-project && mkdir build && cd build && \
+        cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \
+        make -j 8 ; \
+    else echo "using the release compiler"; \
+    fi
+
+RUN if [ "$compiler_version" = "amd-stg-open" ] && [ "$compiler_commit" != "" ]; then \
+        git clone -b "$compiler_version" https://github.com/RadeonOpenCompute/llvm-project.git && \
+        cd llvm-project && git checkout "$compiler_commit" && echo "checking out commit $compiler_commit" && mkdir build && cd build && \
+        cmake -DCMAKE_INSTALL_PREFIX=/opt/rocm/llvm -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=1 -DLLVM_TARGETS_TO_BUILD="AMDGPU;X86" -DLLVM_ENABLE_PROJECTS="clang;lld;compiler-rt" ../llvm && \
+        make -j 8 ; \
+    else echo "using the release compiler"; \
+    fi
+
+
+#ENV HIP_CLANG_PATH='/llvm-project/build/bin'
+#RUN sh -c "echo HIP_CLANG_PATH = '$HIP_CLANG_PATH'"
diff --git a/Jenkinsfile b/Jenkinsfile
new file mode 100644
index 0000000000..668d9a613b
--- /dev/null
+++ b/Jenkinsfile
@@ -0,0 +1,819 @@
+def rocmnode(name) {
+    return 'rocmtest && miopen && ' + name
+}
+
+def show_node_info() {
+    sh """
+        echo "NODE_NAME = \$NODE_NAME"
+        lsb_release -sd
+        uname -r
+        ls /opt/ -la
+    """
+}
+
+def nthreads() {
+    def nproc = sh(returnStdout: true, script: 'nproc')
+    echo "Number of cores: ${nproc}"
+    def n = nproc.toInteger()
+    if (n > 32){
+        n /= 2
+    }
+    if (n > 64){
+        n = 64
+    }
+    echo "Number of threads used for building: ${n}"
+    return n
+}
+
+def runShell(String command){
+    def responseCode = sh returnStatus: true, script: "${command} > tmp.txt"
+    def output = readFile(file: "tmp.txt")
+    return (output != "")
+}
+
+def getDockerImageName(){
+    def img
+    if (params.ROCMVERSION != "5.7"){
+       if (params.COMPILER_VERSION == "") {
+           img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}"
+       }
+       else{
+          if (params.COMPILER_COMMIT == ""){
+             img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}"
+          }
+          else{
+             def commit = "${params.COMPILER_COMMIT}"[0..6]
+             img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}_${commit}"
+          }
+       }
+    }
+    else{
+       if (params.COMPILER_VERSION == "") {
+           img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_rocm${params.ROCMVERSION}"
+       }
+       else{
+          if (params.COMPILER_COMMIT == ""){
+             img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}"
+          }
+          else{
+             def commit = "${params.COMPILER_COMMIT}"[0..6]
+             img = "${env.CK_DOCKERHUB_PRIVATE}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}_${commit}"
+          }
+       }
+    }
+    return img
+}
+
+def check_host() {
+    if ("${env.CK_CCACHE}" != "null"){
+        def CCACHE_SERVER="${env.CK_CCACHE.split(':')[0]}"
+        echo "ccache server: ${CCACHE_SERVER}"
+        sh '''ping -c 1 -p 6379 "${CCACHE_SERVER}" | echo $? > tmp.txt'''
+        def output = readFile(file: "tmp.txt")
+        echo "tmp.txt contents: \$output"
+        return (output != "0")
+    }
+    else{
+        return 1
+    }
+}
+
+def build_compiler(){
+    def compiler
+    if (params.BUILD_COMPILER == "hipcc"){
+        compiler = '/opt/rocm/bin/hipcc'
+    }
+    else{
+        if (params.COMPILER_VERSION == "amd-stg-open" || params.COMPILER_COMMIT != ""){
+            compiler = "/llvm-project/build/bin/clang++"
+        }
+        else{
+            compiler = "/opt/rocm/llvm/bin/clang++"
+        }        
+    }
+    return compiler
+}
+
+def getDockerImage(Map conf=[:]){
+    env.DOCKER_BUILDKIT=1
+    def prefixpath = conf.get("prefixpath", "/opt/rocm") // prefix:/opt/rocm
+    def no_cache = conf.get("no_cache", false)
+    def dockerArgs = "--build-arg BUILDKIT_INLINE_CACHE=1 --build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
+    echo "ccache server: ${env.CK_CCACHE}"
+    if(env.CK_CCACHE)
+    {
+        if(check_host())
+        {
+            echo "FOUND CCACHE SERVER: ${env.CK_CCACHE}"
+        }
+        else 
+        {
+            echo "CCACHE SERVER: ${env.CK_CCACHE} NOT FOUND, got ${check_host} response"
+        }
+        dockerArgs = dockerArgs + " --build-arg CCACHE_SECONDARY_STORAGE='redis://${env.CK_CCACHE}' --build-arg COMPILER_LAUNCHER='ccache' "
+        env.CCACHE_DIR = """/tmp/ccache_store"""
+        env.CCACHE_SECONDARY_STORAGE="""redis://${env.CK_CCACHE}"""
+    }
+    if(no_cache)
+    {
+        dockerArgs = dockerArgs + " --no-cache "
+    }
+    echo "Docker Args: ${dockerArgs}"
+    def image = getDockerImageName()
+    //Check if image exists 
+    def retimage
+    try 
+    {
+        echo "Pulling down image: ${image}"
+        retimage = docker.image("${image}")
+        retimage.pull()
+    }
+    catch(Exception ex)
+    {
+        error "Unable to locate image: ${image}"
+    }
+    return [retimage, image]
+}
+
+def buildDocker(install_prefix){
+    show_node_info()
+    env.DOCKER_BUILDKIT=1
+    checkout scm
+    def image_name = getDockerImageName()
+    echo "Building Docker for ${image_name}"
+    def dockerArgs = "--build-arg BUILDKIT_INLINE_CACHE=1 --build-arg PREFIX=${install_prefix} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
+    echo "ccache server: ${env.CK_CCACHE}"
+    if(env.CK_CCACHE)
+    {
+        if(check_host())
+        {
+            echo "FOUND CCACHE SERVER: ${env.CK_CCACHE}"
+        }
+        else 
+        {
+            echo "CCACHE SERVER: ${env.CK_CCACHE} NOT FOUND, got ${check_host} response"
+        }
+        dockerArgs = dockerArgs + " --build-arg CCACHE_SECONDARY_STORAGE='redis://${env.CK_CCACHE}' --build-arg COMPILER_LAUNCHER='ccache' "
+        env.CCACHE_DIR = """/tmp/ccache_store"""
+        env.CCACHE_SECONDARY_STORAGE="""redis://${env.CK_CCACHE}"""
+    }
+
+    echo "Build Args: ${dockerArgs}"
+    try{
+        if(params.BUILD_DOCKER){
+            //force building the new docker if that parameter is true
+            echo "Building image: ${image_name}"
+            retimage = docker.build("${image_name}", dockerArgs + ' .')
+            retimage.push()
+        }
+        else{
+            echo "Checking for image: ${image_name}"
+            sh "docker manifest inspect --insecure ${image_name}"
+            echo "Image: ${image_name} found!! Skipping building image"
+        }
+    }
+    catch(Exception ex){
+        echo "Unable to locate image: ${image_name}. Building image now"
+        retimage = docker.build("${image_name}", dockerArgs + ' .')
+        retimage.push()
+    }
+}
+
+def cmake_build(Map conf=[:]){
+
+    def compiler = build_compiler()
+    def config_targets = conf.get("config_targets","check")
+    def debug_flags = "-g -fno-omit-frame-pointer -fsanitize=undefined -fno-sanitize-recover=undefined " + conf.get("extradebugflags", "")
+    def build_envs = "CTEST_PARALLEL_LEVEL=4 " + conf.get("build_env","")
+    def prefixpath = conf.get("prefixpath","/opt/rocm")
+    def setup_args = conf.get("setup_args","")
+
+    if (prefixpath != "/usr/local"){
+        setup_args = setup_args + " -DCMAKE_PREFIX_PATH=${prefixpath} "
+    }
+
+    def build_type_debug = (conf.get("build_type",'release') == 'debug')
+
+    //cmake_env can overwrite default CXX variables.
+    def cmake_envs = "CXX=${compiler} CXXFLAGS='-Werror' " + conf.get("cmake_ex_env","")
+
+    def package_build = (conf.get("package_build","") == "true")
+
+    if (package_build == true) {
+        config_targets = "package"
+    }
+
+    if(conf.get("build_install","") == "true")
+    {
+        config_targets = 'install ' + config_targets
+        setup_args = ' -DBUILD_DEV=On -DCMAKE_INSTALL_PREFIX=../install' + setup_args
+    } else{
+        setup_args = ' -DBUILD_DEV=On' + setup_args
+    }
+
+    if(build_type_debug){
+        setup_args = " -DCMAKE_BUILD_TYPE=debug -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}'" + setup_args
+    }else{
+        setup_args = " -DCMAKE_BUILD_TYPE=release" + setup_args
+    }
+    if(env.CK_CCACHE)
+    {
+        setup_args = " -DCMAKE_CXX_COMPILER_LAUNCHER='ccache' -DCMAKE_C_COMPILER_LAUNCHER='ccache' " + setup_args
+    }
+    echo "ccache server: ${env.CK_CCACHE}"
+
+    def pre_setup_cmd = """
+            echo \$HSA_ENABLE_SDMA
+            ulimit -c unlimited
+            rm -rf build
+            mkdir build
+            rm -rf install
+            mkdir install
+            cd build
+        """
+    def setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake ${setup_args}   .. ")
+    // reduce parallelism when compiling, clang uses too much memory
+    def nt = nthreads()
+    def build_cmd = conf.get("build_cmd", "${build_envs} dumb-init make  -j${nt} ${config_targets}")
+    def execute_cmd = conf.get("execute_cmd", "")
+
+    def cmd = conf.get("cmd", """
+            ${pre_setup_cmd}
+            ${setup_cmd}
+            ${build_cmd}
+            ${execute_cmd}
+        """)
+
+    echo cmd
+    sh cmd
+
+    // Only archive from master or develop
+    if (package_build == true && (env.BRANCH_NAME == "develop" || env.BRANCH_NAME == "master")) {
+        archiveArtifacts artifacts: "build/*.deb", allowEmptyArchive: true, fingerprint: true
+    }
+}
+
+def buildHipClangJob(Map conf=[:]){
+        show_node_info()
+
+        env.HSA_ENABLE_SDMA=0
+        checkout scm
+
+        def image = getDockerImageName() 
+        def prefixpath = conf.get("prefixpath", "/opt/rocm")
+
+        // Jenkins is complaining about the render group 
+        def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+        if (conf.get("enforce_xnack_on", false)) {
+            dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
+        }
+        def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
+        if (params.COMPILER_VERSION == "amd-stg-open" || params.COMPILER_COMMIT != ""){
+            dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
+        }
+
+        def variant = env.STAGE_NAME
+
+        def retimage
+        (retimage, image) = getDockerImage(conf)
+
+        gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
+            withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
+                timeout(time: 5, unit: 'HOURS')
+                {
+                    cmake_build(conf)
+                }
+            }
+        }
+        return retimage
+}
+
+def reboot(){
+    build job: 'reboot-slaves', propagate: false , parameters: [string(name: 'server', value: "${env.NODE_NAME}"),]
+}
+
+def buildHipClangJobAndReboot(Map conf=[:]){
+    try{
+        buildHipClangJob(conf)
+    }
+    catch(e){
+        echo "throwing error exception for the stage"
+        echo 'Exception occurred: ' + e.toString()
+        throw e
+    }
+    finally{
+        if (!conf.get("no_reboot", false)) {
+            reboot()
+        }
+    }
+}
+
+def runCKProfiler(Map conf=[:]){
+        show_node_info()
+
+        env.HSA_ENABLE_SDMA=0
+        checkout scm
+
+        def image = getDockerImageName()
+        def prefixpath = conf.get("prefixpath", "/opt/rocm")
+
+        // Jenkins is complaining about the render group 
+        def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+        if (conf.get("enforce_xnack_on", false)) {
+            dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
+        }
+        def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
+        if (params.COMPILER_VERSION == "amd-stg-open" || params.COMPILER_COMMIT != ""){
+            dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
+        }
+
+        def variant = env.STAGE_NAME
+        def retimage
+
+        gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
+            try {
+                (retimage, image) = getDockerImage(conf)
+                withDockerContainer(image: image, args: dockerOpts) {
+                    timeout(time: 5, unit: 'MINUTES'){
+                        sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo | tee clinfo.log'
+                        if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
+                            throw new Exception ("GPU not found")
+                        }
+                        else{
+                            echo "GPU is OK"
+                        }
+                    }
+                }
+            }
+            catch (org.jenkinsci.plugins.workflow.steps.FlowInterruptedException e){
+                echo "The job was cancelled or aborted"
+                throw e
+            }
+            catch(Exception ex) {
+                retimage = docker.build("${image}", dockerArgs + " --no-cache .")
+                withDockerContainer(image: image, args: dockerOpts) {
+                    timeout(time: 5, unit: 'MINUTES'){
+                        sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo | tee clinfo.log'
+                        if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
+                            throw new Exception ("GPU not found")
+                        }
+                        else{
+                            echo "GPU is OK"
+                        }
+                    }
+                }
+            }
+
+            withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
+                timeout(time: 24, unit: 'HOURS')
+                {
+                    //cmake_build(conf)
+                    //instead of building, just unstash the ckProfiler and install it
+                    sh """
+                        rm -rf build
+                        mkdir build
+                    """
+                    dir("build"){
+                        unstash 'ckProfiler.tar.gz'
+                        sh 'tar -xvf ckProfiler.tar.gz'
+                    }
+
+					dir("script"){
+                        if (params.RUN_FULL_QA){
+                            sh "./run_full_performance_tests.sh 1 QA_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME}"
+                            archiveArtifacts "perf_gemm.log"
+                            archiveArtifacts "perf_resnet50_N256.log"
+                            archiveArtifacts "perf_resnet50_N4.log"
+                            archiveArtifacts "perf_batched_gemm.log"
+                            archiveArtifacts "perf_grouped_gemm.log"
+                            archiveArtifacts "perf_conv_fwd.log"
+                            archiveArtifacts "perf_conv_bwd_data.log"
+                            archiveArtifacts "perf_gemm_bilinear.log"
+                            archiveArtifacts "perf_reduction.log"
+                            archiveArtifacts "perf_splitK_gemm_verify.log"
+                            archiveArtifacts "perf_splitK_gemm.log"
+                            archiveArtifacts "perf_onnx_gemm.log"
+                           // stash perf files to master
+                            stash name: "perf_gemm.log"
+                            stash name: "perf_resnet50_N256.log"
+                            stash name: "perf_resnet50_N4.log"
+                            stash name: "perf_batched_gemm.log"
+                            stash name: "perf_grouped_gemm.log"
+                            stash name: "perf_conv_fwd.log"
+                            stash name: "perf_conv_bwd_data.log"
+                            stash name: "perf_gemm_bilinear.log"
+                            stash name: "perf_reduction.log"
+                            stash name: "perf_splitK_gemm.log"
+                            stash name: "perf_onnx_gemm.log"
+                            //we will process results on the master node
+                        }
+                        else{
+                            sh "./run_performance_tests.sh 0 CI_${params.COMPILER_VERSION} ${env.BRANCH_NAME} ${NODE_NAME}"
+                            archiveArtifacts "perf_gemm.log"
+                            archiveArtifacts "perf_resnet50_N256.log"
+                            archiveArtifacts "perf_resnet50_N4.log"
+                            // stash perf files to master
+                            stash name: "perf_gemm.log"
+                            stash name: "perf_resnet50_N256.log"
+                            stash name: "perf_resnet50_N4.log"
+                            //we will process the results on the master node
+                        }
+					}
+                }
+            }
+        }
+        return retimage
+}
+
+def runPerfTest(Map conf=[:]){
+    try{
+        runCKProfiler(conf)
+    }
+    catch(e){
+        echo "throwing error exception in performance tests"
+        echo 'Exception occurred: ' + e.toString()
+        throw e
+    }
+    finally{
+        if (!conf.get("no_reboot", false)) {
+            reboot()
+        }
+    }
+}
+
+def Build_CK(Map conf=[:]){
+        show_node_info()
+
+        env.HSA_ENABLE_SDMA=0
+        checkout scm
+
+        def image = getDockerImageName() 
+        def prefixpath = conf.get("prefixpath", "/opt/rocm")
+
+        // Jenkins is complaining about the render group 
+        def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+        if (conf.get("enforce_xnack_on", false)) {
+            dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
+        }
+        def dockerArgs = "--build-arg PREFIX=${prefixpath} --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' "
+        if (params.COMPILER_VERSION == "amd-stg-open" || params.COMPILER_COMMIT != ""){
+            dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
+        }
+
+        def variant = env.STAGE_NAME
+        def retimage
+        def navi_node = 0
+
+        gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
+            try {
+                (retimage, image) = getDockerImage(conf)
+                withDockerContainer(image: image, args: dockerOpts) {
+                    timeout(time: 5, unit: 'MINUTES'){
+                        sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo | tee clinfo.log'
+                        if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
+                            throw new Exception ("GPU not found")
+                        }
+                        else{
+                            echo "GPU is OK"
+                        }
+                        if ( runShell('grep -n "gfx1030" clinfo.log') || runShell('grep -n "gfx1101" clinfo.log') ){
+                            navi_node = 1
+                        }
+                    }
+                }
+            }
+            catch (org.jenkinsci.plugins.workflow.steps.FlowInterruptedException e){
+                echo "The job was cancelled or aborted"
+                throw e
+            }
+            catch(Exception ex) {
+                retimage = docker.build("${image}", dockerArgs + " --no-cache .")
+                withDockerContainer(image: image, args: dockerOpts) {
+                    timeout(time: 5, unit: 'MINUTES'){
+                        sh 'PATH="/opt/rocm/opencl/bin:/opt/rocm/opencl/bin/x86_64:$PATH" clinfo |tee clinfo.log'
+                        if ( runShell('grep -n "Number of devices:.*. 0" clinfo.log') ){
+                            throw new Exception ("GPU not found")
+                        }
+                        else{
+                            echo "GPU is OK"
+                        }
+                        if ( runShell('grep -n "gfx1030" clinfo.log') || runShell('grep -n "gfx1101" clinfo.log') ){
+                            navi_node = 1
+                        }
+                    }
+                }
+            }
+            withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
+                timeout(time: 24, unit: 'HOURS')
+                {
+                    cmake_build(conf)
+                    dir("build"){
+                        //run tests and examples
+                        sh 'make -j check'
+                        if (navi_node == 0 ){
+                            //we only need the ckProfiler to run the performance tests, so we pack and stash it
+                            //do not stash profiler on Navi nodes
+                           sh 'tar -zcvf ckProfiler.tar.gz bin/ckProfiler'
+                           stash "ckProfiler.tar.gz"
+                        }
+                        if (params.RUN_FULL_QA){
+                           // build deb packages
+                           sh 'make -j package'
+                           archiveArtifacts artifacts: 'composablekernel-ckprofiler_*.deb'
+                           archiveArtifacts artifacts: 'composablekernel-tests_*.deb'
+                           sh 'mv composablekernel-ckprofiler_*.deb ckprofiler_0.2.0_amd64.deb'
+                           stash "ckprofiler_0.2.0_amd64.deb"
+                        }
+                    }
+                }
+            }
+        }
+        return retimage
+}
+
+def Build_CK_and_Reboot(Map conf=[:]){
+    try{
+        Build_CK(conf)
+    }
+    catch(e){
+        echo "throwing error exception while building CK"
+        echo 'Exception occurred: ' + e.toString()
+        throw e
+    }
+    finally{
+        if (!conf.get("no_reboot", false)) {
+            reboot()
+        }
+    }
+}
+
+def process_results(Map conf=[:]){
+    env.HSA_ENABLE_SDMA=0
+    checkout scm
+    def image = getDockerImageName() 
+    def prefixpath = "/opt/rocm"
+
+    // Jenkins is complaining about the render group 
+    def dockerOpts="--cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+    if (conf.get("enforce_xnack_on", false)) {
+        dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
+    }
+
+    def variant = env.STAGE_NAME
+    def retimage
+
+    gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCmSoftwarePlatform', repo: 'composable_kernel') {
+        try {
+            (retimage, image) = getDockerImage(conf)
+        }
+        catch (org.jenkinsci.plugins.workflow.steps.FlowInterruptedException e){
+            echo "The job was cancelled or aborted"
+            throw e
+        }
+    }
+
+    withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
+        timeout(time: 1, unit: 'HOURS'){
+            try{
+                dir("script"){
+                    if (params.RUN_FULL_QA){
+                        // unstash perf files to master
+                        unstash "perf_gemm.log"
+                        unstash "perf_resnet50_N256.log"
+                        unstash "perf_resnet50_N4.log"
+                        unstash "perf_batched_gemm.log"
+                        unstash "perf_grouped_gemm.log"
+                        unstash "perf_conv_fwd.log"
+                        unstash "perf_conv_bwd_data.log"
+                        unstash "perf_gemm_bilinear.log"
+                        unstash "perf_reduction.log"
+                        unstash "perf_splitK_gemm.log"
+                        unstash "perf_onnx_gemm.log"
+                        sh "./process_qa_data.sh"
+                        unstash "ckprofiler_0.2.0_amd64.deb"
+                        sh "sshpass -p ${env.ck_deb_pw} scp -o StrictHostKeyChecking=no ckprofiler_0.2.0_amd64.deb ${env.ck_deb_user}@${env.ck_deb_ip}:/var/www/html/composable_kernel/"
+                    }
+                    else{
+                        // unstash perf files to master
+                        unstash "perf_gemm.log"
+                        unstash "perf_resnet50_N256.log"
+                        unstash "perf_resnet50_N4.log"
+                        sh "./process_perf_data.sh"
+                    }
+                }
+            }
+            catch(e){
+                echo "throwing error exception while processing performance test results"
+                echo 'Exception occurred: ' + e.toString()
+                throw e
+            }
+        }
+    }
+}
+
+//launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
+CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=5.7;COMPILER_VERSION=rc1
+                                              0 21 * * * % ROCMVERSION=5.6;COMPILER_VERSION=;COMPILER_COMMIT=
+                                              0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : ""
+
+pipeline {
+    agent none
+    triggers {
+        parameterizedCron(CRON_SETTINGS)
+    }
+    options {
+        parallelsAlwaysFailFast()
+    }
+    parameters {
+        booleanParam(
+            name: "BUILD_DOCKER",
+            defaultValue: false,
+            description: "Force building docker image (default: false), set to true if docker image needs to be updated.")
+        string(
+            name: 'ROCMVERSION', 
+            defaultValue: '5.6', 
+            description: 'Specify which ROCM version to use: 5.6 (default).')
+        string(
+            name: 'COMPILER_VERSION', 
+            defaultValue: '', 
+            description: 'Specify which version of compiler to use: release, amd-stg-open, or leave blank (default).')
+        string(
+            name: 'COMPILER_COMMIT', 
+            defaultValue: '', 
+            description: 'Specify which commit of compiler branch to use: leave blank to use the latest commit, or use 5541927df00eabd6a110180170eca7785d436ee3 (default) commit of amd-stg-open branch.')
+        string(
+            name: 'BUILD_COMPILER', 
+            defaultValue: 'hipcc', 
+            description: 'Specify whether to build CK with hipcc (default) or with clang.')
+        booleanParam(
+            name: "RUN_FULL_QA",
+            defaultValue: false,
+            description: "Select whether to run small set of performance tests (default) or full QA")
+    }
+    environment{
+        dbuser = "${dbuser}"
+        dbpassword = "${dbpassword}"
+        dbsship = "${dbsship}"
+        dbsshport = "${dbsshport}"
+        dbsshuser = "${dbsshuser}"
+        dbsshpassword = "${dbsshpassword}"
+        status_wrapper_creds = "${status_wrapper_creds}"
+        gerrit_cred="${gerrit_cred}"
+        DOCKER_BUILDKIT = "1"
+    }
+    stages{
+        stage("Build Docker"){
+            //when {
+            //    beforeAgent true
+            //    expression { params.BUILD_DOCKER.toBoolean() }
+            //}
+            parallel{
+                stage('Docker /opt/rocm'){
+                    agent{ label rocmnode("nogpu") }
+                    steps{
+                        buildDocker('/opt/rocm')
+                    }
+                }
+            }
+        }
+        stage("Static checks") {
+            parallel{
+                stage('Clang Format') {
+                    agent{ label rocmnode("nogpu") }
+                    environment{
+                        execute_cmd = "find .. -not -path \'*.git*\' -iname \'*.h\' \
+                                -o -not -path \'*.git*\' -iname \'*.hpp\' \
+                                -o -not -path \'*.git*\' -iname \'*.cpp\' \
+                                -o -iname \'*.h.in\' \
+                                -o -iname \'*.hpp.in\' \
+                                -o -iname \'*.cpp.in\' \
+                                -o -iname \'*.cl\' \
+                                | grep -v 'build/' \
+                                | xargs -n 1 -P 1 -I{} -t sh -c \'clang-format-12 -style=file {} | diff - {}\'"
+                    }
+                    steps{
+                        buildHipClangJobAndReboot(setup_cmd: "", build_cmd: "", execute_cmd: execute_cmd, no_reboot:true)
+                    }
+                }
+            }
+        }
+    
+		stage("Build CK and run Tests")
+        {
+            parallel
+            {
+                stage("Build CK and run Tests on MI100/MI200/MI300")
+                {
+                    when {
+                        beforeAgent true
+                        expression { params.RUN_FULL_QA.toBoolean() }
+                    }
+                    agent{ label rocmnode("gfx908 || gfx90a") }
+                    environment{
+                        setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941" """
+                        execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ 
+                    }
+                    steps{
+                        Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
+                    }
+                }
+                stage("Build CK and run Tests on MI100/MI200")
+                {
+                    when {
+                        beforeAgent true
+                        expression { !params.RUN_FULL_QA.toBoolean() }
+                    }
+                    agent{ label rocmnode("gfx908 || gfx90a") }
+                    environment{
+                        setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a" """
+                        execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ 
+                    }
+                    steps{
+                        Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
+                    }
+                }
+                stage("Build CK and run Tests on Navi21")
+                {
+                    when {
+                        beforeAgent true
+                        expression { !params.RUN_FULL_QA.toBoolean() }
+                    }
+                    agent{ label rocmnode("navi21") }
+                    environment{
+                        setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx1030" -DDL_KERNELS=ON """ 
+                        execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx1030" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
+
+                    }
+                    steps{
+                        Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
+                    }
+                }
+                stage("Build CK and run Tests on Navi32")
+                {
+                    when {
+                        beforeAgent true
+                        expression { !params.RUN_FULL_QA.toBoolean() }
+                    }
+                    agent{ label rocmnode("navi32") }
+                    environment{
+                        setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DDTYPES="fp16;fp32;bf16" -DGPU_TARGETS="gfx1101" """
+                        execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx1101" -DDTYPES="fp16;fp32;bf16" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
+
+                    }
+                    steps{
+                        Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
+                    }
+                }
+            }
+        }
+
+        stage("Performance Tests")
+        {
+            parallel
+            {
+                stage("Run ckProfiler: gfx90*")
+                {
+                    when {
+                        beforeAgent true
+                        expression { !params.RUN_FULL_QA.toBoolean() }
+                    }
+                    options { retry(2) }
+                    agent{ label rocmnode("gfx908 || gfx90a")}
+                    environment{
+                        setup_args = """ -DGPU_TARGETS="gfx908;gfx90a" -DBUILD_DEV=On """
+                   }
+                    steps{
+                        runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
+                    }
+                }
+                stage("Run ckProfiler: gfx90a")
+                {
+                    when {
+                        beforeAgent true
+                        expression { params.RUN_FULL_QA.toBoolean() }
+                    }
+                    options { retry(2) }
+                    agent{ label rocmnode("gfx90a")}
+                    environment{
+                        setup_args = """ -DGPU_TARGETS="gfx90a" -DBUILD_DEV=On """
+                    }
+                    steps{
+                        runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
+                    }
+                }
+            }
+        }
+        stage("Process Performance Test Results")
+        {
+            parallel
+            {
+                stage("Process results"){
+                    agent { label 'mici' }
+                    steps{
+                        process_results()
+                    }
+                }
+            }
+        }
+    }
+}
diff --git a/LICENSE b/LICENSE
new file mode 100644
index 0000000000..e03fddaf78
--- /dev/null
+++ b/LICENSE
@@ -0,0 +1,28 @@
+Copyright (c) 2018-    , Advanced Micro Devices, Inc. (Chao Liu, Jing Zhang)
+Copyright (c) 2019-    , Advanced Micro Devices, Inc. (Letao Qin, Qianfeng Zhang, Liang Huang, Shaojie Wang)
+Copyright (c) 2022-    , Advanced Micro Devices, Inc. (Anthony Chang, Chunyu Lai, Illia Silin, Adam Osewski, Poyen Chen, Jehandad Khan)
+Copyright (c) 2019-2021, Advanced Micro Devices, Inc. (Hanwen Chang)
+Copyright (c) 2019-2020, Advanced Micro Devices, Inc. (Tejash Shah)
+Copyright (c) 2020     , Advanced Micro Devices, Inc. (Xiaoyan Zhou)
+Copyright (c) 2021-2022, Advanced Micro Devices, Inc. (Jianfeng Yan)
+
+SPDX-License-Identifier: MIT
+Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/README.md b/README.md
new file mode 100644
index 0000000000..c2b493db11
--- /dev/null
+++ b/README.md
@@ -0,0 +1,154 @@
+# Composable Kernel
+
+## Methodology
+
+Composable Kernel (CK) library aims to provide a programming model for writing performance critical kernels for machine learning workloads across multiple architectures including GPUs, CPUs, etc, through general purpose kernel languages, like HIP C++.
+
+CK utilizes two concepts to achieve performance portability and code maintainability:
+* A tile-based programming model
+* Algorithm complexity reduction for complex ML operators, using innovative technique we call "Tensor Coordinate Transformation".
+
+![ALT](/docs/data/ck_component.png "CK Components")
+
+## Code Structure
+
+Current CK library are structured into 4 layers:
+* "Templated Tile Operators" layer
+* "Templated Kernel and Invoker" layer
+* "Instantiated Kernel and Invoker" layer
+* "Client API" layer
+
+![ALT](/docs/data/ck_layer.png "CK Layers")
+
+## Documentation
+
+Run the steps below to build documentation locally.
+
+```
+cd docs
+pip3 install -r sphinx/requirements.txt
+python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en . _build/html
+```
+
+## Contributors
+
+The list of developers and contributors is here: [Contributors](/CONTRIBUTORS.md)
+
+## Citation
+
+If you use CK, please use following citations:
+* CK paper will be freely available on arXiv soon: [Realizing Tensor Operators Using Coordinate Transformations and Tile Based Programming](???)
+* [CITATION.cff](/CITATION.cff)
+
+## License
+
+CK is released under the MIT license. [License File](/LICENSE)
+
+
+# Build CK
+
+## Build docker image
+
+```bash
+DOCKER_BUILDKIT=1 docker build -t ck:latest -f Dockerfile .
+```
+Pre-built dockers are available from this public repo: 
+https://hub.docker.com/r/rocm/composable_kernel/tags
+
+## Launch docker
+
+```bash
+docker run                                     \
+-it                                            \
+--privileged                                   \
+--group-add sudo                               \
+-w /root/workspace                             \
+-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace  \
+ck:latest                                      \
+/bin/bash
+```
+
+## Build CK
+
+```bash
+mkdir build && cd build
+
+# Need to specify target ID, example below is for gfx908 and gfx90a
+
+cmake                                                                                             \
+-D CMAKE_PREFIX_PATH=/opt/rocm                                                                    \
+-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                         \
+-D CMAKE_BUILD_TYPE=Release                                                                       \
+-D GPU_TARGETS="gfx908;gfx90a"                                                                    \
+..
+```
+
+If GPU_TARGETS is not set on the cmake command line, CK will be built for all targets supported by the 
+current compiler.
+
+
+Additional cmake flags can be used to significantly speed-up the build:
+
+INSTANCES_ONLY (by default is OFF) must be set to ON in order to build only the instances and library
+while skipping all tests, examples, and profiler. This is useful for libraries that use CK as a dependency.
+
+DTYPES (by default not set) can be set to any subset of "fp64;fp32;fp16;fp8;bf16;int8" to build instances 
+of select data types only. Currently, building of int8 instances is taking a lot of time (the compiler fix is in the works).
+
+DL_KERNELS (by default is OFF) must be set to ON in order to build the gemm_dl and batched_gemm_multi_d_dl 
+instances. Those instances are only needed for the NAVI2x platforms.
+
+### Build examples and tests
+
+```bash
+ make -j examples tests
+ make test
+```
+
+Instructions for running each individual examples are under [example](/example)
+
+
+## Build ckProfiler
+
+```bash
+ make -j ckProfiler
+```
+Instructions for running ckProfiler are under [profiler](/profiler)
+
+## Install CK
+
+```bash
+make install
+```
+
+## Using CK as pre-built kernel library
+
+Instructions for using CK as a pre-built kernel library are under [client_example](/client_example)
+
+## Contributing
+
+When you contribute to Composable Kernel, make sure to run `clang-format` on all the changed files. We highly recommend using git hooks that are managed by the `pre-commit` framework. To install hooks, run:
+
+```bash
+sudo script/install_precommit.sh
+```
+
+This way, `pre-commit` will add the appropriate hooks to your local repository and automatically run `clang-format` (and possibly additional checks) before any commit is created.
+
+If you need to uninstall hooks from the repository, you can do so by running the following command:
+
+```bash
+script/uninstall_precommit.sh
+```
+
+If for any reason, you need to temporarily disable precommit hooks, you can add the `--no-verify` option to the `git commit` command.
+
+## Caveat
+### Kernel Timing and Verification
+
+CK's own kernel timer will warn up kernel once, and then run it multiple times
+to get average kernel time. For some kernels that use atomic add, this will cause
+output buffer to be accumulated multiple times, causing verification failure.
+To work around it, do not use CK's own timer and do verification at the same time.
+CK's own timer and verification in each example and ckProfiler can be enabled or
+disabled from command line.
diff --git a/client_example/01_gemm/CMakeLists.txt b/client_example/01_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..9e741192f9
--- /dev/null
+++ b/client_example/01_gemm/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_gemm gemm.cpp)
+target_link_libraries(client_gemm PRIVATE composable_kernel::device_operations)
diff --git a/client_example/01_gemm/gemm.cpp b/client_example/01_gemm/gemm.cpp
new file mode 100644
index 0000000000..c37f208db1
--- /dev/null
+++ b/client_example/01_gemm/gemm.cpp
@@ -0,0 +1,218 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+using ADataType = F16;
+using BDataType = F16;
+using CDataType = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 7)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideC = std::stoi(argv[6]);
+    }
+    else
+    {
+        printf("arg1 to 6: M, N, K, StrideA, StrideB, StrideC\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem c_device_buf(sizeof(CDataType) * f_matrix_space_size(M, N, StrideC, CLayout{}));
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGemm<ALayout,
+                                                 BLayout,
+                                                 CLayout,
+                                                 ADataType,
+                                                 BDataType,
+                                                 CDataType,
+                                                 ck::tensor_operation::element_wise::PassThrough,
+                                                 ck::tensor_operation::element_wise::PassThrough,
+                                                 ck::tensor_operation::element_wise::PassThrough>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        StrideC,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        StrideC,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt b/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt
new file mode 100644
index 0000000000..ba29520222
--- /dev/null
+++ b/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt
@@ -0,0 +1,27 @@
+add_custom_target(client_gemm_fastgelu_examples)
+
+add_executable(client_gemm_add_add_fastgelu gemm_add_add_fastgelu.cpp)
+target_link_libraries(client_gemm_add_add_fastgelu PRIVATE composable_kernel::device_operations)
+
+add_executable(client_gemm_add_fastgelu gemm_add_fastgelu.cpp)
+target_link_libraries(client_gemm_add_fastgelu PRIVATE composable_kernel::device_operations)
+
+add_executable(client_gemm_fastgelu gemm_fastgelu.cpp)
+target_link_libraries(client_gemm_fastgelu PRIVATE composable_kernel::device_operations)
+
+add_dependencies(client_gemm_fastgelu_examples client_gemm_add_add_fastgelu client_gemm_add_fastgelu
+                 client_gemm_fastgelu)
+
+add_custom_target(client_gemm_fastgelu_generic_examples)
+
+add_executable(client_gemm_add_add_fastgelu_generic gemm_add_add_fastgelu_generic.cpp)
+target_link_libraries(client_gemm_add_add_fastgelu_generic PRIVATE composable_kernel::device_operations)
+
+add_executable(client_gemm_add_fastgelu_generic gemm_add_fastgelu_generic.cpp)
+target_link_libraries(client_gemm_add_fastgelu_generic PRIVATE composable_kernel::device_operations)
+
+add_executable(client_gemm_fastgelu_generic gemm_fastgelu_generic.cpp)
+target_link_libraries(client_gemm_fastgelu_generic PRIVATE composable_kernel::device_operations)
+
+add_dependencies(client_gemm_fastgelu_generic_examples client_gemm_add_add_fastgelu_generic 
+                 client_gemm_add_fastgelu_generic client_gemm_fastgelu_generic)
diff --git a/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp
new file mode 100644
index 0000000000..756889562e
--- /dev/null
+++ b/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp
@@ -0,0 +1,241 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F16;
+using D1DataType = F16;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using ELayout  = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 4096;
+    ck::index_t StrideE  = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 9)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA  = std::stoi(argv[4]);
+        StrideB  = std::stoi(argv[5]);
+        StrideD0 = std::stoi(argv[6]);
+        StrideD1 = std::stoi(argv[7]);
+        StrideE  = std::stoi(argv[8]);
+    }
+    else
+    {
+        printf("arg1 to 8: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
+                                      f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem d1_m_n_device_buf(sizeof(D1DataType) *
+                                      f_matrix_space_size(M, N, StrideD1, D1Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddAddFastGelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp
new file mode 100644
index 0000000000..2ed942f0ad
--- /dev/null
+++ b/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp
@@ -0,0 +1,176 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+#include <stdexcept>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F16;
+using D1DataType = F16;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using ELayout  = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 4096;
+    ck::index_t StrideE  = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 9)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA  = std::stoi(argv[4]);
+        StrideB  = std::stoi(argv[5]);
+        StrideD0 = std::stoi(argv[6]);
+        StrideD1 = std::stoi(argv[7]);
+        StrideE  = std::stoi(argv[8]);
+    }
+    else
+    {
+        printf("arg1 to 8: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
+                                      f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem d1_m_n_device_buf(sizeof(D1DataType) *
+                                      f_matrix_space_size(M, N, StrideD1, D1Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddAddFastGelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    // get generic instance
+    auto& op_ptr = op_ptrs[0];
+
+    std::cout << "Run the generic instance without timing: " << op_ptr->GetTypeString()
+              << std::endl;
+
+    // run the generic instance
+    auto argument_ptr =
+        op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                                               d1_m_n_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    M,
+                                    N,
+                                    K,
+                                    StrideA,
+                                    StrideB,
+                                    std::array<ck::index_t, 2>{StrideD0, StrideD1},
+                                    StrideE,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+    if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+    {
+        invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+    }
+    else
+    {
+        throw std::runtime_error(
+            "Generic instance should be suitable for various input lengths/strides");
+    }
+
+    std::cout << "Done" << std::endl;
+
+    return 0;
+}
diff --git a/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp
new file mode 100644
index 0000000000..8d2a8c234a
--- /dev/null
+++ b/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp
@@ -0,0 +1,233 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddFastGelu;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F16;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using ELayout  = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideE  = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 8)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA  = std::stoi(argv[4]);
+        StrideB  = std::stoi(argv[5]);
+        StrideD0 = std::stoi(argv[6]);
+        StrideE  = std::stoi(argv[7]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideD0, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
+                                      f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddFastGelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 1>{StrideD0},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 1>{StrideD0},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp
new file mode 100644
index 0000000000..644b428fc9
--- /dev/null
+++ b/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp
@@ -0,0 +1,169 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+#include <stdexcept>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddFastGelu;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F16;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using ELayout  = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideE  = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 8)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA  = std::stoi(argv[4]);
+        StrideB  = std::stoi(argv[5]);
+        StrideD0 = std::stoi(argv[6]);
+        StrideE  = std::stoi(argv[7]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideD0, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
+                                      f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddFastGelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    // get generic instance
+    auto& op_ptr = op_ptrs[0];
+
+    std::cout << "Run the generic instance without timing: " << op_ptr->GetTypeString()
+              << std::endl;
+
+    // run the generic instance
+    auto argument_ptr =
+        op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    M,
+                                    N,
+                                    K,
+                                    StrideA,
+                                    StrideB,
+                                    std::array<ck::index_t, 1>{StrideD0},
+                                    StrideE,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+    if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+    {
+        invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+    }
+    else
+    {
+        throw std::runtime_error(
+            "Generic instance should be suitable for various input lengths/strides");
+    }
+
+    std::cout << "Done" << std::endl;
+
+    return 0;
+}
diff --git a/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp
new file mode 100644
index 0000000000..c02df018fd
--- /dev/null
+++ b/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp
@@ -0,0 +1,225 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = FastGelu;
+
+using ADataType = F16;
+using BDataType = F16;
+using EDataType = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideE = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 7)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideE = std::stoi(argv[6]);
+    }
+    else
+    {
+        printf("arg1 to 6: M, N, K, StrideA, StrideB, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::FastGelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        {},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        {},
+                                                        StrideE,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        {},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        {},
+                                                        StrideE,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp
new file mode 100644
index 0000000000..482e93b421
--- /dev/null
+++ b/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp
@@ -0,0 +1,162 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+#include <stdexcept>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = FastGelu;
+
+using ADataType = F16;
+using BDataType = F16;
+using EDataType = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideE = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 7)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideE = std::stoi(argv[6]);
+    }
+    else
+    {
+        printf("arg1 to 6: M, N, K, StrideA, StrideB, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::FastGelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    // get generic instance
+    auto& op_ptr = op_ptrs[0];
+
+    std::cout << "Run the generic instance without timing: " << op_ptr->GetTypeString()
+              << std::endl;
+
+    // run the generic instance
+    auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                    b_device_buf.GetDeviceBuffer(),
+                                                    {},
+                                                    e_device_buf.GetDeviceBuffer(),
+                                                    M,
+                                                    N,
+                                                    K,
+                                                    StrideA,
+                                                    StrideB,
+                                                    {},
+                                                    StrideE,
+                                                    a_element_op,
+                                                    b_element_op,
+                                                    cde_element_op);
+
+    auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+    if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+    {
+        invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+    }
+    else
+    {
+        throw std::runtime_error(
+            "Generic instance should be suitable for various input lengths/strides");
+    }
+
+    std::cout << "Done" << std::endl;
+
+    return 0;
+}
diff --git a/client_example/03_gemm_layernorm/CMakeLists.txt b/client_example/03_gemm_layernorm/CMakeLists.txt
new file mode 100644
index 0000000000..b38698d906
--- /dev/null
+++ b/client_example/03_gemm_layernorm/CMakeLists.txt
@@ -0,0 +1,5 @@
+add_executable(client_gemm_add_add_layernorm_naive gemm_add_add_layernorm_naive.cpp)
+target_link_libraries(client_gemm_add_add_layernorm_naive PRIVATE composable_kernel::device_operations)
+
+add_executable(client_gemm_add_relu_add_layernorm_welford gemm_add_relu_add_layernorm_welford.cpp)
+target_link_libraries(client_gemm_add_relu_add_layernorm_welford PRIVATE composable_kernel::device_operations)
diff --git a/client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp b/client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp
new file mode 100644
index 0000000000..58c91f903b
--- /dev/null
+++ b/client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp
@@ -0,0 +1,275 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ADataType            = F16;
+using BDataType            = F16;
+using BiasDataType         = F32;
+using CDataType            = F16;
+using D0DataType           = F16;
+using ReduceDataType       = F32;
+using GammaDataType        = F16;
+using BetaDataType         = F16;
+using LayerNormOutDataType = F16;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+template <typename gemm_reduce_op_ptr>
+bool RunDeviceGemmMeanSquareMean(gemm_reduce_op_ptr& p_op,
+                                 const void* p_a,
+                                 const void* p_b,
+                                 const void* p_bias,
+                                 const void* p_d0,
+                                 void* p_c,
+                                 void* p_mean,
+                                 void* p_square_mean,
+                                 int M,
+                                 int N,
+                                 int K,
+                                 int StrideA,
+                                 int StrideB,
+                                 int StrideC,
+                                 int StrideD0,
+                                 bool time_kernel)
+{
+    using PassThrough          = ck::tensor_operation::element_wise::PassThrough;
+    using UnaryDivElementOp    = ck::tensor_operation::element_wise::UnaryDivide;
+    using UnarySquareElementOp = ck::tensor_operation::element_wise::UnarySquare;
+
+    auto passOp   = PassThrough{};
+    auto squareOp = UnarySquareElementOp{};
+    auto divOp    = UnaryDivElementOp{N};
+
+    auto argument_ptr =
+        p_op->MakeArgumentPointer(p_a,
+                                  p_b,
+                                  p_bias,
+                                  {p_d0},
+                                  p_c,
+                                  {p_mean, p_square_mean},
+                                  M,
+                                  N,
+                                  K,
+                                  StrideA,
+                                  StrideB,
+                                  StrideC,
+                                  {StrideD0},
+                                  {&passOp, &passOp, &passOp}, // functor for a, b, c
+                                  {&passOp},                   // functor for d0
+                                  {&passOp, &squareOp},        // functor for inputs of reduction
+                                  {&divOp, &divOp});           // functor for outputs of reduction
+
+    if(p_op->IsSupportedArgument(argument_ptr.get()))
+    {
+        auto invoker_ptr = p_op->MakeInvokerPointer();
+
+        // If we evaluate running time of gemm_reduce. The output may wrong.
+        // Because we need to initialize the reduction tensor before runing the kernel.
+        // However we run kernel many times for time_kernel = trie without reinitialize the out
+        // of reduction tensor.
+        float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        if(time_kernel)
+            std::cout << "Gemm + reduce Perf: " << std::setw(10) << ave_time << " ms" << std::endl;
+
+        return true;
+    }
+
+    return false;
+}
+
+template <typename normalize_op_ptr>
+bool RunDeviceNormalize2D(normalize_op_ptr& p_op,
+                          const void* p_x,
+                          const void* p_mean,
+                          const void* p_square_mean,
+                          const void* p_gamma,
+                          const void* p_beta,
+                          void* p_y,
+                          int M,
+                          int N,
+                          int StrideX,
+                          bool time_kernel)
+{
+    std::array<const void*, 5> input = {p_x, p_mean, p_square_mean, p_gamma, p_beta};
+    std::array<void*, 1> output      = {p_y};
+    auto normalize_functor           = ck::tensor_operation::element_wise::Normalize{};
+
+    std::array<ck::index_t, 2> xyLengths = {M, N};
+    std::array<ck::index_t, 2> xyStrides = {StrideX, 1};
+
+    auto argument_ptr = p_op->MakeArgumentPointer(xyLengths,
+                                                  {xyStrides, {1, 0}, {1, 0}, {0, 1}, {0, 1}},
+                                                  {xyStrides},
+                                                  input,
+                                                  output,
+                                                  ck::tensor_operation::element_wise::Normalize{});
+
+    if(p_op->IsSupportedArgument(argument_ptr.get()))
+    {
+        auto invoker_ptr = p_op->MakeInvokerPointer();
+        float ave_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        if(time_kernel)
+            std::cout << "Normalize Perf: " << std::setw(10) << ave_time << " ms" << std::endl;
+
+        return true;
+    }
+
+    return false;
+}
+
+int main()
+{
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA  = 1024;
+    ck::index_t StrideB  = 1024;
+    ck::index_t StrideC  = 1024;
+    ck::index_t StrideD0 = 1024;
+
+    const auto gemm_reduce_ptrs =
+        ck::tensor_operation::device::instance::get_device_gemm_add_add_mean_squaremean_instances<
+            ADataType,
+            BDataType,
+            CDataType,
+            ALayout,
+            BLayout,
+            CLayout>();
+
+    std::cout << "found " << gemm_reduce_ptrs.size()
+              << " gemm_reduceMean_reduceSquareMean instances" << std::endl;
+
+    using NormalizeDeviceOp = ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<CDataType, ReduceDataType, ReduceDataType, GammaDataType, BetaDataType>,
+        ck::Tuple<LayerNormOutDataType>,
+        ck::tensor_operation::element_wise::Normalize,
+        2>;
+
+    const auto normalize_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            NormalizeDeviceOp>::GetInstances();
+
+    std::cout << "found " << normalize_ptrs.size() << " normalize instances" << std::endl;
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem bias_device_buf(sizeof(BiasDataType) * N);
+    SimpleDeviceMem c_device_buf(sizeof(CDataType) * f_matrix_space_size(M, N, StrideC, CLayout{}));
+    SimpleDeviceMem d0_device_buf(sizeof(D0DataType) *
+                                  f_matrix_space_size(M, N, StrideD0, CLayout{}));
+    SimpleDeviceMem reduceMean_device_buf(sizeof(ReduceDataType) * M);
+    SimpleDeviceMem reduceMeanSquare_device_buf(sizeof(ReduceDataType) * M);
+    SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N);
+    SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N);
+    SimpleDeviceMem layerNorm_device_buf(sizeof(LayerNormOutDataType) * M * N);
+
+    bool b_time_kernel           = true;
+    bool b_only_run_first_kernel = true;
+
+    // layernorm => (1) + (2)
+    // (1). c = gemm(a, b), reduce_mean(c), reduce_square_mean(c)
+    // (2). normalize(c, mean, square_mean, gamma, beta)
+    for(auto& gemm_reduce_ptr : gemm_reduce_ptrs)
+    {
+        // run first available kernel
+        if(RunDeviceGemmMeanSquareMean(gemm_reduce_ptr,
+                                       a_device_buf.GetDeviceBuffer(),
+                                       b_device_buf.GetDeviceBuffer(),
+                                       bias_device_buf.GetDeviceBuffer(),
+                                       d0_device_buf.GetDeviceBuffer(),
+                                       c_device_buf.GetDeviceBuffer(),
+                                       reduceMean_device_buf.GetDeviceBuffer(),
+                                       reduceMeanSquare_device_buf.GetDeviceBuffer(),
+                                       M,
+                                       N,
+                                       K,
+                                       StrideA,
+                                       StrideB,
+                                       StrideC,
+                                       StrideD0,
+                                       b_time_kernel))
+        {
+            if(b_only_run_first_kernel)
+                break;
+        }
+        else
+        {
+            std::cout << gemm_reduce_ptr->GetTypeString() << " does not support this problem"
+                      << std::endl;
+        }
+    }
+
+    for(auto& normalize_ptr : normalize_ptrs)
+    {
+        if(RunDeviceNormalize2D(normalize_ptr,
+                                c_device_buf.GetDeviceBuffer(),
+                                reduceMean_device_buf.GetDeviceBuffer(),
+                                reduceMeanSquare_device_buf.GetDeviceBuffer(),
+                                gamma_device_buf.GetDeviceBuffer(),
+                                beta_device_buf.GetDeviceBuffer(),
+                                layerNorm_device_buf.GetDeviceBuffer(),
+                                M,
+                                N,
+                                StrideC,
+                                b_time_kernel))
+        {
+            if(b_only_run_first_kernel)
+                break;
+        }
+        else
+        {
+            std::cout << normalize_ptr->GetTypeString() << " does not support this problem"
+                      << std::endl;
+        }
+    }
+}
diff --git a/client_example/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp b/client_example/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp
new file mode 100644
index 0000000000..3d5fb60048
--- /dev/null
+++ b/client_example/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp
@@ -0,0 +1,244 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using F16 = ck::half_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+// DataType
+using ADataType     = F16;
+using BDataType     = F16;
+using D0DataType    = F16;
+using D1DataType    = F16;
+using GammaDataType = F16;
+using BetaDataType  = F16;
+using HDataType     = F16;
+
+// Layout
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using HLayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddReluAdd;
+using HElementOp   = PassThrough;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}, mMemSize_(mem_size)
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    void SetZero() const { (void)hipMemset(p_mem_, 0, mMemSize_); }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+    std::size_t mMemSize_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA  = K;
+    ck::index_t StrideB  = K;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = N;
+    ck::index_t StrideH  = N;
+
+    float epsilon = 1e-5;
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_device_buf(sizeof(D0DataType) *
+                                  f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem d1_device_buf(sizeof(D1DataType) *
+                                  f_matrix_space_size(M, N, StrideD1, D1Layout{}));
+    SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N);
+    SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N);
+    SimpleDeviceMem h_device_buf(sizeof(HDataType) * f_matrix_space_size(M, N, StrideH, HLayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleDLayernorm<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        HLayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        GammaDataType,
+        BetaDataType,
+        HDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddReluAdd,
+        ck::tensor_operation::element_wise::PassThrough>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+    const auto h_element_op   = HElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+            gamma_device_buf.GetDeviceBuffer(),
+            beta_device_buf.GetDeviceBuffer(),
+            h_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            {StrideD0, StrideD1},
+            StrideH,
+            epsilon,
+            a_element_op,
+            b_element_op,
+            cde_element_op,
+            h_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace_dev(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+            h_device_buf.SetZero();
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                (sizeof(D0DataType) + sizeof(D1DataType) + sizeof(HDataType)) * M * N +
+                (sizeof(GammaDataType) + sizeof(BetaDataType)) * N;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+            gamma_device_buf.GetDeviceBuffer(),
+            beta_device_buf.GetDeviceBuffer(),
+            h_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            {StrideD0, StrideD1},
+            StrideH,
+            epsilon,
+            a_element_op,
+            b_element_op,
+            cde_element_op,
+            h_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace_dev(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+            h_device_buf.SetZero();
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/client_example/04_contraction/CMakeLists.txt b/client_example/04_contraction/CMakeLists.txt
new file mode 100644
index 0000000000..7ffedfeef3
--- /dev/null
+++ b/client_example/04_contraction/CMakeLists.txt
@@ -0,0 +1,15 @@
+add_executable(client_contraction_scale_fp32 contraction_scale_fp32.cpp)
+target_link_libraries(client_contraction_scale_fp32 PRIVATE composable_kernel::device_operations)
+
+add_executable(client_contraction_bilinear_fp32 contraction_bilinear_fp32.cpp)
+target_link_libraries(client_contraction_bilinear_fp32 PRIVATE composable_kernel::device_operations)
+
+add_executable(client_contraction_scale_fp64 contraction_scale_fp64.cpp)
+target_link_libraries(client_contraction_scale_fp64 PRIVATE composable_kernel::device_operations)
+
+add_executable(client_contraction_bilinear_fp64 contraction_bilinear_fp64.cpp)
+target_link_libraries(client_contraction_bilinear_fp64 PRIVATE composable_kernel::device_operations)
+
+add_executable(contraction_g1m2n3k1_add_xdl_fp16 contraction_g1m2n3k1_add_xdl_fp16.cpp)
+target_link_libraries(contraction_g1m2n3k1_add_xdl_fp16 PRIVATE composable_kernel::device_operations)
+
diff --git a/client_example/04_contraction/contraction_bilinear_fp32.cpp b/client_example/04_contraction/contraction_bilinear_fp32.cpp
new file mode 100644
index 0000000000..89f834b982
--- /dev/null
+++ b/client_example/04_contraction/contraction_bilinear_fp32.cpp
@@ -0,0 +1,236 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <numeric>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Bilinear;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F32;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F32;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+
+    float alpha = 1.f;
+    float beta  = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 25)
+    {
+        const ck::index_t M0 = std::stoi(argv[1]);
+        const ck::index_t M1 = std::stoi(argv[2]);
+
+        const ck::index_t N0 = std::stoi(argv[3]);
+        const ck::index_t N1 = std::stoi(argv[4]);
+
+        const ck::index_t K0 = std::stoi(argv[5]);
+        const ck::index_t K1 = std::stoi(argv[6]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[7]), std::stoi(argv[8]), std::stoi(argv[9]), std::stoi(argv[10])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13]), std::stoi(argv[14])};
+
+        d_ms_ns_lengths = {M0, M1, N0, N1};
+        d_ms_ns_strides = {
+            std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17]), std::stoi(argv[18])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21]), std::stoi(argv[22])};
+
+        alpha = std::stof(argv[23]);
+        beta  = std::stof(argv[24]);
+    }
+    else
+    {
+        printf("arg1 to 6: M0, M1, N0, N1, K0, K1\n");
+        printf("arg7 to 10: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg11 to 14: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg15 to 18: Stride_D_M0, Stride_D_M1, Stride_D_N0, Stride_D_N1\n");
+        printf("arg19 to 22: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg23 to 24: alpha, beta\n");
+        exit(0);
+    }
+
+    auto f_tensor_space_size = [](auto lengths, auto strides) {
+        std::size_t space_size = 1;
+        for(std::size_t i = 0; i < lengths.size(); ++i)
+        {
+            space_size += (lengths[i] - 1) * strides[i];
+        }
+        return space_size;
+    };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) *
+                                 f_tensor_space_size(a_ms_ks_lengths, a_ms_ks_strides));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) *
+                                 f_tensor_space_size(b_ns_ks_lengths, b_ns_ks_strides));
+    SimpleDeviceMem d_device_buf(sizeof(DDataType) *
+                                 f_tensor_space_size(d_ms_ns_lengths, d_ms_ns_strides));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) *
+                                 f_tensor_space_size(e_ms_ns_lengths, e_ms_ns_strides));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        ck::Tuple<DDataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Bilinear>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{alpha, beta};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                        b_device_buf.GetDeviceBuffer(),
+                                        std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                        e_device_buf.GetDeviceBuffer(),
+                                        a_ms_ks_lengths,
+                                        a_ms_ks_strides,
+                                        b_ns_ks_lengths,
+                                        b_ns_ks_strides,
+                                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                                        e_ms_ns_lengths,
+                                        e_ms_ns_strides,
+                                        a_element_op,
+                                        b_element_op,
+                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+            std::size_t flop      = std::size_t(2) * M * N * K;
+            std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                    sizeof(DDataType) * M * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return 0;
+}
diff --git a/client_example/04_contraction/contraction_bilinear_fp64.cpp b/client_example/04_contraction/contraction_bilinear_fp64.cpp
new file mode 100644
index 0000000000..1aa3ba7de5
--- /dev/null
+++ b/client_example/04_contraction/contraction_bilinear_fp64.cpp
@@ -0,0 +1,281 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <numeric>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+using F64 = double;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Bilinear;
+
+using ADataType        = F64;
+using BDataType        = F64;
+using AccDataType      = F64;
+using CShuffleDataType = F64;
+using DDataType        = F64;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F64;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+// kknn
+#if 1
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// knnn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{64, 1, 131072, 2048};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// mknn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{128, 1, 245760, 3840};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// mnnn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{128, 1, 245760, 3840};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{64, 1, 131072, 2048};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+#endif
+
+    float alpha = 1.f;
+    float beta  = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 25)
+    {
+        const ck::index_t M0 = std::stoi(argv[1]);
+        const ck::index_t M1 = std::stoi(argv[2]);
+
+        const ck::index_t N0 = std::stoi(argv[3]);
+        const ck::index_t N1 = std::stoi(argv[4]);
+
+        const ck::index_t K0 = std::stoi(argv[5]);
+        const ck::index_t K1 = std::stoi(argv[6]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[7]), std::stoi(argv[8]), std::stoi(argv[9]), std::stoi(argv[10])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13]), std::stoi(argv[14])};
+
+        d_ms_ns_lengths = {M0, M1, N0, N1};
+        d_ms_ns_strides = {
+            std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17]), std::stoi(argv[18])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21]), std::stoi(argv[22])};
+
+        alpha = std::stof(argv[23]);
+        beta  = std::stof(argv[24]);
+    }
+    else
+    {
+        printf("arg1 to 6: M0, M1, N0, N1, K0, K1\n");
+        printf("arg7 to 10: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg11 to 14: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg15 to 18: Stride_D_M0, Stride_D_M1, Stride_D_N0, Stride_D_N1\n");
+        printf("arg19 to 22: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg23 to 24: alpha, beta\n");
+        exit(0);
+    }
+
+    auto f_tensor_space_size = [](auto lengths, auto strides) {
+        std::size_t space_size = 1;
+        for(std::size_t i = 0; i < lengths.size(); ++i)
+        {
+            space_size += (lengths[i] - 1) * strides[i];
+        }
+        return space_size;
+    };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) *
+                                 f_tensor_space_size(a_ms_ks_lengths, a_ms_ks_strides));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) *
+                                 f_tensor_space_size(b_ns_ks_lengths, b_ns_ks_strides));
+    SimpleDeviceMem d_device_buf(sizeof(DDataType) *
+                                 f_tensor_space_size(d_ms_ns_lengths, d_ms_ns_strides));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) *
+                                 f_tensor_space_size(e_ms_ns_lengths, e_ms_ns_strides));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        ck::Tuple<DDataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Bilinear>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{alpha, beta};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                        b_device_buf.GetDeviceBuffer(),
+                                        std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                        e_device_buf.GetDeviceBuffer(),
+                                        a_ms_ks_lengths,
+                                        a_ms_ks_strides,
+                                        b_ns_ks_lengths,
+                                        b_ns_ks_strides,
+                                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                                        e_ms_ns_lengths,
+                                        e_ms_ns_strides,
+                                        a_element_op,
+                                        b_element_op,
+                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+            std::size_t flop      = std::size_t(2) * M * N * K;
+            std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                    sizeof(DDataType) * M * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return 0;
+}
diff --git a/client_example/04_contraction/contraction_g1m2n3k1_add_xdl_fp16.cpp b/client_example/04_contraction/contraction_g1m2n3k1_add_xdl_fp16.cpp
new file mode 100644
index 0000000000..f8ea2258c2
--- /dev/null
+++ b/client_example/04_contraction/contraction_g1m2n3k1_add_xdl_fp16.cpp
@@ -0,0 +1,204 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <numeric>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Add;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimG = 1;
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 3;
+static constexpr ck::index_t NumDimK = 1;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    ck::index_t G0 = 1;
+
+    ck::index_t M0 = 64;
+    ck::index_t M1 = 256;
+
+    ck::index_t N0 = 3;
+    ck::index_t N1 = 12;
+    ck::index_t N2 = 64;
+
+    ck::index_t K0 = 768;
+
+    // A[M0, M1, M2, K0]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, M0, M1, K0};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M0 * M1 * K0, M1 * K0, K0, 1};
+    // B[N0, N1, N2, K0]
+    std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, N0, N1, N2, K0};
+    std::vector<ck::index_t> b_gs_ns_ks_strides{N0 * N1 * N2 * K0, N1 * N2 * K0, N2 * K0, K0, 1};
+
+    // D[N0, M0, N1, M1, N2]
+    std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, M0, M1, N0, N1, N2};
+    std::vector<ck::index_t> d_gs_ms_ns_strides{N0 * N1 * N2, 0, 0, N1 * N2, N2, 1};
+    // E[N0 M0 N1 N2 M1]
+    std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, M0, M1, N0, N1, N2};
+    std::vector<ck::index_t> e_gs_ms_ns_strides{
+        M0 * M1 * N0 * N1 * N2, N1 * N2 * M1, 1, M0 * N1 * N2 * M1, M1 * N2, M1};
+
+    auto f_tensor_space_size = [](auto lengths, auto strides) {
+        std::size_t space_size = 1;
+        for(std::size_t i = 0; i < lengths.size(); ++i)
+        {
+            space_size += (lengths[i] - 1) * strides[i];
+        }
+        return space_size;
+    };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) *
+                                 f_tensor_space_size(a_gs_ms_ks_lengths, a_gs_ms_ks_strides));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) *
+                                 f_tensor_space_size(b_gs_ns_ks_lengths, b_gs_ns_ks_strides));
+    SimpleDeviceMem d_device_buf(sizeof(DDataType) *
+                                 f_tensor_space_size(d_gs_ms_ns_lengths, d_gs_ms_ns_strides));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) *
+                                 f_tensor_space_size(e_gs_ms_ns_lengths, e_gs_ms_ns_strides));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchedContractionMultipleD<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        DsDataType,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Add>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                        b_device_buf.GetDeviceBuffer(),
+                                        std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                        e_device_buf.GetDeviceBuffer(),
+                                        a_gs_ms_ks_lengths,
+                                        a_gs_ms_ks_strides,
+                                        b_gs_ns_ks_lengths,
+                                        b_gs_ns_ks_strides,
+                                        std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
+                                        std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
+                                        e_gs_ms_ns_lengths,
+                                        e_gs_ms_ns_strides,
+                                        a_element_op,
+                                        b_element_op,
+                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_gs_ms_ns_lengths.begin() + NumDimG, NumDimM, 1, std::multiplies<>{});
+
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+            std::size_t flop      = std::size_t(2) * M * N * K;
+            std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                    sizeof(DDataType) * M * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return 0;
+}
diff --git a/client_example/04_contraction/contraction_scale_fp32.cpp b/client_example/04_contraction/contraction_scale_fp32.cpp
new file mode 100644
index 0000000000..ba7b0633c3
--- /dev/null
+++ b/client_example/04_contraction/contraction_scale_fp32.cpp
@@ -0,0 +1,222 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <numeric>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/contraction_scale.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Scale;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F32;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+
+    float scale = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 20)
+    {
+        const ck::index_t M0 = std::stoi(argv[1]);
+        const ck::index_t M1 = std::stoi(argv[2]);
+
+        const ck::index_t N0 = std::stoi(argv[3]);
+        const ck::index_t N1 = std::stoi(argv[4]);
+
+        const ck::index_t K0 = std::stoi(argv[5]);
+        const ck::index_t K1 = std::stoi(argv[6]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[7]), std::stoi(argv[8]), std::stoi(argv[9]), std::stoi(argv[10])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13]), std::stoi(argv[14])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17]), std::stoi(argv[18])};
+
+        scale = std::stof(argv[19]);
+    }
+    else
+    {
+        printf("arg1 to 6: M0, M1, N0, N1, K0, K1\n");
+        printf("arg7 to 10: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg11 to 14: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg15 to 18: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg19: scale\n");
+        exit(0);
+    }
+
+    auto f_tensor_space_size = [](auto lengths, auto strides) {
+        std::size_t space_size = 1;
+        for(std::size_t i = 0; i < lengths.size(); ++i)
+        {
+            space_size += (lengths[i] - 1) * strides[i];
+        }
+        return space_size;
+    };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) *
+                                 f_tensor_space_size(a_ms_ks_lengths, a_ms_ks_strides));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) *
+                                 f_tensor_space_size(b_ns_ks_lengths, b_ns_ks_strides));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) *
+                                 f_tensor_space_size(e_ms_ns_lengths, e_ms_ns_strides));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        ck::Tuple<>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Scale>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{scale};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        std::array<const void*, 0>{},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        a_ms_ks_lengths,
+                                                        a_ms_ks_strides,
+                                                        b_ns_ks_lengths,
+                                                        b_ns_ks_strides,
+                                                        std::array<std::vector<ck::index_t>, 0>{},
+                                                        std::array<std::vector<ck::index_t>, 0>{},
+                                                        e_ms_ns_lengths,
+                                                        e_ms_ns_strides,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return 0;
+}
diff --git a/client_example/04_contraction/contraction_scale_fp64.cpp b/client_example/04_contraction/contraction_scale_fp64.cpp
new file mode 100644
index 0000000000..24e52eb5aa
--- /dev/null
+++ b/client_example/04_contraction/contraction_scale_fp64.cpp
@@ -0,0 +1,270 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <numeric>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/contraction_scale.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+using F64 = double;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Scale;
+
+using ADataType        = F64;
+using BDataType        = F64;
+using AccDataType      = F64;
+using CShuffleDataType = F64;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F64;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+// kkn
+#if 1
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// knn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{64, 1, 131072, 2048};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// mkn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{128, 1, 245760, 3840};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+// mnn
+#elif 0
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{128, 1, 245760, 3840};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{64, 1, 131072, 2048};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+#endif
+
+    float scale = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 20)
+    {
+        const ck::index_t M0 = std::stoi(argv[1]);
+        const ck::index_t M1 = std::stoi(argv[2]);
+
+        const ck::index_t N0 = std::stoi(argv[3]);
+        const ck::index_t N1 = std::stoi(argv[4]);
+
+        const ck::index_t K0 = std::stoi(argv[5]);
+        const ck::index_t K1 = std::stoi(argv[6]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[7]), std::stoi(argv[8]), std::stoi(argv[9]), std::stoi(argv[10])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13]), std::stoi(argv[14])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17]), std::stoi(argv[18])};
+
+        scale = std::stof(argv[19]);
+    }
+    else
+    {
+        printf("arg1 to 6: M0, M1, N0, N1, K0, K1\n");
+        printf("arg7 to 10: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg11 to 14: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg15 to 18: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg19: scale\n");
+        exit(0);
+    }
+
+    auto f_tensor_space_size = [](auto lengths, auto strides) {
+        std::size_t space_size = 1;
+        for(std::size_t i = 0; i < lengths.size(); ++i)
+        {
+            space_size += (lengths[i] - 1) * strides[i];
+        }
+        return space_size;
+    };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) *
+                                 f_tensor_space_size(a_ms_ks_lengths, a_ms_ks_strides));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) *
+                                 f_tensor_space_size(b_ns_ks_lengths, b_ns_ks_strides));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) *
+                                 f_tensor_space_size(e_ms_ns_lengths, e_ms_ns_strides));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        ck::Tuple<>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Scale>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{scale};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        std::array<const void*, 0>{},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        a_ms_ks_lengths,
+                                                        a_ms_ks_strides,
+                                                        b_ns_ks_lengths,
+                                                        b_ns_ks_strides,
+                                                        std::array<std::vector<ck::index_t>, 0>{},
+                                                        std::array<std::vector<ck::index_t>, 0>{},
+                                                        e_ms_ns_lengths,
+                                                        e_ms_ns_strides,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            ck::index_t M = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+            ck::index_t N = ck::accumulate_n<ck::index_t>(
+                e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+            ck::index_t K = ck::accumulate_n<ck::index_t>(
+                a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return 0;
+}
diff --git a/client_example/05_layernorm/CMakeLists.txt b/client_example/05_layernorm/CMakeLists.txt
new file mode 100644
index 0000000000..b582b485d4
--- /dev/null
+++ b/client_example/05_layernorm/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_layernorm2d layernorm2d.cpp)
+target_link_libraries(client_layernorm2d PRIVATE composable_kernel::device_operations)
diff --git a/client_example/05_layernorm/layernorm2d.cpp b/client_example/05_layernorm/layernorm2d.cpp
new file mode 100644
index 0000000000..3ee7cead7b
--- /dev/null
+++ b/client_example/05_layernorm/layernorm2d.cpp
@@ -0,0 +1,171 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/normalization.hpp"
+
+using XDataType       = ck::half_t;
+using GammaDataType   = ck::half_t;
+using BetaDataType    = ck::half_t;
+using YDataType       = ck::half_t;
+using ComputeDataType = float;
+using PassThrough     = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    ck::index_t M      = 1024;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = 1024;
+
+    auto xy_size = (M - 1) * Stride + N;
+
+    SimpleDeviceMem x_device_buf(sizeof(XDataType) * xy_size);
+    SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N);
+    SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N);
+    SimpleDeviceMem y_device_buf(sizeof(YDataType) * xy_size);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
+                                                                       GammaDataType,
+                                                                       BetaDataType,
+                                                                       ComputeDataType,
+                                                                       YDataType,
+                                                                       PassThrough,
+                                                                       Rank,
+                                                                       NumReduceDim>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer({M, N},      // lengths
+                                                        {Stride, 1}, // xStrides
+                                                        {0, 1},      // gammaStrides
+                                                        {0, 1},      // betaStrides
+                                                        {Stride, 1}, // yStrides
+                                                        {1},         // reduceDims
+                                                        1e-4,
+                                                        x_device_buf.GetDeviceBuffer(),
+                                                        gamma_device_buf.GetDeviceBuffer(),
+                                                        beta_device_buf.GetDeviceBuffer(),
+                                                        y_device_buf.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        nullptr,
+                                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte = sizeof(XDataType) * M * N + sizeof(GammaDataType) * N +
+                                   sizeof(BetaDataType) * N + sizeof(YDataType) * M * N;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer({M, N},      // lengths
+                                                        {Stride, 1}, // xStrides
+                                                        {1},         // gammaStrides
+                                                        {1},         // betaStrides
+                                                        {Stride, 1}, // yStrides
+                                                        {1},         // reduceDims
+                                                        1e-4,
+                                                        x_device_buf.GetDeviceBuffer(),
+                                                        gamma_device_buf.GetDeviceBuffer(),
+                                                        beta_device_buf.GetDeviceBuffer(),
+                                                        y_device_buf.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        nullptr,
+                                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/06_softmax/CMakeLists.txt b/client_example/06_softmax/CMakeLists.txt
new file mode 100644
index 0000000000..b38a0fd9e2
--- /dev/null
+++ b/client_example/06_softmax/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_softmax4d softmax4d.cpp)
+target_link_libraries(client_softmax4d PRIVATE composable_kernel::device_operations)
diff --git a/client_example/06_softmax/softmax4d.cpp b/client_example/06_softmax/softmax4d.cpp
new file mode 100644
index 0000000000..2ccad27a88
--- /dev/null
+++ b/client_example/06_softmax/softmax4d.cpp
@@ -0,0 +1,168 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/softmax.hpp"
+
+using InDataType  = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 4;
+constexpr int NumReduceDim = 2;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::vector<ck::index_t> in_lengths{2, 8, 128, 1024};
+    std::vector<ck::index_t> in_strides{8 * 128 * 1024, 128 * 1024, 1024, 1};
+    std::vector<ck::index_t> reduce_dims{2, 3};
+
+    ck::index_t num_elements =
+        std::accumulate(in_lengths.begin(), in_lengths.end(), 1, std::multiplies<ck::index_t>());
+
+    double alpha{2.0};
+    double beta{2.0};
+
+    SimpleDeviceMem in(sizeof(InDataType) * num_elements);
+    SimpleDeviceMem out(sizeof(OutDataType) * num_elements);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceSoftmax<InDataType,
+                                                                 AccDataType,
+                                                                 OutDataType,
+                                                                 PassThrough,
+                                                                 PassThrough,
+                                                                 Rank,
+                                                                 NumReduceDim>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    auto& generic_op_ptr = op_ptrs[0];
+
+    auto generic_argument_ptr = generic_op_ptr->MakeArgumentPointer(in_lengths,
+                                                                    in_strides,
+                                                                    reduce_dims,
+                                                                    alpha,
+                                                                    beta,
+                                                                    in.GetDeviceBuffer(),
+                                                                    out.GetDeviceBuffer(),
+                                                                    PassThrough{},
+                                                                    PassThrough{});
+
+    if(!generic_op_ptr->IsSupportedArgument(generic_argument_ptr.get()))
+    {
+        throw std::runtime_error(
+            "The generic kernel instance should be able to support any input shapes");
+    };
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(in_lengths,
+                                                        in_strides,
+                                                        reduce_dims,
+                                                        alpha,
+                                                        beta,
+                                                        in.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        PassThrough{},
+                                                        PassThrough{});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes = num_elements * sizeof(InDataType) +
+                                    (beta == 0.0f ? 1 : 2) * num_elements * sizeof(OutDataType);
+
+            float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in_lengths,
+                                                        in_strides,
+                                                        reduce_dims,
+                                                        alpha,
+                                                        beta,
+                                                        in.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        PassThrough{},
+                                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/07_grouped_convnd_fwd/CMakeLists.txt b/client_example/07_grouped_convnd_fwd/CMakeLists.txt
new file mode 100644
index 0000000000..fce7e91c1e
--- /dev/null
+++ b/client_example/07_grouped_convnd_fwd/CMakeLists.txt
@@ -0,0 +1,5 @@
+add_executable(client_grouped_conv2d_fwd grouped_conv2d_fwd.cpp)
+target_link_libraries(client_grouped_conv2d_fwd PRIVATE composable_kernel::device_operations)
+
+add_executable(client_grouped_conv1d_fwd grouped_conv1d_fwd.cpp)
+target_link_libraries(client_grouped_conv1d_fwd PRIVATE composable_kernel::device_operations)
diff --git a/client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp b/client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp
new file mode 100644
index 0000000000..70be0101c6
--- /dev/null
+++ b/client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp
@@ -0,0 +1,229 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout    = ck::tensor_layout::convolution::GNWC;
+using WeiLayout   = ck::tensor_layout::convolution::GKXC;
+using OutLayout   = ck::tensor_layout::convolution::GNWK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr ck::index_t NumDimSpatial = 1;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 256;
+static constexpr ck::index_t K             = 192;
+static constexpr ck::index_t C             = 192;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Wo            = 28;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, Wi, C};
+    std::array<ck::index_t, NumDimSpatial + 3> in_strides{0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, X, C};
+    std::array<ck::index_t, NumDimSpatial + 3> wei_strides{0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, Wo, K};
+    std::array<ck::index_t, NumDimSpatial + 3> out_strides{0, 0, 0, 1};
+
+    std::partial_sum(rbegin(in_lengths),
+                     std::prev(rend(in_lengths)),
+                     std::next(rbegin(in_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(wei_lengths),
+                     std::prev(rend(wei_lengths)),
+                     std::next(rbegin(wei_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(out_lengths),
+                     std::prev(rend(out_lengths)),
+                     std::next(rbegin(out_strides)),
+                     std::multiplies<>{});
+
+    // transpose GNWC/GKXC/GNWK to GNCW/GKCX/GNCW
+    std::rotate(rbegin(in_lengths),
+                std::next(rbegin(in_lengths)),
+                std::next(rbegin(in_lengths), NumDimSpatial + 1));
+    std::rotate(rbegin(in_strides),
+                std::next(rbegin(in_strides)),
+                std::next(rbegin(in_strides), NumDimSpatial + 1));
+    std::rotate(rbegin(wei_lengths),
+                std::next(rbegin(wei_lengths)),
+                std::next(rbegin(wei_lengths), NumDimSpatial + 1));
+    std::rotate(rbegin(wei_strides),
+                std::next(rbegin(wei_strides)),
+                std::next(rbegin(wei_strides), NumDimSpatial + 1));
+    std::rotate(rbegin(out_lengths),
+                std::next(rbegin(out_lengths)),
+                std::next(rbegin(out_lengths), NumDimSpatial + 1));
+    std::rotate(rbegin(out_strides),
+                std::next(rbegin(out_strides)),
+                std::next(rbegin(out_strides), NumDimSpatial + 1));
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * G * N * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * G * N * Wo * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                                                 InLayout,
+                                                                                 WeiLayout,
+                                                                                 ck::Tuple<>,
+                                                                                 OutLayout,
+                                                                                 InDataType,
+                                                                                 WeiDataType,
+                                                                                 ck::Tuple<>,
+                                                                                 OutDataType,
+                                                                                 PassThrough,
+                                                                                 PassThrough,
+                                                                                 PassThrough>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = std::size_t(2) * G * N * K * C * Wo * X;
+            std::size_t num_bytes = sizeof(InDataType) * G * N * Wi * C +
+                                    sizeof(WeiDataType) * G * K * X * C +
+                                    sizeof(OutDataType) * G * N * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+}
diff --git a/client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp b/client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp
new file mode 100644
index 0000000000..57a210fa1f
--- /dev/null
+++ b/client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp
@@ -0,0 +1,200 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout    = ck::tensor_layout::convolution::NHWGC;
+using WeiLayout   = ck::tensor_layout::convolution::GKYXC;
+using OutLayout   = ck::tensor_layout::convolution::NHWGK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 256; // batch size
+static constexpr ck::index_t K             = 64;  // output channel
+static constexpr ck::index_t C             = 32;  // input channel (per group)
+static constexpr ck::index_t Y             = 3;   // filter H
+static constexpr ck::index_t X             = 3;   // filter W
+static constexpr ck::index_t Hi            = 28;  // input H
+static constexpr ck::index_t Wi            = 28;  // input W
+static constexpr ck::index_t Ho            = 28;  // output H
+static constexpr ck::index_t Wo            = 28;  // output W
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
+    // However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 5> wei_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> wei_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                                                 InLayout,
+                                                                                 WeiLayout,
+                                                                                 ck::Tuple<>,
+                                                                                 OutLayout,
+                                                                                 InDataType,
+                                                                                 WeiDataType,
+                                                                                 ck::Tuple<>,
+                                                                                 OutDataType,
+                                                                                 PassThrough,
+                                                                                 PassThrough,
+                                                                                 PassThrough>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = std::size_t(2) * G * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes = sizeof(InDataType) * N * Hi * Wi * G * C +
+                                    sizeof(WeiDataType) * G * K * Y * X * C +
+                                    sizeof(OutDataType) * N * Ho * Wo * G * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+}
diff --git a/client_example/08_fused_attention/CMakeLists.txt b/client_example/08_fused_attention/CMakeLists.txt
new file mode 100644
index 0000000000..862b9ed5b7
--- /dev/null
+++ b/client_example/08_fused_attention/CMakeLists.txt
@@ -0,0 +1,5 @@
+add_executable(client_fused_attention fused_attention.cpp)
+target_link_libraries(client_fused_attention PRIVATE composable_kernel::device_operations)
+
+add_executable(client_fused_attention_bias fused_attention_bias.cpp)
+target_link_libraries(client_fused_attention_bias PRIVATE composable_kernel::device_operations)
diff --git a/client_example/08_fused_attention/fused_attention.cpp b/client_example/08_fused_attention/fused_attention.cpp
new file mode 100644
index 0000000000..df6bc11a70
--- /dev/null
+++ b/client_example/08_fused_attention/fused_attention.cpp
@@ -0,0 +1,213 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using AElementOp    = ck::tensor_operation::element_wise::PassThrough;
+using B0ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp    = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr static auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+using ADataType   = ck::half_t;
+using B0DataType  = ck::half_t;
+using B1DataType  = ck::half_t;
+using CDataType   = ck::half_t;
+using AccDataType = float;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    int G0 = 48;
+    int G1 = 16;
+    int M  = 1024;
+    int N  = 1024;
+    int K  = 64;
+    int O  = 64;
+
+    // A layout [G0, M, G1, K]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+    // B0 layout [G0, N, G1, K]
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+    // B1 layout [G0, N, G1, O]
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+    // C layout [G0, M, G1, O]
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * G0 * G1 * M * K);
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) * G0 * G1 * N * K);
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * G0 * G1 * O * N);
+    SimpleDeviceMem c_device_buf(sizeof(CDataType) * G0 * G1 * M * O);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                          1,
+                                                                          1,
+                                                                          1,
+                                                                          1,
+                                                                          ADataType,
+                                                                          B0DataType,
+                                                                          B1DataType,
+                                                                          CDataType,
+                                                                          ck::Tuple<>,
+                                                                          ck::Tuple<>,
+                                                                          AElementOp,
+                                                                          B0ElementOp,
+                                                                          Acc0ElementOp,
+                                                                          B1ElementOp,
+                                                                          CElementOp,
+                                                                          MaskingSpec>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b0_device_buf.GetDeviceBuffer(),
+                                                        b1_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        {}, // p_acc0_biases
+                                                        {}, // p_acc1_biases
+                                                        a_gs_ms_ks_lengths,
+                                                        a_gs_ms_ks_strides,
+                                                        b0_gs_ns_ks_lengths,
+                                                        b0_gs_ns_ks_strides,
+                                                        b1_gs_os_ns_lengths,
+                                                        b1_gs_os_ns_strides,
+                                                        c_gs_ms_os_lengths,
+                                                        c_gs_ms_os_strides,
+                                                        {}, // acc0_biases_gs_ms_ns_lengths
+                                                        {}, // acc0_biases_gs_ms_ns_strides
+                                                        {}, // acc1_biases_gs_ms_os_lengths
+                                                        {}, // acc1_biases_gs_ms_os_strides
+                                                        AElementOp{},
+                                                        B0ElementOp{},
+                                                        Acc0ElementOp{1 / sqrtf(K)},
+                                                        B1ElementOp{},
+                                                        CElementOp{});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * G0 * G1;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                                    G0 * G1;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best instance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b0_device_buf.GetDeviceBuffer(),
+                                                        b1_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        {}, // p_acc0_biases
+                                                        {}, // p_acc1_biases
+                                                        a_gs_ms_ks_lengths,
+                                                        a_gs_ms_ks_strides,
+                                                        b0_gs_ns_ks_lengths,
+                                                        b0_gs_ns_ks_strides,
+                                                        b1_gs_os_ns_lengths,
+                                                        b1_gs_os_ns_strides,
+                                                        c_gs_ms_os_lengths,
+                                                        c_gs_ms_os_strides,
+                                                        {}, // acc0_biases_gs_ms_ns_lengths
+                                                        {}, // acc0_biases_gs_ms_ns_strides
+                                                        {}, // acc1_biases_gs_ms_os_lengths
+                                                        {}, // acc1_biases_gs_ms_os_strides
+                                                        AElementOp{},
+                                                        B0ElementOp{},
+                                                        Acc0ElementOp{1 / sqrtf(K)},
+                                                        B1ElementOp{},
+                                                        CElementOp{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/08_fused_attention/fused_attention_bias.cpp b/client_example/08_fused_attention/fused_attention_bias.cpp
new file mode 100644
index 0000000000..6c9f3bc8f6
--- /dev/null
+++ b/client_example/08_fused_attention/fused_attention_bias.cpp
@@ -0,0 +1,226 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using AElementOp    = ck::tensor_operation::element_wise::PassThrough;
+using B0ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::ScaleAdd;
+using B1ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp    = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr static auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+using ADataType   = ck::half_t;
+using B0DataType  = ck::half_t;
+using B1DataType  = ck::half_t;
+using CDataType   = ck::half_t;
+using D0DataType  = ck::half_t;
+using AccDataType = float;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    int G0 = 48;
+    int G1 = 16;
+    int M  = 1024;
+    int N  = 1024;
+    int K  = 64;
+    int O  = 64;
+
+    // A layout [G0, M, G1, K]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+    // B0 layout [G0, N, G1, K]
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+    // B1 layout [G0, N, G1, O]
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+    // C layout [G0, M, G1, O]
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+    // D layout [G0, M, G1, N]
+    std::vector<ck::index_t> d0_gs_ms_ns_lengths{G0, G1, M, N};
+    std::vector<ck::index_t> d0_gs_ms_ns_strides{M * G1 * N, N, G1 * N, 1};
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * G0 * G1 * M * K);
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) * G0 * G1 * N * K);
+    SimpleDeviceMem d0_device_buf(sizeof(D0DataType) * G0 * G1 * M * N);
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * G0 * G1 * O * N);
+    SimpleDeviceMem c_device_buf(sizeof(CDataType) * G0 * G1 * M * O);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                          1,
+                                                                          1,
+                                                                          1,
+                                                                          1,
+                                                                          ADataType,
+                                                                          B0DataType,
+                                                                          B1DataType,
+                                                                          CDataType,
+                                                                          ck::Tuple<D0DataType>,
+                                                                          ck::Tuple<>,
+                                                                          AElementOp,
+                                                                          B0ElementOp,
+                                                                          Acc0ElementOp,
+                                                                          B1ElementOp,
+                                                                          CElementOp,
+                                                                          MaskingSpec>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b0_device_buf.GetDeviceBuffer(),
+            b1_device_buf.GetDeviceBuffer(),
+            c_device_buf.GetDeviceBuffer(),
+            std::array<void*, 1>{d0_device_buf.GetDeviceBuffer()}, // p_acc0_biases
+            {},                                                    // p_acc1_biases
+            a_gs_ms_ks_lengths,
+            a_gs_ms_ks_strides,
+            b0_gs_ns_ks_lengths,
+            b0_gs_ns_ks_strides,
+            b1_gs_os_ns_lengths,
+            b1_gs_os_ns_strides,
+            c_gs_ms_os_lengths,
+            c_gs_ms_os_strides,
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_strides}, // acc0_biases_gs_ms_ns_strides
+            {},                       // acc1_biases_gs_ms_os_lengths
+            {},                       // acc1_biases_gs_ms_os_strides
+            AElementOp{},
+            B0ElementOp{},
+            Acc0ElementOp{1 / sqrtf(K)},
+            B1ElementOp{},
+            CElementOp{});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * G0 * G1;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O +
+                                     sizeof(D0DataType) * M * N) *
+                                    G0 * G1;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best instance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b0_device_buf.GetDeviceBuffer(),
+            b1_device_buf.GetDeviceBuffer(),
+            c_device_buf.GetDeviceBuffer(),
+            std::array<void*, 1>{d0_device_buf.GetDeviceBuffer()}, // p_acc0_biases
+            {},                                                    // p_acc1_biases
+            a_gs_ms_ks_lengths,
+            a_gs_ms_ks_strides,
+            b0_gs_ns_ks_lengths,
+            b0_gs_ns_ks_strides,
+            b1_gs_os_ns_lengths,
+            b1_gs_os_ns_strides,
+            c_gs_ms_os_lengths,
+            c_gs_ms_os_strides,
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_strides}, // acc0_biases_gs_ms_ns_strides
+            {},                       // acc1_biases_gs_ms_os_lengths
+            {},                       // acc1_biases_gs_ms_os_strides
+            AElementOp{},
+            B0ElementOp{},
+            Acc0ElementOp{1 / sqrtf(K)},
+            B1ElementOp{},
+            CElementOp{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/09_quantization/CMakeLists.txt b/client_example/09_quantization/CMakeLists.txt
new file mode 100644
index 0000000000..ac11aad45d
--- /dev/null
+++ b/client_example/09_quantization/CMakeLists.txt
@@ -0,0 +1,22 @@
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+add_executable(client_conv2d_fwd_bias_tanh_perchannel_quantization conv2d_fwd_bias_tanh_perchannel_quantization.cpp)
+target_link_libraries(client_conv2d_fwd_bias_tanh_perchannel_quantization PRIVATE composable_kernel::device_operations)
+
+add_executable(client_conv2d_fwd_bias_relu_perchannel_quantization conv2d_fwd_bias_relu_perchannel_quantization.cpp)
+target_link_libraries(client_conv2d_fwd_bias_relu_perchannel_quantization PRIVATE composable_kernel::device_operations)
+
+add_executable(client_conv2d_fwd_bias_tanh_perlayer_quantization conv2d_fwd_bias_tanh_perlayer_quantization.cpp)
+target_link_libraries(client_conv2d_fwd_bias_tanh_perlayer_quantization PRIVATE composable_kernel::device_operations)
+
+add_executable(client_conv2d_fwd_bias_relu_perlayer_quantization conv2d_fwd_bias_relu_perlayer_quantization.cpp)
+target_link_libraries(client_conv2d_fwd_bias_relu_perlayer_quantization PRIVATE composable_kernel::device_operations)
+
+add_executable(client_conv2d_fwd_perchannel_quantization conv2d_fwd_perchannel_quantization.cpp)
+target_link_libraries(client_conv2d_fwd_perchannel_quantization PRIVATE composable_kernel::device_operations)
+
+add_executable(client_conv2d_fwd_perlayer_quantization conv2d_fwd_perlayer_quantization.cpp)
+target_link_libraries(client_conv2d_fwd_perlayer_quantization PRIVATE composable_kernel::device_operations)
+
+add_executable(client_gemm_quantization gemm_quantization.cpp)
+target_link_libraries(client_gemm_quantization PRIVATE composable_kernel::device_operations)
+endif()
diff --git a/client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp b/client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
new file mode 100644
index 0000000000..cd504e942e
--- /dev/null
+++ b/client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
@@ -0,0 +1,210 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using BiasDataType         = int32_t;
+using RequantScaleDataType = float;
+using OutDataType          = int8_t;
+
+using InLayout           = ck::tensor_layout::convolution::NHWGC;
+using WeiLayout          = ck::tensor_layout::convolution::GKYXC;
+using BiasLayout         = ck::tensor_layout::convolution::G_K;
+using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
+using OutLayout          = ck::tensor_layout::convolution::NHWGK;
+using PassThrough        = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp       = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 4;
+static constexpr ck::index_t N             = 4;  // batch size
+static constexpr ck::index_t K             = 32; // output channel
+static constexpr ck::index_t C             = 64; // input channel (per group)
+static constexpr ck::index_t Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Hi            = 71; // input H
+static constexpr ck::index_t Wi            = 71; // input W
+static constexpr ck::index_t Ho            = 36; // output H
+static constexpr ck::index_t Wo            = 36; // output W
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
+    // However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
+
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
+    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+        NumDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout, RequantScaleLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<BiasDataType, RequantScaleDataType>,
+        OutDataType,
+        PassThrough,
+        PassThrough,
+        OutElementOp>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                        wei.GetDeviceBuffer(),
+                                        {bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths, requant_scale_lengths},
+                                        {bias_strides, requant_scale_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{ActivationOp{}});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes =
+                G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
+                G * sizeof(BiasDataType) * K + G * sizeof(RequantScaleDataType) * K +
+                G * sizeof(OutDataType) * N * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    // run the best intance
+    if(best_op_id != -1)
+    {
+        std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+                  << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                        wei.GetDeviceBuffer(),
+                                        {bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths, requant_scale_lengths},
+                                        {bias_strides, requant_scale_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{ActivationOp{}});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp b/client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
new file mode 100644
index 0000000000..f4aa3666b1
--- /dev/null
+++ b/client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
@@ -0,0 +1,206 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType   = int8_t;
+using WeiDataType  = int8_t;
+using BiasDataType = int32_t;
+using OutDataType  = int8_t;
+
+using InLayout     = ck::tensor_layout::convolution::NHWGC;
+using WeiLayout    = ck::tensor_layout::convolution::GKYXC;
+using BiasLayout   = ck::tensor_layout::convolution::G_K;
+using OutLayout    = ck::tensor_layout::convolution::NHWGK;
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 4;
+static constexpr ck::index_t N             = 4;    // batch size
+static constexpr ck::index_t K             = 32;   // output channel
+static constexpr ck::index_t C             = 64;   // input channel (per group)
+static constexpr ck::index_t Y             = 3;    // filter H
+static constexpr ck::index_t X             = 3;    // filter W
+static constexpr ck::index_t Hi            = 71;   // input H
+static constexpr ck::index_t Wi            = 71;   // input W
+static constexpr ck::index_t Ho            = 36;   // output H
+static constexpr ck::index_t Wo            = 36;   // output W
+static constexpr float requant_scale       = 0.5f; // requantize qAcc to qz
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
+    // However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
+
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                                    InLayout,
+                                                                    WeiLayout,
+                                                                    ck::Tuple<BiasLayout>,
+                                                                    OutLayout,
+                                                                    InDataType,
+                                                                    WeiDataType,
+                                                                    ck::Tuple<BiasDataType>,
+                                                                    OutDataType,
+                                                                    PassThrough,
+                                                                    PassThrough,
+                                                                    OutElementOp>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                        wei.GetDeviceBuffer(),
+                                        {bias.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths},
+                                        {bias_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{requant_scale, ActivationOp{}});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes =
+                G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
+                G * sizeof(BiasDataType) * K + G * sizeof(OutDataType) * N * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    // run the best intance
+    if(best_op_id != -1)
+    {
+        std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+                  << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                        wei.GetDeviceBuffer(),
+                                        {bias.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths},
+                                        {bias_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{requant_scale, ActivationOp{}});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp b/client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp
new file mode 100644
index 0000000000..ebdbbf52c0
--- /dev/null
+++ b/client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp
@@ -0,0 +1,213 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using BiasDataType         = int32_t;
+using RequantScaleDataType = float;
+using OutDataType          = int8_t;
+
+using InLayout           = ck::tensor_layout::convolution::NHWGC;
+using WeiLayout          = ck::tensor_layout::convolution::GKYXC;
+using BiasLayout         = ck::tensor_layout::convolution::G_K;
+using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
+using OutLayout          = ck::tensor_layout::convolution::NHWGK;
+using PassThrough        = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp       = ck::tensor_operation::element_wise::TanH;
+using OutElementOp =
+    ck::tensor_operation::element_wise::Add_Mul2_Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 4;
+static constexpr ck::index_t N             = 4;    // batch size
+static constexpr ck::index_t K             = 32;   // output channel
+static constexpr ck::index_t C             = 64;   // input channel (per group)
+static constexpr ck::index_t Y             = 3;    // filter H
+static constexpr ck::index_t X             = 3;    // filter W
+static constexpr ck::index_t Hi            = 71;   // input H
+static constexpr ck::index_t Wi            = 71;   // input W
+static constexpr ck::index_t Ho            = 36;   // output H
+static constexpr ck::index_t Wo            = 36;   // output W
+static constexpr float sz_inv              = 0.5f; // inverse of scale_z
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
+    // However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
+
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
+    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+        NumDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout, RequantScaleLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<BiasDataType, RequantScaleDataType>,
+        OutDataType,
+        PassThrough,
+        PassThrough,
+        OutElementOp>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                        wei.GetDeviceBuffer(),
+                                        {bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths, requant_scale_lengths},
+                                        {bias_strides, requant_scale_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{sz_inv, ActivationOp{}});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes =
+                G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
+                G * sizeof(BiasDataType) * K + G * sizeof(RequantScaleDataType) * K +
+                G * sizeof(OutDataType) * N * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    // run the best intance
+    if(best_op_id != -1)
+    {
+        std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+                  << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                        wei.GetDeviceBuffer(),
+                                        {bias.GetDeviceBuffer(), requant_scale.GetDeviceBuffer()},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {bias_lengths, requant_scale_lengths},
+                                        {bias_strides, requant_scale_strides},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{sz_inv, ActivationOp{}});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp b/client_example/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp
new file mode 100644
index 0000000000..9d60baee06
--- /dev/null
+++ b/client_example/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp
@@ -0,0 +1,205 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType   = int8_t;
+using WeiDataType  = int8_t;
+using BiasDataType = int32_t;
+using OutDataType  = int8_t;
+
+using InLayout     = ck::tensor_layout::convolution::NHWGC;
+using WeiLayout    = ck::tensor_layout::convolution::GKYXC;
+using BiasLayout   = ck::tensor_layout::convolution::G_K;
+using OutLayout    = ck::tensor_layout::convolution::NHWGK;
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::TanH;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Mul_Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 4;
+static constexpr ck::index_t N             = 4;    // batch size
+static constexpr ck::index_t K             = 32;   // output channel
+static constexpr ck::index_t C             = 64;   // input channel (per group)
+static constexpr ck::index_t Y             = 3;    // filter H
+static constexpr ck::index_t X             = 3;    // filter W
+static constexpr ck::index_t Hi            = 71;   // input H
+static constexpr ck::index_t Wi            = 71;   // input W
+static constexpr ck::index_t Ho            = 36;   // output H
+static constexpr ck::index_t Wo            = 36;   // output W
+static constexpr float sacc                = 0.5f; //  scale of acc
+static constexpr float sz_inv              = 0.5f; // inverse of scale_z
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
+    // However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> bias_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> bias_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
+
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * G * K);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                                    InLayout,
+                                                                    WeiLayout,
+                                                                    ck::Tuple<BiasLayout>,
+                                                                    OutLayout,
+                                                                    InDataType,
+                                                                    WeiDataType,
+                                                                    ck::Tuple<BiasDataType>,
+                                                                    OutDataType,
+                                                                    PassThrough,
+                                                                    PassThrough,
+                                                                    OutElementOp>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {bias.GetDeviceBuffer()},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {bias_lengths},
+                                                        {bias_strides},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{sacc, sz_inv, ActivationOp{}});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes =
+                G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
+                G * sizeof(BiasDataType) * K + G * sizeof(OutDataType) * N * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    // run the best intance
+    if(best_op_id != -1)
+    {
+        std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+                  << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {bias.GetDeviceBuffer()},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {bias_lengths},
+                                                        {bias_strides},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{sacc, sz_inv, ActivationOp{}});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp b/client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
new file mode 100644
index 0000000000..dd81d9ee6b
--- /dev/null
+++ b/client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
@@ -0,0 +1,203 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using RequantScaleDataType = float;
+using OutDataType          = int8_t;
+
+using InLayout           = ck::tensor_layout::convolution::NHWGC;
+using WeiLayout          = ck::tensor_layout::convolution::GKYXC;
+using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
+using OutLayout          = ck::tensor_layout::convolution::NHWGK;
+using PassThrough        = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp       = PassThrough;
+using OutElementOp       = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 4;
+static constexpr ck::index_t N             = 4;  // batch size
+static constexpr ck::index_t K             = 32; // output channel
+static constexpr ck::index_t C             = 64; // input channel (per group)
+static constexpr ck::index_t Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Hi            = 71; // input H
+static constexpr ck::index_t Wi            = 71; // input W
+static constexpr ck::index_t Ho            = 36; // output H
+static constexpr ck::index_t Wo            = 36; // output W
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
+    // However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> requant_scale_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> requant_scale_strides{K, 0, 1, 0, 0};
+    std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
+
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem requant_scale(sizeof(RequantScaleDataType) * G * K);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                                    InLayout,
+                                                                    WeiLayout,
+                                                                    ck::Tuple<RequantScaleLayout>,
+                                                                    OutLayout,
+                                                                    InDataType,
+                                                                    WeiDataType,
+                                                                    ck::Tuple<RequantScaleDataType>,
+                                                                    OutDataType,
+                                                                    PassThrough,
+                                                                    PassThrough,
+                                                                    OutElementOp>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {requant_scale.GetDeviceBuffer()},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {requant_scale_lengths},
+                                                        {requant_scale_strides},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{ActivationOp{}});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes =
+                G * sizeof(InDataType) * N * Hi * Wi * C + G * sizeof(WeiDataType) * K * Y * X * C +
+                G * sizeof(RequantScaleDataType) * K + G * sizeof(OutDataType) * N * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    // run the best intance
+    if(best_op_id != -1)
+    {
+        std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+                  << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        out.GetDeviceBuffer(),
+                                                        in_lengths,
+                                                        in_strides,
+                                                        weight_lengths,
+                                                        weight_strides,
+                                                        {},
+                                                        {},
+                                                        out_lengths,
+                                                        out_strides,
+                                                        conv_strides,
+                                                        conv_dilations,
+                                                        in_left_pad,
+                                                        in_right_pad,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        OutElementOp{ActivationOp{}});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp b/client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
new file mode 100644
index 0000000000..9c088a21d3
--- /dev/null
+++ b/client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
@@ -0,0 +1,199 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType  = int8_t;
+using WeiDataType = int8_t;
+using OutDataType = int8_t;
+
+using InLayout     = ck::tensor_layout::convolution::NHWGC;
+using WeiLayout    = ck::tensor_layout::convolution::GKYXC;
+using OutLayout    = ck::tensor_layout::convolution::NHWGK;
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 4;
+static constexpr ck::index_t N             = 4;    // batch size
+static constexpr ck::index_t K             = 32;   // output channel
+static constexpr ck::index_t C             = 64;   // input channel (per group)
+static constexpr ck::index_t Y             = 3;    // filter H
+static constexpr ck::index_t X             = 3;    // filter W
+static constexpr ck::index_t Hi            = 71;   // input H
+static constexpr ck::index_t Wi            = 71;   // input W
+static constexpr ck::index_t Ho            = 36;   // output H
+static constexpr ck::index_t Wo            = 36;   // output W
+static constexpr float requant_scale       = 0.5f; // requantize qAcc to qY
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // We have NHWGC/GKYXC/NHWGK (x, weight, y) in memory space
+    // However, CK's API only accept length and stride with order of GNCHW/GKCYX/GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_lengths{G, N, C, Hi, Wi};
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 5> weight_lengths{G, K, C, Y, X};
+    std::array<ck::index_t, 5> weight_strides{K * Y * X * C, Y * X * C, 1, X * C, C};
+    std::array<ck::index_t, 5> out_lengths{G, N, K, Ho, Wo};
+    std::array<ck::index_t, 5> out_strides{C, Ho * Wo * G * C, 1, Wo * G * C, G * C};
+
+    std::array<ck::index_t, 2> in_left_pad{1, 1};
+    std::array<ck::index_t, 2> in_right_pad{1, 1};
+    std::array<ck::index_t, 2> conv_strides{2, 2};
+    std::array<ck::index_t, 2> conv_dilations{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * G * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                                                 InLayout,
+                                                                                 WeiLayout,
+                                                                                 ck::Tuple<>,
+                                                                                 OutLayout,
+                                                                                 InDataType,
+                                                                                 WeiDataType,
+                                                                                 ck::Tuple<>,
+                                                                                 OutDataType,
+                                                                                 PassThrough,
+                                                                                 PassThrough,
+                                                                                 OutElementOp>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                        wei.GetDeviceBuffer(),
+                                        {},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {},
+                                        {},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{requant_scale, ActivationOp{}});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = G * 2 * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes = G * sizeof(InDataType) * N * Hi * Wi * C +
+                                    G * sizeof(WeiDataType) * K * Y * X * C +
+                                    G * sizeof(OutDataType) * N * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id != -1)
+    {
+        std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+                  << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                        wei.GetDeviceBuffer(),
+                                        {},
+                                        out.GetDeviceBuffer(),
+                                        in_lengths,
+                                        in_strides,
+                                        weight_lengths,
+                                        weight_strides,
+                                        {},
+                                        {},
+                                        out_lengths,
+                                        out_strides,
+                                        conv_strides,
+                                        conv_dilations,
+                                        in_left_pad,
+                                        in_right_pad,
+                                        PassThrough{},
+                                        PassThrough{},
+                                        OutElementOp{requant_scale, ActivationOp{}});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/client_example/09_quantization/gemm_quantization.cpp b/client_example/09_quantization/gemm_quantization.cpp
new file mode 100644
index 0000000000..b14e68fa08
--- /dev/null
+++ b/client_example/09_quantization/gemm_quantization.cpp
@@ -0,0 +1,193 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp"
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using ActivationOp = PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
+
+using ADataType = int8_t;
+using BDataType = int8_t;
+using EDataType = int8_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
+
+    float requant_scale = 0.03;
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
+                                                                       BLayout,
+                                                                       ck::Tuple<>,
+                                                                       ELayout,
+                                                                       ADataType,
+                                                                       BDataType,
+                                                                       ck::Tuple<>,
+                                                                       EDataType,
+                                                                       AElementOp,
+                                                                       BElementOp,
+                                                                       CDEElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        {},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        {},
+                                                        StrideE,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        cde_element_op);
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+            std::size_t num_bytes =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id != -1)
+    {
+        std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+                  << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        {},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        {},
+                                                        StrideE,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
\ No newline at end of file
diff --git a/client_example/10_grouped_conv2d_bwd_data/CMakeLists.txt b/client_example/10_grouped_conv2d_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000000..e564f3180d
--- /dev/null
+++ b/client_example/10_grouped_conv2d_bwd_data/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_grouped_conv2d_bwd_data grouped_conv2d_bwd_data.cpp)
+target_link_libraries(client_grouped_conv2d_bwd_data PRIVATE composable_kernel::device_operations)
diff --git a/client_example/10_grouped_conv2d_bwd_data/grouped_conv2d_bwd_data.cpp b/client_example/10_grouped_conv2d_bwd_data/grouped_conv2d_bwd_data.cpp
new file mode 100644
index 0000000000..1b2e8abc20
--- /dev/null
+++ b/client_example/10_grouped_conv2d_bwd_data/grouped_conv2d_bwd_data.cpp
@@ -0,0 +1,226 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout    = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout   = ck::tensor_layout::convolution::GKYXC;
+using OutLayout   = ck::tensor_layout::convolution::GNHWK;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 256;
+static constexpr ck::index_t K             = 192;
+static constexpr ck::index_t C             = 192;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 28;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    std::array<ck::index_t, NumDimSpatial + 3> in_lengths{G, N, Hi, Wi, C};
+    std::array<ck::index_t, NumDimSpatial + 3> in_strides{0, 0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 3> wei_lengths{G, K, Y, X, C};
+    std::array<ck::index_t, NumDimSpatial + 3> wei_strides{0, 0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 3> out_lengths{G, N, Ho, Wo, K};
+    std::array<ck::index_t, NumDimSpatial + 3> out_strides{0, 0, 0, 0, 1};
+
+    std::partial_sum(rbegin(in_lengths),
+                     std::prev(rend(in_lengths)),
+                     std::next(rbegin(in_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(wei_lengths),
+                     std::prev(rend(wei_lengths)),
+                     std::next(rbegin(wei_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(out_lengths),
+                     std::prev(rend(out_lengths)),
+                     std::next(rbegin(out_strides)),
+                     std::multiplies<>{});
+
+    // transpose GNHWC/GKYXC/GNHWK to GNCHW/GKCYX/GNCHW
+    std::rotate(
+        rbegin(in_lengths), std::next(rbegin(in_lengths)), std::next(rbegin(in_lengths), 3));
+    std::rotate(
+        rbegin(in_strides), std::next(rbegin(in_strides)), std::next(rbegin(in_strides), 3));
+    std::rotate(
+        rbegin(wei_lengths), std::next(rbegin(wei_lengths)), std::next(rbegin(wei_lengths), 3));
+    std::rotate(
+        rbegin(wei_strides), std::next(rbegin(wei_strides)), std::next(rbegin(wei_strides), 3));
+    std::rotate(
+        rbegin(out_lengths), std::next(rbegin(out_lengths)), std::next(rbegin(out_lengths), 3));
+    std::rotate(
+        rbegin(out_strides), std::next(rbegin(out_strides)), std::next(rbegin(out_strides), 3));
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * G * N * Hi * Wi * C);
+    SimpleDeviceMem wei(sizeof(WeiDataType) * G * K * Y * X * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * G * N * Ho * Wo * K);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
+                                                                                     OutLayout,
+                                                                                     WeiLayout,
+                                                                                     ck::Tuple<>,
+                                                                                     InLayout,
+                                                                                     OutDataType,
+                                                                                     WeiDataType,
+                                                                                     ck::Tuple<>,
+                                                                                     InDataType,
+                                                                                     PassThrough,
+                                                                                     PassThrough,
+                                                                                     PassThrough>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        in.GetDeviceBuffer(),
+                                                        out_lengths,
+                                                        out_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        in_lengths,
+                                                        in_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = std::size_t(2) * G * N * K * C * Ho * Wo * Y * X;
+            std::size_t num_bytes = sizeof(InDataType) * G * N * Hi * Wi * C +
+                                    sizeof(WeiDataType) * G * K * Y * X * C +
+                                    sizeof(OutDataType) * G * N * Ho * Wo * K;
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(out.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        {},
+                                                        in.GetDeviceBuffer(),
+                                                        out_lengths,
+                                                        out_strides,
+                                                        wei_lengths,
+                                                        wei_strides,
+                                                        {},
+                                                        {},
+                                                        in_lengths,
+                                                        in_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+}
diff --git a/client_example/11_grouped_conv_bwd_weight/CMakeLists.txt b/client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
new file mode 100644
index 0000000000..82162b606a
--- /dev/null
+++ b/client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
@@ -0,0 +1,9 @@
+add_executable(client_grouped_conv1d_bwd_weight_fp16 grouped_conv1d_bwd_weight_fp16.cpp)
+add_executable(client_grouped_conv2d_bwd_weight_fp16 grouped_conv2d_bwd_weight_fp16.cpp)
+add_executable(client_grouped_conv3d_bwd_weight_fp16 grouped_conv3d_bwd_weight_fp16.cpp)
+add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp)
+
+target_link_libraries(client_grouped_conv1d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_conv2d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_conv3d_bwd_weight_fp16 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_operations)
diff --git a/client_example/11_grouped_conv_bwd_weight/common.hpp b/client_example/11_grouped_conv_bwd_weight/common.hpp
new file mode 100644
index 0000000000..4292cded20
--- /dev/null
+++ b/client_example/11_grouped_conv_bwd_weight/common.hpp
@@ -0,0 +1,231 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+template <ck::index_t NumDimSpatial>
+std::size_t GetFlops(const std::array<ck::index_t, NumDimSpatial>& output_lengths,
+                     const std::array<ck::index_t, NumDimSpatial>& filter_lengths)
+{
+    constexpr ck::index_t spatial_offset = 3;
+    const auto C                         = filter_lengths[2];
+    // 2 * G * N * K * C * <output spatial lengths product> * <filter spatial lengths product>
+    return static_cast<std::size_t>(2) * C *
+           std::accumulate(std::begin(output_lengths),
+                           std::end(output_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>()) *
+           std::accumulate(std::begin(filter_lengths) + spatial_offset,
+                           std::end(filter_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>());
+}
+
+template <typename InDataType, ck::index_t NumDimSpatial>
+std::size_t GetInputByte(const std::array<ck::index_t, NumDimSpatial>& input_lengths)
+{
+    // sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
+    return sizeof(InDataType) * (std::accumulate(std::begin(input_lengths),
+                                                 std::end(input_lengths),
+                                                 static_cast<std::size_t>(1),
+                                                 std::multiplies<>()));
+}
+
+template <typename WeiDataType, ck::index_t NumDimSpatial>
+std::size_t GetWeightByte(const std::array<ck::index_t, NumDimSpatial>& filter_lengths)
+{
+    // sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
+    return sizeof(WeiDataType) * (std::accumulate(std::begin(filter_lengths),
+                                                  std::end(filter_lengths),
+                                                  static_cast<std::size_t>(1),
+                                                  std::multiplies<>()));
+}
+
+template <typename OutDataType, ck::index_t NumDimSpatial>
+std::size_t GetOutputByte(const std::array<ck::index_t, NumDimSpatial>& output_lengths)
+{
+    // sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
+    return sizeof(OutDataType) * (std::accumulate(std::begin(output_lengths),
+                                                  std::end(output_lengths),
+                                                  static_cast<std::size_t>(1),
+                                                  std::multiplies<std::size_t>()));
+}
+
+template <ck::index_t NumDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout>
+bool run_grouped_conv_bwd_weight(
+    const std::array<ck::index_t, NumDimSpatial + 3>& input_lengths,
+    const std::array<ck::index_t, NumDimSpatial + 3>& input_strides,
+    const std::array<ck::index_t, NumDimSpatial + 3>& filter_lengths,
+    const std::array<ck::index_t, NumDimSpatial + 3>& weights_strides,
+    const std::array<ck::index_t, NumDimSpatial + 3>& output_lengths,
+    const std::array<ck::index_t, NumDimSpatial + 3>& output_strides,
+    const std::array<ck::index_t, NumDimSpatial>& conv_filter_strides,
+    const std::array<ck::index_t, NumDimSpatial>& conv_filter_dilations,
+    const std::array<ck::index_t, NumDimSpatial>& input_left_pads,
+    const std::array<ck::index_t, NumDimSpatial>& input_right_pads)
+{
+
+    ck::index_t split_k = 2;
+    SimpleDeviceMem in(GetInputByte<InDataType, NumDimSpatial + 3>(input_lengths));
+    SimpleDeviceMem wei(GetWeightByte<WeiDataType, NumDimSpatial + 3>(filter_lengths));
+    SimpleDeviceMem out(GetOutputByte<OutDataType, NumDimSpatial + 3>(output_lengths));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdWeight<NumDimSpatial,
+                                                                              InLayout,
+                                                                              WeiLayout,
+                                                                              OutLayout,
+                                                                              InDataType,
+                                                                              WeiDataType,
+                                                                              OutDataType,
+                                                                              PassThrough,
+                                                                              PassThrough,
+                                                                              PassThrough>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    std::array<ck::index_t, NumDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NumDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NumDimSpatial + 3> b_g_k_c_xs_lengths{};
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        input_lengths,
+                                                        input_strides,
+                                                        filter_lengths,
+                                                        weights_strides,
+                                                        output_lengths,
+                                                        output_strides,
+                                                        conv_filter_strides,
+                                                        conv_filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        split_k);
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = GetFlops<NumDimSpatial + 3>(output_lengths, filter_lengths);
+            std::size_t num_bytes = GetInputByte<InDataType, NumDimSpatial + 3>(input_lengths) +
+                                    GetWeightByte<WeiDataType, NumDimSpatial + 3>(filter_lengths) +
+                                    GetOutputByte<OutDataType, NumDimSpatial + 3>(output_lengths);
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return false;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        input_lengths,
+                                                        input_strides,
+                                                        filter_lengths,
+                                                        weights_strides,
+                                                        output_lengths,
+                                                        output_strides,
+                                                        conv_filter_strides,
+                                                        conv_filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        split_k);
+        auto invoker_ptr  = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return true;
+}
diff --git a/client_example/11_grouped_conv_bwd_weight/grouped_conv1d_bwd_weight_fp16.cpp b/client_example/11_grouped_conv_bwd_weight/grouped_conv1d_bwd_weight_fp16.cpp
new file mode 100644
index 0000000000..e6d427faf4
--- /dev/null
+++ b/client_example/11_grouped_conv_bwd_weight/grouped_conv1d_bwd_weight_fp16.cpp
@@ -0,0 +1,56 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout  = ck::tensor_layout::convolution::GNWC;
+using WeiLayout = ck::tensor_layout::convolution::GKXC;
+using OutLayout = ck::tensor_layout::convolution::GNWK;
+
+static constexpr ck::index_t NumDimSpatial = 1;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 256;
+static constexpr ck::index_t K             = 192;
+static constexpr ck::index_t C             = 192;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Wo            = 28;
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Wi};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, X};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Wo};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{N * Wi * C, Wi* C, 1, C};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{K * X * C, X* C, 1, C};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{N * Wo * K, Wo* K, 1, K};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1};
+
+int main()
+{
+    return run_grouped_conv_bwd_weight<NumDimSpatial,
+                                       InDataType,
+                                       WeiDataType,
+                                       OutDataType,
+                                       InLayout,
+                                       WeiLayout,
+                                       OutLayout>(input_lengths,
+                                                  input_strides,
+                                                  filter_lengths,
+                                                  weights_strides,
+                                                  output_lengths,
+                                                  output_strides,
+                                                  conv_filter_strides,
+                                                  conv_filter_dilations,
+                                                  input_left_pads,
+                                                  input_right_pads)
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}
diff --git a/client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight_fp16.cpp b/client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight_fp16.cpp
new file mode 100644
index 0000000000..4201ea61b4
--- /dev/null
+++ b/client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight_fp16.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout  = ck::tensor_layout::convolution::GNHWC;
+using WeiLayout = ck::tensor_layout::convolution::GKYXC;
+using OutLayout = ck::tensor_layout::convolution::GNHWK;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 32;
+static constexpr ck::index_t N             = 256;
+static constexpr ck::index_t K             = 192;
+static constexpr ck::index_t C             = 192;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 28;
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Hi, Wi};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Y, X};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Ho, Wo};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
+    N * Hi * Wi * C, Hi* Wi* C, 1, Wi* C, C};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
+    K * Y * X * C, Y* X* C, 1, X* C, C};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
+    N * Ho * Wo * K, Ho* Wo* K, 1, Wo* K, K};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
+
+int main()
+{
+    return run_grouped_conv_bwd_weight<NumDimSpatial,
+                                       InDataType,
+                                       WeiDataType,
+                                       OutDataType,
+                                       InLayout,
+                                       WeiLayout,
+                                       OutLayout>(input_lengths,
+                                                  input_strides,
+                                                  filter_lengths,
+                                                  weights_strides,
+                                                  output_lengths,
+                                                  output_strides,
+                                                  conv_filter_strides,
+                                                  conv_filter_dilations,
+                                                  input_left_pads,
+                                                  input_right_pads)
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}
diff --git a/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16.cpp b/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16.cpp
new file mode 100644
index 0000000000..3ae46bcd55
--- /dev/null
+++ b/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16.cpp
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout  = ck::tensor_layout::convolution::GNDHWC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::GNDHWK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 8;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 128;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 3;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 3;
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Z, Y, X};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
+    N * Di * Hi * Wi * C, Di* Hi* Wi* C, 1, Hi* Wi* C, Wi* C, C};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
+    K * Z * Y * X * C, Z* Y* X* C, 1, Y* X* C, X* C, C};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
+    N * Do * Ho * Wo * K, Do* Ho* Wo* K, 1, Ho* Wo* K, Wo* K, K};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+int main()
+{
+    return run_grouped_conv_bwd_weight<NumDimSpatial,
+                                       InDataType,
+                                       WeiDataType,
+                                       OutDataType,
+                                       InLayout,
+                                       WeiLayout,
+                                       OutLayout>(input_lengths,
+                                                  input_strides,
+                                                  filter_lengths,
+                                                  weights_strides,
+                                                  output_lengths,
+                                                  output_strides,
+                                                  conv_filter_strides,
+                                                  conv_filter_dilations,
+                                                  input_left_pads,
+                                                  input_right_pads)
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}
diff --git a/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp32.cpp b/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp32.cpp
new file mode 100644
index 0000000000..2eb869f392
--- /dev/null
+++ b/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp32.cpp
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType  = float;
+using WeiDataType = float;
+using OutDataType = float;
+
+using InLayout  = ck::tensor_layout::convolution::GNDHWC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::GNDHWK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 8;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 128;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 3;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 3;
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Z, Y, X};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
+    N * Di * Hi * Wi * C, Di* Hi* Wi* C, 1, Hi* Wi* C, Wi* C, C};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
+    K * Z * Y * X * C, Z* Y* X* C, 1, Y* X* C, X* C, C};
+static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
+    N * Do * Ho * Wo * K, Do* Ho* Wo* K, 1, Ho* Wo* K, Wo* K, K};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
+static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
+
+int main()
+{
+    return run_grouped_conv_bwd_weight<NumDimSpatial,
+                                       InDataType,
+                                       WeiDataType,
+                                       OutDataType,
+                                       InLayout,
+                                       WeiLayout,
+                                       OutLayout>(input_lengths,
+                                                  input_strides,
+                                                  filter_lengths,
+                                                  weights_strides,
+                                                  output_lengths,
+                                                  output_strides,
+                                                  conv_filter_strides,
+                                                  conv_filter_dilations,
+                                                  input_left_pads,
+                                                  input_right_pads)
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}
diff --git a/client_example/12_elementwise_normalization/CMakeLists.txt b/client_example/12_elementwise_normalization/CMakeLists.txt
new file mode 100644
index 0000000000..1ba0e1279a
--- /dev/null
+++ b/client_example/12_elementwise_normalization/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_elementwise_layernorm2d elementwise_layernorm2d.cpp)
+target_link_libraries(client_elementwise_layernorm2d PRIVATE composable_kernel::device_operations)
diff --git a/client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp b/client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp
new file mode 100644
index 0000000000..bc4a6fe0bf
--- /dev/null
+++ b/client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp
@@ -0,0 +1,175 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp"
+
+using ADataType             = ck::half_t; // Input 1
+using BDataType             = ck::half_t; // Input 2
+using XDataType             = ck::half_t;
+using GammaDataType         = ck::half_t;
+using BetaDataType          = ck::half_t;
+using YDataType             = ck::half_t;
+using AccDataType           = float;
+using XElementwiseOperation = ck::tensor_operation::element_wise::Add;
+using YElementwiseOperation = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    bool time_kernel = true;
+
+    ck::index_t M      = 48 * 256;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = N;
+
+    auto mn_size = (M - 1) * Stride + N;
+
+    SimpleDeviceMem a_dev_buf(sizeof(ADataType) * mn_size);
+    SimpleDeviceMem b_dev_buf(sizeof(BDataType) * mn_size);
+    SimpleDeviceMem gamma_dev_buf(sizeof(GammaDataType) * N);
+    SimpleDeviceMem beta_dev_buf(sizeof(BetaDataType) * N);
+    SimpleDeviceMem y_dev_buf(sizeof(YDataType) * mn_size);
+
+    std::array<const void*, 2> ab_input               = {a_dev_buf.GetDeviceBuffer(),
+                                           b_dev_buf.GetDeviceBuffer()};
+    std::vector<ck::index_t> abStride                 = {Stride, 1};
+    std::array<std::vector<ck::index_t>, 2> abStrides = {abStride, abStride};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwiseNormalization<
+        ck::Tuple<ADataType, BDataType>,
+        GammaDataType,
+        BetaDataType,
+        AccDataType,
+        YDataType,
+        XElementwiseOperation,
+        YElementwiseOperation,
+        Rank,
+        NumReduceDim>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer({M, N}, // lengths
+                                                        abStrides,
+                                                        {0, 1},      // gammaStrides
+                                                        {0, 1},      // betaStrides
+                                                        {Stride, 1}, // yStrides
+                                                        {1},         // reduceDims
+                                                        1e-4,
+                                                        ab_input,
+                                                        gamma_dev_buf.GetDeviceBuffer(),
+                                                        beta_dev_buf.GetDeviceBuffer(),
+                                                        y_dev_buf.GetDeviceBuffer(),
+                                                        XElementwiseOperation{},
+                                                        YElementwiseOperation{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte = sizeof(ADataType) * M * N + sizeof(BDataType) * M * N +
+                                   sizeof(GammaDataType) * N + sizeof(BetaDataType) * N +
+                                   sizeof(YDataType) * M * N;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer({M, N}, // lengths
+                                                        abStrides,
+                                                        {1},         // gammaStrides
+                                                        {1},         // betaStrides
+                                                        {Stride, 1}, // yStrides
+                                                        {1},         // reduceDims
+                                                        1e-4,
+                                                        ab_input,
+                                                        gamma_dev_buf.GetDeviceBuffer(),
+                                                        beta_dev_buf.GetDeviceBuffer(),
+                                                        y_dev_buf.GetDeviceBuffer(),
+                                                        XElementwiseOperation{},
+                                                        YElementwiseOperation{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/13_batchnorm/CMakeLists.txt b/client_example/13_batchnorm/CMakeLists.txt
new file mode 100644
index 0000000000..fc4f9d395c
--- /dev/null
+++ b/client_example/13_batchnorm/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_executable(client_batchnorm_fwd_nhwc batchnorm_fwd_nhwc.cpp)
+add_executable(client_batchnorm_bwd_nhwc batchnorm_bwd_nhwc.cpp)
+add_executable(client_batchnorm_infer_nhwc batchnorm_infer_nhwc.cpp)
+target_link_libraries(client_batchnorm_fwd_nhwc PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_batchnorm_bwd_nhwc PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_batchnorm_infer_nhwc PRIVATE composable_kernel::device_operations)
diff --git a/client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp b/client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp
new file mode 100644
index 0000000000..1ed36e0f50
--- /dev/null
+++ b/client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp
@@ -0,0 +1,207 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp"
+
+using XDataType           = ck::half_t;
+using DxDataType          = float;
+using DyDataType          = float;
+using AccDataType         = float;
+using ScaleDataType       = ck::half_t;
+using DscaleDbiasDataType = float;
+using MeanVarDataType     = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank                  = 4;
+constexpr int NumBatchNormReduceDim = 3;
+
+const double epsilon = std::numeric_limits<float>::epsilon();
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
+    std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
+
+    ck::index_t numXYElement =
+        std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
+                                                             scaleBiasMeanVarLengths.end(),
+                                                             1,
+                                                             std::multiplies<ck::index_t>());
+
+    SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
+    SimpleDeviceMem dy(sizeof(DyDataType) * numXYElement);
+    SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem invVariance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem dx(sizeof(DxDataType) * numXYElement);
+    SimpleDeviceMem dscale(sizeof(DscaleDbiasDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem dbias(sizeof(DscaleDbiasDataType) * numScaleBiasMeanVarElement);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormBwd<XDataType,
+                                                                      DxDataType,
+                                                                      DyDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      DscaleDbiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThrough,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        dy.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        dx.GetDeviceBuffer(),
+                                                        dscale.GetDeviceBuffer(),
+                                                        dbias.GetDeviceBuffer());
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes =
+                numXYElement * (sizeof(XDataType) + sizeof(DyDataType) + sizeof(DxDataType)) +
+                numScaleBiasMeanVarElement *
+                    (sizeof(ScaleDataType) + sizeof(DscaleDbiasDataType) * 2 +
+                     sizeof(MeanVarDataType) * 2);
+
+            float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        // run the best intance
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        dy.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        dx.GetDeviceBuffer(),
+                                                        dscale.GetDeviceBuffer(),
+                                                        dbias.GetDeviceBuffer());
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp b/client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp
new file mode 100644
index 0000000000..f9af011c84
--- /dev/null
+++ b/client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp
@@ -0,0 +1,203 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp"
+
+using XDataType       = float;
+using YDataType       = float;
+using AccDataType     = float;
+using ScaleDataType   = AccDataType;
+using BiasDataType    = AccDataType;
+using MeanVarDataType = AccDataType;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank                  = 4;
+constexpr int NumBatchNormReduceDim = 3;
+
+const double epsilon       = std::numeric_limits<float>::epsilon();
+const double averageFactor = 0.1;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
+    std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
+
+    ck::index_t numXYElement =
+        std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
+                                                             scaleBiasMeanVarLengths.end(),
+                                                             1,
+                                                             std::multiplies<ck::index_t>());
+
+    SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
+    SimpleDeviceMem y(sizeof(YDataType) * numXYElement);
+    SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem invVariance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormFwd<XDataType,
+                                                                      YDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      BiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThrough,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        bias.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        y.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        averageFactor,
+                                                        nullptr,
+                                                        nullptr);
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes =
+                numXYElement * (sizeof(XDataType) + sizeof(YDataType)) +
+                numScaleBiasMeanVarElement * (sizeof(ScaleDataType) + sizeof(BiasDataType) +
+                                              sizeof(MeanVarDataType) + sizeof(MeanVarDataType));
+
+            float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        // run the best intance
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        bias.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        y.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        averageFactor,
+                                                        nullptr,
+                                                        nullptr);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/13_batchnorm/batchnorm_infer_nhwc.cpp b/client_example/13_batchnorm/batchnorm_infer_nhwc.cpp
new file mode 100644
index 0000000000..5e6627ce14
--- /dev/null
+++ b/client_example/13_batchnorm/batchnorm_infer_nhwc.cpp
@@ -0,0 +1,189 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp"
+
+using XDataType       = float;
+using YDataType       = float;
+using ScaleDataType   = float;
+using BiasDataType    = float;
+using MeanVarDataType = float;
+
+constexpr int Rank                  = 4;
+constexpr int NumBatchNormReduceDim = 3;
+
+using Normalize = ck::tensor_operation::element_wise::NormalizeInInfer;
+
+const double epsilon = std::numeric_limits<float>::epsilon();
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
+    std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
+    std::array<int, Rank - NumBatchNormReduceDim> invariantDims{3};
+
+    ck::index_t numXYElement =
+        std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
+                                                             scaleBiasMeanVarLengths.end(),
+                                                             1,
+                                                             std::multiplies<ck::index_t>());
+
+    SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
+    SimpleDeviceMem y(sizeof(YDataType) * numXYElement);
+    SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem variance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+
+    // values in variance need be non-negative
+    (void)hipMemset(
+        variance.GetDeviceBuffer(), 0, sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+
+    std::array<ck::index_t, Rank> aligned_scaleBiasMeanVarStrides{0};
+
+    int i = 0;
+    for(auto dim : invariantDims)
+    {
+        assert(xyLengths[dim] == scaleBiasMeanVarLengths[i]);
+
+        aligned_scaleBiasMeanVarStrides[dim] = scaleBiasMeanVarStrides[i];
+        i++;
+    };
+
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<XDataType, MeanVarDataType, MeanVarDataType, ScaleDataType, BiasDataType>,
+        ck::Tuple<YDataType>,
+        Normalize,
+        Rank>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                        {xyStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides},
+                                                        {xyStrides},
+                                                        {x.GetDeviceBuffer(),
+                                                         mean.GetDeviceBuffer(),
+                                                         variance.GetDeviceBuffer(),
+                                                         scale.GetDeviceBuffer(),
+                                                         bias.GetDeviceBuffer()},
+                                                        {y.GetDeviceBuffer()},
+                                                        Normalize{epsilon});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes =
+                numXYElement * (sizeof(XDataType) + sizeof(YDataType)) +
+                numScaleBiasMeanVarElement * (sizeof(ScaleDataType) + sizeof(BiasDataType) +
+                                              sizeof(MeanVarDataType) + sizeof(MeanVarDataType));
+
+            float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        // run the best intance
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                        {xyStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides,
+                                                         aligned_scaleBiasMeanVarStrides},
+                                                        {xyStrides},
+                                                        {x.GetDeviceBuffer(),
+                                                         mean.GetDeviceBuffer(),
+                                                         variance.GetDeviceBuffer(),
+                                                         scale.GetDeviceBuffer(),
+                                                         bias.GetDeviceBuffer()},
+                                                        {y.GetDeviceBuffer()},
+                                                        Normalize{epsilon});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/14_instance_id/CMakeLists.txt b/client_example/14_instance_id/CMakeLists.txt
new file mode 100644
index 0000000000..87b2a9a0cb
--- /dev/null
+++ b/client_example/14_instance_id/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_batchnorm_fwd_instance_id batchnorm_fwd_instance_id.cpp)
+target_link_libraries(client_batchnorm_fwd_instance_id PRIVATE composable_kernel::device_operations)
diff --git a/client_example/14_instance_id/batchnorm_fwd_instance_id.cpp b/client_example/14_instance_id/batchnorm_fwd_instance_id.cpp
new file mode 100644
index 0000000000..d45782d8e0
--- /dev/null
+++ b/client_example/14_instance_id/batchnorm_fwd_instance_id.cpp
@@ -0,0 +1,206 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp"
+
+using XDataType       = float;
+using YDataType       = float;
+using AccDataType     = float;
+using ScaleDataType   = AccDataType;
+using BiasDataType    = AccDataType;
+using MeanVarDataType = AccDataType;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank                  = 4;
+constexpr int NumBatchNormReduceDim = 3;
+
+const double epsilon       = std::numeric_limits<float>::epsilon();
+const double averageFactor = 0.1;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+// In the actual application, the instance index and name are usually from the perf db
+static int instance_index = -1;
+static std::string instance_name;
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> xyLengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> xyStrides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarLengths{256};
+    std::array<ck::index_t, Rank - NumBatchNormReduceDim> scaleBiasMeanVarStrides{1};
+    std::array<int, NumBatchNormReduceDim> reduceDims{0, 1, 2};
+
+    ck::index_t numXYElement =
+        std::accumulate(xyLengths.begin(), xyLengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t numScaleBiasMeanVarElement = std::accumulate(scaleBiasMeanVarLengths.begin(),
+                                                             scaleBiasMeanVarLengths.end(),
+                                                             1,
+                                                             std::multiplies<ck::index_t>());
+
+    SimpleDeviceMem x(sizeof(XDataType) * numXYElement);
+    SimpleDeviceMem y(sizeof(YDataType) * numXYElement);
+    SimpleDeviceMem scale(sizeof(ScaleDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem bias(sizeof(BiasDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem mean(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+    SimpleDeviceMem invVariance(sizeof(MeanVarDataType) * numScaleBiasMeanVarElement);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormFwd<XDataType,
+                                                                      YDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      BiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThrough,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    bool found          = false;
+    int best_op_index   = -1;
+    float best_ave_time = std::numeric_limits<float>::max();
+
+    // profile device operation instances and save the best performant instance index and instance
+    // name
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                        xyStrides,
+                                                        xyStrides,
+                                                        reduceDims,
+                                                        scaleBiasMeanVarLengths,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        scaleBiasMeanVarStrides,
+                                                        x.GetDeviceBuffer(),
+                                                        scale.GetDeviceBuffer(),
+                                                        bias.GetDeviceBuffer(),
+                                                        epsilon,
+                                                        PassThrough{},
+                                                        y.GetDeviceBuffer(),
+                                                        mean.GetDeviceBuffer(),
+                                                        invVariance.GetDeviceBuffer(),
+                                                        averageFactor,
+                                                        nullptr,
+                                                        nullptr);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            if(ave_time < best_ave_time)
+            {
+                found         = true;
+                best_op_index = i;
+                best_ave_time = ave_time;
+            }
+        }
+    }
+
+    if(found)
+    {
+        instance_index = best_op_index;
+        instance_name  = op_ptrs[instance_index]->GetTypeIdHashCode();
+    };
+
+    // simulate the execution of the operation when the instance index and name are available
+    const auto op_ptrs_2 = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    if(instance_index >= 0 && instance_index < op_ptrs_2.size())
+    {
+        auto& op_ptr = op_ptrs_2[instance_index];
+
+        if(op_ptr->GetTypeIdHashCode() == instance_name)
+        {
+
+            auto argument_ptr = op_ptr->MakeArgumentPointer(xyLengths,
+                                                            xyStrides,
+                                                            xyStrides,
+                                                            reduceDims,
+                                                            scaleBiasMeanVarLengths,
+                                                            scaleBiasMeanVarStrides,
+                                                            scaleBiasMeanVarStrides,
+                                                            scaleBiasMeanVarStrides,
+                                                            x.GetDeviceBuffer(),
+                                                            scale.GetDeviceBuffer(),
+                                                            bias.GetDeviceBuffer(),
+                                                            epsilon,
+                                                            PassThrough{},
+                                                            y.GetDeviceBuffer(),
+                                                            mean.GetDeviceBuffer(),
+                                                            invVariance.GetDeviceBuffer(),
+                                                            averageFactor,
+                                                            nullptr,
+                                                            nullptr);
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+                size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+                SimpleDeviceMem workspace(workspace_sz);
+
+                op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+                float exec_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+                size_t num_bytes = numXYElement * (sizeof(XDataType) + sizeof(YDataType)) +
+                                   numScaleBiasMeanVarElement *
+                                       (sizeof(ScaleDataType) + sizeof(BiasDataType) +
+                                        sizeof(MeanVarDataType) + sizeof(MeanVarDataType));
+
+                float gb_per_sec = num_bytes / 1.E6 / exec_time;
+
+                std::cout << "Kernel execution time: " << std::setw(10) << exec_time
+                          << " ms,  effective data transfer bandwidth: " << gb_per_sec << " GB/s"
+                          << std::endl;
+            }
+        };
+    }
+
+    return 0;
+}
diff --git a/client_example/15_convnd_bwd_data/CMakeLists.txt b/client_example/15_convnd_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000000..8a60a71674
--- /dev/null
+++ b/client_example/15_convnd_bwd_data/CMakeLists.txt
@@ -0,0 +1,5 @@
+add_executable(client_conv3d_bwd_data_fp16 conv3d_bwd_data_fp16.cpp)
+add_executable(client_conv3d_bwd_data_fp32 conv3d_bwd_data_fp32.cpp)
+
+target_link_libraries(client_conv3d_bwd_data_fp16 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_conv3d_bwd_data_fp32 PRIVATE composable_kernel::device_operations)
diff --git a/client_example/15_convnd_bwd_data/common.hpp b/client_example/15_convnd_bwd_data/common.hpp
new file mode 100644
index 0000000000..9799fb73a5
--- /dev/null
+++ b/client_example/15_convnd_bwd_data/common.hpp
@@ -0,0 +1,233 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <string>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+std::size_t GetFlops(ck::index_t N,
+                     ck::index_t K,
+                     ck::index_t C,
+                     const std::vector<ck::index_t>& output_spatial_lengths,
+                     const std::vector<ck::index_t>& weights_spatial_lengths)
+{
+    // 2 * N * K * C * <output spatial lengths product> * <filter spatial lengths product>
+
+    return static_cast<std::size_t>(2) * N * K * C *
+           std::accumulate(std::begin(output_spatial_lengths),
+                           std::end(output_spatial_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>()) *
+           std::accumulate(std::begin(weights_spatial_lengths),
+                           std::end(weights_spatial_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>());
+}
+
+template <typename InDataType>
+std::size_t
+GetInputByte(ck::index_t N, ck::index_t C, const std::vector<ck::index_t>& input_spatial_lengths)
+{
+    // sizeof(InDataType) * (N * C * <input spatial lengths product>) +
+    return sizeof(InDataType) * N * C *
+           std::accumulate(std::begin(input_spatial_lengths),
+                           std::end(input_spatial_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>());
+}
+
+template <typename WeiDataType>
+std::size_t
+GetWeightByte(ck::index_t K, ck::index_t C, const std::vector<ck::index_t>& weights_spatial_lengths)
+{
+    // sizeof(WeiDataType) * (K * C * <filter spatial lengths product>) +
+    return sizeof(WeiDataType) * K * C *
+           std::accumulate(std::begin(weights_spatial_lengths),
+                           std::end(weights_spatial_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>());
+}
+
+template <typename OutDataType>
+std::size_t
+GetOutputByte(ck::index_t N, ck::index_t K, const std::vector<ck::index_t>& output_spatial_lengths)
+{
+    // sizeof(OutDataType) * (N * K * <output spatial lengths product>);
+    return sizeof(OutDataType) * N * K *
+           std::accumulate(std::begin(output_spatial_lengths),
+                           std::end(output_spatial_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<std::size_t>());
+}
+
+template <ck::index_t NumDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout>
+bool run_conv_bwd_data(ck::index_t N,
+                       ck::index_t K,
+                       ck::index_t C,
+                       const std::vector<ck::index_t>& in_spatial_lengths,
+                       const std::vector<ck::index_t>& wei_spatial_lengths,
+                       const std::vector<ck::index_t>& out_spatial_lengths)
+{
+    std::size_t in_mem_size  = GetInputByte<InDataType>(N, C, in_spatial_lengths);
+    std::size_t wei_mem_size = GetWeightByte<WeiDataType>(K, C, wei_spatial_lengths);
+    std::size_t out_mem_size = GetOutputByte<OutDataType>(N, K, out_spatial_lengths);
+
+    SimpleDeviceMem in(in_mem_size);
+    SimpleDeviceMem wei(wei_mem_size);
+    SimpleDeviceMem out(out_mem_size);
+
+    std::vector<ck::index_t> filter_strides(NumDimSpatial, 1);
+    std::vector<ck::index_t> filter_dilations(NumDimSpatial, 1);
+    std::vector<ck::index_t> input_left_pads(NumDimSpatial, 1);
+    std::vector<ck::index_t> input_right_pads(NumDimSpatial, 1);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceConvBwdData<NumDimSpatial,
+                                                                     InLayout,
+                                                                     WeiLayout,
+                                                                     OutLayout,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     PassThrough,
+                                                                     PassThrough,
+                                                                     PassThrough>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    std::size_t flop      = GetFlops(N, K, C, out_spatial_lengths, wei_spatial_lengths);
+    std::size_t num_bytes = in_mem_size + wei_mem_size + out_mem_size;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        N,
+                                                        K,
+                                                        C,
+                                                        in_spatial_lengths,
+                                                        wei_spatial_lengths,
+                                                        out_spatial_lengths,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return false;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        wei.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        N,
+                                                        K,
+                                                        C,
+                                                        in_spatial_lengths,
+                                                        wei_spatial_lengths,
+                                                        out_spatial_lengths,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        PassThrough{},
+                                                        PassThrough{},
+                                                        PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+    return true;
+}
diff --git a/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp16.cpp b/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp16.cpp
new file mode 100644
index 0000000000..5210567241
--- /dev/null
+++ b/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp16.cpp
@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWC;
+using WeiLayout = ck::tensor_layout::convolution::KZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 64;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 28;
+
+int main()
+{
+    return run_conv_bwd_data<NumDimSpatial,
+                             InDataType,
+                             WeiDataType,
+                             OutDataType,
+                             InLayout,
+                             WeiLayout,
+                             OutLayout>(N, K, C, {Di, Hi, Wi}, {Z, Y, X}, {Do, Ho, Wo})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}
diff --git a/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp32.cpp b/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp32.cpp
new file mode 100644
index 0000000000..441bdfe7be
--- /dev/null
+++ b/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp32.cpp
@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType  = float;
+using WeiDataType = float;
+using OutDataType = float;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWC;
+using WeiLayout = ck::tensor_layout::convolution::KZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 64;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 28;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 28;
+
+int main()
+{
+    return run_conv_bwd_data<NumDimSpatial,
+                             InDataType,
+                             WeiDataType,
+                             OutDataType,
+                             InLayout,
+                             WeiLayout,
+                             OutLayout>(N, K, C, {Di, Hi, Wi}, {Z, Y, X}, {Do, Ho, Wo})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}
diff --git a/client_example/15_gemm_add_multiply/CMakeLists.txt b/client_example/15_gemm_add_multiply/CMakeLists.txt
new file mode 100644
index 0000000000..fd2dcf9614
--- /dev/null
+++ b/client_example/15_gemm_add_multiply/CMakeLists.txt
@@ -0,0 +1,3 @@
+
+add_executable(client_gemm_add_multiply gemm_add_multiply.cpp)
+target_link_libraries(client_gemm_add_multiply PRIVATE composable_kernel::device_operations)
\ No newline at end of file
diff --git a/client_example/15_gemm_add_multiply/gemm_add_multiply.cpp b/client_example/15_gemm_add_multiply/gemm_add_multiply.cpp
new file mode 100644
index 0000000000..c74d7c6bd8
--- /dev/null
+++ b/client_example/15_gemm_add_multiply/gemm_add_multiply.cpp
@@ -0,0 +1,241 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddMultiply = ck::tensor_operation::element_wise::AddMultiply;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddMultiply;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F16;
+using D1DataType = F16;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using ELayout  = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 4096;
+    ck::index_t StrideE  = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 9)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA  = std::stoi(argv[4]);
+        StrideB  = std::stoi(argv[5]);
+        StrideD0 = std::stoi(argv[6]);
+        StrideD1 = std::stoi(argv[7]);
+        StrideE  = std::stoi(argv[8]);
+    }
+    else
+    {
+        printf("arg1 to 8: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_m_n_device_buf(sizeof(D0DataType) *
+                                      f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem d1_m_n_device_buf(sizeof(D1DataType) *
+                                      f_matrix_space_size(M, N, StrideD1, D1Layout{}));
+    SimpleDeviceMem e_device_buf(sizeof(EDataType) * f_matrix_space_size(M, N, StrideE, ELayout{}));
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
+                                                          BLayout,
+                                                          ck::Tuple<D0Layout, D1Layout>,
+                                                          ELayout,
+                                                          ADataType,
+                                                          BDataType,
+                                                          ck::Tuple<D0DataType, D1DataType>,
+                                                          EDataType,
+                                                          AElementOp,
+                                                          BElementOp,
+                                                          CDEElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/15_reduce/CMakeLists.txt b/client_example/15_reduce/CMakeLists.txt
new file mode 100644
index 0000000000..d52675ba83
--- /dev/null
+++ b/client_example/15_reduce/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_reduce_nhwc_c reduce_nhwc_c.cpp)
+target_link_libraries(client_reduce_nhwc_c PRIVATE composable_kernel::device_operations)
diff --git a/client_example/15_reduce/reduce_nhwc_c.cpp b/client_example/15_reduce/reduce_nhwc_c.cpp
new file mode 100644
index 0000000000..b45b72f0de
--- /dev/null
+++ b/client_example/15_reduce/reduce_nhwc_c.cpp
@@ -0,0 +1,175 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <numeric>
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp"
+
+using InDataType  = float;
+using OutDataType = float;
+using AccDataType = float;
+using ReduceAdd   = ck::reduce::Add;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using UnaryDivide = ck::tensor_operation::element_wise::UnaryDivide;
+
+constexpr bool PropagateNan = false;
+constexpr bool OutputIndex  = false;
+
+constexpr int Rank         = 4;
+constexpr int NumReduceDim = 3;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    std::array<ck::index_t, Rank> in_lengths{16, 8, 128, 256};
+    std::array<ck::index_t, Rank> in_strides{8 * 128 * 256, 128 * 256, 256, 1};
+    std::array<ck::index_t, Rank - NumReduceDim> out_lengths{256};
+    std::array<ck::index_t, Rank - NumReduceDim> out_strides{1};
+    std::array<int, NumReduceDim> reduce_dims{0, 1, 2};
+
+    ck::index_t num_in_elements =
+        std::accumulate(in_lengths.begin(), in_lengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t num_out_elements =
+        std::accumulate(out_lengths.begin(), out_lengths.end(), 1, std::multiplies<ck::index_t>());
+
+    ck::index_t reduce_length = 1;
+
+    for(auto dim : reduce_dims)
+        reduce_length *= in_lengths[dim];
+
+    double alpha{1.0};
+    double beta{0.0};
+
+    SimpleDeviceMem in(sizeof(InDataType) * num_in_elements);
+    SimpleDeviceMem out(sizeof(OutDataType) * num_out_elements);
+
+    using DeviceOp     = ck::tensor_operation::device::DeviceReduce<InDataType,
+                                                                AccDataType,
+                                                                OutDataType,
+                                                                Rank,
+                                                                NumReduceDim,
+                                                                ReduceAdd,
+                                                                PassThrough,
+                                                                UnaryDivide,
+                                                                PropagateNan,
+                                                                OutputIndex>;
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(in_lengths,
+                                                        in_strides,
+                                                        out_lengths,
+                                                        out_strides,
+                                                        reduce_dims,
+                                                        alpha,
+                                                        beta,
+                                                        in.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        out.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        PassThrough{},
+                                                        UnaryDivide{reduce_length});
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes = num_in_elements * sizeof(InDataType) +
+                                    (beta == 0.0f ? 1 : 2) * num_out_elements * sizeof(OutDataType);
+
+            float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    if(found)
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in_lengths,
+                                                        in_strides,
+                                                        out_lengths,
+                                                        out_strides,
+                                                        reduce_dims,
+                                                        alpha,
+                                                        beta,
+                                                        in.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        out.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        PassThrough{},
+                                                        UnaryDivide{reduce_length});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/16_convnd_fwd/CMakeLists.txt b/client_example/16_convnd_fwd/CMakeLists.txt
new file mode 100644
index 0000000000..e2580a370c
--- /dev/null
+++ b/client_example/16_convnd_fwd/CMakeLists.txt
@@ -0,0 +1,5 @@
+add_executable(client_conv3d_fwd_fp16 conv3d_fwd_fp16.cpp)
+add_executable(client_conv3d_fwd_fp32 conv3d_fwd_fp32.cpp)
+
+target_link_libraries(client_conv3d_fwd_fp16 PRIVATE composable_kernel::device_operations)
+target_link_libraries(client_conv3d_fwd_fp32 PRIVATE composable_kernel::device_operations)
diff --git a/client_example/16_convnd_fwd/common.hpp b/client_example/16_convnd_fwd/common.hpp
new file mode 100644
index 0000000000..449c9466e8
--- /dev/null
+++ b/client_example/16_convnd_fwd/common.hpp
@@ -0,0 +1,300 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <string>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+template <ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetFlops(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths,
+         const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths)
+{
+    // 2 * G * N * K * C * <output spatial lengths product> * <filter spatial lengths product>
+    ck::index_t G = weights_lengths[0];
+    ck::index_t N = output_lengths[1];
+    ck::index_t K = weights_lengths[1];
+    ck::index_t C = weights_lengths[2];
+
+    return static_cast<std::size_t>(2) * G * N * K * C *
+           std::accumulate(std::next(std::begin(output_lengths), NumNonSpatialDim),
+                           std::end(output_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>()) *
+           std::accumulate(std::next(std::begin(weights_lengths), NumNonSpatialDim),
+                           std::end(weights_lengths),
+                           static_cast<std::size_t>(1),
+                           std::multiplies<>());
+}
+
+template <typename InDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetInputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& input_lengths)
+{
+    // sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
+    return sizeof(InDataType) * std::accumulate(std::begin(input_lengths),
+                                                std::end(input_lengths),
+                                                static_cast<std::size_t>(1),
+                                                std::multiplies<>());
+}
+
+template <typename WeiDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetWeightByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& weights_lengths)
+{
+    // sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
+    return sizeof(WeiDataType) * std::accumulate(std::begin(weights_lengths),
+                                                 std::end(weights_lengths),
+                                                 static_cast<std::size_t>(1),
+                                                 std::multiplies<>());
+}
+
+template <typename OutDataType, ck::index_t NumDimSpatial, ck::index_t NumNonSpatialDim = 3>
+std::size_t
+GetOutputByte(const std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>& output_lengths)
+{
+    // sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
+    return sizeof(OutDataType) * std::accumulate(std::begin(output_lengths),
+                                                 std::end(output_lengths),
+                                                 static_cast<std::size_t>(1),
+                                                 std::multiplies<std::size_t>());
+}
+
+template <ck::index_t NumDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          ck::index_t NumNonSpatialDim = 3>
+bool run_grouped_conv_fwd(std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_lengths,
+                          std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_lengths,
+                          std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_lengths)
+{
+    std::size_t in_mem_size  = GetInputByte<InDataType, NumDimSpatial>(in_lengths);
+    std::size_t wei_mem_size = GetWeightByte<WeiDataType, NumDimSpatial>(wei_lengths);
+    std::size_t out_mem_size = GetOutputByte<OutDataType, NumDimSpatial>(out_lengths);
+
+    SimpleDeviceMem in(in_mem_size);
+    SimpleDeviceMem wei(wei_mem_size);
+    SimpleDeviceMem out(out_mem_size);
+
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> in_strides;
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> wei_strides;
+    std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim> out_strides;
+    in_strides.fill(0);
+    wei_strides.fill(0);
+    out_strides.fill(0);
+    in_strides.back()  = 1;
+    wei_strides.back() = 1;
+    out_strides.back() = 1;
+
+    std::partial_sum(rbegin(in_lengths),
+                     std::prev(rend(in_lengths)),
+                     std::next(rbegin(in_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(wei_lengths),
+                     std::prev(rend(wei_lengths)),
+                     std::next(rbegin(wei_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(out_lengths),
+                     std::prev(rend(out_lengths)),
+                     std::next(rbegin(out_strides)),
+                     std::multiplies<>{});
+
+    // transpose NDHWGC/KZYXGC/NDHWGK to GNDHWC/GKZYXC/GNDHWK to GNCDHW/GKCZYX/GNKDHW
+    std::rotate(std::next(rbegin(in_lengths)), std::next(rbegin(in_lengths), 2), rend(in_lengths));
+    std::rotate(rbegin(in_lengths),
+                std::next(rbegin(in_lengths)),
+                std::next(rbegin(in_lengths), NumDimSpatial + 1));
+
+    std::rotate(std::next(rbegin(in_strides)), std::next(rbegin(in_strides), 2), rend(in_strides));
+    std::rotate(rbegin(in_strides),
+                std::next(rbegin(in_strides)),
+                std::next(rbegin(in_strides), NumDimSpatial + 1));
+
+    std::rotate(rbegin(wei_lengths),
+                std::next(rbegin(wei_lengths)),
+                std::next(rbegin(wei_lengths), NumDimSpatial + 1));
+
+    std::rotate(rbegin(wei_strides),
+                std::next(rbegin(wei_strides)),
+                std::next(rbegin(wei_strides), NumDimSpatial + 1));
+
+    std::rotate(
+        std::next(rbegin(out_lengths)), std::next(rbegin(out_lengths), 2), rend(out_lengths));
+    std::rotate(rbegin(out_lengths),
+                std::next(rbegin(out_lengths)),
+                std::next(rbegin(out_lengths), NumDimSpatial + 1));
+
+    std::rotate(
+        std::next(rbegin(out_strides)), std::next(rbegin(out_strides), 2), rend(out_strides));
+    std::rotate(rbegin(out_strides),
+                std::next(rbegin(out_strides)),
+                std::next(rbegin(out_strides), NumDimSpatial + 1));
+
+    std::array<ck::index_t, NumDimSpatial> conv_filter_strides;
+    std::array<ck::index_t, NumDimSpatial> conv_filter_dilations;
+    std::array<ck::index_t, NumDimSpatial> input_left_pads;
+    std::array<ck::index_t, NumDimSpatial> input_right_pads;
+    conv_filter_strides.fill(1);
+    conv_filter_dilations.fill(1);
+    input_left_pads.fill(1);
+    input_right_pads.fill(1);
+
+    std::size_t flop      = GetFlops<NumDimSpatial>(out_lengths, wei_lengths);
+    std::size_t num_bytes = in_mem_size + wei_mem_size + out_mem_size;
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                                                 InLayout,
+                                                                                 WeiLayout,
+                                                                                 ck::Tuple<>,
+                                                                                 OutLayout,
+                                                                                 InDataType,
+                                                                                 WeiDataType,
+                                                                                 ck::Tuple<>,
+                                                                                 OutDataType,
+                                                                                 PassThrough,
+                                                                                 PassThrough,
+                                                                                 PassThrough>;
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            in.GetDeviceBuffer(),
+            wei.GetDeviceBuffer(),
+            std::array<const void*, 0>{},
+            out.GetDeviceBuffer(),
+            in_lengths,
+            in_strides,
+            wei_lengths,
+            wei_strides,
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{{}},
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{{}},
+            out_lengths,
+            out_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            PassThrough{},
+            PassThrough{},
+            PassThrough{});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+                best_tflops     = tflops;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return false;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_tflops
+              << " TFlops, " << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            in.GetDeviceBuffer(),
+            wei.GetDeviceBuffer(),
+            std::array<const void*, 0>{},
+            out.GetDeviceBuffer(),
+            in_lengths,
+            in_strides,
+            wei_lengths,
+            wei_strides,
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{{}},
+            std::array<std::array<ck::index_t, NumDimSpatial + NumNonSpatialDim>, 0>{{}},
+            out_lengths,
+            out_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            PassThrough{},
+            PassThrough{},
+            PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+    return true;
+}
diff --git a/client_example/16_convnd_fwd/conv3d_fwd_fp16.cpp b/client_example/16_convnd_fwd/conv3d_fwd_fp16.cpp
new file mode 100644
index 0000000000..d4455df628
--- /dev/null
+++ b/client_example/16_convnd_fwd/conv3d_fwd_fp16.cpp
@@ -0,0 +1,44 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWGK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 64;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 3;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 3;
+
+int main()
+{
+    return run_grouped_conv_fwd<NumDimSpatial,
+                                InDataType,
+                                WeiDataType,
+                                OutDataType,
+                                InLayout,
+                                WeiLayout,
+                                OutLayout>(
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}
diff --git a/client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp b/client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp
new file mode 100644
index 0000000000..7e8c98b603
--- /dev/null
+++ b/client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp
@@ -0,0 +1,44 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType  = float;
+using WeiDataType = float;
+using OutDataType = float;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWGC;
+using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
+using OutLayout = ck::tensor_layout::convolution::NDHWGK;
+
+static constexpr ck::index_t NumDimSpatial = 3;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 64;
+static constexpr ck::index_t K             = 128;
+static constexpr ck::index_t C             = 64;
+static constexpr ck::index_t Z             = 3;
+static constexpr ck::index_t Y             = 3;
+static constexpr ck::index_t X             = 3;
+static constexpr ck::index_t Di            = 28;
+static constexpr ck::index_t Hi            = 28;
+static constexpr ck::index_t Wi            = 3;
+static constexpr ck::index_t Do            = 28;
+static constexpr ck::index_t Ho            = 28;
+static constexpr ck::index_t Wo            = 3;
+
+int main()
+{
+    return run_grouped_conv_fwd<NumDimSpatial,
+                                InDataType,
+                                WeiDataType,
+                                OutDataType,
+                                InLayout,
+                                WeiLayout,
+                                OutLayout>(
+               {N, Di, Hi, Wi, G, C}, {G, K, Z, Y, X, C}, {N, Do, Ho, Wo, G, K})
+               ? EXIT_SUCCESS
+               : EXIT_FAILURE;
+}
diff --git a/client_example/17_grouped_gemm_fastgelu/CMakeLists.txt b/client_example/17_grouped_gemm_fastgelu/CMakeLists.txt
new file mode 100644
index 0000000000..659e6769d8
--- /dev/null
+++ b/client_example/17_grouped_gemm_fastgelu/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_grouped_gemm_fastgelu grouped_gemm_fastgelu.cpp)
+target_link_libraries(client_grouped_gemm_fastgelu PRIVATE composable_kernel::device_operations)
\ No newline at end of file
diff --git a/client_example/17_grouped_gemm_fastgelu/grouped_gemm_fastgelu.cpp b/client_example/17_grouped_gemm_fastgelu/grouped_gemm_fastgelu.cpp
new file mode 100644
index 0000000000..7ba3224fc3
--- /dev/null
+++ b/client_example/17_grouped_gemm_fastgelu/grouped_gemm_fastgelu.cpp
@@ -0,0 +1,232 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+#include <random>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using DsDataType = ck::Tuple<>;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = FastGelu;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    std::mt19937 gen(19391);
+    std::uniform_int_distribution<> distrib(1, 10);
+    int group_count = distrib(gen);
+
+    std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        Ms.push_back(256 + 256 * distrib(gen));
+        Ns.push_back(256 + 256 * distrib(gen));
+        Ks.push_back(128 + 128 * distrib(gen));
+
+        StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
+        StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
+        StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, e_dev_bufs;
+
+    a_dev_bufs.reserve(group_count);
+    b_dev_bufs.reserve(group_count);
+    e_dev_bufs.reserve(group_count);
+
+    std::vector<const void*> p_a, p_b;
+    std::vector<void*> p_e;
+
+    p_a.reserve(group_count);
+    p_b.reserve(group_count);
+    p_e.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        a_dev_bufs.emplace_back(sizeof(ADataType) *
+                                f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
+        b_dev_bufs.emplace_back(sizeof(BDataType) *
+                                f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
+        e_dev_bufs.emplace_back(sizeof(EDataType) *
+                                f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
+
+        gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideEs[i], {}});
+
+        p_a.push_back(a_dev_bufs[i].GetDeviceBuffer());
+        p_b.push_back(b_dev_bufs[i].GetDeviceBuffer());
+        p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
+                                                                     BLayout,
+                                                                     DsLayout,
+                                                                     ELayout,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     DsDataType,
+                                                                     EDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     CDEElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    auto p_ds = std::vector<std::array<const void*, 0>>{};
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        SimpleDeviceMem gemm_desc_workspace(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = 0, num_btype = 0;
+            for(std::size_t j = 0; j < gemm_descs.size(); ++j)
+            {
+                flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
+
+                num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
+                             sizeof(EDataType) * Ms[j] * Ns[j];
+            }
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    if(found)
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        SimpleDeviceMem gemm_desc_workspace(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/18_groupnorm/CMakeLists.txt b/client_example/18_groupnorm/CMakeLists.txt
new file mode 100644
index 0000000000..17c88cb61b
--- /dev/null
+++ b/client_example/18_groupnorm/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_groupnorm_swish groupnorm_swish.cpp)
+target_link_libraries(client_groupnorm_swish PRIVATE composable_kernel::device_operations)
diff --git a/client_example/18_groupnorm/groupnorm_swish.cpp b/client_example/18_groupnorm/groupnorm_swish.cpp
new file mode 100644
index 0000000000..df0a9ceec6
--- /dev/null
+++ b/client_example/18_groupnorm/groupnorm_swish.cpp
@@ -0,0 +1,202 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/normalization_swish.hpp"
+
+using XDataType       = ck::half_t;
+using GammaDataType   = float;
+using BetaDataType    = float;
+using YDataType       = ck::half_t;
+using ComputeDataType = float;
+using Swish           = ck::tensor_operation::element_wise::Swish;
+
+constexpr int Rank         = 5;
+constexpr int NumReduceDim = 3;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    ck::index_t N = 32;
+    ck::index_t H = 16;
+    ck::index_t W = 16;
+    ck::index_t G = 64;
+    ck::index_t C = 128;
+
+    std::size_t xy_size         = N * H * W * G * C;
+    std::size_t gamma_beta_size = G * C;
+
+    std::vector<ck::index_t> xy_strides         = {H * W * G * C, W * G * C, G * C, C, 1};
+    std::vector<ck::index_t> gamma_beta_strides = {0, 0, 0, C, 1};
+
+    SimpleDeviceMem x_device_buf(sizeof(XDataType) * xy_size);
+    SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_beta_size);
+    SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * gamma_beta_size);
+    SimpleDeviceMem y_device_buf(sizeof(YDataType) * xy_size);
+
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
+                                                                       GammaDataType,
+                                                                       BetaDataType,
+                                                                       ComputeDataType,
+                                                                       YDataType,
+                                                                       Swish,
+                                                                       Rank,
+                                                                       NumReduceDim>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto& generic_op_ptr = op_ptrs[0];
+
+    auto generic_argument_ptr =
+        generic_op_ptr->MakeArgumentPointer({N, H, W, G, C},    // lengths
+                                            xy_strides,         // xStrides
+                                            gamma_beta_strides, // gammaStrides
+                                            gamma_beta_strides, // betaStrides
+                                            xy_strides,         // yStrides
+                                            {1, 2, 4},          // reduceDims
+                                            1e-6,
+                                            x_device_buf.GetDeviceBuffer(),
+                                            gamma_device_buf.GetDeviceBuffer(),
+                                            beta_device_buf.GetDeviceBuffer(),
+                                            y_device_buf.GetDeviceBuffer(),
+                                            nullptr,
+                                            nullptr,
+                                            Swish{});
+
+    if(!generic_op_ptr->IsSupportedArgument(generic_argument_ptr.get()))
+    {
+        throw std::runtime_error(
+            "The generic kernel instance should be able to support any input shapes");
+    };
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer({N, H, W, G, C},    // lengths
+                                                        xy_strides,         // xStrides
+                                                        gamma_beta_strides, // gammaStrides
+                                                        gamma_beta_strides, // betaStrides
+                                                        xy_strides,         // yStrides
+                                                        {1, 2, 4},          // reduceDims
+                                                        1e-6,
+                                                        x_device_buf.GetDeviceBuffer(),
+                                                        gamma_device_buf.GetDeviceBuffer(),
+                                                        beta_device_buf.GetDeviceBuffer(),
+                                                        y_device_buf.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        nullptr,
+                                                        Swish{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte =
+                sizeof(XDataType) * xy_size + sizeof(GammaDataType) * gamma_beta_size +
+                sizeof(BetaDataType) * gamma_beta_size + sizeof(YDataType) * xy_size;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    // run the best intance
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer({N, H, W, G, C},    // lengths
+                                                        xy_strides,         // xStrides
+                                                        gamma_beta_strides, // gammaStrides
+                                                        gamma_beta_strides, // betaStrides
+                                                        xy_strides,         // yStrides
+                                                        {1, 2, 4},          // reduceDims
+                                                        1e-6,
+                                                        x_device_buf.GetDeviceBuffer(),
+                                                        gamma_device_buf.GetDeviceBuffer(),
+                                                        beta_device_buf.GetDeviceBuffer(),
+                                                        y_device_buf.GetDeviceBuffer(),
+                                                        nullptr,
+                                                        nullptr,
+                                                        Swish{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/19_pool/CMakeLists.txt b/client_example/19_pool/CMakeLists.txt
new file mode 100644
index 0000000000..d4e2e6d4dc
--- /dev/null
+++ b/client_example/19_pool/CMakeLists.txt
@@ -0,0 +1,11 @@
+add_executable(client_max_pool2d_fwd max_pool2d_fwd.cpp)
+target_link_libraries(client_max_pool2d_fwd PRIVATE composable_kernel::device_operations)
+
+add_executable(client_max_pool2d_bwd max_pool2d_bwd.cpp)
+target_link_libraries(client_max_pool2d_bwd PRIVATE composable_kernel::device_operations)
+
+add_executable(client_avg_pool3d_fwd avg_pool3d_fwd.cpp)
+target_link_libraries(client_avg_pool3d_fwd PRIVATE composable_kernel::device_operations)
+
+add_executable(client_avg_pool3d_bwd avg_pool3d_bwd.cpp)
+target_link_libraries(client_avg_pool3d_bwd PRIVATE composable_kernel::device_operations)
diff --git a/client_example/19_pool/avg_pool3d_bwd.cpp b/client_example/19_pool/avg_pool3d_bwd.cpp
new file mode 100644
index 0000000000..686d1da3ad
--- /dev/null
+++ b/client_example/19_pool/avg_pool3d_bwd.cpp
@@ -0,0 +1,191 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp"
+
+using DOutDataType = ck::half_t;
+using DInDataType  = ck::half_t;
+
+using DOutLayout = ck::tensor_layout::convolution::NDHWC;
+using DInLayout  = ck::tensor_layout::convolution::NDHWC;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}, mMemSize_(mem_size)
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    void SetZero() const { (void)hipMemset(p_mem_, 0, mMemSize_); }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+    std::size_t mMemSize_;
+};
+
+int main(int argc, char* argv[])
+{
+    ck::index_t N                 = 2;
+    ck::index_t C                 = 32;
+    ck::index_t Z                 = 2;
+    ck::index_t Y                 = 2;
+    ck::index_t X                 = 2;
+    ck::index_t Di                = 30;
+    ck::index_t Hi                = 30;
+    ck::index_t Wi                = 30;
+    ck::index_t window_stride_d   = 2;
+    ck::index_t window_stride_h   = 2;
+    ck::index_t window_stride_w   = 2;
+    ck::index_t window_dilation_d = 1;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
+    ck::index_t in_left_pad_d     = 1;
+    ck::index_t in_left_pad_h     = 1;
+    ck::index_t in_left_pad_w     = 1;
+    ck::index_t in_right_pad_d    = 1;
+    ck::index_t in_right_pad_h    = 1;
+    ck::index_t in_right_pad_w    = 1;
+
+    const ck::index_t Zs = (Z - 1) * window_dilation_d + 1;
+    const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
+    const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
+    ck::index_t Do       = (Di + in_left_pad_d + in_right_pad_d - Zs) / window_stride_d + 1;
+    ck::index_t Ho       = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
+    ck::index_t Wo       = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
+
+    // Pool API only support the order of NCDHW
+    std::vector<ck::index_t> in_length              = {N, C, Di, Hi, Wi};
+    std::vector<ck::index_t> out_length             = {N, C, Do, Ho, Wo};
+    std::vector<ck::index_t> window_spatial_lengths = {Z, Y, X};
+    std::vector<ck::index_t> window_strides = {window_stride_d, window_stride_h, window_stride_w};
+    std::vector<ck::index_t> window_dilations{
+        window_dilation_d, window_dilation_h, window_dilation_w};
+    std::vector<ck::index_t> input_left_pads  = {in_left_pad_d, in_left_pad_h, in_left_pad_w};
+    std::vector<ck::index_t> input_right_pads = {in_right_pad_d, in_right_pad_h, in_right_pad_w};
+
+    std::size_t in_tensor_size  = N * C * Di * Hi * Wi;
+    std::size_t out_tensor_size = N * C * Do * Ho * Wo;
+
+    // tensor layout = NDHWC
+    std::vector<ck::index_t> in_tensor_stride  = {Di * C * Hi * Wi, 1, C * Hi * Wi, Wi * C, C};
+    std::vector<ck::index_t> out_tensor_stride = {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C};
+
+    SimpleDeviceMem dout_device_buf(sizeof(DOutDataType) * out_tensor_size);
+    SimpleDeviceMem din_device_buf(sizeof(DInDataType) * in_tensor_size);
+
+    using DeviceOp = ck::tensor_operation::device::
+        DeviceAvgPoolBwd<3, DOutDataType, DInDataType, DOutLayout, DInLayout>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
+            static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+            out_length,
+            in_length,
+            out_tensor_stride,
+            in_tensor_stride,
+            window_spatial_lengths,
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            din_device_buf.SetZero();
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes =
+                in_tensor_size * sizeof(DInDataType) + out_tensor_size * sizeof(DOutDataType);
+
+            float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    // run the best intance
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
+            static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+            out_length,
+            in_length,
+            out_tensor_stride,
+            in_tensor_stride,
+            window_spatial_lengths,
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            din_device_buf.SetZero();
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/19_pool/avg_pool3d_fwd.cpp b/client_example/19_pool/avg_pool3d_fwd.cpp
new file mode 100644
index 0000000000..db8e0569d7
--- /dev/null
+++ b/client_example/19_pool/avg_pool3d_fwd.cpp
@@ -0,0 +1,214 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
+
+using InDataType    = ck::half_t;
+using OutDataType   = ck::half_t;
+using IndexDataType = int32_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWC;
+using OutLayout = ck::tensor_layout::convolution::NDHWC;
+
+constexpr ck::index_t InOutRank  = 5;
+constexpr ck::index_t WindowRank = 3;
+#if 0
+constexpr auto ReduceOpId  = ck::ReduceTensorOp::MAX;
+constexpr bool OutputIndex = false;
+#else
+constexpr auto ReduceOpId  = ck::ReduceTensorOp::AVG;
+constexpr bool OutputIndex = false;
+#endif
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    ck::index_t N                 = 2;
+    ck::index_t C                 = 32;
+    ck::index_t Z                 = 2;
+    ck::index_t Y                 = 2;
+    ck::index_t X                 = 2;
+    ck::index_t Di                = 30;
+    ck::index_t Hi                = 30;
+    ck::index_t Wi                = 30;
+    ck::index_t window_stride_d   = 2;
+    ck::index_t window_stride_h   = 2;
+    ck::index_t window_stride_w   = 2;
+    ck::index_t window_dilation_d = 1;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
+    ck::index_t in_left_pad_d     = 1;
+    ck::index_t in_left_pad_h     = 1;
+    ck::index_t in_left_pad_w     = 1;
+    ck::index_t in_right_pad_d    = 1;
+    ck::index_t in_right_pad_h    = 1;
+    ck::index_t in_right_pad_w    = 1;
+
+    const ck::index_t Zs = (Z - 1) * window_dilation_d + 1;
+    const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
+    const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
+    ck::index_t Do       = (Di + in_left_pad_d + in_right_pad_d - Zs) / window_stride_d + 1;
+    ck::index_t Ho       = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
+    ck::index_t Wo       = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
+
+    // Pool API only support the order of NCDHW
+    std::vector<ck::index_t> in_length              = {N, C, Di, Hi, Wi};
+    std::vector<ck::index_t> out_length             = {N, C, Do, Ho, Wo};
+    std::vector<ck::index_t> window_spatial_lengths = {Z, Y, X};
+    std::vector<ck::index_t> window_strides = {window_stride_d, window_stride_h, window_stride_w};
+    std::vector<ck::index_t> window_dilations{
+        window_dilation_d, window_dilation_h, window_dilation_w};
+    std::vector<ck::index_t> input_left_pads  = {in_left_pad_d, in_left_pad_h, in_left_pad_w};
+    std::vector<ck::index_t> input_right_pads = {in_right_pad_d, in_right_pad_h, in_right_pad_w};
+
+    std::size_t in_tensor_size  = N * C * Di * Hi * Wi;
+    std::size_t out_tensor_size = N * C * Do * Ho * Wo;
+
+    // tensor layout = NDHWC
+    std::vector<ck::index_t> in_tensor_stride  = {Di * C * Hi * Wi, 1, C * Hi * Wi, Wi * C, C};
+    std::vector<ck::index_t> out_tensor_stride = {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C};
+
+    SimpleDeviceMem in_device_buf(sizeof(InDataType) * in_tensor_size);
+    SimpleDeviceMem out_device_buf(sizeof(OutDataType) * out_tensor_size);
+    SimpleDeviceMem out_indices_device_buf(sizeof(IndexDataType) * out_tensor_size);
+
+    using DeviceOp = ck::tensor_operation::device::DevicePoolFwd<InOutRank,
+                                                                 WindowRank,
+                                                                 InDataType,
+                                                                 OutDataType,
+                                                                 IndexDataType,
+                                                                 InLayout,
+                                                                 OutLayout,
+                                                                 ReduceOpId,
+                                                                 OutputIndex>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
+            in_length,
+            window_spatial_lengths,
+            out_length,
+            in_tensor_stride,
+            out_tensor_stride,
+            out_tensor_stride,
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads,
+            {2, 3, 4});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes =
+                in_tensor_size * sizeof(InDataType) + out_tensor_size * sizeof(OutDataType);
+
+            if constexpr(OutputIndex)
+                num_bytes += out_tensor_size * sizeof(IndexDataType);
+
+            float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    // run the best intance
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
+            in_length,
+            window_spatial_lengths,
+            out_length,
+            in_tensor_stride,
+            out_tensor_stride,
+            out_tensor_stride,
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads,
+            {2, 3, 4});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/19_pool/max_pool2d_bwd.cpp b/client_example/19_pool/max_pool2d_bwd.cpp
new file mode 100644
index 0000000000..53ece7425f
--- /dev/null
+++ b/client_example/19_pool/max_pool2d_bwd.cpp
@@ -0,0 +1,280 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
+#include "ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
+#include "ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp"
+
+using InDataType    = ck::half_t;
+using OutDataType   = ck::half_t;
+using DOutDataType  = ck::half_t;
+using DInDataType   = ck::half_t;
+using IndexDataType = int32_t;
+
+// We use pool3d to implement pool2d in this example
+using InLayout  = ck::tensor_layout::convolution::NDHWC;
+using OutLayout = ck::tensor_layout::convolution::NDHWC;
+
+constexpr ck::index_t InOutRank  = 5;
+constexpr ck::index_t WindowRank = 3;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+void TransformPool2dparamToPool3d(std::vector<ck::index_t>& input_lengths,
+                                  std::vector<ck::index_t>& window_lengths,
+                                  std::vector<ck::index_t>& output_lengths,
+                                  std::vector<ck::index_t>& input_stride,
+                                  std::vector<ck::index_t>& output_stride,
+                                  std::vector<ck::index_t>& indices_stride,
+                                  std::vector<ck::index_t>& window_strides,
+                                  std::vector<ck::index_t>& window_dilations,
+                                  std::vector<ck::index_t>& input_left_pads,
+                                  std::vector<ck::index_t>& input_right_pads,
+                                  std::vector<ck::index_t>& pooling_dims)
+{
+    // NCHW to NCDHW
+    input_lengths.insert(input_lengths.begin() + 2, 1);
+    output_lengths.insert(output_lengths.begin() + 2, 1);
+    input_stride.insert(input_stride.begin() + 2, 0);
+    output_stride.insert(output_stride.begin() + 2, 0);
+    indices_stride.insert(indices_stride.begin() + 2, 0);
+
+    // YX to ZYX
+    window_lengths.insert(window_lengths.begin(), 1);
+    window_strides.insert(window_strides.begin(), 0);
+    window_dilations.insert(window_dilations.begin(), 0);
+    input_left_pads.insert(input_left_pads.begin(), 0);
+    input_right_pads.insert(input_right_pads.begin(), 0);
+
+    pooling_dims = {2, 3, 4};
+}
+
+int main(int argc, char* argv[])
+{
+    ck::index_t N                 = 2;
+    ck::index_t C                 = 32;
+    ck::index_t Y                 = 2;
+    ck::index_t X                 = 2;
+    ck::index_t Hi                = 30;
+    ck::index_t Wi                = 30;
+    ck::index_t window_stride_h   = 2;
+    ck::index_t window_stride_w   = 2;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
+    ck::index_t in_left_pad_h     = 1;
+    ck::index_t in_left_pad_w     = 1;
+    ck::index_t in_right_pad_h    = 1;
+    ck::index_t in_right_pad_w    = 1;
+
+    const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
+    const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
+    ck::index_t Ho       = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
+    ck::index_t Wo       = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
+
+    // Pool API only support the order of NCHW
+    std::vector<ck::index_t> in_length              = {N, C, Hi, Wi};
+    std::vector<ck::index_t> out_length             = {N, C, Ho, Wo};
+    std::vector<ck::index_t> window_spatial_lengths = {Y, X};
+    std::vector<ck::index_t> window_strides         = {window_stride_h, window_stride_w};
+    std::vector<ck::index_t> window_dilations       = {window_dilation_h, window_dilation_w};
+    std::vector<ck::index_t> input_left_pads        = {in_left_pad_h, in_left_pad_w};
+    std::vector<ck::index_t> input_right_pads       = {in_right_pad_h, in_right_pad_w};
+    std::vector<ck::index_t> pooling_dims           = {2, 3};
+
+    std::size_t in_tensor_size  = N * C * Hi * Wi;
+    std::size_t out_tensor_size = N * C * Ho * Wo;
+
+    // tensor layout = NHWC
+    std::vector<ck::index_t> in_tensor_stride  = {C * Hi * Wi, 1, Wi * C, C};
+    std::vector<ck::index_t> out_tensor_stride = {C * Ho * Wo, 1, Wo * C, C};
+
+    TransformPool2dparamToPool3d(in_length,
+                                 window_spatial_lengths,
+                                 out_length,
+                                 in_tensor_stride,
+                                 out_tensor_stride,
+                                 out_tensor_stride,
+                                 window_strides,
+                                 window_dilations,
+                                 input_left_pads,
+                                 input_right_pads,
+                                 pooling_dims);
+
+    SimpleDeviceMem in_device_buf(sizeof(InDataType) * in_tensor_size);
+    SimpleDeviceMem out_device_buf(sizeof(OutDataType) * out_tensor_size);
+    SimpleDeviceMem indices_device_buf(sizeof(IndexDataType) * out_tensor_size);
+    SimpleDeviceMem dout_device_buf(sizeof(DOutDataType) * out_tensor_size);
+    SimpleDeviceMem din_device_buf(sizeof(DInDataType) * in_tensor_size);
+
+    // Generate index data from max pool forward
+    {
+        using MaxPoolFwdDeviceOp =
+            ck::tensor_operation::device::DevicePoolFwd<InOutRank,
+                                                        WindowRank,
+                                                        InDataType,
+                                                        OutDataType,
+                                                        IndexDataType,
+                                                        InLayout,
+                                                        OutLayout,
+                                                        ck::ReduceTensorOp::MAX,
+                                                        true>;
+
+        const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            MaxPoolFwdDeviceOp>::GetInstances();
+
+        auto& op_ptr = op_ptrs[0];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
+            in_length,
+            window_spatial_lengths,
+            out_length,
+            in_tensor_stride,
+            out_tensor_stride,
+            out_tensor_stride,
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads,
+            pooling_dims);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+    }
+
+    // Run MaxPool bwd
+    using MaxPoolBwdDeviceOp =
+        ck::tensor_operation::device::DeviceMaxPoolBwd<DOutDataType, IndexDataType, DInDataType>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        MaxPoolBwdDeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<InDataType*>(dout_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
+            static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+            out_tensor_size,
+            in_tensor_size,
+            window_spatial_lengths,
+            window_strides,
+            window_dilations);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes = in_tensor_size * sizeof(DInDataType) +
+                                    out_tensor_size * sizeof(IndexDataType) +
+                                    out_tensor_size * sizeof(DOutDataType);
+
+            float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << "GB / s,"
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    // run the best intance
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<InDataType*>(dout_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
+            static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+            out_tensor_size,
+            in_tensor_size,
+            window_spatial_lengths,
+            window_strides,
+            window_dilations);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+            SimpleDeviceMem workspace(workspace_sz);
+
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/19_pool/max_pool2d_fwd.cpp b/client_example/19_pool/max_pool2d_fwd.cpp
new file mode 100644
index 0000000000..84b818a60f
--- /dev/null
+++ b/client_example/19_pool/max_pool2d_fwd.cpp
@@ -0,0 +1,248 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
+
+using InDataType    = ck::half_t;
+using OutDataType   = ck::half_t;
+using IndexDataType = int32_t;
+
+// We use pool3d to implement pool2d in this example
+using InLayout  = ck::tensor_layout::convolution::NDHWC;
+using OutLayout = ck::tensor_layout::convolution::NDHWC;
+
+constexpr ck::index_t InOutRank  = 5;
+constexpr ck::index_t WindowRank = 3;
+#if 1
+constexpr auto ReduceOpId  = ck::ReduceTensorOp::MAX;
+constexpr bool OutputIndex = true;
+#else
+constexpr auto ReduceOpId  = ck::ReduceTensorOp::AVG;
+constexpr bool OutputIndex = false;
+#endif
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+void TransformPool2dparamToPool3d(std::vector<ck::index_t>& input_lengths,
+                                  std::vector<ck::index_t>& window_lengths,
+                                  std::vector<ck::index_t>& output_lengths,
+                                  std::vector<ck::index_t>& input_stride,
+                                  std::vector<ck::index_t>& output_stride,
+                                  std::vector<ck::index_t>& indices_stride,
+                                  std::vector<ck::index_t>& window_strides,
+                                  std::vector<ck::index_t>& window_dilations,
+                                  std::vector<ck::index_t>& input_left_pads,
+                                  std::vector<ck::index_t>& input_right_pads,
+                                  std::vector<ck::index_t>& pooling_dims)
+{
+    // NCHW to NCDHW
+    input_lengths.insert(input_lengths.begin() + 2, 1);
+    output_lengths.insert(output_lengths.begin() + 2, 1);
+    input_stride.insert(input_stride.begin() + 2, 0);
+    output_stride.insert(output_stride.begin() + 2, 0);
+    indices_stride.insert(indices_stride.begin() + 2, 0);
+
+    // YX to ZYX
+    window_lengths.insert(window_lengths.begin(), 1);
+    window_strides.insert(window_strides.begin(), 0);
+    window_dilations.insert(window_dilations.begin(), 0);
+    input_left_pads.insert(input_left_pads.begin(), 0);
+    input_right_pads.insert(input_right_pads.begin(), 0);
+
+    pooling_dims = {2, 3, 4};
+}
+
+int main(int argc, char* argv[])
+{
+    ck::index_t N                 = 2;
+    ck::index_t C                 = 32;
+    ck::index_t Y                 = 2;
+    ck::index_t X                 = 2;
+    ck::index_t Hi                = 30;
+    ck::index_t Wi                = 30;
+    ck::index_t window_stride_h   = 2;
+    ck::index_t window_stride_w   = 2;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
+    ck::index_t in_left_pad_h     = 1;
+    ck::index_t in_left_pad_w     = 1;
+    ck::index_t in_right_pad_h    = 1;
+    ck::index_t in_right_pad_w    = 1;
+
+    const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
+    const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
+    ck::index_t Ho       = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
+    ck::index_t Wo       = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
+
+    // Pool API only support the order of NCHW
+    std::vector<ck::index_t> in_length              = {N, C, Hi, Wi};
+    std::vector<ck::index_t> out_length             = {N, C, Ho, Wo};
+    std::vector<ck::index_t> window_spatial_lengths = {Y, X};
+    std::vector<ck::index_t> window_strides         = {window_stride_h, window_stride_w};
+    std::vector<ck::index_t> window_dilations       = {window_dilation_h, window_dilation_w};
+    std::vector<ck::index_t> input_left_pads        = {in_left_pad_h, in_left_pad_w};
+    std::vector<ck::index_t> input_right_pads       = {in_right_pad_h, in_right_pad_w};
+    std::vector<ck::index_t> pooling_dims           = {2, 3};
+
+    std::size_t in_tensor_size  = N * C * Hi * Wi;
+    std::size_t out_tensor_size = N * C * Ho * Wo;
+
+    // tensor layout = NHWC
+    std::vector<ck::index_t> in_tensor_stride  = {C * Hi * Wi, 1, Wi * C, C};
+    std::vector<ck::index_t> out_tensor_stride = {C * Ho * Wo, 1, Wo * C, C};
+
+    TransformPool2dparamToPool3d(in_length,
+                                 window_spatial_lengths,
+                                 out_length,
+                                 in_tensor_stride,
+                                 out_tensor_stride,
+                                 out_tensor_stride,
+                                 window_strides,
+                                 window_dilations,
+                                 input_left_pads,
+                                 input_right_pads,
+                                 pooling_dims);
+
+    SimpleDeviceMem in_device_buf(sizeof(InDataType) * in_tensor_size);
+    SimpleDeviceMem out_device_buf(sizeof(OutDataType) * out_tensor_size);
+    SimpleDeviceMem out_indices_device_buf(sizeof(IndexDataType) * out_tensor_size);
+
+    using DeviceOp = ck::tensor_operation::device::DevicePoolFwd<InOutRank,
+                                                                 WindowRank,
+                                                                 InDataType,
+                                                                 OutDataType,
+                                                                 IndexDataType,
+                                                                 InLayout,
+                                                                 OutLayout,
+                                                                 ReduceOpId,
+                                                                 OutputIndex>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
+            in_length,
+            window_spatial_lengths,
+            out_length,
+            in_tensor_stride,
+            out_tensor_stride,
+            out_tensor_stride,
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads,
+            pooling_dims);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes =
+                in_tensor_size * sizeof(InDataType) + out_tensor_size * sizeof(OutDataType);
+
+            if constexpr(OutputIndex)
+                num_bytes += out_tensor_size * sizeof(IndexDataType);
+
+            float gb_per_sec = num_bytes / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    // run the best intance
+    if(found)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_op_name << std::endl;
+
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
+            in_length,
+            window_spatial_lengths,
+            out_length,
+            in_tensor_stride,
+            out_tensor_stride,
+            out_tensor_stride,
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads,
+            pooling_dims);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/20_image_to_column/CMakeLists.txt b/client_example/20_image_to_column/CMakeLists.txt
new file mode 100644
index 0000000000..80edcd0416
--- /dev/null
+++ b/client_example/20_image_to_column/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_image_to_column image_to_column.cpp)
+target_link_libraries(client_image_to_column PRIVATE composable_kernel::device_operations)
diff --git a/client_example/20_image_to_column/image_to_column.cpp b/client_example/20_image_to_column/image_to_column.cpp
new file mode 100644
index 0000000000..ace4c1a681
--- /dev/null
+++ b/client_example/20_image_to_column/image_to_column.cpp
@@ -0,0 +1,167 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/image_to_column.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType  = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout = ck::tensor_layout::convolution::GNHWC;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 32; // batch size
+static constexpr ck::index_t C             = 32; // input channel (per group)
+static constexpr ck::index_t Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Hi            = 28; // input H
+static constexpr ck::index_t Wi            = 28; // input W
+static constexpr ck::index_t Ho            = 28; // output H
+static constexpr ck::index_t Wo            = 28; // output W
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+
+    std::array<ck::index_t, 2> in_spatial_lengths{Hi, Wi};
+    std::array<ck::index_t, 2> wei_spatial_lengths{Y, X};
+    std::array<ck::index_t, 2> out_spatial_lengths{Ho, Wo};
+
+    // We have NHWGC in memory space (G is dummy)
+    // However, CK's API only accept length and stride with order of GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 2> out_strides{Y * X * C, 1};
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * Y * X * C);
+
+    using DeviceOp = ck::tensor_operation::device::
+        DeviceImageToColumn<NumDimSpatial, InLayout, InDataType, OutDataType>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        N,
+                                                        C,
+                                                        in_spatial_lengths,
+                                                        out_spatial_lengths,
+                                                        wei_spatial_lengths,
+                                                        in_strides,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads);
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes = sizeof(InDataType) * N * Hi * Wi * G * C +
+                                    sizeof(OutDataType) * N * Ho * Wo * Y * X * C;
+
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(avg_time < best_avg_time)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_gb_per_sec
+              << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        N,
+                                                        C,
+                                                        in_spatial_lengths,
+                                                        out_spatial_lengths,
+                                                        wei_spatial_lengths,
+                                                        in_strides,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+}
diff --git a/client_example/20_splitk_gemm/CMakeLists.txt b/client_example/20_splitk_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..a60bada473
--- /dev/null
+++ b/client_example/20_splitk_gemm/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_splitK_gemm splitK_gemm_fp16_f8.cpp)
+target_link_libraries(client_splitK_gemm PRIVATE composable_kernel::device_operations)
diff --git a/client_example/20_splitk_gemm/splitK_gemm_fp16_f8.cpp b/client_example/20_splitk_gemm/splitK_gemm_fp16_f8.cpp
new file mode 100644
index 0000000000..94a57cd029
--- /dev/null
+++ b/client_example/20_splitk_gemm/splitK_gemm_fp16_f8.cpp
@@ -0,0 +1,225 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <vector>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp"
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+using ADataType = F8;
+using BDataType = F16;
+using CDataType = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    ck::index_t KBatch = 1;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 8)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+        K = std::stoi(argv[3]);
+
+        StrideA = std::stoi(argv[4]);
+        StrideB = std::stoi(argv[5]);
+        StrideC = std::stoi(argv[6]);
+
+        KBatch = std::stoi(argv[7]);
+    }
+    else
+    {
+        printf("arg1 to 7: M, N, K, StrideA, StrideB, StrideC, KBatch\n");
+        exit(0);
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem c_device_buf(sizeof(CDataType) * f_matrix_space_size(M, N, StrideC, CLayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmSplitK<
+        ALayout,
+        BLayout,
+        CLayout,
+        ADataType,
+        BDataType,
+        CDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        StrideC,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        c_element_op,
+                                                        KBatch);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        c_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        StrideC,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        c_element_op,
+                                                        KBatch);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
diff --git a/client_example/21_grouped_gemm_bias/CMakeLists.txt b/client_example/21_grouped_gemm_bias/CMakeLists.txt
new file mode 100644
index 0000000000..a2abd15731
--- /dev/null
+++ b/client_example/21_grouped_gemm_bias/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(client_grouped_gemm_fixed_nk_bias_fp16 grouped_gemm_fixed_nk_bias_fp16.cpp)
+target_link_libraries(client_grouped_gemm_fixed_nk_bias_fp16 PRIVATE composable_kernel::device_operations)
diff --git a/client_example/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp b/client_example/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp
new file mode 100644
index 0000000000..3b6dd9a2a9
--- /dev/null
+++ b/client_example/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp
@@ -0,0 +1,244 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+#include <random>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_bias.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using D0DataType = F32;
+using DsDataType = ck::Tuple<D0DataType>;
+using EDataType  = F32;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Add;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
+
+    int sum_of_m = 0;
+
+    Ms = {167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
+
+    int group_count = Ms.size();
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        Ns.push_back(768);
+        Ks.push_back(4608);
+
+        StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
+        StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
+        StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
+
+        sum_of_m += Ms[i];
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, d0_dev_bufs, e_dev_bufs;
+
+    a_dev_bufs.reserve(group_count);
+    b_dev_bufs.reserve(group_count);
+    d0_dev_bufs.reserve(group_count);
+    e_dev_bufs.reserve(group_count);
+
+    std::vector<void*> p_e;
+
+    p_e.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GroupedGemmKernelArgument<1>>
+        grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        a_dev_bufs.emplace_back(sizeof(ADataType) *
+                                f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
+        b_dev_bufs.emplace_back(sizeof(BDataType) *
+                                f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
+        d0_dev_bufs.emplace_back(sizeof(D0DataType) *
+                                 f_matrix_space_size(Ms[i], Ns[i], 0, D0Layout{}));
+        e_dev_bufs.emplace_back(sizeof(EDataType) *
+                                f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
+
+        gemm_descs.push_back({sum_of_m, Ns[i], Ks[i], 1, StrideBs[i], 1, {0}});
+
+        p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
+
+        grouped_gemm_kernel_args_.push_back(
+            {a_dev_bufs[i].GetDeviceBuffer(),
+             b_dev_bufs[i].GetDeviceBuffer(),
+             std::array<const void*, 1>{d0_dev_bufs[i].GetDeviceBuffer()},
+             e_dev_bufs[i].GetDeviceBuffer(),
+             Ms[i],
+             Ns[i],
+             Ks[i],
+             StrideAs[i],
+             StrideBs[i],
+             std::array<ck::index_t, 1>{0},
+             StrideEs[i]});
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
+                                                                            BLayout,
+                                                                            DsLayout,
+                                                                            ELayout,
+                                                                            ADataType,
+                                                                            BDataType,
+                                                                            DsDataType,
+                                                                            EDataType,
+                                                                            AElementOp,
+                                                                            BElementOp,
+                                                                            CDEElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    std::vector<const void*> p_a = {}, p_b = {};
+    std::vector<std::array<const void*, 1>> p_ds = {};
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        SimpleDeviceMem grouped_gemm_kernel_args_dev(
+            op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
+
+        SimpleDeviceMem grouped_gemm_workspace_dev(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        hipGetErrorString(hipMemcpy(grouped_gemm_kernel_args_dev.GetDeviceBuffer(),
+                                    grouped_gemm_kernel_args_.data(),
+                                    op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
+                                    hipMemcpyHostToDevice));
+
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(),
+                                    grouped_gemm_workspace_dev.GetDeviceBuffer());
+
+        op_ptr->SetDeviceKernelArgs(argument_ptr.get(),
+                                    grouped_gemm_kernel_args_dev.GetDeviceBuffer());
+
+        op_ptr->SetKBatch(argument_ptr.get(), 2);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = 0, num_btype = 0;
+            for(std::size_t j = 0; j < gemm_descs.size(); ++j)
+            {
+                flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
+
+                num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
+                             sizeof(EDataType) * Ms[j] * Ns[j];
+            }
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return 0;
+}
diff --git a/client_example/CMakeLists.txt b/client_example/CMakeLists.txt
new file mode 100644
index 0000000000..eb793b3cbd
--- /dev/null
+++ b/client_example/CMakeLists.txt
@@ -0,0 +1,61 @@
+cmake_minimum_required(VERSION 3.15)
+project(ck_app)
+add_compile_options(-std=c++17)
+
+if (DTYPES)
+    add_definitions(-DDTYPES)
+    if (DTYPES MATCHES "int8")
+        add_definitions(-DCK_ENABLE_INT8)
+        if(NOT DEFINED ${CK_ENABLE_INT8})
+            set(CK_ENABLE_INT8 "ON")
+        endif()
+    endif()
+    if (DTYPES MATCHES "fp8")
+        add_definitions(-DCK_ENABLE_FP8)
+        if(NOT DEFINED ${CK_ENABLE_FP8})
+            set(CK_ENABLE_FP8 "ON")
+        endif()
+    endif()
+    if (DTYPES MATCHES "fp16")
+        add_definitions(-DCK_ENABLE_FP16)
+        if(NOT DEFINED ${CK_ENABLE_FP16})
+            set(CK_ENABLE_FP16 "ON")
+        endif()
+    endif()
+    if (DTYPES MATCHES "fp32")
+        add_definitions(-DCK_ENABLE_FP32)
+        if(NOT DEFINED ${CK_ENABLE_FP32})
+            set(CK_ENABLE_FP32 "ON")
+        endif()
+    endif()
+    if (DTYPES MATCHES "fp64")
+        add_definitions(-DCK_ENABLE_FP64)
+        if(NOT DEFINED ${CK_ENABLE_FP64})
+            set(CK_ENABLE_FP64 "ON")
+        endif()
+    endif()
+    if (DTYPES MATCHES "bf16")
+        add_definitions(-DCK_ENABLE_BF16)
+        if(NOT DEFINED ${CK_ENABLE_BF16})
+            set(CK_ENABLE_BF16 "ON")
+        endif()
+    endif()
+    message("DTYPES macro set to ${DTYPES}")
+else()
+    add_definitions(-DCK_ENABLE_INT8 -DCK_ENABLE_FP8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_BF16)
+    if(NOT DEFINED ${CK_ENABLE_ALL_DTYPES})
+        set(CK_ENABLE_ALL_DTYPES "ON")
+    endif()
+endif()
+
+find_package(composable_kernel COMPONENTS device_operations)
+find_package(hip REQUIRED PATHS /opt/rocm)
+message(STATUS "Build with HIP ${hip_VERSION}")
+
+# add all example subdir
+file(GLOB dir_list LIST_DIRECTORIES true *)
+FOREACH(subdir ${dir_list})
+    IF(IS_DIRECTORY "${subdir}" AND (NOT "${subdir}" MATCHES "build"))
+        add_subdirectory(${subdir})
+    ENDIF()
+ENDFOREACH()
diff --git a/client_example/README.md b/client_example/README.md
new file mode 100644
index 0000000000..64a7130d53
--- /dev/null
+++ b/client_example/README.md
@@ -0,0 +1,21 @@
+##
+Client application links to CK library, and therefore CK library needs to be installed before building client applications.
+
+
+## Build
+```bash
+mkdir -p client_example/build
+cd client_example/build
+```
+
+```bash
+cmake                                                                 \
+-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                             \
+-D CMAKE_PREFIX_PATH="/opt/rocm;${PATH_TO_CK_INSTALL_DIRECTORY}"      \
+..
+```
+
+### Build client example
+```bash
+ make -j 
+```
diff --git a/cmake/Analyzers.cmake b/cmake/Analyzers.cmake
new file mode 100644
index 0000000000..1bf1a52c68
--- /dev/null
+++ b/cmake/Analyzers.cmake
@@ -0,0 +1,34 @@
+################################################################################
+# 
+# MIT License
+# 
+# Copyright (c) 2017 Advanced Micro Devices, Inc.
+# 
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+# 
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+# 
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+# 
+################################################################################
+
+if(NOT TARGET analyze)
+    add_custom_target(analyze)
+endif()
+
+function(mark_as_analyzer)
+    add_dependencies(analyze ${ARGN})
+endfunction()
+
diff --git a/cmake/ClangTidy.cmake b/cmake/ClangTidy.cmake
new file mode 100644
index 0000000000..01b348c458
--- /dev/null
+++ b/cmake/ClangTidy.cmake
@@ -0,0 +1,162 @@
+################################################################################
+#
+# MIT License
+#
+# Copyright (c) 2017 Advanced Micro Devices, Inc.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+#
+################################################################################
+include(CMakeParseArguments)
+include(Analyzers)
+
+get_filename_component(CLANG_TIDY_EXE_HINT "${CMAKE_CXX_COMPILER}" PATH)
+
+find_program(CLANG_TIDY_EXE
+    NAMES
+        clang-tidy
+        clang-tidy-5.0
+        clang-tidy-4.0
+        clang-tidy-3.9
+        clang-tidy-3.8
+        clang-tidy-3.7
+        clang-tidy-3.6
+        clang-tidy-3.5
+    HINTS
+        ${CLANG_TIDY_EXE_HINT}
+    PATH_SUFFIXES
+        compiler/bin
+    PATHS
+        /opt/rocm/llvm/bin
+        /opt/rocm/hcc
+        /usr/local/opt/llvm/bin
+)
+
+function(find_clang_tidy_version VAR)
+    execute_process(COMMAND ${CLANG_TIDY_EXE} -version OUTPUT_VARIABLE VERSION_OUTPUT)
+    separate_arguments(VERSION_OUTPUT_LIST UNIX_COMMAND "${VERSION_OUTPUT}")
+    list(FIND VERSION_OUTPUT_LIST "version" VERSION_INDEX)
+    if(VERSION_INDEX GREATER 0)
+        math(EXPR VERSION_INDEX "${VERSION_INDEX} + 1")
+        list(GET VERSION_OUTPUT_LIST ${VERSION_INDEX} VERSION)
+        set(${VAR} ${VERSION} PARENT_SCOPE)
+    else()
+        set(${VAR} "0.0" PARENT_SCOPE)
+    endif()
+
+endfunction()
+
+if( NOT CLANG_TIDY_EXE )
+    message( STATUS "Clang tidy not found" )
+    set(CLANG_TIDY_VERSION "0.0")
+else()
+    find_clang_tidy_version(CLANG_TIDY_VERSION)
+    message( STATUS "Clang tidy found: ${CLANG_TIDY_VERSION}")
+endif()
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+set(CLANG_TIDY_FIXIT_DIR ${CMAKE_BINARY_DIR}/fixits)
+file(MAKE_DIRECTORY ${CLANG_TIDY_FIXIT_DIR})
+set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES ${CLANG_TIDY_FIXIT_DIR})
+
+macro(enable_clang_tidy)
+    set(options ANALYZE_TEMPORARY_DTORS ALL)
+    set(oneValueArgs HEADER_FILTER)
+    set(multiValueArgs CHECKS ERRORS EXTRA_ARGS)
+
+    cmake_parse_arguments(PARSE "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+    string(REPLACE ";" "," CLANG_TIDY_CHECKS "${PARSE_CHECKS}")
+    string(REPLACE ";" "," CLANG_TIDY_ERRORS "${PARSE_ERRORS}")
+    set(CLANG_TIDY_EXTRA_ARGS)
+    foreach(ARG ${PARSE_EXTRA_ARGS})
+        list(APPEND CLANG_TIDY_EXTRA_ARGS "-extra-arg=${ARG}")
+    endforeach()
+
+    set(CLANG_TIDY_ALL)
+    if(PARSE_ALL)
+        set(CLANG_TIDY_ALL ALL)
+    endif()
+
+    message(STATUS "Clang tidy checks: ${CLANG_TIDY_CHECKS}")
+
+    if (${PARSE_ANALYZE_TEMPORARY_DTORS})
+        set(CLANG_TIDY_ANALYZE_TEMPORARY_DTORS "-analyze-temporary-dtors")
+    endif()
+
+    if (${CLANG_TIDY_VERSION} VERSION_LESS "3.9.0")
+        set(CLANG_TIDY_ERRORS_ARG "")
+    else()
+        set(CLANG_TIDY_ERRORS_ARG "-warnings-as-errors='${CLANG_TIDY_ERRORS}'")
+    endif()
+
+    if (${CLANG_TIDY_VERSION} VERSION_LESS "3.9.0")
+        set(CLANG_TIDY_QUIET_ARG "")
+    else()
+        set(CLANG_TIDY_QUIET_ARG "-quiet")
+    endif()
+
+    if(PARSE_HEADER_FILTER)
+        string(REPLACE "$" "$$" CLANG_TIDY_HEADER_FILTER "${PARSE_HEADER_FILTER}")
+    else()
+        set(CLANG_TIDY_HEADER_FILTER ".*")
+    endif()
+
+    set(CLANG_TIDY_COMMAND
+        ${CLANG_TIDY_EXE}
+        ${CLANG_TIDY_QUIET_ARG}
+        -p ${CMAKE_BINARY_DIR}
+        -checks='${CLANG_TIDY_CHECKS}'
+        ${CLANG_TIDY_ERRORS_ARG}
+        ${CLANG_TIDY_EXTRA_ARGS}
+        ${CLANG_TIDY_ANALYZE_TEMPORARY_DTORS}
+        -header-filter='${CLANG_TIDY_HEADER_FILTER}'
+    )
+    add_custom_target(tidy ${CLANG_TIDY_ALL})
+    mark_as_analyzer(tidy)
+    add_custom_target(tidy-base)
+    add_custom_target(tidy-make-fixit-dir COMMAND ${CMAKE_COMMAND} -E make_directory ${CLANG_TIDY_FIXIT_DIR})
+    add_custom_target(tidy-rm-fixit-dir COMMAND ${CMAKE_COMMAND} -E remove_directory ${CLANG_TIDY_FIXIT_DIR})
+    add_dependencies(tidy-make-fixit-dir tidy-rm-fixit-dir)
+    add_dependencies(tidy-base tidy-make-fixit-dir)
+endmacro()
+
+function(clang_tidy_check TARGET)
+    get_target_property(SOURCES ${TARGET} SOURCES)
+    # TODO: Use generator expressions instead
+    # COMMAND ${CLANG_TIDY_COMMAND} $<TARGET_PROPERTY:${TARGET},SOURCES>
+    # COMMAND ${CLANG_TIDY_COMMAND} $<JOIN:$<TARGET_PROPERTY:${TARGET},SOURCES>, >
+    foreach(SOURCE ${SOURCES})
+        if((NOT "${SOURCE}" MATCHES "(h|hpp|hxx)$") AND (NOT "${SOURCE}" MATCHES "TARGET_OBJECTS"))
+            string(MAKE_C_IDENTIFIER "${SOURCE}" tidy_file)
+            set(tidy_target tidy-target-${TARGET}-${tidy_file})
+            add_custom_target(${tidy_target}
+                # for some targets clang-tidy not able to get information from .clang-tidy
+                DEPENDS ${SOURCE}
+                COMMAND ${CLANG_TIDY_COMMAND} "-config=\{CheckOptions: \[\{key: bugprone-reserved-identifier.AllowedIdentifiers,value: __HIP_PLATFORM_HCC__\; __HIP_ROCclr__\}\]\}" ${SOURCE} "-export-fixes=${CLANG_TIDY_FIXIT_DIR}/${TARGET}-${tidy_file}.yaml"
+                WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+                COMMENT "clang-tidy: Running clang-tidy on target ${SOURCE}..."
+            )
+            add_dependencies(${tidy_target} ${TARGET})
+            add_dependencies(${tidy_target} tidy-base)
+            add_dependencies(tidy ${tidy_target})
+        endif()
+    endforeach()
+endfunction()
+
diff --git a/cmake/CppCheck.cmake b/cmake/CppCheck.cmake
new file mode 100644
index 0000000000..797dcf4b4d
--- /dev/null
+++ b/cmake/CppCheck.cmake
@@ -0,0 +1,130 @@
+################################################################################
+# 
+# MIT License
+# 
+# Copyright (c) 2017 Advanced Micro Devices, Inc.
+# 
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+# 
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+# 
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+# 
+################################################################################
+
+include(CMakeParseArguments)
+include(ProcessorCount)
+include(Analyzers)
+
+find_program(CPPCHECK_EXE 
+    NAMES 
+        cppcheck
+    PATHS
+        /opt/rocm/bin
+)
+
+ProcessorCount(CPPCHECK_JOBS)
+
+set(CPPCHECK_BUILD_DIR ${CMAKE_BINARY_DIR}/cppcheck-build)
+file(MAKE_DIRECTORY ${CPPCHECK_BUILD_DIR})
+set_property(DIRECTORY APPEND PROPERTY ADDITIONAL_MAKE_CLEAN_FILES ${CPPCHECK_BUILD_DIR})
+
+macro(enable_cppcheck)
+    set(options FORCE)
+    set(oneValueArgs)
+    set(multiValueArgs CHECKS SUPPRESS DEFINE UNDEFINE INCLUDE SOURCES)
+
+    cmake_parse_arguments(PARSE "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+    string(REPLACE ";" "," CPPCHECK_CHECKS "${PARSE_CHECKS}")
+    string(REPLACE ";" "\n" CPPCHECK_SUPPRESS "${PARSE_SUPPRESS};*:/usr/*")
+    file(WRITE ${CMAKE_BINARY_DIR}/cppcheck-supressions "${CPPCHECK_SUPPRESS}")
+    set(CPPCHECK_DEFINES)
+    foreach(DEF ${PARSE_DEFINE})
+        set(CPPCHECK_DEFINES "${CPPCHECK_DEFINES} -D${DEF}")
+    endforeach()
+
+    set(CPPCHECK_UNDEFINES)
+    foreach(DEF ${PARSE_UNDEFINE})
+        set(CPPCHECK_UNDEFINES "${CPPCHECK_UNDEFINES} -U${DEF}")
+    endforeach()
+
+    set(CPPCHECK_INCLUDES)
+    foreach(INC ${PARSE_INCLUDE})
+        set(CPPCHECK_INCLUDES "${CPPCHECK_INCLUDES} -I${INC}")
+    endforeach()
+
+    # set(CPPCHECK_FORCE)
+    set(CPPCHECK_FORCE "--project=${CMAKE_BINARY_DIR}/compile_commands.json")
+    if(PARSE_FORCE)
+        set(CPPCHECK_FORCE --force)
+    endif()
+
+    set(SOURCES)
+    set(GLOBS)
+    foreach(SOURCE ${PARSE_SOURCES})
+        get_filename_component(ABS_SOURCE ${SOURCE} ABSOLUTE)
+        if(EXISTS ${ABS_SOURCE})
+            if(IS_DIRECTORY ${ABS_SOURCE})
+                set(GLOBS "${GLOBS} ${ABS_SOURCE}/*.cpp ${ABS_SOURCE}/*.hpp ${ABS_SOURCE}/*.cxx ${ABS_SOURCE}/*.c ${ABS_SOURCE}/*.h")
+            else()
+                set(SOURCES "${SOURCES} ${ABS_SOURCE}")
+            endif()
+        else()
+            set(GLOBS "${GLOBS} ${ABS_SOURCE}")
+        endif()
+    endforeach()
+
+    file(WRITE ${CMAKE_BINARY_DIR}/cppcheck.cmake "
+        file(GLOB_RECURSE GSRCS ${GLOBS})
+        set(CPPCHECK_COMMAND
+            ${CPPCHECK_EXE}
+            -q
+            # -v
+            # --report-progress
+            ${CPPCHECK_FORCE}
+            --cppcheck-build-dir=${CPPCHECK_BUILD_DIR}
+            --platform=native
+            --template=gcc
+            --error-exitcode=1
+            -j ${CPPCHECK_JOBS}
+            ${CPPCHECK_DEFINES}
+            ${CPPCHECK_UNDEFINES}
+            ${CPPCHECK_INCLUDES}
+            --enable=${CPPCHECK_CHECKS}
+            --inline-suppr
+            --suppressions-list=${CMAKE_BINARY_DIR}/cppcheck-supressions
+            ${SOURCES} \${GSRCS}
+        )
+        string(REPLACE \";\" \" \" CPPCHECK_SHOW_COMMAND \"\${CPPCHECK_COMMAND}\")
+        message(\"\${CPPCHECK_SHOW_COMMAND}\")
+        execute_process(
+            COMMAND \${CPPCHECK_COMMAND}
+            WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+            RESULT_VARIABLE RESULT
+        )
+        if(NOT RESULT EQUAL 0)
+            message(FATAL_ERROR \"Cppcheck failed\")
+        endif()
+")
+
+    add_custom_target(cppcheck
+        COMMAND ${CMAKE_COMMAND} -P ${CMAKE_BINARY_DIR}/cppcheck.cmake
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        COMMENT "cppcheck: Running cppcheck..."
+    )
+    mark_as_analyzer(cppcheck)
+endmacro()
+
+
diff --git a/cmake/DoxygenDoc.cmake b/cmake/DoxygenDoc.cmake
new file mode 100644
index 0000000000..2e3669fcdf
--- /dev/null
+++ b/cmake/DoxygenDoc.cmake
@@ -0,0 +1,355 @@
+################################################################################
+# 
+# MIT License
+# 
+# Copyright (c) 2017 Advanced Micro Devices, Inc.
+# 
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+# 
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+# 
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+# 
+################################################################################
+include(CMakeParseArguments)
+include(MainDoc)
+
+find_program(DOXYGEN_EXECUTABLE NAMES doxygen
+    PATH_SUFFIXES bin
+    DOC "Doxygen documentation generator"
+)
+mark_as_advanced(DOXYGEN_EXECUTABLE)
+
+find_path(DOT_EXECUTABLE NAMES dot
+    PATH_SUFFIXES bin
+    DOC "Graphviz"
+)
+mark_as_advanced(DOT_EXECUTABLE)
+
+set(DOXYGEN_ARGS
+ABBREVIATE_BRIEF
+ALIASES
+ALLEXTERNALS
+ALLOW_UNICODE_NAMES
+ALPHABETICAL_INDEX
+ALWAYS_DETAILED_SEC
+AUTOLINK_SUPPORT
+BINARY_TOC
+BRIEF_MEMBER_DESC
+BUILTIN_STL_SUPPORT
+CALLER_GRAPH
+CALL_GRAPH
+CASE_SENSE_NAMES
+CHM_FILE
+CHM_INDEX_ENCODING
+CITE_BIB_FILES
+CLANG_ASSISTED_PARSING
+CLANG_OPTIONS
+CLASS_DIAGRAMS
+CLASS_GRAPH
+COLLABORATION_GRAPH
+COLS_IN_ALPHA_INDEX
+COMPACT_LATEX
+COMPACT_RTF
+CPP_CLI_SUPPORT
+CREATE_SUBDIRS
+DIAFILE_DIRS
+DIA_PATH
+DIRECTORY_GRAPH
+DISABLE_INDEX
+DISTRIBUTE_GROUP_DOC
+DOCBOOK_OUTPUT
+DOCBOOK_PROGRAMLISTING
+DOCSET_BUNDLE_ID
+DOCSET_FEEDNAME
+DOCSET_PUBLISHER_ID
+DOCSET_PUBLISHER_NAME
+DOTFILE_DIRS
+DOT_CLEANUP
+DOT_FONTNAME
+DOT_FONTPATH
+DOT_FONTSIZE
+DOT_GRAPH_MAX_NODES
+DOT_IMAGE_FORMAT
+DOT_MULTI_TARGETS
+DOT_NUM_THREADS
+# DOT_PATH
+DOT_TRANSPARENT
+DOXYFILE_ENCODING
+ECLIPSE_DOC_ID
+ENABLED_SECTIONS
+ENABLE_PREPROCESSING
+ENUM_VALUES_PER_LINE
+EXAMPLE_PATH
+EXAMPLE_PATTERNS
+EXAMPLE_RECURSIVE
+EXCLUDE
+EXCLUDE_PATTERNS
+EXCLUDE_SYMBOLS
+EXCLUDE_SYMLINKS
+EXPAND_AS_DEFINED
+EXPAND_ONLY_PREDEF
+EXTENSION_MAPPING
+EXTERNAL_GROUPS
+EXTERNAL_PAGES
+EXTERNAL_SEARCH
+EXTERNAL_SEARCH_ID
+EXTRACT_ALL
+EXTRACT_ANON_NSPACES
+EXTRACT_LOCAL_CLASSES
+EXTRACT_LOCAL_METHODS
+EXTRACT_PACKAGE
+EXTRACT_PRIVATE
+EXTRACT_STATIC
+EXTRA_PACKAGES
+EXTRA_SEARCH_MAPPINGS
+EXT_LINKS_IN_WINDOW
+FILE_PATTERNS
+FILE_VERSION_FILTER
+FILTER_PATTERNS
+FILTER_SOURCE_FILES
+FILTER_SOURCE_PATTERNS
+FORCE_LOCAL_INCLUDES
+FORMULA_FONTSIZE
+FORMULA_TRANSPARENT
+FULL_PATH_NAMES
+GENERATE_AUTOGEN_DEF
+GENERATE_BUGLIST
+GENERATE_CHI
+GENERATE_DEPRECATEDLIST
+GENERATE_DOCBOOK
+GENERATE_DOCSET
+GENERATE_ECLIPSEHELP
+GENERATE_HTML
+GENERATE_HTMLHELP
+GENERATE_LATEX
+GENERATE_LEGEND
+GENERATE_MAN
+GENERATE_PERLMOD
+GENERATE_QHP
+GENERATE_RTF
+GENERATE_TAGFILE
+GENERATE_TESTLIST
+GENERATE_TODOLIST
+GENERATE_TREEVIEW
+GENERATE_XML
+GRAPHICAL_HIERARCHY
+GROUP_GRAPHS
+GROUP_NESTED_COMPOUNDS
+# HAVE_DOT
+HHC_LOCATION
+HIDE_COMPOUND_REFERENCE
+HIDE_FRIEND_COMPOUNDS
+HIDE_IN_BODY_DOCS
+HIDE_SCOPE_NAMES
+HIDE_UNDOC_CLASSES
+HIDE_UNDOC_MEMBERS
+HIDE_UNDOC_RELATIONS
+HTML_COLORSTYLE_GAMMA
+HTML_COLORSTYLE_HUE
+HTML_COLORSTYLE_SAT
+HTML_DYNAMIC_SECTIONS
+HTML_EXTRA_FILES
+HTML_EXTRA_STYLESHEET
+HTML_FILE_EXTENSION
+HTML_FOOTER
+HTML_HEADER
+HTML_INDEX_NUM_ENTRIES
+HTML_OUTPUT
+HTML_STYLESHEET
+HTML_TIMESTAMP
+IDL_PROPERTY_SUPPORT
+IGNORE_PREFIX
+IMAGE_PATH
+INCLUDED_BY_GRAPH
+INCLUDE_FILE_PATTERNS
+INCLUDE_GRAPH
+INCLUDE_PATH
+INHERIT_DOCS
+INLINE_GROUPED_CLASSES
+INLINE_INFO
+INLINE_INHERITED_MEMB
+INLINE_SIMPLE_STRUCTS
+INLINE_SOURCES
+INPUT
+INPUT_ENCODING
+INPUT_FILTER
+INTERACTIVE_SVG
+INTERNAL_DOCS
+JAVADOC_AUTOBRIEF
+LATEX_BATCHMODE
+LATEX_BIB_STYLE
+LATEX_CMD_NAME
+LATEX_EXTRA_FILES
+LATEX_EXTRA_STYLESHEET
+LATEX_FOOTER
+LATEX_HEADER
+LATEX_HIDE_INDICES
+LATEX_OUTPUT
+LATEX_SOURCE_CODE
+LATEX_TIMESTAMP
+LAYOUT_FILE
+LOOKUP_CACHE_SIZE
+MACRO_EXPANSION
+MAKEINDEX_CMD_NAME
+MAN_EXTENSION
+MAN_LINKS
+MAN_OUTPUT
+MAN_SUBDIR
+MARKDOWN_SUPPORT
+MATHJAX_CODEFILE
+MATHJAX_EXTENSIONS
+MATHJAX_FORMAT
+MATHJAX_RELPATH
+MAX_DOT_GRAPH_DEPTH
+MAX_INITIALIZER_LINES
+MSCFILE_DIRS
+MSCGEN_PATH
+MULTILINE_CPP_IS_BRIEF
+OPTIMIZE_FOR_FORTRAN
+OPTIMIZE_OUTPUT_FOR_C
+OPTIMIZE_OUTPUT_JAVA
+OPTIMIZE_OUTPUT_VHDL
+OUTPUT_DIRECTORY
+OUTPUT_LANGUAGE
+PAPER_TYPE
+PDF_HYPERLINKS
+PERLMOD_LATEX
+PERLMOD_MAKEVAR_PREFIX
+PERLMOD_PRETTY
+PERL_PATH
+PLANTUML_CFG_FILE
+PLANTUML_INCLUDE_PATH
+PLANTUML_JAR_PATH
+PREDEFINED
+PROJECT_BRIEF
+PROJECT_LOGO
+PROJECT_NAME
+PROJECT_NUMBER
+QCH_FILE
+QHG_LOCATION
+QHP_CUST_FILTER_ATTRS
+QHP_CUST_FILTER_NAME
+QHP_NAMESPACE
+QHP_SECT_FILTER_ATTRS
+QHP_VIRTUAL_FOLDER
+QT_AUTOBRIEF
+QUIET
+RECURSIVE
+REFERENCED_BY_RELATION
+REFERENCES_LINK_SOURCE
+REFERENCES_RELATION
+REPEAT_BRIEF
+RTF_EXTENSIONS_FILE
+RTF_HYPERLINKS
+RTF_OUTPUT
+RTF_SOURCE_CODE
+RTF_STYLESHEET_FILE
+SEARCHDATA_FILE
+SEARCHENGINE
+SEARCHENGINE_URL
+SEARCH_INCLUDES
+SEPARATE_MEMBER_PAGES
+SERVER_BASED_SEARCH
+SHORT_NAMES
+SHOW_FILES
+SHOW_GROUPED_MEMB_INC
+SHOW_INCLUDE_FILES
+SHOW_NAMESPACES
+SHOW_USED_FILES
+SIP_SUPPORT
+SKIP_FUNCTION_MACROS
+SORT_BRIEF_DOCS
+SORT_BY_SCOPE_NAME
+SORT_GROUP_NAMES
+SORT_MEMBERS_CTORS_1ST
+SORT_MEMBER_DOCS
+SOURCE_BROWSER
+SOURCE_TOOLTIPS
+STRICT_PROTO_MATCHING
+STRIP_CODE_COMMENTS
+STRIP_FROM_INC_PATH
+STRIP_FROM_PATH
+SUBGROUPING
+TAB_SIZE
+TAGFILES
+TCL_SUBST
+TEMPLATE_RELATIONS
+TOC_EXPAND
+TOC_INCLUDE_HEADINGS
+TREEVIEW_WIDTH
+TYPEDEF_HIDES_STRUCT
+UML_LIMIT_NUM_FIELDS
+UML_LOOK
+USE_HTAGS
+USE_MATHJAX
+USE_MDFILE_AS_MAINPAGE
+USE_PDFLATEX
+VERBATIM_HEADERS
+WARNINGS
+WARN_AS_ERROR
+WARN_FORMAT
+WARN_IF_DOC_ERROR
+WARN_IF_UNDOCUMENTED
+WARN_LOGFILE
+WARN_NO_PARAMDOC
+XML_OUTPUT
+XML_PROGRAMLISTING
+)
+
+set(DOXYGEN_CONFIG_FILE "${CMAKE_CURRENT_BINARY_DIR}/doxygen/doxygen.conf" CACHE PATH "Path to generated doxygen configuration file")
+
+function(add_doxygen_doc)
+    set(options)
+    set(oneValueArgs)
+    set(multiValueArgs DEPENDS ${DOXYGEN_ARGS})
+
+    cmake_parse_arguments(PARSE "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+
+    file(WRITE ${DOXYGEN_CONFIG_FILE} "# Auto-generated doxygen configuration file\n")
+
+    foreach(ARG ${DOXYGEN_ARGS})
+        if(PARSE_${ARG})
+            string(REPLACE ";" " " ARG_VALUE ${PARSE_${ARG}})
+            file(APPEND ${DOXYGEN_CONFIG_FILE} "\n${ARG} = ${ARG_VALUE}\n")
+        endif()
+    endforeach()
+
+    if(PARSE_OUTPUT_DIRECTORY)
+        if(NOT EXISTS ${PARSE_OUTPUT_DIRECTORY})
+            file(MAKE_DIRECTORY ${PARSE_OUTPUT_DIRECTORY})
+        endif()
+    endif()
+
+    if(DOT_EXECUTABLE)
+        file(APPEND ${DOXYGEN_CONFIG_FILE} "\nDOT_PATH = \"${DOT_EXECUTABLE}\"\n")
+        file(APPEND ${DOXYGEN_CONFIG_FILE} "\nHAVE_DOT = YES\n")
+    else()
+        file(APPEND ${DOXYGEN_CONFIG_FILE} "\nHAVE_DOT = NO\n")
+    endif()
+
+    add_custom_target(doxygen
+        ${DOXYGEN_EXECUTABLE} ${DOXYGEN_CONFIG_FILE}
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        COMMENT "Building documentation with doxygen"
+    )
+    if(PARSE_OUTPUT_DIRECTORY)
+        clean_doc_output(${PARSE_OUTPUT_DIRECTORY})
+    endif()
+    mark_as_doc(doxygen)
+    if(PARSE_DEPENDS)
+        add_dependencies(doxygen ${PARSE_DEPENDS})
+    endif()
+endfunction()
diff --git a/cmake/EnableCompilerWarnings.cmake b/cmake/EnableCompilerWarnings.cmake
new file mode 100644
index 0000000000..0d9d759d02
--- /dev/null
+++ b/cmake/EnableCompilerWarnings.cmake
@@ -0,0 +1,114 @@
+################################################################################
+#
+# MIT License
+#
+# Copyright (c) 2017 Advanced Micro Devices, Inc.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+#
+################################################################################
+# - Enable warning all for gcc/clang or use /W4 for visual studio
+
+## Strict warning level
+if (MSVC)
+    # Use the highest warning level for visual studio.
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /w")
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /w")
+    # set(CMAKE_CXX_WARNING_LEVEL 4)
+    # if (CMAKE_CXX_FLAGS MATCHES "/W[0-4]")
+    #     string(REGEX REPLACE "/W[0-4]" "/W4" CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
+    # else ()
+    #     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /W4")
+    # endif ()
+
+    # set(CMAKE_C_WARNING_LEVEL 4)
+    # if (CMAKE_C_FLAGS MATCHES "/W[0-4]")
+    #     string(REGEX REPLACE "/W[0-4]" "/W4" CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
+    # else ()
+    #     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /W4")
+    # endif ()
+
+else()
+    foreach(COMPILER C CXX)
+        set(CMAKE_COMPILER_WARNINGS)
+        # use -Wall for gcc and clang
+        list(APPEND CMAKE_COMPILER_WARNINGS
+            -Wall
+            -Wextra
+            -Wcomment
+            -Wendif-labels
+            -Wformat
+            -Winit-self
+            -Wreturn-type
+            -Wsequence-point
+            # Shadow is broken on gcc when using lambdas
+            # -Wshadow
+            -Wswitch
+            -Wtrigraphs
+            -Wundef
+            -Wuninitialized
+            -Wunreachable-code
+            -Wunused
+            -Wno-reserved-identifier
+            -Werror
+            -Wno-option-ignored
+            -Wsign-compare
+            -Wno-extra-semi-stmt
+        )
+        if (CMAKE_${COMPILER}_COMPILER_ID MATCHES "Clang")
+            list(APPEND CMAKE_COMPILER_WARNINGS
+                -Weverything
+                -Wno-c++98-compat
+                -Wno-c++98-compat-pedantic
+                -Wno-conversion
+                -Wno-double-promotion
+                -Wno-exit-time-destructors
+                -Wno-extra-semi
+                -Wno-float-conversion
+                -Wno-gnu-anonymous-struct
+                -Wno-gnu-zero-variadic-macro-arguments
+                -Wno-missing-prototypes
+                -Wno-nested-anon-types
+                -Wno-padded
+                -Wno-return-std-move-in-c++11
+                -Wno-shorten-64-to-32
+                -Wno-sign-conversion
+                -Wno-unknown-warning-option
+                -Wno-unused-command-line-argument
+                -Wno-weak-vtables
+                -Wno-covered-switch-default
+                -Wno-unused-lambda-capture
+                -Wno-unsafe-buffer-usage
+            )
+        else()
+            if (CMAKE_${COMPILER}_COMPILER_ID MATCHES "GNU" AND ${COMPILER} MATCHES "CXX")
+                # cmake 3.5.2 does not support >=.
+                if(NOT CMAKE_CXX_COMPILER_VERSION VERSION_LESS "6.1")
+                    list(APPEND CMAKE_COMPILER_WARNINGS
+                        -Wno-ignored-attributes)
+                endif()
+            endif()
+            list(APPEND CMAKE_COMPILER_WARNINGS
+                -Wno-missing-field-initializers
+                -Wno-deprecated-declarations
+            )
+        endif()
+        add_definitions(${CMAKE_COMPILER_WARNINGS})
+    endforeach()
+endif ()
diff --git a/cmake/TargetFlags.cmake b/cmake/TargetFlags.cmake
new file mode 100644
index 0000000000..4f83fb5d39
--- /dev/null
+++ b/cmake/TargetFlags.cmake
@@ -0,0 +1,50 @@
+
+function(get_target_property2 VAR TARGET PROPERTY)
+    get_target_property(_pflags ${TARGET} ${PROPERTY})
+    if(_pflags)
+        set(${VAR} ${_pflags} PARENT_SCOPE)
+    else()
+        set(${VAR} "" PARENT_SCOPE)
+    endif()
+endfunction()
+
+
+macro(append_flags FLAGS TARGET PROPERTY PREFIX)
+    get_target_property2(_pflags ${TARGET} ${PROPERTY})
+    foreach(FLAG ${_pflags})
+        if(TARGET ${FLAG})
+            target_flags(_pflags2 ${FLAG})
+            string(APPEND ${FLAGS} " ${_pflags2}")
+        else()
+            string(APPEND ${FLAGS} " ${PREFIX}${FLAG}")
+        endif()
+    endforeach()
+endmacro()
+
+macro(append_link_flags FLAGS TARGET PROPERTY)
+    get_target_property2(_pflags ${TARGET} ${PROPERTY})
+    foreach(FLAG ${_pflags})
+        if(TARGET ${FLAG})
+            target_flags(_pflags2 ${FLAG})
+            string(APPEND ${FLAGS} " ${_pflags2}")
+        elseif(FLAG MATCHES "^-.*")
+            string(APPEND ${FLAGS} " ${FLAG}")
+        elseif(EXISTS ${FLAG})
+            string(APPEND ${FLAGS} " ${FLAG}")
+        else()
+            string(APPEND ${FLAGS} " -l${FLAG}")
+        endif()
+    endforeach()
+endmacro()
+
+function(target_flags FLAGS TARGET)
+    set(_flags)
+    append_flags(_flags ${TARGET} "INTERFACE_COMPILE_OPTIONS" "")
+    append_flags(_flags ${TARGET} "INTERFACE_COMPILE_DEFINITIONS" "-D")
+    append_flags(_flags ${TARGET} "INTERFACE_INCLUDE_DIRECTORIES" "-isystem ")
+    append_flags(_flags ${TARGET} "INTERFACE_LINK_DIRECTORIES" "-L ")
+    append_flags(_flags ${TARGET} "INTERFACE_LINK_OPTIONS" "")
+    append_link_flags(_flags ${TARGET} "INTERFACE_LINK_LIBRARIES" "")
+    # message("_flags: ${_flags}")
+    set(${FLAGS} ${_flags} PARENT_SCOPE)
+endfunction()
diff --git a/cmake/googletest.cmake b/cmake/googletest.cmake
new file mode 100644
index 0000000000..d6577ac33e
--- /dev/null
+++ b/cmake/googletest.cmake
@@ -0,0 +1,50 @@
+include(FetchContent)
+
+set(GOOGLETEST_DIR "" CACHE STRING "Location of local GoogleTest repo to build against")
+
+if(GOOGLETEST_DIR)
+  set(FETCHCONTENT_SOURCE_DIR_GOOGLETEST ${GOOGLETEST_DIR} CACHE STRING "GoogleTest source directory override")
+endif()
+
+message(STATUS "Fetching GoogleTest")
+
+list(APPEND GTEST_CMAKE_CXX_FLAGS
+     -Wno-undef
+     -Wno-reserved-identifier
+     -Wno-global-constructors
+     -Wno-missing-noreturn
+     -Wno-disabled-macro-expansion
+     -Wno-used-but-marked-unused
+     -Wno-switch-enum
+     -Wno-zero-as-null-pointer-constant
+     -Wno-unused-member-function
+     -Wno-comma
+     -Wno-old-style-cast
+     -Wno-deprecated
+     -Wno-unsafe-buffer-usage
+)
+message(STATUS "Suppressing googltest warnings with flags: ${GTEST_CMAKE_CXX_FLAGS}")
+
+FetchContent_Declare(
+  googletest
+  GIT_REPOSITORY https://github.com/google/googletest.git
+  GIT_TAG        b85864c64758dec007208e56af933fc3f52044ee
+)
+
+# Will be necessary for windows build
+# set(gtest_force_shared_crt ON CACHE BOOL "" FORCE)
+FetchContent_GetProperties(googletest)
+if(NOT googletest_POPULATED)
+  FetchContent_Populate(googletest)
+  add_subdirectory(${googletest_SOURCE_DIR} ${googletest_BINARY_DIR} EXCLUDE_FROM_ALL)
+endif()
+
+target_compile_options(gtest PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
+target_compile_options(gtest_main PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
+target_compile_options(gmock PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
+target_compile_options(gmock_main PRIVATE ${GTEST_CMAKE_CXX_FLAGS})
+
+set_target_properties(gtest PROPERTIES POSITION_INDEPENDENT_CODE ON)
+set_target_properties(gtest_main PROPERTIES POSITION_INDEPENDENT_CODE ON)
+set_target_properties(gmock PROPERTIES POSITION_INDEPENDENT_CODE ON)
+set_target_properties(gmock_main PROPERTIES POSITION_INDEPENDENT_CODE ON)
diff --git a/dev-requirements.txt b/dev-requirements.txt
new file mode 100644
index 0000000000..9e7b9f01e1
--- /dev/null
+++ b/dev-requirements.txt
@@ -0,0 +1,3 @@
+ROCmSoftwarePlatform/rocm-recipes
+RadeonOpenCompute/rocm-cmake@04f694df2a8dc9d7e35fa4dee4ba5fa407ec04f8 --build
+danmar/cppcheck@2.9
\ No newline at end of file
diff --git a/docs/API_Reference_Guide.rst b/docs/API_Reference_Guide.rst
new file mode 100644
index 0000000000..f21d43c593
--- /dev/null
+++ b/docs/API_Reference_Guide.rst
@@ -0,0 +1,52 @@
+
+*******************
+API Reference Guide
+*******************
+
+=================
+Introduction
+=================
+
+This document contains details of the APIs for the Composable Kernel (CK) library and introduces
+some of the key design principles that are used to write new classes that extend CK functionality.
+
+=================
+Using CK API
+=================
+
+This section describes how to use the CK library API.
+
+=================
+CK Datatypes
+=================
+
+-----------------
+DeviceMem
+-----------------
+
+.. doxygenstruct:: DeviceMem
+
+---------------------------
+Kernels For Flashattention
+---------------------------
+
+The Flashattention algorithm is defined in :cite:t:`dao2022flashattention`. This sections lists
+the classes that are used in the CK GPU implementation of Flashattention.
+
+**Gridwise classes**
+
+.. doxygenstruct:: ck::GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
+
+**Blockwise classes**
+
+.. doxygenstruct:: ck::ThreadGroupTensorSliceTransfer_v4r1
+
+.. doxygenstruct:: ck::BlockwiseGemmXdlops_v2
+
+.. doxygenstruct:: ck::BlockwiseSoftmax
+
+**Threadwise classes**
+
+.. doxygenstruct:: ck::ThreadwiseTensorSliceTransfer_StaticToStatic
+
+.. bibliography::
diff --git a/docs/Contributors_Guide.rst b/docs/Contributors_Guide.rst
new file mode 100644
index 0000000000..b2ddff398c
--- /dev/null
+++ b/docs/Contributors_Guide.rst
@@ -0,0 +1,8 @@
+===================
+Contributor's Guide
+===================
+
+Pull-request guidelines
+=======================
+
+[TODO]
diff --git a/docs/Supported_Primitives_Guide.rst b/docs/Supported_Primitives_Guide.rst
new file mode 100644
index 0000000000..3462283d90
--- /dev/null
+++ b/docs/Supported_Primitives_Guide.rst
@@ -0,0 +1,76 @@
+==========================
+Supported Primitives Guide
+==========================
+
+This document contains details of supported primitives in Composable Kernel (CK). In contrast to the
+API Reference Guide, the Supported Primitives Guide is an introduction to the math which underpins
+the algorithms implemented in CK.
+
+------------
+Softmax
+------------
+
+For vectors :math:`x^{(1)}, x^{(2)}, \ldots, x^{(T)}` of size :math:`B` we can decompose the
+softmax of concatenated :math:`x = [ x^{(1)}\ | \ \ldots \ | \ x^{(T)} ]` as,
+
+.. math::
+   :nowrap:
+
+   \begin{align}
+      m(x) & = m( [ x^{(1)}\ | \ \ldots \ | \ x^{(T)} ] ) = \max( m(x^{(1)}),\ldots, m(x^{(T)}) )  \\
+      f(x) & = [\exp( m(x^{(1)}) - m(x) ) f( x^{(1)} )\ | \ \ldots \ | \ \exp( m(x^{(T)}) - m(x) ) f( x^{(T)} )] \\
+      z(x) & = \exp( m(x^{(1)}) - m(x) )\ z(x^{(1)}) + \ldots + \exp( m(x^{(T)}) - m(x) )\ z(x^{(1)}) \\
+      \operatorname{softmax}(x) &= f(x)\ / \ z(x)
+   \end{align}
+
+where :math:`f(x^{(j)}) = \exp( x^{(j)} - m(x^{(j)}) )` is of size :math:`B` and
+:math:`z(x^{(j)}) = f(x_1^{(j)})+ \ldots+ f(x_B^{(j)})` is a scalar.
+
+For a matrix :math:`X` composed of :math:`T_r \times T_c` tiles, :math:`X_{ij}`, of size
+:math:`B_r \times B_c` we can compute the row-wise softmax as follows.
+
+For :math:`j` from :math:`1` to :math:`T_c`, and :math:`i` from :math:`1` to :math:`T_r` calculate,
+
+.. math::
+   :nowrap:
+
+   \begin{align}
+      \tilde{m}_{ij}   &= \operatorname{rowmax}( X_{ij} ) \\
+      \tilde{P}_{ij}   &= \exp(X_{ij} - \tilde{m}_{ij} ) \\
+      \tilde{z}_{ij}   &= \operatorname{rowsum}( P_{ij} ) \\
+   \end{align}
+
+If :math:`j=1`, initialize running max, running sum, and the first column block of the output,
+
+.. math::
+   :nowrap:
+
+   \begin{align}
+      m_i            &= \tilde{m}_{i1} \\
+      z_i            &= \tilde{z}_{i1} \\
+      \tilde{Y}_{i1} &= \diag(\tilde{z}_{ij})^{-1} \tilde{P}_{i1}
+   \end{align}
+
+Else if :math:`j>1`,
+
+1. Update running max, running sum and column blocks :math:`k=1` to :math:`k=j-1`
+
+.. math::
+   :nowrap:
+
+   \begin{align}
+      m^{new}_i &= \max(m_i, \tilde{m}_{ij} ) \\
+      z^{new}_i &= \exp(m_i - m^{new}_i)\ z_i + \exp( \tilde{m}_{ij} - m^{new}_i )\ \tilde{z}_{ij}  \\
+      Y_{ik}    &= \diag(z^{new}_{i})^{-1} \diag(z_{i}) \exp(m_i - m^{new}_i)\ Y_{ik}
+   \end{align}
+
+2. Initialize column block :math:`j` of output and reset running max and running sum variables:
+
+.. math::
+   :nowrap:
+
+   \begin{align}
+      \tilde{Y}_{ij} &= \diag(z^{new}_{i})^{-1} \exp(\tilde{m}_{ij} - m^{new}_i ) \tilde{P}_{ij} \\
+      z_i            &= z^{new}_i \\
+      m_i            &= m^{new}_i \\
+   \end{align}
diff --git a/docs/conf.py b/docs/conf.py
new file mode 100644
index 0000000000..0de590da1a
--- /dev/null
+++ b/docs/conf.py
@@ -0,0 +1,36 @@
+# Configuration file for the Sphinx documentation builder.
+#
+# This file only contains a selection of the most common options. For a full
+# list see the documentation:
+# https://www.sphinx-doc.org/en/master/usage/configuration.html
+
+import subprocess
+
+from rocm_docs import ROCmDocs
+
+
+name = "Composable Kernel"
+get_version = r'sed -n -e "s/^rocm_setup_version(.* \([0-9\.]\{1,\}\).*/\1/p" ../CMakeLists.txt'
+version = subprocess.getoutput(get_version)
+if len(version) > 0:
+    name = f"{name} {version}"
+
+external_toc_path = "./sphinx/_toc.yml"
+
+docs_core = ROCmDocs(f"{name} Documentation")
+docs_core.run_doxygen(doxygen_root="doxygen", doxygen_path="doxygen/docBin/xml")
+docs_core.setup()
+
+mathjax3_config = {
+'tex': {
+    'macros': {
+        'diag': '\\operatorname{diag}',
+        }
+    }
+}
+
+for sphinx_var in ROCmDocs.SPHINX_VARS:
+    globals()[sphinx_var] = getattr(docs_core, sphinx_var)
+
+extensions += ['sphinxcontrib.bibtex']
+bibtex_bibfiles = ['refs.bib']
diff --git a/docs/data/ck_component.png b/docs/data/ck_component.png
new file mode 100644
index 0000000000..db892331d7
Binary files /dev/null and b/docs/data/ck_component.png differ
diff --git a/docs/data/ck_layer.png b/docs/data/ck_layer.png
new file mode 100644
index 0000000000..117a1b3a0e
Binary files /dev/null and b/docs/data/ck_layer.png differ
diff --git a/docs/dockerhub.rst b/docs/dockerhub.rst
new file mode 100644
index 0000000000..66ec91096e
--- /dev/null
+++ b/docs/dockerhub.rst
@@ -0,0 +1,99 @@
+===================
+CK Docker Hub
+===================
+
+-------------------------------------
+Why do I need this?
+-------------------------------------
+
+To make our lives easier and bring Composable Kernel dependencies together, we recommend using
+docker images that can be found on `Docker Hub <https://hub.docker.com/r/rocm/composable_kernel>`_.
+
+-------------------------------------
+So what is Composable Kernel?
+-------------------------------------
+
+Composable Kernel (CK) library aims to provide a programming model for writing performance critical
+kernels for machine learning workloads across multiple architectures including GPUs, CPUs, etc,
+through general purpose kernel languages, like HIP C++.
+
+To get the CK library::
+
+    git clone https://github.com/ROCmSoftwarePlatform/composable_kernel.git
+
+
+run a docker container::
+
+    docker run                                                            \
+    -it                                                                   \
+    --privileged                                                          \
+    --group-add sudo                                                      \
+    -w /root/workspace                                                    \
+    -v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace                         \
+    rocm/composable_kernel:ck_ub20.04_rocm5.6                             \
+    /bin/bash
+
+and build the CK::
+
+    mkdir build && cd build
+    # Need to specify target ID, example below is for gfx908 and gfx90a
+    cmake                                                                                             \
+    -D CMAKE_PREFIX_PATH=/opt/rocm                                                                    \
+    -D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                         \
+    -D CMAKE_CXX_FLAGS="-O3"                                                                          \
+    -D CMAKE_BUILD_TYPE=Release                                                                       \
+    -D GPU_TARGETS="gfx908;gfx90a"                                                                    \
+    ..
+
+and::
+
+    make -j examples tests
+
+To run all the test cases including tests and examples run::
+
+    make test
+
+We can also run specific examples or tests like::
+
+    ./bin/example_gemm_xdl_fp16
+    ./bin/test_gemm_fp16
+
+For more details visit `CK github repository <https://github.com/ROCmSoftwarePlatform/composable_kernel>`_,
+`CK examples <https://github.com/ROCmSoftwarePlatform/composable_kernel/tree/develop/example)>`_,
+`even more CK examples <https://github.com/ROCmSoftwarePlatform/composable_kernel/tree/develop/client_example>`_.
+
+-------------------------------------
+And what is inside?
+-------------------------------------
+
+The docker images have everything you need for running CK including:
+
+* `ROCm <https://www.amd.com/en/graphics/servers-solutions-rocm>`_
+* `CMake <https://cmake.org/>`_
+* `Compiler <https://github.com/RadeonOpenCompute/llvm-project>`_
+
+-------------------------------------
+Which image is right for me?
+-------------------------------------
+
+Let's take a look at the image naming, for example ``ck_ub20.04_rocm5.6``. The image specs are:
+
+* ``ck`` - made for running Composable Kernel;
+* ``ub20.04`` - based on Ubuntu 20.04;
+* ``rocm5.6`` - ROCm platform version 5.6.
+
+So just pick the right image for your project dependencies and you're all set.
+
+-------------------------------------
+DIY starts here
+-------------------------------------
+
+If you need to customize a docker image or just can't stop tinkering, feel free to adjust the
+`Dockerfile <https://github.com/ROCmSoftwarePlatform/composable_kernel/blob/develop/Dockerfile>`_
+for your needs.
+
+-------------------------------------
+License
+-------------------------------------
+
+CK is released under the MIT `license <https://github.com/ROCmSoftwarePlatform/composable_kernel/blob/develop/LICENSE>`_.
diff --git a/docs/doxygen/Doxyfile b/docs/doxygen/Doxyfile
new file mode 100644
index 0000000000..1084f94c81
--- /dev/null
+++ b/docs/doxygen/Doxyfile
@@ -0,0 +1,2455 @@
+# Doxyfile 1.8.10
+
+# This file describes the settings to be used by the documentation system
+# doxygen (www.doxygen.org) for a project.
+#
+# All text after a double hash (##) is considered a comment and is placed in
+# front of the TAG it is preceding.
+#
+# All text after a single hash (#) is considered a comment and will be ignored.
+# The format is:
+# TAG = value [value, ...]
+# For lists, items can also be appended using:
+# TAG += value [value, ...]
+# Values that contain spaces should be placed between quotes (\" \").
+
+#---------------------------------------------------------------------------
+# Project related configuration options
+#---------------------------------------------------------------------------
+
+# This tag specifies the encoding used for all characters in the config file
+# that follow. The default is UTF-8 which is also the encoding used for all text
+# before the first occurrence of this tag. Doxygen uses libiconv (or the iconv
+# built into libc) for the transcoding. See http://www.gnu.org/software/libiconv
+# for the list of possible encodings.
+# The default value is: UTF-8.
+
+DOXYFILE_ENCODING      = UTF-8
+
+# The PROJECT_NAME tag is a single word (or a sequence of words surrounded by
+# double-quotes, unless you are using Doxywizard) that should identify the
+# project for which the documentation is generated. This name is used in the
+# title of most generated pages and in a few other places.
+# The default value is: My Project.
+
+PROJECT_NAME           = "ck"
+
+# The PROJECT_NUMBER tag can be used to enter a project or revision number. This
+# could be handy for archiving the generated documentation or if some version
+# control system is used.
+
+PROJECT_NUMBER         = v3.0.1.0
+
+# Using the PROJECT_BRIEF tag one can provide an optional one line description
+# for a project that appears at the top of each page and should give viewer a
+# quick idea about the purpose of the project. Keep the description short.
+
+PROJECT_BRIEF          = "prototype interfaces compatible with ROCm platform and HiP"
+
+# With the PROJECT_LOGO tag one can specify a logo or an icon that is included
+# in the documentation. The maximum height of the logo should not exceed 55
+# pixels and the maximum width should not exceed 200 pixels. Doxygen will copy
+# the logo to the output directory.
+
+PROJECT_LOGO           = 
+
+# The OUTPUT_DIRECTORY tag is used to specify the (relative or absolute) path
+# into which the generated documentation will be written. If a relative path is
+# entered, it will be relative to the location where doxygen was started. If
+# left blank the current directory will be used.
+
+OUTPUT_DIRECTORY       = docBin
+
+# If the CREATE_SUBDIRS tag is set to YES then doxygen will create 4096 sub-
+# directories (in 2 levels) under the output directory of each output format and
+# will distribute the generated files over these directories. Enabling this
+# option can be useful when feeding doxygen a huge amount of source files, where
+# putting all generated files in the same directory would otherwise causes
+# performance problems for the file system.
+# The default value is: NO.
+
+CREATE_SUBDIRS         = NO
+
+# If the ALLOW_UNICODE_NAMES tag is set to YES, doxygen will allow non-ASCII
+# characters to appear in the names of generated files. If set to NO, non-ASCII
+# characters will be escaped, for example _xE3_x81_x84 will be used for Unicode
+# U+3044.
+# The default value is: NO.
+
+ALLOW_UNICODE_NAMES    = NO
+
+# The OUTPUT_LANGUAGE tag is used to specify the language in which all
+# documentation generated by doxygen is written. Doxygen will use this
+# information to generate all constant output in the proper language.
+# Possible values are: Afrikaans, Arabic, Armenian, Brazilian, Catalan, Chinese,
+# Chinese-Traditional, Croatian, Czech, Danish, Dutch, English (United States),
+# Esperanto, Farsi (Persian), Finnish, French, German, Greek, Hungarian,
+# Indonesian, Italian, Japanese, Japanese-en (Japanese with English messages),
+# Korean, Korean-en (Korean with English messages), Latvian, Lithuanian,
+# Macedonian, Norwegian, Persian (Farsi), Polish, Portuguese, Romanian, Russian,
+# Serbian, Serbian-Cyrillic, Slovak, Slovene, Spanish, Swedish, Turkish,
+# Ukrainian and Vietnamese.
+# The default value is: English.
+
+OUTPUT_LANGUAGE        = English
+
+# If the BRIEF_MEMBER_DESC tag is set to YES, doxygen will include brief member
+# descriptions after the members that are listed in the file and class
+# documentation (similar to Javadoc). Set to NO to disable this.
+# The default value is: YES.
+
+BRIEF_MEMBER_DESC      = YES
+
+# If the REPEAT_BRIEF tag is set to YES, doxygen will prepend the brief
+# description of a member or function before the detailed description
+#
+# Note: If both HIDE_UNDOC_MEMBERS and BRIEF_MEMBER_DESC are set to NO, the
+# brief descriptions will be completely suppressed.
+# The default value is: YES.
+
+REPEAT_BRIEF           = YES
+
+# This tag implements a quasi-intelligent brief description abbreviator that is
+# used to form the text in various listings. Each string in this list, if found
+# as the leading text of the brief description, will be stripped from the text
+# and the result, after processing the whole list, is used as the annotated
+# text. Otherwise, the brief description is used as-is. If left blank, the
+# following values are used ($name is automatically replaced with the name of
+# the entity):The $name class, The $name widget, The $name file, is, provides,
+# specifies, contains, represents, a, an and the.
+
+ABBREVIATE_BRIEF       = "The $name class" \
+                         "The $name widget" \
+                         "The $name file" \
+                         is \
+                         provides \
+                         specifies \
+                         contains \
+                         represents \
+                         a \
+                         an \
+                         the
+
+# If the ALWAYS_DETAILED_SEC and REPEAT_BRIEF tags are both set to YES then
+# doxygen will generate a detailed section even if there is only a brief
+# description.
+# The default value is: NO.
+
+ALWAYS_DETAILED_SEC    = NO
+
+# If the INLINE_INHERITED_MEMB tag is set to YES, doxygen will show all
+# inherited members of a class in the documentation of that class as if those
+# members were ordinary class members. Constructors, destructors and assignment
+# operators of the base classes will not be shown.
+# The default value is: NO.
+
+INLINE_INHERITED_MEMB  = NO
+
+# If the FULL_PATH_NAMES tag is set to YES, doxygen will prepend the full path
+# before files name in the file list and in the header files. If set to NO the
+# shortest path that makes the file name unique will be used
+# The default value is: YES.
+
+FULL_PATH_NAMES        = YES
+
+# The STRIP_FROM_PATH tag can be used to strip a user-defined part of the path.
+# Stripping is only done if one of the specified strings matches the left-hand
+# part of the path. The tag can be used to show relative paths in the file list.
+# If left blank the directory from which doxygen is run is used as the path to
+# strip.
+#
+# Note that you can specify absolute paths here, but also relative paths, which
+# will be relative from the directory where doxygen is started.
+# This tag requires that the tag FULL_PATH_NAMES is set to YES.
+
+STRIP_FROM_PATH        =
+
+# The STRIP_FROM_INC_PATH tag can be used to strip a user-defined part of the
+# path mentioned in the documentation of a class, which tells the reader which
+# header file to include in order to use a class. If left blank only the name of
+# the header file containing the class definition is used. Otherwise one should
+# specify the list of include paths that are normally passed to the compiler
+# using the -I flag.
+
+STRIP_FROM_INC_PATH    =
+
+# If the SHORT_NAMES tag is set to YES, doxygen will generate much shorter (but
+# less readable) file names. This can be useful is your file systems doesn't
+# support long names like on DOS, Mac, or CD-ROM.
+# The default value is: NO.
+
+SHORT_NAMES            = NO
+
+# If the JAVADOC_AUTOBRIEF tag is set to YES then doxygen will interpret the
+# first line (until the first dot) of a Javadoc-style comment as the brief
+# description. If set to NO, the Javadoc-style will behave just like regular Qt-
+# style comments (thus requiring an explicit @brief command for a brief
+# description.)
+# The default value is: NO.
+
+JAVADOC_AUTOBRIEF      = NO
+
+# If the QT_AUTOBRIEF tag is set to YES then doxygen will interpret the first
+# line (until the first dot) of a Qt-style comment as the brief description. If
+# set to NO, the Qt-style will behave just like regular Qt-style comments (thus
+# requiring an explicit \brief command for a brief description.)
+# The default value is: NO.
+
+QT_AUTOBRIEF           = NO
+
+# The MULTILINE_CPP_IS_BRIEF tag can be set to YES to make doxygen treat a
+# multi-line C++ special comment block (i.e. a block of //! or /// comments) as
+# a brief description. This used to be the default behavior. The new default is
+# to treat a multi-line C++ comment block as a detailed description. Set this
+# tag to YES if you prefer the old behavior instead.
+#
+# Note that setting this tag to YES also means that rational rose comments are
+# not recognized any more.
+# The default value is: NO.
+
+MULTILINE_CPP_IS_BRIEF = NO
+
+# If the INHERIT_DOCS tag is set to YES then an undocumented member inherits the
+# documentation from any documented member that it re-implements.
+# The default value is: YES.
+
+INHERIT_DOCS           = YES
+
+# If the SEPARATE_MEMBER_PAGES tag is set to YES then doxygen will produce a new
+# page for each member. If set to NO, the documentation of a member will be part
+# of the file/class/namespace that contains it.
+# The default value is: NO.
+
+SEPARATE_MEMBER_PAGES  = NO
+
+# The TAB_SIZE tag can be used to set the number of spaces in a tab. Doxygen
+# uses this value to replace tabs by spaces in code fragments.
+# Minimum value: 1, maximum value: 16, default value: 4.
+
+TAB_SIZE               = 4
+
+# This tag can be used to specify a number of aliases that act as commands in
+# the documentation. An alias has the form:
+# name=value
+# For example adding
+# "sideeffect=@par Side Effects:\n"
+# will allow you to put the command \sideeffect (or @sideeffect) in the
+# documentation, which will result in a user-defined paragraph with heading
+# "Side Effects:". You can put \n's in the value part of an alias to insert
+# newlines.
+
+ALIASES                =
+
+# This tag can be used to specify a number of word-keyword mappings (TCL only).
+# A mapping has the form "name=value". For example adding "class=itcl::class"
+# will allow you to use the command class in the itcl::class meaning.
+
+TCL_SUBST              =
+
+# Set the OPTIMIZE_OUTPUT_FOR_C tag to YES if your project consists of C sources
+# only. Doxygen will then generate output that is more tailored for C. For
+# instance, some of the names that are used will be different. The list of all
+# members will be omitted, etc.
+# The default value is: NO.
+
+OPTIMIZE_OUTPUT_FOR_C  = NO
+
+# Set the OPTIMIZE_OUTPUT_JAVA tag to YES if your project consists of Java or
+# Python sources only. Doxygen will then generate output that is more tailored
+# for that language. For instance, namespaces will be presented as packages,
+# qualified scopes will look different, etc.
+# The default value is: NO.
+
+OPTIMIZE_OUTPUT_JAVA   = NO
+
+# Set the OPTIMIZE_FOR_FORTRAN tag to YES if your project consists of Fortran
+# sources. Doxygen will then generate output that is tailored for Fortran.
+# The default value is: NO.
+
+OPTIMIZE_FOR_FORTRAN   = NO
+
+# Set the OPTIMIZE_OUTPUT_VHDL tag to YES if your project consists of VHDL
+# sources. Doxygen will then generate output that is tailored for VHDL.
+# The default value is: NO.
+
+OPTIMIZE_OUTPUT_VHDL   = NO
+
+# Doxygen selects the parser to use depending on the extension of the files it
+# parses. With this tag you can assign which parser to use for a given
+# extension. Doxygen has a built-in mapping, but you can override or extend it
+# using this tag. The format is ext=language, where ext is a file extension, and
+# language is one of the parsers supported by doxygen: IDL, Java, Javascript,
+# C#, C, C++, D, PHP, Objective-C, Python, Fortran (fixed format Fortran:
+# FortranFixed, free formatted Fortran: FortranFree, unknown formatted Fortran:
+# Fortran. In the later case the parser tries to guess whether the code is fixed
+# or free formatted code, this is the default for Fortran type files), VHDL. For
+# instance to make doxygen treat .inc files as Fortran files (default is PHP),
+# and .f files as C (default is Fortran), use: inc=Fortran f=C.
+#
+# Note: For files without extension you can use no_extension as a placeholder.
+#
+# Note that for custom extensions you also need to set FILE_PATTERNS otherwise
+# the files are not read by doxygen.
+
+EXTENSION_MAPPING      =
+
+# If the MARKDOWN_SUPPORT tag is enabled then doxygen pre-processes all comments
+# according to the Markdown format, which allows for more readable
+# documentation. See http://daringfireball.net/projects/markdown/ for details.
+# The output of markdown processing is further processed by doxygen, so you can
+# mix doxygen, HTML, and XML commands with Markdown formatting. Disable only in
+# case of backward compatibilities issues.
+# The default value is: YES.
+
+MARKDOWN_SUPPORT       = YES
+
+# When enabled doxygen tries to link words that correspond to documented
+# classes, or namespaces to their corresponding documentation. Such a link can
+# be prevented in individual cases by putting a % sign in front of the word or
+# globally by setting AUTOLINK_SUPPORT to NO.
+# The default value is: YES.
+
+AUTOLINK_SUPPORT       = YES
+
+# If you use STL classes (i.e. std::string, std::vector, etc.) but do not want
+# to include (a tag file for) the STL sources as input, then you should set this
+# tag to YES in order to let doxygen match functions declarations and
+# definitions whose arguments contain STL classes (e.g. func(std::string);
+# versus func(std::string) {}). This also make the inheritance and collaboration
+# diagrams that involve STL classes more complete and accurate.
+# The default value is: NO.
+
+BUILTIN_STL_SUPPORT    = YES
+
+# If you use Microsoft's C++/CLI language, you should set this option to YES to
+# enable parsing support.
+# The default value is: NO.
+
+CPP_CLI_SUPPORT        = NO
+
+# Set the SIP_SUPPORT tag to YES if your project consists of sip (see:
+# http://www.riverbankcomputing.co.uk/software/sip/intro) sources only. Doxygen
+# will parse them like normal C++ but will assume all classes use public instead
+# of private inheritance when no explicit protection keyword is present.
+# The default value is: NO.
+
+SIP_SUPPORT            = NO
+
+# For Microsoft's IDL there are propget and propput attributes to indicate
+# getter and setter methods for a property. Setting this option to YES will make
+# doxygen to replace the get and set methods by a property in the documentation.
+# This will only work if the methods are indeed getting or setting a simple
+# type. If this is not the case, or you want to show the methods anyway, you
+# should set this option to NO.
+# The default value is: YES.
+
+IDL_PROPERTY_SUPPORT   = YES
+
+# If member grouping is used in the documentation and the DISTRIBUTE_GROUP_DOC
+# tag is set to YES then doxygen will reuse the documentation of the first
+# member in the group (if any) for the other members of the group. By default
+# all members of a group must be documented explicitly.
+# The default value is: NO.
+
+DISTRIBUTE_GROUP_DOC   = YES
+
+# If one adds a struct or class to a group and this option is enabled, then also
+# any nested class or struct is added to the same group. By default this option
+# is disabled and one has to add nested compounds explicitly via \ingroup.
+# The default value is: NO.
+
+GROUP_NESTED_COMPOUNDS = NO
+
+# Set the SUBGROUPING tag to YES to allow class member groups of the same type
+# (for instance a group of public functions) to be put as a subgroup of that
+# type (e.g. under the Public Functions section). Set it to NO to prevent
+# subgrouping. Alternatively, this can be done per class using the
+# \nosubgrouping command.
+# The default value is: YES.
+
+SUBGROUPING            = YES
+
+# When the INLINE_GROUPED_CLASSES tag is set to YES, classes, structs and unions
+# are shown inside the group in which they are included (e.g. using \ingroup)
+# instead of on a separate page (for HTML and Man pages) or section (for LaTeX
+# and RTF).
+#
+# Note that this feature does not work in combination with
+# SEPARATE_MEMBER_PAGES.
+# The default value is: NO.
+
+INLINE_GROUPED_CLASSES = NO
+
+# When the INLINE_SIMPLE_STRUCTS tag is set to YES, structs, classes, and unions
+# with only public data fields or simple typedef fields will be shown inline in
+# the documentation of the scope in which they are defined (i.e. file,
+# namespace, or group documentation), provided this scope is documented. If set
+# to NO, structs, classes, and unions are shown on a separate page (for HTML and
+# Man pages) or section (for LaTeX and RTF).
+# The default value is: NO.
+
+INLINE_SIMPLE_STRUCTS  = NO
+
+# When TYPEDEF_HIDES_STRUCT tag is enabled, a typedef of a struct, union, or
+# enum is documented as struct, union, or enum with the name of the typedef. So
+# typedef struct TypeS {} TypeT, will appear in the documentation as a struct
+# with name TypeT. When disabled the typedef will appear as a member of a file,
+# namespace, or class. And the struct will be named TypeS. This can typically be
+# useful for C code in case the coding convention dictates that all compound
+# types are typedef'ed and only the typedef is referenced, never the tag name.
+# The default value is: NO.
+
+TYPEDEF_HIDES_STRUCT   = YES
+
+# The size of the symbol lookup cache can be set using LOOKUP_CACHE_SIZE. This
+# cache is used to resolve symbols given their name and scope. Since this can be
+# an expensive process and often the same symbol appears multiple times in the
+# code, doxygen keeps a cache of pre-resolved symbols. If the cache is too small
+# doxygen will become slower. If the cache is too large, memory is wasted. The
+# cache size is given by this formula: 2^(16+LOOKUP_CACHE_SIZE). The valid range
+# is 0..9, the default is 0, corresponding to a cache size of 2^16=65536
+# symbols. At the end of a run doxygen will report the cache usage and suggest
+# the optimal cache size from a speed point of view.
+# Minimum value: 0, maximum value: 9, default value: 0.
+
+LOOKUP_CACHE_SIZE      = 0
+
+#---------------------------------------------------------------------------
+# Build related configuration options
+#---------------------------------------------------------------------------
+
+# If the EXTRACT_ALL tag is set to YES, doxygen will assume all entities in
+# documentation are documented, even if no documentation was available. Private
+# class members and static file members will be hidden unless the
+# EXTRACT_PRIVATE respectively EXTRACT_STATIC tags are set to YES.
+# Note: This will also disable the warnings about undocumented members that are
+# normally produced when WARNINGS is set to YES.
+# The default value is: NO.
+
+EXTRACT_ALL            = YES
+
+# If the EXTRACT_PRIVATE tag is set to YES, all private members of a class will
+# be included in the documentation.
+# The default value is: NO.
+
+EXTRACT_PRIVATE        = NO
+
+# If the EXTRACT_PACKAGE tag is set to YES, all members with package or internal
+# scope will be included in the documentation.
+# The default value is: NO.
+
+EXTRACT_PACKAGE        = NO
+
+# If the EXTRACT_STATIC tag is set to YES, all static members of a file will be
+# included in the documentation.
+# The default value is: NO.
+
+EXTRACT_STATIC         = NO
+
+# If the EXTRACT_LOCAL_CLASSES tag is set to YES, classes (and structs) defined
+# locally in source files will be included in the documentation. If set to NO,
+# only classes defined in header files are included. Does not have any effect
+# for Java sources.
+# The default value is: YES.
+
+EXTRACT_LOCAL_CLASSES  = YES
+
+# This flag is only useful for Objective-C code. If set to YES, local methods,
+# which are defined in the implementation section but not in the interface are
+# included in the documentation. If set to NO, only methods in the interface are
+# included.
+# The default value is: NO.
+
+EXTRACT_LOCAL_METHODS  = NO
+
+# If this flag is set to YES, the members of anonymous namespaces will be
+# extracted and appear in the documentation as a namespace called
+# 'anonymous_namespace{file}', where file will be replaced with the base name of
+# the file that contains the anonymous namespace. By default anonymous namespace
+# are hidden.
+# The default value is: NO.
+
+EXTRACT_ANON_NSPACES   = NO
+
+# If the HIDE_UNDOC_MEMBERS tag is set to YES, doxygen will hide all
+# undocumented members inside documented classes or files. If set to NO these
+# members will be included in the various overviews, but no documentation
+# section is generated. This option has no effect if EXTRACT_ALL is enabled.
+# The default value is: NO.
+
+HIDE_UNDOC_MEMBERS     = NO
+
+# If the HIDE_UNDOC_CLASSES tag is set to YES, doxygen will hide all
+# undocumented classes that are normally visible in the class hierarchy. If set
+# to NO, these classes will be included in the various overviews. This option
+# has no effect if EXTRACT_ALL is enabled.
+# The default value is: NO.
+
+HIDE_UNDOC_CLASSES     = NO
+
+# If the HIDE_FRIEND_COMPOUNDS tag is set to YES, doxygen will hide all friend
+# (class|struct|union) declarations. If set to NO, these declarations will be
+# included in the documentation.
+# The default value is: NO.
+
+HIDE_FRIEND_COMPOUNDS  = NO
+
+# If the HIDE_IN_BODY_DOCS tag is set to YES, doxygen will hide any
+# documentation blocks found inside the body of a function. If set to NO, these
+# blocks will be appended to the function's detailed documentation block.
+# The default value is: NO.
+
+HIDE_IN_BODY_DOCS      = NO
+
+# The INTERNAL_DOCS tag determines if documentation that is typed after a
+# \internal command is included. If the tag is set to NO then the documentation
+# will be excluded. Set it to YES to include the internal documentation.
+# The default value is: NO.
+
+INTERNAL_DOCS          = NO
+
+# If the CASE_SENSE_NAMES tag is set to NO then doxygen will only generate file
+# names in lower-case letters. If set to YES, upper-case letters are also
+# allowed. This is useful if you have classes or files whose names only differ
+# in case and if your file system supports case sensitive file names. Windows
+# and Mac users are advised to set this option to NO.
+# The default value is: system dependent.
+
+CASE_SENSE_NAMES       = NO
+
+# If the HIDE_SCOPE_NAMES tag is set to NO then doxygen will show members with
+# their full class and namespace scopes in the documentation. If set to YES, the
+# scope will be hidden.
+# The default value is: NO.
+
+HIDE_SCOPE_NAMES       = NO
+
+# If the HIDE_COMPOUND_REFERENCE tag is set to NO (default) then doxygen will
+# append additional text to a page's title, such as Class Reference. If set to
+# YES the compound reference will be hidden.
+# The default value is: NO.
+
+HIDE_COMPOUND_REFERENCE= NO
+
+# If the SHOW_INCLUDE_FILES tag is set to YES then doxygen will put a list of
+# the files that are included by a file in the documentation of that file.
+# The default value is: YES.
+
+SHOW_INCLUDE_FILES     = YES
+
+# If the SHOW_GROUPED_MEMB_INC tag is set to YES then Doxygen will add for each
+# grouped member an include statement to the documentation, telling the reader
+# which file to include in order to use the member.
+# The default value is: NO.
+
+SHOW_GROUPED_MEMB_INC  = NO
+
+# If the FORCE_LOCAL_INCLUDES tag is set to YES then doxygen will list include
+# files with double quotes in the documentation rather than with sharp brackets.
+# The default value is: NO.
+
+FORCE_LOCAL_INCLUDES   = NO
+
+# If the INLINE_INFO tag is set to YES then a tag [inline] is inserted in the
+# documentation for inline members.
+# The default value is: YES.
+
+INLINE_INFO            = YES
+
+# If the SORT_MEMBER_DOCS tag is set to YES then doxygen will sort the
+# (detailed) documentation of file and class members alphabetically by member
+# name. If set to NO, the members will appear in declaration order.
+# The default value is: YES.
+
+SORT_MEMBER_DOCS       = YES
+
+# If the SORT_BRIEF_DOCS tag is set to YES then doxygen will sort the brief
+# descriptions of file, namespace and class members alphabetically by member
+# name. If set to NO, the members will appear in declaration order. Note that
+# this will also influence the order of the classes in the class list.
+# The default value is: NO.
+
+SORT_BRIEF_DOCS        = NO
+
+# If the SORT_MEMBERS_CTORS_1ST tag is set to YES then doxygen will sort the
+# (brief and detailed) documentation of class members so that constructors and
+# destructors are listed first. If set to NO the constructors will appear in the
+# respective orders defined by SORT_BRIEF_DOCS and SORT_MEMBER_DOCS.
+# Note: If SORT_BRIEF_DOCS is set to NO this option is ignored for sorting brief
+# member documentation.
+# Note: If SORT_MEMBER_DOCS is set to NO this option is ignored for sorting
+# detailed member documentation.
+# The default value is: NO.
+
+SORT_MEMBERS_CTORS_1ST = NO
+
+# If the SORT_GROUP_NAMES tag is set to YES then doxygen will sort the hierarchy
+# of group names into alphabetical order. If set to NO the group names will
+# appear in their defined order.
+# The default value is: NO.
+
+SORT_GROUP_NAMES       = NO
+
+# If the SORT_BY_SCOPE_NAME tag is set to YES, the class list will be sorted by
+# fully-qualified names, including namespaces. If set to NO, the class list will
+# be sorted only by class name, not including the namespace part.
+# Note: This option is not very useful if HIDE_SCOPE_NAMES is set to YES.
+# Note: This option applies only to the class list, not to the alphabetical
+# list.
+# The default value is: NO.
+
+SORT_BY_SCOPE_NAME     = NO
+
+# If the STRICT_PROTO_MATCHING option is enabled and doxygen fails to do proper
+# type resolution of all parameters of a function it will reject a match between
+# the prototype and the implementation of a member function even if there is
+# only one candidate or it is obvious which candidate to choose by doing a
+# simple string match. By disabling STRICT_PROTO_MATCHING doxygen will still
+# accept a match between prototype and implementation in such cases.
+# The default value is: NO.
+
+STRICT_PROTO_MATCHING  = NO
+
+# The GENERATE_TODOLIST tag can be used to enable (YES) or disable (NO) the todo
+# list. This list is created by putting \todo commands in the documentation.
+# The default value is: YES.
+
+GENERATE_TODOLIST      = YES
+
+# The GENERATE_TESTLIST tag can be used to enable (YES) or disable (NO) the test
+# list. This list is created by putting \test commands in the documentation.
+# The default value is: YES.
+
+GENERATE_TESTLIST      = YES
+
+# The GENERATE_BUGLIST tag can be used to enable (YES) or disable (NO) the bug
+# list. This list is created by putting \bug commands in the documentation.
+# The default value is: YES.
+
+GENERATE_BUGLIST       = YES
+
+# The GENERATE_DEPRECATEDLIST tag can be used to enable (YES) or disable (NO)
+# the deprecated list. This list is created by putting \deprecated commands in
+# the documentation.
+# The default value is: YES.
+
+GENERATE_DEPRECATEDLIST= YES
+
+# The ENABLED_SECTIONS tag can be used to enable conditional documentation
+# sections, marked by \if <section_label> ... \endif and \cond <section_label>
+# ... \endcond blocks.
+
+ENABLED_SECTIONS       =
+
+# The MAX_INITIALIZER_LINES tag determines the maximum number of lines that the
+# initial value of a variable or macro / define can have for it to appear in the
+# documentation. If the initializer consists of more lines than specified here
+# it will be hidden. Use a value of 0 to hide initializers completely. The
+# appearance of the value of individual variables and macros / defines can be
+# controlled using \showinitializer or \hideinitializer command in the
+# documentation regardless of this setting.
+# Minimum value: 0, maximum value: 10000, default value: 30.
+
+MAX_INITIALIZER_LINES  = 30
+
+# Set the SHOW_USED_FILES tag to NO to disable the list of files generated at
+# the bottom of the documentation of classes and structs. If set to YES, the
+# list will mention the files that were used to generate the documentation.
+# The default value is: YES.
+
+SHOW_USED_FILES        = YES
+
+# Set the SHOW_FILES tag to NO to disable the generation of the Files page. This
+# will remove the Files entry from the Quick Index and from the Folder Tree View
+# (if specified).
+# The default value is: YES.
+
+SHOW_FILES             = YES
+
+# Set the SHOW_NAMESPACES tag to NO to disable the generation of the Namespaces
+# page. This will remove the Namespaces entry from the Quick Index and from the
+# Folder Tree View (if specified).
+# The default value is: YES.
+
+SHOW_NAMESPACES        = YES
+
+# The FILE_VERSION_FILTER tag can be used to specify a program or script that
+# doxygen should invoke to get the current version for each file (typically from
+# the version control system). Doxygen will invoke the program by executing (via
+# popen()) the command command input-file, where command is the value of the
+# FILE_VERSION_FILTER tag, and input-file is the name of an input file provided
+# by doxygen. Whatever the program writes to standard output is used as the file
+# version. For an example see the documentation.
+
+FILE_VERSION_FILTER    =
+
+# The LAYOUT_FILE tag can be used to specify a layout file which will be parsed
+# by doxygen. The layout file controls the global structure of the generated
+# output files in an output format independent way. To create the layout file
+# that represents doxygen's defaults, run doxygen with the -l option. You can
+# optionally specify a file name after the option, if omitted DoxygenLayout.xml
+# will be used as the name of the layout file.
+#
+# Note that if you run doxygen from a directory containing a file called
+# DoxygenLayout.xml, doxygen will parse it automatically even if the LAYOUT_FILE
+# tag is left empty.
+
+LAYOUT_FILE            =
+
+# The CITE_BIB_FILES tag can be used to specify one or more bib files containing
+# the reference definitions. This must be a list of .bib files. The .bib
+# extension is automatically appended if omitted. This requires the bibtex tool
+# to be installed. See also http://en.wikipedia.org/wiki/BibTeX for more info.
+# For LaTeX the style of the bibliography can be controlled using
+# LATEX_BIB_STYLE. To use this feature you need bibtex and perl available in the
+# search path. See also \cite for info how to create references.
+
+CITE_BIB_FILES         =
+
+#---------------------------------------------------------------------------
+# Configuration options related to warning and progress messages
+#---------------------------------------------------------------------------
+
+# The QUIET tag can be used to turn on/off the messages that are generated to
+# standard output by doxygen. If QUIET is set to YES this implies that the
+# messages are off.
+# The default value is: NO.
+
+QUIET                  = NO
+
+# The WARNINGS tag can be used to turn on/off the warning messages that are
+# generated to standard error (stderr) by doxygen. If WARNINGS is set to YES
+# this implies that the warnings are on.
+#
+# Tip: Turn warnings on while writing the documentation.
+# The default value is: YES.
+
+WARNINGS               = YES
+
+# If the WARN_IF_UNDOCUMENTED tag is set to YES then doxygen will generate
+# warnings for undocumented members. If EXTRACT_ALL is set to YES then this flag
+# will automatically be disabled.
+# The default value is: YES.
+
+WARN_IF_UNDOCUMENTED   = YES
+
+# If the WARN_IF_DOC_ERROR tag is set to YES, doxygen will generate warnings for
+# potential errors in the documentation, such as not documenting some parameters
+# in a documented function, or documenting parameters that don't exist or using
+# markup commands wrongly.
+# The default value is: YES.
+
+WARN_IF_DOC_ERROR      = YES
+
+# This WARN_NO_PARAMDOC option can be enabled to get warnings for functions that
+# are documented, but have no documentation for their parameters or return
+# value. If set to NO, doxygen will only warn about wrong or incomplete
+# parameter documentation, but not about the absence of documentation.
+# The default value is: NO.
+
+WARN_NO_PARAMDOC       = NO
+
+# The WARN_FORMAT tag determines the format of the warning messages that doxygen
+# can produce. The string should contain the $file, $line, and $text tags, which
+# will be replaced by the file and line number from which the warning originated
+# and the warning text. Optionally the format may contain $version, which will
+# be replaced by the version of the file (if it could be obtained via
+# FILE_VERSION_FILTER)
+# The default value is: $file:$line: $text.
+
+WARN_FORMAT            = "$file:$line: $text"
+
+# The WARN_LOGFILE tag can be used to specify a file to which warning and error
+# messages should be written. If left blank the output is written to standard
+# error (stderr).
+
+WARN_LOGFILE           =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the input files
+#---------------------------------------------------------------------------
+
+# The INPUT tag is used to specify the files and/or directories that contain
+# documented source files. You may enter file names like myfile.cpp or
+# directories like /usr/src/myproject. Separate the files or directories with
+# spaces. See also FILE_PATTERNS and EXTENSION_MAPPING
+# Note: If this tag is empty the current directory is searched.
+
+INPUT                  = ../../include/ck/tensor_operation/gpu/grid \
+                         ../../include/ck/tensor_operation/gpu/block \
+                         ../../include/ck/tensor_operation/gpu/thread \
+                         ../../library/include/ck/library/utility
+
+# This tag can be used to specify the character encoding of the source files
+# that doxygen parses. Internally doxygen uses the UTF-8 encoding. Doxygen uses
+# libiconv (or the iconv built into libc) for the transcoding. See the libiconv
+# documentation (see: http://www.gnu.org/software/libiconv) for the list of
+# possible encodings.
+# The default value is: UTF-8.
+
+INPUT_ENCODING         = UTF-8
+
+# If the value of the INPUT tag contains directories, you can use the
+# FILE_PATTERNS tag to specify one or more wildcard patterns (like *.cpp and
+# *.h) to filter out the source-files in the directories.
+#
+# Note that for custom extensions or not directly supported extensions you also
+# need to set EXTENSION_MAPPING for the extension otherwise the files are not
+# read by doxygen.
+#
+# If left blank the following patterns are tested:*.c, *.cc, *.cxx, *.cpp,
+# *.c++, *.java, *.ii, *.ixx, *.ipp, *.i++, *.inl, *.idl, *.ddl, *.odl, *.h,
+# *.hh, *.hxx, *.hpp, *.h++, *.cs, *.d, *.php, *.php4, *.php5, *.phtml, *.inc,
+# *.m, *.markdown, *.md, *.mm, *.dox, *.py, *.f90, *.f, *.for, *.tcl, *.vhd,
+# *.vhdl, *.ucf, *.qsf, *.as and *.js.
+
+FILE_PATTERNS          = *.c \
+                         *.cc \
+                         *.cxx \
+                         *.cpp \
+                         *.c++ \
+                         *.java \
+                         *.ii \
+                         *.ixx \
+                         *.ipp \
+                         *.i++ \
+                         *.inl \
+                         *.idl \
+                         *.ddl \
+                         *.odl \
+                         *.h \
+                         *.hh \
+                         *.hxx \
+                         *.hpp \
+                         *.h++ \
+                         *.cs \
+                         *.d \
+                         *.php \
+                         *.php4 \
+                         *.php5 \
+                         *.phtml \
+                         *.inc \
+                         *.m \
+                         *.markdown \
+                         *.md \
+                         *.mm \
+                         *.dox \
+                         *.py \
+                         *.tcl \
+                         *.vhd \
+                         *.vhdl \
+                         *.ucf \
+                         *.qsf \
+                         *.as \
+                         *.js
+
+# The RECURSIVE tag can be used to specify whether or not subdirectories should
+# be searched for input files as well.
+# The default value is: NO.
+
+RECURSIVE              = YES
+
+# The EXCLUDE tag can be used to specify files and/or directories that should be
+# excluded from the INPUT source files. This way you can easily exclude a
+# subdirectory from a directory tree whose root is specified with the INPUT tag.
+#
+# Note that relative paths are relative to the directory from which doxygen is
+# run.
+
+EXCLUDE                =
+
+# The EXCLUDE_SYMLINKS tag can be used to select whether or not files or
+# directories that are symbolic links (a Unix file system feature) are excluded
+# from the input.
+# The default value is: NO.
+
+EXCLUDE_SYMLINKS       = NO
+
+# If the value of the INPUT tag contains directories, you can use the
+# EXCLUDE_PATTERNS tag to specify one or more wildcard patterns to exclude
+# certain files from those directories.
+#
+# Note that the wildcards are matched against the file with absolute path, so to
+# exclude all test directories for example use the pattern */test/*
+
+EXCLUDE_PATTERNS       =
+
+# The EXCLUDE_SYMBOLS tag can be used to specify one or more symbol names
+# (namespaces, classes, functions, etc.) that should be excluded from the
+# output. The symbol name can be a fully qualified name, a word, or if the
+# wildcard * is used, a substring. Examples: ANamespace, AClass,
+# AClass::ANamespace, ANamespace::*Test
+#
+# Note that the wildcards are matched against the file with absolute path, so to
+# exclude all test directories use the pattern */test/*
+
+EXCLUDE_SYMBOLS        =
+
+# The EXAMPLE_PATH tag can be used to specify one or more files or directories
+# that contain example code fragments that are included (see the \include
+# command).
+
+EXAMPLE_PATH           =
+
+# If the value of the EXAMPLE_PATH tag contains directories, you can use the
+# EXAMPLE_PATTERNS tag to specify one or more wildcard pattern (like *.cpp and
+# *.h) to filter out the source-files in the directories. If left blank all
+# files are included.
+
+EXAMPLE_PATTERNS       = *
+
+# If the EXAMPLE_RECURSIVE tag is set to YES then subdirectories will be
+# searched for input files to be used with the \include or \dontinclude commands
+# irrespective of the value of the RECURSIVE tag.
+# The default value is: NO.
+
+EXAMPLE_RECURSIVE      = NO
+
+# The IMAGE_PATH tag can be used to specify one or more files or directories
+# that contain images that are to be included in the documentation (see the
+# \image command).
+
+IMAGE_PATH             =
+
+# The INPUT_FILTER tag can be used to specify a program that doxygen should
+# invoke to filter for each input file. Doxygen will invoke the filter program
+# by executing (via popen()) the command:
+#
+# <filter> <input-file>
+#
+# where <filter> is the value of the INPUT_FILTER tag, and <input-file> is the
+# name of an input file. Doxygen will then use the output that the filter
+# program writes to standard output. If FILTER_PATTERNS is specified, this tag
+# will be ignored.
+#
+# Note that the filter must not add or remove lines; it is applied before the
+# code is scanned, but not when the output code is generated. If lines are added
+# or removed, the anchors will not be placed correctly.
+
+INPUT_FILTER           =
+
+# The FILTER_PATTERNS tag can be used to specify filters on a per file pattern
+# basis. Doxygen will compare the file name with each pattern and apply the
+# filter if there is a match. The filters are a list of the form: pattern=filter
+# (like *.cpp=my_cpp_filter). See INPUT_FILTER for further information on how
+# filters are used. If the FILTER_PATTERNS tag is empty or if none of the
+# patterns match the file name, INPUT_FILTER is applied.
+
+FILTER_PATTERNS        =
+
+# If the FILTER_SOURCE_FILES tag is set to YES, the input filter (if set using
+# INPUT_FILTER) will also be used to filter the input files that are used for
+# producing the source files to browse (i.e. when SOURCE_BROWSER is set to YES).
+# The default value is: NO.
+
+FILTER_SOURCE_FILES    = NO
+
+# The FILTER_SOURCE_PATTERNS tag can be used to specify source filters per file
+# pattern. A pattern will override the setting for FILTER_PATTERN (if any) and
+# it is also possible to disable source filtering for a specific pattern using
+# *.ext= (so without naming a filter).
+# This tag requires that the tag FILTER_SOURCE_FILES is set to YES.
+
+FILTER_SOURCE_PATTERNS =
+
+# If the USE_MDFILE_AS_MAINPAGE tag refers to the name of a markdown file that
+# is part of the input, its contents will be placed on the main page
+# (index.html). This can be useful if you have a project on for instance GitHub
+# and want to reuse the introduction page also for the doxygen output.
+
+USE_MDFILE_AS_MAINPAGE = ../README.md
+
+#---------------------------------------------------------------------------
+# Configuration options related to source browsing
+#---------------------------------------------------------------------------
+
+# If the SOURCE_BROWSER tag is set to YES then a list of source files will be
+# generated. Documented entities will be cross-referenced with these sources.
+#
+# Note: To get rid of all source code in the generated output, make sure that
+# also VERBATIM_HEADERS is set to NO.
+# The default value is: NO.
+
+SOURCE_BROWSER         = NO
+
+# Setting the INLINE_SOURCES tag to YES will include the body of functions,
+# classes and enums directly into the documentation.
+# The default value is: NO.
+
+INLINE_SOURCES         = NO
+
+# Setting the STRIP_CODE_COMMENTS tag to YES will instruct doxygen to hide any
+# special comment blocks from generated source code fragments. Normal C, C++ and
+# Fortran comments will always remain visible.
+# The default value is: YES.
+
+STRIP_CODE_COMMENTS    = YES
+
+# If the REFERENCED_BY_RELATION tag is set to YES then for each documented
+# function all documented functions referencing it will be listed.
+# The default value is: NO.
+
+REFERENCED_BY_RELATION = NO
+
+# If the REFERENCES_RELATION tag is set to YES then for each documented function
+# all documented entities called/used by that function will be listed.
+# The default value is: NO.
+
+REFERENCES_RELATION    = NO
+
+# If the REFERENCES_LINK_SOURCE tag is set to YES and SOURCE_BROWSER tag is set
+# to YES then the hyperlinks from functions in REFERENCES_RELATION and
+# REFERENCED_BY_RELATION lists will link to the source code. Otherwise they will
+# link to the documentation.
+# The default value is: YES.
+
+REFERENCES_LINK_SOURCE = YES
+
+# If SOURCE_TOOLTIPS is enabled (the default) then hovering a hyperlink in the
+# source code will show a tooltip with additional information such as prototype,
+# brief description and links to the definition and documentation. Since this
+# will make the HTML file larger and loading of large files a bit slower, you
+# can opt to disable this feature.
+# The default value is: YES.
+# This tag requires that the tag SOURCE_BROWSER is set to YES.
+
+SOURCE_TOOLTIPS        = YES
+
+# If the USE_HTAGS tag is set to YES then the references to source code will
+# point to the HTML generated by the htags(1) tool instead of doxygen built-in
+# source browser. The htags tool is part of GNU's global source tagging system
+# (see http://www.gnu.org/software/global/global.html). You will need version
+# 4.8.6 or higher.
+#
+# To use it do the following:
+# - Install the latest version of global
+# - Enable SOURCE_BROWSER and USE_HTAGS in the config file
+# - Make sure the INPUT points to the root of the source tree
+# - Run doxygen as normal
+#
+# Doxygen will invoke htags (and that will in turn invoke gtags), so these
+# tools must be available from the command line (i.e. in the search path).
+#
+# The result: instead of the source browser generated by doxygen, the links to
+# source code will now point to the output of htags.
+# The default value is: NO.
+# This tag requires that the tag SOURCE_BROWSER is set to YES.
+
+USE_HTAGS              = NO
+
+# If the VERBATIM_HEADERS tag is set the YES then doxygen will generate a
+# verbatim copy of the header file for each class for which an include is
+# specified. Set to NO to disable this.
+# See also: Section \class.
+# The default value is: YES.
+
+VERBATIM_HEADERS       = YES
+
+# If the CLANG_ASSISTED_PARSING tag is set to YES then doxygen will use the
+# clang parser (see: http://clang.llvm.org/) for more accurate parsing at the
+# cost of reduced performance. This can be particularly helpful with template
+# rich C++ code for which doxygen's built-in parser lacks the necessary type
+# information.
+# Note: The availability of this option depends on whether or not doxygen was
+# compiled with the --with-libclang option.
+# The default value is: NO.
+
+CLANG_ASSISTED_PARSING = NO
+
+# If clang assisted parsing is enabled you can provide the compiler with command
+# line options that you would normally use when invoking the compiler. Note that
+# the include paths will already be set by doxygen for the files and directories
+# specified with INPUT and INCLUDE_PATH.
+# This tag requires that the tag CLANG_ASSISTED_PARSING is set to YES.
+
+CLANG_OPTIONS          =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the alphabetical class index
+#---------------------------------------------------------------------------
+
+# If the ALPHABETICAL_INDEX tag is set to YES, an alphabetical index of all
+# compounds will be generated. Enable this if the project contains a lot of
+# classes, structs, unions or interfaces.
+# The default value is: YES.
+
+ALPHABETICAL_INDEX     = YES
+
+# The COLS_IN_ALPHA_INDEX tag can be used to specify the number of columns in
+# which the alphabetical index list will be split.
+# Minimum value: 1, maximum value: 20, default value: 5.
+# This tag requires that the tag ALPHABETICAL_INDEX is set to YES.
+
+COLS_IN_ALPHA_INDEX    = 5
+
+# In case all classes in a project start with a common prefix, all classes will
+# be put under the same header in the alphabetical index. The IGNORE_PREFIX tag
+# can be used to specify a prefix (or a list of prefixes) that should be ignored
+# while generating the index headers.
+# This tag requires that the tag ALPHABETICAL_INDEX is set to YES.
+
+IGNORE_PREFIX          =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the HTML output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_HTML tag is set to YES, doxygen will generate HTML output
+# The default value is: YES.
+
+GENERATE_HTML          = YES
+
+# The HTML_OUTPUT tag is used to specify where the HTML docs will be put. If a
+# relative path is entered the value of OUTPUT_DIRECTORY will be put in front of
+# it.
+# The default directory is: html.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_OUTPUT            = html
+
+# The HTML_FILE_EXTENSION tag can be used to specify the file extension for each
+# generated HTML page (for example: .htm, .php, .asp).
+# The default value is: .html.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_FILE_EXTENSION    = .html
+
+# The HTML_HEADER tag can be used to specify a user-defined HTML header file for
+# each generated HTML page. If the tag is left blank doxygen will generate a
+# standard header.
+#
+# To get valid HTML the header file that includes any scripts and style sheets
+# that doxygen needs, which is dependent on the configuration options used (e.g.
+# the setting GENERATE_TREEVIEW). It is highly recommended to start with a
+# default header using
+# doxygen -w html new_header.html new_footer.html new_stylesheet.css
+# YourConfigFile
+# and then modify the file new_header.html. See also section "Doxygen usage"
+# for information on how to generate the default header that doxygen normally
+# uses.
+# Note: The header is subject to change so you typically have to regenerate the
+# default header when upgrading to a newer version of doxygen. For a description
+# of the possible markers and block names see the documentation.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_HEADER            =
+
+# The HTML_FOOTER tag can be used to specify a user-defined HTML footer for each
+# generated HTML page. If the tag is left blank doxygen will generate a standard
+# footer. See HTML_HEADER for more information on how to generate a default
+# footer and what special commands can be used inside the footer. See also
+# section "Doxygen usage" for information on how to generate the default footer
+# that doxygen normally uses.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_FOOTER            =
+
+# The HTML_STYLESHEET tag can be used to specify a user-defined cascading style
+# sheet that is used by each HTML page. It can be used to fine-tune the look of
+# the HTML output. If left blank doxygen will generate a default style sheet.
+# See also section "Doxygen usage" for information on how to generate the style
+# sheet that doxygen normally uses.
+# Note: It is recommended to use HTML_EXTRA_STYLESHEET instead of this tag, as
+# it is more robust and this tag (HTML_STYLESHEET) will in the future become
+# obsolete.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_STYLESHEET        =
+
+# The HTML_EXTRA_STYLESHEET tag can be used to specify additional user-defined
+# cascading style sheets that are included after the standard style sheets
+# created by doxygen. Using this option one can overrule certain style aspects.
+# This is preferred over using HTML_STYLESHEET since it does not replace the
+# standard style sheet and is therefore more robust against future updates.
+# Doxygen will copy the style sheet files to the output directory.
+# Note: The order of the extra style sheet files is of importance (e.g. the last
+# style sheet in the list overrules the setting of the previous ones in the
+# list). For an example see the documentation.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_EXTRA_STYLESHEET  =
+
+# The HTML_EXTRA_FILES tag can be used to specify one or more extra images or
+# other source files which should be copied to the HTML output directory. Note
+# that these files will be copied to the base HTML output directory. Use the
+# $relpath^ marker in the HTML_HEADER and/or HTML_FOOTER files to load these
+# files. In the HTML_STYLESHEET file, use the file name only. Also note that the
+# files will be copied as-is; there are no commands or markers available.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_EXTRA_FILES       =
+
+# The HTML_COLORSTYLE_HUE tag controls the color of the HTML output. Doxygen
+# will adjust the colors in the style sheet and background images according to
+# this color. Hue is specified as an angle on a colorwheel, see
+# http://en.wikipedia.org/wiki/Hue for more information. For instance the value
+# 0 represents red, 60 is yellow, 120 is green, 180 is cyan, 240 is blue, 300
+# purple, and 360 is red again.
+# Minimum value: 0, maximum value: 359, default value: 220.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_COLORSTYLE_HUE    = 220
+
+# The HTML_COLORSTYLE_SAT tag controls the purity (or saturation) of the colors
+# in the HTML output. For a value of 0 the output will use grayscales only. A
+# value of 255 will produce the most vivid colors.
+# Minimum value: 0, maximum value: 255, default value: 100.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_COLORSTYLE_SAT    = 100
+
+# The HTML_COLORSTYLE_GAMMA tag controls the gamma correction applied to the
+# luminance component of the colors in the HTML output. Values below 100
+# gradually make the output lighter, whereas values above 100 make the output
+# darker. The value divided by 100 is the actual gamma applied, so 80 represents
+# a gamma of 0.8, The value 220 represents a gamma of 2.2, and 100 does not
+# change the gamma.
+# Minimum value: 40, maximum value: 240, default value: 80.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_COLORSTYLE_GAMMA  = 80
+
+# If the HTML_TIMESTAMP tag is set to YES then the footer of each generated HTML
+# page will contain the date and time when the page was generated. Setting this
+# to YES can help to show when doxygen was last run and thus if the
+# documentation is up to date.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_TIMESTAMP         = NO
+
+# If the HTML_DYNAMIC_SECTIONS tag is set to YES then the generated HTML
+# documentation will contain sections that can be hidden and shown after the
+# page has loaded.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_DYNAMIC_SECTIONS  = NO
+
+# With HTML_INDEX_NUM_ENTRIES one can control the preferred number of entries
+# shown in the various tree structured indices initially; the user can expand
+# and collapse entries dynamically later on. Doxygen will expand the tree to
+# such a level that at most the specified number of entries are visible (unless
+# a fully collapsed tree already exceeds this amount). So setting the number of
+# entries 1 will produce a full collapsed tree by default. 0 is a special value
+# representing an infinite number of entries and will result in a full expanded
+# tree by default.
+# Minimum value: 0, maximum value: 9999, default value: 100.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_INDEX_NUM_ENTRIES = 100
+
+# If the GENERATE_DOCSET tag is set to YES, additional index files will be
+# generated that can be used as input for Apple's Xcode 3 integrated development
+# environment (see: http://developer.apple.com/tools/xcode/), introduced with
+# OSX 10.5 (Leopard). To create a documentation set, doxygen will generate a
+# Makefile in the HTML output directory. Running make will produce the docset in
+# that directory and running make install will install the docset in
+# ~/Library/Developer/Shared/Documentation/DocSets so that Xcode will find it at
+# startup. See http://developer.apple.com/tools/creatingdocsetswithdoxygen.html
+# for more information.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+GENERATE_DOCSET        = NO
+
+# This tag determines the name of the docset feed. A documentation feed provides
+# an umbrella under which multiple documentation sets from a single provider
+# (such as a company or product suite) can be grouped.
+# The default value is: Doxygen generated docs.
+# This tag requires that the tag GENERATE_DOCSET is set to YES.
+
+DOCSET_FEEDNAME        = "Doxygen generated docs"
+
+# This tag specifies a string that should uniquely identify the documentation
+# set bundle. This should be a reverse domain-name style string, e.g.
+# com.mycompany.MyDocSet. Doxygen will append .docset to the name.
+# The default value is: org.doxygen.Project.
+# This tag requires that the tag GENERATE_DOCSET is set to YES.
+
+DOCSET_BUNDLE_ID       = org.doxygen.Project
+
+# The DOCSET_PUBLISHER_ID tag specifies a string that should uniquely identify
+# the documentation publisher. This should be a reverse domain-name style
+# string, e.g. com.mycompany.MyDocSet.documentation.
+# The default value is: org.doxygen.Publisher.
+# This tag requires that the tag GENERATE_DOCSET is set to YES.
+
+DOCSET_PUBLISHER_ID    = org.doxygen.Publisher
+
+# The DOCSET_PUBLISHER_NAME tag identifies the documentation publisher.
+# The default value is: Publisher.
+# This tag requires that the tag GENERATE_DOCSET is set to YES.
+
+DOCSET_PUBLISHER_NAME  = Publisher
+
+# If the GENERATE_HTMLHELP tag is set to YES then doxygen generates three
+# additional HTML index files: index.hhp, index.hhc, and index.hhk. The
+# index.hhp is a project file that can be read by Microsoft's HTML Help Workshop
+# (see: http://www.microsoft.com/en-us/download/details.aspx?id=21138) on
+# Windows.
+#
+# The HTML Help Workshop contains a compiler that can convert all HTML output
+# generated by doxygen into a single compiled HTML file (.chm). Compiled HTML
+# files are now used as the Windows 98 help format, and will replace the old
+# Windows help format (.hlp) on all Windows platforms in the future. Compressed
+# HTML files also contain an index, a table of contents, and you can search for
+# words in the documentation. The HTML workshop also contains a viewer for
+# compressed HTML files.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+GENERATE_HTMLHELP      = NO
+
+# The CHM_FILE tag can be used to specify the file name of the resulting .chm
+# file. You can add a path in front of the file if the result should not be
+# written to the html output directory.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+CHM_FILE               =
+
+# The HHC_LOCATION tag can be used to specify the location (absolute path
+# including file name) of the HTML help compiler (hhc.exe). If non-empty,
+# doxygen will try to run the HTML help compiler on the generated index.hhp.
+# The file has to be specified with full path.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+HHC_LOCATION           =
+
+# The GENERATE_CHI flag controls if a separate .chi index file is generated
+# (YES) or that it should be included in the master .chm file (NO).
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+GENERATE_CHI           = NO
+
+# The CHM_INDEX_ENCODING is used to encode HtmlHelp index (hhk), content (hhc)
+# and project file content.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+CHM_INDEX_ENCODING     =
+
+# The BINARY_TOC flag controls whether a binary table of contents is generated
+# (YES) or a normal table of contents (NO) in the .chm file. Furthermore it
+# enables the Previous and Next buttons.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+BINARY_TOC             = NO
+
+# The TOC_EXPAND flag can be set to YES to add extra items for group members to
+# the table of contents of the HTML help documentation and to the tree view.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+TOC_EXPAND             = NO
+
+# If the GENERATE_QHP tag is set to YES and both QHP_NAMESPACE and
+# QHP_VIRTUAL_FOLDER are set, an additional index file will be generated that
+# can be used as input for Qt's qhelpgenerator to generate a Qt Compressed Help
+# (.qch) of the generated HTML documentation.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+GENERATE_QHP           = NO
+
+# If the QHG_LOCATION tag is specified, the QCH_FILE tag can be used to specify
+# the file name of the resulting .qch file. The path specified is relative to
+# the HTML output folder.
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QCH_FILE               =
+
+# The QHP_NAMESPACE tag specifies the namespace to use when generating Qt Help
+# Project output. For more information please see Qt Help Project / Namespace
+# (see: http://qt-project.org/doc/qt-4.8/qthelpproject.html#namespace).
+# The default value is: org.doxygen.Project.
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHP_NAMESPACE          = org.doxygen.Project
+
+# The QHP_VIRTUAL_FOLDER tag specifies the namespace to use when generating Qt
+# Help Project output. For more information please see Qt Help Project / Virtual
+# Folders (see: http://qt-project.org/doc/qt-4.8/qthelpproject.html#virtual-
+# folders).
+# The default value is: doc.
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHP_VIRTUAL_FOLDER     = doc
+
+# If the QHP_CUST_FILTER_NAME tag is set, it specifies the name of a custom
+# filter to add. For more information please see Qt Help Project / Custom
+# Filters (see: http://qt-project.org/doc/qt-4.8/qthelpproject.html#custom-
+# filters).
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHP_CUST_FILTER_NAME   =
+
+# The QHP_CUST_FILTER_ATTRS tag specifies the list of the attributes of the
+# custom filter to add. For more information please see Qt Help Project / Custom
+# Filters (see: http://qt-project.org/doc/qt-4.8/qthelpproject.html#custom-
+# filters).
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHP_CUST_FILTER_ATTRS  =
+
+# The QHP_SECT_FILTER_ATTRS tag specifies the list of the attributes this
+# project's filter section matches. Qt Help Project / Filter Attributes (see:
+# http://qt-project.org/doc/qt-4.8/qthelpproject.html#filter-attributes).
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHP_SECT_FILTER_ATTRS  =
+
+# The QHG_LOCATION tag can be used to specify the location of Qt's
+# qhelpgenerator. If non-empty doxygen will try to run qhelpgenerator on the
+# generated .qhp file.
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHG_LOCATION           =
+
+# If the GENERATE_ECLIPSEHELP tag is set to YES, additional index files will be
+# generated, together with the HTML files, they form an Eclipse help plugin. To
+# install this plugin and make it available under the help contents menu in
+# Eclipse, the contents of the directory containing the HTML and XML files needs
+# to be copied into the plugins directory of eclipse. The name of the directory
+# within the plugins directory should be the same as the ECLIPSE_DOC_ID value.
+# After copying Eclipse needs to be restarted before the help appears.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+GENERATE_ECLIPSEHELP   = NO
+
+# A unique identifier for the Eclipse help plugin. When installing the plugin
+# the directory name containing the HTML and XML files should also have this
+# name. Each documentation set should have its own identifier.
+# The default value is: org.doxygen.Project.
+# This tag requires that the tag GENERATE_ECLIPSEHELP is set to YES.
+
+ECLIPSE_DOC_ID         = org.doxygen.Project
+
+# If you want full control over the layout of the generated HTML pages it might
+# be necessary to disable the index and replace it with your own. The
+# DISABLE_INDEX tag can be used to turn on/off the condensed index (tabs) at top
+# of each HTML page. A value of NO enables the index and the value YES disables
+# it. Since the tabs in the index contain the same information as the navigation
+# tree, you can set this option to YES if you also set GENERATE_TREEVIEW to YES.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+DISABLE_INDEX          = NO
+
+# The GENERATE_TREEVIEW tag is used to specify whether a tree-like index
+# structure should be generated to display hierarchical information. If the tag
+# value is set to YES, a side panel will be generated containing a tree-like
+# index structure (just like the one that is generated for HTML Help). For this
+# to work a browser that supports JavaScript, DHTML, CSS and frames is required
+# (i.e. any modern browser). Windows users are probably better off using the
+# HTML help feature. Via custom style sheets (see HTML_EXTRA_STYLESHEET) one can
+# further fine-tune the look of the index. As an example, the default style
+# sheet generated by doxygen has an example that shows how to put an image at
+# the root of the tree instead of the PROJECT_NAME. Since the tree basically has
+# the same information as the tab index, you could consider setting
+# DISABLE_INDEX to YES when enabling this option.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+GENERATE_TREEVIEW      = NO
+
+# The ENUM_VALUES_PER_LINE tag can be used to set the number of enum values that
+# doxygen will group on one line in the generated HTML documentation.
+#
+# Note that a value of 0 will completely suppress the enum values from appearing
+# in the overview section.
+# Minimum value: 0, maximum value: 20, default value: 4.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+ENUM_VALUES_PER_LINE   = 1
+
+# If the treeview is enabled (see GENERATE_TREEVIEW) then this tag can be used
+# to set the initial width (in pixels) of the frame in which the tree is shown.
+# Minimum value: 0, maximum value: 1500, default value: 250.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+TREEVIEW_WIDTH         = 250
+
+# If the EXT_LINKS_IN_WINDOW option is set to YES, doxygen will open links to
+# external symbols imported via tag files in a separate window.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+EXT_LINKS_IN_WINDOW    = NO
+
+# Use this tag to change the font size of LaTeX formulas included as images in
+# the HTML documentation. When you change the font size after a successful
+# doxygen run you need to manually remove any form_*.png images from the HTML
+# output directory to force them to be regenerated.
+# Minimum value: 8, maximum value: 50, default value: 10.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+FORMULA_FONTSIZE       = 10
+
+# Use the FORMULA_TRANPARENT tag to determine whether or not the images
+# generated for formulas are transparent PNGs. Transparent PNGs are not
+# supported properly for IE 6.0, but are supported on all modern browsers.
+#
+# Note that when changing this option you need to delete any form_*.png files in
+# the HTML output directory before the changes have effect.
+# The default value is: YES.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+FORMULA_TRANSPARENT    = YES
+
+# Enable the USE_MATHJAX option to render LaTeX formulas using MathJax (see
+# http://www.mathjax.org) which uses client side Javascript for the rendering
+# instead of using pre-rendered bitmaps. Use this if you do not have LaTeX
+# installed or if you want to formulas look prettier in the HTML output. When
+# enabled you may also need to install MathJax separately and configure the path
+# to it using the MATHJAX_RELPATH option.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+USE_MATHJAX            = YES
+
+# When MathJax is enabled you can set the default output format to be used for
+# the MathJax output. See the MathJax site (see:
+# http://docs.mathjax.org/en/latest/output.html) for more details.
+# Possible values are: HTML-CSS (which is slower, but has the best
+# compatibility), NativeMML (i.e. MathML) and SVG.
+# The default value is: HTML-CSS.
+# This tag requires that the tag USE_MATHJAX is set to YES.
+
+MATHJAX_FORMAT         = HTML-CSS
+
+# When MathJax is enabled you need to specify the location relative to the HTML
+# output directory using the MATHJAX_RELPATH option. The destination directory
+# should contain the MathJax.js script. For instance, if the mathjax directory
+# is located at the same level as the HTML output directory, then
+# MATHJAX_RELPATH should be ../mathjax. The default value points to the MathJax
+# Content Delivery Network so you can quickly see the result without installing
+# MathJax. However, it is strongly recommended to install a local copy of
+# MathJax from http://www.mathjax.org before deployment.
+# The default value is: http://cdn.mathjax.org/mathjax/latest.
+# This tag requires that the tag USE_MATHJAX is set to YES.
+
+MATHJAX_RELPATH        = http://cdn.mathjax.org/mathjax/latest
+
+# The MATHJAX_EXTENSIONS tag can be used to specify one or more MathJax
+# extension names that should be enabled during MathJax rendering. For example
+# MATHJAX_EXTENSIONS = TeX/AMSmath TeX/AMSsymbols
+# This tag requires that the tag USE_MATHJAX is set to YES.
+
+MATHJAX_EXTENSIONS     =
+
+# The MATHJAX_CODEFILE tag can be used to specify a file with javascript pieces
+# of code that will be used on startup of the MathJax code. See the MathJax site
+# (see: http://docs.mathjax.org/en/latest/output.html) for more details. For an
+# example see the documentation.
+# This tag requires that the tag USE_MATHJAX is set to YES.
+
+MATHJAX_CODEFILE       =
+
+# When the SEARCHENGINE tag is enabled doxygen will generate a search box for
+# the HTML output. The underlying search engine uses javascript and DHTML and
+# should work on any modern browser. Note that when using HTML help
+# (GENERATE_HTMLHELP), Qt help (GENERATE_QHP), or docsets (GENERATE_DOCSET)
+# there is already a search function so this one should typically be disabled.
+# For large projects the javascript based search engine can be slow, then
+# enabling SERVER_BASED_SEARCH may provide a better solution. It is possible to
+# search using the keyboard; to jump to the search box use <access key> + S
+# (what the <access key> is depends on the OS and browser, but it is typically
+# <CTRL>, <ALT>/<option>, or both). Inside the search box use the <cursor down
+# key> to jump into the search results window, the results can be navigated
+# using the <cursor keys>. Press <Enter> to select an item or <escape> to cancel
+# the search. The filter options can be selected when the cursor is inside the
+# search box by pressing <Shift>+<cursor down>. Also here use the <cursor keys>
+# to select a filter and <Enter> or <escape> to activate or cancel the filter
+# option.
+# The default value is: YES.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+SEARCHENGINE           = YES
+
+# When the SERVER_BASED_SEARCH tag is enabled the search engine will be
+# implemented using a web server instead of a web client using Javascript. There
+# are two flavors of web server based searching depending on the EXTERNAL_SEARCH
+# setting. When disabled, doxygen will generate a PHP script for searching and
+# an index file used by the script. When EXTERNAL_SEARCH is enabled the indexing
+# and searching needs to be provided by external tools. See the section
+# "External Indexing and Searching" for details.
+# The default value is: NO.
+# This tag requires that the tag SEARCHENGINE is set to YES.
+
+SERVER_BASED_SEARCH    = NO
+
+# When EXTERNAL_SEARCH tag is enabled doxygen will no longer generate the PHP
+# script for searching. Instead the search results are written to an XML file
+# which needs to be processed by an external indexer. Doxygen will invoke an
+# external search engine pointed to by the SEARCHENGINE_URL option to obtain the
+# search results.
+#
+# Doxygen ships with an example indexer (doxyindexer) and search engine
+# (doxysearch.cgi) which are based on the open source search engine library
+# Xapian (see: http://xapian.org/).
+#
+# See the section "External Indexing and Searching" for details.
+# The default value is: NO.
+# This tag requires that the tag SEARCHENGINE is set to YES.
+
+EXTERNAL_SEARCH        = NO
+
+# The SEARCHENGINE_URL should point to a search engine hosted by a web server
+# which will return the search results when EXTERNAL_SEARCH is enabled.
+#
+# Doxygen ships with an example indexer (doxyindexer) and search engine
+# (doxysearch.cgi) which are based on the open source search engine library
+# Xapian (see: http://xapian.org/). See the section "External Indexing and
+# Searching" for details.
+# This tag requires that the tag SEARCHENGINE is set to YES.
+
+SEARCHENGINE_URL       =
+
+# When SERVER_BASED_SEARCH and EXTERNAL_SEARCH are both enabled the unindexed
+# search data is written to a file for indexing by an external tool. With the
+# SEARCHDATA_FILE tag the name of this file can be specified.
+# The default file is: searchdata.xml.
+# This tag requires that the tag SEARCHENGINE is set to YES.
+
+SEARCHDATA_FILE        = searchdata.xml
+
+# When SERVER_BASED_SEARCH and EXTERNAL_SEARCH are both enabled the
+# EXTERNAL_SEARCH_ID tag can be used as an identifier for the project. This is
+# useful in combination with EXTRA_SEARCH_MAPPINGS to search through multiple
+# projects and redirect the results back to the right project.
+# This tag requires that the tag SEARCHENGINE is set to YES.
+
+EXTERNAL_SEARCH_ID     =
+
+# The EXTRA_SEARCH_MAPPINGS tag can be used to enable searching through doxygen
+# projects other than the one defined by this configuration file, but that are
+# all added to the same external search index. Each project needs to have a
+# unique id set via EXTERNAL_SEARCH_ID. The search mapping then maps the id of
+# to a relative location where the documentation can be found. The format is:
+# EXTRA_SEARCH_MAPPINGS = tagname1=loc1 tagname2=loc2 ...
+# This tag requires that the tag SEARCHENGINE is set to YES.
+
+EXTRA_SEARCH_MAPPINGS  =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the LaTeX output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_LATEX tag is set to YES, doxygen will generate LaTeX output.
+# The default value is: YES.
+
+GENERATE_LATEX         = NO
+
+# The LATEX_OUTPUT tag is used to specify where the LaTeX docs will be put. If a
+# relative path is entered the value of OUTPUT_DIRECTORY will be put in front of
+# it.
+# The default directory is: latex.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_OUTPUT           = latex
+
+# The LATEX_CMD_NAME tag can be used to specify the LaTeX command name to be
+# invoked.
+#
+# Note that when enabling USE_PDFLATEX this option is only used for generating
+# bitmaps for formulas in the HTML output, but not in the Makefile that is
+# written to the output directory.
+# The default file is: latex.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_CMD_NAME         = latex
+
+# The MAKEINDEX_CMD_NAME tag can be used to specify the command name to generate
+# index for LaTeX.
+# The default file is: makeindex.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+MAKEINDEX_CMD_NAME     = makeindex
+
+# If the COMPACT_LATEX tag is set to YES, doxygen generates more compact LaTeX
+# documents. This may be useful for small projects and may help to save some
+# trees in general.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+COMPACT_LATEX          = NO
+
+# The PAPER_TYPE tag can be used to set the paper type that is used by the
+# printer.
+# Possible values are: a4 (210 x 297 mm), letter (8.5 x 11 inches), legal (8.5 x
+# 14 inches) and executive (7.25 x 10.5 inches).
+# The default value is: a4.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+PAPER_TYPE             = a4
+
+# The EXTRA_PACKAGES tag can be used to specify one or more LaTeX package names
+# that should be included in the LaTeX output. The package can be specified just
+# by its name or with the correct syntax as to be used with the LaTeX
+# \usepackage command. To get the times font for instance you can specify :
+# EXTRA_PACKAGES=times or EXTRA_PACKAGES={times}
+# To use the option intlimits with the amsmath package you can specify:
+# EXTRA_PACKAGES=[intlimits]{amsmath}
+# If left blank no extra packages will be included.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+EXTRA_PACKAGES         =
+
+# The LATEX_HEADER tag can be used to specify a personal LaTeX header for the
+# generated LaTeX document. The header should contain everything until the first
+# chapter. If it is left blank doxygen will generate a standard header. See
+# section "Doxygen usage" for information on how to let doxygen write the
+# default header to a separate file.
+#
+# Note: Only use a user-defined header if you know what you are doing! The
+# following commands have a special meaning inside the header: $title,
+# $datetime, $date, $doxygenversion, $projectname, $projectnumber,
+# $projectbrief, $projectlogo. Doxygen will replace $title with the empty
+# string, for the replacement values of the other commands the user is referred
+# to HTML_HEADER.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_HEADER           =
+
+# The LATEX_FOOTER tag can be used to specify a personal LaTeX footer for the
+# generated LaTeX document. The footer should contain everything after the last
+# chapter. If it is left blank doxygen will generate a standard footer. See
+# LATEX_HEADER for more information on how to generate a default footer and what
+# special commands can be used inside the footer.
+#
+# Note: Only use a user-defined footer if you know what you are doing!
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_FOOTER           =
+
+# The LATEX_EXTRA_STYLESHEET tag can be used to specify additional user-defined
+# LaTeX style sheets that are included after the standard style sheets created
+# by doxygen. Using this option one can overrule certain style aspects. Doxygen
+# will copy the style sheet files to the output directory.
+# Note: The order of the extra style sheet files is of importance (e.g. the last
+# style sheet in the list overrules the setting of the previous ones in the
+# list).
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_EXTRA_STYLESHEET =
+
+# The LATEX_EXTRA_FILES tag can be used to specify one or more extra images or
+# other source files which should be copied to the LATEX_OUTPUT output
+# directory. Note that the files will be copied as-is; there are no commands or
+# markers available.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_EXTRA_FILES      =
+
+# If the PDF_HYPERLINKS tag is set to YES, the LaTeX that is generated is
+# prepared for conversion to PDF (using ps2pdf or pdflatex). The PDF file will
+# contain links (just like the HTML output) instead of page references. This
+# makes the output suitable for online browsing using a PDF viewer.
+# The default value is: YES.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+PDF_HYPERLINKS         = YES
+
+# If the USE_PDFLATEX tag is set to YES, doxygen will use pdflatex to generate
+# the PDF file directly from the LaTeX files. Set this option to YES, to get a
+# higher quality PDF documentation.
+# The default value is: YES.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+USE_PDFLATEX           = YES
+
+# If the LATEX_BATCHMODE tag is set to YES, doxygen will add the \batchmode
+# command to the generated LaTeX files. This will instruct LaTeX to keep running
+# if errors occur, instead of asking the user for help. This option is also used
+# when generating formulas in HTML.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_BATCHMODE        = NO
+
+# If the LATEX_HIDE_INDICES tag is set to YES then doxygen will not include the
+# index chapters (such as File Index, Compound Index, etc.) in the output.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_HIDE_INDICES     = NO
+
+# If the LATEX_SOURCE_CODE tag is set to YES then doxygen will include source
+# code with syntax highlighting in the LaTeX output.
+#
+# Note that which sources are shown also depends on other settings such as
+# SOURCE_BROWSER.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_SOURCE_CODE      = NO
+
+# The LATEX_BIB_STYLE tag can be used to specify the style to use for the
+# bibliography, e.g. plainnat, or ieeetr. See
+# http://en.wikipedia.org/wiki/BibTeX and \cite for more info.
+# The default value is: plain.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_BIB_STYLE        = plain
+
+#---------------------------------------------------------------------------
+# Configuration options related to the RTF output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_RTF tag is set to YES, doxygen will generate RTF output. The
+# RTF output is optimized for Word 97 and may not look too pretty with other RTF
+# readers/editors.
+# The default value is: NO.
+
+GENERATE_RTF           = NO
+
+# The RTF_OUTPUT tag is used to specify where the RTF docs will be put. If a
+# relative path is entered the value of OUTPUT_DIRECTORY will be put in front of
+# it.
+# The default directory is: rtf.
+# This tag requires that the tag GENERATE_RTF is set to YES.
+
+RTF_OUTPUT             = rtf
+
+# If the COMPACT_RTF tag is set to YES, doxygen generates more compact RTF
+# documents. This may be useful for small projects and may help to save some
+# trees in general.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_RTF is set to YES.
+
+COMPACT_RTF            = NO
+
+# If the RTF_HYPERLINKS tag is set to YES, the RTF that is generated will
+# contain hyperlink fields. The RTF file will contain links (just like the HTML
+# output) instead of page references. This makes the output suitable for online
+# browsing using Word or some other Word compatible readers that support those
+# fields.
+#
+# Note: WordPad (write) and others do not support links.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_RTF is set to YES.
+
+RTF_HYPERLINKS         = NO
+
+# Load stylesheet definitions from file. Syntax is similar to doxygen's config
+# file, i.e. a series of assignments. You only have to provide replacements,
+# missing definitions are set to their default value.
+#
+# See also section "Doxygen usage" for information on how to generate the
+# default style sheet that doxygen normally uses.
+# This tag requires that the tag GENERATE_RTF is set to YES.
+
+RTF_STYLESHEET_FILE    =
+
+# Set optional variables used in the generation of an RTF document. Syntax is
+# similar to doxygen's config file. A template extensions file can be generated
+# using doxygen -e rtf extensionFile.
+# This tag requires that the tag GENERATE_RTF is set to YES.
+
+RTF_EXTENSIONS_FILE    =
+
+# If the RTF_SOURCE_CODE tag is set to YES then doxygen will include source code
+# with syntax highlighting in the RTF output.
+#
+# Note that which sources are shown also depends on other settings such as
+# SOURCE_BROWSER.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_RTF is set to YES.
+
+RTF_SOURCE_CODE        = NO
+
+#---------------------------------------------------------------------------
+# Configuration options related to the man page output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_MAN tag is set to YES, doxygen will generate man pages for
+# classes and files.
+# The default value is: NO.
+
+GENERATE_MAN           = NO
+
+# The MAN_OUTPUT tag is used to specify where the man pages will be put. If a
+# relative path is entered the value of OUTPUT_DIRECTORY will be put in front of
+# it. A directory man3 will be created inside the directory specified by
+# MAN_OUTPUT.
+# The default directory is: man.
+# This tag requires that the tag GENERATE_MAN is set to YES.
+
+MAN_OUTPUT             = man
+
+# The MAN_EXTENSION tag determines the extension that is added to the generated
+# man pages. In case the manual section does not start with a number, the number
+# 3 is prepended. The dot (.) at the beginning of the MAN_EXTENSION tag is
+# optional.
+# The default value is: .3.
+# This tag requires that the tag GENERATE_MAN is set to YES.
+
+MAN_EXTENSION          = .3
+
+# The MAN_SUBDIR tag determines the name of the directory created within
+# MAN_OUTPUT in which the man pages are placed. If defaults to man followed by
+# MAN_EXTENSION with the initial . removed.
+# This tag requires that the tag GENERATE_MAN is set to YES.
+
+MAN_SUBDIR             =
+
+# If the MAN_LINKS tag is set to YES and doxygen generates man output, then it
+# will generate one additional man file for each entity documented in the real
+# man page(s). These additional files only source the real man page, but without
+# them the man command would be unable to find the correct page.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_MAN is set to YES.
+
+MAN_LINKS              = NO
+
+#---------------------------------------------------------------------------
+# Configuration options related to the XML output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_XML tag is set to YES, doxygen will generate an XML file that
+# captures the structure of the code including all documentation.
+# The default value is: NO.
+
+GENERATE_XML           = YES
+
+# The XML_OUTPUT tag is used to specify where the XML pages will be put. If a
+# relative path is entered the value of OUTPUT_DIRECTORY will be put in front of
+# it.
+# The default directory is: xml.
+# This tag requires that the tag GENERATE_XML is set to YES.
+
+XML_OUTPUT             = xml
+
+# If the XML_PROGRAMLISTING tag is set to YES, doxygen will dump the program
+# listings (including syntax highlighting and cross-referencing information) to
+# the XML output. Note that enabling this will significantly increase the size
+# of the XML output.
+# The default value is: YES.
+# This tag requires that the tag GENERATE_XML is set to YES.
+
+XML_PROGRAMLISTING     = YES
+
+#---------------------------------------------------------------------------
+# Configuration options related to the DOCBOOK output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_DOCBOOK tag is set to YES, doxygen will generate Docbook files
+# that can be used to generate PDF.
+# The default value is: NO.
+
+GENERATE_DOCBOOK       = NO
+
+# The DOCBOOK_OUTPUT tag is used to specify where the Docbook pages will be put.
+# If a relative path is entered the value of OUTPUT_DIRECTORY will be put in
+# front of it.
+# The default directory is: docbook.
+# This tag requires that the tag GENERATE_DOCBOOK is set to YES.
+
+DOCBOOK_OUTPUT         = docbook
+
+# If the DOCBOOK_PROGRAMLISTING tag is set to YES, doxygen will include the
+# program listings (including syntax highlighting and cross-referencing
+# information) to the DOCBOOK output. Note that enabling this will significantly
+# increase the size of the DOCBOOK output.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_DOCBOOK is set to YES.
+
+DOCBOOK_PROGRAMLISTING = NO
+
+#---------------------------------------------------------------------------
+# Configuration options for the AutoGen Definitions output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_AUTOGEN_DEF tag is set to YES, doxygen will generate an
+# AutoGen Definitions (see http://autogen.sf.net) file that captures the
+# structure of the code including all documentation. Note that this feature is
+# still experimental and incomplete at the moment.
+# The default value is: NO.
+
+GENERATE_AUTOGEN_DEF   = NO
+
+#---------------------------------------------------------------------------
+# Configuration options related to the Perl module output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_PERLMOD tag is set to YES, doxygen will generate a Perl module
+# file that captures the structure of the code including all documentation.
+#
+# Note that this feature is still experimental and incomplete at the moment.
+# The default value is: NO.
+
+GENERATE_PERLMOD       = NO
+
+# If the PERLMOD_LATEX tag is set to YES, doxygen will generate the necessary
+# Makefile rules, Perl scripts and LaTeX code to be able to generate PDF and DVI
+# output from the Perl module output.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_PERLMOD is set to YES.
+
+PERLMOD_LATEX          = NO
+
+# If the PERLMOD_PRETTY tag is set to YES, the Perl module output will be nicely
+# formatted so it can be parsed by a human reader. This is useful if you want to
+# understand what is going on. On the other hand, if this tag is set to NO, the
+# size of the Perl module output will be much smaller and Perl will parse it
+# just the same.
+# The default value is: YES.
+# This tag requires that the tag GENERATE_PERLMOD is set to YES.
+
+PERLMOD_PRETTY         = YES
+
+# The names of the make variables in the generated doxyrules.make file are
+# prefixed with the string contained in PERLMOD_MAKEVAR_PREFIX. This is useful
+# so different doxyrules.make files included by the same Makefile don't
+# overwrite each other's variables.
+# This tag requires that the tag GENERATE_PERLMOD is set to YES.
+
+PERLMOD_MAKEVAR_PREFIX =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the preprocessor
+#---------------------------------------------------------------------------
+
+# If the ENABLE_PREPROCESSING tag is set to YES, doxygen will evaluate all
+# C-preprocessor directives found in the sources and include files.
+# The default value is: YES.
+
+ENABLE_PREPROCESSING   = YES
+
+# If the MACRO_EXPANSION tag is set to YES, doxygen will expand all macro names
+# in the source code. If set to NO, only conditional compilation will be
+# performed. Macro expansion can be done in a controlled way by setting
+# EXPAND_ONLY_PREDEF to YES.
+# The default value is: NO.
+# This tag requires that the tag ENABLE_PREPROCESSING is set to YES.
+
+MACRO_EXPANSION        = YES
+
+# If the EXPAND_ONLY_PREDEF and MACRO_EXPANSION tags are both set to YES then
+# the macro expansion is limited to the macros specified with the PREDEFINED and
+# EXPAND_AS_DEFINED tags.
+# The default value is: NO.
+# This tag requires that the tag ENABLE_PREPROCESSING is set to YES.
+
+EXPAND_ONLY_PREDEF     = YES
+
+# If the SEARCH_INCLUDES tag is set to YES, the include files in the
+# INCLUDE_PATH will be searched if a #include is found.
+# The default value is: YES.
+# This tag requires that the tag ENABLE_PREPROCESSING is set to YES.
+
+SEARCH_INCLUDES        = NO
+
+# The INCLUDE_PATH tag can be used to specify one or more directories that
+# contain include files that are not input files but should be processed by the
+# preprocessor.
+# This tag requires that the tag SEARCH_INCLUDES is set to YES.
+
+INCLUDE_PATH           =
+
+# You can use the INCLUDE_FILE_PATTERNS tag to specify one or more wildcard
+# patterns (like *.h and *.hpp) to filter out the header-files in the
+# directories. If left blank, the patterns specified with FILE_PATTERNS will be
+# used.
+# This tag requires that the tag ENABLE_PREPROCESSING is set to YES.
+
+INCLUDE_FILE_PATTERNS  =
+
+# The PREDEFINED tag can be used to specify one or more macro names that are
+# defined before the preprocessor is started (similar to the -D option of e.g.
+# gcc). The argument of the tag is a list of macros of the form: name or
+# name=definition (no spaces). If the definition and the "=" are omitted, "=1"
+# is assumed. To prevent a macro definition from being undefined via #undef or
+# recursively expanded use the := operator instead of the = operator.
+# This tag requires that the tag ENABLE_PREPROCESSING is set to YES.
+
+PREDEFINED             = __attribute__(x)= \
+                         __inline=
+
+# If the MACRO_EXPANSION and EXPAND_ONLY_PREDEF tags are set to YES then this
+# tag can be used to specify a list of macro names that should be expanded. The
+# macro definition that is found in the sources will be used. Use the PREDEFINED
+# tag if you want to use a different macro definition that overrules the
+# definition found in the source code.
+# This tag requires that the tag ENABLE_PREPROCESSING is set to YES.
+
+EXPAND_AS_DEFINED      =
+
+# If the SKIP_FUNCTION_MACROS tag is set to YES then doxygen's preprocessor will
+# remove all references to function-like macros that are alone on a line, have
+# an all uppercase name, and do not end with a semicolon. Such function macros
+# are typically used for boiler-plate code, and will confuse the parser if not
+# removed.
+# The default value is: YES.
+# This tag requires that the tag ENABLE_PREPROCESSING is set to YES.
+
+SKIP_FUNCTION_MACROS   = YES
+
+#---------------------------------------------------------------------------
+# Configuration options related to external references
+#---------------------------------------------------------------------------
+
+# The TAGFILES tag can be used to specify one or more tag files. For each tag
+# file the location of the external documentation should be added. The format of
+# a tag file without this location is as follows:
+# TAGFILES = file1 file2 ...
+# Adding location for the tag files is done as follows:
+# TAGFILES = file1=loc1 "file2 = loc2" ...
+# where loc1 and loc2 can be relative or absolute paths or URLs. See the
+# section "Linking to external documentation" for more information about the use
+# of tag files.
+# Note: Each tag file must have a unique name (where the name does NOT include
+# the path). If a tag file is not located in the directory in which doxygen is
+# run, you must also specify the path to the tagfile here.
+
+TAGFILES               =
+
+# When a file name is specified after GENERATE_TAGFILE, doxygen will create a
+# tag file that is based on the input files it reads. See section "Linking to
+# external documentation" for more information about the usage of tag files.
+
+GENERATE_TAGFILE       =
+
+# If the ALLEXTERNALS tag is set to YES, all external class will be listed in
+# the class index. If set to NO, only the inherited external classes will be
+# listed.
+# The default value is: NO.
+
+ALLEXTERNALS           = NO
+
+# If the EXTERNAL_GROUPS tag is set to YES, all external groups will be listed
+# in the modules index. If set to NO, only the current project's groups will be
+# listed.
+# The default value is: YES.
+
+EXTERNAL_GROUPS        = YES
+
+# If the EXTERNAL_PAGES tag is set to YES, all external pages will be listed in
+# the related pages index. If set to NO, only the current project's pages will
+# be listed.
+# The default value is: YES.
+
+EXTERNAL_PAGES         = YES
+
+# The PERL_PATH should be the absolute path and name of the perl script
+# interpreter (i.e. the result of 'which perl').
+# The default file (with absolute path) is: /usr/bin/perl.
+
+PERL_PATH              = /usr/bin/perl
+
+#---------------------------------------------------------------------------
+# Configuration options related to the dot tool
+#---------------------------------------------------------------------------
+
+# If the CLASS_DIAGRAMS tag is set to YES, doxygen will generate a class diagram
+# (in HTML and LaTeX) for classes with base or super classes. Setting the tag to
+# NO turns the diagrams off. Note that this option also works with HAVE_DOT
+# disabled, but it is recommended to install and use dot, since it yields more
+# powerful graphs.
+# The default value is: YES.
+
+CLASS_DIAGRAMS         = NO
+
+# You can define message sequence charts within doxygen comments using the \msc
+# command. Doxygen will then run the mscgen tool (see:
+# http://www.mcternan.me.uk/mscgen/)) to produce the chart and insert it in the
+# documentation. The MSCGEN_PATH tag allows you to specify the directory where
+# the mscgen tool resides. If left empty the tool is assumed to be found in the
+# default search path.
+
+MSCGEN_PATH            =
+
+# You can include diagrams made with dia in doxygen documentation. Doxygen will
+# then run dia to produce the diagram and insert it in the documentation. The
+# DIA_PATH tag allows you to specify the directory where the dia binary resides.
+# If left empty dia is assumed to be found in the default search path.
+
+DIA_PATH               =
+
+# If set to YES the inheritance and collaboration graphs will hide inheritance
+# and usage relations if the target is undocumented or is not a class.
+# The default value is: YES.
+
+HIDE_UNDOC_RELATIONS   = YES
+
+# If you set the HAVE_DOT tag to YES then doxygen will assume the dot tool is
+# available from the path. This tool is part of Graphviz (see:
+# http://www.graphviz.org/), a graph visualization toolkit from AT&T and Lucent
+# Bell Labs. The other options in this section have no effect if this option is
+# set to NO
+# The default value is: NO.
+
+HAVE_DOT               = NO
+
+# The DOT_NUM_THREADS specifies the number of dot invocations doxygen is allowed
+# to run in parallel. When set to 0 doxygen will base this on the number of
+# processors available in the system. You can set it explicitly to a value
+# larger than 0 to get control over the balance between CPU load and processing
+# speed.
+# Minimum value: 0, maximum value: 32, default value: 0.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_NUM_THREADS        = 0
+
+# When you want a differently looking font in the dot files that doxygen
+# generates you can specify the font name using DOT_FONTNAME. You need to make
+# sure dot is able to find the font, which can be done by putting it in a
+# standard location or by setting the DOTFONTPATH environment variable or by
+# setting DOT_FONTPATH to the directory containing the font.
+# The default value is: Helvetica.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_FONTNAME           = Helvetica
+
+# The DOT_FONTSIZE tag can be used to set the size (in points) of the font of
+# dot graphs.
+# Minimum value: 4, maximum value: 24, default value: 10.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_FONTSIZE           = 10
+
+# By default doxygen will tell dot to use the default font as specified with
+# DOT_FONTNAME. If you specify a different font using DOT_FONTNAME you can set
+# the path where dot can find it using this tag.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_FONTPATH           =
+
+# If the CLASS_GRAPH tag is set to YES then doxygen will generate a graph for
+# each documented class showing the direct and indirect inheritance relations.
+# Setting this tag to YES will force the CLASS_DIAGRAMS tag to NO.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+CLASS_GRAPH            = YES
+
+# If the COLLABORATION_GRAPH tag is set to YES then doxygen will generate a
+# graph for each documented class showing the direct and indirect implementation
+# dependencies (inheritance, containment, and class references variables) of the
+# class with other documented classes.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+COLLABORATION_GRAPH    = YES
+
+# If the GROUP_GRAPHS tag is set to YES then doxygen will generate a graph for
+# groups, showing the direct groups dependencies.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+GROUP_GRAPHS           = YES
+
+# If the UML_LOOK tag is set to YES, doxygen will generate inheritance and
+# collaboration diagrams in a style similar to the OMG's Unified Modeling
+# Language.
+# The default value is: NO.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+UML_LOOK               = NO
+
+# If the UML_LOOK tag is enabled, the fields and methods are shown inside the
+# class node. If there are many fields or methods and many nodes the graph may
+# become too big to be useful. The UML_LIMIT_NUM_FIELDS threshold limits the
+# number of items for each type to make the size more manageable. Set this to 0
+# for no limit. Note that the threshold may be exceeded by 50% before the limit
+# is enforced. So when you set the threshold to 10, up to 15 fields may appear,
+# but if the number exceeds 15, the total amount of fields shown is limited to
+# 10.
+# Minimum value: 0, maximum value: 100, default value: 10.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+UML_LIMIT_NUM_FIELDS   = 10
+
+# If the TEMPLATE_RELATIONS tag is set to YES then the inheritance and
+# collaboration graphs will show the relations between templates and their
+# instances.
+# The default value is: NO.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+TEMPLATE_RELATIONS     = NO
+
+# If the INCLUDE_GRAPH, ENABLE_PREPROCESSING and SEARCH_INCLUDES tags are set to
+# YES then doxygen will generate a graph for each documented file showing the
+# direct and indirect include dependencies of the file with other documented
+# files.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+INCLUDE_GRAPH          = YES
+
+# If the INCLUDED_BY_GRAPH, ENABLE_PREPROCESSING and SEARCH_INCLUDES tags are
+# set to YES then doxygen will generate a graph for each documented file showing
+# the direct and indirect include dependencies of the file with other documented
+# files.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+INCLUDED_BY_GRAPH      = YES
+
+# If the CALL_GRAPH tag is set to YES then doxygen will generate a call
+# dependency graph for every global function or class method.
+#
+# Note that enabling this option will significantly increase the time of a run.
+# So in most cases it will be better to enable call graphs for selected
+# functions only using the \callgraph command. Disabling a call graph can be
+# accomplished by means of the command \hidecallgraph.
+# The default value is: NO.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+CALL_GRAPH             = NO
+
+# If the CALLER_GRAPH tag is set to YES then doxygen will generate a caller
+# dependency graph for every global function or class method.
+#
+# Note that enabling this option will significantly increase the time of a run.
+# So in most cases it will be better to enable caller graphs for selected
+# functions only using the \callergraph command. Disabling a caller graph can be
+# accomplished by means of the command \hidecallergraph.
+# The default value is: NO.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+CALLER_GRAPH           = NO
+
+# If the GRAPHICAL_HIERARCHY tag is set to YES then doxygen will graphical
+# hierarchy of all classes instead of a textual one.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+GRAPHICAL_HIERARCHY    = YES
+
+# If the DIRECTORY_GRAPH tag is set to YES then doxygen will show the
+# dependencies a directory has on other directories in a graphical way. The
+# dependency relations are determined by the #include relations between the
+# files in the directories.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DIRECTORY_GRAPH        = YES
+
+# The DOT_IMAGE_FORMAT tag can be used to set the image format of the images
+# generated by dot. For an explanation of the image formats see the section
+# output formats in the documentation of the dot tool (Graphviz (see:
+# http://www.graphviz.org/)).
+# Note: If you choose svg you need to set HTML_FILE_EXTENSION to xhtml in order
+# to make the SVG files visible in IE 9+ (other browsers do not have this
+# requirement).
+# Possible values are: png, jpg, gif, svg, png:gd, png:gd:gd, png:cairo,
+# png:cairo:gd, png:cairo:cairo, png:cairo:gdiplus, png:gdiplus and
+# png:gdiplus:gdiplus.
+# The default value is: png.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_IMAGE_FORMAT       = png
+
+# If DOT_IMAGE_FORMAT is set to svg, then this option can be set to YES to
+# enable generation of interactive SVG images that allow zooming and panning.
+#
+# Note that this requires a modern browser other than Internet Explorer. Tested
+# and working are Firefox, Chrome, Safari, and Opera.
+# Note: For IE 9+ you need to set HTML_FILE_EXTENSION to xhtml in order to make
+# the SVG files visible. Older versions of IE do not have SVG support.
+# The default value is: NO.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+INTERACTIVE_SVG        = NO
+
+# The DOT_PATH tag can be used to specify the path where the dot tool can be
+# found. If left blank, it is assumed the dot tool can be found in the path.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_PATH               =
+
+# The DOTFILE_DIRS tag can be used to specify one or more directories that
+# contain dot files that are included in the documentation (see the \dotfile
+# command).
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOTFILE_DIRS           =
+
+# The MSCFILE_DIRS tag can be used to specify one or more directories that
+# contain msc files that are included in the documentation (see the \mscfile
+# command).
+
+MSCFILE_DIRS           =
+
+# The DIAFILE_DIRS tag can be used to specify one or more directories that
+# contain dia files that are included in the documentation (see the \diafile
+# command).
+
+DIAFILE_DIRS           =
+
+# When using plantuml, the PLANTUML_JAR_PATH tag should be used to specify the
+# path where java can find the plantuml.jar file. If left blank, it is assumed
+# PlantUML is not used or called during a preprocessing step. Doxygen will
+# generate a warning when it encounters a \startuml command in this case and
+# will not generate output for the diagram.
+
+PLANTUML_JAR_PATH      =
+
+# When using plantuml, the specified paths are searched for files specified by
+# the !include statement in a plantuml block.
+
+PLANTUML_INCLUDE_PATH  =
+
+# The DOT_GRAPH_MAX_NODES tag can be used to set the maximum number of nodes
+# that will be shown in the graph. If the number of nodes in a graph becomes
+# larger than this value, doxygen will truncate the graph, which is visualized
+# by representing a node as a red box. Note that doxygen if the number of direct
+# children of the root node in a graph is already larger than
+# DOT_GRAPH_MAX_NODES then the graph will not be shown at all. Also note that
+# the size of a graph can be further restricted by MAX_DOT_GRAPH_DEPTH.
+# Minimum value: 0, maximum value: 10000, default value: 50.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_GRAPH_MAX_NODES    = 50
+
+# The MAX_DOT_GRAPH_DEPTH tag can be used to set the maximum depth of the graphs
+# generated by dot. A depth value of 3 means that only nodes reachable from the
+# root by following a path via at most 3 edges will be shown. Nodes that lay
+# further from the root node will be omitted. Note that setting this option to 1
+# or 2 may greatly reduce the computation time needed for large code bases. Also
+# note that the size of a graph can be further restricted by
+# DOT_GRAPH_MAX_NODES. Using a depth of 0 means no depth restriction.
+# Minimum value: 0, maximum value: 1000, default value: 0.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+MAX_DOT_GRAPH_DEPTH    = 0
+
+# Set the DOT_TRANSPARENT tag to YES to generate images with a transparent
+# background. This is disabled by default, because dot on Windows does not seem
+# to support this out of the box.
+#
+# Warning: Depending on the platform used, enabling this option may lead to
+# badly anti-aliased labels on the edges of a graph (i.e. they become hard to
+# read).
+# The default value is: NO.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_TRANSPARENT        = NO
+
+# Set the DOT_MULTI_TARGETS tag to YES to allow dot to generate multiple output
+# files in one run (i.e. multiple -o and -T options on the command line). This
+# makes dot run faster, but since only newer versions of dot (>1.8.10) support
+# this, this feature is disabled by default.
+# The default value is: NO.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_MULTI_TARGETS      = NO
+
+# If the GENERATE_LEGEND tag is set to YES doxygen will generate a legend page
+# explaining the meaning of the various boxes and arrows in the dot generated
+# graphs.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+GENERATE_LEGEND        = YES
+
+# If the DOT_CLEANUP tag is set to YES, doxygen will remove the intermediate dot
+# files that are used to generate the various graphs.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_CLEANUP            = YES
diff --git a/docs/index.rst b/docs/index.rst
new file mode 100644
index 0000000000..51c0c862ae
--- /dev/null
+++ b/docs/index.rst
@@ -0,0 +1,55 @@
+============================
+Composable Kernel User Guide
+============================
+
+------------
+Introduction
+------------
+
+This document contains instructions for installing, using, and contributing to Composable Kernel (CK).
+
+-----------
+Methodology
+-----------
+
+Composable Kernel (CK) library aims to provide a programming model for writing performance critical
+kernels for machine learning workloads across multiple architectures including GPUs, CPUs, etc,
+through general purpose kernel languages, like HIP C++.
+
+CK utilizes two concepts to achieve performance portability and code maintainability:
+
+* A tile-based programming model
+* Algorithm complexity reduction for complex ML operators, using innovative technique we call
+  "Tensor Coordinate Transformation".
+
+.. image:: data/ck_component.png
+   :alt: CK Components
+
+--------------
+Code Structure
+--------------
+
+Current CK library are structured into 4 layers:
+
+* "Templated Tile Operators" layer
+* "Templated Kernel and Invoker" layer
+* "Instantiated Kernel and Invoker" layer
+* "Client API" layer
+
+.. image:: data/ck_layer.png
+   :alt: CK Layers
+   
+Documentation Roadmap
+^^^^^^^^^^^^^^^^^^^^^
+The following is a list of CK documents in the suggested reading order:
+
+.. toctree::
+   :maxdepth: 5
+   :caption: Contents:
+   :numbered:
+
+   tutorial_hello_world
+   dockerhub
+   Supported_Primitives_Guide
+   API_Reference_Guide
+   Contributors_Guide
diff --git a/docs/license.rst b/docs/license.rst
new file mode 100644
index 0000000000..ddb544496e
--- /dev/null
+++ b/docs/license.rst
@@ -0,0 +1,6 @@
+=======
+License
+=======
+
+.. include:: ../LICENSE
+   :literal:
diff --git a/docs/refs.bib b/docs/refs.bib
new file mode 100644
index 0000000000..3c6d775a7e
--- /dev/null
+++ b/docs/refs.bib
@@ -0,0 +1,7 @@
+
+@article{dao2022flashattention,
+  title={Flashattention: Fast and memory-efficient exact attention with io-awareness},
+  author={Dao, Tri and Fu, Daniel Y and Ermon, Stefano and Rudra, Atri and R{\'e}, Christopher},
+  journal={arXiv preprint arXiv:2205.14135},
+  year={2022}
+}
diff --git a/docs/sphinx/_toc.yml.in b/docs/sphinx/_toc.yml.in
new file mode 100644
index 0000000000..83dd1e7b1a
--- /dev/null
+++ b/docs/sphinx/_toc.yml.in
@@ -0,0 +1,10 @@
+# Anywhere {branch} is used, the branch name will be substituted.
+# These comments will also be removed.
+defaults:
+  numbered: False
+  maxdepth: 6
+root: index
+subtrees:
+  - caption: About
+    entries:
+      - file: license
diff --git a/docs/sphinx/requirements.in b/docs/sphinx/requirements.in
new file mode 100644
index 0000000000..a12a00a2b2
--- /dev/null
+++ b/docs/sphinx/requirements.in
@@ -0,0 +1,2 @@
+rocm-docs-core>=0.20.0
+sphinxcontrib-bibtex==2.5.0
diff --git a/docs/sphinx/requirements.txt b/docs/sphinx/requirements.txt
new file mode 100644
index 0000000000..d5f67eeb58
--- /dev/null
+++ b/docs/sphinx/requirements.txt
@@ -0,0 +1,159 @@
+#
+# This file is autogenerated by pip-compile with Python 3.8
+# by the following command:
+#
+#    pip-compile requirements.in
+#
+accessible-pygments==0.0.3
+    # via pydata-sphinx-theme
+alabaster==0.7.13
+    # via sphinx
+babel==2.12.1
+    # via
+    #   pydata-sphinx-theme
+    #   sphinx
+beautifulsoup4==4.11.2
+    # via pydata-sphinx-theme
+breathe==4.34.0
+    # via rocm-docs-core
+certifi==2022.12.7
+    # via requests
+cffi==1.15.1
+    # via
+    #   cryptography
+    #   pynacl
+charset-normalizer==3.1.0
+    # via requests
+click==8.1.3
+    # via sphinx-external-toc
+cryptography==40.0.2
+    # via pyjwt
+deprecated==1.2.13
+    # via pygithub
+docutils==0.16
+    # via
+    #   breathe
+    #   myst-parser
+    #   pybtex-docutils
+    #   pydata-sphinx-theme
+    #   sphinx
+    #   sphinxcontrib-bibtex
+fastjsonschema==2.18.0
+    # via rocm-docs-core
+gitdb==4.0.10
+    # via gitpython
+gitpython==3.1.31
+    # via rocm-docs-core
+idna==3.4
+    # via requests
+imagesize==1.4.1
+    # via sphinx
+jinja2==3.1.2
+    # via
+    #   myst-parser
+    #   sphinx
+latexcodec==2.0.1
+    # via pybtex
+markdown-it-py==2.2.0
+    # via
+    #   mdit-py-plugins
+    #   myst-parser
+markupsafe==2.1.2
+    # via jinja2
+mdit-py-plugins==0.3.5
+    # via myst-parser
+mdurl==0.1.2
+    # via markdown-it-py
+myst-parser==1.0.0
+    # via rocm-docs-core
+packaging==23.0
+    # via
+    #   pydata-sphinx-theme
+    #   sphinx
+pybtex==0.24.0
+    # via
+    #   pybtex-docutils
+    #   sphinxcontrib-bibtex
+pybtex-docutils==1.0.2
+    # via sphinxcontrib-bibtex
+pycparser==2.21
+    # via cffi
+pydata-sphinx-theme==0.13.3
+    # via
+    #   rocm-docs-core
+    #   sphinx-book-theme
+pygithub==1.58.2
+    # via rocm-docs-core
+pygments==2.14.0
+    # via
+    #   accessible-pygments
+    #   pydata-sphinx-theme
+    #   sphinx
+pyjwt[crypto]==2.6.0
+    # via pygithub
+pynacl==1.5.0
+    # via pygithub
+pyyaml==6.0
+    # via
+    #   myst-parser
+    #   pybtex
+    #   rocm-docs-core
+    #   sphinx-external-toc
+requests==2.28.2
+    # via
+    #   pygithub
+    #   sphinx
+rocm-docs-core>=0.20.0
+    # via -r requirements.in
+six==1.16.0
+    # via
+    #   latexcodec
+    #   pybtex
+smmap==5.0.0
+    # via gitdb
+snowballstemmer==2.2.0
+    # via sphinx
+soupsieve==2.4
+    # via beautifulsoup4
+sphinx==5.3.0
+    # via
+    #   breathe
+    #   myst-parser
+    #   pydata-sphinx-theme
+    #   rocm-docs-core
+    #   sphinx-book-theme
+    #   sphinx-copybutton
+    #   sphinx-design
+    #   sphinx-external-toc
+    #   sphinx-notfound-page
+    #   sphinxcontrib-bibtex
+sphinx-book-theme==1.0.1
+    # via rocm-docs-core
+sphinx-copybutton==0.5.1
+    # via rocm-docs-core
+sphinx-design==0.3.0
+    # via rocm-docs-core
+sphinx-external-toc==0.3.1
+    # via rocm-docs-core
+sphinx-notfound-page==0.8.3
+    # via rocm-docs-core
+sphinxcontrib-applehelp==1.0.4
+    # via sphinx
+sphinxcontrib-bibtex==2.5.0
+    # via -r requirements.in
+sphinxcontrib-devhelp==1.0.2
+    # via sphinx
+sphinxcontrib-htmlhelp==2.0.1
+    # via sphinx
+sphinxcontrib-jsmath==1.0.1
+    # via sphinx
+sphinxcontrib-qthelp==1.0.3
+    # via sphinx
+sphinxcontrib-serializinghtml==1.1.5
+    # via sphinx
+typing-extensions==4.5.0
+    # via pydata-sphinx-theme
+urllib3==1.26.15
+    # via requests
+wrapt==1.15.0
+    # via deprecated
diff --git a/docs/tutorial_hello_world.rst b/docs/tutorial_hello_world.rst
new file mode 100644
index 0000000000..bfb197e085
--- /dev/null
+++ b/docs/tutorial_hello_world.rst
@@ -0,0 +1,210 @@
+===============
+CK Hello world
+===============
+
+-------------------------------------
+Motivation
+-------------------------------------
+
+This tutorial is aimed at engineers dealing with artificial intelligence and machine learning who
+would like to optimize their pipelines and squeeze every performance drop by adding Composable
+Kernel (CK) library to their projects. We would like to make the CK library approachable so
+the tutorial is not based on the latest release and doesn't have all the bleeding edge features,
+but it will be reproducible now and forever.
+
+During this tutorial we will have an introduction to the CK library, we will build it and run some
+examples and tests, so to say we will run a "Hello world" example. In future tutorials we will go
+in depth and breadth and get familiar with other tools and ways to integrate CK into your project.
+
+-------------------------------------
+Description
+-------------------------------------
+
+Modern AI technology solves more and more problems in all imaginable fields, but crafting fast and
+efficient workflows is still challenging. CK is one of the tools to make AI heavy lifting as fast
+and efficient as possible. CK is a collection of optimized AI operator kernels and tools to create
+new ones. The library has components required for majority of modern neural networks architectures
+including matrix multiplication, convolution, contraction, reduction, attention modules, variety of
+activation functions, fused operators and many more.
+
+So how do we (almost) reach the speed of light? CK acceleration abilities are based on:
+
+* Layered structure.
+* Tile-based computation model.
+* Tensor coordinate transformation.
+* Hardware acceleration use.
+* Support of low precision data types including fp16, bf16, int8 and int4.
+
+If you are excited and need more technical details and benchmarking results - read this awesome
+`blog post <https://community.amd.com/t5/instinct-accelerators/amd-composable-kernel-library-efficient-fused-kernels-for-ai/ba-p/553224>`_.
+
+For more details visit our `github repository <https://github.com/ROCmSoftwarePlatform/composable_kernel>`_.
+
+-------------------------------------
+Hardware targets
+-------------------------------------
+
+CK library fully supports `gfx908` and `gfx90a` GPU architectures and only some operators are
+supported for `gfx1030`. Let's check the hardware you have at hand and decide on the target
+GPU architecture.
+
+==========     =========
+GPU Target     AMD GPU
+==========     =========
+gfx908 	       Radeon Instinct MI100
+gfx90a 	       Radeon Instinct MI210, MI250, MI250X
+gfx1030        Radeon PRO V620, W6800, W6800X, W6800X Duo, W6900X, RX 6800, RX 6800 XT, RX 6900 XT, RX 6900 XTX, RX 6950 XT
+==========     =========
+
+There are also `cloud options <https://aws.amazon.com/ec2/instance-types/g4/>`_ you can find if
+you don't have an AMD GPU at hand.
+
+-------------------------------------
+Build the library
+-------------------------------------
+
+First let's clone the library and rebase to the tested version::
+
+    git clone https://github.com/ROCmSoftwarePlatform/composable_kernel.git
+    cd composable_kernel/
+    git checkout tutorial_hello_world
+
+To make our lives easier we prepared
+`docker images <https://hub.docker.com/r/rocm/composable_kernel>`_ with all the necessary
+dependencies. Pick the right image and create a container. In this tutorial we use
+``rocm/composable_kernel:ck_ub20.04_rocm5.6`` image, it is based on Ubuntu 20.04 and
+ROCm v5.6.
+
+If your current folder is ``${HOME}``, start the docker container with::
+
+    docker run  \
+    -it  \
+    --privileged  \
+    --group-add sudo  \
+    -w /root/workspace  \
+    -v ${HOME}:/root/workspace  \
+    rocm/composable_kernel:ck_ub20.04_rocm5.6  \
+    /bin/bash
+
+If your current folder is different from ``${HOME}``, adjust the line ``-v ${HOME}:/root/workspace``
+to fit your folder structure.
+
+Inside the docker container current folder is ``~/workspace``, library path is
+``~/workspace/composable_kernel``, navigate to the library::
+
+    cd composable_kernel/
+
+Create and go to the ``build`` directory::
+
+    mkdir build && cd build
+
+In the previous section we talked about target GPU architecture. Once you decide which one is right
+for you, run CMake using the right ``GPU_TARGETS`` flag::
+
+    cmake  \
+    -D CMAKE_PREFIX_PATH=/opt/rocm  \
+    -D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc  \
+    -D CMAKE_CXX_FLAGS="-O3"  \
+    -D CMAKE_BUILD_TYPE=Release  \
+    -D BUILD_DEV=OFF  \
+    -D GPU_TARGETS="gfx908;gfx90a;gfx1030" ..
+
+If everything went well the CMake run will end up with::
+
+    -- Configuring done
+    -- Generating done
+    -- Build files have been written to: "/root/workspace/composable_kernel/build"
+
+Finally, we can build examples and tests::
+
+    make -j examples tests
+
+If everything is smooth, you'll see::
+
+    Scanning dependencies of target tests
+    [100%] Built target tests
+
+---------------------------
+Run examples and tests
+---------------------------
+
+Examples are listed as test cases as well, so we can run all examples and tests with::
+
+    ctest
+
+You can check the list of all tests by running::
+
+    ctest -N
+
+We can also run them separately, here is a separate example execution::
+
+    ./bin/example_gemm_xdl_fp16 1 1 1
+
+The arguments ``1 1 1`` mean that we want to run this example in the mode: verify results with CPU,
+initialize matrices with integers and benchmark the kernel execution. You can play around with
+these parameters and see how output and execution results change.
+
+If everything goes well and you have a device based on `gfx908` or `gfx90a` architecture you should see
+something like::
+
+    a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+    b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+    c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+    launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
+    Warm up 1 time
+    Start running 10 times...
+    Perf: 1.10017 ms, 117.117 TFlops, 87.6854 GB/s, DeviceGemmXdl<256, 256, 128, 4, 8, 32, 32, 4, 2> NumPrefetch: 1, LoopScheduler: Default, PipelineVersion: v1
+
+Meanwhile, running it on a `gfx1030` device should result in::
+
+    a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+    b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+    c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+    DeviceGemmXdl<256, 256, 128, 4, 8, 32, 32, 4, 2> NumPrefetch: 1, LoopScheduler: Default, PipelineVersion: v1 does not support this problem
+
+But don't panic, some of the operators are supported on `gfx1030` architecture, so you can run a
+separate example like::
+
+    ./bin/example_gemm_dl_fp16 1 1 1
+
+and it should result in something nice similar to::
+
+    a_m_k: dim 2, lengths {3840, 4096}, strides {1, 4096}
+    b_k_n: dim 2, lengths {4096, 4096}, strides {4096, 1}
+    c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+    arg.a_grid_desc_k0_m0_m1_k1_{2048, 3840, 2}
+    arg.b_grid_desc_k0_n0_n1_k1_{2048, 4096, 2}
+    arg.c_grid_desc_m_n_{ 3840, 4096}
+    launch_and_time_kernel: grid_dim {960, 1, 1}, block_dim {256, 1, 1}
+    Warm up 1 time
+    Start running 10 times...
+    Perf: 3.65695 ms, 35.234 TFlops, 26.3797 GB/s, DeviceGemmDl<256, 128, 128, 16, 2, 4, 4, 1>
+
+.. note::
+
+    There was a new CMake flag ``DL_KERNELS`` added in the latest versions of CK. If you use one of
+    the newest versions of the library and do not see the above results when running
+    ``example_gemm_dl_fp16``, it might be necessary to add ``-D DL_KERNELS=ON`` to your CMake command
+    in order to build the operators supported on the `gfx1030` architecture.
+
+We can also run a separate test::
+
+    ctest -R test_gemm_fp16
+
+If everything goes well you should see something like::
+
+    Start 121: test_gemm_fp16
+    1/1 Test #121: test_gemm_fp16 ...................   Passed   51.81 sec
+
+    100% tests passed, 0 tests failed out of 1
+
+-----------
+Summary
+-----------
+
+In this tutorial we took the first look at the Composable Kernel library, built it on your system
+and ran some examples and tests. Stay tuned, in the next tutorial we will run kernels with different
+configs to find out the best one for your hardware and task.
+
+P.S.: Don't forget to switch off the cloud instance if you have launched one, you can find better
+ways to spend your money for sure!
diff --git a/example/01_gemm/CMakeLists.txt b/example/01_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..a671877f68
--- /dev/null
+++ b/example/01_gemm/CMakeLists.txt
@@ -0,0 +1,74 @@
+if(DL_KERNELS)
+  add_custom_target(example_gemm_dl)
+
+  add_example_executable(example_gemm_dl_fp32 gemm_dl_fp32.cpp)
+  add_dependencies(example_gemm_dl example_gemm_dl_fp32)
+  if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_gemm_dl_fp16 gemm_dl_fp16.cpp)
+    add_dependencies(example_gemm_dl example_gemm_dl_fp16)
+    add_example_executable(example_gemm_dl_dpp8_fp16 gemm_dl_dpp8_fp16.cpp)
+    add_dependencies(example_gemm_dl example_gemm_dl_dpp8_fp16)
+  endif()
+  if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+    add_example_executable(example_gemm_dl_int8 gemm_dl_int8.cpp)
+    add_dependencies(example_gemm_dl example_gemm_dl_int8)
+  endif()
+
+  if(USE_BITINT_EXTENSION_INT4)
+    add_example_executable(example_gemm_dl_int4 gemm_dl_int4.cpp)
+    add_dependencies(example_gemm_dl example_gemm_dl_int4)
+  endif(USE_BITINT_EXTENSION_INT4)
+endif()
+
+add_custom_target(example_gemm_xdl)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_gemm_xdl_fp16 gemm_xdl_fp16.cpp)
+  add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_fp16)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)
+  add_example_executable(example_gemm_xdl_skip_b_lds_fp16 gemm_xdl_skip_b_lds_fp16.cpp)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_skip_b_lds_fp16)
+
+  if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
+    add_custom_target(example_gemm_wmma)
+    add_example_executable(example_gemm_wmma_fp16 gemm_wmma_fp16.cpp)
+    add_dependencies(example_gemm_wmma example_gemm_wmma_fp16)
+  endif()
+
+endif()
+
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_gemm_xdl_bf16 gemm_xdl_bf16.cpp)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_bf16)
+
+  add_example_executable(example_gemm_xdl_bf16_rtn gemm_xdl_bf16_rtn.cpp)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_bf16_rtn)
+endif()
+
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+  add_example_executable(example_gemm_xdl_int8 gemm_xdl_int8.cpp)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_int8)
+endif()
+
+if(USE_BITINT_EXTENSION_INT4)
+  add_example_executable(example_gemm_xdl_int4 gemm_xdl_int4.cpp)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_int4)
+endif(USE_BITINT_EXTENSION_INT4)
+
+if(DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
+  # FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
+  add_example_executable_no_testing(example_gemm_xdl_fp64 gemm_xdl_fp64.cpp)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_fp64)
+endif()
+
+add_example_executable(example_gemm_xdl_streamk gemm_xdl_streamk.cpp)
+
+if(DTYPES MATCHES "fp8" OR NOT DEFINED DTYPES)
+  if(GPU_TARGETS MATCHES "gfx940" OR GPU_TARGETS MATCHES "gfx941" OR GPU_TARGETS MATCHES "gfx942")
+    add_example_executable(example_gemm_xdl_f8 gemm_xdl_f8.cpp)
+    add_dependencies(example_gemm_xdl example_gemm_xdl_f8)
+  endif()
+endif()
+
+add_example_executable(example_gemm_xdl_fp16_f8 gemm_xdl_fp16_f8.cpp)
+add_dependencies(example_gemm_xdl example_gemm_xdl_fp16_f8)
diff --git a/example/01_gemm/README.md b/example/01_gemm/README.md
new file mode 100644
index 0000000000..226783b03b
--- /dev/null
+++ b/example/01_gemm/README.md
@@ -0,0 +1,23 @@
+# Instructions for ```example_gemm_xdl```
+
+## Run ```example_gemm_xdl```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: run kernel # of times (>1)
+./bin/example_gemm_xdl 0 1 5
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
+arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
+arg.c_grid_desc_m_n_{ 3840, 4096}
+launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
+Warm up
+Start running 5 times...
+Perf: 1.19685 ms, 107.657 TFlops, 78.8501 GB/s
+```
diff --git a/example/01_gemm/common.hpp b/example/01_gemm/common.hpp
new file mode 100644
index 0000000000..7fd15b2833
--- /dev/null
+++ b/example/01_gemm/common.hpp
@@ -0,0 +1,160 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <numeric>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+struct ProblemSize final
+{
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+};
+
+struct ProblemSizeStreamK final
+{
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    ck::index_t NumSKBlocks = -1;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+template <typename ProblemType>
+bool parse_cmd_args(int, char*[], ProblemType&, ExecutionConfig&)
+{
+    return false;
+}
+
+template <>
+bool parse_cmd_args<ProblemSize>(int argc,
+                                 char* argv[],
+                                 ProblemSize& problem_size,
+                                 ExecutionConfig& config)
+{
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        problem_size.M = std::stoi(argv[4]);
+        problem_size.N = std::stoi(argv[5]);
+        problem_size.K = std::stoi(argv[6]);
+
+        problem_size.StrideA = std::stoi(argv[7]);
+        problem_size.StrideB = std::stoi(argv[8]);
+        problem_size.StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
+                  << std::endl
+                  << "arg3: time kernel (0=no, 1=yes)" << std::endl
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
+template <>
+bool parse_cmd_args<ProblemSizeStreamK>(int argc,
+                                        char* argv[],
+                                        ProblemSizeStreamK& problem_size,
+                                        ExecutionConfig& config)
+{
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc >= 10)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        problem_size.M = std::stoi(argv[4]);
+        problem_size.N = std::stoi(argv[5]);
+        problem_size.K = std::stoi(argv[6]);
+
+        problem_size.StrideA = std::stoi(argv[7]);
+        problem_size.StrideB = std::stoi(argv[8]);
+        problem_size.StrideC = std::stoi(argv[9]);
+
+        if(argc >= 11)
+        {
+            problem_size.NumSKBlocks = std::stoi(argv[10]);
+        }
+    }
+    else
+    {
+        std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
+                  << std::endl
+                  << "arg3: time kernel (0=no, 1=yes)" << std::endl
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC" << std::endl
+                  << "arg10: NumSKBlocks(optional)" << std::endl;
+        return false;
+    }
+
+    return true;
+}
diff --git a/example/01_gemm/gemm_dl_dpp8_fp16.cpp b/example/01_gemm/gemm_dl_dpp8_fp16.cpp
new file mode 100644
index 0000000000..ea0ba39076
--- /dev/null
+++ b/example/01_gemm/gemm_dl_dpp8_fp16.cpp
@@ -0,0 +1,37 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl_dpp8.hpp"
+
+using ADataType   = ck::half_t;
+using BDataType   = ck::half_t;
+using CDataType   = ck::half_t;
+using AccDataType = float;
+
+using ALayout = Col;
+using BLayout = Row;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDlDpp8
+// ######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   128,   128,    16,  2,          1,          8,      8,       S<8, 8>,       S<4, 1>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,               5,                  4>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/example/01_gemm/gemm_dl_fp16.cpp b/example/01_gemm/gemm_dl_fp16.cpp
new file mode 100644
index 0000000000..b5fecb9752
--- /dev/null
+++ b/example/01_gemm/gemm_dl_fp16.cpp
@@ -0,0 +1,37 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+
+using ADataType   = ck::half_t;
+using BDataType   = ck::half_t;
+using CDataType   = ck::half_t;
+using AccDataType = float;
+
+using ALayout = Col;
+using BLayout = Row;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDl
+// ######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,               5,                  4>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/example/01_gemm/gemm_dl_fp32.cpp b/example/01_gemm/gemm_dl_fp32.cpp
new file mode 100644
index 0000000000..212b72f2a6
--- /dev/null
+++ b/example/01_gemm/gemm_dl_fp32.cpp
@@ -0,0 +1,37 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+
+using ADataType   = float;
+using BDataType   = float;
+using CDataType   = float;
+using AccDataType = float;
+
+using ALayout = Col;
+using BLayout = Row;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDl
+// ######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 1>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<2, 1, 4, 1>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 1>, S<0, 1, 2, 3, 4, 5>,               5,                  4>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/example/01_gemm/gemm_dl_int4.cpp b/example/01_gemm/gemm_dl_int4.cpp
new file mode 100644
index 0000000000..e55ae14013
--- /dev/null
+++ b/example/01_gemm/gemm_dl_int4.cpp
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+#error Should compile this file with ck::int4_t support
+#endif
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+
+using ADataType       = ck::int4_t;
+using BDataType       = ck::int4_t;
+using CDataType       = ck::int4_t;
+using KernelADataType = int8_t;
+using KernelBDataType = int8_t;
+using KernelCDataType = int8_t;
+using AccDataType     = int32_t;
+
+using ALayout = Col;
+using BLayout = Row;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDl
+// ######|           AData|           BData|           CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|            Type|            Type|            Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|                |                |                |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|                |                |                |            |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < KernelADataType, KernelBDataType, KernelCDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 4>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,               5,                  4>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#define BUILD_INT4_EXAMPLE
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/example/01_gemm/gemm_dl_int8.cpp b/example/01_gemm/gemm_dl_int8.cpp
new file mode 100644
index 0000000000..1840390aa9
--- /dev/null
+++ b/example/01_gemm/gemm_dl_int8.cpp
@@ -0,0 +1,37 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+
+using ADataType   = int8_t;
+using BDataType   = int8_t;
+using CDataType   = int8_t;
+using AccDataType = int32_t;
+
+using ALayout = Col;
+using BLayout = Row;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDl
+// ######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,               5,                  4>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/example/01_gemm/gemm_wmma_fp16.cpp b/example/01_gemm/gemm_wmma_fp16.cpp
new file mode 100644
index 0000000000..b11fe76ab2
--- /dev/null
+++ b/example/01_gemm/gemm_wmma_fp16.cpp
@@ -0,0 +1,38 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp"
+
+using ADataType        = ck::half_t;
+using BDataType        = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = float;
+using CDataType        = ck::half_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmWmma_CShuffle
+// ######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     AccData|         CShuffle|           A|           B|           C|              GEMM| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer|MRepeat|NRepeat|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+// ######|        |        |        |      Type|      Type|      Type|        Type|         DataType| Elementwise| Elementwise| Elementwise|    Spacialization|  Size| Block| Block| Block|    | WMMA| WMMA|       |       |   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN|MWmmaPerWave|NWmmaPerWave|        _MBlock_MWaveMPerWmma| ScalarPerVector|
+// ######|        |        |        |          |          |          |            |                 |   Operation|   Operation|   Operation|                  |      |      |      |      |    |     |     |       |       | Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|        _NBlock_NWaveNPerWmma|  _NWaveNPerWmma|
+// ######|        |        |        |          |          |          |            |                 |            |            |            |                  |      |      |      |      |    |     |     |       |       |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmMNKPadding,   256,   128,   256,     8,   8,   16,   16,      4,      4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,              S<1, 32, 1,  8>,               8, 1>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/example/01_gemm/gemm_xdl_bf16.cpp b/example/01_gemm/gemm_xdl_bf16.cpp
new file mode 100644
index 0000000000..3cac55ef47
--- /dev/null
+++ b/example/01_gemm/gemm_xdl_bf16.cpp
@@ -0,0 +1,38 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+
+using ADataType        = ck::bhalf_t;
+using BDataType        = ck::bhalf_t;
+using CDataType        = ck::bhalf_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::bhalf_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
+// ######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     AccData|         CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+// ######|        |        |        |      Type|      Type|      Type|        Type|         DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+// ######|        |        |        |          |          |          |            |                 |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+// ######|        |        |        |          |          |          |            |                 |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/example/01_gemm/gemm_xdl_bf16_rtn.cpp b/example/01_gemm/gemm_xdl_bf16_rtn.cpp
new file mode 100644
index 0000000000..cc14dcb8eb
--- /dev/null
+++ b/example/01_gemm/gemm_xdl_bf16_rtn.cpp
@@ -0,0 +1,39 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/utility/type_convert.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+
+using ADataType        = ck::bhalf_t;
+using BDataType        = ck::bhalf_t;
+using CDataType        = ck::bhalf_t;
+using AccDataType      = float;
+using CShuffleDataType = float;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = ck::tensor_operation::element_wise::ConvertBF16RTN;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
+// ######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     AccData|         CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+// ######|        |        |        |      Type|      Type|      Type|        Type|         DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+// ######|        |        |        |          |          |          |            |                 |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+// ######|        |        |        |          |          |          |            |                 |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/example/01_gemm/gemm_xdl_f8.cpp b/example/01_gemm/gemm_xdl_f8.cpp
new file mode 100644
index 0000000000..1015926777
--- /dev/null
+++ b/example/01_gemm/gemm_xdl_f8.cpp
@@ -0,0 +1,38 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+
+using ADataType        = ck::f8_t;
+using BDataType        = ck::f8_t;
+using CDataType        = ck::f8_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::f8_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
+// ######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     AccData|         CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+// ######|        |        |        |      Type|      Type|      Type|        Type|         DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+// ######|        |        |        |          |          |          |            |                 |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+// ######|        |        |        |          |          |          |            |                 |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 64, 1, 4>,              16>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/example/01_gemm/gemm_xdl_fp16.cpp b/example/01_gemm/gemm_xdl_fp16.cpp
new file mode 100644
index 0000000000..4d72737142
--- /dev/null
+++ b/example/01_gemm/gemm_xdl_fp16.cpp
@@ -0,0 +1,54 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+
+using ADataType        = ck::half_t;
+using BDataType        = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = float;
+using CDataType        = ck::half_t;
+
+using F16 = ck::half_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance0 = ck::tensor_operation::device::DeviceGemmXdl
+// ######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+// ######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise|Spacialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+// ######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+// ######|          |          |          |            |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+         < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>;
+// // clang-format on
+
+// clang-format off
+using DeviceGemmInstance1 = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
+// ######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     AccData|         CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+// ######|        |        |        |      Type|      Type|      Type|        Type|         DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+// ######|        |        |        |          |          |          |            |                 |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+// ######|        |        |        |          |          |          |            |                 |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, 
+         ck::make_default_loop_scheduler(),
+         ck::PipelineVersion::v2>;
+// clang-format on
+
+using DeviceGemmInstance = DeviceGemmInstance1;
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/example/01_gemm/gemm_xdl_fp16_f8.cpp b/example/01_gemm/gemm_xdl_fp16_f8.cpp
new file mode 100644
index 0000000000..d3cf3d397a
--- /dev/null
+++ b/example/01_gemm/gemm_xdl_fp16_f8.cpp
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+
+using ADataType        = ck::f8_t;
+using BDataType        = ck::half_t;
+using CDataType        = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::half_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto LoopSched   = ck::make_default_loop_scheduler();
+static constexpr auto PipelineVer = ck::PipelineVersion::v1;
+using ComputeType                 = ck::half_t;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
+// ######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     AccData|         CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|      Loop|    Pipeline|  ComputeType|
+// ######|        |        |        |      Type|      Type|      Type|        Type|         DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector| Scheduler|     Version|             |
+// ######|        |        |        |          |          |          |            |                 |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|          |            |             |
+// ######|        |        |        |          |          |          |            |                 |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |          |            |             |
+        < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,  LoopSched, PipelineVer, ComputeType>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/example/01_gemm/gemm_xdl_fp64.cpp b/example/01_gemm/gemm_xdl_fp64.cpp
new file mode 100644
index 0000000000..8361576299
--- /dev/null
+++ b/example/01_gemm/gemm_xdl_fp64.cpp
@@ -0,0 +1,46 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+
+using ADataType   = double;
+using BDataType   = double;
+using CDataType   = double;
+using AccDataType = double;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
+// ######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+// ######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise|Spacialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+// ######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+// ######|          |          |          |            |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+#if 0
+         < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,   GemmDefault,    64,    32,    32,     4,  1,   16,   16,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,               7,               1>;
+#else
+         < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,   GemmDefault,   256,   128,   128,     4,  2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>;
+#endif
+    // clang-format on
+
+    using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                            BDataType,
+                                                                            CDataType,
+                                                                            AccDataType,
+                                                                            AElementOp,
+                                                                            BElementOp,
+                                                                            CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/example/01_gemm/gemm_xdl_int4.cpp b/example/01_gemm/gemm_xdl_int4.cpp
new file mode 100644
index 0000000000..f6238c7aa5
--- /dev/null
+++ b/example/01_gemm/gemm_xdl_int4.cpp
@@ -0,0 +1,46 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+#error Should compile this file with ck::int4_t support
+#endif
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+
+using ADataType        = ck::int4_t;
+using BDataType        = ck::int4_t;
+using CDataType        = ck::int4_t;
+using KernelADataType  = int8_t;
+using KernelBDataType  = int8_t;
+using KernelCDataType  = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int8_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
+// ######| ALayout| BLayout| CLayout|           AData|           BData|           CData|     AccData|         CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+// ######|        |        |        |            Type|            Type|            Type|        Type|         DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+// ######|        |        |        |                |                |                |            |                 |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+// ######|        |        |        |                |                |                |            |                 |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < ALayout, BLayout, CLayout, KernelADataType, KernelBDataType, KernelCDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,          1,          1,           1,               S<1, 64, 1, 4>,              16>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#define BUILD_INT4_EXAMPLE
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/example/01_gemm/gemm_xdl_int8.cpp b/example/01_gemm/gemm_xdl_int8.cpp
new file mode 100644
index 0000000000..cc03200b9d
--- /dev/null
+++ b/example/01_gemm/gemm_xdl_int8.cpp
@@ -0,0 +1,38 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+
+using ADataType        = int8_t;
+using BDataType        = int8_t;
+using CDataType        = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int8_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
+// ######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     AccData|         CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+// ######|        |        |        |      Type|      Type|      Type|        Type|         DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+// ######|        |        |        |          |          |          |            |                 |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+// ######|        |        |        |          |          |          |            |                 |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 64, 1, 4>,              16>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp b/example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp
new file mode 100644
index 0000000000..4a0c23cf44
--- /dev/null
+++ b/example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp
@@ -0,0 +1,249 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp"
+
+#include "ck/library/utility/literals.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+#define USING_SKIP_LDS 1
+
+// clang-format off
+#if USING_SKIP_LDS
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSkipBLds
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BThreadTransfer|  BBlock| CThreadTransfer| CThreadTransfer|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Spacialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|       SrcScalar|  buffer| SrcDstVectorDim|       DstScalar|
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |       PerVector|  size  |                |       PerVector|
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |        |                |                |
+#if 0       
+                    <   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    16,   64,     4,  8,   16,   16,    1,    1,     S<16, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,            8,       8,        7,               1>;
+using ADataType   = ck::half_t;
+using BDataType   = ck::half_t;
+using CDataType   = ck::half_t;
+using AccDataType = float;
+#else  
+                    <   F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    16,   64,     4,  4,   16,   16,    1,    1,     S<16, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,             4,      4,        7,               1>;
+using ADataType   = float;
+using BDataType   = float;
+using CDataType   = float;
+using AccDataType = float;
+#endif
+
+#else
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Spacialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+                   <   F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    16,   64,     4,  4,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,              1,      2>;
+using ADataType   = float;
+using BDataType   = float;
+using CDataType   = float;
+using AccDataType = float;
+
+#endif
+    // clang-format on
+
+    using ReferenceGemmInstance = ck::tensor_operation::host::
+        ReferenceGemm<ADataType, BDataType, CDataType, float, AElementOp, BElementOp, CElementOp>;
+
+template <typename DataType>
+std::ostream& show_2d_matrix(std::ostream& os, Tensor<DataType>& matrix)
+{
+    os << "[" << std::endl;
+    for(size_t x = 0; x < matrix.mDesc.GetLengths()[0]; x++)
+    {
+        os << "[";
+        for(size_t y = 0; y < matrix.mDesc.GetLengths()[1]; y++)
+        {
+            os << std::setw(5) << static_cast<float>(matrix(x, y));
+        }
+        os << "]" << std::endl;
+    }
+    os << "]";
+    return os;
+}
+int main(int argc, char* argv[])
+{
+    bool do_verification = 0;
+    int init_method      = 0;
+    bool time_kernel     = false;
+
+    // GEMM shape
+#if 1
+    ck::index_t M = 16;
+    ck::index_t N = 64 * 120;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideC = N;
+#else
+    ck::index_t M = 16;
+    ck::index_t N = 16;
+    ck::index_t K = 32;
+
+    ck::index_t StrideA = 8;
+    ck::index_t StrideB = 8;
+    ck::index_t StrideC = 16;
+#endif
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    default:
+        // a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+    }
+
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+                                      static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+                                      static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      StrideC,
+                                      a_element_op,
+                                      b_element_op,
+                                      c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+#if 0
+        {
+            show_2d_matrix(std::cout << "a : ", a_m_k) << std::endl;
+            show_2d_matrix(std::cout << "b: ", b_k_n) << std::endl;
+            show_2d_matrix(std::cout << "c_device: ", c_m_n_device_result) << std::endl;
+            show_2d_matrix(std::cout << "c_host  :", c_m_n_host_result) << std::endl;
+        }
+#endif
+        ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+    }
+
+    return 0;
+}
diff --git a/example/01_gemm/gemm_xdl_streamk.cpp b/example/01_gemm/gemm_xdl_streamk.cpp
new file mode 100644
index 0000000000..7d433b6145
--- /dev/null
+++ b/example/01_gemm/gemm_xdl_streamk.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_streamk.hpp"
+
+using ADataType        = ck::half_t;
+using BDataType        = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = float;
+using CDataType        = ck::half_t;
+
+using F16 = ck::half_t;
+
+using ALayout = Row;
+using BLayout = Row;
+// using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+// clang-format off
+using DeviceGemmStreamK = ck::tensor_operation::device::DeviceGemmXdlStreamK
+// ######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|  Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+// ######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise|   Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+// ######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|       |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+// ######|          |          |          |            |        |        |        |            |            |            |       |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+    < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,      S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+
+          // < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    256,   256,   128,     4,   8,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,         1,           1,           1,               S<1, 32, 1, 8>,              8>;
+      //  < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    128,   32,   64,     4,   8,  32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>;
+       // < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    128,   32,   128,     4,   8,  32,   32,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,         1,           1,           1,               S<1, 32, 1, 4>,              8>;
+
+
+
+// // clang-format on
+// clang-format on
+
+using DeviceGemmInstance = DeviceGemmStreamK;
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_streamk_example(argc, argv); }
diff --git a/example/01_gemm/gemm_xdl_wavelet_fp16.cpp b/example/01_gemm/gemm_xdl_wavelet_fp16.cpp
new file mode 100644
index 0000000000..b0f963fee5
--- /dev/null
+++ b/example/01_gemm/gemm_xdl_wavelet_fp16.cpp
@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp"
+
+using ADataType        = ck::half_t;
+using BDataType        = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = float;
+using CDataType        = ck::half_t;
+
+using F16 = ck::half_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_WaveletModel_CShuffle
+    // clang-format off
+// ######| ALayout| BLayout| CLayout|     AData|     BData|     AccData|         CShuffle|     CData|           A|           B|           C|           GEMM| NumGemmK| ABBlockTransfer|       BlockGemm|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+// ######|        |        |        |      Type|      Type|        Type|         DataType|      Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| ThreadGroupSize| ThreadGroupSize| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+// ######|        |        |        |          |          |            |                 |          |   Operation|   Operation|   Operation|               |    Stage|                |                |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+// ######|        |        |        |          |          |            |                 |          |            |            |            |               |         |                |                |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < ALayout, BLayout, CLayout, ADataType, BDataType, AccDataType,              F16, CDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,             256,             256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,                S<1, 32, 1,8>,               8>;
+// clang-format on
+
+using DeviceGemmInstance = DeviceGemmInstance;
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
diff --git a/example/01_gemm/run_gemm_example.inc b/example/01_gemm/run_gemm_example.inc
new file mode 100644
index 0000000000..7be2539d90
--- /dev/null
+++ b/example/01_gemm/run_gemm_example.inc
@@ -0,0 +1,264 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/device_gemm_streamk.hpp"
+
+template <typename ProblemType>
+bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
+{
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+#endif
+
+    using namespace ck::literals;
+
+    auto M       = problem_size.M;
+    auto N       = problem_size.N;
+    auto K       = problem_size.K;
+    auto StrideA = problem_size.StrideA;
+    auto StrideB = problem_size.StrideB;
+    auto StrideC = problem_size.StrideC;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    auto f_get_default_stride =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(stride == 0)
+            {
+                // give a chance if stride is zero, return a default packed stride
+                if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+                {
+                    return col;
+                }
+                else
+                {
+                    return row;
+                }
+            }
+            else
+                return stride;
+        };
+
+    StrideA = f_get_default_stride(M, K, StrideA, ALayout{});
+    StrideB = f_get_default_stride(K, N, StrideB, BLayout{});
+    StrideC = f_get_default_stride(M, N, StrideC, CLayout{});
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+
+    switch(config.init_method)
+    {
+    case 0:
+        ck::utils::FillConstant<ADataType>{static_cast<ADataType>(1.f)}(a_m_k);
+        ck::utils::FillConstant<BDataType>{static_cast<BDataType>(1.f)}(b_k_n);
+        break;
+    case 1:
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
+        ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
+        break;
+    default:
+        ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
+        ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
+    }
+
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+
+#ifdef BUILD_INT4_EXAMPLE
+    DeviceMem a_m_k_device_buf(sizeof(KernelADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(KernelBDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(KernelCDataType) *
+                               c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    const Tensor<KernelADataType> a_m_k_converted(a_m_k);
+    const Tensor<KernelBDataType> b_k_n_converted(b_k_n);
+
+    a_m_k_device_buf.ToDevice(a_m_k_converted.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n_converted.mData.data());
+#else
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
+#endif
+    DeviceMem workspace;
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+
+    using BaseStreamK = ck::tensor_operation::device::DeviceGemmStreamK<ALayout,
+                                                                        BLayout,
+                                                                        CLayout,
+                                                                        ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CElementOp>;
+
+    // do GEMM
+    auto gemm      = DeviceGemmInstance{};
+    auto invoker   = gemm.MakeInvoker();
+    float ave_time = 0;
+
+    if constexpr(std::is_same<ProblemType, ProblemSize>::value &&
+                 !std::is_base_of<BaseStreamK, DeviceGemmInstance>::value)
+    {
+        auto argument = gemm.MakeArgument(
+#ifdef BUILD_INT4_EXAMPLE
+            static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+            static_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+            static_cast<KernelCDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+#else
+            static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+            static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+            static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+#endif
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            StrideC,
+            a_element_op,
+            b_element_op,
+            c_element_op);
+
+        if(!gemm.IsSupportedArgument(argument))
+        {
+            std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+            return true;
+        }
+
+        ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+    }
+    else if constexpr(std::is_same<ProblemType, ProblemSizeStreamK>::value &&
+                      std::is_base_of<BaseStreamK, DeviceGemmInstance>::value)
+    {
+        auto argument = gemm.MakeArgument(
+#ifdef BUILD_INT4_EXAMPLE
+            static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+            static_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+            static_cast<KernelCDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+#else
+            static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+            static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+            static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+#endif
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            StrideC,
+            a_element_op,
+            b_element_op,
+            c_element_op,
+            problem_size.NumSKBlocks);
+
+        if(!gemm.IsSupportedArgument(argument))
+        {
+            std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+            return true;
+        }
+
+        std::size_t workspace_size = gemm.GetWorkSpaceSize(&argument);
+        if(workspace_size != 0)
+        {
+            workspace.Realloc(workspace_size);
+            gemm.SetWorkSpacePointer(&argument, workspace.GetDeviceBuffer());
+        }
+
+        ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+#if 0
+        // TODO!!!!!
+        if(workspace_size != 0){
+            float * ws_ptr = reinterpret_cast<float*>(malloc(workspace_size));
+            size_t ws_dwords = workspace_size / sizeof(float);
+            workspace.FromDevice(ws_ptr);
+
+            for(size_t i = 0; i < ws_dwords; i++) {
+                uint32_t rere = reinterpret_cast<uint32_t*>(ws_ptr)[i];
+                printf("%4lu : %f(0x%08x)\n", i, ws_ptr[i], rere);
+            }
+            free(ws_ptr);
+        }
+#endif
+    }
+
+    std::size_t flop = 2_uz * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+#ifdef BUILD_INT4_EXAMPLE
+        Tensor<CDataType> c_m_n_device_result_converted(c_m_n_host_result.mDesc);
+
+        c_m_n_device_buf.FromDevice(c_m_n_device_result_converted.mData.data());
+
+        c_m_n_device_result = c_m_n_device_result_converted.CopyAsType<CDataType>();
+
+        return ck::utils::check_err(c_m_n_device_result_converted, c_m_n_host_result);
+#else
+        c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+#endif
+    }
+
+    return true;
+}
+
+bool run_gemm_example(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    return !parse_cmd_args(argc, argv, problem_size, config) || run_gemm(problem_size, config);
+}
+
+bool run_gemm_streamk_example(int argc, char* argv[])
+{
+    ProblemSizeStreamK problem_size;
+    ExecutionConfig config;
+
+    return !parse_cmd_args(argc, argv, problem_size, config) || run_gemm(problem_size, config);
+}
diff --git a/example/02_gemm_bilinear/CMakeLists.txt b/example/02_gemm_bilinear/CMakeLists.txt
new file mode 100644
index 0000000000..9989e45e0c
--- /dev/null
+++ b/example/02_gemm_bilinear/CMakeLists.txt
@@ -0,0 +1,19 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+list(APPEND gpu_list1 gfx1100 gfx1101 gfx1102)
+list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
+    add_example_executable(example_gemm_bilinear_wmma_fp16 gemm_bilinear_wmma_fp16.cpp)
+   set(target 1)
+ endif()
+endforeach()
+
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
+    add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp)
+   set(target 1)
+ endif()
+endforeach()
+endif()
diff --git a/example/02_gemm_bilinear/README.md b/example/02_gemm_bilinear/README.md
new file mode 100644
index 0000000000..9eb87e1e34
--- /dev/null
+++ b/example/02_gemm_bilinear/README.md
@@ -0,0 +1,28 @@
+# Instructions for ```example_gemm_bilinear_xdl_fp16```
+
+## Run ```example_gemm_bilinear_xdl_fp16```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE
+#arg11 to 12: alpha, beta
+./bin/example_gemm_bilinear_xdl_fp16 1 1 1 3840 4096 4096 4096 4096 4096 4096 0.5 0.5
+```
+Result (MI100 @ 1502Mhz, 184.6TFlops peak FP16)
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+c0_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
+arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
+arg.c0_grid_desc_m_n_{ 3840, 4096}
+arg.c_grid_desc_m_n_{ 3840, 4096}
+launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
+Warm up
+Start running 1 times...
+Perf: 0.936965 ms, 137.517 TFlops, 102.959 GB/s
+error: 0
+max_diff: 0, 558.5, 558.5
+```
diff --git a/example/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp b/example/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp
new file mode 100644
index 0000000000..877792d740
--- /dev/null
+++ b/example/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp
@@ -0,0 +1,304 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+struct AlphaBetaAdd
+{
+    AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
+
+    template <typename E, typename C, typename D>
+    __host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::half_t, float, ck::half_t>(
+        ck::half_t& e, const float& c, const ck::half_t& d) const
+    {
+        e = ck::type_convert<ck::half_t>(alpha_ * c + beta_ * ck::type_convert<float>(d));
+    };
+
+    float alpha_;
+    float beta_;
+};
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F16;
+using EDataType        = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using DLayout = Row;
+using ELayout = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AlphaBetaAdd;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceOpInstance =
+    ck::tensor_operation::device::DeviceGemmMultipleD_Wmma_CShuffle<ALayout,
+                                                                    BLayout,
+                                                                    ck::Tuple<DLayout>,
+                                                                    ELayout,
+                                                                    ADataType,
+                                                                    BDataType,
+                                                                    ck::Tuple<DDataType>,
+                                                                    EDataType,
+                                                                    AccDataType,
+                                                                    CShuffleDataType,
+                                                                    AElementOp,
+                                                                    BElementOp,
+                                                                    CDEElementOp,
+                                                                    GemmSpec,
+                                                                    256,
+                                                                    128,
+                                                                    256,
+                                                                    8,
+                                                                    8,
+                                                                    16,
+                                                                    16,
+                                                                    4,
+                                                                    4,
+                                                                    S<4, 64, 1>,
+                                                                    S<1, 0, 2>,
+                                                                    S<1, 0, 2>,
+                                                                    2,
+                                                                    8,
+                                                                    8,
+                                                                    true,
+                                                                    S<4, 64, 1>,
+                                                                    S<1, 0, 2>,
+                                                                    S<1, 0, 2>,
+                                                                    2,
+                                                                    8,
+                                                                    8,
+                                                                    true,
+                                                                    1,
+                                                                    1,
+                                                                    S<1, 32, 1, 8>,
+                                                                    8>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideD = 4096;
+    ck::index_t StrideE = 4096;
+
+    float alpha = 1.0f;
+    float beta  = 1.0f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        alpha = std::stof(argv[4]);
+        beta  = std::stof(argv[5]);
+    }
+    else if(argc == 13)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[10]);
+
+        alpha = std::stof(argv[11]);
+        beta  = std::stof(argv[12]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, alpha, "
+               "beta\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DLayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d_device_buf.ToDevice(d_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{alpha, beta};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               std::array<ck::index_t, 1>{StrideD},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp b/example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp
new file mode 100644
index 0000000000..c3e6ef7d5d
--- /dev/null
+++ b/example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp
@@ -0,0 +1,306 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+struct AlphaBetaAdd
+{
+    AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
+
+    template <typename E, typename C, typename D>
+    __host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::half_t, float, ck::half_t>(
+        ck::half_t& e, const float& c, const ck::half_t& d) const
+    {
+        e = ck::type_convert<ck::half_t>(alpha_ * c + beta_ * ck::type_convert<float>(d));
+    };
+
+    float alpha_;
+    float beta_;
+};
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F16;
+using EDataType        = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using DLayout = Row;
+using ELayout = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AlphaBetaAdd;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceOpInstance =
+    ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<ALayout,
+                                                                   BLayout,
+                                                                   ck::Tuple<DLayout>,
+                                                                   ELayout,
+                                                                   ADataType,
+                                                                   BDataType,
+                                                                   AccDataType,
+                                                                   CShuffleDataType,
+                                                                   ck::Tuple<DDataType>,
+                                                                   EDataType,
+                                                                   AElementOp,
+                                                                   BElementOp,
+                                                                   CDEElementOp,
+                                                                   GemmSpec,
+                                                                   1,
+                                                                   256,
+                                                                   256,
+                                                                   128,
+                                                                   32,
+                                                                   8,
+                                                                   8,
+                                                                   32,
+                                                                   32,
+                                                                   4,
+                                                                   2,
+                                                                   S<4, 64, 1>,
+                                                                   S<1, 0, 2>,
+                                                                   S<1, 0, 2>,
+                                                                   2,
+                                                                   8,
+                                                                   8,
+                                                                   1,
+                                                                   S<4, 64, 1>,
+                                                                   S<1, 0, 2>,
+                                                                   S<1, 0, 2>,
+                                                                   2,
+                                                                   8,
+                                                                   8,
+                                                                   1,
+                                                                   1,
+                                                                   1,
+                                                                   S<1, 32, 1, 8>,
+                                                                   8>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideD = 4096;
+    ck::index_t StrideE = 4096;
+
+    float alpha = 1.0f;
+    float beta  = 1.0f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        alpha = std::stof(argv[4]);
+        beta  = std::stof(argv[5]);
+    }
+    else if(argc == 13)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[10]);
+
+        alpha = std::stof(argv[11]);
+        beta  = std::stof(argv[12]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, alpha, "
+               "beta\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DLayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d_device_buf.ToDevice(d_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{alpha, beta};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               std::array<ck::index_t, 1>{StrideD},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/03_gemm_bias_relu/CMakeLists.txt b/example/03_gemm_bias_relu/CMakeLists.txt
new file mode 100644
index 0000000000..a247a052cb
--- /dev/null
+++ b/example/03_gemm_bias_relu/CMakeLists.txt
@@ -0,0 +1,10 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+    add_example_executable(example_gemm_bias_relu_xdl_fp16 gemm_bias_relu_xdl_fp16.cpp)
+   set(target 1)
+ endif()
+endforeach()
+endif()
diff --git a/example/03_gemm_bias_relu/README.md b/example/03_gemm_bias_relu/README.md
new file mode 100644
index 0000000000..f28a9a071c
--- /dev/null
+++ b/example/03_gemm_bias_relu/README.md
@@ -0,0 +1,10 @@
+# Instructions for ```example_gemm_bias_relu_xdl_fp16```
+
+## Run ```example_gemm_bias_relu_xdl_fp16```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE
+./bin/example_gemm_bias_relu_xdl_fp16 1 1 1 3840 4096 4096 4096 4096 4096
+```
diff --git a/example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp b/example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp
new file mode 100644
index 0000000000..dffeff2337
--- /dev/null
+++ b/example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp
@@ -0,0 +1,283 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// C = A * B
+// E = Relu(C + D);
+struct AddRelu
+{
+    __host__ __device__ void
+    operator()(ck::half_t& e, const ck::half_t& c, const ck::half_t& d) const
+    {
+        const ck::half_t x = c + d;
+
+        e = x > 0 ? x : 0;
+    }
+};
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using EDataType        = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using DLayout = Row;
+using ELayout = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddRelu;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceOpInstance =
+    ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<ALayout,
+                                                                   BLayout,
+                                                                   ck::Tuple<DLayout>,
+                                                                   ELayout,
+                                                                   ADataType,
+                                                                   BDataType,
+                                                                   AccDataType,
+                                                                   CShuffleDataType,
+                                                                   ck::Tuple<DDataType>,
+                                                                   EDataType,
+                                                                   AElementOp,
+                                                                   BElementOp,
+                                                                   CDEElementOp,
+                                                                   GemmSpec,
+                                                                   1,
+                                                                   256,
+                                                                   256,
+                                                                   128,
+                                                                   32,
+                                                                   8,
+                                                                   8,
+                                                                   32,
+                                                                   32,
+                                                                   4,
+                                                                   2,
+                                                                   S<4, 64, 1>,
+                                                                   S<1, 0, 2>,
+                                                                   S<1, 0, 2>,
+                                                                   2,
+                                                                   8,
+                                                                   8,
+                                                                   1,
+                                                                   S<4, 64, 1>,
+                                                                   S<1, 0, 2>,
+                                                                   S<1, 0, 2>,
+                                                                   2,
+                                                                   8,
+                                                                   8,
+                                                                   1,
+                                                                   1,
+                                                                   1,
+                                                                   S<1, 32, 1, 8>,
+                                                                   8>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideE = 4096;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, 0, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d_device_buf.ToDevice(d_m_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+
+    auto invoker = device_op.MakeInvoker();
+
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               std::array<ck::index_t, 1>{0},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(EDataType) * M * N + sizeof(EDataType) * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    if(do_verification)
+    {
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        Tensor<AccDataType> c_m_n(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
+            }
+        }
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/04_gemm_add_add_fastgelu/CMakeLists.txt b/example/04_gemm_add_add_fastgelu/CMakeLists.txt
new file mode 100644
index 0000000000..15ec62c89f
--- /dev/null
+++ b/example/04_gemm_add_add_fastgelu/CMakeLists.txt
@@ -0,0 +1,28 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+    add_custom_target(example_gemm_add_add_fastgelu_xdl)
+    if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_gemm_add_add_fastgelu_xdl_bf16 gemm_add_add_fastgelu_xdl_bf16.cpp)
+      add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_bf16)
+    endif()
+    if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_gemm_add_add_fastgelu_xdl_fp16 gemm_add_add_fastgelu_xdl_fp16.cpp)
+      add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp16)
+    endif()
+    if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+      add_example_executable(example_gemm_add_add_fastgelu_xdl_fp32 gemm_add_add_fastgelu_xdl_fp32.cpp)
+      add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_fp32)
+    endif()
+    if(USE_BITINT_EXTENSION_INT4)
+       add_example_executable(example_gemm_add_add_fastgelu_xdl_int4 gemm_add_add_fastgelu_xdl_int4.cpp)
+       add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int4)
+    endif(USE_BITINT_EXTENSION_INT4)
+    if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_gemm_add_add_fastgelu_xdl_int8 gemm_add_add_fastgelu_xdl_int8.cpp)
+      add_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int8)
+    endif()
+   set(target 1)
+ endif()
+endforeach()
\ No newline at end of file
diff --git a/example/04_gemm_add_add_fastgelu/README.md b/example/04_gemm_add_add_fastgelu/README.md
new file mode 100644
index 0000000000..08a55fb9a3
--- /dev/null
+++ b/example/04_gemm_add_add_fastgelu/README.md
@@ -0,0 +1,23 @@
+# Instructions for ```example_gemm_add_add_fastgelu_xdl_fp16```
+
+## Run ```example_gemm_add_add_fastgelu_xdl_fp16```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, StrideE"
+./bin/example_gemm_add_add_fastgelu_xdl_fp16 1 1 1
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+d0_m_n: dim 2, lengths {3840, 4096}, strides {0, 1}
+d1_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+e_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 1.26914 ms, 101.525 TFlops, 100.804 GB/s, DeviceGemmMultipleD_Xdl_CShuffle<256, 256, 128, 32, 8, 8>
+```
diff --git a/example/04_gemm_add_add_fastgelu/common.hpp b/example/04_gemm_add_add_fastgelu/common.hpp
new file mode 100644
index 0000000000..91d17df95f
--- /dev/null
+++ b/example/04_gemm_add_add_fastgelu/common.hpp
@@ -0,0 +1,106 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <cstddef>
+#include <iostream>
+#include <stdexcept>
+#include <string>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+using BF16 = ck::bhalf_t;
+using F16  = ck::half_t;
+using F32  = float;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+using I4 = ck::int4_t;
+#endif
+using I8  = int8_t;
+using I32 = int32_t;
+
+struct ProblemSize final
+{
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 4096;
+    ck::index_t StrideE  = 4096;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+inline bool
+parse_cmd_args(int argc, char* argv[], ProblemSize& problem_size, ExecutionConfig& config)
+{
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 12)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        problem_size.M = std::stoi(argv[4]);
+        problem_size.N = std::stoi(argv[5]);
+        problem_size.K = std::stoi(argv[6]);
+
+        problem_size.StrideA  = std::stoi(argv[7]);
+        problem_size.StrideB  = std::stoi(argv[8]);
+        problem_size.StrideD0 = std::stoi(argv[9]);
+        problem_size.StrideD1 = std::stoi(argv[10]);
+        problem_size.StrideE  = std::stoi(argv[11]);
+    }
+    else
+    {
+        std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
+                  << std::endl
+                  << "arg3: time kernel (0=no, 1=yes)" << std::endl
+                  << "arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, "
+                     "StrideE"
+                  << std::endl;
+        return false;
+    }
+
+    return true;
+}
diff --git a/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp
new file mode 100644
index 0000000000..e630f67837
--- /dev/null
+++ b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp
@@ -0,0 +1,48 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType        = BF16;
+using BDataType        = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType  = F32; // C matrix doesn't exsit in GPU memory, this is used for host verification
+using D0DataType = BF16;
+using D1DataType = BF16;
+using DsDataType = ck::Tuple<D0DataType, D1DataType>;
+using EDataType  = BF16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_add_add_fastgelu_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
diff --git a/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp
new file mode 100644
index 0000000000..71f6677bae
--- /dev/null
+++ b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp
@@ -0,0 +1,48 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType  = F32; // C matrix doesn't exsit in GPU memory, this is used for host verification
+using D0DataType = F16;
+using D1DataType = F16;
+using DsDataType = ck::Tuple<D0DataType, D1DataType>;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_add_add_fastgelu_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
diff --git a/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp
new file mode 100644
index 0000000000..4665c3932f
--- /dev/null
+++ b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp
@@ -0,0 +1,47 @@
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType  = F32; // C matrix doesn't exsit in GPU memory, this is used for host verification
+using D0DataType = F32;
+using D1DataType = F32;
+using DsDataType = ck::Tuple<D0DataType, D1DataType>;
+using EDataType  = F32;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 32, 1, 8>,               4>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_add_add_fastgelu_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
diff --git a/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp
new file mode 100644
index 0000000000..f206bbeb41
--- /dev/null
+++ b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+#error Should compile this file with ck::int4_t support
+#endif
+
+#include "common.hpp"
+
+using ADataType        = I4;
+using BDataType        = I4;
+using AccDataType      = I32;
+using CShuffleDataType = I32;
+using CDataType  = I32; // C matrix doesn't exsit in GPU memory, this is used for host verification
+using D0DataType = I4;
+using D1DataType = I4;
+using DsDataType = ck::Tuple<D0DataType, D1DataType>;
+using EDataType  = I4;
+
+using KernelADataType  = I8;
+using KernelBDataType  = I8;
+using KernelD0DataType = I8;
+using KernelD1DataType = I8;
+using KernelDsDataType = ck::Tuple<KernelD0DataType, KernelD1DataType>;
+using KernelEDataType  = I8;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle
+//######| ALayout| BLayout| DsLayout| ELayout|           AData|           BData|     AccData|         CShuffle|           DsData|           EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |            Type|            Type|        Type|         DataType|             Type|            Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |                |                |            |                 |                 |                |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |                |                |            |                 |                 |                |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, KernelADataType, KernelBDataType, AccDataType, CShuffleDataType, KernelDsDataType, KernelEDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 64, 1, 4>,              16>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#define BUILD_INT4_EXAMPLE
+#include "run_gemm_add_add_fastgelu_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
diff --git a/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp
new file mode 100644
index 0000000000..e46483ab38
--- /dev/null
+++ b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp
@@ -0,0 +1,48 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType        = I8;
+using BDataType        = I8;
+using AccDataType      = I32;
+using CShuffleDataType = I32;
+using CDataType  = I32; // C matrix doesn't exsit in GPU memory, this is used for host verification
+using D0DataType = I8;
+using D1DataType = I8;
+using DsDataType = ck::Tuple<D0DataType, D1DataType>;
+using EDataType  = I8;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAddFastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 64, 1, 4>,              16>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_add_add_fastgelu_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_add_fastgelu_example(argc, argv); }
diff --git a/example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc b/example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc
new file mode 100644
index 0000000000..cb3147bcd7
--- /dev/null
+++ b/example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc
@@ -0,0 +1,166 @@
+#pragma once
+
+bool run_gemm_add_add_fastgelu(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+#endif
+    using namespace ck::literals;
+
+    auto& [M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE] = problem_size;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<
+#ifdef BUILD_INT4_EXAMPLE
+        KernelEDataType
+#else
+        EDataType
+#endif
+        >
+        e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<KernelADataType> a_m_k_converted(a_m_k);
+    const Tensor<KernelBDataType> b_k_n_converted(b_k_n);
+    const Tensor<KernelD0DataType> d0_m_n_converted(d0_m_n);
+    const Tensor<KernelD1DataType> d1_m_n_converted(d1_m_n);
+
+    a_device_buf.ToDevice(a_m_k_converted.mData.data());
+    b_device_buf.ToDevice(b_k_n_converted.mData.data());
+    d0_device_buf.ToDevice(d0_m_n_converted.mData.data());
+    d1_device_buf.ToDevice(d1_m_n_converted.mData.data());
+#else
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+#endif
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               {StrideD0, StrideD1},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = 2_uz * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(D0DataType) * N + sizeof(D1DataType) * M * N +
+                            sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << device_op.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        Tensor<CDataType> c_m_n({M, N});
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+#ifdef BUILD_INT4_EXAMPLE
+        const Tensor<EDataType> e_m_n_device_result_converted(e_m_n_device_result);
+
+        return ck::utils::check_err(e_m_n_device_result_converted, e_m_n_host_result);
+#else
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+#endif
+    }
+
+    return true;
+}
+
+bool run_gemm_add_add_fastgelu_example(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    return !parse_cmd_args(argc, argv, problem_size, config) ||
+           run_gemm_add_add_fastgelu(problem_size, config);
+}
diff --git a/example/09_convnd_fwd/CMakeLists.txt b/example/09_convnd_fwd/CMakeLists.txt
new file mode 100644
index 0000000000..1af1e6c858
--- /dev/null
+++ b/example/09_convnd_fwd/CMakeLists.txt
@@ -0,0 +1,35 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+    if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+      add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp)
+    endif()
+    if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
+    endif()
+    if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp)
+    endif()
+    if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
+    endif()
+    # FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
+    if(DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
+      add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
+    endif()
+   set(target 1)
+ endif()
+endforeach()
+
+if(DL_KERNELS)
+  if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_convnd_fwd_dl_fp16 convnd_fwd_dl_fp16.cpp)
+  endif()
+  if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_convnd_fwd_dl_fp32 convnd_fwd_dl_fp32.cpp)
+  endif()
+  if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+    add_example_executable(example_convnd_fwd_dl_int8 convnd_fwd_dl_int8.cpp)
+  endif()
+endif()
diff --git a/example/09_convnd_fwd/README.md b/example/09_convnd_fwd/README.md
new file mode 100644
index 0000000000..9ab5fee549
--- /dev/null
+++ b/example/09_convnd_fwd/README.md
@@ -0,0 +1,32 @@
+# Instructions for ```example_convnd_fwd_xdl```
+
+## Run ```example_convnd_fwd_xdl```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: run kernel # of times (>1)
+#arg4: N spatial dimensions (default 2)
+#Following arguments (depending on number of spatial dims):
+# N, K, C, 
+# <filter spatial dimensions>, (ie Y, X for 2D)
+# <input image spatial dimensions>, (ie Hi, Wi for 2D)
+# <strides>, (ie Sy, Sx for 2D)
+# <dilations>, (ie Dy, Dx for 2D)
+# <left padding>, (ie LeftPy, LeftPx for 2D)
+# <right padding>, (ie RightPy, RightPx for 2D)
+./bin/example_convnd_fwd_xdl 0 1 100
+```
+
+Result (MI100 @ 1087Mhz, 33.4TFlops peak FP32)
+```
+input: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
+weights: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
+output: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
+arg.a_grid_desc_k0_m_k1_{432, 165888, 4}
+arg.b_grid_desc_k0_n_k1_{432, 256, 4}
+arg.c_grid_desc_m_n_{ 165888, 256}
+launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
+Warm up
+Start running 100 times...
+Perf: 4.43736 ms, 33.0753 TFlops, 150.357 GB/s
+```
diff --git a/example/09_convnd_fwd/convnd_fwd_common.hpp b/example/09_convnd_fwd/convnd_fwd_common.hpp
new file mode 100644
index 0000000000..109b8f9ee3
--- /dev/null
+++ b/example/09_convnd_fwd/convnd_fwd_common.hpp
@@ -0,0 +1,172 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+void print_helper_msg()
+{
+    std::cout << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          int init_method,
+                          bool time_kernel,
+                          const ck::utils::conv::ConvParam& conv_param,
+                          const HostTensorDescriptor& in_g_n_c_wis_desc,
+                          const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op,
+                          const OutElementOp& out_element_op)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                      wei_device_buf.GetDeviceBuffer(),
+                                      std::array<const void*, 0>{},
+                                      out_device_buf.GetDeviceBuffer(),
+                                      a_g_n_c_wis_lengths,
+                                      a_g_n_c_wis_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                      e_g_n_k_wos_lengths,
+                                      e_g_n_k_wos_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      in_element_op,
+                                      wei_element_op,
+                                      out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  out_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        return ck::utils::check_err(
+            out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return true;
+}
diff --git a/example/09_convnd_fwd/convnd_fwd_dl_common.hpp b/example/09_convnd_fwd/convnd_fwd_dl_common.hpp
new file mode 100644
index 0000000000..aeddd4fc59
--- /dev/null
+++ b/example/09_convnd_fwd/convnd_fwd_dl_common.hpp
@@ -0,0 +1,196 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+void print_helper_msg()
+{
+    std::cout << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename DsDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd_dl(bool do_verification,
+                             int init_method,
+                             bool time_kernel,
+                             const ck::utils::conv::ConvParam& conv_param,
+                             const HostTensorDescriptor& in_g_n_c_wis_desc,
+                             const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                             const HostTensorDescriptor& out_g_n_k_wos_desc,
+                             const InElementOp& in_element_op,
+                             const WeiElementOp& wei_element_op,
+                             const OutElementOp& out_element_op)
+{
+    using DDataType = ck::remove_cvref_t<ck::tuple_element_t<0, DsDataType>>;
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<DDataType> bias(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 3});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 3});
+        bias.GenerateTensorValue(GeneratorTensor_2<DDataType>{-2, 3});
+        break;
+    case 2:
+        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+        bias.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
+        wei.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{-1});
+        bias.GenerateTensorValue(GeneratorTensor_1<DDataType>{1});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(DDataType) * bias.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    bias_device_buf.ToDevice(bias.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), d_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), d_g_n_k_wos_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(
+        in_device_buf.GetDeviceBuffer(),
+        wei_device_buf.GetDeviceBuffer(),
+        std::array<const void*, 1>{bias_device_buf.GetDeviceBuffer()},
+        out_device_buf.GetDeviceBuffer(),
+        a_g_n_c_wis_lengths,
+        a_g_n_c_wis_strides,
+        b_g_k_c_xs_lengths,
+        b_g_k_c_xs_strides,
+        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d_g_n_k_wos_lengths}},
+        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{{d_g_n_k_wos_strides}},
+        e_g_n_k_wos_lengths,
+        e_g_n_k_wos_strides,
+        conv_filter_strides,
+        conv_filter_dilations,
+        input_left_pads,
+        input_right_pads,
+        in_element_op,
+        wei_element_op,
+        out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        std::cout << "wrong! device_conv with the specified compilation parameters does not "
+                     "support this Conv problem"
+                  << std::endl;
+        return true;
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<
+            NDimSpatial,
+            InDataType,
+            WeiDataType,
+            OutDataType,
+            InElementOp,
+            WeiElementOp,
+            ck::tensor_operation::element_wise::PassThrough>();
+
+        auto ref_invoker = ref_conv.MakeInvoker();
+        auto ref_argument =
+            ref_conv.MakeArgument(in,
+                                  wei,
+                                  out_host,
+                                  conv_param.conv_filter_strides_,
+                                  conv_param.conv_filter_dilations_,
+                                  conv_param.input_left_pads_,
+                                  conv_param.input_right_pads_,
+                                  in_element_op,
+                                  wei_element_op,
+                                  ck::tensor_operation::element_wise::PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        // cde_elementwise
+        out_host.ForEach(
+            [&](auto&, auto idx) { out_element_op(out_host(idx), out_host(idx), bias(idx)); });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        return ck::utils::check_err(
+            out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return true;
+}
diff --git a/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp b/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp
new file mode 100644
index 0000000000..564fadcbf8
--- /dev/null
+++ b/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_dl_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using AccDataType = float;
+using DsDataType  = ck::Tuple<ck::half_t>;
+using OutDataType = ck::half_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+// clang-format off
+using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
+// ######|        NDim|     InData|     WeiData|   MultpleD|     OutData|     AccData| InLayout| WeiLayout|            MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|     Spatial|       Type|        Type|       Type|        Type|        Type|         |          |               Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|            |           |            |           |            |            |         |          |                     |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|            |           |            |           |            |            |         |          |                     |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < NDimSpatial, InDataType, WeiDataType, DsDataType, OutDataType, AccDataType, InLayout, WeiLayout, ck::Tuple<OutLayout>, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,       ConvSpec,    GemmPadingSpec,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>;
+// clang-format on
+
+#include "run_convnd_fwd_dl_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_dl_example(argc, argv) ? 0 : 1; }
diff --git a/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp b/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp
new file mode 100644
index 0000000000..0bd90c9992
--- /dev/null
+++ b/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_dl_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType  = float;
+using WeiDataType = float;
+using AccDataType = float;
+using DsDataType  = ck::Tuple<float>;
+using OutDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+// clang-format off
+using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
+// ######|        NDim|     InData|     WeiData|   MultpleD|     OutData|     AccData| InLayout| WeiLayout|            MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|     Spatial|       Type|        Type|       Type|        Type|        Type|         |          |               Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|            |           |            |           |            |            |         |          |                     |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|            |           |            |           |            |            |         |          |                     |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < NDimSpatial, InDataType, WeiDataType, DsDataType, OutDataType, AccDataType, InLayout, WeiLayout, ck::Tuple<OutLayout>, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,       ConvSpec,    GemmPadingSpec,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 4>;
+// clang-format on
+
+#include "run_convnd_fwd_dl_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_dl_example(argc, argv) ? 0 : 1; }
diff --git a/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp b/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp
new file mode 100644
index 0000000000..9b0c8b31f1
--- /dev/null
+++ b/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_dl_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType  = int8_t;
+using WeiDataType = int8_t;
+using AccDataType = int32_t;
+using DsDataType  = ck::Tuple<int8_t>;
+using OutDataType = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+// clang-format off
+using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
+// ######|        NDim|     InData|     WeiData|   MultpleD|     OutData|     AccData| InLayout| WeiLayout|            MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+// ######|     Spatial|       Type|        Type|       Type|        Type|        Type|         |          |               Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+// ######|            |           |            |           |            |            |         |          |                     |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+// ######|            |           |            |           |            |            |         |          |                     |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+         < NDimSpatial, InDataType, WeiDataType, DsDataType, OutDataType, AccDataType, InLayout, WeiLayout, ck::Tuple<OutLayout>, OutLayout,  InElementOp,  WeiElementOp,  OutElementOp,       ConvSpec,    GemmPadingSpec,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5,                 4>;
+// clang-format on
+
+#include "run_convnd_fwd_dl_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_dl_example(argc, argv) ? 0 : 1; }
diff --git a/example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp b/example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp
new file mode 100644
index 0000000000..74cf91d160
--- /dev/null
+++ b/example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType       = ck::bhalf_t;
+using WeiDataType      = ck::bhalf_t;
+using AccDataType      = float;
+using CShuffleDataType = float;
+using OutDataType      = ck::bhalf_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        32,          // KPerBlock
+        8,           // AK1
+        8,           // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        8,           // ABlockTransferSrcScalarPerVector
+        8,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        8,           // BBlockTransferSrcScalarPerVector
+        8,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 32, 1, 8>,
+        8>;
+
+#include "run_convnd_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp b/example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
new file mode 100644
index 0000000000..f6d69bafd4
--- /dev/null
+++ b/example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType       = ck::half_t;
+using WeiDataType      = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::half_t;
+using OutDataType      = ck::half_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::UnaryConvert;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        32,          // KPerBlock
+        8,           // AK1
+        8,           // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        8,           // ABlockTransferSrcScalarPerVector
+        8,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        8,           // BBlockTransferSrcScalarPerVector
+        8,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 32, 1, 8>,
+        8>;
+
+#include "run_convnd_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp b/example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp
new file mode 100644
index 0000000000..6c3171f615
--- /dev/null
+++ b/example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType       = float;
+using WeiDataType      = float;
+using AccDataType      = float;
+using CShuffleDataType = float;
+using OutDataType      = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        16,          // KPerBlock
+        4,           // AK1
+        4,           // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        4,           // ABlockTransferSrcScalarPerVector
+        4,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        4,           // BBlockTransferSrcScalarPerVector
+        4,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 16, 1, 16>,
+        4>;
+
+#include "run_convnd_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp b/example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp
new file mode 100644
index 0000000000..9977a496d2
--- /dev/null
+++ b/example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType       = double;
+using WeiDataType      = double;
+using AccDataType      = double;
+using CShuffleDataType = double;
+using OutDataType      = double;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        128,         // NPerBlock
+        8,           // KPerBlock
+        2,           // AK1
+        2,           // BK1
+        16,          // MPerXdl
+        16,          // NPerXdl
+        4,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        2,           // ABlockTransferSrcScalarPerVector
+        2,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        2,           // BBlockTransferSrcScalarPerVector
+        2,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 16, 1, 16>,
+        1>;
+
+#include "run_convnd_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp b/example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
new file mode 100644
index 0000000000..bf084b3cc0
--- /dev/null
+++ b/example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_fwd_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+using InDataType       = int8_t;
+using WeiDataType      = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int8_t;
+using OutDataType      = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        64,          // KPerBlock
+        16,          // AK1
+        16,          // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        16,          // ABlockTransferSrcScalarPerVector
+        16,          // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        16,          // BBlockTransferSrcScalarPerVector
+        16,          // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 64, 1, 4>,
+        16>;
+
+#include "run_convnd_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc b/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc
new file mode 100644
index 0000000000..6474df1c35
--- /dev/null
+++ b/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc
@@ -0,0 +1,98 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+bool run_convnd_fwd_dl_example(int argc, char* argv[])
+{
+    print_helper_msg();
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::utils::conv::ConvParam conv_param{
+        2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        do_verification                   = std::stoi(argv[1]);
+        init_method                       = std::stoi(argv[2]);
+        time_kernel                       = std::stoi(argv[3]);
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+
+        conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
+    }
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    const auto run = [&](auto ndim_spatial, auto in_layout, auto wei_layout, auto out_layout) {
+        constexpr ck::index_t ndim_spatial_value = ndim_spatial.value;
+        std::cout << "ndim_spatial_value: " << ndim_spatial_value << std::endl;
+
+        using InLayout  = decltype(in_layout);
+        using WeiLayout = decltype(wei_layout);
+        using OutLayout = decltype(out_layout);
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_grouped_conv_fwd_dl<
+            ndim_spatial_value,
+            InDataType,
+            WeiDataType,
+            DsDataType,
+            OutDataType,
+            InElementOp,
+            WeiElementOp,
+            OutElementOp,
+            DeviceGroupedConvNDFwdInstance<ndim_spatial_value, InLayout, WeiLayout, OutLayout>>(
+            do_verification,
+            init_method,
+            time_kernel,
+            conv_param,
+            in_g_n_c_wis_desc,
+            wei_g_k_c_xs_desc,
+            out_g_n_k_wos_desc,
+            in_element_op,
+            wei_element_op,
+            out_element_op);
+    };
+
+    namespace ctc = ck::tensor_layout::convolution;
+
+    if(conv_param.num_dim_spatial_ == 1)
+    {
+        return run(ck::Number<1>{}, ctc::GNWC{}, ctc::GKXC{}, ctc::GNWK{});
+    }
+    else if(conv_param.num_dim_spatial_ == 2)
+    {
+        return run(ck::Number<2>{}, ctc::GNHWC{}, ctc::GKYXC{}, ctc::GNHWK{});
+    }
+    else if(conv_param.num_dim_spatial_ == 3)
+    {
+        return run(ck::Number<3>{}, ctc::GNDHWC{}, ctc::GKZYXC{}, ctc::GNDHWK{});
+    }
+
+    return true;
+}
diff --git a/example/09_convnd_fwd/run_convnd_fwd_example.inc b/example/09_convnd_fwd/run_convnd_fwd_example.inc
new file mode 100644
index 0000000000..49852ff667
--- /dev/null
+++ b/example/09_convnd_fwd/run_convnd_fwd_example.inc
@@ -0,0 +1,97 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+bool run_convnd_fwd_example(int argc, char* argv[])
+{
+    print_helper_msg();
+
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::utils::conv::ConvParam conv_param{
+        2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        do_verification                   = std::stoi(argv[1]);
+        init_method                       = std::stoi(argv[2]);
+        time_kernel                       = std::stoi(argv[3]);
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+
+        conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
+    }
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    const auto run = [&](auto ndim_spatial, auto in_layout, auto wei_layout, auto out_layout) {
+        constexpr ck::index_t ndim_spatial_value = ndim_spatial.value;
+
+        using InLayout  = decltype(in_layout);
+        using WeiLayout = decltype(wei_layout);
+        using OutLayout = decltype(out_layout);
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_grouped_conv_fwd<
+            ndim_spatial_value,
+            InDataType,
+            WeiDataType,
+            OutDataType,
+            InElementOp,
+            WeiElementOp,
+            OutElementOp,
+            DeviceGroupedConvNDFwdInstance<ndim_spatial_value, InLayout, WeiLayout, OutLayout>>(
+            do_verification,
+            init_method,
+            time_kernel,
+            conv_param,
+            in_g_n_c_wis_desc,
+            wei_g_k_c_xs_desc,
+            out_g_n_k_wos_desc,
+            in_element_op,
+            wei_element_op,
+            out_element_op);
+    };
+
+    namespace ctc = ck::tensor_layout::convolution;
+
+    if(conv_param.num_dim_spatial_ == 1)
+    {
+        return run(ck::Number<1>{}, ctc::GNWC{}, ctc::GKXC{}, ctc::GNWK{});
+    }
+    else if(conv_param.num_dim_spatial_ == 2)
+    {
+        return run(ck::Number<2>{}, ctc::GNHWC{}, ctc::GKYXC{}, ctc::GNHWK{});
+    }
+    else if(conv_param.num_dim_spatial_ == 3)
+    {
+        return run(ck::Number<3>{}, ctc::GNDHWC{}, ctc::GKZYXC{}, ctc::GNDHWK{});
+    }
+
+    return true;
+}
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt b/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
new file mode 100644
index 0000000000..e7d941ae6b
--- /dev/null
+++ b/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
@@ -0,0 +1,28 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_custom_target(example_convnd_fwd_reduce_xdl)
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+    add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
+    add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp)
+    add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_bf16)
+   endif()
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable_no_testing(example_convnd_fwd_max_xdl_fp16 convnd_fwd_max_xdl_fp16.cpp)
+    add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp16)
+   endif()
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_convnd_fwd_max_xdl_fp32 convnd_fwd_max_xdl_fp32.cpp)
+    add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp32)
+   endif()
+   if(USE_BITINT_EXTENSION_INT4)
+      add_example_executable(example_convnd_fwd_max_xdl_int4 convnd_fwd_max_xdl_int4.cpp)
+      add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int4)
+   endif(USE_BITINT_EXTENSION_INT4)
+   set(target 1)
+ endif()
+endforeach()
\ No newline at end of file
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp
new file mode 100644
index 0000000000..137b0d1ff0
--- /dev/null
+++ b/example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp
@@ -0,0 +1,159 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cassert>
+#include <cstdint>
+#include <cstdlib>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <type_traits>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using BF16 = ck::bhalf_t;
+using FP16 = ck::half_t;
+using FP32 = float;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+using I4 = ck::int4_t;
+#endif
+using I8  = std::int8_t;
+using I32 = std::int32_t;
+
+template <typename ALay, typename BLay, typename DELay, typename RLay>
+struct LayoutSetting
+{
+    using ALayout  = ALay;
+    using BLayout  = BLay;
+    using DELayout = DELay;
+    using RLayout  = RLay;
+};
+
+template <ck::index_t NDimSpatial>
+struct LayoutSettingSelector;
+
+namespace ctl = ck::tensor_layout::convolution;
+
+template <>
+struct LayoutSettingSelector<1> final : LayoutSetting<ctl::GNWC, ctl::GKXC, ctl::GNWK, ctl::GNW>
+{
+};
+
+template <>
+struct LayoutSettingSelector<2> final : LayoutSetting<ctl::GNHWC, ctl::GKYXC, ctl::GNHWK, ctl::GNHW>
+{
+};
+
+template <>
+struct LayoutSettingSelector<3> final
+    : LayoutSetting<ctl::GNDHWC, ctl::GKZYXC, ctl::GNDHWK, ctl::GNDHW>
+{
+};
+
+template <ck::index_t NDimSpatial>
+using ALayout = typename LayoutSettingSelector<NDimSpatial>::ALayout;
+
+template <ck::index_t NDimSpatial>
+using BLayout = typename LayoutSettingSelector<NDimSpatial>::BLayout;
+
+template <ck::index_t NDimSpatial>
+using DELayout = typename LayoutSettingSelector<NDimSpatial>::DELayout;
+
+template <ck::index_t NDimSpatial>
+using RLayout = typename LayoutSettingSelector<NDimSpatial>::RLayout;
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+inline void print_help_msg()
+{
+    std::cerr << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+inline bool parse_cmd_args(int argc,
+                           char* argv[],
+                           ck::utils::conv::ConvParam& problem_size,
+                           ExecutionConfig& config)
+{
+    constexpr int num_execution_config_args =
+        3; // arguments for do_verification, init_method, time_kernel
+    constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
+
+    constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
+    constexpr int threshold_to_catch_all_args =
+        threshold_to_catch_partial_args + num_conv_param_leading_args;
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    // catch only ExecutionConfig arguments
+    else if(argc == threshold_to_catch_partial_args)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    // catch both ExecutionConfig & ConvParam arguments
+    else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+        problem_size                      = ck::utils::conv::parse_conv_param(
+            num_dim_spatial, threshold_to_catch_partial_args, argv);
+    }
+    else
+    {
+        print_help_msg();
+        return false;
+    }
+
+    return true;
+}
+
+inline HostTensorDescriptor
+make_r0_host_tensor_descriptor(const ck::utils::conv::ConvParam& problem_size)
+{
+    std::vector<ck::index_t> dimensions{problem_size.G_, problem_size.N_};
+
+    ck::ranges::copy(problem_size.output_spatial_lengths_, std::back_inserter(dimensions));
+
+    return HostTensorDescriptor(dimensions);
+}
+
+template <typename Lengths, typename Strides>
+void unpack_host_tensor_descriptor(const HostTensorDescriptor& descriptor,
+                                   Lengths& lengths,
+                                   Strides& strides)
+{
+    assert(size(descriptor.GetLengths()) == size(lengths));
+    std::copy_n(begin(descriptor.GetLengths()), size(descriptor.GetLengths()), begin(lengths));
+
+    assert(size(descriptor.GetStrides()) == size(strides));
+    std::copy_n(begin(descriptor.GetStrides()), size(descriptor.GetStrides()), begin(strides));
+}
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp
new file mode 100644
index 0000000000..4ccacb0bce
--- /dev/null
+++ b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp
@@ -0,0 +1,18 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType         = BF16;
+using BDataType         = BF16;
+using AccDataType       = FP32;
+using CShuffleDataType  = FP32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = BF16;
+using ReduceAccDataType = FP32;
+using R0DataType        = FP32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+#include "run_convnd_fwd_max_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp
new file mode 100644
index 0000000000..bf495725e8
--- /dev/null
+++ b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp
@@ -0,0 +1,18 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType         = FP16;
+using BDataType         = FP16;
+using AccDataType       = FP32;
+using CShuffleDataType  = FP32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = FP16;
+using ReduceAccDataType = FP32;
+using R0DataType        = FP32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+#include "run_convnd_fwd_max_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp
new file mode 100644
index 0000000000..5848785673
--- /dev/null
+++ b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp
@@ -0,0 +1,18 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType         = FP32;
+using BDataType         = FP32;
+using AccDataType       = FP32;
+using CShuffleDataType  = FP32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = FP32;
+using ReduceAccDataType = FP32;
+using R0DataType        = FP32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+#include "run_convnd_fwd_max_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp
new file mode 100644
index 0000000000..bf7127502f
--- /dev/null
+++ b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+#error Should compile this file with ck::int4_t support
+#endif
+
+#define BUILD_INT4_EXAMPLE
+
+#include "common.hpp"
+
+using ADataType         = I4;
+using BDataType         = I4;
+using KernelADataType   = I8;
+using KernelBDataType   = I8;
+using AccDataType       = I32;
+using CShuffleDataType  = I32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = I32;
+using ReduceAccDataType = I32;
+using R0DataType        = I32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+#include "run_convnd_fwd_max_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp
new file mode 100644
index 0000000000..3e1694cbe8
--- /dev/null
+++ b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp
@@ -0,0 +1,18 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using ADataType         = I8;
+using BDataType         = I8;
+using AccDataType       = I32;
+using CShuffleDataType  = I32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = I32;
+using ReduceAccDataType = I32;
+using R0DataType        = I32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+#include "run_convnd_fwd_max_example.inc"
+
+int main(int argc, char* argv[]) { return !run_convnd_fwd_max_example(argc, argv); }
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc b/example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc
new file mode 100644
index 0000000000..cebfeb51d6
--- /dev/null
+++ b/example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc
@@ -0,0 +1,307 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough>;
+using RsElementOp  = ck::Tuple<PassThrough>;
+
+// ReduceOp
+using RsThreadReduceOp = ck::Tuple<ck::reduce::Max>;
+
+using RsGlobalReduceOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+template <ck::index_t NDimSpatial>
+using DeviceInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
+//######| NDimSpatial|              ALayout|              BLayout|              DELayout|              RLayout|           AData|           BData|     AccData|         CShuffle|     DsData|     EData|     ReduceAccData|     RsData|           A|           B|          CDE|          Qs|          Rs|           Thread|           Global|           Conv|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|    CDRThreadTransfer|                  CDE|    RThreadTransfer|
+//######|            |                     |                     |                      |                     |            Type|            Type|        Type|         DataType|       Type|      Type|              Type|       Type| Elementwise| Elementwise|  Elementwise| Elementwise| Elementwise|           Reduce|           Reduce|            Fwd|Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|       ClusterLengths| ReduceThreadTransfer| DstScalarPerVector|
+//######|            |                     |                     |                      |                     |                |                |            |                 |           |          |                  |           |   Operation|   Operation|    Operation|   Operation|   Operation|        Operation|        Operation| Specialization|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _MPerBlock_NPerBlock|      ScalarPerVector|         _MPerBlock|
+//######|            |                     |                     |                      |                     |                |                |            |                 |           |          |                  |           |            |            |             |            |            |                 |                 |               |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |           _NPerBlock|                   |
+#ifdef BUILD_INT4_EXAMPLE
+        < NDimSpatial, ALayout<NDimSpatial>, BLayout<NDimSpatial>, DELayout<NDimSpatial>, RLayout<NDimSpatial>, KernelADataType, KernelBDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType,  AElementOp,  BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp,       ConvSpec,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<64, 4>,                    4,                  1>;
+#else
+        < NDimSpatial, ALayout<NDimSpatial>, BLayout<NDimSpatial>, DELayout<NDimSpatial>, RLayout<NDimSpatial>,       ADataType,       BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType,  AElementOp,  BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp,       ConvSpec,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<64, 4>,                    4,                  1>;
+#endif
+
+template <ck::index_t NDimSpatial>
+using HostInstance = ck::tensor_operation::host::ReferenceConvFwd
+        <NDimSpatial, ADataType, BDataType, EDataType, AElementOp, BElementOp, PassThrough>;
+// clang-format on
+
+template <ck::index_t NDimSpatial>
+bool run_convnd_fwd_max(const ck::utils::conv::ConvParam& problem_size,
+                        const ExecutionConfig& config)
+{
+    static_assert(1 <= NDimSpatial && NDimSpatial <= 3, "Unsupported NDimSpatial");
+
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+#endif
+
+    const auto conv_input_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<ALayout<NDimSpatial>>(
+            problem_size);
+
+    const auto conv_weight_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<BLayout<NDimSpatial>>(
+            problem_size);
+
+    const auto conv_output_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<DELayout<NDimSpatial>>(
+            problem_size);
+
+    const auto r0_desc = make_r0_host_tensor_descriptor(problem_size);
+
+    Tensor<ADataType> conv_input(conv_input_g_n_c_wis_desc);
+    Tensor<BDataType> conv_weight(conv_weight_g_k_c_xs_desc);
+    Tensor<EDataType> conv_output_device(conv_output_g_n_k_wos_desc);
+    Tensor<R0DataType> r0_device(r0_desc);
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-8, 7}(conv_input);
+        ck::utils::FillUniformDistributionIntegerValue<BDataType>{-8, 7}(conv_weight);
+        break;
+    default:
+        ck::utils::FillUniformDistribution<ADataType>{-5, 5}(conv_input);
+        ck::utils::FillUniformDistribution<BDataType>{-5, 5}(conv_weight);
+    }
+
+    DeviceMem conv_input_device_buf(sizeof(ADataType) * conv_input.mDesc.GetElementSpaceSize());
+    DeviceMem conv_weight_device_buf(sizeof(BDataType) * conv_weight.mDesc.GetElementSpaceSize());
+    DeviceMem conv_output_device_buf(sizeof(EDataType) *
+                                     conv_output_device.mDesc.GetElementSpaceSize());
+    DeviceMem r0_device_buf(sizeof(R0DataType) * r0_device.mDesc.GetElementSpaceSize());
+
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<KernelADataType> conv_input_converted(conv_input);
+    const Tensor<KernelBDataType> conv_weight_converted(conv_weight);
+
+    conv_input_device_buf.ToDevice(conv_input_converted.mData.data());
+    conv_weight_device_buf.ToDevice(conv_weight_converted.mData.data());
+#else
+    conv_input_device_buf.ToDevice(conv_input.mData.data());
+    conv_weight_device_buf.ToDevice(conv_weight.mData.data());
+#endif
+
+    std::array<ck::index_t, NDimSpatial + 3> conv_input_g_n_c_wis_lengths{},
+        conv_input_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> conv_weight_g_k_c_xs_lengths{},
+        conv_weight_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> conv_output_g_n_k_wos_lengths{},
+        conv_output_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 2> r0_lengths{}, r0_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{}, conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{}, input_right_pads{};
+
+    unpack_host_tensor_descriptor(
+        conv_input_g_n_c_wis_desc, conv_input_g_n_c_wis_lengths, conv_input_g_n_c_wis_strides);
+    unpack_host_tensor_descriptor(
+        conv_weight_g_k_c_xs_desc, conv_weight_g_k_c_xs_lengths, conv_weight_g_k_c_xs_strides);
+    unpack_host_tensor_descriptor(
+        conv_output_g_n_k_wos_desc, conv_output_g_n_k_wos_lengths, conv_output_g_n_k_wos_strides);
+    unpack_host_tensor_descriptor(r0_desc, r0_lengths, r0_strides);
+
+    ck::ranges::copy(problem_size.conv_filter_strides_, begin(conv_filter_strides));
+    ck::ranges::copy(problem_size.conv_filter_dilations_, begin(conv_filter_dilations));
+    ck::ranges::copy(problem_size.input_left_pads_, begin(input_left_pads));
+    ck::ranges::copy(problem_size.input_right_pads_, begin(input_right_pads));
+
+    // run Conv + Reduction on device
+    auto conv     = DeviceInstance<NDimSpatial>{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(conv_input_device_buf.GetDeviceBuffer(),
+                                      conv_weight_device_buf.GetDeviceBuffer(),
+                                      std::array<const void*, 0>{},
+                                      conv_output_device_buf.GetDeviceBuffer(),
+                                      {r0_device_buf.GetDeviceBuffer()},
+                                      conv_input_g_n_c_wis_lengths,
+                                      conv_input_g_n_c_wis_strides,
+                                      conv_weight_g_k_c_xs_lengths,
+                                      conv_weight_g_k_c_xs_strides,
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
+                                      conv_output_g_n_k_wos_lengths,
+                                      conv_output_g_n_k_wos_strides,
+                                      r0_lengths,
+                                      r0_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      AElementOp{},
+                                      BElementOp{},
+                                      CDEElementOp{},
+                                      QsElementOp{},
+                                      RsElementOp{});
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        std::cerr << "wrong! device_conv with the specified compilation parameters does "
+                     "not support this Conv problem"
+                  << std::endl;
+        return false;
+    }
+
+    const float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    const std::size_t flop      = problem_size.GetFlops();
+    const std::size_t num_btype = problem_size.GetByte<ADataType, BDataType, EDataType>();
+
+    const float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    const float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        Tensor<EDataType> conv_output_host(conv_output_g_n_k_wos_desc);
+
+        // run Conv + Reduction on host
+        auto ref_conv     = HostInstance<NDimSpatial>{};
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(conv_input,
+                                                  conv_weight,
+                                                  conv_output_host,
+                                                  problem_size.conv_filter_strides_,
+                                                  problem_size.conv_filter_dilations_,
+                                                  problem_size.input_left_pads_,
+                                                  problem_size.input_right_pads_,
+                                                  AElementOp{},
+                                                  BElementOp{},
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        Tensor<R0DataType> r0_host(r0_device.mDesc);
+
+        auto reduce0_op = RsThreadReduceOp{}[ck::Number<0>{}];
+
+        auto& output_dims = conv_output_g_n_k_wos_desc.GetLengths();
+
+        if constexpr(NDimSpatial == 1)
+        {
+            for(std::size_t g = 0; g < output_dims[0]; ++g)
+            {
+                for(std::size_t n = 0; n < output_dims[1]; ++n)
+                {
+                    for(std::size_t w = 0; w < output_dims[3]; ++w)
+                    {
+                        auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+                        for(std::size_t k = 0; k < output_dims[2]; ++k)
+                        {
+
+                            auto e_val =
+                                ck::type_convert<ReduceAccDataType>(conv_output_host(g, n, k, w));
+                            reduce0_op(reduce0_acc, e_val);
+                        }
+                        r0_host(g, n, w) = ck::type_convert<R0DataType>(reduce0_acc);
+                    }
+                }
+            }
+        }
+        else if constexpr(NDimSpatial == 2)
+        {
+            for(std::size_t g = 0; g < output_dims[0]; ++g)
+            {
+                for(std::size_t n = 0; n < output_dims[1]; ++n)
+                {
+                    for(std::size_t h = 0; h < output_dims[3]; ++h)
+                    {
+                        for(std::size_t w = 0; w < output_dims[4]; ++w)
+                        {
+                            auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+                            for(std::size_t k = 0; k < output_dims[2]; ++k)
+                            {
+
+                                auto e_val = ck::type_convert<ReduceAccDataType>(
+                                    conv_output_host(g, n, k, h, w));
+                                reduce0_op(reduce0_acc, e_val);
+                            }
+                            r0_host(g, n, h, w) = ck::type_convert<R0DataType>(reduce0_acc);
+                        }
+                    }
+                }
+            }
+        }
+        else if constexpr(NDimSpatial == 3)
+        {
+            for(std::size_t g = 0; g < output_dims[0]; ++g)
+            {
+                for(std::size_t n = 0; n < output_dims[1]; ++n)
+                {
+                    for(std::size_t d = 0; d < output_dims[3]; ++d)
+                    {
+                        for(std::size_t h = 0; h < output_dims[4]; ++h)
+                        {
+                            for(std::size_t w = 0; w < output_dims[5]; ++w)
+                            {
+                                auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+                                for(std::size_t k = 0; k < output_dims[2]; ++k)
+                                {
+
+                                    auto e_val = ck::type_convert<ReduceAccDataType>(
+                                        conv_output_host(g, n, k, d, h, w));
+                                    reduce0_op(reduce0_acc, e_val);
+                                }
+                                r0_host(g, n, d, h, w) = ck::type_convert<R0DataType>(reduce0_acc);
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        conv_output_device_buf.FromDevice(conv_output_device.mData.data());
+        r0_device_buf.FromDevice(r0_device.mData.data());
+
+        return ck::utils::check_err(conv_output_device,
+                                    conv_output_host,
+                                    "Error: incorrect results! (Matrix E)",
+                                    1e-5f,
+                                    1e-4f) &&
+               ck::utils::check_err(
+                   r0_device, r0_host, "Error: incorrect results! (Matrix R0)", 1e-5f, 1e-4f);
+    }
+
+    return true;
+}
+
+bool run_convnd_fwd_max_example(int argc, char* argv[])
+{
+    ck::utils::conv::ConvParam problem_size{
+        2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
+    ExecutionConfig config;
+
+    if(!parse_cmd_args(argc, argv, problem_size, config))
+    {
+        return false;
+    }
+
+    switch(problem_size.num_dim_spatial_)
+    {
+    case 1: return run_convnd_fwd_max<1>(problem_size, config);
+    case 2: return run_convnd_fwd_max<2>(problem_size, config);
+    case 3: return run_convnd_fwd_max<3>(problem_size, config);
+    }
+
+    return false;
+}
diff --git a/example/12_reduce/CMakeLists.txt b/example/12_reduce/CMakeLists.txt
new file mode 100644
index 0000000000..6e58ed9338
--- /dev/null
+++ b/example/12_reduce/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_example_executable(example_reduce_blockwise reduce_blockwise.cpp)
+add_example_executable(example_reduce_multiblock_atomic_add reduce_multiblock_atomic_add.cpp)
+add_example_executable(example_reduce_blockwise_two_call reduce_blockwise_two_call.cpp)
diff --git a/example/12_reduce/README.md b/example/12_reduce/README.md
new file mode 100644
index 0000000000..76d28527bb
--- /dev/null
+++ b/example/12_reduce/README.md
@@ -0,0 +1,62 @@
+# Instructions for ```example_reduce_blockwise```
+
+## Run ```example_reduce_blockwise```
+```bash
+# -D <xxx> : input 3d/4d/5d tensor lengths
+# -R <xxx> : reduce dimension ids
+# -v <x> :   verification (0=no, 1=yes)
+#arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64, 7: int4)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+./bin/example_reduce_blockwise -D 16,64,32,960 -v 1 0 2 1
+```
+
+Result
+```
+./bin/example_reduce_blockwise -D 16,64,32,960 -v 1 0 2 1
+launch_and_time_kernel: grid_dim {240, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 0.238063 ms, 264.285 GB/s, DeviceReduceBlockWise<256,M_C4_S1,K_C64_S1,InSrcVectorDim_0_InSrcVectorSize_1_OutDstVectorSize_1>
+```
+
+## Run ```example_reduce_multiblock_atomic_add```
+```bash
+# -D <xxx> : input 3d/4d/5d tensor lengths
+# -R <xxx> : reduce dimension ids
+# -v <x> :   verification (0=no, 1=yes)
+#arg1: data type (0: fp32, 1: fp64)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+./bin/example_reduce_multiblock_atomic_add -D 16,64,32,960 -v 1 0 2 0
+```
+
+Result
+```
+./bin/example_reduce_multiblock_atomic_add -D 16,64,32,960 -v 1 0 2 0
+Perf: 0 ms, inf GB/s, DeviceReduceMultiBlock<256,M_C4_S1,K_C64_S1,InSrcVectorDim_0_InSrcVectorSize_1_OutDstVectorSize_1>
+echo $?
+0
+```
+
+# Instructions for ```example_reduce_blockwise_two_call```
+
+## Run ```example_reduce_blockwise_two_call```
+```bash
+#arg1:  verification (0=no, 1=yes(
+#arg2:  initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3:  time kernel (0=no, 1=yes)
+./bin/example_reduce_blockwise_two_call 1 2 1
+```
+
+Result
+```
+./bin/example_reduce_blockwise_two_call 1 2 1
+launch_and_time_kernel: grid_dim {204800, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {6400, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 2.1791 ms, 771.42 GB/s, DeviceReduceBlockWise<256,M_C32_S1,K_C8_S1,InSrcVectorDim_1_InSrcVectorSize_1_OutDstVectorSize_1> => DeviceReduceBlockWise<256,M_C256_S1,K_C1_S1,InSrcVectorDim_1_InSrcVectorSize_1_OutDstVectorSize_1>
+```
diff --git a/example/12_reduce/reduce_blockwise.cpp b/example/12_reduce/reduce_blockwise.cpp
new file mode 100644
index 0000000000..9a736d4cfa
--- /dev/null
+++ b/example/12_reduce/reduce_blockwise.cpp
@@ -0,0 +1,299 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/utility/reduction_enums.hpp"
+#include "reduce_blockwise_impl.hpp"
+#include "reduce_example_common.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths = {16, 64, 32, 960};
+    std::vector<int> reduceDims   = {0, 1, 2};
+    std::vector<float> scales     = {1.0f, 0.0f};
+
+    bool do_verification = true;
+    int data_type        = 1;
+    int init_method      = 2;
+    bool time_kernel     = true;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
+                  << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of to-reduce dimensions"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
+                     "comparing with the host-based reduction"
+                  << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64, 7: int4)"
+                  << std::endl;
+        std::cout << "Arg2 -- init method (0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg3 -- time kernel (0=no, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:v:l:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 3 > argc)
+        {
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+        };
+
+        data_type   = std::atoi(argv[optind++]);
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(scales.empty())
+        {
+            scales.push_back(1.0f);
+            scales.push_back(0.0f);
+        };
+
+        return (0);
+    };
+};
+
+template <typename InOutDataType,
+          typename AccDataType,
+          ReduceTensorOp ReduceOpId,
+          index_t PropagateNan,
+          index_t OutputIndex>
+bool reduce_blockwise_test(bool do_verification,
+                           int init_method,
+                           bool time_kernel,
+                           const std::vector<size_t>& inLengths,
+                           const std::vector<int>& reduceDims,
+                           float alpha,
+                           float beta)
+{
+    bool matched = false;
+    int result   = 0;
+
+    const auto tuple_object = reduce_shape_instances{};
+
+    static_for<0, std::tuple_size<reduce_shape_instances>::value, 1>{}([&](auto i) {
+        if(matched)
+            return;
+
+        using ShapeType = remove_cvref_t<decltype(std::get<i>(tuple_object))>;
+
+        if(ShapeType::Rank_ != inLengths.size() || ShapeType::NumReduceDim_ != reduceDims.size())
+            return;
+
+        std::array<int, ShapeType::NumReduceDim_> arrReduceDims;
+
+        ck::ranges::copy(reduceDims, arrReduceDims.begin());
+
+        result = reduce_blockwise_impl<InOutDataType,
+                                       AccDataType,
+                                       ReduceOpId,
+                                       ShapeType::Rank_,
+                                       ShapeType::NumReduceDim_,
+                                       PropagateNan,
+                                       OutputIndex>(
+            do_verification, init_method, time_kernel, inLengths, arrReduceDims, alpha, beta);
+
+        matched = true;
+    });
+
+    return (result == 0) ? true : false;
+};
+
+constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::AVG;
+constexpr bool PropagateNan         = true;
+constexpr bool OutputIndex          = false;
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        SimpleAppArgs arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            pass = reduce_blockwise_test<ck::half_t, float, ReduceOpId, PropagateNan, OutputIndex>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+        else if(arg.data_type == 1)
+        {
+            pass = reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+        else if(arg.data_type == 3)
+        {
+            pass = reduce_blockwise_test<int8_t, float, ReduceOpId, PropagateNan, OutputIndex>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+        else if(arg.data_type == 5)
+        {
+            pass = reduce_blockwise_test<ck::bhalf_t, float, ReduceOpId, PropagateNan, OutputIndex>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+        else if(arg.data_type == 6)
+        {
+            pass = reduce_blockwise_test<double, double, ReduceOpId, PropagateNan, OutputIndex>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        else if(arg.data_type == 7)
+        {
+            pass = reduce_blockwise_test<int4_t, int32_t, ReduceTensorOp::AVG, false, false>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+
+            pass = pass && reduce_blockwise_test<int4_t, int8_t, ReduceTensorOp::MAX, false, false>(
+                               arg.do_verification,
+                               arg.init_method,
+                               arg.time_kernel,
+                               arg.inLengths,
+                               arg.reduceDims,
+                               arg.scales[0],
+                               arg.scales[1]);
+        }
+#endif
+    }
+    else
+    {
+        // for testing half_t
+        pass =
+            pass && reduce_blockwise_test<ck::half_t, float, ReduceOpId, PropagateNan, OutputIndex>(
+                        true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing float
+        pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
+                           true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing double
+        pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
+                           true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing bhalf_t
+        pass = pass &&
+               reduce_blockwise_test<ck::bhalf_t, float, ReduceOpId, PropagateNan, OutputIndex>(
+                   true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing int8_t
+        pass =
+            pass && reduce_blockwise_test<int8_t, int32_t, ReduceOpId, PropagateNan, OutputIndex>(
+                        true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        // for testing int4_t using AVG operation
+        pass = pass && reduce_blockwise_test<int4_t, int32_t, ReduceTensorOp::AVG, false, false>(
+                           true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing int4_t using MAX operation
+        pass = pass && reduce_blockwise_test<int4_t, int8_t, ReduceTensorOp::MAX, false, false>(
+                           true, 2, true, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+#endif
+        // for testing 3D input
+        pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
+                           true, 2, true, {16, 64, 960}, {0, 1}, 1.0f, 0.0f);
+
+        // for testing 5D input
+        pass = pass && reduce_blockwise_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
+                           true, 2, true, {16, 64, 32, 2, 960}, {0, 1, 2, 3}, 1.0f, 0.0f);
+    };
+
+    return (pass ? 0 : 1);
+};
diff --git a/example/12_reduce/reduce_blockwise_impl.hpp b/example/12_reduce/reduce_blockwise_impl.hpp
new file mode 100644
index 0000000000..7f8394a730
--- /dev/null
+++ b/example/12_reduce/reduce_blockwise_impl.hpp
@@ -0,0 +1,356 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_reduce.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+
+#include "reduce_example_common.hpp"
+
+template <typename InOutDataType,
+          typename AccDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          ck::index_t Rank,
+          ck::index_t NumReduceDim,
+          bool PropagateNan,
+          bool OutputIndex>
+int reduce_blockwise_impl(bool do_verification,
+                          int init_method,
+                          bool time_kernel,
+                          const std::vector<size_t>& inLengths,
+                          const std::array<int, NumReduceDim>& reduceDims,
+                          float alpha,
+                          float beta)
+
+{
+    using namespace ck;
+    using namespace ck::tensor_operation::device;
+
+    constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
+
+    constexpr bool op_support_indices =
+        (ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
+         ReduceOpId == ReduceTensorOp::AMAX);
+
+    constexpr bool invalid_reduce_1 = OutputIndex && !op_support_indices;
+
+    // 1) If InOutDataType is half_t, must use half_t as AccDataType for indexable reduction
+    // operations 2) If InOutDataType is half_t, must use float as AccDataType for non-indexable
+    // reduction operations
+    constexpr bool invalid_reduce_2 =
+        std::is_same<InOutDataType, half_t>::value &&
+        ((!op_support_indices && !std::is_same<AccDataType, float>::value) ||
+         (op_support_indices && !std::is_same<AccDataType, half_t>::value));
+
+    // 1) If InOutDataType is float, must use float as AccDataType for indexable reduction
+    // operations
+    constexpr bool invalid_reduce_3 =
+        std::is_same<InOutDataType, float>::value &&
+        (op_support_indices && !std::is_same<AccDataType, float>::value);
+
+    // 1) If InOutDataType is int8_t or int4_t, must use int8_t as AccDataType for indexable
+    // reduction operations 2) If InOutDataType is int8_t or int4_t, must use int32_t as AccDataType
+    // for non-indexable reduction operations
+    constexpr bool invalid_reduce_4 =
+        std::is_same<InOutDataType, int8_t>::value &&
+        ((!op_support_indices && !std::is_same<AccDataType, int32_t>::value) ||
+         (op_support_indices && !std::is_same<AccDataType, int8_t>::value));
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    constexpr bool invalid_reduce_4_2 =
+        std::is_same<InOutDataType, int4_t>::value &&
+        ((!op_support_indices && !std::is_same<AccDataType, int32_t>::value) ||
+         (op_support_indices && !std::is_same<AccDataType, int8_t>::value));
+#endif
+
+    // 1) If InOutDataType is int8_t or int4_t, the supported operation must be either indexable
+    // operations or ADD/AVG
+    constexpr bool invalid_reduce_5 = std::is_same<InOutDataType, int8_t>::value &&
+                                      (!op_support_indices && ReduceOpId != ReduceTensorOp::ADD &&
+                                       ReduceOpId != ReduceTensorOp::AVG);
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    constexpr bool invalid_reduce_5_2 = std::is_same<InOutDataType, int4_t>::value &&
+                                        (!op_support_indices && ReduceOpId != ReduceTensorOp::ADD &&
+                                         ReduceOpId != ReduceTensorOp::AVG);
+#endif
+
+    // 1) If InOutDataType is bhalf_t, must use float as AccDataType for all reduction operations
+    constexpr bool invalid_reduce_6 =
+        std::is_same<InOutDataType, bhalf_t>::value && !std::is_same<AccDataType, float>::value;
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    constexpr bool invalid_reduce =
+        (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 || invalid_reduce_4 ||
+         invalid_reduce_5 || invalid_reduce_6 || invalid_reduce_4_2 || invalid_reduce_5_2);
+#else
+    constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 ||
+                                     invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6);
+#endif
+
+    if constexpr(invalid_reduce)
+    {
+        std::cerr << "The reduction setting is invalid, exiting!" << std::endl;
+        return (-1);
+    };
+
+    using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
+    using InElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+    using AccElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    using InOutDataTypeInDevice = typename std::
+        conditional<std::is_same<InOutDataType, int4_t>::value, int8_t, InOutDataType>::type;
+#else
+    using InOutDataTypeInDevice   = InOutDataType;
+#endif
+
+    using DeviceReduceInstance =
+        ck::tensor_operation::device::DeviceReduceMultiBlock<InOutDataTypeInDevice,
+                                                             AccDataType,
+                                                             InOutDataTypeInDevice,
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             ReduceOperation,
+                                                             InElementwiseOperation,
+                                                             AccElementwiseOperation,
+                                                             InMemoryDataOperationEnum::Set,
+                                                             PropagateNan,
+                                                             OutputIndex,
+                                                             false, // HaveIndexInputIfOutputIndex
+                                                             256,   // BlockSize
+                                                             4,     // MThreadClusterSize
+                                                             64,    // KThreadClusterSize
+                                                             1,     // MThreadSliceSize
+                                                             1,     // KThreadSliceSize
+                                                             0,     // InSrcVectorDim
+                                                             1,     // InSrceVectorSize
+                                                             1>;    // OutDstVectorSize
+
+    Tensor<InOutDataType> in(inLengths);
+
+    std::vector<size_t> outLengths;
+
+    auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
+
+    if(invariantDims.empty())
+        outLengths.push_back(1);
+    else
+        for(auto dim : invariantDims)
+            outLengths.push_back(inLengths[dim]);
+
+    Tensor<InOutDataType> out_ref(outLengths);
+    Tensor<InOutDataType> out(outLengths);
+    Tensor<int> out_indices_ref(outLengths);
+    Tensor<int> out_indices(outLengths);
+
+    auto inStrides  = in.mDesc.GetStrides();
+    auto outStrides = out.mDesc.GetStrides();
+
+    size_t invariant_total_length = out.mDesc.GetElementSize();
+    size_t reduce_total_length    = in.mDesc.GetElementSize() / invariant_total_length;
+
+    std::size_t num_thread = 1;
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            in.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            break;
+        case 2:
+            in.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            in.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0},
+                                            num_thread);
+        }
+
+        if(beta != 0.0f)
+            for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
+                out.mData[i] = out_ref.mData[i];
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem in_dev(sizeof(InOutDataTypeInDevice) * in.mDesc.GetElementSpaceSize());
+    DeviceMem out_dev(sizeof(InOutDataTypeInDevice) * out.mDesc.GetElementSpaceSize());
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    if(std::is_same<InOutDataType, int4_t>::value)
+    {
+        std::vector<InOutDataTypeInDevice> tmp_buf(in.mData.size());
+
+        std::copy_n(in.mData.data(), in.mData.size(), tmp_buf.data());
+        in_dev.ToDevice(tmp_buf.data());
+    }
+    else
+#endif
+        in_dev.ToDevice(in.mData.data());
+
+    if(beta != 0.0f)
+    {
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        if(std::is_same<InOutDataType, int4_t>::value)
+        {
+            std::vector<InOutDataTypeInDevice> tmp_buf(in.mData.size());
+
+            std::copy_n(out.mData.data(), out.mData.size(), tmp_buf.data());
+            out_dev.ToDevice(tmp_buf.data());
+        }
+        else
+#endif
+            out_dev.ToDevice(out.mData.data());
+    };
+
+    size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int32_t) : 0;
+
+    DeviceMem out_index_dev(indicesSizeInBytes);
+
+    InElementwiseOperation in_elementwise_op;
+    AccElementwiseOperation acc_elementwise_op;
+
+    std::tie(in_elementwise_op, acc_elementwise_op) =
+        reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+            static_cast<int32_t>(reduce_total_length));
+
+    std::array<index_t, Rank> arrInLengths;
+    std::array<index_t, Rank> arrInStrides;
+    std::array<index_t, NumOutDim> arrOutLengths;
+    std::array<index_t, NumOutDim> arrOutStrides;
+
+    ck::ranges::copy(inLengths, arrInLengths.begin());
+    ck::ranges::copy(inStrides, arrInStrides.begin());
+    ck::ranges::copy(outLengths, arrOutLengths.begin());
+    ck::ranges::copy(outStrides, arrOutStrides.begin());
+
+    if(do_verification)
+    {
+        using ReferenceReduceInstance =
+            ck::tensor_operation::host::ReferenceReduce<InOutDataType,
+                                                        AccDataType,
+                                                        InOutDataType,
+                                                        Rank,
+                                                        NumReduceDim,
+                                                        ReduceOperation,
+                                                        InElementwiseOperation,
+                                                        AccElementwiseOperation,
+                                                        PropagateNan,
+                                                        OutputIndex>;
+
+        auto reduce_ref = ReferenceReduceInstance{};
+
+        auto argument_ptr_ref = reduce_ref.MakeArgumentPointer(arrInLengths,
+                                                               arrInStrides,
+                                                               arrOutLengths,
+                                                               arrOutStrides,
+                                                               reduceDims,
+                                                               static_cast<double>(alpha),
+                                                               static_cast<double>(beta),
+                                                               in.mData.data(),
+                                                               nullptr,
+                                                               out_ref.mData.data(),
+                                                               out_indices_ref.mData.data(),
+                                                               in_elementwise_op,
+                                                               acc_elementwise_op);
+
+        if(!reduce_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters not supported by the reduce reference, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = reduce_ref.MakeInvokerPointer();
+
+        invoker_ptr_ref->Run(argument_ptr_ref.get());
+    };
+
+    auto reduce = DeviceReduceInstance{};
+
+    auto argument_ptr = reduce.MakeArgumentPointer(arrInLengths,
+                                                   arrInStrides,
+                                                   arrOutLengths,
+                                                   arrOutStrides,
+                                                   reduceDims,
+                                                   static_cast<double>(alpha),
+                                                   static_cast<double>(beta),
+                                                   in_dev.GetDeviceBuffer(),
+                                                   nullptr,
+                                                   out_dev.GetDeviceBuffer(),
+                                                   out_index_dev.GetDeviceBuffer(),
+                                                   in_elementwise_op,
+                                                   acc_elementwise_op);
+
+    if(!reduce.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cerr << "The runtime parameters not supported by the DeviceReduce instance, exiting!"
+                  << std::endl;
+
+        return (-2);
+    };
+
+    std::string reduce_name = reduce.GetTypeString();
+
+    auto invoker_ptr = reduce.MakeInvokerPointer();
+
+    float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InOutDataType) +
+                            invariant_total_length * sizeof(InOutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+    std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        if(std::is_same<InOutDataType, int4_t>::value)
+        {
+            std::vector<InOutDataTypeInDevice> tmp_buf(out.mData.size());
+
+            out_dev.FromDevice(tmp_buf.data());
+
+            std::copy_n(tmp_buf.data(), out.mData.size(), out.mData.data());
+        }
+        else
+#endif
+            out_dev.FromDevice(out.mData.data());
+
+        pass = pass && ck::utils::check_err(out, out_ref);
+
+        if(OutputIndex)
+        {
+            out_index_dev.FromDevice(out_indices.mData.data());
+            pass = pass && ck::utils::check_err(out_indices, out_indices_ref);
+        };
+    };
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/12_reduce/reduce_blockwise_two_call.cpp b/example/12_reduce/reduce_blockwise_two_call.cpp
new file mode 100644
index 0000000000..eb8b5c76d3
--- /dev/null
+++ b/example/12_reduce/reduce_blockwise_two_call.cpp
@@ -0,0 +1,319 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <sstream>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_reduce.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+using InOutDataType = ck::half_t;
+using InOutDataType = ck::half_t;
+using AccDataType   = float;
+
+constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::NORM2;
+constexpr bool PropagateNan         = true;
+constexpr bool OutputIndex          = false;
+
+using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
+using InElementwiseOperation =
+    typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+using AccElementwiseOperation =
+    typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+
+using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+using DeviceReduceInstance_1 = DeviceReduceMultiBlock<InOutDataType,
+                                                      AccDataType,
+                                                      InOutDataType,
+                                                      5, // Rank
+                                                      1, // NumReduceDim
+                                                      ReduceOperation,
+                                                      InElementwiseOperation,
+                                                      PassThroughOp,
+                                                      InMemoryDataOperationEnum::Set,
+                                                      PropagateNan,
+                                                      OutputIndex,
+                                                      false, // HaveIndexInputIfOutputIndex
+                                                      256,
+                                                      32,
+                                                      8,
+                                                      1,
+                                                      1,
+                                                      1, // vector dim
+                                                      1,
+                                                      1>;
+
+using DeviceReduceInstance_2 = DeviceReduceMultiBlock<InOutDataType,
+                                                      AccDataType,
+                                                      InOutDataType,
+                                                      4, // Rank
+                                                      1, // NumReduceDim
+                                                      ReduceOperation,
+                                                      PassThroughOp,
+                                                      AccElementwiseOperation,
+                                                      InMemoryDataOperationEnum::Set,
+                                                      PropagateNan,
+                                                      OutputIndex,
+                                                      false, // HaveIndexInputIfOutputIndex
+                                                      256,
+                                                      128,
+                                                      2,
+                                                      1,
+                                                      1,
+                                                      1, // vector dim
+                                                      1,
+                                                      1>;
+
+static bool do_verify;
+static int init_method;
+static float alpha;
+static float beta;
+static bool time_kernel;
+
+int main(int argc, char* argv[])
+{
+    // used by the device reduction
+    const std::array<int, 1> reduceDims_1 = {4};
+    // const std::array<int, 4> invariantDims_1 = {0, 1, 2, 3};
+
+    const std::array<int, 1> reduceDims_2 = {3};
+    // const std::array<int, 3> invariantDims_2 = {0, 1, 2};
+
+    // used by the host reduction
+    const std::array<int, 2> reduceDims = {3, 4};
+    // const std::array<int, 3> invariantDims = {0, 1, 2};
+
+    const std::vector<size_t> inLengths_1 = {64, 320, 80, 4, 128};
+
+    // input lengths of the second reduction, which is also the output lengths of the first
+    // reduction
+    const std::vector<size_t> inLengths_2 = {64, 320, 80, 4};
+
+    const std::vector<size_t> outLengths = {64, 320, 80};
+
+    if(argc == 1)
+    {
+        do_verify   = true;
+        init_method = 2;
+        time_kernel = true;
+    }
+    else if(argc == 4)
+    {
+        do_verify   = static_cast<bool>(argv[1]);
+        init_method = atoi(argv[2]);
+        time_kernel = static_cast<bool>(atoi(argv[3]));
+    }
+    else
+    {
+        std::ostringstream ostr;
+
+        ostr << "Wrong parameter! " << std::endl
+             << "Usage: " << argv[0] << "[verify 0/1] init_method time_kernel" << std::endl;
+
+        throw std::runtime_error(ostr.str());
+    };
+
+    alpha = 1.0f;
+    beta  = 0.0f;
+
+    Tensor<InOutDataType> in_1(inLengths_1);
+
+    Tensor<InOutDataType> out_ref(outLengths);
+    Tensor<InOutDataType> in_2(inLengths_2); // also the output tensor of the first reduction
+    Tensor<InOutDataType> out(outLengths);
+
+    auto inStrides_1 = in_1.mDesc.GetStrides();
+    auto inStrides_2 = in_2.mDesc.GetStrides();
+    auto outStrides  = out.mDesc.GetStrides();
+
+    size_t invariant_total_length = out.mDesc.GetElementSize();
+    size_t reduce_total_length    = in_1.mDesc.GetElementSize() / invariant_total_length;
+
+    std::size_t num_thread = 1;
+
+    if(do_verify)
+    {
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            in_1.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            break;
+        case 2:
+            in_1.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            in_1.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0},
+                                            num_thread);
+        }
+
+        if(beta != 0.0f)
+            for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
+                out.mData[i] = out_ref.mData[i];
+    };
+
+    DeviceMem in_1_dev(sizeof(InOutDataType) * in_1.mDesc.GetElementSpaceSize());
+    DeviceMem in_2_dev(sizeof(InOutDataType) * in_2.mDesc.GetElementSpaceSize());
+    DeviceMem out_dev(sizeof(InOutDataType) * out.mDesc.GetElementSpaceSize());
+
+    in_1_dev.ToDevice(in_1.mData.data());
+
+    if(beta != 0.0f)
+        out_dev.ToDevice(out.mData.data());
+
+    InElementwiseOperation in_elementwise_op;
+    AccElementwiseOperation acc_elementwise_op;
+
+    std::tie(in_elementwise_op, acc_elementwise_op) =
+        reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+            static_cast<int32_t>(reduce_total_length));
+
+    std::array<index_t, 5> arrInLengths_1;
+    std::array<index_t, 5> arrInStrides_1;
+    std::array<index_t, 4> arrInLengths_2;
+    std::array<index_t, 4> arrInStrides_2;
+    std::array<index_t, 3> arrOutLengths;
+    std::array<index_t, 3> arrOutStrides;
+
+    ck::ranges::copy(inLengths_1, arrInLengths_1.begin());
+    ck::ranges::copy(inStrides_1, arrInStrides_1.begin());
+    ck::ranges::copy(inLengths_2, arrInLengths_2.begin());
+    ck::ranges::copy(inStrides_2, arrInStrides_2.begin());
+    ck::ranges::copy(outLengths, arrOutLengths.begin());
+    ck::ranges::copy(outStrides, arrOutStrides.begin());
+
+    if(do_verify)
+    {
+        using ReferenceReduceInstance =
+            ck::tensor_operation::host::ReferenceReduce<InOutDataType,
+                                                        AccDataType,
+                                                        InOutDataType,
+                                                        5,
+                                                        2,
+                                                        ReduceOperation,
+                                                        InElementwiseOperation,
+                                                        AccElementwiseOperation,
+                                                        PropagateNan,
+                                                        OutputIndex>;
+
+        auto reduce_ref = ReferenceReduceInstance{};
+
+        auto argument_ptr_ref = reduce_ref.MakeArgumentPointer(arrInLengths_1,
+                                                               arrInStrides_1,
+                                                               arrOutLengths,
+                                                               arrOutStrides,
+                                                               reduceDims,
+                                                               static_cast<double>(alpha),
+                                                               static_cast<double>(beta),
+                                                               in_1.mData.data(),
+                                                               nullptr,
+                                                               out_ref.mData.data(),
+                                                               nullptr,
+                                                               in_elementwise_op,
+                                                               acc_elementwise_op);
+
+        if(!reduce_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters not supported by the reduce reference, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = reduce_ref.MakeInvokerPointer();
+
+        invoker_ptr_ref->Run(argument_ptr_ref.get());
+    };
+
+    auto reduce_1 = DeviceReduceInstance_1{};
+
+    auto argument_ptr_1 = reduce_1.MakeArgumentPointer(arrInLengths_1,
+                                                       arrInStrides_1,
+                                                       arrInLengths_2,
+                                                       arrInStrides_2,
+                                                       reduceDims_1,
+                                                       1.0,
+                                                       0.0,
+                                                       in_1_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       in_2_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       in_elementwise_op,
+                                                       PassThroughOp{});
+
+    if(!reduce_1.IsSupportedArgument(argument_ptr_1.get()))
+    {
+        std::cout << "The runtime parameters seems supported by the DeviceReduce instance, exiting!"
+                  << std::endl;
+    };
+
+    auto invoker_ptr_1 = reduce_1.MakeInvokerPointer();
+
+    auto reduce_2 = DeviceReduceInstance_2{};
+
+    auto argument_ptr_2 = reduce_2.MakeArgumentPointer(arrInLengths_2,
+                                                       arrInStrides_2,
+                                                       arrOutLengths,
+                                                       arrOutStrides,
+                                                       reduceDims_2,
+                                                       static_cast<double>(alpha),
+                                                       static_cast<double>(beta),
+                                                       in_2_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       out_dev.GetDeviceBuffer(),
+                                                       nullptr,
+                                                       PassThroughOp{},
+                                                       acc_elementwise_op);
+
+    if(!reduce_2.IsSupportedArgument(argument_ptr_2.get()))
+    {
+        std::cout
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+    };
+
+    auto invoker_ptr_2 = reduce_2.MakeInvokerPointer();
+
+    float avg_time_1 = invoker_ptr_1->Run(argument_ptr_1.get(), StreamConfig{nullptr, time_kernel});
+    float avg_time_2 = invoker_ptr_2->Run(argument_ptr_2.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InOutDataType) +
+                            invariant_total_length * sizeof(InOutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / (avg_time_1 + avg_time_2);
+
+    std::cout << "Perf: " << avg_time_1 + avg_time_2 << " ms, " << gb_per_sec << " GB/s, "
+              << reduce_1.GetTypeString() << " => " << reduce_2.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verify)
+    {
+        out_dev.FromDevice(out.mData.data());
+        pass = pass && ck::utils::check_err(out, out_ref);
+    };
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/12_reduce/reduce_example_common.hpp b/example/12_reduce/reduce_example_common.hpp
new file mode 100644
index 0000000000..5f9a48804a
--- /dev/null
+++ b/example/12_reduce/reduce_example_common.hpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+
+template <int Rank, int NumReduceDim>
+static inline std::array<int, Rank - NumReduceDim>
+get_invariant_dims(const std::array<int, NumReduceDim>& reduceDims)
+{
+    int reduceFlag = 0;
+
+    // flag the bits for the reduceDims
+    for(int i = 0; i < NumReduceDim; i++)
+    {
+        reduceFlag |= 1 << reduceDims[i];
+    };
+
+    std::array<int, Rank - NumReduceDim> invariantDims;
+
+    // collect invariant dimensions
+    int dim = 0;
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) == 0)
+        {
+            invariantDims[dim] = i;
+            dim++;
+        };
+
+    return invariantDims;
+};
+
+template <ck::index_t Rank, ck::index_t NumReduceDim>
+struct ReduceShape
+{
+    static constexpr ck::index_t Rank_         = Rank;
+    static constexpr ck::index_t NumReduceDim_ = NumReduceDim;
+};
+
+using reduce_shape_instances = std::tuple<ReduceShape<3, 1>,
+                                          ReduceShape<3, 2>,
+                                          ReduceShape<4, 1>,
+                                          ReduceShape<4, 2>,
+                                          ReduceShape<4, 3>,
+                                          ReduceShape<5, 1>,
+                                          ReduceShape<5, 2>,
+                                          ReduceShape<5, 3>,
+                                          ReduceShape<5, 4>>;
diff --git a/example/12_reduce/reduce_multiblock_atomic_add.cpp b/example/12_reduce/reduce_multiblock_atomic_add.cpp
new file mode 100644
index 0000000000..120e3f0595
--- /dev/null
+++ b/example/12_reduce/reduce_multiblock_atomic_add.cpp
@@ -0,0 +1,216 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/utility/reduction_enums.hpp"
+#include "reduce_multiblock_atomic_add_impl.hpp"
+#include "reduce_example_common.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths = {16, 64, 32, 960};
+    std::vector<int> reduceDims   = {0, 1, 2};
+    std::vector<float> scales     = {1.0f, 0.0f};
+
+    bool do_verification = true;
+    int data_type        = 1;
+    int init_method      = 2;
+    bool time_kernel     = true;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
+                  << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of to-reduce dimensions"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
+                     "comparing with the host-based reduction"
+                  << std::endl;
+        std::cout << "Arg1: data type (0: fp32, 1: fp64)" << std::endl;
+        std::cout << "Arg2 -- init method (0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg3 -- time kernel (0=no, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:v:l:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 3 > argc)
+        {
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+        };
+
+        data_type   = std::atoi(argv[optind++]);
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(scales.empty())
+        {
+            scales.push_back(1.0f);
+            scales.push_back(0.0f);
+        };
+
+        return (0);
+    };
+};
+
+template <typename InOutDataType,
+          typename AccDataType,
+          ReduceTensorOp ReduceOpId,
+          index_t PropagateNan>
+bool reduce_multiblock_atomic_add_test(bool do_verification,
+                                       int init_method,
+                                       bool time_kernel,
+                                       const std::vector<size_t>& inLengths,
+                                       const std::vector<int>& reduceDims,
+                                       float alpha,
+                                       float beta)
+{
+    bool matched = false;
+    int result   = 0;
+
+    const auto tuple_object = reduce_shape_instances{};
+
+    static_for<0, std::tuple_size<reduce_shape_instances>::value, 1>{}([&](auto i) {
+        if(matched)
+            return;
+
+        using ShapeType = remove_cvref_t<decltype(std::get<i>(tuple_object))>;
+
+        if(ShapeType::Rank_ != inLengths.size() || ShapeType::NumReduceDim_ != reduceDims.size())
+            return;
+
+        std::array<int, ShapeType::NumReduceDim_> a_reduceDims;
+
+        ck::ranges::copy(reduceDims, a_reduceDims.begin());
+
+        result = reduce_multiblock_atomic_add_impl<InOutDataType,
+                                                   AccDataType,
+                                                   ReduceOpId,
+                                                   ShapeType::Rank_,
+                                                   ShapeType::NumReduceDim_,
+                                                   PropagateNan>(
+            do_verification, init_method, time_kernel, inLengths, a_reduceDims, alpha, beta);
+
+        matched = true;
+    });
+
+    return (result == 0) ? true : false;
+};
+
+constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::AVG;
+constexpr bool PropagateNan         = true;
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        SimpleAppArgs arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            pass = reduce_multiblock_atomic_add_test<float, float, ReduceOpId, PropagateNan>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+        else if(arg.data_type == 1)
+        {
+            pass = reduce_multiblock_atomic_add_test<double, double, ReduceOpId, PropagateNan>(
+                arg.do_verification,
+                arg.init_method,
+                arg.time_kernel,
+                arg.inLengths,
+                arg.reduceDims,
+                arg.scales[0],
+                arg.scales[1]);
+        }
+    }
+    else
+    {
+        // for testing float
+        pass = pass && reduce_multiblock_atomic_add_test<float, float, ReduceOpId, PropagateNan>(
+                           true, 2, false, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing double
+        pass = pass && reduce_multiblock_atomic_add_test<double, double, ReduceOpId, PropagateNan>(
+                           true, 2, false, {16, 64, 32, 960}, {0, 1, 2}, 1.0f, 0.0f);
+
+        // for testing 3D input
+        pass = pass && reduce_multiblock_atomic_add_test<float, float, ReduceOpId, PropagateNan>(
+                           true, 2, false, {16, 64, 960}, {0, 1}, 1.0f, 0.0f);
+
+        // for testing 5D input
+        pass = pass && reduce_multiblock_atomic_add_test<float, float, ReduceOpId, PropagateNan>(
+                           true, 2, false, {16, 64, 32, 2, 960}, {0, 1, 2, 3}, 1.0f, 0.0f);
+    };
+
+    return (pass ? 0 : 1);
+};
diff --git a/example/12_reduce/reduce_multiblock_atomic_add_impl.hpp b/example/12_reduce/reduce_multiblock_atomic_add_impl.hpp
new file mode 100644
index 0000000000..fed6218644
--- /dev/null
+++ b/example/12_reduce/reduce_multiblock_atomic_add_impl.hpp
@@ -0,0 +1,251 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_reduce.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+
+#include "reduce_example_common.hpp"
+
+template <typename InOutDataType,
+          typename AccDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          ck::index_t Rank,
+          ck::index_t NumReduceDim,
+          bool PropagateNan>
+int reduce_multiblock_atomic_add_impl(bool do_verification,
+                                      int init_method,
+                                      bool time_kernel,
+                                      const std::vector<size_t>& inLengths,
+                                      const std::array<int, NumReduceDim>& reduceDims,
+                                      float alpha,
+                                      float beta)
+
+{
+    using namespace ck;
+    using namespace ck::tensor_operation::device;
+
+    constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
+
+    constexpr bool op_support_atomic_add =
+        (ReduceOpId == ReduceTensorOp::ADD || ReduceOpId == ReduceTensorOp::AVG);
+
+    constexpr bool invalid_reduce_1 = !op_support_atomic_add;
+    constexpr bool invalid_reduce_2 =
+        !(std::is_same<InOutDataType, float>::value || std::is_same<InOutDataType, double>::value);
+
+    constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2);
+
+    if(invalid_reduce)
+    {
+        std::cerr << "The reduction setting is invalid, exiting!" << std::endl;
+        return (-1);
+    };
+
+    using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
+    using InElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+    using AccElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+
+    using DeviceReduceInstance =
+        ck::tensor_operation::device::DeviceReduceMultiBlock<InOutDataType,
+                                                             AccDataType,
+                                                             InOutDataType,
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             ReduceOperation,
+                                                             InElementwiseOperation,
+                                                             AccElementwiseOperation,
+                                                             InMemoryDataOperationEnum::AtomicAdd,
+                                                             PropagateNan,
+                                                             false,
+                                                             false, // HaveIndexInputIfOutputIndex
+                                                             256,
+                                                             4,
+                                                             64,
+                                                             1,
+                                                             1,
+                                                             0,
+                                                             1,
+                                                             1>;
+
+    Tensor<InOutDataType> in(inLengths);
+
+    std::vector<size_t> outLengths;
+
+    auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
+
+    if(invariantDims.empty())
+        outLengths.push_back(1);
+    else
+        for(auto dim : invariantDims)
+            outLengths.push_back(inLengths[dim]);
+
+    Tensor<InOutDataType> out_ref(outLengths);
+    Tensor<InOutDataType> out(outLengths);
+
+    auto inStrides  = in.mDesc.GetStrides();
+    auto outStrides = out.mDesc.GetStrides();
+
+    size_t invariant_total_length = out.mDesc.GetElementSize();
+    size_t reduce_total_length    = in.mDesc.GetElementSize() / invariant_total_length;
+
+    std::size_t num_thread = 1;
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            in.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
+            break;
+        case 2:
+            in.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            in.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0},
+                                            num_thread);
+        }
+
+        if(beta != 0.0f)
+            for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
+                out.mData[i] = out_ref.mData[i];
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem in_dev(sizeof(InOutDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem out_dev(sizeof(InOutDataType) * out.mDesc.GetElementSpaceSize());
+
+    in_dev.ToDevice(in.mData.data());
+
+    if(beta != 0.0f)
+        out_dev.ToDevice(out.mData.data());
+
+    InElementwiseOperation in_elementwise_op;
+    AccElementwiseOperation acc_elementwise_op;
+
+    std::tie(in_elementwise_op, acc_elementwise_op) =
+        reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+            static_cast<int32_t>(reduce_total_length));
+
+    std::array<index_t, Rank> arrInLengths;
+    std::array<index_t, Rank> arrInStrides;
+    std::array<index_t, NumOutDim> arrOutLengths;
+    std::array<index_t, NumOutDim> arrOutStrides;
+
+    ck::ranges::copy(inLengths, arrInLengths.begin());
+    ck::ranges::copy(inStrides, arrInStrides.begin());
+    ck::ranges::copy(outLengths, arrOutLengths.begin());
+    ck::ranges::copy(outStrides, arrOutStrides.begin());
+
+    if(do_verification)
+    {
+        using ReferenceReduceInstance =
+            ck::tensor_operation::host::ReferenceReduce<InOutDataType,
+                                                        AccDataType,
+                                                        InOutDataType,
+                                                        Rank,
+                                                        NumReduceDim,
+                                                        ReduceOperation,
+                                                        InElementwiseOperation,
+                                                        AccElementwiseOperation,
+                                                        PropagateNan,
+                                                        false>;
+
+        auto reduce_ref = ReferenceReduceInstance{};
+
+        auto argument_ptr_ref = reduce_ref.MakeArgumentPointer(arrInLengths,
+                                                               arrInStrides,
+                                                               arrOutLengths,
+                                                               arrOutStrides,
+                                                               reduceDims,
+                                                               static_cast<double>(alpha),
+                                                               static_cast<double>(beta),
+                                                               in.mData.data(),
+                                                               nullptr,
+                                                               out_ref.mData.data(),
+                                                               nullptr,
+                                                               in_elementwise_op,
+                                                               acc_elementwise_op);
+
+        if(!reduce_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters not supported by the reduce reference, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = reduce_ref.MakeInvokerPointer();
+
+        invoker_ptr_ref->Run(argument_ptr_ref.get());
+    };
+
+    auto reduce = DeviceReduceInstance{};
+
+    auto argument_ptr = reduce.MakeArgumentPointer(arrInLengths,
+                                                   arrInStrides,
+                                                   arrOutLengths,
+                                                   arrOutStrides,
+                                                   reduceDims,
+                                                   static_cast<double>(alpha),
+                                                   static_cast<double>(beta),
+                                                   in_dev.GetDeviceBuffer(),
+                                                   nullptr,
+                                                   out_dev.GetDeviceBuffer(),
+                                                   nullptr,
+                                                   in_elementwise_op,
+                                                   acc_elementwise_op);
+
+    if(!reduce.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cerr << "The runtime parameters not supported by the DeviceReduce instance, exiting!"
+                  << std::endl;
+
+        return (-2);
+    };
+
+    std::string reduce_name = reduce.GetTypeString();
+
+    auto invoker_ptr = reduce.MakeInvokerPointer();
+
+    float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InOutDataType) +
+                            invariant_total_length * sizeof(InOutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+    std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        out_dev.FromDevice(out.mData.data());
+        pass = pass && ck::utils::check_err(out, out_ref);
+    };
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/13_pool2d_fwd/CMakeLists.txt b/example/13_pool2d_fwd/CMakeLists.txt
new file mode 100644
index 0000000000..d0f356757b
--- /dev/null
+++ b/example/13_pool2d_fwd/CMakeLists.txt
@@ -0,0 +1,6 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_pool2d_fwd_fp16 pool2d_fwd_fp16.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_pool2d_fwd_fp32 pool2d_fwd_fp32.cpp)
+endif()
diff --git a/example/13_pool2d_fwd/README.md b/example/13_pool2d_fwd/README.md
new file mode 100644
index 0000000000..9b017734e9
--- /dev/null
+++ b/example/13_pool2d_fwd/README.md
@@ -0,0 +1,41 @@
+# Instructions for ```example_pool2d_fwd``` Examples
+
+## Run ```example_pool2d_fwd_fp16```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, RightPx
+./bin/example_pool2d_fwd_fp16 1 1 1
+```
+
+Result 
+```
+in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
+out_n_c_ho_wo: dim 4, lengths {128, 192, 36, 36}, strides {248832, 1, 6912, 192}
+launch_and_time_kernel: grid_dim {124416, 1, 1}, block_dim {64, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 0.397436 ms, 1.44252 TFlops, 783.713 GB/s
+```
+
+## Run ```example_pool2d_fwd_fp32```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, RightPx
+./bin/example_pool2d_fwd_fp32 1 1 1
+```
+
+
+Result 
+```
+./bin/example_pool2d_fwd_fp32 1 1 1
+in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
+out_n_c_ho_wo: dim 4, lengths {128, 192, 36, 36}, strides {248832, 1, 6912, 192}
+launch_and_time_kernel: grid_dim {124416, 1, 1}, block_dim {64, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 1.01823 ms, 0.563045 TFlops, 611.8 GB/s
+```
diff --git a/example/13_pool2d_fwd/pool2d_fwd_common.hpp b/example/13_pool2d_fwd/pool2d_fwd_common.hpp
new file mode 100644
index 0000000000..3ce08fd2af
--- /dev/null
+++ b/example/13_pool2d_fwd/pool2d_fwd_common.hpp
@@ -0,0 +1,197 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp"
+
+template <typename InDataType,
+          typename OutDataType,
+          typename ComputeDataType,
+          typename IndexDataType,
+          typename InLayout,
+          typename OutLayout,
+          ck::ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool OutputIndex>
+bool pool_test(bool do_verification,
+               int init_method,
+               bool time_kernel,
+               ck::index_t N,
+               ck::index_t C,
+               ck::index_t Y,
+               ck::index_t X,
+               ck::index_t Hi,
+               ck::index_t Wi,
+               ck::index_t window_stride_h,
+               ck::index_t window_stride_w,
+               ck::index_t window_dilation_h,
+               ck::index_t window_dilation_w,
+               ck::index_t in_left_pad_h,
+               ck::index_t in_left_pad_w,
+               ck::index_t in_right_pad_h,
+               ck::index_t in_right_pad_w)
+{
+    using DevicePoolFwdInstance =
+        ck::tensor_operation::device::DevicePool2dFwd_NHWC_NHWC<InDataType,
+                                                                OutDataType,
+                                                                IndexDataType,
+                                                                ComputeDataType,
+                                                                ReduceOpId,
+                                                                OutputIndex,
+                                                                64, // BlockSize
+                                                                64, // ReduceMThreadClusterSize
+                                                                1,  // ReduceKThreadClusterSize
+                                                                4,  // ReduceMThreadSliceSize
+                                                                1,  // ReduceKThreadSliceSize
+                                                                1>; // InSrcOutDstVectorSize
+
+    const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
+    const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
+    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
+    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
+
+    const std::vector<ck::index_t> window_spatial_lengths{Y, X};
+    const std::vector<ck::index_t> window_strides{window_stride_h, window_stride_w};
+    const std::vector<ck::index_t> window_dilations{window_dilation_h, window_dilation_w};
+    const std::vector<ck::index_t> input_left_pads{in_left_pad_h, in_left_pad_w};
+    const std::vector<ck::index_t> input_right_pads{in_right_pad_h, in_right_pad_w};
+
+    // tensor layout
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
+            using namespace ck::literals;
+
+            if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value)
+            {
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, H * W, W, 1_uz});
+            }
+            else if constexpr(ck::is_same<decltype(layout),
+                                          ck::tensor_layout::convolution::NHWC>::value)
+            {
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, 1_uz, W * C_, C_});
+            }
+        };
+
+    Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
+    Tensor<OutDataType> out_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+    Tensor<IndexDataType> out_indices_n_c_ho_wo_host(
+        f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+    Tensor<OutDataType> out_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+    Tensor<IndexDataType> out_indices_n_c_ho_wo_device(
+        f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "out_n_c_ho_wo: " << out_n_c_ho_wo_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}); break;
+    case 2: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
+    default: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) *
+                             out_n_c_ho_wo_device.mDesc.GetElementSpaceSize());
+    DeviceMem out_indices_device_buf(sizeof(IndexDataType) *
+                                     out_indices_n_c_ho_wo_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+
+    auto pool         = DevicePoolFwdInstance{};
+    auto invoker_ptr  = pool.MakeInvokerPointer();
+    auto argument_ptr = pool.MakeArgumentPointer(
+        static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+        static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
+        {N, C, Hi, Wi},
+        {Y, X},
+        {N, C, Ho, Wo},
+        {C * Hi * Wi, 1, Wi * C, C},
+        {C * Ho * Wo, 1, Wo * C, C},
+        {C * Ho * Wo, 1, Wo * C, C},
+        window_strides,
+        window_dilations,
+        input_left_pads,
+        input_right_pads,
+        {2, 3});
+
+    if(!pool.IsSupportedArgument(argument_ptr.get()))
+    {
+        throw std::runtime_error("wrong! device_op with the specified compilation parameters does "
+                                 "not support this problem");
+    }
+
+    float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * N * C * Ho * Wo * Y * X;
+
+    std::size_t num_btype =
+        sizeof(InDataType) * (N * C * Hi * Wi) + sizeof(OutDataType) * (N * C * Ho * Wo);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+              << " GB / s " << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        using ReferencePoolingFwdInstance =
+            ck::tensor_operation::host::ReferencePoolingFwd<4,
+                                                            2,
+                                                            InDataType,
+                                                            OutDataType,
+                                                            ComputeDataType,
+                                                            IndexDataType,
+                                                            ReduceOpId,
+                                                            PropagateNan,
+                                                            OutputIndex>;
+
+        auto ref_pooling          = ReferencePoolingFwdInstance{};
+        auto ref_pooling_invoker  = ref_pooling.MakeInvoker();
+        auto ref_pooling_argument = ref_pooling.MakeArgument(in_n_c_hi_wi,
+                                                             out_n_c_ho_wo_host,
+                                                             out_indices_n_c_ho_wo_host,
+                                                             window_spatial_lengths,
+                                                             window_strides,
+                                                             window_dilations,
+                                                             input_left_pads,
+                                                             input_right_pads);
+
+        ref_pooling_invoker.Run(ref_pooling_argument);
+
+        out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
+
+        pass = pass && ck::utils::check_err(out_n_c_ho_wo_device, out_n_c_ho_wo_host);
+
+        if constexpr(OutputIndex)
+        {
+            out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data());
+
+            pass = pass &&
+                   ck::utils::check_err(out_indices_n_c_ho_wo_device, out_indices_n_c_ho_wo_host);
+        };
+    }
+
+    return (pass);
+};
diff --git a/example/13_pool2d_fwd/pool2d_fwd_fp16.cpp b/example/13_pool2d_fwd/pool2d_fwd_fp16.cpp
new file mode 100644
index 0000000000..d767e92248
--- /dev/null
+++ b/example/13_pool2d_fwd/pool2d_fwd_fp16.cpp
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "pool2d_fwd_common.hpp"
+
+using InDataType      = ck::half_t;
+using OutDataType     = ck::half_t;
+using ComputeDataType = float;
+
+using IndexDataType = int32_t;
+
+using InLayout  = ck::tensor_layout::convolution::NHWC;
+using OutLayout = ck::tensor_layout::convolution::NHWC;
+
+#if 1
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+#else
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+#endif
+
+static constexpr bool OutputIndex  = false;
+static constexpr bool PropagateNan = false;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification;
+    int init_method;
+    bool time_kernel;
+
+    // Pool shape
+    ck::index_t N                 = 128;
+    ck::index_t C                 = 192;
+    ck::index_t Y                 = 3;
+    ck::index_t X                 = 3;
+    ck::index_t Hi                = 71;
+    ck::index_t Wi                = 71;
+    ck::index_t window_stride_h   = 2;
+    ck::index_t window_stride_w   = 2;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
+    ck::index_t in_left_pad_h     = 1;
+    ck::index_t in_left_pad_w     = 1;
+    ck::index_t in_right_pad_h    = 1;
+    ck::index_t in_right_pad_w    = 1;
+
+    if(argc == 1)
+    {
+        do_verification = true;
+        init_method     = 1;
+        time_kernel     = true;
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+    }
+    else if(argc == 18)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+
+        N                 = std::stoi(argv[4]);
+        C                 = std::stoi(argv[5]);
+        Y                 = std::stoi(argv[6]);
+        X                 = std::stoi(argv[7]);
+        Hi                = std::stoi(argv[8]);
+        Wi                = std::stoi(argv[9]);
+        window_stride_h   = std::stoi(argv[10]);
+        window_stride_w   = std::stoi(argv[11]);
+        window_dilation_h = std::stoi(argv[12]);
+        window_dilation_w = std::stoi(argv[13]);
+        in_left_pad_h     = std::stoi(argv[14]);
+        in_left_pad_w     = std::stoi(argv[15]);
+        in_right_pad_h    = std::stoi(argv[16]);
+        in_right_pad_w    = std::stoi(argv[17]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
+               "RightPx\n");
+        exit(0);
+    }
+
+    bool pass = pool_test<InDataType,
+                          OutDataType,
+                          ComputeDataType,
+                          IndexDataType,
+                          InLayout,
+                          OutLayout,
+                          ReduceOpId,
+                          PropagateNan,
+                          OutputIndex>(do_verification,
+                                       init_method,
+                                       time_kernel,
+                                       N,
+                                       C,
+                                       Y,
+                                       X,
+                                       Hi,
+                                       Wi,
+                                       window_stride_h,
+                                       window_stride_w,
+                                       window_dilation_h,
+                                       window_dilation_w,
+                                       in_left_pad_h,
+                                       in_left_pad_w,
+                                       in_right_pad_h,
+                                       in_right_pad_w);
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/13_pool2d_fwd/pool2d_fwd_fp32.cpp b/example/13_pool2d_fwd/pool2d_fwd_fp32.cpp
new file mode 100644
index 0000000000..2621500ef1
--- /dev/null
+++ b/example/13_pool2d_fwd/pool2d_fwd_fp32.cpp
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "pool2d_fwd_common.hpp"
+
+using InDataType      = float;
+using OutDataType     = float;
+using ComputeDataType = float;
+
+using IndexDataType = int32_t;
+
+using InLayout  = ck::tensor_layout::convolution::NHWC;
+using OutLayout = ck::tensor_layout::convolution::NHWC;
+
+#if 1
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+#else
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+#endif
+
+static constexpr bool OutputIndex  = false;
+static constexpr bool PropagateNan = false;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification;
+    int init_method;
+    bool time_kernel;
+
+    // Pool shape
+    ck::index_t N                 = 128;
+    ck::index_t C                 = 192;
+    ck::index_t Y                 = 3;
+    ck::index_t X                 = 3;
+    ck::index_t Hi                = 71;
+    ck::index_t Wi                = 71;
+    ck::index_t window_stride_h   = 2;
+    ck::index_t window_stride_w   = 2;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
+    ck::index_t in_left_pad_h     = 1;
+    ck::index_t in_left_pad_w     = 1;
+    ck::index_t in_right_pad_h    = 1;
+    ck::index_t in_right_pad_w    = 1;
+
+    if(argc == 1)
+    {
+        do_verification = true;
+        init_method     = 1;
+        time_kernel     = true;
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+    }
+    else if(argc == 18)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = static_cast<bool>(std::stoi(argv[3]));
+
+        N                 = std::stoi(argv[4]);
+        C                 = std::stoi(argv[5]);
+        Y                 = std::stoi(argv[6]);
+        X                 = std::stoi(argv[7]);
+        Hi                = std::stoi(argv[8]);
+        Wi                = std::stoi(argv[9]);
+        window_stride_h   = std::stoi(argv[10]);
+        window_stride_w   = std::stoi(argv[11]);
+        window_dilation_h = std::stoi(argv[12]);
+        window_dilation_w = std::stoi(argv[13]);
+        in_left_pad_h     = std::stoi(argv[14]);
+        in_left_pad_w     = std::stoi(argv[15]);
+        in_right_pad_h    = std::stoi(argv[16]);
+        in_right_pad_w    = std::stoi(argv[17]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
+               "RightPx\n");
+        exit(0);
+    }
+
+    bool pass = pool_test<InDataType,
+                          OutDataType,
+                          ComputeDataType,
+                          IndexDataType,
+                          InLayout,
+                          OutLayout,
+                          ReduceOpId,
+                          PropagateNan,
+                          OutputIndex>(do_verification,
+                                       init_method,
+                                       time_kernel,
+                                       N,
+                                       C,
+                                       Y,
+                                       X,
+                                       Hi,
+                                       Wi,
+                                       window_stride_h,
+                                       window_stride_w,
+                                       window_dilation_h,
+                                       window_dilation_w,
+                                       in_left_pad_h,
+                                       in_left_pad_w,
+                                       in_right_pad_h,
+                                       in_right_pad_w);
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/14_gemm_quantization/CMakeLists.txt b/example/14_gemm_quantization/CMakeLists.txt
new file mode 100644
index 0000000000..3b3ad80dd8
--- /dev/null
+++ b/example/14_gemm_quantization/CMakeLists.txt
@@ -0,0 +1,17 @@
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+# dlops
+if(DL_KERNELS)
+  add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
+endif()
+
+# xdlops
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_example_executable(example_gemm_xdl_bias_relu_quantization_int8 gemm_xdl_bias_relu_quantization_int8.cpp)
+   add_example_executable(example_gemm_xdl_quantization_int8 gemm_xdl_quantization_int8.cpp)
+   set(target 1)
+ endif()
+endforeach()
+endif()
\ No newline at end of file
diff --git a/example/14_gemm_quantization/gemm_dl_quantization_int8.cpp b/example/14_gemm_quantization/gemm_dl_quantization_int8.cpp
new file mode 100644
index 0000000000..2585072dfe
--- /dev/null
+++ b/example/14_gemm_quantization/gemm_dl_quantization_int8.cpp
@@ -0,0 +1,204 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using I8  = int8_t;
+using I32 = int32_t;
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using ActivationOp = PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
+
+using ADataType        = I8;
+using BDataType        = I8;
+using AccDataType      = I32;
+using CShuffleDataType = I32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = I8;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Dl<
+    ALayout,
+    BLayout,
+    DsLayout,
+    ELayout,
+    ADataType,
+    BDataType,
+    AccDataType,
+    DsDataType,
+    EDataType,
+    AElementOp,
+    BElementOp,
+    CDEElementOp,
+    GemmDefault,
+    256,                 // BlockSize
+    128,                 // MPerBlock
+    128,                 // NPerBlock
+    16,                  // K0PerBlock
+    4,                   // K1
+    4,                   // M1PerThread
+    4,                   // N1PerThread
+    1,                   // KPerThread
+    S<8, 2>,             // M1N1ThreadClusterM1Xs
+    S<8, 2>,             // M1N1ThreadClusterN1Xs
+    S<8, 1, 1, 4>,       // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
+    S<2, 1, 128, 1>,     // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
+    S<1, 2, 0, 3>,       // ABlockTransferThreadClusterArrangeOrder
+    S<1, 2, 0, 3>,       // ABlockTransferSrcAccessOrder
+    S<4, 1, 1, 4>,       // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
+    S<1, 2, 0, 3>,       // ABlockTransferSrcVectorTensorContiguousDimOrder
+    S<1, 1, 1, 4>,       // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
+    S<8, 1, 1, 4>,       // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
+    S<2, 1, 128, 1>,     // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
+    S<1, 2, 0, 3>,       // BBlockTransferThreadClusterArrangeOrder
+    S<1, 2, 0, 3>,       // BBlockTransferSrcAccessOrder
+    S<4, 1, 1, 4>,       // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
+    S<1, 2, 0, 3>,       // BBlockTransferSrcVectorTensorContiguousDimOrder
+    S<1, 1, 1, 4>,       // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
+    S<0, 1, 2, 3, 4, 5>, // CThreadTransferSrcDstAccessOrder
+    5,                   // CThreadTransferSrcDstVectorDim
+    4>;                  // CThreadTransferDstScalarPerVector
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, EDataType, float, PassThrough, PassThrough, CDEElementOp>;
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
+
+    float requant_scale = 0.03;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({stride, 1_uz}));
+            }
+            else
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({1_uz, stride}));
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+    b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                      b_device_buf.GetDeviceBuffer(),
+                                      {},
+                                      e_device_buf.GetDeviceBuffer(),
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      {},
+                                      StrideE,
+                                      a_element_op,
+                                      b_element_op,
+                                      cde_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, e_m_n_host_result, a_element_op, b_element_op, cde_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp b/example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp
new file mode 100644
index 0000000000..aa3e011695
--- /dev/null
+++ b/example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp
@@ -0,0 +1,235 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using I8  = int8_t;
+using I32 = int32_t;
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using CDEElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
+
+using ADataType        = I8;
+using BDataType        = I8;
+using AccDataType      = I32;
+using CShuffleDataType = I32;
+using BiasDataType     = I32;
+using DsDataType       = ck::Tuple<BiasDataType>;
+using EDataType        = I8;
+
+using ALayout    = Row;
+using BLayout    = Col;
+using BiasLayout = Row;
+using DsLayout   = ck::Tuple<BiasLayout>;
+using ELayout    = Row;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<
+     ALayout,
+     BLayout,
+     DsLayout,
+     ELayout,
+     ADataType,
+     BDataType,
+     AccDataType,
+     CShuffleDataType,
+     DsDataType,
+     EDataType,
+     PassThrough,                // AElementwiseOperation,
+     PassThrough,                // BElementwiseOperation,
+     CDEElementOp,               // CDEElementwiseOperation,
+     GemmDefault,                // GemmSpecialization GemmSpec,
+     1,                          // NumGemmKPrefetchStage,
+     256,                        // BlockSize,
+     256,                        // MPerBlock,
+     128,                        // NPerBlock,
+     64,                         // KPerBlock,
+     16,                         // AK1,
+     16,                         // BK1,
+     32,                         // MPerXDL,
+     32,                         // NPerXDL,
+     4,                          // MXdlPerWave,
+     2,                          // NXdlPerWave,
+     S<4, 64, 1>,                // ABlockTransferThreadClusterLengths_AK0_M_AK1,
+     S<1, 0, 2>,                 // ABlockTransferThreadClusterArrangeOrder,
+     S<1, 0, 2>,                 // ABlockTransferSrcAccessOrder,
+     2,                          // index_t ABlockTransferSrcVectorDim,
+     16,                         // index_t ABlockTransferSrcScalarPerVector,
+     16,                         // index_t ABlockTransferDstScalarPerVector_AK1,
+     1,                          // bool ABlockLdsExtraM,
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder,
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder,
+     2,                          // index_t BBlockTransferSrcVectorDim,
+     8,                          // index_t BBlockTransferSrcScalarPerVector,
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1,
+     1,                          // bool BBlockLdsExtraN,
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle,
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle,
+     S<1, 64, 1, 4>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+     8>;                         // index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        AccDataType,
+                                                                        AccDataType,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        PassThrough>;
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA    = 1024;
+    ck::index_t StrideB    = 1024;
+    ck::index_t StrideBias = 0;
+    ck::index_t StrideE    = 1024;
+
+    float requant_scale = 0.03;
+
+    auto f_host_tensor_descriptor2d =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({stride, 1_uz}));
+            }
+            else
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({1_uz, stride}));
+            }
+        };
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({len}),
+                                    std::vector<std::size_t>({stride}));
+    };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<BiasDataType> bias_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "bias_n: " << bias_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-128, 127});
+    b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-128, 127});
+    bias_n.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-128, 127});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    bias_device_buf.ToDevice(bias_n.mData.data());
+
+    auto a_element_op   = PassThrough{};
+    auto b_element_op   = PassThrough{};
+    auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                      b_device_buf.GetDeviceBuffer(),
+                                      {bias_device_buf.GetDeviceBuffer()},
+                                      e_device_buf.GetDeviceBuffer(),
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      {StrideBias},
+                                      StrideE,
+                                      a_element_op,
+                                      b_element_op,
+                                      cde_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), bias_n(n));
+            }
+        }
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp b/example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp
new file mode 100644
index 0000000000..4b207df5c6
--- /dev/null
+++ b/example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp
@@ -0,0 +1,207 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using I8  = int8_t;
+using I32 = int32_t;
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using ActivationOp = PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
+
+using ADataType        = I8;
+using BDataType        = I8;
+using AccDataType      = I32;
+using CShuffleDataType = I32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = I8;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmMultipleD_Xdl_CShuffle<
+     ALayout,
+     BLayout,
+     DsLayout,
+     ELayout,
+     ADataType,
+     BDataType,
+     AccDataType,
+     CShuffleDataType,
+     DsDataType,
+     EDataType,
+     PassThrough,                // AElementwiseOperation,
+     PassThrough,                // BElementwiseOperation,
+     CDEElementOp,               // CDEElementwiseOperation,
+     GemmDefault,                // GemmSpecialization GemmSpec,
+     1,                          // NumGemmKPrefetchStage,
+     256,                        // BlockSize,
+     256,                        // MPerBlock,
+     128,                        // NPerBlock,
+     64,                         // KPerBlock,
+     16,                         // AK1,
+     16,                         // BK1,
+     32,                         // MPerXDL,
+     32,                         // NPerXDL,
+     4,                          // MXdlPerWave,
+     2,                          // NXdlPerWave,
+     S<4, 64, 1>,                // ABlockTransferThreadClusterLengths_AK0_M_AK1,
+     S<1, 0, 2>,                 // ABlockTransferThreadClusterArrangeOrder,
+     S<1, 0, 2>,                 // ABlockTransferSrcAccessOrder,
+     2,                          // index_t ABlockTransferSrcVectorDim,
+     16,                         // index_t ABlockTransferSrcScalarPerVector,
+     16,                         // index_t ABlockTransferDstScalarPerVector_AK1,
+     1,                          // bool ABlockLdsExtraM,
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder,
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder,
+     2,                          // index_t BBlockTransferSrcVectorDim,
+     8,                          // index_t BBlockTransferSrcScalarPerVector,
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1,
+     1,                          // bool BBlockLdsExtraN,
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle,
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle,
+     S<1, 64, 1, 4>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+     16>;                        // index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, EDataType, float, PassThrough, PassThrough, CDEElementOp>;
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
+
+    float requant_scale = 0.03;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({stride, 1_uz}));
+            }
+            else
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({1_uz, stride}));
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-128, 127});
+    b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-128, 127});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    auto a_element_op   = PassThrough{};
+    auto b_element_op   = PassThrough{};
+    auto cde_element_op = CDEElementOp{requant_scale, ActivationOp{}};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                      b_device_buf.GetDeviceBuffer(),
+                                      {},
+                                      e_device_buf.GetDeviceBuffer(),
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      {},
+                                      StrideE,
+                                      a_element_op,
+                                      b_element_op,
+                                      cde_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, e_m_n_host_result, a_element_op, b_element_op, cde_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/15_grouped_gemm/CMakeLists.txt b/example/15_grouped_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..2f880af3cf
--- /dev/null
+++ b/example/15_grouped_gemm/CMakeLists.txt
@@ -0,0 +1,31 @@
+add_custom_target(example_grouped_gemm_xdl)
+
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+  add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
+  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp32)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
+  add_example_executable(example_grouped_gemm_multiple_d_dl_fp16 grouped_gemm_multiple_d_dl_fp16.cpp)
+  add_example_executable(example_grouped_gemm_xdl_splitk_fp16 grouped_gemm_xdl_splitk_fp16.cpp)
+  add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp16 grouped_gemm_xdl_fixed_nk_fp16.cpp)
+  add_example_executable(example_grouped_gemm_xdl_fixed_nk_bias_fp16 grouped_gemm_xdl_fixed_nk_bias_fp16.cpp)
+  add_dependencies(example_grouped_gemm_xdl 
+                   example_grouped_gemm_xdl_fp16
+                   example_grouped_gemm_multiple_d_dl_fp16
+                   example_grouped_gemm_xdl_splitk_fp16
+                   example_grouped_gemm_xdl_fixed_nk_fp16
+                   example_grouped_gemm_xdl_fixed_nk_bias_fp16)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_grouped_gemm_xdl_bfp16 grouped_gemm_xdl_bfp16.cpp)
+  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_bfp16)
+endif()
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+  add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
+  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
+endif()
+if(USE_BITINT_EXTENSION_INT4)
+  add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
+  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)
+endif()
diff --git a/example/15_grouped_gemm/README.md b/example/15_grouped_gemm/README.md
new file mode 100644
index 0000000000..c83b23e08c
--- /dev/null
+++ b/example/15_grouped_gemm/README.md
@@ -0,0 +1,25 @@
+# Instructions for ```example_grouped_gemm_xdl```
+
+## Run ```example_grouped_gemm_xdl```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: run kernel # of times (>1)
+./bin/example_grouped_gemm_xdl_fp16 0 1 5
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+gemm[0] a_m_k: dim 2, lengths {256, 64}, strides {64, 1} b_k_n: dim 2, lengths {64, 128}, strides {1, 64} c_m_n: dim 2, lengths {256, 128}, strides {128, 1}
+gemm[1] a_m_k: dim 2, lengths {512, 128}, strides {128, 1} b_k_n: dim 2, lengths {128, 256}, strides {1, 128} c_m_n: dim 2, lengths {512, 256}, strides {256, 1}
+gemm[2] a_m_k: dim 2, lengths {768, 192}, strides {192, 1} b_k_n: dim 2, lengths {192, 384}, strides {1, 192} c_m_n: dim 2, lengths {768, 384}, strides {384, 1}
+gemm[3] a_m_k: dim 2, lengths {1024, 256}, strides {256, 1} b_k_n: dim 2, lengths {256, 512}, strides {1, 256} c_m_n: dim 2, lengths {1024, 512}, strides {512, 1}
+group: 0 arg.a_grid_desc_k0_m_k1_{8, 256, 8}, arg.b_grid_desc_k0_n_k1_{8, 128, 8}, arg.c_grid_desc_m_n_{ 256, 128}
+group: 1 arg.a_grid_desc_k0_m_k1_{16, 512, 8}, arg.b_grid_desc_k0_n_k1_{16, 256, 8}, arg.c_grid_desc_m_n_{ 512, 256}
+group: 2 arg.a_grid_desc_k0_m_k1_{24, 768, 8}, arg.b_grid_desc_k0_n_k1_{24, 384, 8}, arg.c_grid_desc_m_n_{ 768, 384}
+group: 3 arg.a_grid_desc_k0_m_k1_{32, 1024, 8}, arg.b_grid_desc_k0_n_k1_{32, 512, 8}, arg.c_grid_desc_m_n_{ 1024, 512}
+launch_and_time_kernel: grid_dim {30, 1, 1}, block_dim {256, 1, 1} 
+Warm up
+Start running 5 times...
+Perf: 0.037887 ms, 11.0706 TFlops, 90.8132 GB/s, DeviceGroupedGemmXdl<256, 256, 128, 4, 8, 32, 32, 4, 2>
+```
diff --git a/example/15_grouped_gemm/grouped_gemm_multiple_d_dl_fp16.cpp b/example/15_grouped_gemm/grouped_gemm_multiple_d_dl_fp16.cpp
new file mode 100644
index 0000000000..3e1f7f0893
--- /dev/null
+++ b/example/15_grouped_gemm/grouped_gemm_multiple_d_dl_fp16.cpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cstddef>
+#include <initializer_list>
+#include <iostream>
+#include <numeric>
+#include <stdexcept>
+#include <string>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/utility/sequence.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType   = F16;
+using BDataType   = F16;
+using AccDataType = F32;
+using DsDataType  = ck::Tuple<>;
+using EDataType   = F16;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::
+        //   ##################| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|     DsData|     EData|           A|           B|          CDE|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //   ##################|        |        |         |        |      Type|      Type|        Type|       Type|      Type| Elementwise| Elementwise|  Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //   ##################|        |        |         |        |          |          |            |           |          |   Operation|   Operation|    Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //   ##################|        |        |         |        |          |          |            |           |          |            |            |             |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+  DeviceGroupedGemmMultipleD_Dl< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>;
+// clang-format on
+
+#include "run_grouped_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_bfp16.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_bfp16.cpp
new file mode 100644
index 0000000000..680cee1f81
--- /dev/null
+++ b/example/15_grouped_gemm/grouped_gemm_xdl_bfp16.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = BF16;
+using BDataType        = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = BF16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+#include "run_grouped_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp
new file mode 100644
index 0000000000..a193fc39ba
--- /dev/null
+++ b/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp
@@ -0,0 +1,353 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = F32;
+using DsDataType       = ck::Tuple<D0DataType>;
+using EDataType        = F32;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+using ELayout  = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+
+using CDEElementOp = Add;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl_Fixed_NK
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   128,    16,   128,    32,   8,   8,   16,   16,    1,    4,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              4>;
+// clang-format on
+
+struct ProblemSize final
+{
+    std::vector<ck::index_t> Ms;
+    std::vector<ck::index_t> Ns;
+    std::vector<ck::index_t> Ks;
+
+    std::vector<ck::index_t> stride_As;
+    std::vector<ck::index_t> stride_Bs;
+    std::vector<ck::index_t> stride_Cs;
+
+    ck::index_t group_count;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    int k_batch          = 1;
+};
+
+bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    auto group_count = problem_size.group_count;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    int sum_of_m = 0;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    std::vector<Tensor<BDataType>> b_tensors;
+    std::vector<Tensor<D0DataType>> d0_tensors;
+    std::vector<Tensor<EDataType>> c_host_tensors;
+    std::vector<Tensor<EDataType>> c_device_tensors;
+
+    a_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    d0_tensors.reserve(group_count);
+    c_host_tensors.reserve(group_count);
+    c_device_tensors.reserve(group_count);
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+
+    std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, d0_tensors_device,
+        c_tensors_device;
+
+    a_tensors_device.reserve(group_count);
+    b_tensors_device.reserve(group_count);
+    d0_tensors_device.reserve(group_count);
+    c_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(int i = 0; i < group_count; i++)
+    {
+        sum_of_m += problem_size.Ms[i];
+        a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ks[i], problem_size.stride_As[i], ALayout{})));
+        b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
+            problem_size.Ks[i], problem_size.Ns[i], problem_size.stride_Bs[i], BLayout{})));
+        d0_tensors.push_back(Tensor<D0DataType>(
+            f_host_tensor_descriptor(problem_size.Ms[i], problem_size.Ns[i], 0, ELayout{})));
+        c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
+        c_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
+        std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
+                  << " b_k_n: " << b_tensors[i].mDesc << " d_m_n: " << d0_tensors[i].mDesc
+                  << " c_m_n: " << c_device_tensors[i].mDesc << std::endl;
+
+        flop += std::size_t(2) * problem_size.Ms[i] * problem_size.Ks[i] * problem_size.Ns[i];
+        num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
+                     sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(D0DataType) * d0_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSize();
+
+        switch(config.init_method)
+        {
+        case 0: break;
+        case 1:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            break;
+        case 2:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+            break;
+        default:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        }
+
+        d0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+    }
+
+    using GroupedGemmKernelArgument = ck::tensor_operation::device::GroupedGemmKernelArgument<1>;
+
+    std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(ADataType) * sum_of_m * problem_size.Ks[i]));
+
+        b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(BDataType) * problem_size.Ns[i] * problem_size.Ks[i]));
+
+        d0_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(D0DataType) * problem_size.Ns[i]));
+
+        c_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
+
+        a_tensors_device[i]->ToDevice(a_tensors[i].mData.data(),
+                                      a_tensors[i].mDesc.GetElementSpaceSize() * sizeof(ADataType));
+        b_tensors_device[i]->ToDevice(b_tensors[i].mData.data(),
+                                      b_tensors[i].mDesc.GetElementSpaceSize() * sizeof(BDataType));
+        d0_tensors_device[i]->ToDevice(d0_tensors[i].mData.data());
+        c_tensors_device[i]->SetZero();
+
+        gemm_descs.push_back({sum_of_m,
+                              problem_size.Ns[i],
+                              problem_size.Ks[i],
+                              1,
+                              problem_size.stride_Bs[i],
+                              1,
+                              {0}});
+
+        grouped_gemm_kernel_args_.push_back(
+            {a_tensors_device[i]->GetDeviceBuffer(),
+             b_tensors_device[i]->GetDeviceBuffer(),
+             std::array<const void*, 1>{d0_tensors_device[i]->GetDeviceBuffer()},
+             c_tensors_device[i]->GetDeviceBuffer(),
+             problem_size.Ms[i],
+             problem_size.Ns[i],
+             problem_size.Ks[i],
+             problem_size.stride_As[i],
+             problem_size.stride_Bs[i],
+             std::array<ck::index_t, 1>{0},
+             problem_size.stride_Cs[i]});
+    }
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    std::vector<const void*> p_As                = {};
+    std::vector<const void*> p_Bs                = {};
+    std::vector<std::array<const void*, 1>> p_Ds = {};
+    std::vector<void*> p_Cs                      = {};
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(
+        p_As, p_Bs, p_Ds, p_Cs, gemm_descs, a_element_op, b_element_op, cde_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    DeviceMem gemm_workspace_dev(gemm.GetWorkSpaceSize(&argument));
+    gemm.SetWorkSpacePointer(&argument, gemm_workspace_dev.GetDeviceBuffer());
+
+    DeviceMem gemm_kernel_args_dev(gemm.GetDeviceKernelArgSize(&argument));
+    hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
+                              grouped_gemm_kernel_args_.data(),
+                              gemm.GetDeviceKernelArgSize(&argument),
+                              hipMemcpyHostToDevice));
+
+    gemm.SetDeviceKernelArgs(argument, gemm_kernel_args_dev.GetDeviceBuffer());
+    gemm.SetKBatch(argument, config.k_batch);
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    if(config.time_kernel)
+    {
+        float ave_time   = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+        float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << gemm.GetTypeString() << std::endl;
+    }
+
+    bool pass = true;
+    if(config.do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                EDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data(),
+                                            c_device_tensors[i].mDesc.GetElementSize() *
+                                                sizeof(EDataType));
+
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
+                                                      b_tensors[i],
+                                                      c_host_tensors[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      PassThrough{});
+
+            ref_invoker.Run(ref_argument);
+
+            for(int m = 0; m < problem_size.Ms[i]; ++m)
+            {
+                for(int n = 0; n < problem_size.Ns[i]; ++n)
+                {
+                    cde_element_op(
+                        c_host_tensors[i](m, n), c_host_tensors[i](m, n), d0_tensors[i](m, n));
+                }
+            }
+
+            pass &= ck::utils::check_err(c_device_tensors[i], c_host_tensors[i]);
+        }
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    problem_size.Ms = {0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0};
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ns.push_back(768);
+        problem_size.Ks.push_back(4608);
+
+        problem_size.stride_As.push_back(problem_size.Ks[i]);
+        problem_size.stride_Bs.push_back(problem_size.Ns[i]);
+        problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+    }
+
+    if(argc == 5)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+        config.k_batch         = std::stoi(argv[4]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        printf("arg4: k_batch (>0)\n");
+        exit(0);
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16.cpp
new file mode 100644
index 0000000000..89d4789c12
--- /dev/null
+++ b/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16.cpp
@@ -0,0 +1,329 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F32;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl_Fixed_NK
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>;
+// clang-format on
+
+struct ProblemSize final
+{
+    std::vector<ck::index_t> Ms;
+    std::vector<ck::index_t> Ns;
+    std::vector<ck::index_t> Ks;
+
+    std::vector<ck::index_t> stride_As;
+    std::vector<ck::index_t> stride_Bs;
+    std::vector<ck::index_t> stride_Cs;
+
+    ck::index_t group_count;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    int k_batch          = 1;
+    bool time_kernel     = false;
+};
+
+bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    auto group_count = problem_size.group_count;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+    std::vector<void*> p_Cs;
+
+    gemm_descs.reserve(group_count);
+
+    int sum_of_m = 0;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    std::vector<Tensor<BDataType>> b_tensors;
+    std::vector<Tensor<EDataType>> c_host_tensors;
+    std::vector<Tensor<EDataType>> c_device_tensors;
+
+    a_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    c_host_tensors.reserve(group_count);
+    c_device_tensors.reserve(group_count);
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+
+    std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, c_tensors_device;
+
+    a_tensors_device.reserve(group_count);
+    b_tensors_device.reserve(group_count);
+    c_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(int i = 0; i < group_count; i++)
+    {
+        sum_of_m += problem_size.Ms[i];
+        a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ks[i], problem_size.stride_As[i], ALayout{})));
+        b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
+            problem_size.Ks[i], problem_size.Ns[i], problem_size.stride_Bs[i], BLayout{})));
+        c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
+        c_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
+        std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
+                  << " b_k_n: " << b_tensors[i].mDesc << " c_m_n: " << c_device_tensors[i].mDesc
+                  << std::endl;
+
+        flop += std::size_t(2) * problem_size.Ms[i] * problem_size.Ks[i] * problem_size.Ns[i];
+        num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
+                     sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSize();
+
+        switch(config.init_method)
+        {
+        case 0: break;
+        case 1:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            break;
+        case 2:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+            break;
+        default:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        }
+    }
+
+    using GroupedGemmKernelArgument = ck::tensor_operation::device::GroupedGemmKernelArgument<>;
+
+    std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(ADataType) * sum_of_m * problem_size.Ks[i]));
+
+        b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(BDataType) * problem_size.Ns[i] * problem_size.Ks[i]));
+
+        c_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
+
+        a_tensors_device[i]->ToDevice(a_tensors[i].mData.data(),
+                                      a_tensors[i].mDesc.GetElementSpaceSize() * sizeof(ADataType));
+        b_tensors_device[i]->ToDevice(b_tensors[i].mData.data(),
+                                      b_tensors[i].mDesc.GetElementSpaceSize() * sizeof(BDataType));
+        c_tensors_device[i]->SetZero();
+
+        p_Cs.push_back(c_tensors_device[i]->GetDeviceBuffer());
+
+        gemm_descs.push_back({sum_of_m,
+                              problem_size.Ns[i],
+                              problem_size.Ks[i],
+                              1,
+                              problem_size.stride_Bs[i],
+                              1,
+                              {}});
+
+        grouped_gemm_kernel_args_.push_back({a_tensors_device[i]->GetDeviceBuffer(),
+                                             b_tensors_device[i]->GetDeviceBuffer(),
+                                             {},
+                                             c_tensors_device[i]->GetDeviceBuffer(),
+                                             problem_size.Ms[i],
+                                             problem_size.Ns[i],
+                                             problem_size.Ks[i],
+                                             problem_size.stride_As[i],
+                                             problem_size.stride_Bs[i],
+                                             {},
+                                             problem_size.stride_Cs[i]});
+    }
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    std::vector<const void*> p_As                = {};
+    std::vector<const void*> p_Bs                = {};
+    std::vector<std::array<const void*, 0>> p_Ds = {};
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(
+        p_As, p_Bs, p_Ds, p_Cs, gemm_descs, a_element_op, b_element_op, c_element_op);
+
+    DeviceMem gemm_arg_dev_mem(gemm.GetDeviceKernelArgSize(&argument));
+    DeviceMem gemm_workspace_dev(gemm.GetWorkSpaceSize(&argument));
+
+    gemm.SetWorkSpacePointer(&argument, gemm_workspace_dev.GetDeviceBuffer());
+
+    hip_check_error(hipMemcpy(gemm_arg_dev_mem.GetDeviceBuffer(),
+                              grouped_gemm_kernel_args_.data(),
+                              gemm.GetDeviceKernelArgSize(&argument),
+                              hipMemcpyHostToDevice));
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    gemm.SetDeviceKernelArgs(argument, gemm_arg_dev_mem.GetDeviceBuffer());
+    gemm.SetKBatch(argument, config.k_batch);
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    if(config.time_kernel)
+    {
+        float ave_time   = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+        float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << gemm.GetTypeString() << std::endl;
+    }
+
+    bool pass = true;
+    if(config.do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                EDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CDEElementOp>;
+
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data(),
+                                            c_device_tensors[i].mDesc.GetElementSize() *
+                                                sizeof(EDataType));
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
+                                                      b_tensors[i],
+                                                      c_host_tensors[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      c_element_op);
+
+            ref_invoker.Run(ref_argument);
+
+            pass &= ck::utils::check_err(c_device_tensors[i], c_host_tensors[i]);
+        }
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    problem_size.Ms = {
+        167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ns.push_back(768);
+        problem_size.Ks.push_back(4608);
+
+        problem_size.stride_As.push_back(problem_size.Ks[i]);
+        problem_size.stride_Bs.push_back(problem_size.Ks[i]);
+        problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+    }
+
+    if(argc == 5)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+        config.k_batch         = std::stoi(argv[4]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        printf("arg4: k_batch (> 0)\n");
+        exit(0);
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
new file mode 100644
index 0000000000..90a12bc1dd
--- /dev/null
+++ b/example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+#include "run_grouped_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp
new file mode 100644
index 0000000000..28b0fcd0ce
--- /dev/null
+++ b/example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F32;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 32, 1, 8>,               4>;
+// clang-format on
+
+#include "run_grouped_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp
new file mode 100644
index 0000000000..60c4a71a35
--- /dev/null
+++ b/example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = ck::int4_t;
+using BDataType        = ck::int4_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int32_t;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = ck::int4_t;
+
+using KernelADataType = int8_t;
+using KernelBDataType = int8_t;
+using KernelEDataType = int8_t;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
+    // clang-format off
+        < ALayout,              //ALayout
+          BLayout,              //BLayout
+          DsLayout,             //DsLayout
+          ELayout,              //ELayout
+          KernelADataType,      //ADataType    
+          KernelBDataType,      //BDataType   
+          AccDataType,          //AccDataType
+          CShuffleDataType,     //CShuffleDataType
+          DsDataType,           //DsDataType
+          KernelEDataType,      //EDataType
+          AElementOp,           //AElementwiseOperation
+          BElementOp,           //BElementwiseOperation
+          CDEElementOp,         //CDEElementwiseOperation
+          GemmDefault,          //GEMMSpecialization
+          1,                    // NumGemmKPrefetchStage
+          256,                  // BlockSize
+          256,                  // MPerBlock
+          128,                  // NPerBlock
+          64,                   // KPerBlock
+          16,                   // AK1
+          16,                   // BK1
+          32,                   // MPerXdl
+          32,                   // NPerXdl
+          4,                    // MXdlPerWave
+          2,                    // NXdlPerWave
+          S<4, 64, 1>,          // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+          S<1, 0, 2>,           // ABlockTransfer ThreadCluster ArrangeOrder
+          S<1, 0, 2>,           // ABlockTransfer SrcAccessOrder
+          2,                    // ABlockTransfer SrcVectorDim
+          16,                   // ABlockTransfer SrcScalarPerVector
+          16,                   // ABlockTransfer DstScalarPerVector_K1
+          1,                    // ABlockLdsExtraM
+          S<4, 64, 1>,          // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+          S<1, 0, 2>,           // BBlockTransfer ThreadCluster ArrangeOrder
+          S<1, 0, 2>,           // BBlockTransfer SrcAccessOrder
+          2,                    // BBlockTransfer SrcVectorDim
+          16,                   // BBlockTransfer SrcScalarPerVector
+          16,                   // BBlockTransfer DstScalarPerVector_K1
+          1,                    // BBlockLdsExtraN
+          1,                    // CShuffleMXdlPerWavePerShuffle
+          1,                    // CShuffleNXdlPerWavePerShuffle
+          S<1, 64, 1, 4>,       // CBlockTransferClusterLengths_MBlock_MWaveMPerXdl_NBlock_NWaveNPerXdl
+          16>;                  // CBlockTransferScalarPerVector_NWaveNPerXdl
+// clang-format on
+
+#define BUILD_INT4_EXAMPLE
+#include "run_grouped_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp
new file mode 100644
index 0000000000..0c96ef56d3
--- /dev/null
+++ b/example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = int8_t;
+using BDataType        = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int8_t;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = int8_t;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>;
+// clang-format on
+
+#include "run_grouped_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp
new file mode 100644
index 0000000000..743ab96be6
--- /dev/null
+++ b/example/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp
@@ -0,0 +1,97 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemmXdlSplitKCShuffle
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+#include "run_grouped_gemm_example.inc"
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    problem_size.Ms = {
+        167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ns.push_back(768);
+        problem_size.Ks.push_back(4608);
+
+        problem_size.stride_As.push_back(problem_size.Ks[i]);
+        problem_size.stride_Bs.push_back(problem_size.Ks[i]);
+        problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+    }
+
+    if(argc == 4)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        exit(0);
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}
diff --git a/example/15_grouped_gemm/run_grouped_gemm_example.inc b/example/15_grouped_gemm/run_grouped_gemm_example.inc
new file mode 100644
index 0000000000..320870e0de
--- /dev/null
+++ b/example/15_grouped_gemm/run_grouped_gemm_example.inc
@@ -0,0 +1,266 @@
+#pragma once
+
+struct ProblemSize final
+{
+    std::vector<ck::index_t> Ms;
+    std::vector<ck::index_t> Ns;
+    std::vector<ck::index_t> Ks;
+
+    std::vector<ck::index_t> stride_As;
+    std::vector<ck::index_t> stride_Bs;
+    std::vector<ck::index_t> stride_Cs;
+
+    ck::index_t group_count;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+    static_assert(sizeof(ADataType) == sizeof(KernelADataType));
+    static_assert(sizeof(BDataType) == sizeof(KernelBDataType));
+    static_assert(sizeof(EDataType) == sizeof(KernelEDataType));
+#endif
+    int group_count = problem_size.group_count;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+    std::vector<const void*> p_a, p_b;
+    std::vector<void*> p_c;
+
+    gemm_descs.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        int M = problem_size.Ms[i];
+        int N = problem_size.Ns[i];
+        int K = problem_size.Ks[i];
+
+        int stride_A = problem_size.stride_As[i];
+        int stride_B = problem_size.stride_Bs[i];
+        int stride_C = problem_size.stride_Cs[i];
+
+        gemm_descs.push_back({M, N, K, stride_A, stride_B, stride_C, {}});
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    std::vector<Tensor<BDataType>> b_tensors;
+    std::vector<Tensor<EDataType>> c_host_tensors;
+#ifdef BUILD_INT4_EXAMPLE
+    std::vector<Tensor<KernelEDataType>> c_device_tensors;
+#else
+    std::vector<Tensor<EDataType>> c_device_tensors;
+#endif
+
+    a_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    c_host_tensors.reserve(group_count);
+    c_device_tensors.reserve(group_count);
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+
+    std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, c_tensors_device;
+
+    a_tensors_device.reserve(group_count);
+    b_tensors_device.reserve(group_count);
+    c_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(std::size_t i = 0; i < gemm_descs.size(); i++)
+    {
+        a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].K_, gemm_descs[i].stride_A_, ALayout{})));
+        b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].K_, gemm_descs[i].N_, gemm_descs[i].stride_B_, BLayout{})));
+        c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
+#ifdef BUILD_INT4_EXAMPLE
+        c_device_tensors.push_back(Tensor<KernelEDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
+#else
+        c_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
+#endif
+        std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
+                  << " b_k_n: " << b_tensors[i].mDesc << " c_m_n: " << c_device_tensors[i].mDesc
+                  << std::endl;
+
+        flop += std::size_t(2) * gemm_descs[i].M_ * gemm_descs[i].K_ * gemm_descs[i].N_;
+        num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
+                     sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSize();
+
+        switch(config.init_method)
+        {
+        case 0: break;
+        case 1:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            break;
+        case 2:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+            break;
+        default:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        }
+    }
+
+    for(std::size_t i = 0; i < gemm_descs.size(); i++)
+    {
+        a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpaceSize()));
+        b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(BDataType) * b_tensors[i].mDesc.GetElementSpaceSize()));
+        c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSpaceSize()));
+
+#ifdef BUILD_INT4_EXAMPLE
+        const Tensor<KernelADataType> a_converted(a_tensors[i]);
+        const Tensor<KernelBDataType> b_converted(b_tensors[i]);
+
+        a_tensors_device[i]->ToDevice(a_converted.mData.data());
+        b_tensors_device[i]->ToDevice(b_converted.mData.data());
+#else
+        a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
+        b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
+        c_tensors_device[i]->SetZero();
+#endif
+
+        p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
+        p_b.push_back(b_tensors_device[i]->GetDeviceBuffer());
+        p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
+    }
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    std::vector<std::array<const void*, 0>> p_Ds = {};
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(
+        p_a, p_b, p_Ds, p_c, gemm_descs, a_element_op, b_element_op, c_element_op);
+
+    DeviceMem gemm_desc_workspace(gemm.GetWorkSpaceSize(&argument));
+
+    gemm.SetWorkSpacePointer(&argument, gemm_desc_workspace.GetDeviceBuffer());
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    bool pass = true;
+    if(config.do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                EDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CDEElementOp>;
+
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data());
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
+                                                      b_tensors[i],
+                                                      c_host_tensors[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      c_element_op);
+
+            ref_invoker.Run(ref_argument);
+
+#ifdef BUILD_INT4_EXAMPLE
+            const Tensor<EDataType> c_device_result_converted(c_device_tensors[i]);
+            pass &= ck::utils::check_err(c_device_result_converted, c_host_tensors[i]);
+
+#else
+            pass &= ck::utils::check_err(c_device_tensors[i], c_host_tensors[i]);
+#endif
+        }
+    }
+
+    if(config.time_kernel)
+    {
+        float ave_time   = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+        float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << gemm.GetTypeString() << std::endl;
+    }
+
+    return pass;
+}
+
+bool run_grouped_gemm_example(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ms.push_back(256 + 256 * i);
+        problem_size.Ns.push_back(128 + 128 * i);
+        problem_size.Ks.push_back(128 + 64 * i);
+
+        problem_size.stride_As.push_back(problem_size.Ks[i]);
+        problem_size.stride_Bs.push_back(problem_size.Ks[i]);
+        problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+    }
+
+    if(argc == 4)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        exit(0);
+    }
+
+    return run_grouped_gemm(problem_size, config);
+}
diff --git a/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt b/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
new file mode 100644
index 0000000000..00786d34a3
--- /dev/null
+++ b/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
@@ -0,0 +1,47 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_custom_target(example_gemm_reduce_xdl)
+   add_custom_target(example_gemm_reduce_xdl_max)
+   add_custom_target(example_gemm_reduce_xdl_mean_meansquare)
+   add_custom_target(example_gemm_add_add_mean_meansquare_xdl)
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_gemm_max_xdl_fp16 gemm_max_xdl_fp16.cpp)
+    add_example_executable(example_gemm_add_add_mean_meansquare_xdl_fp16 gemm_add_add_mean_meansquare_xdl_fp16.cpp)
+    add_example_executable(example_gemm_mean_meansquare_xdl_fp16 gemm_mean_meansquare_xdl_fp16.cpp)
+    add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp16)
+    add_dependencies(example_gemm_add_add_mean_meansquare_xdl example_gemm_add_add_mean_meansquare_xdl_fp16)
+    add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp16)
+   endif()
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+    add_example_executable(example_gemm_max_xdl_int8 gemm_max_xdl_int8.cpp)
+    add_example_executable(example_gemm_add_addsquare_xdl_int8 gemm_add_addsquare_xdl_int8.cpp)
+    add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int8)
+    add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_add_addsquare_xdl_int8)
+   endif()
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_gemm_max_xdl_fp32 gemm_max_xdl_fp32.cpp)
+    add_example_executable(example_gemm_mean_meansquare_xdl_fp32 gemm_mean_meansquare_xdl_fp32.cpp)
+    add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_fp32)
+    add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_fp32)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_gemm_max_xdl_bf16 gemm_max_xdl_bf16.cpp)
+    add_example_executable(example_gemm_mean_meansquare_xdl_bf16 gemm_mean_meansquare_xdl_bf16.cpp)
+    add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_bf16)
+    add_dependencies(example_gemm_reduce_xdl_mean_meansquare example_gemm_mean_meansquare_xdl_bf16)
+   endif()
+   
+   add_dependencies(example_gemm_reduce_xdl
+                 example_gemm_reduce_xdl_mean_meansquare
+                 example_gemm_reduce_xdl_max
+                 example_gemm_add_add_mean_meansquare_xdl)
+
+   if(USE_BITINT_EXTENSION_INT4)
+      add_example_executable(example_gemm_max_xdl_int4 gemm_max_xdl_int4.cpp)
+      add_dependencies(example_gemm_reduce_xdl_max example_gemm_max_xdl_int4)
+   endif()
+   set(target 1)
+ endif()
+endforeach()
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp
new file mode 100644
index 0000000000..2f6533d448
--- /dev/null
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp
@@ -0,0 +1,276 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+// DataType
+using ADataType         = F16;
+using BDataType         = F16;
+using GemmAccDataType   = F32;
+using CShuffleDataType  = F32;
+using D0DataType        = F16;
+using D1DataType        = F16;
+using DsDataType        = ck::Tuple<D0DataType, D1DataType>;
+using EDataType         = F16;
+using ReduceAccDataType = F32;
+using R0DataType        = F32;
+using R1DataType        = F32;
+using RsDataType        = ck::Tuple<R0DataType, R1DataType>;
+
+// Layout
+using ALayout  = Row;
+using BLayout  = Col;
+using D1Layout = Row;
+using ELayout  = D1Layout;
+
+// Elementwise op
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using AddAdd       = ck::tensor_operation::element_wise::AddAdd;
+using Square       = ck::tensor_operation::element_wise::UnarySquare;
+using Div          = ck::tensor_operation::element_wise::UnaryDivide;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddAdd;
+using QsElementOp  = ck::Tuple<PassThrough, Square>;
+using RsElementOp  = ck::Tuple<Div, Div>;
+
+// ReduceOp
+using R0ThreadReduceOp = ck::reduce::Add;
+using R1ThreadReduceOp = ck::reduce::Add;
+using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
+
+static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+//######| ALayout| BLayout| ELayout|     AData|     BData|     GemmAccData|         CShuffle|     DsData|     EData|     ReduceAccData|     RsData|           A|           B|          CDE|          Qs|          Rs|           Thread|           Global|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|    CDRThreadTransfer|                  CDE|    RThreadTransfer|
+//######|        |        |        |      Type|      Type|            Type|         DataType|       Type|      Type|              Type|       Type| Elementwise| Elementwise|  Elementwise| Elementwise| Elementwise|           Reduce|           Reduce| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|       ClusterLengths| ReduceThreadTransfer| DstScalarPerVector|
+//######|        |        |        |          |          |                |                 |           |          |                  |           |   Operation|   Operation|    Operation|   Operation|   Operation|        Operation|        Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _MPerBlock_NPerBlock|      ScalarPerVector|         _MPerBlock|
+//######|        |        |        |          |          |                |                 |           |          |                  |           |            |            |             |            |            |                 |                 |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |           _NPerBlock|                   |
+        < ALayout, BLayout, ELayout, ADataType, BDataType, GemmAccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType,  AElementOp,  BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<64, 4>,                    4,                  1>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        EDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+template <typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType,
+          typename R0DataType,
+          typename R1DataType>
+void DumpPerf(float ave_time, int M, int N, int K)
+{
+    std::size_t flop          = std::size_t(2) * M * N * K + std::size_t(2) * M * N;
+    std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                sizeof(D0DataType) * M * N + sizeof(D1DataType) * M * N +
+                                sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
+                                sizeof(R1DataType) * M;
+
+    float tflops          = static_cast<float>(flop) / 1.E9 / ave_time;
+    float gemm_gb_per_sec = gemm_num_byte / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gemm_gb_per_sec
+              << " GB/s, " << std::endl;
+}
+
+auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+    return HostTensorDescriptor({len}, {stride});
+};
+
+auto f_host_tensor_descriptor2d =
+    [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({row, col}, {stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({row, col}, {1_uz, stride});
+        }
+    };
+
+int main()
+{
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA  = 1024;
+    ck::index_t StrideB  = 1024;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 1024;
+    ck::index_t StrideE  = 1024;
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor2d(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> r0_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<R1DataType> r1_m(f_host_tensor_descriptor1d(M, 1));
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
+    b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
+    d0_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-1, 1});
+    d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-1, 1});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem r0_device_buf(sizeof(R0DataType) * r0_m.mDesc.GetElementSpaceSize());
+    DeviceMem r1_device_buf(sizeof(R1DataType) * r1_m.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_device_buf.ToDevice(d0_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+    auto qs_element_op  = QsElementOp{};
+    auto rs_element_op  = RsElementOp{N, N};
+
+    // Prepare GEMM, mean, mean_square
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               {r0_device_buf.GetDeviceBuffer(), r1_device_buf.GetDeviceBuffer()},
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               {StrideD0, StrideD1},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op,
+                               qs_element_op,
+                               rs_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    // init reducetion buffer to 0
+    r0_device_buf.SetZero();
+    r1_device_buf.SetZero();
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    bool do_verification = true;
+    bool pass            = true;
+
+    if(do_verification)
+    {
+        auto I0 = ck::Number<0>{};
+        auto I1 = ck::Number<1>{};
+
+        Tensor<EDataType> e_m_n_host(e_m_n.mDesc);
+        Tensor<R0DataType> r0_m_host(r0_m.mDesc);
+        Tensor<R1DataType> r1_m_host(r1_m.mDesc);
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, e_m_n_host, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        auto reduce0_op = R0ThreadReduceOp{};
+        auto reduce1_op = R1ThreadReduceOp{};
+
+        for(int m = 0; m < M; ++m)
+        {
+            auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+            auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
+
+            for(int n = 0; n < N; ++n)
+            {
+                ReduceAccDataType square_e_val;
+
+                auto e_val  = ck::type_convert<GemmAccDataType>(e_m_n_host(m, n));
+                auto d0_val = ck::type_convert<GemmAccDataType>(d0_n(n));
+                auto d1_val = ck::type_convert<GemmAccDataType>(d1_m_n(m, n));
+                cde_element_op(e_val, e_val, d0_val, d1_val);
+                e_m_n_host(m, n) = ck::type_convert<EDataType>(e_val);
+
+                auto e_val_reduce = ck::type_convert<ReduceAccDataType>(e_val);
+                qs_element_op[I1](square_e_val, e_val_reduce);
+
+                reduce0_op(reduce0_acc, e_val_reduce);
+                reduce1_op(reduce1_acc, square_e_val);
+            }
+
+            rs_element_op[I0](reduce0_acc, reduce0_acc);
+            rs_element_op[I1](reduce1_acc, reduce1_acc);
+            r0_m_host(m) = ck::type_convert<R0DataType>(reduce0_acc);
+            r1_m_host(m) = ck::type_convert<R1DataType>(reduce1_acc);
+        }
+
+        e_device_buf.FromDevice(e_m_n.mData.data());
+        r0_device_buf.FromDevice(r0_m.mData.data());
+        r1_device_buf.FromDevice(r1_m.mData.data());
+
+        pass = ck::utils::check_err(e_m_n, e_m_n_host, "Error: Incorrect results c", 1e-2, 1e-2);
+        pass &= ck::utils::check_err(r0_m, r0_m_host, "Error: Incorrect results d0", 1e-2, 1e-2);
+        pass &= ck::utils::check_err(r1_m, r1_m_host, "Error: Incorrect results d1", 1e-2, 1e-2);
+    }
+
+    bool time_kernel = true;
+    if(time_kernel)
+    {
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+        DumpPerf<ADataType, BDataType, D0DataType, D1DataType, EDataType, R0DataType, R1DataType>(
+            ave_time, M, N, K);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp
new file mode 100644
index 0000000000..b28e7f85d3
--- /dev/null
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp
@@ -0,0 +1,364 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// DataType
+using ADataType         = INT8;
+using BDataType         = INT8;
+using GemmAccDataType   = INT32;
+using CShuffleDataType  = INT32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = INT8;
+using ReduceAccDataType = INT32;
+using R0DataType        = INT32;
+using R1DataType        = INT32;
+using RsDataType        = ck::Tuple<R0DataType, R1DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using Square       = ck::tensor_operation::element_wise::UnarySquare;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough, Square>;
+using RsElementOp  = ck::Tuple<PassThrough, PassThrough>;
+
+// ReduceOp
+using R0ThreadReduceOp = ck::reduce::Add;
+using R1ThreadReduceOp = ck::reduce::Add;
+using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
+
+static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         64,                        // KPerBlock
+         16,                        // AK1
+         16,                        // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         16,                        // ABlockTransfer SrcScalarPerVector
+         16,                        // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         16,                        // BBlockTransfer SrcScalarPerVector
+         16,                        // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+using namespace ck::literals;
+
+template <typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename R0DataType,
+          typename R1DataType,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp,
+          typename QsElementOp,
+          typename RsElementOp,
+          typename RsThreadReduceOp,
+          typename ReduceAccDataType,
+          typename DeviceOpInstance,
+          typename ReferenceGemmInstance>
+bool run_gemm_reduce_add_addsquare_xdl(ck::index_t M,
+                                       ck::index_t N,
+                                       ck::index_t K,
+                                       ck::index_t StrideA,
+                                       ck::index_t StrideB,
+                                       ck::index_t StrideE,
+                                       bool do_verification,
+                                       int init_method,
+                                       bool time_kernel)
+{
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    auto f_host_tensor_descriptor2d =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> r0_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<R1DataType> r1_m(f_host_tensor_descriptor1d(M, 1));
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
+        ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
+        break;
+    default:
+        ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
+        ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem r0_device_buf(sizeof(R0DataType) * r0_m.mDesc.GetElementSpaceSize());
+    DeviceMem r1_device_buf(sizeof(R1DataType) * r1_m.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+    auto qs_element_op  = QsElementOp{};
+    auto rs_element_op  = RsElementOp{};
+
+    // Prepare GEMM, add, add_square
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               {},
+                               e_device_buf.GetDeviceBuffer(),
+                               {r0_device_buf.GetDeviceBuffer(), r1_device_buf.GetDeviceBuffer()},
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               {},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op,
+                               qs_element_op,
+                               rs_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    // init reducetion buffer to 0
+    r0_device_buf.SetZero();
+    r1_device_buf.SetZero();
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        auto I0 = ck::Number<0>{};
+        auto I1 = ck::Number<1>{};
+
+        Tensor<ReduceAccDataType> e_m_n_host(e_m_n.mDesc);
+        Tensor<R0DataType> r0_m_host(r0_m.mDesc);
+        Tensor<R1DataType> r1_m_host(r1_m.mDesc);
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, e_m_n_host, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        auto reduce0_op = RsThreadReduceOp{}[I0];
+        auto reduce1_op = RsThreadReduceOp{}[I1];
+
+        for(int m = 0; m < M; ++m)
+        {
+            auto reduce0_acc = reduce0_op.template GetIdentityValue<ReduceAccDataType>();
+            auto reduce1_acc = reduce1_op.template GetIdentityValue<ReduceAccDataType>();
+
+            for(int n = 0; n < N; ++n)
+            {
+                ReduceAccDataType square_e_val;
+                auto e_val = ck::type_convert<ReduceAccDataType>(e_m_n_host(m, n));
+                qs_element_op[I1](square_e_val, e_val);
+
+                reduce0_op(reduce0_acc, e_val);
+                reduce1_op(reduce1_acc, square_e_val);
+            }
+
+            r0_m_host(m) = ck::type_convert<R0DataType>(reduce0_acc);
+            r1_m_host(m) = ck::type_convert<R1DataType>(reduce1_acc);
+        }
+        e_device_buf.FromDevice(e_m_n.mData.data());
+
+        Tensor<EDataType> e_m_n_host_converted(e_m_n_host);
+
+        pass = ck::utils::check_err(
+            e_m_n, e_m_n_host_converted, "Error: Incorrect results c", 1e-2, 1e-2);
+
+        r0_device_buf.FromDevice(r0_m.mData.data());
+        r1_device_buf.FromDevice(r1_m.mData.data());
+
+        pass &= ck::utils::check_err(r0_m, r0_m_host, "Error: Incorrect results d0", 1e-2, 1e-2);
+        pass &= ck::utils::check_err(r1_m, r1_m_host, "Error: Incorrect results d1", 1e-2, 1e-2);
+
+        if(pass)
+        {
+            std::cout << "Success!" << std::endl;
+        }
+    }
+
+    if(time_kernel)
+    {
+        float ave_time            = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+        std::size_t flop          = 2_uz * M * N * K + 3_uz * M * N;
+        std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                    sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
+                                    sizeof(R1DataType) * M;
+
+        float tflops          = static_cast<float>(flop) / 1.E9 / ave_time;
+        float gemm_gb_per_sec = gemm_num_byte / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gemm_gb_per_sec
+                  << " GB/s, " << std::endl;
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1152;
+    ck::index_t K = 512;
+
+    ck::index_t StrideA = 512;
+    ck::index_t StrideB = 512;
+    ck::index_t StrideE = 1152;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return !run_gemm_reduce_add_addsquare_xdl<ADataType,
+                                              BDataType,
+                                              EDataType,
+                                              R0DataType,
+                                              R1DataType,
+                                              ALayout,
+                                              BLayout,
+                                              ELayout,
+                                              AElementOp,
+                                              BElementOp,
+                                              CDEElementOp,
+                                              QsElementOp,
+                                              RsElementOp,
+                                              RsThreadReduceOp,
+                                              ReduceAccDataType,
+                                              DeviceOpInstance,
+                                              ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_bf16.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_bf16.cpp
new file mode 100644
index 0000000000..b30ce2c48a
--- /dev/null
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_bf16.cpp
@@ -0,0 +1,167 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// DataType
+using ADataType         = BF16;
+using BDataType         = BF16;
+using GemmAccDataType   = F32;
+using CShuffleDataType  = F32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = BF16;
+using ReduceAccDataType = F32;
+using R0DataType        = F32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough>;
+using RsElementOp  = ck::Tuple<PassThrough>;
+
+// ReduceOp
+using RsThreadReduceOp = ck::Tuple<ck::reduce::Max>;
+using RsGlobalReduceOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         32,                        // KPerBlock
+         8,                         // AK1
+         8,                         // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         8,                         // ABlockTransfer SrcScalarPerVector
+         8,                         // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         8,                         // BBlockTransfer SrcScalarPerVector
+         8,                         // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1152;
+    ck::index_t K = 256;
+
+    ck::index_t StrideA = 256;
+    ck::index_t StrideB = 256;
+    ck::index_t StrideE = 1152;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return run_gemm_reduce_max_xdl<ADataType,
+                                   BDataType,
+                                   EDataType,
+                                   R0DataType,
+                                   ALayout,
+                                   BLayout,
+                                   ELayout,
+                                   AElementOp,
+                                   BElementOp,
+                                   CDEElementOp,
+                                   QsElementOp,
+                                   RsElementOp,
+                                   RsThreadReduceOp,
+                                   ReduceAccDataType,
+                                   DeviceOpInstance,
+                                   ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp
new file mode 100644
index 0000000000..31e2efd6f6
--- /dev/null
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp
@@ -0,0 +1,167 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// DataType
+using ADataType         = F16;
+using BDataType         = F16;
+using GemmAccDataType   = F32;
+using CShuffleDataType  = F32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = F16;
+using ReduceAccDataType = F32;
+using R0DataType        = F32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough>;
+using RsElementOp  = ck::Tuple<PassThrough>;
+
+// ReduceOp
+using RsThreadReduceOp = ck::Tuple<ck::reduce::Max>;
+using RsGlobalReduceOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         32,                        // KPerBlock
+         8,                         // AK1
+         8,                         // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         8,                         // ABlockTransfer SrcScalarPerVector
+         8,                         // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         8,                         // BBlockTransfer SrcScalarPerVector
+         8,                         // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return run_gemm_reduce_max_xdl<ADataType,
+                                   BDataType,
+                                   EDataType,
+                                   R0DataType,
+                                   ALayout,
+                                   BLayout,
+                                   ELayout,
+                                   AElementOp,
+                                   BElementOp,
+                                   CDEElementOp,
+                                   QsElementOp,
+                                   RsElementOp,
+                                   RsThreadReduceOp,
+                                   ReduceAccDataType,
+                                   DeviceOpInstance,
+                                   ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp
new file mode 100644
index 0000000000..d3c7c1d99c
--- /dev/null
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp
@@ -0,0 +1,166 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// DataType
+using ADataType         = F32;
+using BDataType         = F32;
+using GemmAccDataType   = F32;
+using CShuffleDataType  = F32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = F32;
+using ReduceAccDataType = F32;
+using R0DataType        = F32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough>;
+using RsElementOp  = ck::Tuple<PassThrough>;
+
+// ReduceOp
+using RsThreadReduceOp = ck::Tuple<ck::reduce::Max>;
+using RsGlobalReduceOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         16,                        // KPerBlock
+         4,                         // AK1
+         4,                         // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         4,                         // ABlockTransfer SrcScalarPerVector
+         4,                         // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         4,                         // BBlockTransfer SrcScalarPerVector
+         4,                         // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: Measure kernel execution time (1=ON, 0=Off)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n");
+        exit(0);
+    }
+
+    return run_gemm_reduce_max_xdl<ADataType,
+                                   BDataType,
+                                   EDataType,
+                                   R0DataType,
+                                   ALayout,
+                                   BLayout,
+                                   ELayout,
+                                   AElementOp,
+                                   BElementOp,
+                                   CDEElementOp,
+                                   QsElementOp,
+                                   RsElementOp,
+                                   RsThreadReduceOp,
+                                   ReduceAccDataType,
+                                   DeviceOpInstance,
+                                   ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp
new file mode 100644
index 0000000000..9a4a6bc6e1
--- /dev/null
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp
@@ -0,0 +1,172 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType         = INT4;
+using ADataKernelType   = INT8;
+using BDataType         = INT4;
+using BDataKernelType   = INT8;
+using GemmAccDataType   = INT32;
+using CShuffleDataType  = INT32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = INT4;
+using EDataKernelType   = INT8;
+using ReduceAccDataType = INT32;
+using R0DataType        = INT32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough>;
+using RsElementOp  = ck::Tuple<PassThrough>;
+
+// ReduceOp
+using RsThreadReduceOp = ck::Tuple<ck::reduce::Max>;
+using RsGlobalReduceOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataKernelType,           // ADataType
+         BDataKernelType,           // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataKernelType,           // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         64,                        // KPerBlock
+         16,                        // AK1
+         16,                        // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         16,                        // ABlockTransfer SrcScalarPerVector
+         16,                        // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         16,                        // BBlockTransfer SrcScalarPerVector
+         16,                        // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1152;
+    ck::index_t K = 256;
+
+    ck::index_t StrideA = 256;
+    ck::index_t StrideB = 256;
+    ck::index_t StrideE = 1152;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return run_gemm_reduce_max_xdl<ADataType,
+                                   BDataType,
+                                   EDataType,
+                                   R0DataType,
+                                   ALayout,
+                                   BLayout,
+                                   ELayout,
+                                   AElementOp,
+                                   BElementOp,
+                                   CDEElementOp,
+                                   QsElementOp,
+                                   RsElementOp,
+                                   RsThreadReduceOp,
+                                   ReduceAccDataType,
+                                   DeviceOpInstance,
+                                   ReferenceGemmInstance,
+                                   ADataKernelType,
+                                   BDataKernelType,
+                                   EDataKernelType>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp
new file mode 100644
index 0000000000..1a8457a8bf
--- /dev/null
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp
@@ -0,0 +1,166 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType         = INT8;
+using BDataType         = INT8;
+using GemmAccDataType   = INT32;
+using CShuffleDataType  = INT32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = INT8;
+using ReduceAccDataType = INT32;
+using R0DataType        = INT32;
+using RsDataType        = ck::Tuple<R0DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough>;
+using RsElementOp  = ck::Tuple<PassThrough>;
+
+// ReduceOp
+using RsThreadReduceOp = ck::Tuple<ck::reduce::Max>;
+using RsGlobalReduceOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         64,                        // KPerBlock
+         16,                        // AK1
+         16,                        // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         16,                        // ABlockTransfer SrcScalarPerVector
+         16,                        // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         16,                        // BBlockTransfer SrcScalarPerVector
+         16,                        // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1152;
+    ck::index_t K = 512;
+
+    ck::index_t StrideA = 512;
+    ck::index_t StrideB = 512;
+    ck::index_t StrideE = 1152;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return run_gemm_reduce_max_xdl<ADataType,
+                                   BDataType,
+                                   EDataType,
+                                   R0DataType,
+                                   ALayout,
+                                   BLayout,
+                                   ELayout,
+                                   AElementOp,
+                                   BElementOp,
+                                   CDEElementOp,
+                                   QsElementOp,
+                                   RsElementOp,
+                                   RsThreadReduceOp,
+                                   ReduceAccDataType,
+                                   DeviceOpInstance,
+                                   ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp
new file mode 100644
index 0000000000..5c2706c79a
--- /dev/null
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp
@@ -0,0 +1,174 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// DataType
+using ADataType         = BF16;
+using BDataType         = BF16;
+using GemmAccDataType   = F32;
+using CShuffleDataType  = F32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = BF16;
+using ReduceAccDataType = F32;
+using R0DataType        = F32;
+using R1DataType        = F32;
+using RsDataType        = ck::Tuple<R0DataType, R1DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using Square       = ck::tensor_operation::element_wise::UnarySquare;
+using Div          = ck::tensor_operation::element_wise::UnaryDivide;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough, Square>;
+using RsElementOp  = ck::Tuple<Div, Div>;
+
+// ReduceOp
+using R0ThreadReduceOp = ck::reduce::Add;
+using R1ThreadReduceOp = ck::reduce::Add;
+using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
+
+static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         32,                        // KPerBlock
+         8,                         // AK1
+         8,                         // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         8,                         // ABlockTransfer SrcScalarPerVector
+         8,                         // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         8,                         // BBlockTransfer SrcScalarPerVector
+         8,                         // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1152;
+    ck::index_t K = 192;
+
+    ck::index_t StrideA = 192;
+    ck::index_t StrideB = 192;
+    ck::index_t StrideE = 1152;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return !run_gemm_reduce_mean_meansquare_xdl<ADataType,
+                                                BDataType,
+                                                EDataType,
+                                                R0DataType,
+                                                R1DataType,
+                                                ALayout,
+                                                BLayout,
+                                                ELayout,
+                                                AElementOp,
+                                                BElementOp,
+                                                CDEElementOp,
+                                                QsElementOp,
+                                                RsElementOp,
+                                                RsThreadReduceOp,
+                                                ReduceAccDataType,
+                                                DeviceOpInstance,
+                                                ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp
new file mode 100644
index 0000000000..c119e24370
--- /dev/null
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp
@@ -0,0 +1,174 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// DataType
+using ADataType         = F16;
+using BDataType         = F16;
+using GemmAccDataType   = F32;
+using CShuffleDataType  = F32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = F16;
+using ReduceAccDataType = F32;
+using R0DataType        = F32;
+using R1DataType        = F32;
+using RsDataType        = ck::Tuple<R0DataType, R1DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using Square       = ck::tensor_operation::element_wise::UnarySquare;
+using Div          = ck::tensor_operation::element_wise::UnaryDivide;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough, Square>;
+using RsElementOp  = ck::Tuple<Div, Div>;
+
+// ReduceOp
+using R0ThreadReduceOp = ck::reduce::Add;
+using R1ThreadReduceOp = ck::reduce::Add;
+using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
+
+static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         32,                        // KPerBlock
+         8,                         // AK1
+         8,                         // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         8,                         // ABlockTransfer SrcScalarPerVector
+         8,                         // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         8,                         // BBlockTransfer SrcScalarPerVector
+         8,                         // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return !run_gemm_reduce_mean_meansquare_xdl<ADataType,
+                                                BDataType,
+                                                EDataType,
+                                                R0DataType,
+                                                R1DataType,
+                                                ALayout,
+                                                BLayout,
+                                                ELayout,
+                                                AElementOp,
+                                                BElementOp,
+                                                CDEElementOp,
+                                                QsElementOp,
+                                                RsElementOp,
+                                                RsThreadReduceOp,
+                                                ReduceAccDataType,
+                                                DeviceOpInstance,
+                                                ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp
new file mode 100644
index 0000000000..0f5e588383
--- /dev/null
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp
@@ -0,0 +1,174 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_reduce_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// DataType
+using ADataType         = F32;
+using BDataType         = F32;
+using GemmAccDataType   = F32;
+using CShuffleDataType  = F32;
+using DsDataType        = ck::Tuple<>;
+using EDataType         = F32;
+using ReduceAccDataType = F32;
+using R0DataType        = F32;
+using R1DataType        = F32;
+using RsDataType        = ck::Tuple<R0DataType, R1DataType>;
+
+// Layout
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+// Elementwise op
+using Square       = ck::tensor_operation::element_wise::UnarySquare;
+using Div          = ck::tensor_operation::element_wise::UnaryDivide;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough, Square>;
+using RsElementOp  = ck::Tuple<Div, Div>;
+
+// ReduceOp
+using R0ThreadReduceOp = ck::reduce::Add;
+using R1ThreadReduceOp = ck::reduce::Add;
+using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
+
+static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+        <ALayout,                   // ALayout
+         BLayout,                   // BLayout
+         ELayout,                   // ELayout
+         ADataType,                 // ADataType
+         BDataType,                 // BDataType
+         GemmAccDataType,           // GemmAccDataType
+         CShuffleDataType,          // CShuffleDataType
+         DsDataType,                // DsDataType
+         EDataType,                 // EDataType
+         ReduceAccDataType,         // ReduceAccDataType
+         RsDataType,                // RsDataType
+         AElementOp,                // AElementwiseOperation
+         BElementOp,                // BElementwiseOperation
+         CDEElementOp,              // CDE ElementwiseOperation
+         QsElementOp,               // Qs Elementwise Operation
+         RsElementOp,               // Rs Elementwise Operation
+         RsThreadReduceOp,          // Thread Reduce Operation
+         RsGlobalReduceOp,          // Global Reduce Operation
+         GemmDefault,               // GEMM Specialization
+         1,                         // NumGemmKPrefetchStage
+         256,                       // BlockSize
+         256,                       // MPerBlock
+         128,                       // NPerBlock
+         16,                        // KPerBlock
+         4,                         // AK1
+         4,                         // BK1
+         32,                        // MPerXdl
+         32,                        // NPerXdl
+         4,                         // MXdlPerWave
+         2,                         // NXdlPerWave
+         S<4, 64, 1>,               // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<1, 0, 2>,                // ABlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // ABlockTransfer SrcAccessOrder
+         2,                         // ABlockTransfer SrcVectorDim
+         4,                         // ABlockTransfer SrcScalarPerVector
+         4,                         // ABlockTransfer DstScalarPerVector_K1
+         1,                         // ABlockLdsExtraM
+         S<4, 64, 1>,               // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<1, 0, 2>,                // BBlockTransfer ThreadCluster ArrangeOrder
+         S<1, 0, 2>,                // BBlockTransfer SrcAccessOrder
+         2,                         // BBlockTransfer SrcVectorDim
+         4,                         // BBlockTransfer SrcScalarPerVector
+         4,                         // BBlockTransfer DstScalarPerVector_K1
+         1,                         // BBlockLdsExtraN
+         1,                         // CShuffleMXdlPerWavePerShuffle
+         1,                         // CShuffleNXdlPerWavePerShuffle
+         S<64, 4>,                  // CD Reduce Thread Transfer ClusterLengths _MPerBlock_NPerBlock
+         4,                         // CDE ReduceThreadTransfer ScalarPerVector _NPerBlock
+         1>;                        // RThread DstScalarPerVector _MPerBlock
+// clang-format on
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        ReduceAccDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideE = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << " arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << " arg3: Measure kernel execution time (1=ON, 0=Off)\n"
+                  << " arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideE\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return !run_gemm_reduce_mean_meansquare_xdl<ADataType,
+                                                BDataType,
+                                                EDataType,
+                                                R0DataType,
+                                                R1DataType,
+                                                ALayout,
+                                                BLayout,
+                                                ELayout,
+                                                AElementOp,
+                                                BElementOp,
+                                                CDEElementOp,
+                                                QsElementOp,
+                                                RsElementOp,
+                                                RsThreadReduceOp,
+                                                ReduceAccDataType,
+                                                DeviceOpInstance,
+                                                ReferenceGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideE, do_verification, init_method, time_kernel);
+}
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp b/example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp
new file mode 100644
index 0000000000..1bea1bcf3e
--- /dev/null
+++ b/example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp
@@ -0,0 +1,491 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/host_utility/io.hpp"
+#include "ck/stream_config.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using F16         = ck::half_t;
+using BF16        = ck::bhalf_t;
+using F32         = float;
+using F64         = double;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+using INT4 = ck::int4_t;
+#endif
+using INT8  = std::int8_t;
+using INT32 = std::int32_t;
+
+template <typename ADataType, typename BDataType, typename EDataType, typename R0DataType>
+void DumpGemmReduceMaxPerf(float ave_time, int M, int N, int K)
+{
+    using namespace ck::literals;
+
+    std::size_t flop          = 2_uz * M * N * K;
+    std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                sizeof(EDataType) * M * N + sizeof(R0DataType) * M;
+
+    float tflops          = static_cast<float>(flop) / 1.E9 / ave_time;
+    float gemm_gb_per_sec = gemm_num_byte / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gemm_gb_per_sec
+              << " GB/s, " << std::endl;
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename R0DataType,
+          typename R1DataType>
+void DumpGemmReduceMeanSquareMeanPerf(float ave_time, int M, int N, int K)
+{
+    using namespace ck::literals;
+
+    std::size_t flop          = 2_uz * M * N * K + M * (3_uz * N + 2_uz);
+    std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
+                                sizeof(R1DataType) * M;
+
+    float tflops          = static_cast<float>(flop) / 1.E9 / ave_time;
+    float gemm_gb_per_sec = gemm_num_byte / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gemm_gb_per_sec
+              << " GB/s, " << std::endl;
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename R0DataType,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp,
+          typename QsElementOp,
+          typename RsElementOp,
+          typename RsThreadReduceOp,
+          typename ReduceAccDataType,
+          typename DeviceOpInstance,
+          typename ReferenceGemmInstance,
+          typename ADataKernelType = ADataType,
+          typename BDataKernelType = BDataType,
+          typename EDataKernelType = EDataType>
+auto run_gemm_reduce_max_xdl(ck::index_t M,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t StrideA,
+                             ck::index_t StrideB,
+                             ck::index_t StrideE,
+                             bool do_verification,
+                             int init_method,
+                             bool time_kernel)
+{
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+    static_assert(sizeof(ADataType) == sizeof(ADataKernelType));
+    static_assert(sizeof(BDataType) == sizeof(BDataKernelType));
+    static_assert(sizeof(EDataType) == sizeof(EDataKernelType));
+#endif
+    using namespace ck::literals;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    auto f_host_tensor_descriptor2d =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<EDataKernelType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> r0_m(f_host_tensor_descriptor1d(M, 1));
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
+        ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
+        break;
+    default:
+        ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
+        ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataKernelType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataKernelType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataKernelType) * e_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem r0_device_buf(sizeof(R0DataType) * r0_m.mDesc.GetElementSpaceSize());
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    if constexpr(std::is_same_v<ADataType, ck::int4_t>)
+    {
+        Tensor<ADataKernelType> a_m_k_converted = a_m_k.template CopyAsType<ADataKernelType>();
+        Tensor<BDataKernelType> b_k_n_converted = b_k_n.template CopyAsType<BDataKernelType>();
+
+        a_device_buf.ToDevice(a_m_k_converted.mData.data());
+        b_device_buf.ToDevice(b_k_n_converted.mData.data());
+    }
+    else
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    {
+        a_device_buf.ToDevice(a_m_k.mData.data());
+        b_device_buf.ToDevice(b_k_n.mData.data());
+    }
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+    auto qs_element_op  = QsElementOp{};
+    auto rs_element_op  = RsElementOp{};
+
+    // Prepare GEMM, max
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument  = device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                           b_device_buf.GetDeviceBuffer(),
+                                           {},
+                                           e_device_buf.GetDeviceBuffer(),
+                                           {r0_device_buf.GetDeviceBuffer()},
+                                           M,
+                                           N,
+                                           K,
+                                           StrideA,
+                                           StrideB,
+                                           {},
+                                           StrideE,
+                                           a_element_op,
+                                           b_element_op,
+                                           cde_element_op,
+                                           qs_element_op,
+                                           rs_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    // [CAUTION]: launch_and_time_kernel will not initialize D.
+    // If we evaluate kernel multiple time but without initialize D. Verification will fail
+    r0_device_buf.SetValue(ck::NumericLimits<R0DataType>::Lowest());
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        auto I0 = ck::Number<0>{};
+
+        Tensor<ReduceAccDataType> e_m_n_host(e_m_n.mDesc);
+        Tensor<R0DataType> r0_m_host(r0_m.mDesc);
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, e_m_n_host, a_element_op, b_element_op, cde_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        auto reduce0_op = RsThreadReduceOp{}[I0];
+
+        for(int m = 0; m < M; ++m)
+        {
+            auto reduce0_acc = reduce0_op.template GetIdentityValue<ReduceAccDataType>();
+
+            for(int n = 0; n < N; ++n)
+            {
+                auto e_val = e_m_n_host(m, n);
+                reduce0_op(reduce0_acc, e_val);
+            };
+
+            r0_m_host(m) = ck::type_convert<R0DataType>(reduce0_acc);
+        }
+
+        e_device_buf.FromDevice(e_m_n.mData.data());
+        Tensor<EDataType> e_m_n_host_converted(e_m_n_host);
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        if constexpr(std::is_same_v<ADataType, ck::int4_t>)
+        {
+            Tensor<EDataType> e_m_n_device_converted(e_m_n);
+            pass = ck::utils::check_err(e_m_n_device_converted,
+                                        e_m_n_host_converted,
+                                        "Error: Incorrect results c",
+                                        1e-2,
+                                        1e-2);
+        }
+        else
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        {
+            pass = ck::utils::check_err(
+                e_m_n, e_m_n_host_converted, "Error: Incorrect results c", 1e-2, 1e-2);
+        }
+
+        r0_device_buf.FromDevice(r0_m.mData.data());
+        pass &= ck::utils::check_err(r0_m, r0_m_host, "Error: Incorrect results d0", 1e-2, 1e-2);
+
+        if(pass)
+        {
+            std::cout << "Success!" << std::endl;
+        }
+    }
+
+    if(time_kernel)
+    {
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+        DumpGemmReduceMaxPerf<ADataType, BDataType, EDataType, R0DataType>(ave_time, M, N, K);
+    }
+
+    return pass ? 0 : 1;
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename R0DataType,
+          typename R1DataType,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp,
+          typename QsElementOp,
+          typename RsElementOp,
+          typename RsThreadReduceOp,
+          typename ReduceAccDataType,
+          typename DeviceOpInstance,
+          typename ReferenceGemmInstance,
+          typename ADataKernelType = ADataType,
+          typename BDataKernelType = BDataType,
+          typename EDataKernelType = EDataType>
+bool run_gemm_reduce_mean_meansquare_xdl(ck::index_t M,
+                                         ck::index_t N,
+                                         ck::index_t K,
+                                         ck::index_t StrideA,
+                                         ck::index_t StrideB,
+                                         ck::index_t StrideE,
+                                         bool do_verification,
+                                         int init_method,
+                                         bool time_kernel)
+{
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+    static_assert(sizeof(ADataType) == sizeof(ADataKernelType));
+    static_assert(sizeof(BDataType) == sizeof(BDataKernelType));
+    static_assert(sizeof(EDataType) == sizeof(EDataKernelType));
+#endif
+    using namespace ck::literals;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    auto f_host_tensor_descriptor2d =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<EDataKernelType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> r0_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<R1DataType> r1_m(f_host_tensor_descriptor1d(M, 1));
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
+        ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
+        break;
+    default:
+        ck::utils::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
+        ck::utils::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataKernelType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataKernelType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataKernelType) * e_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem r0_device_buf(sizeof(R0DataType) * r0_m.mDesc.GetElementSpaceSize());
+    DeviceMem r1_device_buf(sizeof(R1DataType) * r1_m.mDesc.GetElementSpaceSize());
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    if constexpr(std::is_same_v<ADataType, ck::int4_t>)
+    {
+        Tensor<ADataKernelType> a_m_k_converted = a_m_k.template CopyAsType<ADataKernelType>();
+        Tensor<BDataKernelType> b_k_n_converted = b_k_n.template CopyAsType<BDataKernelType>();
+
+        a_device_buf.ToDevice(a_m_k_converted.mData.data());
+        b_device_buf.ToDevice(b_k_n_converted.mData.data());
+    }
+    else
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    {
+        a_device_buf.ToDevice(a_m_k.mData.data());
+        b_device_buf.ToDevice(b_k_n.mData.data());
+    }
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+    auto qs_element_op  = QsElementOp{};
+    auto rs_element_op  = RsElementOp{N, N};
+
+    // Prepare GEMM, mean, mean_square
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               {},
+                               e_device_buf.GetDeviceBuffer(),
+                               {r0_device_buf.GetDeviceBuffer(), r1_device_buf.GetDeviceBuffer()},
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               {},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op,
+                               qs_element_op,
+                               rs_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    // init reducetion buffer to 0
+    r0_device_buf.SetZero();
+    r1_device_buf.SetZero();
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        auto I0 = ck::Number<0>{};
+        auto I1 = ck::Number<1>{};
+
+        Tensor<ReduceAccDataType> e_m_n_host(e_m_n.mDesc);
+        Tensor<R0DataType> r0_m_host(r0_m.mDesc);
+        Tensor<R1DataType> r1_m_host(r1_m.mDesc);
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, e_m_n_host, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        auto reduce0_op = RsThreadReduceOp{}[I0];
+        auto reduce1_op = RsThreadReduceOp{}[I1];
+
+        for(int m = 0; m < M; ++m)
+        {
+            auto reduce0_acc = reduce0_op.template GetIdentityValue<ReduceAccDataType>();
+            auto reduce1_acc = reduce1_op.template GetIdentityValue<ReduceAccDataType>();
+
+            for(int n = 0; n < N; ++n)
+            {
+                ReduceAccDataType square_e_val;
+                auto e_val = ck::type_convert<ReduceAccDataType>(e_m_n_host(m, n));
+                qs_element_op[I1](square_e_val, e_val);
+
+                reduce0_op(reduce0_acc, e_val);
+                reduce1_op(reduce1_acc, square_e_val);
+            }
+
+            rs_element_op[I0](reduce0_acc, reduce0_acc);
+            rs_element_op[I1](reduce1_acc, reduce1_acc);
+            r0_m_host(m) = ck::type_convert<R0DataType>(reduce0_acc);
+            r1_m_host(m) = ck::type_convert<R1DataType>(reduce1_acc);
+        }
+        e_device_buf.FromDevice(e_m_n.mData.data());
+        Tensor<EDataType> e_m_n_host_converted(e_m_n_host);
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        if constexpr(std::is_same_v<ADataType, ck::int4_t>)
+        {
+            Tensor<EDataType> e_m_n_device_converted(e_m_n);
+            pass = ck::utils::check_err(e_m_n_device_converted,
+                                        e_m_n_host_converted,
+                                        "Error: Incorrect results c",
+                                        1e-2,
+                                        1e-2);
+        }
+        else
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        {
+            pass = ck::utils::check_err(
+                e_m_n, e_m_n_host_converted, "Error: Incorrect results c", 1e-2, 1e-2);
+        }
+
+        r0_device_buf.FromDevice(r0_m.mData.data());
+        r1_device_buf.FromDevice(r1_m.mData.data());
+
+        pass &= ck::utils::check_err(r0_m, r0_m_host, "Error: Incorrect results d0", 1e-2, 1e-2);
+        pass &= ck::utils::check_err(r1_m, r1_m_host, "Error: Incorrect results d1", 1e-2, 1e-2);
+
+        if(pass)
+        {
+            std::cout << "Success!" << std::endl;
+        }
+    }
+
+    if(time_kernel)
+    {
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+        DumpGemmReduceMeanSquareMeanPerf<ADataType, BDataType, EDataType, R0DataType, R1DataType>(
+            ave_time, M, N, K);
+    }
+
+    return pass;
+}
diff --git a/example/17_convnd_bwd_data/CMakeLists.txt b/example/17_convnd_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000000..e187bd4337
--- /dev/null
+++ b/example/17_convnd_bwd_data/CMakeLists.txt
@@ -0,0 +1,15 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_example_executable(example_convnd_bwd_data_xdl_fp16 convnd_bwd_data_xdl_fp16.cpp)
+   target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility)
+   set(target 1)
+ endif()
+endforeach()
+  if(DL_KERNELS)
+    add_example_executable(example_convnd_bwd_data_dl_fp16 convnd_bwd_data_dl_fp16.cpp)
+    target_link_libraries(example_convnd_bwd_data_dl_fp16 PRIVATE utility)
+  endif()
+endif()
diff --git a/example/17_convnd_bwd_data/README.md b/example/17_convnd_bwd_data/README.md
new file mode 100644
index 0000000000..b5c8281ed8
--- /dev/null
+++ b/example/17_convnd_bwd_data/README.md
@@ -0,0 +1,47 @@
+# Instructions for ```example_convnd_bwd_data_xdl```
+
+## Run ```example_example_convnd_bwd_data_xdl```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: run kernel # of times (>1)
+#arg4: num_dim_spatial(1|2|3)
+#arg5 to ...: N, K, C, [Z,] [Y,] X, [Di,] [Hi,] Wi, S[z,] [Sy,] Sx, [Dz,] [Dy,] Dx, [LeftPz,] [LeftPy,] LeftPx, [RightPy,] [RightPy,] RightPx
+./bin/example_convnd_bwd_data_xdl 0 1 5 
+```
+
+Result
+```
+in_n_c_hi_wi: dim 4, lengths {128, 128, 71, 71}, strides {645248, 1, 9088, 128}
+wei_k_c_y_x: dim 4, lengths {256, 128, 3, 3}, strides {1152, 1, 384, 128}
+out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
+arg.a_grid_desc_k0_m_k1_container_{128, 175232, 8}
+arg.b_grid_desc_k0_n_k1_container_{128, 128, 8}
+arg.c_grid_desc_m_n_container_{ 175232, 128}
+arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 2, 2, 2, 4, 2 ) 
+launch_and_time_kernel: grid_dim {1369, 1, 1}, block_dim {256, 1, 1} 
+Warm up
+Start running 1 times...
+arg.a_grid_desc_k0_m_k1_container_{64, 175232, 8}
+arg.b_grid_desc_k0_n_k1_container_{64, 128, 8}
+arg.c_grid_desc_m_n_container_{ 175232, 128}
+arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 2, 2, 2, 4, 2 ) 
+launch_and_time_kernel: grid_dim {1369, 1, 1}, block_dim {256, 1, 1} 
+Warm up
+Start running 1 times...
+arg.a_grid_desc_k0_m_k1_container_{64, 175232, 8}
+arg.b_grid_desc_k0_n_k1_container_{64, 128, 8}
+arg.c_grid_desc_m_n_container_{ 175232, 128}
+arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 2, 2, 2, 4, 2 ) 
+launch_and_time_kernel: grid_dim {1369, 1, 1}, block_dim {256, 1, 1} 
+Warm up
+Start running 1 times...
+arg.a_grid_desc_k0_m_k1_container_{32, 175232, 8}
+arg.b_grid_desc_k0_n_k1_container_{32, 128, 8}
+arg.c_grid_desc_m_n_container_{ 175232, 128}
+arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 2, 2, 2, 4, 2 ) 
+launch_and_time_kernel: grid_dim {1369, 1, 1}, block_dim {256, 1, 1} 
+Warm up
+Start running 1 times...
+Perf: 1.40031 ms, 69.8734 TFlops, 179.037 GB/s
+```
diff --git a/example/17_convnd_bwd_data/convnd_bwd_data_common.hpp b/example/17_convnd_bwd_data/convnd_bwd_data_common.hpp
new file mode 100644
index 0000000000..4a9d16c5c3
--- /dev/null
+++ b/example/17_convnd_bwd_data/convnd_bwd_data_common.hpp
@@ -0,0 +1,153 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
+
+void print_helper_msg()
+{
+    std::cout << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNdBwdDataInstance>
+int run_conv_bwd_data(bool do_verification,
+                      int init_method,
+                      bool time_kernel,
+                      const ck::utils::conv::ConvParam& conv_param,
+                      const HostTensorDescriptor& in_g_n_c_wis_desc,
+                      const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                      const HostTensorDescriptor& out_g_n_k_wos_desc,
+                      const InElementOp& in_element_op,
+                      const WeiElementOp& wei_element_op,
+                      const OutElementOp& out_element_op)
+{
+    Tensor<InDataType> in_host(in_g_n_c_wis_desc);
+    Tensor<InDataType> in_device(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> out(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in_host.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "out: " << out.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    case 2:
+        out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+        break;
+    default:
+        out.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
+        wei.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_device.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
+
+    out_device_buf.ToDevice(out.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+
+    // reset input to zero
+    in_device_buf.SetZero();
+
+    // do GEMM
+    auto conv    = DeviceConvNdBwdDataInstance{};
+    auto invoker = conv.MakeInvoker();
+    auto argument =
+        conv.MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                 static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                 static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                 conv_param.N_,
+                                 conv_param.K_,
+                                 conv_param.C_,
+                                 conv_param.input_spatial_lengths_,
+                                 conv_param.filter_spatial_lengths_,
+                                 conv_param.GetOutputSpatialLengths(),
+                                 conv_param.conv_filter_strides_,
+                                 conv_param.conv_filter_dilations_,
+                                 conv_param.input_left_pads_,
+                                 conv_param.input_right_pads_,
+                                 in_element_op,
+                                 wei_element_op,
+                                 out_element_op);
+
+    if(!conv.IsSupportedArgument(argument.get()))
+    {
+        std::cout << "Not support,please check parameters or device";
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<NDimSpatial,
+                                                                         InDataType,
+                                                                         WeiDataType,
+                                                                         OutDataType,
+                                                                         InElementOp,
+                                                                         WeiElementOp,
+                                                                         OutElementOp>();
+
+        auto ref_invoker = ref_conv.MakeInvoker();
+
+        auto ref_argument = ref_conv.MakeArgument(in_host,
+                                                  wei,
+                                                  out,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        in_device_buf.FromDevice(in_device.mData.data());
+
+        return ck::utils::check_err(in_device, in_host) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp b/example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp
new file mode 100644
index 0000000000..6b84eaba47
--- /dev/null
+++ b/example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp
@@ -0,0 +1,180 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_bwd_data_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+template <ck::index_t NDimSpatial>
+// clang-format off
+using DeviceConvNdBwdDataInstance = ck::tensor_operation::device::DeviceConvNdBwdDataNwcKxcNwk_Dl<
+//        ######|       NDim|     InData|     WeiData|     OutData|     AccData|           In|           Wei|           Out|    Convolution| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+//        ######|    Spatial|       Type|        Type|        Type|        Type|  Elementwise|   Elementwise|   Elementwise|        Forward|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+//        ######|           |           |            |            |            |    Operation|     Operation|     Operation| Specialization|      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+//        ######|           |           |            |            |            |             |              |              |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+                 NDimSpatial, InDataType, WeiDataType, OutDataType, AccDataType,  InElementOp,  WeiElementOp,  OutElementOp, ConvBwdDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<1, 1, 8, 2>,      S<16, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 8, 1>,      S<0, 3, 1, 2>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    namespace ctc = ck::tensor_layout::convolution;
+
+    print_helper_msg();
+
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::utils::conv::ConvParam conv_param{
+        2, 1, 128, 256, 256, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        do_verification                   = std::stoi(argv[1]);
+        init_method                       = std::stoi(argv[2]);
+        time_kernel                       = std::stoi(argv[3]);
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+
+        conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
+    }
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    if(conv_param.num_dim_spatial_ == 1)
+    {
+        using InLayout  = ctc::GNWC;
+        using WeiLayout = ctc::GKXC;
+        using OutLayout = ctc::GNWK;
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_conv_bwd_data<1,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementOp,
+                                 WeiElementOp,
+                                 OutElementOp,
+                                 DeviceConvNdBwdDataInstance<1>>(do_verification,
+                                                                 init_method,
+                                                                 time_kernel,
+                                                                 conv_param,
+                                                                 in_g_n_c_wis_desc,
+                                                                 wei_g_k_c_xs_desc,
+                                                                 out_g_n_k_wos_desc,
+                                                                 in_element_op,
+                                                                 wei_element_op,
+                                                                 out_element_op);
+    }
+    else if(conv_param.num_dim_spatial_ == 2)
+    {
+        using InLayout  = ctc::GNHWC;
+        using WeiLayout = ctc::GKYXC;
+        using OutLayout = ctc::GNHWK;
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_conv_bwd_data<2,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementOp,
+                                 WeiElementOp,
+                                 OutElementOp,
+                                 DeviceConvNdBwdDataInstance<2>>(do_verification,
+                                                                 init_method,
+                                                                 time_kernel,
+                                                                 conv_param,
+                                                                 in_g_n_c_wis_desc,
+                                                                 wei_g_k_c_xs_desc,
+                                                                 out_g_n_k_wos_desc,
+                                                                 in_element_op,
+                                                                 wei_element_op,
+                                                                 out_element_op);
+    }
+    else if(conv_param.num_dim_spatial_ == 3)
+    {
+        using InLayout  = ctc::GNDHWC;
+        using WeiLayout = ctc::GKZYXC;
+        using OutLayout = ctc::GNDHWK;
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_conv_bwd_data<3,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementOp,
+                                 WeiElementOp,
+                                 OutElementOp,
+                                 DeviceConvNdBwdDataInstance<3>>(do_verification,
+                                                                 init_method,
+                                                                 time_kernel,
+                                                                 conv_param,
+                                                                 in_g_n_c_wis_desc,
+                                                                 wei_g_k_c_xs_desc,
+                                                                 out_g_n_k_wos_desc,
+                                                                 in_element_op,
+                                                                 wei_element_op,
+                                                                 out_element_op);
+    }
+
+    return 0;
+}
diff --git a/example/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp b/example/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp
new file mode 100644
index 0000000000..c9989c60ac
--- /dev/null
+++ b/example/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp
@@ -0,0 +1,207 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "convnd_bwd_data_common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+template <ck::index_t NDimSpatial>
+using DeviceConvNdBwdDataInstance = ck::tensor_operation::device::DeviceConvNdBwdDataNwcKxcNwk_Xdl<
+    NDimSpatial,    // NDimSpatial
+    InDataType,     // InDataType
+    WeiDataType,    // WeiDataType
+    OutDataType,    // OutDataType
+    AccDataType,    // AccDataType
+    InElementOp,    // InElementwiseOperation
+    WeiElementOp,   // WeiElementwiseOperation
+    OutElementOp,   // OutElementwiseOperation
+    ConvBwdDefault, // ConvolutionBackwardDataSpecialization
+    256,            // BlockSize
+    128,            // MPerBlock
+    128,            // NPerBlock
+    4,              // K0PerBlock
+    8,              // K1
+    32,             // MPerXdl
+    32,             // NPerXdl
+    2,              // MXdlPerWave
+    2,              // NXdlPerWave
+    S<4, 64, 1>,    // ABlockTransferThreadClusterLengths_K0_M_K1
+    S<1, 0, 2>,     // ABlockTransferThreadClusterArrangeOrder
+    S<1, 0, 2>,     // ABlockTransferSrcAccessOrder
+    2,              // ABlockTransferSrcVectorDim
+    8,              // ABlockTransferSrcScalarPerVector
+    8,              // ABlockTransferDstScalarPerVector_K1
+    true,           // ABlockLdsAddExtraM
+    S<4, 64, 1>,    // BBlockTransferThreadClusterLengths_K0_N_K1
+    S<2, 0, 1>,     // BBlockTransferThreadClusterArrangeOrder
+    S<0, 2, 1>,     // BBlockTransferSrcAccessOrder
+    1,              // BBlockTransferSrcVectorDim
+    2,              // BBlockTransferSrcScalarPerVector
+    8,              // BBlockTransferDstScalarPerVector_K1
+    true,           // BBlockLdsAddExtraN
+    7,
+    1>; // GemmCThreadTransferDstScalarPerVector
+
+int main(int argc, char* argv[])
+{
+    namespace ctc = ck::tensor_layout::convolution;
+
+    print_helper_msg();
+
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::utils::conv::ConvParam conv_param{
+        2, 1, 128, 256, 256, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}};
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        do_verification                   = std::stoi(argv[1]);
+        init_method                       = std::stoi(argv[2]);
+        time_kernel                       = std::stoi(argv[3]);
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+
+        conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
+    }
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    if(conv_param.num_dim_spatial_ == 1)
+    {
+        using InLayout  = ctc::GNWC;
+        using WeiLayout = ctc::GKXC;
+        using OutLayout = ctc::GNWK;
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_conv_bwd_data<1,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementOp,
+                                 WeiElementOp,
+                                 OutElementOp,
+                                 DeviceConvNdBwdDataInstance<1>>(do_verification,
+                                                                 init_method,
+                                                                 time_kernel,
+                                                                 conv_param,
+                                                                 in_g_n_c_wis_desc,
+                                                                 wei_g_k_c_xs_desc,
+                                                                 out_g_n_k_wos_desc,
+                                                                 in_element_op,
+                                                                 wei_element_op,
+                                                                 out_element_op);
+    }
+    else if(conv_param.num_dim_spatial_ == 2)
+    {
+        using InLayout  = ctc::GNHWC;
+        using WeiLayout = ctc::GKYXC;
+        using OutLayout = ctc::GNHWK;
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_conv_bwd_data<2,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementOp,
+                                 WeiElementOp,
+                                 OutElementOp,
+                                 DeviceConvNdBwdDataInstance<2>>(do_verification,
+                                                                 init_method,
+                                                                 time_kernel,
+                                                                 conv_param,
+                                                                 in_g_n_c_wis_desc,
+                                                                 wei_g_k_c_xs_desc,
+                                                                 out_g_n_k_wos_desc,
+                                                                 in_element_op,
+                                                                 wei_element_op,
+                                                                 out_element_op);
+    }
+    else if(conv_param.num_dim_spatial_ == 3)
+    {
+        using InLayout  = ctc::GNDHWC;
+        using WeiLayout = ctc::GKZYXC;
+        using OutLayout = ctc::GNDHWK;
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        return run_conv_bwd_data<3,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementOp,
+                                 WeiElementOp,
+                                 OutElementOp,
+                                 DeviceConvNdBwdDataInstance<3>>(do_verification,
+                                                                 init_method,
+                                                                 time_kernel,
+                                                                 conv_param,
+                                                                 in_g_n_c_wis_desc,
+                                                                 wei_g_k_c_xs_desc,
+                                                                 out_g_n_k_wos_desc,
+                                                                 in_element_op,
+                                                                 wei_element_op,
+                                                                 out_element_op);
+    }
+
+    return 0;
+}
diff --git a/example/18_batched_gemm_reduce/CMakeLists.txt b/example/18_batched_gemm_reduce/CMakeLists.txt
new file mode 100644
index 0000000000..a1bb398af0
--- /dev/null
+++ b/example/18_batched_gemm_reduce/CMakeLists.txt
@@ -0,0 +1,10 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_example_executable(example_batched_gemm_reduce_xdl_fp16 batched_gemm_reduce_xdl_fp16.cpp)
+   set(target 1)
+ endif()
+endforeach()
+endif()
diff --git a/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp b/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
new file mode 100644
index 0000000000..e363dc5c12
--- /dev/null
+++ b/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
@@ -0,0 +1,311 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using ADataType         = F16;
+using BDataType         = F16;
+using CDataType         = F16;
+using ReduceAccDataType = F32;
+using ReduceDataType    = F32;
+using ReducePtrsGlobal  = ck::Tuple<ReduceDataType*, ReduceDataType*>;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+using ReduceOp0  = ck::reduce::Add;
+using ReduceOp1  = ck::reduce::Add;
+using ReduceOps  = ck::Tuple<ReduceOp0, ReduceOp1>;
+
+using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
+using UnarySquareElementOp  = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps    = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
+using ReduceOutElementOps   = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
+
+using ReduceGlobalMemOps =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmSpecialization =
+    ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceBatchedGemmReduceInstance = ck::tensor_operation::device::DeviceBatchedGemmReduce_Xdl_CShuffle
+//######| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         DData|           A|           B|           C|         Dxs|     DxsInEleOp|     DxsAccEleOp|            D|               GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+//######|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|    Type Tuple| Elementwise| Elementwise| Elementwise|      Reduce|               |                |   MemoryData|     Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+//######|        |        |        |     |      |      |         |         |          |              |   Operation|   Operation|   Operation|   Operation|               |                |    Operation|                   |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+//######|        |        |        |     |      |      |         |         |          |              |            |            |            |            |               |                |             |                   |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        <     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal,  AElementOp,  BElementOp,  CElementOp, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceGlobalMemOps, GemmSpecialization,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>;
+// clang-format on
+
+using ReferenceBatchedGemmInstance =
+    ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                     BDataType,
+                                                     CDataType,
+                                                     ReduceAccDataType,
+                                                     AElementOp,
+                                                     BElementOp,
+                                                     CElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 2048;
+    ck::index_t N = 1920;
+    ck::index_t K = 2048;
+
+    ck::index_t StrideA = 2048;
+    ck::index_t StrideB = 2048;
+    ck::index_t StrideC = 1920;
+
+    ck::index_t BatchCount = 4;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 11)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+
+        BatchCount = std::stoi(argv[10]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        printf("arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, BatchCount\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {row * stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {col * stride, 1_uz, stride});
+        }
+    };
+
+    Tensor<ADataType> a_g_m_k(f_host_tensor_descriptor(BatchCount, M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_g_k_n(f_host_tensor_descriptor(BatchCount, K, N, StrideB, BLayout{}));
+
+    Tensor<CDataType> c_g_m_n_host_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> d0_g_m_host_result({BatchCount, M});
+    Tensor<ReduceDataType> d1_g_m_host_result({BatchCount, M});
+
+    Tensor<CDataType> c_g_m_n_device_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> d0_g_m_device_result({BatchCount, M});
+    Tensor<ReduceDataType> d1_g_m_device_result({BatchCount, M});
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
+    std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
+    std::cout << "d0_g_m: " << d0_g_m_host_result.mDesc << std::endl;
+    std::cout << "d1_g_m: " << d1_g_m_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
+                                 d0_g_m_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
+                                 d1_g_m_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_g_m_k.mData.data());
+    b_device_buf.ToDevice(b_g_k_n.mData.data());
+
+    auto a_element_op                     = AElementOp{};
+    auto b_element_op                     = BElementOp{};
+    auto c_element_op                     = CElementOp{};
+    std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
+
+    auto passthrough                            = UnaryIdenticElementOp{};
+    auto square                                 = UnarySquareElementOp{};
+    std::array<void*, 2> reduce_in_element_ops  = {&passthrough, &square};
+    std::array<void*, 2> reduce_out_element_ops = {&passthrough, &passthrough};
+
+    std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
+                                      reduce1_device_buf.GetDeviceBuffer()};
+
+    // do GEMM
+    auto batched_gemm = DeviceBatchedGemmReduceInstance{};
+    auto invoker      = batched_gemm.MakeInvoker();
+    auto argument     = batched_gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                              b_device_buf.GetDeviceBuffer(),
+                                              nullptr,
+                                              {},
+                                              c_device_buf.GetDeviceBuffer(),
+                                              p_reduces,
+                                              M,
+                                              N,
+                                              K,
+                                              StrideA,
+                                              StrideB,
+                                              StrideC,
+                                              {},
+                                              gemm_element_ops,
+                                              {},
+                                              reduce_in_element_ops,
+                                              reduce_out_element_ops,
+                                              BatchCount);
+
+    if(!batched_gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    // init DO, D1 to 0
+    reduce0_device_buf.SetZero();
+    reduce1_device_buf.SetZero();
+
+    // if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result
+    // will not be correct. need to set time_kernel = false for correctness test
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = std::size_t(2) * BatchCount * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * BatchCount * M * K +
+                            sizeof(BDataType) * BatchCount * K * N +
+                            sizeof(CDataType) * BatchCount * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << batched_gemm.GetTypeString() << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
+        reduce0_device_buf.FromDevice(d0_g_m_device_result.mData.data());
+        reduce1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
+
+        auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
+        auto ref_invoker      = ref_batched_gemm.MakeInvoker();
+
+        auto ref_argument = ref_batched_gemm.MakeArgument(
+            a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        auto reduce0_op = ReduceOp0{};
+        auto reduce1_op = ReduceOp1{};
+
+        for(int batch = 0; batch < BatchCount; ++batch)
+        {
+            for(int m = 0; m < M; ++m)
+            {
+                auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+                auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
+
+                for(int n = 0; n < N; ++n)
+                {
+                    auto c_val =
+                        ck::type_convert<ReduceAccDataType>(c_g_m_n_host_result(batch, m, n));
+                    ReduceAccDataType d0_val;
+                    ReduceAccDataType d1_val;
+
+                    UnaryIdenticElementOp{}(d0_val, c_val);
+                    UnarySquareElementOp{}(d1_val, c_val);
+                    reduce0_op(reduce0_acc, d0_val);
+                    reduce1_op(reduce1_acc, d1_val);
+                }
+
+                d0_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce0_acc);
+                d1_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce1_acc);
+            }
+        }
+
+        pass = ck::utils::check_err(
+                   c_g_m_n_host_result, c_g_m_n_device_result, "Error: Incorrect results c") &&
+               ck::utils::check_err(d0_g_m_device_result,
+                                    d0_g_m_host_result,
+                                    "Error: Incorrect results! D0",
+                                    1e-4,
+                                    1e-5) &&
+               ck::utils::check_err(d1_g_m_device_result,
+                                    d1_g_m_host_result,
+                                    "Error: Incorrect results! D1",
+                                    1e-3,
+                                    1e-5);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/example/19_binary_elementwise/CMakeLists.txt b/example/19_binary_elementwise/CMakeLists.txt
new file mode 100644
index 0000000000..39646e0ab5
--- /dev/null
+++ b/example/19_binary_elementwise/CMakeLists.txt
@@ -0,0 +1,4 @@
+add_example_executable(example_broadcast_add_2d_amn_bn broadcast_add_2d_amn_bn.cpp)
+add_example_executable(example_broadcast_add_3d_am_bmnk broadcast_add_3d_am_bmnk.cpp)
+add_example_executable(example_elementwise_add_1d elementwise_add_1d.cpp)
+add_example_executable(example_elementwise_add_4d elementwise_add_4d.cpp)
\ No newline at end of file
diff --git a/example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp b/example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
new file mode 100644
index 0000000000..24c8d82f67
--- /dev/null
+++ b/example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
@@ -0,0 +1,136 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType = F16;
+using CDataType  = F16;
+
+using Add = ck::tensor_operation::element_wise::Add;
+
+using DeviceElementwiseAddInstance =
+    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ABDataType, ABDataType>,
+                                                        ck::Tuple<CDataType>,
+                                                        Add,
+                                                        2,
+                                                        8,
+                                                        ck::Sequence<8, 8>,
+                                                        ck::Sequence<8>>;
+
+template <typename HostTensorA,
+          typename HostTensorB,
+          typename HostTensorC,
+          typename Functor,
+          int broadcastDim>
+void host_broadcast2D(
+    HostTensorC& C, const HostTensorA& A, const HostTensorB& B, int M, int N, Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0))>;
+
+    for(int m = 0; m < M; ++m)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            auto Amn  = A(m, n);
+            ctype Cmn = 0;
+            if constexpr(broadcastDim == 0)
+            {
+                auto Bn = B(n);
+                functor(Cmn, Amn, Bn);
+            }
+            else
+            {
+                auto Bm = B(m);
+                functor(Cmn, Amn, Bm);
+            }
+            C(m, n) = Cmn;
+        }
+    }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    ck::index_t M      = 1024;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = 1024;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
+        using namespace ck::literals;
+
+        return HostTensorDescriptor({row, col}, {stride, 1_uz});
+    };
+
+    Tensor<ABDataType> a_m_n(f_host_tensor_descriptor2d(M, N, Stride));
+    Tensor<ABDataType> b_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, Stride));
+
+    a_m_n.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b_n.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_m_n_device_buf(sizeof(ABDataType) * a_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem b_n_device_buf(sizeof(ABDataType) * b_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n.mDesc.GetElementSpaceSize());
+
+    a_m_n_device_buf.ToDevice(a_m_n.mData.data());
+    b_n_device_buf.ToDevice(b_n.mData.data());
+
+    std::array<const void*, 2> input = {a_m_n_device_buf.GetDeviceBuffer(),
+                                        b_n_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {c_m_n_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 2> abc_lengths = {M, N};
+    std::array<ck::index_t, 2> a_strides   = {Stride, 1};
+    std::array<ck::index_t, 2> b_strides   = {0, 1};
+    std::array<ck::index_t, 2> c_strides   = {Stride, 1};
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(
+        abc_lengths, {a_strides, b_strides}, {c_strides}, input, output, Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_m_n_device_buf.FromDevice(c_m_n.mData.data());
+        Tensor<CDataType> host_c_m_n(f_host_tensor_descriptor2d(M, N, Stride));
+
+        host_broadcast2D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add, 0>(
+            host_c_m_n, a_m_n, b_n, M, N, Add{});
+
+        pass &= ck::utils::check_err(c_m_n, host_c_m_n, "Error: Incorrect results c", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp b/example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
new file mode 100644
index 0000000000..3c04c56140
--- /dev/null
+++ b/example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
@@ -0,0 +1,120 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType = F16;
+using CDataType  = F16;
+
+using Add = ck::tensor_operation::element_wise::Add;
+
+using DeviceElementwiseAddInstance =
+    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ABDataType, ABDataType>,
+                                                        ck::Tuple<CDataType>,
+                                                        Add,
+                                                        3,
+                                                        8,
+                                                        ck::Sequence<1, 8>,
+                                                        ck::Sequence<8>>;
+
+template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
+void host_broadcast3D_am_bmnk(HostTensorC& C,
+                              const HostTensorA& A,
+                              const HostTensorB& B,
+                              const std::vector<std::size_t>& shape,
+                              Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0))>;
+
+    for(std::size_t m = 0; m < shape[0]; ++m)
+        for(std::size_t n = 0; n < shape[1]; ++n)
+            for(std::size_t k = 0; k < shape[2]; ++k)
+            {
+                auto a_val  = A(m);
+                auto b_val  = B(m, n, k);
+                ctype c_val = 0;
+                functor(c_val, a_val, b_val);
+                C(m, n, k) = c_val;
+            }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    std::vector<std::size_t> mnk = {4, 16, 32};
+    ck::index_t M                = mnk[0];
+
+    Tensor<ABDataType> a_m({M});
+    Tensor<ABDataType> b_m_n_k(mnk);
+    Tensor<CDataType> c_m_n_k(mnk);
+
+    a_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b_m_n_k.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_m_device_buf(sizeof(ABDataType) * a_m.mDesc.GetElementSpaceSize());
+    DeviceMem b_m_n_k_device_buf(sizeof(ABDataType) * b_m_n_k.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_k_device_buf(sizeof(CDataType) * c_m_n_k.mDesc.GetElementSpaceSize());
+
+    a_m_device_buf.ToDevice(a_m.mData.data());
+    b_m_n_k_device_buf.ToDevice(b_m_n_k.mData.data());
+
+    std::array<const void*, 2> input = {a_m_device_buf.GetDeviceBuffer(),
+                                        b_m_n_k_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {c_m_n_k_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 3> abc_lengths;
+    std::array<ck::index_t, 3> a_strides = {1, 0, 0};
+    std::array<ck::index_t, 3> b_strides;
+    std::array<ck::index_t, 3> c_strides;
+
+    ck::ranges::copy(mnk, abc_lengths.begin());
+    ck::ranges::copy(b_m_n_k.mDesc.GetStrides(), b_strides.begin());
+    ck::ranges::copy(c_m_n_k.mDesc.GetStrides(), c_strides.begin());
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(
+        abc_lengths, {a_strides, b_strides}, {c_strides}, input, output, Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_m_n_k_device_buf.FromDevice(c_m_n_k.mData.data());
+        Tensor<CDataType> host_c_m_n_k(mnk);
+
+        host_broadcast3D_am_bmnk<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>(
+            host_c_m_n_k, a_m, b_m_n_k, mnk, Add{});
+
+        pass &=
+            ck::utils::check_err(c_m_n_k, host_c_m_n_k, "Error: Incorrect results c", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/example/19_binary_elementwise/elementwise_add_1d.cpp b/example/19_binary_elementwise/elementwise_add_1d.cpp
new file mode 100644
index 0000000000..1ac09641a1
--- /dev/null
+++ b/example/19_binary_elementwise/elementwise_add_1d.cpp
@@ -0,0 +1,111 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType = F16;
+using CDataType  = F16;
+
+using Add = ck::tensor_operation::element_wise::Add;
+
+using DeviceElementwiseAddInstance =
+    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ABDataType, ABDataType>,
+                                                        ck::Tuple<CDataType>,
+                                                        Add,
+                                                        1,
+                                                        8,
+                                                        ck::Sequence<8, 8>,
+                                                        ck::Sequence<8>>;
+
+template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
+void host_elementwise1D(
+    HostTensorC& C, const HostTensorA& A, const HostTensorB& B, int M, Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0))>;
+
+    for(int m = 0; m < M; ++m)
+    {
+        auto Am  = A(m);
+        auto Bm  = B(m);
+        ctype Cm = 0;
+        functor(Cm, Am, Bm);
+        C(m) = Cm;
+    }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    ck::index_t M = 1024;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    Tensor<ABDataType> a_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<ABDataType> b_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<CDataType> c_m(f_host_tensor_descriptor1d(M, 1));
+
+    a_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_m_device_buf(sizeof(ABDataType) * a_m.mDesc.GetElementSpaceSize());
+    DeviceMem b_m_device_buf(sizeof(ABDataType) * b_m.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_device_buf(sizeof(CDataType) * c_m.mDesc.GetElementSpaceSize());
+
+    a_m_device_buf.ToDevice(a_m.mData.data());
+    b_m_device_buf.ToDevice(b_m.mData.data());
+
+    std::array<const void*, 2> input = {a_m_device_buf.GetDeviceBuffer(),
+                                        b_m_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {c_m_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 1> abc_lengths = {M};
+    std::array<ck::index_t, 1> a_strides   = {1};
+    std::array<ck::index_t, 1> b_strides   = {1};
+    std::array<ck::index_t, 1> c_strides   = {1};
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(
+        abc_lengths, {a_strides, b_strides}, {c_strides}, input, output, Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_m_device_buf.FromDevice(c_m.mData.data());
+        Tensor<CDataType> host_c_m(f_host_tensor_descriptor1d(M, 1));
+
+        host_elementwise1D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>(
+            host_c_m, a_m, b_m, M, Add{});
+
+        pass &= ck::utils::check_err(c_m, host_c_m, "Error: Incorrect results c", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/example/19_binary_elementwise/elementwise_add_4d.cpp b/example/19_binary_elementwise/elementwise_add_4d.cpp
new file mode 100644
index 0000000000..e571aa8468
--- /dev/null
+++ b/example/19_binary_elementwise/elementwise_add_4d.cpp
@@ -0,0 +1,120 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType = F16;
+using CDataType  = F16;
+
+using Add = ck::tensor_operation::element_wise::Add;
+
+using DeviceElementwiseAddInstance =
+    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ABDataType, ABDataType>,
+                                                        ck::Tuple<CDataType>,
+                                                        Add,
+                                                        4,
+                                                        8,
+                                                        ck::Sequence<8, 8>,
+                                                        ck::Sequence<8>>;
+
+template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
+void host_elementwise4D(HostTensorC& C,
+                        const HostTensorA& A,
+                        const HostTensorB& B,
+                        const std::vector<std::size_t>& shape,
+                        Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0, 0, 0))>;
+
+    for(std::size_t n = 0; n < shape[0]; ++n)
+        for(std::size_t c = 0; c < shape[1]; ++c)
+            for(std::size_t h = 0; h < shape[2]; ++h)
+                for(std::size_t w = 0; w < shape[3]; ++w)
+                {
+                    auto a_val  = A(n, c, h, w);
+                    auto b_val  = B(n, c, h, w);
+                    ctype c_val = 0;
+                    functor(c_val, a_val, b_val);
+                    C(n, c, h, w) = c_val;
+                }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    std::vector<std::size_t> nchw = {4, 16, 32, 32};
+
+    Tensor<ABDataType> a(nchw);
+    Tensor<ABDataType> b(nchw);
+    Tensor<CDataType> c(nchw);
+
+    a.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_device_buf(sizeof(ABDataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(ABDataType) * b.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+    b_device_buf.ToDevice(b.mData.data());
+
+    std::array<const void*, 2> input = {a_device_buf.GetDeviceBuffer(),
+                                        b_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {c_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 4> abc_lengths;
+    std::array<ck::index_t, 4> a_strides;
+    std::array<ck::index_t, 4> b_strides;
+    std::array<ck::index_t, 4> c_strides;
+
+    ck::ranges::copy(nchw, abc_lengths.begin());
+    ck::ranges::copy(a.mDesc.GetStrides(), a_strides.begin());
+    ck::ranges::copy(b.mDesc.GetStrides(), b_strides.begin());
+    ck::ranges::copy(c.mDesc.GetStrides(), c_strides.begin());
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(
+        abc_lengths, {a_strides, b_strides}, {c_strides}, input, output, Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c.mData.data());
+        Tensor<CDataType> host_c(nchw);
+
+        host_elementwise4D<Tensor<ABDataType>, Tensor<ABDataType>, Tensor<CDataType>, Add>(
+            host_c, a, b, nchw, Add{});
+
+        pass &= ck::utils::check_err(c, host_c, "Error: Incorrect results c", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/example/20_grouped_conv_bwd_weight/CMakeLists.txt b/example/20_grouped_conv_bwd_weight/CMakeLists.txt
new file mode 100644
index 0000000000..d649567ed2
--- /dev/null
+++ b/example/20_grouped_conv_bwd_weight/CMakeLists.txt
@@ -0,0 +1,24 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_custom_target(example_grouped_conv_bwd_weight)
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp)
+    add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_grouped_conv_bwd_weight_xdl_bf16 grouped_conv_bwd_weight_xdl_bf16.cpp)
+    add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_bf16)
+   endif()
+   set(target 1)
+ endif()
+endforeach()
+
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  if(DL_KERNELS)
+    add_custom_target(example_grouped_conv_bwd_weight_dl)
+    add_example_executable(example_grouped_conv_bwd_weight_dl_fp16 grouped_conv_bwd_weight_dl_fp16.cpp)
+    add_dependencies(example_grouped_conv_bwd_weight_dl example_grouped_conv_bwd_weight_dl_fp16)
+  endif()
+endif()
\ No newline at end of file
diff --git a/example/20_grouped_conv_bwd_weight/common.hpp b/example/20_grouped_conv_bwd_weight/common.hpp
new file mode 100644
index 0000000000..15727495f0
--- /dev/null
+++ b/example/20_grouped_conv_bwd_weight/common.hpp
@@ -0,0 +1,137 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <iostream>
+#include <iterator>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp"
+
+using BF16 = ck::bhalf_t;
+using F16  = ck::half_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+template <typename InputLay, typename WeightLay, typename OutputLay>
+struct CommonLayoutSetting
+{
+    using InputLayout  = InputLay;
+    using WeightLayout = WeightLay;
+    using OutputLayout = OutputLay;
+};
+
+template <ck::index_t NDimSpatial>
+struct CommonLayoutSettingSelector;
+
+namespace ctl = ck::tensor_layout::convolution;
+
+template <>
+struct CommonLayoutSettingSelector<1> final : CommonLayoutSetting<ctl::GNWC, ctl::GKXC, ctl::GNWK>
+{
+};
+
+template <>
+struct CommonLayoutSettingSelector<2> final
+    : CommonLayoutSetting<ctl::GNHWC, ctl::GKYXC, ctl::GNHWK>
+{
+};
+
+template <>
+struct CommonLayoutSettingSelector<3> final
+    : CommonLayoutSetting<ctl::GNDHWC, ctl::GKZYXC, ctl::GNDHWK>
+{
+};
+
+template <ck::index_t NDimSpatial>
+using InputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::InputLayout;
+
+template <ck::index_t NDimSpatial>
+using WeightLayout = typename CommonLayoutSettingSelector<NDimSpatial>::WeightLayout;
+
+template <ck::index_t NDimSpatial>
+using OutputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::OutputLayout;
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+#define DefaultConvParam                                                      \
+    ck::utils::conv::ConvParam                                                \
+    {                                                                         \
+        2, 4, 1, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, { 1, 1 } \
+    }
+
+inline void print_help_msg()
+{
+    std::cerr << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+inline bool parse_cmd_args(int argc,
+                           char* argv[],
+                           ExecutionConfig& config,
+                           ck::utils::conv::ConvParam& conv_param)
+{
+    constexpr int num_execution_config_args =
+        3; // arguments for do_verification, init_method, time_kernel
+    constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
+
+    constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
+    constexpr int threshold_to_catch_all_args =
+        threshold_to_catch_partial_args + num_conv_param_leading_args;
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    // catch only ExecutionConfig arguments
+    else if(argc == threshold_to_catch_partial_args)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    // catch both ExecutionConfig & ConvParam arguments
+    else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+        conv_param                        = ck::utils::conv::parse_conv_param(
+            num_dim_spatial, threshold_to_catch_partial_args, argv);
+    }
+    else
+    {
+        print_help_msg();
+        return false;
+    }
+
+    return true;
+}
diff --git a/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp
new file mode 100644
index 0000000000..375c309e1c
--- /dev/null
+++ b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_dl.hpp"
+
+using InDataType  = F16;
+using WeiDataType = F16;
+using OutDataType = F16;
+using AccDataType = F32;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = PassThrough;
+
+template <ck::index_t NDimSpatial>
+using DeviceConvBwdWeightInstance =
+    ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl<
+        NDimSpatial,          // NDimSpatial
+        InDataType,           // InDataType
+        WeiDataType,          // WeiDataType
+        OutDataType,          // OutDataType
+        AccDataType,          // AccDataType
+        InElementOp,          // InElementwiseOperation
+        WeiElementOp,         // WeiElementwiseOperation
+        OutElementOp,         // OutElementwiseOperation
+        ConvBwdWeightDefault, // ConvBackwardWeightSpecialization
+        256,                  // BlockSize
+        128,                  // MPerBlock
+        128,                  // NPerBlock
+        16,                   // K0PerBlock
+        2,                    // K1
+        4,                    // M1PerThread
+        4,                    // N1PerThread
+        1,                    // KPerThread
+        S<8, 2>,              // M1N1ThreadClusterM1Xs
+        S<8, 2>,              // M1N1ThreadClusterN1Xs
+        S<1, 8, 1, 1, 2>,     // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
+        S<1, 2, 1, 128, 1>,   // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
+        S<0, 2, 3, 1, 4>,     // ABlockTransferThreadClusterArrangeOrder
+        S<0, 2, 3, 1, 4>,     // ABlockTransferSrcAccessOrder
+        S<1, 1, 1, 1, 1>,     // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
+        S<0, 2, 3, 1, 4>,     // ABlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 1, 1>,     // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
+        S<1, 1, 1, 8, 2>,     // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
+        S<1, 16, 1, 16, 1>,   // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
+        S<0, 1, 4, 2, 3>,     // BBlockTransferThreadClusterArrangeOrder
+        S<0, 1, 4, 2, 3>,     // BBlockTransferSrcAccessOrder
+        S<1, 1, 1, 8, 1>,     // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
+        S<0, 1, 4, 2, 3>,     // BBlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 1, 2>,     // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
+        S<0, 1, 2, 3, 4, 5>,  // CThreadTransferSrcDstAccessOrder
+        5,                    // CThreadTransferSrcDstVectorDim
+        4>;                   // CThreadTransferDstScalarPerVector
+
+#include "run_grouped_conv_bwd_weight_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
diff --git a/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp
new file mode 100644
index 0000000000..31e277b5c7
--- /dev/null
+++ b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp
@@ -0,0 +1,72 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
+
+using InDataType = BF16;
+// bf16 kernel use fp32 atomic add to accumulate Weight tensor into global memory
+using WeiDataType = F32;
+using OutDataType = BF16;
+using AccDataType = F32;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = PassThrough;
+
+template <ck::index_t NDimSpatial>
+using DeviceConvBwdWeightInstance =
+    ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Xdl_CShuffle<
+        NDimSpatial,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GNWC,
+                                      ck::tensor_layout::convolution::GNHWC,
+                                      ck::tensor_layout::convolution::GNDHWC>>,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GKXC,
+                                      ck::tensor_layout::convolution::GKYXC,
+                                      ck::tensor_layout::convolution::GKZYXC>>,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GNWK,
+                                      ck::tensor_layout::convolution::GNHWK,
+                                      ck::tensor_layout::convolution::GNDHWK>>,
+        InDataType,           // InDataType
+        WeiDataType,          // WeiDataType
+        OutDataType,          // OutDataType
+        AccDataType,          // AccDataType
+        InElementOp,          // InElementwiseOperation
+        WeiElementOp,         // WeiElementwiseOperation
+        OutElementOp,         // OutElementwiseOperation
+        ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
+        256,                  // BlockSize
+        128,                  // MPerBlock
+        128,                  // NPerBlock
+        4,                    // K0PerBlock
+        8,                    // K1
+        32,                   // MPerXdl
+        32,                   // NPerXdl
+        2,                    // MXdlPerWave
+        2,                    // NXdlPerWave
+        S<1, 4, 16, 4>,       // ABlockTransferThreadClusterLengths_K0_M_K1
+        S<0, 3, 1, 2>,        // ABlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1, 3>,        // ABlockTransferSrcAccessOrder
+        2,                    // ABlockTransferSrcVectorDim
+        8,                    // ABlockTransferSrcScalarPerVector
+        2,                    // ABlockTransferDstScalarPerVector_K1
+        true,                 // ABlockLdsAddExtraM
+        S<1, 4, 16, 4>,       // BBlockTransferThreadClusterLengths_K0_N_K1
+        S<0, 3, 1, 2>,        // BBlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1, 3>,        // BBlockTransferSrcAccessOrder
+        2,                    // BBlockTransferSrcVectorDim
+        8,                    // BBlockTransferSrcScalarPerVector
+        2,                    // BBlockTransferDstScalarPerVector_K1
+        true,                 // BBlockLdsAddExtraN
+        1,                    // CShuffleMXdlPerWavePerShuffle
+        1,                    // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 4>,       // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
+
+#include "run_grouped_conv_bwd_weight_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
diff --git a/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp
new file mode 100644
index 0000000000..69c831cc54
--- /dev/null
+++ b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp
@@ -0,0 +1,71 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
+
+using InDataType  = F16;
+using WeiDataType = F16;
+using OutDataType = F16;
+using AccDataType = F32;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = PassThrough;
+
+template <ck::index_t NDimSpatial>
+using DeviceConvBwdWeightInstance =
+    ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Xdl_CShuffle<
+        NDimSpatial,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GNWC,
+                                      ck::tensor_layout::convolution::GNHWC,
+                                      ck::tensor_layout::convolution::GNDHWC>>,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GKXC,
+                                      ck::tensor_layout::convolution::GKYXC,
+                                      ck::tensor_layout::convolution::GKZYXC>>,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GNWK,
+                                      ck::tensor_layout::convolution::GNHWK,
+                                      ck::tensor_layout::convolution::GNDHWK>>,
+        InDataType,           // InDataType
+        WeiDataType,          // WeiDataType
+        OutDataType,          // OutDataType
+        AccDataType,          // AccDataType
+        InElementOp,          // InElementwiseOperation
+        WeiElementOp,         // WeiElementwiseOperation
+        OutElementOp,         // OutElementwiseOperation
+        ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
+        256,                  // BlockSize
+        128,                  // MPerBlock
+        128,                  // NPerBlock
+        4,                    // K0PerBlock
+        8,                    // K1
+        32,                   // MPerXdl
+        32,                   // NPerXdl
+        2,                    // MXdlPerWave
+        2,                    // NXdlPerWave
+        S<1, 4, 16, 4>,       // ABlockTransferThreadClusterLengths_K0_M_K1
+        S<0, 3, 1, 2>,        // ABlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1, 3>,        // ABlockTransferSrcAccessOrder
+        2,                    // ABlockTransferSrcVectorDim
+        8,                    // ABlockTransferSrcScalarPerVector
+        2,                    // ABlockTransferDstScalarPerVector_K1
+        true,                 // ABlockLdsAddExtraM
+        S<1, 4, 16, 4>,       // BBlockTransferThreadClusterLengths_K0_N_K1
+        S<0, 3, 1, 2>,        // BBlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1, 3>,        // BBlockTransferSrcAccessOrder
+        2,                    // BBlockTransferSrcVectorDim
+        8,                    // BBlockTransferSrcScalarPerVector
+        2,                    // BBlockTransferDstScalarPerVector_K1
+        true,                 // BBlockLdsAddExtraN
+        1,                    // CShuffleMXdlPerWavePerShuffle
+        1,                    // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 4>,       // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
+
+#include "run_grouped_conv_bwd_weight_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_bwd_weight_example(argc, argv); }
diff --git a/example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc b/example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
new file mode 100644
index 0000000000..29ce0324ab
--- /dev/null
+++ b/example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
@@ -0,0 +1,184 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+template <ck::index_t NDimSpatial>
+using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
+                                                                                     InDataType,
+                                                                                     WeiDataType,
+                                                                                     OutDataType,
+                                                                                     InElementOp,
+                                                                                     WeiElementOp,
+                                                                                     OutElementOp>;
+
+template <ck::index_t NDimSpatial>
+bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
+                                 const ck::utils::conv::ConvParam& conv_param)
+{
+    ck::index_t split_k;
+    // Set split_k = 2 for xdl op, split_k = 1 for dl
+    // Dl op doesn't support split_k > 1
+    // TODO: Add Dl op split_k > 1 support
+    if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
+         ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
+         ck::get_device_name() == "gfx1102"))
+    {
+        split_k = 2;
+    }
+    else
+    {
+        split_k = 1;
+    }
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<
+            InputLayout<NDimSpatial>>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<
+            WeightLayout<NDimSpatial>>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<
+            OutputLayout<NDimSpatial>>(conv_param);
+
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei_host_result(wei_g_k_c_xs_desc);
+    Tensor<WeiDataType> wei_device_result(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> out(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei_host_result.mDesc << std::endl;
+    std::cout << "out: " << out.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    out_device_buf.ToDevice(out.mData.data());
+
+    // init to 0
+    wei_device_buf.SetZero();
+
+    std::array<ck::index_t, NDimSpatial + 3> input_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> input_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> filter_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> weights_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> output_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> output_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto range_copy = [](const auto& from, auto to) { std::copy(begin(from), end(from), to); };
+
+    range_copy(in_g_n_c_wis_desc.GetLengths(), begin(input_lengths));
+    range_copy(in_g_n_c_wis_desc.GetStrides(), begin(input_strides));
+    range_copy(wei_g_k_c_xs_desc.GetLengths(), begin(filter_lengths));
+    range_copy(wei_g_k_c_xs_desc.GetStrides(), begin(weights_strides));
+    range_copy(out_g_n_k_wos_desc.GetLengths(), begin(output_lengths));
+    range_copy(out_g_n_k_wos_desc.GetStrides(), begin(output_strides));
+    range_copy(conv_param.conv_filter_strides_, begin(conv_filter_strides));
+    range_copy(conv_param.conv_filter_dilations_, begin(conv_filter_dilations));
+    range_copy(conv_param.input_left_pads_, begin(input_left_pads));
+    range_copy(conv_param.input_right_pads_, begin(input_right_pads));
+
+    // do GEMM
+    auto conv     = DeviceConvBwdWeightInstance<NDimSpatial>{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                      static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                      static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                      input_lengths,
+                                      input_strides,
+                                      filter_lengths,
+                                      weights_strides,
+                                      output_lengths,
+                                      output_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      InElementOp{},
+                                      WeiElementOp{},
+                                      OutElementOp{},
+                                      split_k);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        std::cerr << "wrong! device_conv with the specified compilation parameters does "
+                     "not support this Conv problem"
+                  << std::endl;
+        return true;
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
+
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+    std::cerr << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl
+              << "DeviceOp: " << conv.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        auto ref_conv     = HostConvBwdWeightInstance<NDimSpatial>{};
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei_host_result,
+                                                  out,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  InElementOp{},
+                                                  WeiElementOp{},
+                                                  OutElementOp{});
+
+        ref_invoker.Run(ref_argument);
+
+        wei_device_buf.FromDevice(wei_device_result.mData.data());
+
+        return ck::utils::check_err(wei_device_result.mData, wei_host_result.mData);
+    }
+
+    return true;
+}
+
+bool run_grouped_conv_bwd_weight_example(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+
+    if(!parse_cmd_args(argc, argv, config, conv_param))
+    {
+        return false;
+    }
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1: return run_grouped_conv_bwd_weight<1>(config, conv_param);
+    case 2: return run_grouped_conv_bwd_weight<2>(config, conv_param);
+    case 3: return run_grouped_conv_bwd_weight<3>(config, conv_param);
+    }
+
+    return false;
+}
diff --git a/example/21_gemm_layernorm/CMakeLists.txt b/example/21_gemm_layernorm/CMakeLists.txt
new file mode 100644
index 0000000000..6a6735efd6
--- /dev/null
+++ b/example/21_gemm_layernorm/CMakeLists.txt
@@ -0,0 +1,13 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_welford_fp16 gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp)
+   add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_naive_fp16 gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp)
+   add_example_executable(example_gemm_layernorm_xdl_naive_fp16 gemm_layernorm_xdl_naive_fp16.cpp)
+   add_example_executable(example_gemm_xdl_layernorm_naive_single_kernel_fp16 gemm_xdl_layernorm_naive_single_kernel_fp16.cpp)
+   set(target 1)
+ endif()
+endforeach()
+endif()
diff --git a/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp b/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp
new file mode 100644
index 0000000000..96d04dcb37
--- /dev/null
+++ b/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp
@@ -0,0 +1,405 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+// DataType
+using ADataType                = F16;
+using BDataType                = F16;
+using GemmAccDataType          = F32;
+using CShuffleDataType         = F32;
+using D0DataType               = F16;
+using D1DataType               = F16;
+using DsDataType               = ck::Tuple<D0DataType, D1DataType>;
+using EDataType                = F16;
+using ReduceAccDataType        = F32;
+using R0DataType               = F32;
+using R1DataType               = F32;
+using RsDataType               = ck::Tuple<R0DataType, R1DataType>;
+using GammaDataType            = F16;
+using BetaDataType             = F16;
+using LayerNormOutDataType     = F16;
+using NormalizeComputeDataType = F32;
+
+// Layout
+using ALayout  = Row;
+using BLayout  = Col;
+using D1Layout = Row;
+using ELayout  = D1Layout;
+
+// Elementwise op
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd   = ck::tensor_operation::element_wise::AddReluAdd;
+using Square       = ck::tensor_operation::element_wise::UnarySquare;
+using Div          = ck::tensor_operation::element_wise::UnaryDivide;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddReluAdd;
+using QsElementOp  = ck::Tuple<PassThrough, Square>;
+using RsElementOp  = ck::Tuple<Div, Div>;
+
+// ReduceOp
+using R0ThreadReduceOp = ck::reduce::Add;
+using R1ThreadReduceOp = ck::reduce::Add;
+using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
+
+static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+//######| ALayout| BLayout| ELayout|     AData|     BData|     GemmAccData|         CShuffle|     DsData|     EData|     ReduceAccData|     RsData|           A|           B|          CDE|          Qs|          Rs|           Thread|           Global|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|    CDRThreadTransfer|                  CDE|    RThreadTransfer|
+//######|        |        |        |      Type|      Type|            Type|         DataType|       Type|      Type|              Type|       Type| Elementwise| Elementwise|  Elementwise| Elementwise| Elementwise|           Reduce|           Reduce| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|       ClusterLengths| ReduceThreadTransfer| DstScalarPerVector|
+//######|        |        |        |          |          |                |                 |           |          |                  |           |   Operation|   Operation|    Operation|   Operation|   Operation|        Operation|        Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _MPerBlock_NPerBlock|      ScalarPerVector|         _MPerBlock|
+//######|        |        |        |          |          |                |                 |           |          |                  |           |            |            |             |            |            |                 |                 |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |           _NPerBlock|                   |
+        < ALayout, BLayout, ELayout, ADataType, BDataType, GemmAccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType,  AElementOp,  BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<64, 4>,                    4,                  1>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        EDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+using NormalizeFunctor = ck::tensor_operation::element_wise::Normalize;
+
+// A:x, B:E[x], C:E[x^2], D:Gamma, E:Beta , F:y
+using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
+    ck::Tuple<EDataType,
+              R0DataType,
+              R1DataType,
+              GammaDataType,
+              BetaDataType>,         // x(gemm_out), mean, meansquare, gamma, beta
+    ck::Tuple<LayerNormOutDataType>, // y
+    NormalizeFunctor,
+    2,
+    8,                           // MPerthread
+    ck::Sequence<8, 1, 1, 8, 8>, // scalarPerVector: x(gemm_out), mean, meansquare, gamma, beta
+    ck::Sequence<8>>;            // scalarPerVector: y(layerNorm_out)
+
+auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+    return HostTensorDescriptor({len}, {stride});
+};
+
+auto f_host_tensor_descriptor2d =
+    [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+        using namespace ck::literals;
+
+        if constexpr(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({row, col}, {stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({row, col}, {1_uz, stride});
+        }
+    };
+
+void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
+                         const Tensor<ADataType>& a_m_k,
+                         const Tensor<BDataType>& b_k_n,
+                         const Tensor<D0DataType>& bias_n,
+                         const Tensor<D1DataType>& d1_m_n,
+                         const Tensor<GammaDataType>& gamma_n,
+                         const Tensor<BetaDataType>& beta_n,
+                         AElementOp a_element_op,
+                         BElementOp b_element_op,
+                         CDEElementOp cde_element_op,
+                         int M,
+                         int N)
+{
+
+    int StrideE = N;
+    Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> mean_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<R1DataType> meanSquare_m(f_host_tensor_descriptor1d(M, 1));
+    auto averageOpInst = Div{N};
+
+    auto ref_gemm    = ReferenceGemmInstance{};
+    auto ref_invoker = ref_gemm.MakeInvoker();
+
+    auto ref_argument =
+        ref_gemm.MakeArgument(a_m_k, b_k_n, e_m_n, a_element_op, b_element_op, PassThrough{});
+
+    ref_invoker.Run(ref_argument);
+
+    // c = activation(c + bias) + c1_functor(c1)
+    for(int m = 0; m < M; ++m)
+        for(int n = 0; n < N; ++n)
+        {
+            auto acc = ck::type_convert<GemmAccDataType>(e_m_n(m, n));
+            cde_element_op(e_m_n(m, n), acc, bias_n(n), d1_m_n(m, n));
+        }
+
+    // reduce_mean and reduce_square_mean
+    auto r0Op = R0ThreadReduceOp{};
+    auto r1Op = R1ThreadReduceOp{};
+    for(int m = 0; m < M; ++m)
+    {
+        auto mean_acc        = r0Op.GetIdentityValue<ReduceAccDataType>();
+        auto mean_square_acc = r1Op.GetIdentityValue<ReduceAccDataType>();
+
+        for(int n = 0; n < N; ++n)
+        {
+            auto e_val                     = ck::type_convert<ReduceAccDataType>(e_m_n(m, n));
+            ReduceAccDataType square_e_val = 0;
+            Square{}(square_e_val, e_val);
+
+            r0Op(mean_acc, e_val);
+            r1Op(mean_square_acc, square_e_val);
+        }
+
+        averageOpInst(mean_acc, mean_acc);
+        averageOpInst(mean_square_acc, mean_square_acc);
+        mean_m(m)       = ck::type_convert<R0DataType>(mean_acc);
+        meanSquare_m(m) = ck::type_convert<R1DataType>(mean_square_acc);
+    }
+
+    // LayerNorm
+    auto layerNormInst = NormalizeFunctor{};
+    for(int m = 0; m < M; ++m)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            LayerNormOutDataType out_val = 0;
+            layerNormInst(out_val, e_m_n(m, n), mean_m(m), meanSquare_m(m), gamma_n(n), beta_n(n));
+            out_m_n(m, n) = out_val;
+        }
+    }
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename R0DataType,
+          typename R1DataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename NormalizeDataType>
+void DumpGemmLayerNormPerf(float gemm_reduce_time, float normalize_time, int M, int N, int K)
+{
+    std::size_t gemm_flop     = std::size_t(2) * M * N * K + std::size_t(2) * M * N;
+    std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                sizeof(EDataType) * M * N + sizeof(D0DataType) * M * N +
+                                sizeof(D0DataType) * M * N + sizeof(R0DataType) * M +
+                                sizeof(R1DataType) * M;
+
+    std::size_t normalize_num_byte = sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
+                                     sizeof(R1DataType) * M + sizeof(GammaDataType) * N +
+                                     sizeof(BetaDataType) * N + sizeof(NormalizeDataType) * M * N;
+
+    float tflops               = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
+    float gemm_gb_per_sec      = gemm_num_byte / 1.E6 / gemm_reduce_time;
+    float normalize_gb_per_sec = normalize_num_byte / 1.E6 / normalize_time;
+
+    std::cout << "gemm + reduce_mean + reduce_square_mean Perf: " << gemm_reduce_time << " ms, "
+              << tflops << " TFlops, " << gemm_gb_per_sec << " GB/s, " << std::endl;
+
+    std::cout << "5-ary elementwise Perf: " << normalize_time << " ms, " << normalize_gb_per_sec
+              << " GB/s, " << std::endl;
+}
+
+int main()
+{
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA  = 1024;
+    ck::index_t StrideB  = 1024;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 1024;
+    ck::index_t StrideE  = 1024;
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> bias_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor2d(M, N, StrideD1, ELayout{}));
+    Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> r0_Mean_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<R1DataType> r1_MeanSquare_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<GammaDataType> gamma_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<BetaDataType> beta_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<LayerNormOutDataType> layerNorm_m_n(
+        f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
+    b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
+    bias_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-1, 1});
+    d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-5, 5});
+    gamma_n.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-1, 1});
+    beta_n.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-1, 1});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(D0DataType) * bias_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem r0_Mean_device_buf(sizeof(R0DataType) * r0_Mean_m.mDesc.GetElementSpaceSize());
+    DeviceMem r1_MeanSquare_device_buf(sizeof(R1DataType) *
+                                       r1_MeanSquare_m.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_n.mDesc.GetElementSpaceSize());
+    DeviceMem beta_device_buf(sizeof(BetaDataType) * beta_n.mDesc.GetElementSpaceSize());
+    DeviceMem layerNorm_device_buf(sizeof(LayerNormOutDataType) *
+                                   layerNorm_m_n.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    bias_device_buf.ToDevice(bias_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+    gamma_device_buf.ToDevice(gamma_n.mData.data());
+    beta_device_buf.ToDevice(beta_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+    auto qs_element_op  = QsElementOp{};
+    auto rs_element_op  = RsElementOp{N, N};
+
+    // Prepare GEMM, mean, mean_square
+    auto gemmReduce          = DeviceOpInstance{};
+    auto gemmReduce_invoker  = gemmReduce.MakeInvoker();
+    auto gemmReduce_argument = gemmReduce.MakeArgument(
+        a_device_buf.GetDeviceBuffer(),
+        b_device_buf.GetDeviceBuffer(),
+        {bias_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+        e_device_buf.GetDeviceBuffer(),
+        {r0_Mean_device_buf.GetDeviceBuffer(), r1_MeanSquare_device_buf.GetDeviceBuffer()},
+        M,
+        N,
+        K,
+        StrideA,
+        StrideB,
+        {StrideD0, StrideD1},
+        StrideE,
+        a_element_op,
+        b_element_op,
+        cde_element_op,
+        qs_element_op,
+        rs_element_op);
+
+    if(!gemmReduce.IsSupportedArgument(gemmReduce_argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    // init reducetion buffer to 0
+    r0_Mean_device_buf.SetZero();
+    r1_MeanSquare_device_buf.SetZero();
+
+    // Prepare LayerNorm
+    std::array<const void*, 5> input = {e_device_buf.GetDeviceBuffer(),
+                                        r0_Mean_device_buf.GetDeviceBuffer(),
+                                        r1_MeanSquare_device_buf.GetDeviceBuffer(),
+                                        gamma_device_buf.GetDeviceBuffer(),
+                                        beta_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {layerNorm_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 2> xyLengths = {M, N};
+    std::array<ck::index_t, 2> xyStrides = {StrideE, 1};
+
+    auto normalize         = DeviceNormalizeInstance{};
+    auto normalize_invoker = normalize.MakeInvoker();
+    auto normalize_argument_ptr =
+        normalize.MakeArgumentPointer(xyLengths,
+                                      {xyStrides, {1, 0}, {1, 0}, {0, 1}, {0, 1}},
+                                      {xyStrides},
+                                      input,
+                                      output,
+                                      NormalizeFunctor{});
+
+    if(!normalize.IsSupportedArgument(normalize_argument_ptr.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device, exiting!");
+    }
+
+    // run kernel
+    gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, false});
+    normalize_invoker.Run(normalize_argument_ptr.get(), StreamConfig{nullptr, false});
+
+    bool pass = true;
+    {
+        // verification
+        Tensor<LayerNormOutDataType> host_layerNorm_m_n(
+            f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+
+        host_gemm_layernorm(host_layerNorm_m_n,
+                            a_m_k,
+                            b_k_n,
+                            bias_n,
+                            d1_m_n,
+                            gamma_n,
+                            beta_n,
+                            a_element_op,
+                            b_element_op,
+                            cde_element_op,
+                            M,
+                            N);
+
+        layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data());
+        pass &= ck::utils::check_err(layerNorm_m_n,
+                                     host_layerNorm_m_n,
+                                     "Error: Incorrect results layerNorm_m_n",
+                                     1e-2,
+                                     1e-2);
+    }
+
+    {
+        // evaluate kernel perf
+        bool time_kernel = true;
+
+        float gemm_reduce_mean_reduce_square_mean_ave_time =
+            gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, time_kernel});
+        float normalize_ave_time =
+            normalize_invoker.Run(normalize_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        if(time_kernel)
+            DumpGemmLayerNormPerf<ADataType,
+                                  BDataType,
+                                  EDataType,
+                                  D0DataType,
+                                  D1DataType,
+                                  R0DataType,
+                                  R1DataType,
+                                  GammaDataType,
+                                  BetaDataType,
+                                  LayerNormOutDataType>(
+                gemm_reduce_mean_reduce_square_mean_ave_time, normalize_ave_time, M, N, K);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp b/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp
new file mode 100644
index 0000000000..fc58ca19f8
--- /dev/null
+++ b/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp
@@ -0,0 +1,263 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+// DataType
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = F16;
+using D1DataType       = F16;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType>;
+using EMeanVarDataType = F16;
+using GammaDataType    = F16;
+using BetaDataType     = F16;
+using HDataType        = F16;
+
+// Layout
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using HLayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddReluAdd;
+using HElementOp   = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle
+//######| ALayout| BLayout| DsLayout| HLayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EMeanVarData|     GammaData|     BetaData|     HData|           A|           B|          CDE|            H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|           PostShuffle|     PostShuffle|            Layernorm|       Layernorm|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|             Type|          Type|         Type|      Type| Elementwise| Elementwise|  Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|  ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|
+//######|        |        |         |        |          |          |            |                 |           |                 |              |             |          |   Operation|   Operation|    Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                  _M_N|            _M_N|                 _M_N|              _M|
+//######|        |        |         |        |          |          |            |                 |           |                 |              |             |          |            |            |             |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                      |                |                     |                |
+        < ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType,  AElementOp,  BElementOp, CDEElementOp,   HElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<32, 8>,               8,             S<8, 32>,               8>;
+// clang-format on
+
+auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+    return HostTensorDescriptor({len}, {stride});
+};
+
+auto f_host_tensor_descriptor2d =
+    [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+        using namespace ck::literals;
+
+        if constexpr(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({row, col}, {stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({row, col}, {1_uz, stride});
+        }
+    };
+
+void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
+                         const Tensor<ADataType>& a_m_k,
+                         const Tensor<BDataType>& b_k_n,
+                         const Tensor<D0DataType>& bias_n,
+                         const Tensor<D1DataType>& d1_m_n,
+                         const Tensor<GammaDataType>& gamma_n,
+                         const Tensor<BetaDataType>& beta_n,
+                         AElementOp a_element_op,
+                         BElementOp b_element_op,
+                         CDEElementOp cde_element_op,
+                         HElementOp h_element_op,
+                         int M,
+                         int N,
+                         AccDataType epsilon = 1e-5)
+{
+    using ReferenceGemm = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                    BDataType,
+                                                                    AccDataType,
+                                                                    AccDataType,
+                                                                    AElementOp,
+                                                                    BElementOp,
+                                                                    PassThrough>;
+
+    using ReferenceLayernorm = ck::tensor_operation::host::ReferenceLayernorm<EMeanVarDataType,
+                                                                              GammaDataType,
+                                                                              BetaDataType,
+                                                                              HDataType,
+                                                                              AccDataType,
+                                                                              HElementOp,
+                                                                              2,
+                                                                              1>;
+
+    Tensor<EMeanVarDataType> e_m_n(HostTensorDescriptor{M, N});
+    Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
+
+    auto ref_gemm         = ReferenceGemm{};
+    auto ref_gemm_invoker = ref_gemm.MakeInvoker();
+
+    auto ref_gemm_argument =
+        ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+    ref_gemm_invoker.Run(ref_gemm_argument);
+
+    for(int n = 0; n < N; ++n)
+    {
+        AccDataType bias = static_cast<AccDataType>(bias_n(n));
+        for(int m = 0; m < M; ++m)
+        {
+            AccDataType e  = static_cast<AccDataType>(e_m_n(m, n));
+            AccDataType d1 = static_cast<AccDataType>(d1_m_n(m, n));
+            cde_element_op(e, c_m_n(m, n), bias, d1);
+            e_m_n(m, n) = static_cast<EMeanVarDataType>(e);
+        }
+    }
+
+    ReferenceLayernorm ref_layernorm;
+    auto ref_layernorm_invoker = ref_layernorm.MakeInvoker();
+
+    auto ref_layernorm_argument = ref_layernorm.MakeArgument(
+        e_m_n, gamma_n, beta_n, h_m_n, h_element_op, {M, N}, {1}, epsilon);
+    ref_layernorm_invoker.Run(ref_layernorm_argument);
+}
+
+int main()
+{
+    bool do_verification = true;
+
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA  = K;
+    ck::index_t StrideB  = K;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = N;
+    ck::index_t StrideH  = N;
+
+    float epsilon = 1e-5;
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor2d(M, N, StrideD1, D1Layout{}));
+    Tensor<GammaDataType> gamma_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<BetaDataType> beta_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<HDataType> h_m_n(f_host_tensor_descriptor2d(M, N, StrideH, HLayout{}));
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
+    b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
+    d0_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-1, 1});
+    d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-1, 1});
+    gamma_n.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-1, 1});
+    beta_n.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-1, 1});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_n.mDesc.GetElementSpaceSize());
+    DeviceMem beta_device_buf(sizeof(BetaDataType) * beta_n.mDesc.GetElementSpaceSize());
+    DeviceMem h_device_buf(sizeof(HDataType) * h_m_n.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_device_buf.ToDevice(d0_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+    gamma_device_buf.ToDevice(gamma_n.mData.data());
+    beta_device_buf.ToDevice(beta_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+    auto h_element_op   = HElementOp{};
+
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+                               gamma_device_buf.GetDeviceBuffer(),
+                               beta_device_buf.GetDeviceBuffer(),
+                               h_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               {StrideD0, StrideD1},
+                               StrideH,
+                               epsilon,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op,
+                               h_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    size_t workspace_sz = device_op.GetWorkSpaceSize(&argument);
+    DeviceMem workspace_dev(workspace_sz);
+    device_op.SetWorkSpacePointer(&argument, workspace_dev.GetDeviceBuffer());
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        Tensor<HDataType> h_m_n_host(HostTensorDescriptor{M, N});
+        host_gemm_layernorm(h_m_n_host,
+                            a_m_k,
+                            b_k_n,
+                            d0_n,
+                            d1_m_n,
+                            gamma_n,
+                            beta_n,
+                            a_element_op,
+                            b_element_op,
+                            cde_element_op,
+                            h_element_op,
+                            M,
+                            N,
+                            epsilon);
+
+        h_device_buf.FromDevice(h_m_n.mData.data());
+        pass &=
+            ck::utils::check_err(h_m_n, h_m_n_host, "Error: Incorrect results h_m_n", 1e-2, 1e-2);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp b/example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp
new file mode 100644
index 0000000000..bd1d6932ac
--- /dev/null
+++ b/example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp
@@ -0,0 +1,374 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+// DataType
+using ADataType                = F16;
+using BDataType                = F16;
+using GemmAccDataType          = F32;
+using CShuffleDataType         = F32;
+using DsDataType               = ck::Tuple<>;
+using EDataType                = F16;
+using ReduceAccDataType        = F32;
+using R0DataType               = F32;
+using R1DataType               = F32;
+using RsDataType               = ck::Tuple<R0DataType, R1DataType>;
+using GammaDataType            = F16;
+using BetaDataType             = F16;
+using LayerNormOutDataType     = F16;
+using NormalizeComputeDataType = F32;
+
+// Layout
+using ALayout  = Row;
+using BLayout  = Col;
+using D1Layout = Row;
+using ELayout  = D1Layout;
+
+// Elementwise op
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using Square       = ck::tensor_operation::element_wise::UnarySquare;
+using Div          = ck::tensor_operation::element_wise::UnaryDivide;
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+using QsElementOp  = ck::Tuple<PassThrough, Square>;
+using RsElementOp  = ck::Tuple<Div, Div>;
+
+// ReduceOp
+using R0ThreadReduceOp = ck::reduce::Add;
+using R1ThreadReduceOp = ck::reduce::Add;
+using RsThreadReduceOp = ck::Tuple<R0ThreadReduceOp, R1ThreadReduceOp>;
+
+static constexpr auto R0GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+static constexpr auto R1GlobalReduceOp = ck::InMemoryDataOperationEnum::AtomicAdd;
+using RsGlobalReduceOp = ck::InMemoryDataOperationEnumSequence<R0GlobalReduceOp, R1GlobalReduceOp>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+//######| ALayout| BLayout| ELayout|     AData|     BData|     GemmAccData|         CShuffle|     DsData|     EData|     ReduceAccData|     RsData|           A|           B|          CDE|          Qs|          Rs|           Thread|           Global|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|    CDRThreadTransfer|                  CDE|    RThreadTransfer|
+//######|        |        |        |      Type|      Type|            Type|         DataType|       Type|      Type|              Type|       Type| Elementwise| Elementwise|  Elementwise| Elementwise| Elementwise|           Reduce|           Reduce| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|       ClusterLengths| ReduceThreadTransfer| DstScalarPerVector|
+//######|        |        |        |          |          |                |                 |           |          |                  |           |   Operation|   Operation|    Operation|   Operation|   Operation|        Operation|        Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _MPerBlock_NPerBlock|      ScalarPerVector|         _MPerBlock|
+//######|        |        |        |          |          |                |                 |           |          |                  |           |            |            |             |            |            |                 |                 |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |           _NPerBlock|                   |
+        < ALayout, BLayout, ELayout, ADataType, BDataType, GemmAccDataType, CShuffleDataType, DsDataType, EDataType, ReduceAccDataType, RsDataType,  AElementOp,  BElementOp, CDEElementOp, QsElementOp, RsElementOp, RsThreadReduceOp, RsGlobalReduceOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<64, 4>,                    4,                  1>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        EDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+using NormalizeFunctor = ck::tensor_operation::element_wise::Normalize;
+
+// A:x, B:E[x], C:E[x^2], D:Gamma, E:Beta , F:y
+using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
+    ck::Tuple<EDataType,
+              R0DataType,
+              R1DataType,
+              GammaDataType,
+              BetaDataType>,         // x(gemm_out), mean,
+                                     // meansquare,
+                                     // gamma, beta
+    ck::Tuple<LayerNormOutDataType>, // y
+    NormalizeFunctor,
+    2,
+    8,                           // MPerthread
+    ck::Sequence<8, 1, 1, 8, 8>, // scalarPerVector: x(gemm_out), mean, meansquare, gamma, beta
+    ck::Sequence<8>>;            // scalarPerVector: y(layerNorm_out)
+
+auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+    return HostTensorDescriptor({len}, {stride});
+};
+
+auto f_host_tensor_descriptor2d =
+    [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+        using namespace ck::literals;
+
+        if constexpr(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({row, col}, {stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({row, col}, {1_uz, stride});
+        }
+    };
+
+void host_gemm_layernorm(Tensor<LayerNormOutDataType>& out_m_n,
+                         const Tensor<ADataType>& a_m_k,
+                         const Tensor<BDataType>& b_k_n,
+                         const Tensor<GammaDataType>& gamma_n,
+                         const Tensor<BetaDataType>& beta_n,
+                         AElementOp a_element_op,
+                         BElementOp b_element_op,
+                         CDEElementOp c_element_op,
+                         int M,
+                         int N)
+{
+    int StrideE = N;
+    Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> mean_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<R1DataType> meanSquare_m(f_host_tensor_descriptor1d(M, 1));
+    auto averageOpInst = Div{N};
+
+    auto ref_gemm    = ReferenceGemmInstance{};
+    auto ref_invoker = ref_gemm.MakeInvoker();
+
+    auto ref_argument =
+        ref_gemm.MakeArgument(a_m_k, b_k_n, e_m_n, a_element_op, b_element_op, c_element_op);
+
+    ref_invoker.Run(ref_argument);
+
+    // reduce_mean and reduce_square_mean
+    auto r0Op = R0ThreadReduceOp{};
+    auto r1Op = R1ThreadReduceOp{};
+    for(int m = 0; m < M; ++m)
+    {
+        auto mean_acc        = r0Op.GetIdentityValue<ReduceAccDataType>();
+        auto mean_square_acc = r1Op.GetIdentityValue<ReduceAccDataType>();
+
+        for(int n = 0; n < N; ++n)
+        {
+            auto e_val                     = ck::type_convert<ReduceAccDataType>(e_m_n(m, n));
+            ReduceAccDataType square_e_val = 0;
+            Square{}(square_e_val, e_val);
+
+            r0Op(mean_acc, e_val);
+            r1Op(mean_square_acc, square_e_val);
+        }
+
+        averageOpInst(mean_acc, mean_acc);
+        averageOpInst(mean_square_acc, mean_square_acc);
+        mean_m(m)       = ck::type_convert<R0DataType>(mean_acc);
+        meanSquare_m(m) = ck::type_convert<R1DataType>(mean_square_acc);
+    }
+
+    // LayerNorm
+    auto layerNormInst = NormalizeFunctor{};
+    for(int m = 0; m < M; ++m)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            LayerNormOutDataType out_val = 0;
+            layerNormInst(out_val, e_m_n(m, n), mean_m(m), meanSquare_m(m), gamma_n(n), beta_n(n));
+            out_m_n(m, n) = out_val;
+        }
+    }
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename R0DataType,
+          typename R1DataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename NormalizeDataType>
+void DumpGemmLayerNormPerf(float gemm_reduce_time, float normalize_time, int M, int N, int K)
+{
+    std::size_t gemm_flop     = std::size_t(2) * M * N * K;
+    std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
+                                sizeof(R1DataType) * M;
+
+    std::size_t normalize_num_btye = sizeof(EDataType) * M * N + sizeof(R0DataType) * M +
+                                     sizeof(R1DataType) * M + sizeof(GammaDataType) * N +
+                                     sizeof(BetaDataType) * N + sizeof(NormalizeDataType) * M * N;
+
+    float tflops               = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
+    float gemm_gb_per_sec      = gemm_num_byte / 1.E6 / gemm_reduce_time;
+    float normalize_gb_per_sec = normalize_num_btye / 1.E6 / normalize_time;
+
+    std::cout << "gemm + reduce_mean + reduce_square_mean Perf: " << gemm_reduce_time << " ms, "
+              << tflops << " TFlops, " << gemm_gb_per_sec << " GB/s, " << std::endl;
+
+    std::cout << "5-ary elementwise Perf: " << normalize_time << " ms, " << normalize_gb_per_sec
+              << " GB/s, " << std::endl;
+}
+
+int main()
+{
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideE = 1024;
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<EDataType> e_m_n(f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+    Tensor<R0DataType> r0_Mean_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<R1DataType> r1_MeanSquare_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<GammaDataType> gamma_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<BetaDataType> beta_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<LayerNormOutDataType> layerNorm_m_n(
+        f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
+    b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
+    gamma_n.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-1, 1});
+    beta_n.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-1, 1});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem r0_Mean_device_buf(sizeof(R0DataType) * r0_Mean_m.mDesc.GetElementSpaceSize());
+    DeviceMem r1_MeanSquare_device_buf(sizeof(R1DataType) *
+                                       r1_MeanSquare_m.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_n.mDesc.GetElementSpaceSize());
+    DeviceMem beta_device_buf(sizeof(BetaDataType) * beta_n.mDesc.GetElementSpaceSize());
+    DeviceMem layerNorm_device_buf(sizeof(LayerNormOutDataType) *
+                                   layerNorm_m_n.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    gamma_device_buf.ToDevice(gamma_n.mData.data());
+    beta_device_buf.ToDevice(beta_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+    auto qs_element_op  = QsElementOp{};
+    auto rs_element_op  = RsElementOp{N, N};
+
+    // Prepare GEMM, mean, mean_square
+    auto gemmReduce          = DeviceOpInstance{};
+    auto gemmReduce_invoker  = gemmReduce.MakeInvoker();
+    auto gemmReduce_argument = gemmReduce.MakeArgument(
+        a_device_buf.GetDeviceBuffer(),
+        b_device_buf.GetDeviceBuffer(),
+        {},
+        e_device_buf.GetDeviceBuffer(),
+        {r0_Mean_device_buf.GetDeviceBuffer(), r1_MeanSquare_device_buf.GetDeviceBuffer()},
+        M,
+        N,
+        K,
+        StrideA,
+        StrideB,
+        {},
+        StrideE,
+        a_element_op,
+        b_element_op,
+        cde_element_op,
+        qs_element_op,
+        rs_element_op);
+
+    if(!gemmReduce.IsSupportedArgument(gemmReduce_argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    r0_Mean_device_buf.SetZero();
+    r1_MeanSquare_device_buf.SetZero();
+
+    // Prepare LayerNorm
+    std::array<const void*, 5> input = {e_device_buf.GetDeviceBuffer(),
+                                        r0_Mean_device_buf.GetDeviceBuffer(),
+                                        r1_MeanSquare_device_buf.GetDeviceBuffer(),
+                                        gamma_device_buf.GetDeviceBuffer(),
+                                        beta_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {layerNorm_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 2> xyLengths = {M, N};
+    std::array<ck::index_t, 2> xyStrides = {StrideE, 1};
+
+    auto normalize         = DeviceNormalizeInstance{};
+    auto normalize_invoker = normalize.MakeInvoker();
+    auto normalize_argument_ptr =
+        normalize.MakeArgumentPointer(xyLengths,
+                                      {xyStrides, {1, 0}, {1, 0}, {0, 1}, {0, 1}},
+                                      {xyStrides},
+                                      input,
+                                      output,
+                                      NormalizeFunctor{});
+
+    if(!normalize.IsSupportedArgument(normalize_argument_ptr.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device, exiting");
+    }
+
+    // run kernel
+    gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, false});
+    normalize_invoker.Run(normalize_argument_ptr.get(), StreamConfig{nullptr, false});
+
+    bool pass = true;
+    {
+        // verification
+        Tensor<LayerNormOutDataType> host_layerNorm_m_n(
+            f_host_tensor_descriptor2d(M, N, StrideE, ELayout{}));
+
+        host_gemm_layernorm(host_layerNorm_m_n,
+                            a_m_k,
+                            b_k_n,
+                            gamma_n,
+                            beta_n,
+                            a_element_op,
+                            b_element_op,
+                            cde_element_op,
+                            M,
+                            N);
+
+        layerNorm_device_buf.FromDevice(layerNorm_m_n.mData.data());
+        pass &= ck::utils::check_err(
+            layerNorm_m_n, host_layerNorm_m_n, "Error: Incorrect results d1", 1e-3, 1e-3);
+    }
+
+    {
+        // evaluate kernel perf
+        bool time_kernel = true;
+
+        float gemm_reduce_mean_reduce_square_mean_ave_time =
+            gemmReduce_invoker.Run(gemmReduce_argument, StreamConfig{nullptr, time_kernel});
+        float normalize_ave_time =
+            normalize_invoker.Run(normalize_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        if(time_kernel)
+            DumpGemmLayerNormPerf<ADataType,
+                                  BDataType,
+                                  EDataType,
+                                  R0DataType,
+                                  R1DataType,
+                                  GammaDataType,
+                                  BetaDataType,
+                                  LayerNormOutDataType>(
+                gemm_reduce_mean_reduce_square_mean_ave_time, normalize_ave_time, M, N, K);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp b/example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp
new file mode 100644
index 0000000000..90d80f9f03
--- /dev/null
+++ b/example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp
@@ -0,0 +1,287 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+// This example demonstrate a single kernel that runs GEMM layer and laynorm in one fused kernel
+//
+// The GEMM + Layernorm implementation is a specialized kernel which allows fusing both layers
+// together given the condition GEMM extents N of MNK is spanned by a single workgroup. For example,
+// a kernel configured with NPerBlock = 128 allows to operate on all GEMM sizes if N <= 128
+//
+// D = Layernorm(acc_element_op(A * B + broadcast(bias)) + add) * broadcast(gamma) + broadcast(beta)
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using CDataType        = F16;
+using C0DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+struct Relu
+{
+    template <typename OutT, typename InT>
+    __host__ __device__ void operator()(OutT& y, const InT& x) const
+    {
+        y = x > 0 ? x : 0;
+    }
+};
+
+using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+// Elementwise operation that operates on the output of matrix multiplication
+// i.e., AccElementOp(A * B + bias)
+using AccElementOp = Relu;
+// Elementwise operation that operates on the output of layer normalization
+using CElementOp = Relu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmLayerNorm_Xdl_CShuffle
+//######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     C0Data|     GemmAcc|         CShuffle|   ReduceAcc|           A|           B|          Acc|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce|     CReduceThreadCopy|
+//######|        |        |        |      Type|      Type|      Type|       Type|    DataType|         DataType|    DataType| Elementwise| Elementwise|  Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector|
+//######|        |        |        |          |          |          |           |            |                 |            |   Operation|   Operation|    Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|            _NPerBlock|
+//######|        |        |        |          |          |          |           |            |                 |            |            |            |             |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                      |
+        <     Row,     Col,     Row, ADataType, BDataType, CDataType, C0DataType, AccDataType, CShuffleDataType, AccDataType,  AElementOp,  BElementOp, AccElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           2,               S<1, 32, 1, 8>,               8,             S<64, 4>,                     4>;
+// clang-format on
+
+using ReferenceInstance = ck::tensor_operation::host::ReferenceGemmLayernorm<ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             C0DataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             AccElementOp,
+                                                                             CElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 128;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 128;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if constexpr(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<AccDataType> acc_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<C0DataType> c0_n_bias({N});
+    Tensor<C0DataType> c0_m_n_add(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<C0DataType> c0_n_gamma({N});
+    Tensor<C0DataType> c0_n_beta({N});
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+    std::cout << "c0_n_bias: " << c0_n_bias.mDesc << std::endl;
+    std::cout << "c0_m_n_add: " << c0_m_n_add.mDesc << std::endl;
+    std::cout << "c0_n_gamma: " << c0_n_gamma.mDesc << std::endl;
+    std::cout << "c0_n_beta: " << c0_n_beta.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+        b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+    }
+
+    c0_n_bias.GenerateTensorValue(GeneratorTensor_2<C0DataType>{-5, 5});
+    c0_m_n_add.GenerateTensorValue(GeneratorTensor_2<C0DataType>{-5, 5});
+    c0_n_gamma.GenerateTensorValue(GeneratorTensor_2<C0DataType>{0, 2});
+    c0_n_beta.GenerateTensorValue(GeneratorTensor_2<C0DataType>{0, 5});
+    c_m_n_host_result.GenerateTensorValue(GeneratorTensor_1<CDataType>{0});
+    acc_m_n_host_result.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem c0_bias_buf(sizeof(C0DataType) * c0_n_bias.mDesc.GetElementSpaceSize());
+    DeviceMem c0_add_buf(sizeof(C0DataType) * c0_m_n_add.mDesc.GetElementSpaceSize());
+    DeviceMem c0_gamma_buf(sizeof(C0DataType) * c0_n_gamma.mDesc.GetElementSpaceSize());
+    DeviceMem c0_beta_buf(sizeof(C0DataType) * c0_n_beta.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    c0_bias_buf.ToDevice(c0_n_bias.mData.data());
+    c0_add_buf.ToDevice(c0_m_n_add.mData.data());
+    c0_gamma_buf.ToDevice(c0_n_gamma.mData.data());
+    c0_beta_buf.ToDevice(c0_n_beta.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto acc_element_op = AccElementOp{};
+    auto c_element_op   = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                      static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                      static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                      static_cast<C0DataType*>(c0_add_buf.GetDeviceBuffer()),
+                                      static_cast<C0DataType*>(c0_bias_buf.GetDeviceBuffer()),
+                                      static_cast<C0DataType*>(c0_gamma_buf.GetDeviceBuffer()),
+                                      static_cast<C0DataType*>(c0_beta_buf.GetDeviceBuffer()),
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      StrideC,
+                                      a_element_op,
+                                      b_element_op,
+                                      acc_element_op,
+                                      c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    // extra 6MN flops due to: bias + add + gamma + beta + norm_sub + norm_div,
+    // excluding reduction steps
+    std::size_t flop = std::size_t(2) * M * N * K + std::size_t(6) * M * N;
+    // extra MN and 3N due to c0_add (MxN), bias (1xN), gamma (1xN), beta (1xN)
+    std::size_t bytes = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                        sizeof(CDataType) * 2 * M * N + sizeof(C0DataType) * 3 * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = bytes / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+        auto ref_gemm    = ReferenceInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(a_m_k,
+                                                  b_k_n,
+                                                  c_m_n_host_result,
+                                                  c0_n_bias,
+                                                  c0_m_n_add,
+                                                  c0_n_gamma,
+                                                  c0_n_beta,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  acc_element_op,
+                                                  c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        if constexpr(std::is_same<CShuffleDataType, F32>::value)
+        {
+            pass &= ck::utils::check_err(
+                c_m_n_device_result, c_m_n_host_result, "Error: Incorrect results c");
+        }
+        else if constexpr(std::is_same<CShuffleDataType, F16>::value)
+        {
+            pass &= ck::utils::check_err(
+                c_m_n_device_result, c_m_n_host_result, "Error: Incorrect results c", 1e-2, 1e-2);
+        }
+    }
+    return pass ? 0 : 1;
+}
diff --git a/example/22_cgemm/CMakeLists.txt b/example/22_cgemm/CMakeLists.txt
new file mode 100644
index 0000000000..854f07fda6
--- /dev/null
+++ b/example/22_cgemm/CMakeLists.txt
@@ -0,0 +1,22 @@
+add_custom_target(example_cgemm_xdl)
+
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)
+  add_dependencies(example_cgemm_xdl example_cgemm_xdl_bf16)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
+  add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp16)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+add_example_executable(example_cgemm_xdl_fp32 cgemm_xdl_fp32.cpp)
+add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp32)
+endif()
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+  add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp)
+  add_dependencies(example_cgemm_xdl example_cgemm_xdl_int8)
+endif()
+if(USE_BITINT_EXTENSION_INT4)
+  add_example_executable(example_cgemm_xdl_int4 cgemm_xdl_int4.cpp)
+  add_dependencies(example_cgemm_xdl example_cgemm_xdl_int4)
+endif()
diff --git a/example/22_cgemm/cgemm_xdl_bf16.cpp b/example/22_cgemm/cgemm_xdl_bf16.cpp
new file mode 100644
index 0000000000..fa4482a984
--- /dev/null
+++ b/example/22_cgemm/cgemm_xdl_bf16.cpp
@@ -0,0 +1,132 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "cgemm_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType   = BF16;
+using BDataType   = BF16;
+using CDataType   = BF16;
+using AccDataType = F32;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
+// clang-format off
+using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
+    <ALayout,                    // typename ALayout
+     BLayout,                    // typename BLayout
+     CLayout,                    // typename CLayout
+     ADataType,                  // typename ADataType
+     BDataType,                  // typename BDataType
+     CDataType,                  // typename CDataType
+     AccDataType,                // typename GemmAccDataType
+     CDataType,                  // typename CShuffleDataType
+     PassThrough,                // typename AElementwiseOperation
+     PassThrough,                // typename BElementwiseOperation
+     PassThrough,                // typename CElementwiseOperation
+     GemmDefault,                // GemmSpecialization GemmSpec
+     1,                          // index_t NumGemmKPrefetchStage
+     256,                        // index_t BlockSize
+     256,                        // index_t MPerBlock
+     128,                        // index_t NPerBlock
+     32,                         // index_t KPerBlock
+     8,                          // index_t AK1
+     8,                          // index_t BK1
+     32,                         // index_t MPerXDL
+     32,                         // index_t NPerXDL
+     4,                          // index_t MXdlPerWave
+     2,                          // index_t NXdlPerWave
+     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
+     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder
+     2,                          // index_t ABlockTransferSrcVectorDim
+     8,                          // index_t ABlockTransferSrcScalarPerVector
+     8,                          // index_t ABlockTransferDstScalarPerVector_AK1
+     1,                          // index_t ABlockLdsExtraM
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder
+     2,                          // index_t BBlockTransferSrcVectorDim
+     8,                          // index_t BBlockTransferSrcScalarPerVector
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1
+     1,                          // index_t BBlockLdsExtraN
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle
+     S<1, 32, 1, 8>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+     8>;                         // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // CGEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 416;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << "arg3: run kernel # of times (>1)\n"
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
+                  << std::endl;
+        exit(0);
+    }
+
+    return !run_cgemm_xdl<ADataType,
+                          BDataType,
+                          CDataType,
+                          ALayout,
+                          BLayout,
+                          CLayout,
+                          PassThrough,
+                          PassThrough,
+                          PassThrough,
+                          DeviceCGemmInstance,
+                          ReferenceCGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
+}
diff --git a/example/22_cgemm/cgemm_xdl_common.hpp b/example/22_cgemm/cgemm_xdl_common.hpp
new file mode 100644
index 0000000000..26137a7c2e
--- /dev/null
+++ b/example/22_cgemm/cgemm_xdl_common.hpp
@@ -0,0 +1,254 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/stream_config.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16   = ck::half_t;
+using F32   = float;
+using BF16  = ck::bhalf_t;
+using INT8  = std::int8_t;
+using INT32 = std::int32_t;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+using INT4 = ck::int4_t;
+#endif
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename DeviceCGemmInstance,
+          typename ReferenceCGemmInstance,
+          typename KernelADataType = ADataType,
+          typename KernelBDataType = BDataType,
+          typename KernelCDataType = CDataType>
+bool run_cgemm_xdl(ck::index_t M,
+                   ck::index_t N,
+                   ck::index_t K,
+                   ck::index_t StrideA,
+                   ck::index_t StrideB,
+                   ck::index_t StrideC,
+                   bool do_verification,
+                   int init_method,
+                   bool time_kernel)
+{
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t),
+                  "sizeof ck::int4_t and int8_t is different!");
+    static_assert(sizeof(ADataType) == sizeof(KernelADataType),
+                  "sizeof ADataType and KernelADataType is different!");
+    static_assert(sizeof(BDataType) == sizeof(KernelBDataType),
+                  "sizeof BDataType and KernelBDataType is different!");
+    static_assert(sizeof(CDataType) == sizeof(KernelCDataType),
+                  "sizeof CDataType and KernelCDataType is different!");
+#endif
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k_real(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<ADataType> a_m_k_imag(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n_real(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<BDataType> b_k_n_imag(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<KernelCDataType> c_m_n_real_device_result(
+        f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<KernelCDataType> c_m_n_imag_device_result(
+        f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k_real: " << a_m_k_real.mDesc << std::endl;
+    std::cout << "a_m_k_imag: " << a_m_k_imag.mDesc << std::endl;
+    std::cout << "b_k_n_real: " << b_k_n_real.mDesc << std::endl;
+    std::cout << "b_k_n_imag: " << b_k_n_imag.mDesc << std::endl;
+    std::cout << "c_m_n_real: " << c_m_n_real_device_result.mDesc << std::endl;
+    std::cout << "c_m_n_imag: " << c_m_n_imag_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k_real.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        a_m_k_imag.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n_real.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        b_k_n_imag.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        break;
+    default:
+        a_m_k_real.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
+        a_m_k_imag.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
+        b_k_n_real.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        b_k_n_imag.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    auto cgemm = DeviceCGemmInstance{};
+
+    DeviceMem a_m_k_real_device_buf(sizeof(KernelADataType) *
+                                    a_m_k_real.mDesc.GetElementSpaceSize());
+    DeviceMem a_m_k_imag_device_buf(sizeof(KernelADataType) *
+                                    a_m_k_imag.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_real_device_buf(sizeof(KernelBDataType) *
+                                    b_k_n_real.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_imag_device_buf(sizeof(KernelBDataType) *
+                                    b_k_n_imag.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_real_device_buf(sizeof(KernelCDataType) *
+                                    c_m_n_real_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_imag_device_buf(sizeof(KernelCDataType) *
+                                    c_m_n_imag_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem workspace_device_buf(cgemm.GetWorkspaceSize(M, N, K, StrideA, StrideB, StrideC));
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    if constexpr(std::is_same_v<ADataType, ck::int4_t>)
+    {
+        Tensor<KernelADataType> a_m_k_real_converted(a_m_k_real);
+        Tensor<KernelADataType> a_m_k_imag_converted(a_m_k_imag);
+        Tensor<KernelBDataType> b_k_n_real_converted(b_k_n_real);
+        Tensor<KernelBDataType> b_k_n_imag_converted(b_k_n_imag);
+
+        a_m_k_real_device_buf.ToDevice(a_m_k_real_converted.mData.data());
+        a_m_k_imag_device_buf.ToDevice(a_m_k_imag_converted.mData.data());
+        b_k_n_real_device_buf.ToDevice(b_k_n_real_converted.mData.data());
+        b_k_n_imag_device_buf.ToDevice(b_k_n_imag_converted.mData.data());
+    }
+    else
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    {
+        a_m_k_real_device_buf.ToDevice(a_m_k_real.mData.data());
+        a_m_k_imag_device_buf.ToDevice(a_m_k_imag.mData.data());
+        b_k_n_real_device_buf.ToDevice(b_k_n_real.mData.data());
+        b_k_n_imag_device_buf.ToDevice(b_k_n_imag.mData.data());
+    }
+
+    auto a_element_op = AElementwiseOperation{};
+    auto b_element_op = BElementwiseOperation{};
+    auto c_element_op = CElementwiseOperation{};
+
+    // do GEMM
+    auto invoker = cgemm.MakeInvoker();
+    auto argument =
+        cgemm.MakeArgument(static_cast<KernelADataType*>(a_m_k_real_device_buf.GetDeviceBuffer()),
+                           static_cast<KernelADataType*>(a_m_k_imag_device_buf.GetDeviceBuffer()),
+                           static_cast<KernelBDataType*>(b_k_n_real_device_buf.GetDeviceBuffer()),
+                           static_cast<KernelBDataType*>(b_k_n_imag_device_buf.GetDeviceBuffer()),
+                           static_cast<KernelCDataType*>(c_m_n_real_device_buf.GetDeviceBuffer()),
+                           static_cast<KernelCDataType*>(c_m_n_imag_device_buf.GetDeviceBuffer()),
+                           static_cast<KernelCDataType*>(workspace_device_buf.GetDeviceBuffer()),
+                           M,
+                           N,
+                           K,
+                           StrideA,
+                           StrideB,
+                           StrideC,
+                           a_element_op,
+                           b_element_op,
+                           c_element_op);
+
+    if(!cgemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_cgemm with the specified compilation parameters does "
+            "not support this CGEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(8) * M * N * K;
+    std::size_t num_btype =
+        std::size_t(2) *
+        (sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N);
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << cgemm.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        Tensor<CDataType> c_m_n_real_host_result(
+            f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+        Tensor<CDataType> c_m_n_imag_host_result(
+            f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+        auto ref_cgemm    = ReferenceCGemmInstance{};
+        auto ref_invoker  = ref_cgemm.MakeInvoker();
+        auto ref_argument = ref_cgemm.MakeArgument(a_m_k_real,
+                                                   a_m_k_imag,
+                                                   b_k_n_real,
+                                                   b_k_n_imag,
+                                                   c_m_n_real_host_result,
+                                                   c_m_n_imag_host_result,
+                                                   a_element_op,
+                                                   b_element_op,
+                                                   c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        c_m_n_real_device_buf.FromDevice(c_m_n_real_device_result.mData.data());
+        c_m_n_imag_device_buf.FromDevice(c_m_n_imag_device_result.mData.data());
+
+        bool result = true;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        if constexpr(std::is_same_v<ADataType, ck::int4_t>)
+        {
+            const Tensor<CDataType> c_m_n_real_device_result_converted(c_m_n_real_device_result);
+            const Tensor<CDataType> c_m_n_imag_device_result_converted(c_m_n_imag_device_result);
+
+            result = ck::utils::check_err(c_m_n_real_device_result_converted,
+                                          c_m_n_real_host_result,
+                                          "Verification error: incorrect results in real part!",
+                                          1e-2f,
+                                          1e-1f);
+            result = result && ck::utils::check_err(
+                                   c_m_n_imag_device_result_converted,
+                                   c_m_n_imag_host_result,
+                                   "Verification error: incorrect results in imaginary part!",
+                                   1e-2f,
+                                   1e-1f);
+        }
+        else
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+        {
+            result = ck::utils::check_err(c_m_n_real_device_result,
+                                          c_m_n_real_host_result,
+                                          "Verification error: incorrect results in real part!",
+                                          1e-2f,
+                                          1e-1f);
+            result = result && ck::utils::check_err(
+                                   c_m_n_imag_device_result,
+                                   c_m_n_imag_host_result,
+                                   "Verification error: incorrect results in imaginary part!",
+                                   1e-2f,
+                                   1e-1f);
+        }
+
+        return result;
+    }
+    return true;
+}
diff --git a/example/22_cgemm/cgemm_xdl_fp16.cpp b/example/22_cgemm/cgemm_xdl_fp16.cpp
new file mode 100644
index 0000000000..89a581e865
--- /dev/null
+++ b/example/22_cgemm/cgemm_xdl_fp16.cpp
@@ -0,0 +1,131 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "cgemm_xdl_common.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType        = F16;
+using BDataType        = F16;
+using CDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
+// clang-format off
+using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
+    <ALayout,                    // typename ALayout
+     BLayout,                    // typename BLayout
+     CLayout,                    // typename CLayout
+     ADataType,                  // typename ADataType
+     BDataType,                  // typename BDataType
+     CDataType,                  // typename CDataType
+     AccDataType,                // typename GemmAccDataType
+     CShuffleDataType,           // typename CShuffleDataType
+     PassThrough,                // typename AElementwiseOperation
+     PassThrough,                // typename BElementwiseOperation
+     PassThrough,                // typename CElementwiseOperation
+     GemmDefault,                // GemmSpecialization GemmSpec
+     1,                          // index_t NumGemmKPrefetchStage
+     256,                        // index_t BlockSize
+     256,                        // index_t MPerBlock
+     128,                        // index_t NPerBlock
+     32,                         // index_t KPerBlock
+     8,                          // index_t AK1
+     8,                          // index_t BK1
+     32,                         // index_t MPerXDL
+     32,                         // index_t NPerXDL
+     4,                          // index_t MXdlPerWave
+     2,                          // index_t NXdlPerWave
+     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
+     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder
+     2,                          // index_t ABlockTransferSrcVectorDim
+     8,                          // index_t ABlockTransferSrcScalarPerVector
+     8,                          // index_t ABlockTransferDstScalarPerVector_AK1
+     1,                          // index_t ABlockLdsExtraM
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder
+     2,                          // index_t BBlockTransferSrcVectorDim
+     8,                          // index_t BBlockTransferSrcScalarPerVector
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1
+     1,                          // index_t BBlockLdsExtraN
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle
+     S<1, 32, 1, 8>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+     8>;                         // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // CGEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << "arg3: run kernel # of times (>1)\n"
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
+                  << std::endl;
+        exit(0);
+    }
+
+    return !run_cgemm_xdl<ADataType,
+                          BDataType,
+                          CDataType,
+                          ALayout,
+                          BLayout,
+                          CLayout,
+                          PassThrough,
+                          PassThrough,
+                          PassThrough,
+                          DeviceCGemmInstance,
+                          ReferenceCGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
+}
diff --git a/example/22_cgemm/cgemm_xdl_fp32.cpp b/example/22_cgemm/cgemm_xdl_fp32.cpp
new file mode 100644
index 0000000000..cf96599599
--- /dev/null
+++ b/example/22_cgemm/cgemm_xdl_fp32.cpp
@@ -0,0 +1,132 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "cgemm_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType   = F32;
+using BDataType   = F32;
+using CDataType   = F32;
+using AccDataType = F32;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
+// clang-format off
+using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
+    <ALayout,                    // typename ALayout
+     BLayout,                    // typename BLayout
+     CLayout,                    // typename CLayout
+     ADataType,                  // typename ADataType
+     BDataType,                  // typename BDataType
+     CDataType,                  // typename CDataType
+     AccDataType,                // typename GemmAccDataType
+     CDataType,                  // typename CShuffleDataType
+     PassThrough,                // typename AElementwiseOperation
+     PassThrough,                // typename BElementwiseOperation
+     PassThrough,                // typename CElementwiseOperation
+     GemmDefault,                // GemmSpecialization GemmSpec
+     1,                          // index_t NumGemmKPrefetchStage
+     256,                        // index_t BlockSize
+     256,                        // index_t MPerBlock
+     128,                        // index_t NPerBlock
+     16,                         // index_t KPerBlock
+     4,                          // index_t AK1
+     4,                          // index_t BK1
+     32,                         // index_t MPerXDL
+     32,                         // index_t NPerXDL
+     4,                          // index_t MXdlPerWave
+     2,                          // index_t NXdlPerWave
+     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
+     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder
+     2,                          // index_t ABlockTransferSrcVectorDim
+     4,                          // index_t ABlockTransferSrcScalarPerVector
+     4,                          // index_t ABlockTransferDstScalarPerVector_AK1
+     1,                          // index_t ABlockLdsExtraM
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder
+     2,                          // index_t BBlockTransferSrcVectorDim
+     4,                          // index_t BBlockTransferSrcScalarPerVector
+     4,                          // index_t BBlockTransferDstScalarPerVector_BK1
+     1,                          // index_t BBlockLdsExtraN
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle
+     S<1, 16, 1, 16>,            // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+     4>;                         // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // CGEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << "arg3: run kernel # of times (>1)\n"
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
+                  << std::endl;
+        exit(0);
+    }
+
+    return !run_cgemm_xdl<ADataType,
+                          BDataType,
+                          CDataType,
+                          ALayout,
+                          BLayout,
+                          CLayout,
+                          PassThrough,
+                          PassThrough,
+                          PassThrough,
+                          DeviceCGemmInstance,
+                          ReferenceCGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
+}
diff --git a/example/22_cgemm/cgemm_xdl_int4.cpp b/example/22_cgemm/cgemm_xdl_int4.cpp
new file mode 100644
index 0000000000..f69cc2b3cc
--- /dev/null
+++ b/example/22_cgemm/cgemm_xdl_int4.cpp
@@ -0,0 +1,140 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "cgemm_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType        = INT4;
+using BDataType        = INT4;
+using CDataType        = INT4;
+using AccDataType      = INT32;
+using CShuffleDataType = INT32;
+
+using KernelADataType = INT8;
+using KernelBDataType = INT8;
+using KernelCDataType = INT8;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
+// clang-format off
+using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
+    <ALayout,                    // typename ALayout
+     BLayout,                    // typename BLayout
+     CLayout,                    // typename CLayout
+     KernelADataType,            // typename ADataType
+     KernelBDataType,            // typename BDataType
+     KernelCDataType,            // typename CDataType
+     AccDataType,                // typename GemmAccDataType
+     CShuffleDataType,           // typename CShuffleDataType
+     PassThrough,                // typename AElementwiseOperation
+     PassThrough,                // typename BElementwiseOperation
+     PassThrough,                // typename CElementwiseOperation
+     GemmDefault,                // GemmSpecialization GemmSpec
+     1,                          // index_t NumGemmKPrefetchStage
+     256,                        // index_t BlockSize
+     256,                        // index_t MPerBlock
+     128,                        // index_t NPerBlock
+     64,                         // index_t KPerBlock
+     16,                         // index_t AK1
+     16,                         // index_t BK1
+     32,                         // index_t MPerXDL
+     32,                         // index_t NPerXDL
+     4,                          // index_t MXdlPerWave
+     2,                          // index_t NXdlPerWave
+     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
+     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder
+     2,                          // index_t ABlockTransferSrcVectorDim
+     16,                         // index_t ABlockTransferSrcScalarPerVector
+     16,                         // index_t ABlockTransferDstScalarPerVector_AK1
+     1,                          // index_t ABlockLdsExtraM
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder
+     2,                          // index_t BBlockTransferSrcVectorDim
+     8,                          // index_t BBlockTransferSrcScalarPerVector
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1
+     1,                          // index_t BBlockLdsExtraN
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle
+     S<1, 64, 1, 4>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+     16>;                        // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // CGEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1152;
+    ck::index_t K = 512;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideC = N;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << "arg3: time kernel (0=no, 1=yes)\n"
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
+                  << std::endl;
+        exit(EXIT_SUCCESS);
+    }
+
+    return !run_cgemm_xdl<ADataType,
+                          BDataType,
+                          CDataType,
+                          ALayout,
+                          BLayout,
+                          CLayout,
+                          PassThrough,
+                          PassThrough,
+                          PassThrough,
+                          DeviceCGemmInstance,
+                          ReferenceCGemmInstance,
+                          KernelADataType,
+                          KernelBDataType,
+                          KernelCDataType>(
+        M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
+}
diff --git a/example/22_cgemm/cgemm_xdl_int8.cpp b/example/22_cgemm/cgemm_xdl_int8.cpp
new file mode 100644
index 0000000000..c4835b853e
--- /dev/null
+++ b/example/22_cgemm/cgemm_xdl_int8.cpp
@@ -0,0 +1,132 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "cgemm_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType   = INT8;
+using BDataType   = INT8;
+using CDataType   = INT8;
+using AccDataType = INT32;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
+// clang-format off
+using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
+    <ALayout,                    // typename ALayout
+     BLayout,                    // typename BLayout
+     CLayout,                    // typename CLayout
+     ADataType,                  // typename ADataType
+     BDataType,                  // typename BDataType
+     CDataType,                  // typename CDataType
+     AccDataType,                // typename GemmAccDataType
+     CDataType,                  // typename CShuffleDataType
+     PassThrough,                // typename AElementwiseOperation
+     PassThrough,                // typename BElementwiseOperation
+     PassThrough,                // typename CElementwiseOperation
+     GemmDefault,                // GemmSpecialization GemmSpec
+     1,                          // index_t NumGemmKPrefetchStage
+     256,                        // index_t BlockSize
+     256,                        // index_t MPerBlock
+     128,                        // index_t NPerBlock
+     64,                         // index_t KPerBlock
+     16,                         // index_t AK1
+     16,                         // index_t BK1
+     32,                         // index_t MPerXDL
+     32,                         // index_t NPerXDL
+     4,                          // index_t MXdlPerWave
+     2,                          // index_t NXdlPerWave
+     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
+     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder
+     2,                          // index_t ABlockTransferSrcVectorDim
+     16,                         // index_t ABlockTransferSrcScalarPerVector
+     16,                         // index_t ABlockTransferDstScalarPerVector_AK1
+     1,                          // index_t ABlockLdsExtraM
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder
+     2,                          // index_t BBlockTransferSrcVectorDim
+     8,                          // index_t BBlockTransferSrcScalarPerVector
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1
+     1,                          // index_t BBlockLdsExtraN
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle
+     S<1, 64, 1, 4>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+     16>;                        // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // CGEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << "arg3: run kernel # of times (>1)\n"
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
+                  << std::endl;
+        exit(0);
+    }
+
+    return !run_cgemm_xdl<ADataType,
+                          BDataType,
+                          CDataType,
+                          ALayout,
+                          BLayout,
+                          CLayout,
+                          PassThrough,
+                          PassThrough,
+                          PassThrough,
+                          DeviceCGemmInstance,
+                          ReferenceCGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
+}
diff --git a/example/23_softmax/CMakeLists.txt b/example/23_softmax/CMakeLists.txt
new file mode 100644
index 0000000000..dafe65521a
--- /dev/null
+++ b/example/23_softmax/CMakeLists.txt
@@ -0,0 +1 @@
+add_example_executable(example_softmax_blockwise softmax_blockwise.cpp)
\ No newline at end of file
diff --git a/example/23_softmax/README.md b/example/23_softmax/README.md
new file mode 100644
index 0000000000..37c43e9b55
--- /dev/null
+++ b/example/23_softmax/README.md
@@ -0,0 +1,18 @@
+# Instructions for ```example_softmax_blockwise```
+
+## Run ```example_softmax_blockwise```
+```bash
+# -D <xxx> : input 3-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+#arg1: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg2: time kernel (0=no, 1=yes)
+example_softmax_blockwise -D 4,128,2048 -v 1 1 1
+```
+
+Result
+```
+launch_and_time_kernel: grid_dim {64, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 0.0242877 ms, 259.039 GB/s, DeviceReduceSoftmax<256,M_C8_S1,K_C32_S8,InSrcVectorDim_1_InSrcVectorSize_8_OutDstVectorSize_8>
+```
diff --git a/example/23_softmax/softmax_blockwise.cpp b/example/23_softmax/softmax_blockwise.cpp
new file mode 100644
index 0000000000..d09e434bcf
--- /dev/null
+++ b/example/23_softmax/softmax_blockwise.cpp
@@ -0,0 +1,264 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+using namespace ck::tensor_operation::device;
+
+using InDataType  = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 3;
+constexpr int NumReduceDim = 1;
+
+using DeviceInstance = DeviceSoftmaxImpl<InDataType,
+                                         AccDataType,
+                                         OutDataType,
+                                         PassThrough, // InElementwiseOperation
+                                         PassThrough, // AccElementwiseOperation
+                                         Rank,
+                                         NumReduceDim,
+                                         256, // BlockSize
+                                         8,   // ClusterM
+                                         32,  // ClusterK
+                                         1,   // SliceM
+                                         8,   // SliceK
+                                         1,   // SrcVecDim (0=M, 1=K)
+                                         8,   // SrcScalarPerVector
+                                         8>;  // OutScalarPerVector
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths = {8, 128, 2048};
+    std::vector<double> scales    = {2.0, 2.0};
+
+    bool do_verification = true;
+    int init_method      = 2;
+    bool time_kernel     = true;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
+                     "comparing with the host-based reduction"
+                  << std::endl;
+        std::cout << "Arg1 -- init method (0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg2 -- time kernel (0=no, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:l:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 2 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(scales.empty())
+        {
+            scales.push_back(1.0f);
+            scales.push_back(0.0f);
+        };
+
+        return (0);
+    };
+};
+
+int main(int argc, char* argv[])
+{
+    // Example: batched gemm C[G, M, N] applies max/sum reduction along N internally
+    const std::vector<int> invariantDims{0, 1};
+    const std::vector<int> reduceDims{2};
+
+    SimpleAppArgs args;
+
+    if(argc > 1)
+    {
+        if(args.processArgs(argc, argv) < 0)
+            return (-1);
+    };
+
+    Tensor<InDataType> in(args.inLengths);
+    Tensor<OutDataType> out_ref(args.inLengths);
+    Tensor<OutDataType> out(args.inLengths);
+
+    auto inStrides  = in.mDesc.GetStrides();
+    auto outStrides = out.mDesc.GetStrides();
+
+    double alpha = args.scales[0];
+    double beta  = args.scales[1];
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "out: " << out.mDesc << std::endl;
+
+    std::size_t num_thread = 1;
+
+    if(args.do_verification)
+    {
+        switch(args.init_method)
+        {
+        case 0: break;
+        case 1:
+            in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1}, num_thread);
+            break;
+        case 2:
+            in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0}, num_thread);
+            if(beta != 0.0f)
+                out_ref.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-5.0, 5.0}, num_thread);
+        }
+
+        if(beta != 0.0f)
+            for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
+                out.mData[i] = out_ref.mData[i];
+    };
+    // std::cout << "beta = " << beta << std::endl;
+    // LogRangeAsType<float>(std::cout << "tensor in: " , in.mData, ",") << std::endl;
+    // LogRangeAsType<float>(std::cout << "tensor prior out: " , out.mData, ",") << std::endl;
+
+    // these buffers are usually provided by the user application
+    DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
+
+    in_dev.ToDevice(in.mData.data());
+
+    if(beta != 0.0f)
+        out_dev.ToDevice(out.mData.data());
+
+    if(args.do_verification)
+    {
+        using ReferenceInstance =
+            ck::tensor_operation::host::ReferenceSoftmax<InDataType, OutDataType, AccDataType>;
+        ReferenceInstance ref;
+        auto ref_arg = ref.MakeArgument(in, out_ref, alpha, beta, reduceDims);
+        auto invoker = ref.MakeInvoker();
+        invoker.Run(ref_arg);
+        // LogRangeAsType<float>(std::cout << "tensor out_ref: ", out_ref.mData, ",") << std::endl;
+    };
+
+    std::vector<ck::index_t> i_inLengths;
+    std::vector<ck::index_t> i_inStrides;
+
+    i_inLengths.assign(args.inLengths.begin(), args.inLengths.end());
+    i_inStrides.assign(inStrides.begin(), inStrides.end());
+
+    auto device_instance = DeviceInstance{};
+
+    std::cout << i_inLengths.size() << ", " << i_inStrides.size() << std::endl;
+
+    auto argument_ptr = device_instance.MakeArgumentPointer(i_inLengths,
+                                                            i_inStrides,
+                                                            reduceDims,
+                                                            alpha,
+                                                            beta,
+                                                            in_dev.GetDeviceBuffer(),
+                                                            out_dev.GetDeviceBuffer(),
+                                                            PassThrough{},
+                                                            PassThrough{});
+
+    if(!device_instance.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+        return 1;
+    };
+
+    std::string instance_name = device_instance.GetTypeString();
+
+    auto invoker_ptr = device_instance.MakeInvokerPointer();
+
+    bool pass = true;
+    if(args.do_verification)
+    {
+        invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        out_dev.FromDevice(out.mData.data());
+        // LogRangeAsType<float>(std::cout << "tensor out: " , out.mData, ",") << std::endl;
+        pass = pass && ck::utils::check_err(out, out_ref);
+    };
+
+    float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, args.time_kernel});
+
+    std::size_t num_bytes =
+        in.mDesc.GetElementSize() * sizeof(InDataType) +
+        (beta == 0.0f ? 1 : 2) * out.mDesc.GetElementSize() * sizeof(OutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+    std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << instance_name
+              << std::endl;
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/24_batched_gemm/CMakeLists.txt b/example/24_batched_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..48a3b58ff5
--- /dev/null
+++ b/example/24_batched_gemm/CMakeLists.txt
@@ -0,0 +1,21 @@
+add_custom_target(example_batched_gemm_xdl)
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+  add_example_executable(example_batched_gemm_xdl_fp32 batched_gemm_xdl_fp32.cpp)
+  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp32)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_batched_gemm_xdl_fp16 batched_gemm_xdl_fp16.cpp)
+  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp16)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+  add_example_executable(example_batched_gemm_xdl_bfp16 batched_gemm_xdl_bfp16.cpp)
+  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_bfp16)
+endif()
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+  add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp)
+  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int8)
+endif()
+if(USE_BITINT_EXTENSION_INT4)
+  add_example_executable(example_batched_gemm_xdl_int4 batched_gemm_xdl_int4.cpp)
+  add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int4)
+endif()
diff --git a/example/24_batched_gemm/batched_gemm_xdl_bfp16.cpp b/example/24_batched_gemm/batched_gemm_xdl_bfp16.cpp
new file mode 100644
index 0000000000..c684c13d0d
--- /dev/null
+++ b/example/24_batched_gemm/batched_gemm_xdl_bfp16.cpp
@@ -0,0 +1,59 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = BF16;
+using BDataType        = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = BF16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+#include "run_batched_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
diff --git a/example/24_batched_gemm/batched_gemm_xdl_fp16.cpp b/example/24_batched_gemm/batched_gemm_xdl_fp16.cpp
new file mode 100644
index 0000000000..d1985f9af5
--- /dev/null
+++ b/example/24_batched_gemm/batched_gemm_xdl_fp16.cpp
@@ -0,0 +1,59 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+#include "run_batched_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
diff --git a/example/24_batched_gemm/batched_gemm_xdl_fp32.cpp b/example/24_batched_gemm/batched_gemm_xdl_fp32.cpp
new file mode 100644
index 0000000000..a92a04dbe6
--- /dev/null
+++ b/example/24_batched_gemm/batched_gemm_xdl_fp32.cpp
@@ -0,0 +1,58 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F32;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 32, 1, 8>,               4>;
+// clang-format on
+
+#include "run_batched_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
diff --git a/example/24_batched_gemm/batched_gemm_xdl_int4.cpp b/example/24_batched_gemm/batched_gemm_xdl_int4.cpp
new file mode 100644
index 0000000000..5e82cfe324
--- /dev/null
+++ b/example/24_batched_gemm/batched_gemm_xdl_int4.cpp
@@ -0,0 +1,99 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = ck::int4_t;
+using BDataType        = ck::int4_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int32_t;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = ck::int4_t;
+
+using KernelADataType = int8_t;
+using KernelBDataType = int8_t;
+using KernelEDataType = int8_t;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
+    // clang-format off
+        < ALayout,              //ALayout
+          BLayout,              //BLayout
+          DsLayout,             //DsLayout
+          ELayout,              //ELayout
+          KernelADataType,      //ADataType    
+          KernelBDataType,      //BDataType   
+          AccDataType,          //AccDataType
+          CShuffleDataType,     //CShuffleDataType
+          DsDataType,           //DsDataType
+          KernelEDataType,      //EDataType
+          AElementOp,           //AElementwiseOperation
+          BElementOp,           //BElementwiseOperation
+          CDEElementOp,         //CDEElementwiseOperation
+          GemmDefault,          //GEMMSpecialization
+          1,                    // NumGemmKPrefetchStage
+          256,                  // BlockSize
+          256,                  // MPerBlock
+          128,                  // NPerBlock
+          64,                   // KPerBlock
+          16,                   // AK1
+          16,                   // BK1
+          32,                   // MPerXdl
+          32,                   // NPerXdl
+          4,                    // MXdlPerWave
+          2,                    // NXdlPerWave
+          S<4, 64, 1>,          // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+          S<1, 0, 2>,           // ABlockTransfer ThreadCluster ArrangeOrder
+          S<1, 0, 2>,           // ABlockTransfer SrcAccessOrder
+          2,                    // ABlockTransfer SrcVectorDim
+          16,                   // ABlockTransfer SrcScalarPerVector
+          16,                   // ABlockTransfer DstScalarPerVector_K1
+          1,                    // ABlockLdsExtraM
+          S<4, 64, 1>,          // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+          S<1, 0, 2>,           // BBlockTransfer ThreadCluster ArrangeOrder
+          S<1, 0, 2>,           // BBlockTransfer SrcAccessOrder
+          2,                    // BBlockTransfer SrcVectorDim
+          16,                   // BBlockTransfer SrcScalarPerVector
+          16,                   // BBlockTransfer DstScalarPerVector_K1
+          1,                    // BBlockLdsExtraN
+          1,                    // CShuffleMXdlPerWavePerShuffle
+          1,                    // CShuffleNXdlPerWavePerShuffle
+          S<1, 64, 1, 4>,       // CBlockTransferClusterLengths_MBlock_MWaveMPerXdl_NBlock_NWaveNPerXdl
+          16>;                  // CBlockTransferScalarPerVector_NWaveNPerXdl
+// clang-format on
+
+#define BUILD_INT4_EXAMPLE
+#include "run_batched_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
diff --git a/example/24_batched_gemm/batched_gemm_xdl_int8.cpp b/example/24_batched_gemm/batched_gemm_xdl_int8.cpp
new file mode 100644
index 0000000000..ad22227af5
--- /dev/null
+++ b/example/24_batched_gemm/batched_gemm_xdl_int8.cpp
@@ -0,0 +1,56 @@
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = int8_t;
+using BDataType        = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int8_t;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = int8_t;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiD_Xdl
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16>;
+// clang-format on
+
+#include "run_batched_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_batched_gemm_example(argc, argv); }
diff --git a/example/24_batched_gemm/run_batched_gemm_example.inc b/example/24_batched_gemm/run_batched_gemm_example.inc
new file mode 100644
index 0000000000..21934add31
--- /dev/null
+++ b/example/24_batched_gemm/run_batched_gemm_example.inc
@@ -0,0 +1,240 @@
+#include <random>
+
+#pragma once
+
+struct ProblemSize final
+{
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t stride_A = K;
+    ck::index_t stride_B = K;
+    ck::index_t stride_C = N;
+
+    ck::index_t batch_stride_A = M * K;
+    ck::index_t batch_stride_B = K * N;
+    ck::index_t batch_stride_C = M * N;
+
+    ck::index_t batch_count = 16;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+bool run_batched_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    using namespace ck::literals;
+
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+    static_assert(sizeof(ADataType) == sizeof(KernelADataType));
+    static_assert(sizeof(BDataType) == sizeof(KernelBDataType));
+    static_assert(sizeof(EDataType) == sizeof(KernelEDataType));
+#endif
+
+    auto& [M,
+           N,
+           K,
+           stride_A,
+           stride_B,
+           stride_C,
+           batch_stride_A,
+           batch_stride_B,
+           batch_stride_C,
+           batch_count] = problem_size;
+
+    // GEMM shape
+    auto f_host_tensor_descriptor = [](std::size_t batch_count_,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({batch_count_, row, col}, {batch_stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count_, row, col}, {batch_stride, 1_uz, stride});
+        }
+    };
+
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, ALayout{}));
+    Tensor<BDataType> b_g_k_n(
+        f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
+#ifdef BUILD_INT4_EXAMPLE
+    Tensor<KernelEDataType> e_g_m_n_device_result(
+        f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
+#else
+    Tensor<EDataType> e_g_m_n_device_result(
+        f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
+#endif
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
+    std::cout << "e_g_m_n: " << e_g_m_n_device_result.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(EDataType) * e_g_m_n_device_result.mDesc.GetElementSpaceSize());
+
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<KernelADataType> a_g_m_k_converted(a_g_m_k);
+    const Tensor<KernelBDataType> b_g_k_n_converted(b_g_k_n);
+
+    a_device_buf.ToDevice(a_g_m_k_converted.mData.data());
+    b_device_buf.ToDevice(b_g_k_n_converted.mData.data());
+#else
+    a_device_buf.ToDevice(a_g_m_k.mData.data());
+    b_device_buf.ToDevice(b_g_k_n.mData.data());
+#endif
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                      b_device_buf.GetDeviceBuffer(),
+                                      {},
+                                      c_device_buf.GetDeviceBuffer(),
+                                      M,
+                                      N,
+                                      K,
+                                      batch_count,
+                                      stride_A,
+                                      stride_B,
+                                      {},
+                                      stride_C,
+                                      batch_stride_A,
+                                      batch_stride_B,
+                                      {},
+                                      batch_stride_C,
+                                      a_element_op,
+                                      b_element_op,
+                                      cde_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+    bool pass = true;
+
+    if(config.do_verification)
+    {
+        c_device_buf.FromDevice(e_g_m_n_device_result.mData.data());
+
+        using ReferenceBatchedGemmInstance =
+            ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                             BDataType,
+                                                             EDataType,
+                                                             AccDataType,
+                                                             AElementOp,
+                                                             BElementOp,
+                                                             CDEElementOp>;
+
+        auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
+        auto ref_invoker      = ref_batched_gemm.MakeInvoker();
+
+        Tensor<EDataType> e_g_m_n_host_result(
+            f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
+
+        auto ref_argument = ref_batched_gemm.MakeArgument(
+            a_g_m_k, b_g_k_n, e_g_m_n_host_result, a_element_op, b_element_op, cde_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+#ifdef BUILD_INT4_EXAMPLE
+        const Tensor<EDataType> e_device_result_converted(e_g_m_n_device_result);
+        pass &= ck::utils::check_err(e_device_result_converted, e_g_m_n_host_result);
+
+#else
+        pass = ck::utils::check_err(
+            e_g_m_n_device_result, e_g_m_n_host_result, "Error: Incorrect results c");
+#endif
+    }
+
+    if(config.time_kernel)
+    {
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+        std::size_t flop      = std::size_t(2) * batch_count * M * N * K;
+        std::size_t num_btype = sizeof(ADataType) * batch_count * M * K +
+                                sizeof(BDataType) * batch_count * K * N +
+                                sizeof(EDataType) * batch_count * M * N;
+
+        float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << gemm.GetTypeString() << std::endl;
+    }
+
+    return pass ? 0 : 1;
+}
+
+bool run_batched_gemm_example(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    std::mt19937 gen(11939);
+    std::uniform_int_distribution<int> dis(0, 15);
+
+    problem_size.M = 256 * (dis(gen) + 1);
+    problem_size.N = 128 * (dis(gen) + 1);
+    problem_size.K = 64 * (dis(gen) + 2);
+
+    problem_size.stride_A = problem_size.K;
+    problem_size.stride_B = problem_size.K;
+    problem_size.stride_C = problem_size.N;
+
+    problem_size.batch_stride_A = problem_size.M * problem_size.K;
+    problem_size.batch_stride_B = problem_size.K * problem_size.N;
+    problem_size.batch_stride_C = problem_size.M * problem_size.N;
+
+    problem_size.batch_count = 16;
+
+    if(argc == 4)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        exit(0);
+    }
+
+    return run_batched_gemm(problem_size, config);
+}
diff --git a/example/25_gemm_bias_e_permute/CMakeLists.txt b/example/25_gemm_bias_e_permute/CMakeLists.txt
new file mode 100644
index 0000000000..eb274b2338
--- /dev/null
+++ b/example/25_gemm_bias_e_permute/CMakeLists.txt
@@ -0,0 +1,4 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_gemm_bias_e_permute_g1m3n2k1_xdl_fp16 gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp)
+    add_example_executable(example_gemm_bias_e_permute_g1m2n3k1_xdl_fp16 gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp)
+endif()
diff --git a/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp b/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp
new file mode 100644
index 0000000000..420a7cf74f
--- /dev/null
+++ b/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp
@@ -0,0 +1,397 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimG = 1;
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 3;
+static constexpr ck::index_t NumDimK = 1;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //############################################| NumDimG| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           Gemm|              A|              B|             DE| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //############################################|        |        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Spacialization| Spacialization| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //############################################|        |        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |               |               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //############################################|        |        |        |        |      |      |        |         |           |      |             |            |             |               |               |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK,   F16,   F16,     F32,      F16, DsDataType,   F16,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1, 32, 1, 4>,               8>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimG == 1 && NumDimM == 2 && NumDimN == 3 && NumDimK == 1, bool> =
+              false>
+struct ReferenceContraction_G1_M2_N3_K1 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_gs_ms_ks,
+                 const Tensor<BDataType>& b_gs_ns_ks,
+                 Tensor<EDataType>& e_gs_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_gs_ms_ks_{a_gs_ms_ks},
+              b_gs_ns_ks_{b_gs_ns_ks},
+              e_gs_ms_ns_{e_gs_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_gs_ms_ks_;
+        const Tensor<BDataType>& b_gs_ns_ks_;
+        Tensor<EDataType>& e_gs_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_G1_M2_N3_K1::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_gs_ms_ns = [&](auto g0, auto m0, auto m1, auto n0, auto n1, auto n2) {
+                const int K0 = arg.a_gs_ms_ks_.mDesc.GetLengths()[3];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    AccDataType v_a;
+                    AccDataType v_b;
+
+                    arg.a_element_op_(
+                        v_a, ck::type_convert<const AccDataType>(arg.a_gs_ms_ks_(g0, m0, m1, k0)));
+                    arg.b_element_op_(
+                        v_b,
+                        ck::type_convert<const AccDataType>(arg.b_gs_ns_ks_(g0, n0, n1, n2, k0)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_gs_ms_ns_(g0, m0, m1, n0, n1, n2) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_gs_ms_ns,
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[3],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[4],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[5])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_gs_ms_ks,
+                             const Tensor<BDataType>& b_gs_ns_ks,
+                             Tensor<EDataType>& e_gs_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{
+            a_gs_ms_ks, b_gs_ns_ks, e_gs_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_M3_N2_K1"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::index_t G0 = 1;
+
+    ck::index_t M0 = 4;
+    ck::index_t M1 = 256;
+
+    ck::index_t N0 = 4;
+    ck::index_t N1 = 16;
+    ck::index_t N2 = 32;
+
+    ck::index_t K0 = 256;
+
+    // A[M0, M1, M2, K0]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, M0, M1, K0};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M0 * M1 * K0, M1 * K0, K0, 1};
+    // B[N0, N1, K0]
+    std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, N0, N1, N2, K0};
+    std::vector<ck::index_t> b_gs_ns_ks_strides{N0 * N1 * N2 * K0, N1 * N2 * K0, N2 * K0, K0, 1};
+
+    // D[N0, M0, N1, M1, N2]
+    std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, M0, M1, N0, N1, N2};
+    std::vector<ck::index_t> d_gs_ms_ns_strides{N0 * N1 * N2, 0, 0, N1 * N2, N2, 1};
+    // E[N0, M0, N1, M1, N2]
+    std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, M0, M1, N0, N1, N2};
+    std::vector<ck::index_t> e_gs_ms_ns_strides{
+        M0 * M1 * N0 * N1 * N2, N1 * M1 * N2, N2, M0 * N1 * M1 * N2, M1 * N2, 1};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<BDataType> b_gs_ns_ks(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
+    Tensor<DDataType> d_gs_ms_ns(d_gs_ms_ns_lengths, d_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_device_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
+    std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
+    std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) *
+                           e_gs_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_gs_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_gs_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_gs_ms_ks_lengths,
+                                    a_gs_ms_ks_strides,
+                                    b_gs_ns_ks_lengths,
+                                    b_gs_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
+                                    e_gs_ms_ns_lengths,
+                                    e_gs_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t M = ck::accumulate_n<ck::index_t>(
+        e_gs_ms_ns_lengths.begin() + NumDimG, NumDimM, 1, std::multiplies<>{});
+
+    std::size_t N = ck::accumulate_n<ck::index_t>(
+        e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    std::size_t K = ck::accumulate_n<ck::index_t>(
+        a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+    std::size_t flop      = std::size_t(2) * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(DDataType) * M * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_gs_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+        using ReferenceOpInstance = ReferenceContraction_G1_M2_N3_K1<NumDimM,
+                                                                     NumDimN,
+                                                                     NumDimK,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CShuffleDataType,
+                                                                     AccDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(a_gs_ms_ks,
+                                                  b_gs_ns_ks,
+                                                  c_gs_ms_ns_host_result,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t g0 = 0; g0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[0]; ++g0)
+        {
+            for(size_t m0 = 0; m0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[1]; ++m0)
+            {
+                for(size_t m1 = 0; m1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[2]; ++m1)
+                {
+                    for(size_t n0 = 0; n0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[3]; ++n0)
+                    {
+                        for(size_t n1 = 0; n1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[4]; ++n1)
+                        {
+                            for(size_t n2 = 0; n2 < e_gs_ms_ns_host_result.mDesc.GetLengths()[5];
+                                ++n2)
+                            {
+                                cde_element_op(e_gs_ms_ns_host_result(g0, m0, m1, n0, n1, n2),
+                                               c_gs_ms_ns_host_result(g0, m0, m1, n0, n1, n2),
+                                               d_gs_ms_ns(g0, m0, m1, n0, n1, n2));
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_gs_ms_ns_device_result, e_gs_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp b/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp
new file mode 100644
index 0000000000..9d606db205
--- /dev/null
+++ b/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp
@@ -0,0 +1,398 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimG = 1;
+static constexpr ck::index_t NumDimM = 3;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 1;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //############################################| NumDimG| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           Gemm|              A|              B|             DE| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //############################################|        |        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Spacialization| Spacialization| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //############################################|        |        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |               |               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //############################################|        |        |        |        |      |      |        |         |           |      |             |            |             |               |               |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK,   F16,   F16,     F32,      F16, DsDataType,   F16,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1, 32, 1, 4>,               8>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+template <ck::index_t NumDimG,
+          ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimG == 1 && NumDimM == 3 && NumDimN == 2 && NumDimK == 1, bool> =
+              false>
+struct ReferenceContraction_G1_M3_N2_K1 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_gs_ms_ks,
+                 const Tensor<BDataType>& b_gs_ns_ks,
+                 Tensor<EDataType>& e_gs_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_gs_ms_ks_{a_gs_ms_ks},
+              b_gs_ns_ks_{b_gs_ns_ks},
+              e_gs_ms_ns_{e_gs_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_gs_ms_ks_;
+        const Tensor<BDataType>& b_gs_ns_ks_;
+        Tensor<EDataType>& e_gs_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_G1_M3_N2_K1::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_gs_ms_ns = [&](auto g0, auto m0, auto m1, auto m2, auto n0, auto n1) {
+                const int K0 = arg.a_gs_ms_ks_.mDesc.GetLengths()[4];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    AccDataType v_a;
+                    AccDataType v_b;
+
+                    arg.a_element_op_(
+                        v_a,
+                        ck::type_convert<const AccDataType>(arg.a_gs_ms_ks_(g0, m0, m1, m2, k0)));
+                    arg.b_element_op_(
+                        v_b, ck::type_convert<const AccDataType>(arg.b_gs_ns_ks_(g0, n0, n1, k0)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_gs_ms_ns_(g0, m0, m1, m2, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_gs_ms_ns,
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[3],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[4],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[5])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_gs_ms_ks,
+                             const Tensor<BDataType>& b_gs_ns_ks,
+                             Tensor<EDataType>& e_gs_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{
+            a_gs_ms_ks, b_gs_ns_ks, e_gs_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_G1_M3_N2_K1"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::index_t G0 = 1;
+
+    ck::index_t M0 = 4;
+    ck::index_t M1 = 8;
+    ck::index_t M2 = 256;
+
+    ck::index_t N0 = 32;
+    ck::index_t N1 = 128;
+
+    ck::index_t K0 = 1024;
+
+    // A[M0, M1, M2, K0]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, M0, M1, M2, K0};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M0 * M1 * M2 * K0, M1 * M2 * K0, M2 * K0, K0, 1};
+
+    // B[N0, N1, K0]
+    std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, N0, N1, K0};
+    std::vector<ck::index_t> b_gs_ns_ks_strides{N0 * N1 * K0, N1 * K0, K0, 1};
+
+    // D[M0, N0, M1, N1, M2]
+    std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, M0, M1, M2, N0, N1};
+    std::vector<ck::index_t> d_gs_ms_ns_strides{N0 * N1, 0, 0, 0, N1, 1};
+
+    // E[M1, M0, N0, M1, N1]
+    std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, M0, M1, M2, N0, N1};
+    std::vector<ck::index_t> e_gs_ms_ns_strides{
+        M0 * M1 * M2 * N1 * N0, N0 * M1 * N1, N1, M0 * N0 * M1 * N1, M1 * N1, 1};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<BDataType> b_gs_ns_ks(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
+    Tensor<DDataType> d_gs_ms_ns(d_gs_ms_ns_lengths, d_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_device_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
+    std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
+    std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) *
+                           e_gs_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_gs_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_gs_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_gs_ms_ks_lengths,
+                                    a_gs_ms_ks_strides,
+                                    b_gs_ns_ks_lengths,
+                                    b_gs_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
+                                    e_gs_ms_ns_lengths,
+                                    e_gs_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t M =
+        ck::accumulate_n<ck::index_t>(e_gs_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+    ck::index_t N = ck::accumulate_n<ck::index_t>(
+        e_gs_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    ck::index_t K = ck::accumulate_n<ck::index_t>(
+        a_gs_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+    std::size_t flop      = std::size_t(2) * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(DDataType) * M * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_gs_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+        using ReferenceOpInstance = ReferenceContraction_G1_M3_N2_K1<NumDimG,
+                                                                     NumDimM,
+                                                                     NumDimN,
+                                                                     NumDimK,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CShuffleDataType,
+                                                                     AccDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(a_gs_ms_ks,
+                                                  b_gs_ns_ks,
+                                                  c_gs_ms_ns_host_result,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t g0 = 0; g0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[0]; ++g0)
+        {
+            for(size_t m0 = 0; m0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[1]; ++m0)
+            {
+                for(size_t m1 = 0; m1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[2]; ++m1)
+                {
+                    for(size_t m2 = 0; m2 < e_gs_ms_ns_host_result.mDesc.GetLengths()[3]; ++m2)
+                    {
+                        for(size_t n0 = 0; n0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[4]; ++n0)
+                        {
+                            for(size_t n1 = 0; n1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[5];
+                                ++n1)
+                            {
+                                cde_element_op(e_gs_ms_ns_host_result(g0, m0, m1, m2, n0, n1),
+                                               c_gs_ms_ns_host_result(g0, m0, m1, m2, n0, n1),
+                                               d_gs_ms_ns(g0, m0, m1, m2, n0, n1));
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_gs_ms_ns_device_result, e_gs_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/26_contraction/CMakeLists.txt b/example/26_contraction/CMakeLists.txt
new file mode 100644
index 0000000000..6cab88b13e
--- /dev/null
+++ b/example/26_contraction/CMakeLists.txt
@@ -0,0 +1,8 @@
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_contraction_bilinear_xdl_fp32 contraction_bilinear_xdl_fp32.cpp)
+    add_example_executable(example_contraction_scale_xdl_fp32 contraction_scale_xdl_fp32.cpp)
+endif()
+if(DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
+    add_example_executable(example_contraction_bilinear_xdl_fp64 contraction_bilinear_xdl_fp64.cpp)
+    add_example_executable(example_contraction_scale_xdl_fp64 contraction_scale_xdl_fp64.cpp)
+endif()
diff --git a/example/26_contraction/README.md b/example/26_contraction/README.md
new file mode 100644
index 0000000000..c88d93cf83
--- /dev/null
+++ b/example/26_contraction/README.md
@@ -0,0 +1,20 @@
+# Instructions for ```example_contraction_bilinear_xdl_fp32```
+
+## Run
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+./bin/example_contraction_bilinear_xdl_fp32 1 1 1
+```
+
+Result (MI100 @ dynammic freq, 46TFlops peak FP32)
+```
+a_ms_ks: dim 4, lengths {30, 128, 32, 64}, strides {524288, 4096, 128, 1}
+b_ks_ns: dim 4, lengths {32, 64, 32, 64}, strides {128, 1, 524288, 4096}
+c_ms_ns: dim 4, lengths {30, 128, 32, 64}, strides {524288, 4096, 128, 1}
+launch_and_time_kernel: grid_dim {240, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 0.843286 ms, 38.1985 TFlops, 94.5014 GB/s, DeviceContractionMultipleD_Xdl_CShuffle<256, 256, 128, 16, 4, 4>
+```
diff --git a/example/26_contraction/contraction_bilinear_xdl_fp32.cpp b/example/26_contraction/contraction_bilinear_xdl_fp32.cpp
new file mode 100644
index 0000000000..78522160c8
--- /dev/null
+++ b/example/26_contraction/contraction_bilinear_xdl_fp32.cpp
@@ -0,0 +1,293 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F32;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F32;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>;
+
+using DeviceOpInstanceKNNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>;
+
+using DeviceOpInstanceMKNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>;
+
+using DeviceOpInstanceMNNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+
+    float alpha = 1.f;
+    float beta  = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 28)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t M0 = std::stoi(argv[4]);
+        const ck::index_t M1 = std::stoi(argv[5]);
+
+        const ck::index_t N0 = std::stoi(argv[6]);
+        const ck::index_t N1 = std::stoi(argv[7]);
+
+        const ck::index_t K0 = std::stoi(argv[8]);
+        const ck::index_t K1 = std::stoi(argv[9]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[10]), std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[14]), std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17])};
+
+        d_ms_ns_lengths = {M0, M1, N0, N1};
+        d_ms_ns_strides = {
+            std::stoi(argv[18]), std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[22]), std::stoi(argv[23]), std::stoi(argv[24]), std::stoi(argv[25])};
+
+        alpha = std::stof(argv[26]);
+        beta  = std::stof(argv[27]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 7: M0, M1, N0, N1, K0, K1\n");
+        printf("arg10 to 13: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg14 to 17: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg18 to 21: Stride_D_M0, Stride_D_M1, Stride_D_N0, Stride_D_N1\n");
+        printf("arg22 to 25: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg26 to 27: alpha, beta\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_ms_ks(a_ms_ks_lengths, a_ms_ks_strides);
+    Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
+    Tensor<EDataType> d_ms_ns(d_ms_ns_lengths, d_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+    std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl;
+    std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
+    std::cout << "d_ms_ns: " << d_ms_ns.mDesc << std::endl;
+    std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{alpha, beta};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_ms_ks_lengths,
+                                    a_ms_ks_strides,
+                                    b_ns_ks_lengths,
+                                    b_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                                    e_ms_ns_lengths,
+                                    e_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t M =
+        ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+    ck::index_t N = ck::accumulate_n<ck::index_t>(
+        e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    ck::index_t K = ck::accumulate_n<ck::index_t>(
+        a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+    std::size_t flop      = std::size_t(2) * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(DDataType) * M * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+        using ReferenceOpInstance =
+            ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
+                                                                      NumDimN,
+                                                                      NumDimK,
+                                                                      ADataType,
+                                                                      BDataType,
+                                                                      CShuffleDataType,
+                                                                      AccDataType,
+                                                                      AElementOp,
+                                                                      BElementOp>;
+
+        auto ref_op      = ReferenceOpInstance{};
+        auto ref_invoker = ref_op.MakeInvoker();
+
+        auto ref_argument =
+            ref_op.MakeArgument(a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
+                {
+                    for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
+                    {
+                        cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
+                                       c_ms_ns_host_result(m0, m1, n0, n1),
+                                       d_ms_ns(m0, m1, n0, n1));
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/26_contraction/contraction_bilinear_xdl_fp64.cpp b/example/26_contraction/contraction_bilinear_xdl_fp64.cpp
new file mode 100644
index 0000000000..6cceed5bc1
--- /dev/null
+++ b/example/26_contraction/contraction_bilinear_xdl_fp64.cpp
@@ -0,0 +1,293 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F64 = double;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F64;
+using BDataType        = F64;
+using AccDataType      = F64;
+using CShuffleDataType = F64;
+using DDataType        = F64;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F64;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,         1,           1,           1,              S<1, 16, 1, 16>,               1>;
+
+using DeviceOpInstanceKNNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   2,   1,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>;
+
+using DeviceOpInstanceMKNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   1,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,         1,           1,           1,              S<1, 16, 1, 16>,               1>;
+
+using DeviceOpInstanceMNNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   1,   1,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+
+    float alpha = 1.f;
+    float beta  = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 28)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t M0 = std::stoi(argv[4]);
+        const ck::index_t M1 = std::stoi(argv[5]);
+
+        const ck::index_t N0 = std::stoi(argv[6]);
+        const ck::index_t N1 = std::stoi(argv[7]);
+
+        const ck::index_t K0 = std::stoi(argv[8]);
+        const ck::index_t K1 = std::stoi(argv[9]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[10]), std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[14]), std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17])};
+
+        d_ms_ns_lengths = {M0, M1, N0, N1};
+        d_ms_ns_strides = {
+            std::stoi(argv[18]), std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[22]), std::stoi(argv[23]), std::stoi(argv[24]), std::stoi(argv[25])};
+
+        alpha = std::stof(argv[26]);
+        beta  = std::stof(argv[27]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 7: M0, M1, N0, N1, K0, K1\n");
+        printf("arg10 to 13: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg14 to 17: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg18 to 21: Stride_D_M0, Stride_D_M1, Stride_D_N0, Stride_D_N1\n");
+        printf("arg22 to 25: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg26 to 27: alpha, beta\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_ms_ks(a_ms_ks_lengths, a_ms_ks_strides);
+    Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
+    Tensor<EDataType> d_ms_ns(d_ms_ns_lengths, d_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+    std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl;
+    std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
+    std::cout << "d_ms_ns: " << d_ms_ns.mDesc << std::endl;
+    std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{alpha, beta};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_ms_ks_lengths,
+                                    a_ms_ks_strides,
+                                    b_ns_ks_lengths,
+                                    b_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                                    e_ms_ns_lengths,
+                                    e_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t M =
+        ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+    ck::index_t N = ck::accumulate_n<ck::index_t>(
+        e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    ck::index_t K = ck::accumulate_n<ck::index_t>(
+        a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+    std::size_t flop      = std::size_t(2) * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(DDataType) * M * N + sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+        using ReferenceOpInstance =
+            ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
+                                                                      NumDimN,
+                                                                      NumDimK,
+                                                                      ADataType,
+                                                                      BDataType,
+                                                                      CShuffleDataType,
+                                                                      AccDataType,
+                                                                      AElementOp,
+                                                                      BElementOp>;
+
+        auto ref_op      = ReferenceOpInstance{};
+        auto ref_invoker = ref_op.MakeInvoker();
+
+        auto ref_argument =
+            ref_op.MakeArgument(a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
+                {
+                    for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
+                    {
+                        cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
+                                       c_ms_ns_host_result(m0, m1, n0, n1),
+                                       d_ms_ns(m0, m1, n0, n1));
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/26_contraction/contraction_scale_xdl_fp32.cpp b/example/26_contraction/contraction_scale_xdl_fp32.cpp
new file mode 100644
index 0000000000..1574f5d18f
--- /dev/null
+++ b/example/26_contraction/contraction_scale_xdl_fp32.cpp
@@ -0,0 +1,276 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F32;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// clang-format off
+using DeviceOpInstanceKKN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>;
+
+using DeviceOpInstanceKNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>;
+
+using DeviceOpInstanceMKN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>;
+
+using DeviceOpInstanceMNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKN;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+
+    float scale = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 23)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t M0 = std::stoi(argv[4]);
+        const ck::index_t M1 = std::stoi(argv[5]);
+
+        const ck::index_t N0 = std::stoi(argv[6]);
+        const ck::index_t N1 = std::stoi(argv[7]);
+
+        const ck::index_t K0 = std::stoi(argv[8]);
+        const ck::index_t K1 = std::stoi(argv[9]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[10]), std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[14]), std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[18]), std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21])};
+
+        scale = std::stof(argv[22]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M0, M1, N0, N1, K0, K1\n");
+        printf("arg10 to 13: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg14 to 17: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg18 to 21: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg22: scale\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_ms_ks(a_ms_ks_lengths, a_ms_ks_strides);
+    Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
+    Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+    std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl;
+    std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
+    std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_ns_ks.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{scale};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 0>{},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_ms_ks_lengths,
+                                    a_ms_ks_strides,
+                                    b_ns_ks_lengths,
+                                    b_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 0>{},
+                                    std::array<std::vector<ck::index_t>, 0>{},
+                                    e_ms_ns_lengths,
+                                    e_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t M =
+        ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+    ck::index_t N = ck::accumulate_n<ck::index_t>(
+        e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    ck::index_t K = ck::accumulate_n<ck::index_t>(
+        a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + +sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+        using ReferenceOpInstance =
+            ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
+                                                                      NumDimN,
+                                                                      NumDimK,
+                                                                      ADataType,
+                                                                      BDataType,
+                                                                      CShuffleDataType,
+                                                                      AccDataType,
+                                                                      AElementOp,
+                                                                      BElementOp>;
+
+        auto ref_op      = ReferenceOpInstance{};
+        auto ref_invoker = ref_op.MakeInvoker();
+
+        Tensor<float> empty_tensor(std::vector<ck::index_t>{}, std::vector<ck::index_t>{});
+        auto ref_argument =
+            ref_op.MakeArgument(a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
+                {
+                    for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
+                    {
+                        cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
+                                       c_ms_ns_host_result(m0, m1, n0, n1));
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/26_contraction/contraction_scale_xdl_fp64.cpp b/example/26_contraction/contraction_scale_xdl_fp64.cpp
new file mode 100644
index 0000000000..3dacc70887
--- /dev/null
+++ b/example/26_contraction/contraction_scale_xdl_fp64.cpp
@@ -0,0 +1,276 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F64 = double;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F64;
+using BDataType        = F64;
+using AccDataType      = F64;
+using CShuffleDataType = F64;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F64;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// clang-format off
+using DeviceOpInstanceKKN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,         1,           1,           1,              S<1, 16, 1, 16>,               1>;
+
+using DeviceOpInstanceKNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   2,   1,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>;
+
+using DeviceOpInstanceMKN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   1,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,         1,           1,           1,              S<1, 16, 1, 16>,               1>;
+
+using DeviceOpInstanceMNN = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F64,   F64,     F64,      F64, DsDataType,   F64,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   1,   1,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKN;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+
+    float scale = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 23)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t M0 = std::stoi(argv[4]);
+        const ck::index_t M1 = std::stoi(argv[5]);
+
+        const ck::index_t N0 = std::stoi(argv[6]);
+        const ck::index_t N1 = std::stoi(argv[7]);
+
+        const ck::index_t K0 = std::stoi(argv[8]);
+        const ck::index_t K1 = std::stoi(argv[9]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[10]), std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[14]), std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[18]), std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21])};
+
+        scale = std::stof(argv[22]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M0, M1, N0, N1, K0, K1\n");
+        printf("arg10 to 13: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg14 to 17: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg18 to 21: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg22: scale\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_ms_ks(a_ms_ks_lengths, a_ms_ks_strides);
+    Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
+    Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+    std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl;
+    std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
+    std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_ns_ks.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{scale};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 0>{},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_ms_ks_lengths,
+                                    a_ms_ks_strides,
+                                    b_ns_ks_lengths,
+                                    b_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 0>{},
+                                    std::array<std::vector<ck::index_t>, 0>{},
+                                    e_ms_ns_lengths,
+                                    e_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t M =
+        ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+    ck::index_t N = ck::accumulate_n<ck::index_t>(
+        e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    ck::index_t K = ck::accumulate_n<ck::index_t>(
+        a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + +sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+        using ReferenceOpInstance =
+            ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
+                                                                      NumDimN,
+                                                                      NumDimK,
+                                                                      ADataType,
+                                                                      BDataType,
+                                                                      CShuffleDataType,
+                                                                      AccDataType,
+                                                                      AElementOp,
+                                                                      BElementOp>;
+
+        auto ref_op      = ReferenceOpInstance{};
+        auto ref_invoker = ref_op.MakeInvoker();
+
+        Tensor<float> empty_tensor(std::vector<ck::index_t>{}, std::vector<ck::index_t>{});
+        auto ref_argument =
+            ref_op.MakeArgument(a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
+                {
+                    for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
+                    {
+                        cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
+                                       c_ms_ns_host_result(m0, m1, n0, n1));
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/27_layernorm/CMakeLists.txt b/example/27_layernorm/CMakeLists.txt
new file mode 100644
index 0000000000..9cb2cd0766
--- /dev/null
+++ b/example/27_layernorm/CMakeLists.txt
@@ -0,0 +1,4 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_layernorm_fp16 layernorm_fp16.cpp)
+    add_example_executable(example_layernorm_splitk_fp16 layernorm_splitk_fp16.cpp)
+endif()
diff --git a/example/27_layernorm/common.hpp b/example/27_layernorm/common.hpp
new file mode 100644
index 0000000000..62a71713df
--- /dev/null
+++ b/example/27_layernorm/common.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
diff --git a/example/27_layernorm/layernorm_fp16.cpp b/example/27_layernorm/layernorm_fp16.cpp
new file mode 100644
index 0000000000..bb8b954f0a
--- /dev/null
+++ b/example/27_layernorm/layernorm_fp16.cpp
@@ -0,0 +1,39 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using XDataType       = ck::half_t;
+using GammaDataType   = ck::half_t;
+using BetaDataType    = ck::half_t;
+using YDataType       = ck::half_t;
+using ComputeDataType = float;
+using PassThrough     = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+using DeviceInstance =
+    ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
+                                                          GammaDataType,
+                                                          BetaDataType,
+                                                          ComputeDataType,
+                                                          YDataType,
+                                                          PassThrough,
+                                                          Rank,
+                                                          NumReduceDim,
+                                                          256, // BlockSize
+                                                          8,   // ClusterM
+                                                          32,  // ClusterK
+                                                          1,   // SliceM
+                                                          8,   // SliceK
+                                                          1,   // XYVectorDim (0=M, 1=K)
+                                                          8,   // SrcScalarPerVector
+                                                          1,   // GammaVecDim (0=M, 1=K)
+                                                          8,   // GammaScalarPerVector
+                                                          1,   // BetaVecDim (0=M, 1=K)
+                                                          8,   // BetaScalarPerVector
+                                                          8>;  // OutScalarPerVector
+#include "run_layernorm_example.inc"
+
+int main() { return run_groupnorm_example<DeviceInstance>(); }
diff --git a/example/27_layernorm/layernorm_splitk_fp16.cpp b/example/27_layernorm/layernorm_splitk_fp16.cpp
new file mode 100644
index 0000000000..e0378d028b
--- /dev/null
+++ b/example/27_layernorm/layernorm_splitk_fp16.cpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using XDataType       = ck::half_t;
+using GammaDataType   = ck::half_t;
+using BetaDataType    = ck::half_t;
+using YDataType       = ck::half_t;
+using ComputeDataType = float;
+using PassThrough     = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+using DeviceInstance =
+    ck::tensor_operation::device::DeviceNormalizationSplitKImpl<XDataType,
+                                                                GammaDataType,
+                                                                BetaDataType,
+                                                                ComputeDataType,
+                                                                YDataType,
+                                                                PassThrough,
+                                                                Rank,
+                                                                NumReduceDim,
+                                                                256, // BlockSize
+                                                                8,   // ClusterM
+                                                                32,  // ClusterK
+                                                                1,   // SliceM
+                                                                8,   // SliceK
+                                                                1,   // XYVectorDim (0=M, 1=K)
+                                                                8,   // XScalarPerVector
+                                                                1,   // GammaVecDim (0=M, 1=K)
+                                                                8,   // GammaScalarPerVector
+                                                                1,   // BetaVecDim (0=M, 1=K)
+                                                                8,   // BetaScalarPerVector
+                                                                8>;  // YScalarPerVector
+
+#include "run_layernorm_example.inc"
+
+int main() { return run_groupnorm_example<DeviceInstance>(); }
diff --git a/example/27_layernorm/run_layernorm_example.inc b/example/27_layernorm/run_layernorm_example.inc
new file mode 100644
index 0000000000..95200b540a
--- /dev/null
+++ b/example/27_layernorm/run_layernorm_example.inc
@@ -0,0 +1,96 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+template <typename DeviceInstance>
+int run_groupnorm_example()
+{
+    bool time_kernel = false;
+
+    ck::index_t M      = 1024;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = N;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
+        using namespace ck::literals;
+
+        return HostTensorDescriptor({row, col}, {stride, 1_uz});
+    };
+
+    Tensor<XDataType> x(f_host_tensor_descriptor2d(M, N, Stride));
+    Tensor<GammaDataType> gamma(f_host_tensor_descriptor1d(N, 1));
+    Tensor<BetaDataType> beta(f_host_tensor_descriptor1d(N, 1));
+    Tensor<YDataType> y(f_host_tensor_descriptor2d(M, N, Stride));
+
+    x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
+    gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{0.0, 1.0});
+    beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{0.0, 1.0});
+
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    beta_dev.ToDevice(beta.mData.data());
+
+    auto device_instance = DeviceInstance{};
+    auto argument_ptr    = device_instance.MakeArgumentPointer(
+        {M, N},
+        std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
+        {0, 1},
+        {0, 1},
+        std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
+        {1},
+        1e-4,
+        x_dev.GetDeviceBuffer(),
+        gamma_dev.GetDeviceBuffer(),
+        beta_dev.GetDeviceBuffer(),
+        y_dev.GetDeviceBuffer(),
+        nullptr,
+        nullptr,
+        PassThrough{});
+
+    if(!device_instance.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters are not supported" << std::endl;
+        return 1;
+    };
+
+    size_t workspace_sz = device_instance.GetWorkSpaceSize(argument_ptr.get());
+    DeviceMem workspace_dev(workspace_sz);
+    device_instance.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+    auto invoker_ptr = device_instance.MakeInvokerPointer();
+    invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    bool pass = true;
+    {
+        Tensor<YDataType> host_y(f_host_tensor_descriptor2d(M, N, Stride));
+        using ReferenceInstance = ck::tensor_operation::host::ReferenceLayernorm<XDataType,
+                                                                                 GammaDataType,
+                                                                                 BetaDataType,
+                                                                                 YDataType,
+                                                                                 ComputeDataType,
+                                                                                 PassThrough,
+                                                                                 Rank,
+                                                                                 NumReduceDim>;
+
+        ReferenceInstance ref;
+        auto ref_argument =
+            ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, {M, N}, {1}, 1e-4);
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+
+        y_dev.FromDevice(y.mData.data());
+        pass &= ck::utils::check_err(y, host_y, "Error: Incorrect results", 1e-3, 1e-3);
+    }
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/28_grouped_gemm_bias_e_permute/CMakeLists.txt b/example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
new file mode 100644
index 0000000000..2fda1f62a9
--- /dev/null
+++ b/example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
@@ -0,0 +1,3 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_grouped_gemm_bias_e_permute_xdl_fp16 grouped_gemm_bias_e_permute_xdl_fp16.cpp)
+endif()
diff --git a/example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp b/example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp
new file mode 100644
index 0000000000..24e9b1d9b7
--- /dev/null
+++ b/example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp
@@ -0,0 +1,466 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimM = 3;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 1;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //############################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           Gemm|              A|              B|             DE| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //############################################|        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Spacialization| Spacialization| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //############################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |               |               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //############################################|        |        |        |      |      |        |         |           |      |             |            |             |               |               |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F16,   F16,     F32,      F16, DsDataType,   F16,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1, 32, 1, 4>,               8>;
+// clang-format on
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimM == 3 && NumDimN == 2 && NumDimK == 1, bool> = false>
+struct ReferenceContraction_M3_N2_K1 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_ms_ks,
+                 const Tensor<BDataType>& b_ns_ks,
+                 Tensor<EDataType>& e_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_ms_ks_{a_ms_ks},
+              b_ns_ks_{b_ns_ks},
+              e_ms_ns_{e_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_ms_ks_;
+        const Tensor<BDataType>& b_ns_ks_;
+        Tensor<EDataType>& e_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_M3_N2_K1::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto m0, auto m1, auto m2, auto n0, auto n1) {
+                const int K0 = arg.a_ms_ks_.mDesc.GetLengths()[3];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    AccDataType v_a;
+                    AccDataType v_b;
+
+                    arg.a_element_op_(
+                        v_a, ck::type_convert<const AccDataType>(arg.a_ms_ks_(m0, m1, m2, k0)));
+                    arg.b_element_op_(
+                        v_b, ck::type_convert<const AccDataType>(arg.b_ns_ks_(n0, n1, k0)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_ms_ns_(m0, m1, m2, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.e_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[3],
+                                       arg.e_ms_ns_.mDesc.GetLengths()[4])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_ms_ks,
+                             const Tensor<BDataType>& b_ns_ks,
+                             Tensor<EDataType>& e_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{a_ms_ks, b_ns_ks, e_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_M3_N2_K1"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        exit(0);
+    }
+
+    std::size_t group_count = rand() % 16 + 1;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::ContractionDesc<1>> contraction_descs;
+    std::vector<const void*> p_a, p_b;
+    std::vector<std::array<const void*, 1>> p_ds;
+    std::vector<void*> p_c;
+
+    contraction_descs.reserve(group_count);
+
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        int M0 = 4 * (rand() % 4 + 1);
+        int M1 = 4 * (rand() % 4 + 1);
+        int M2 = 256;
+
+        int N0 = 4 * (rand() % 4 + 1);
+        int N1 = 128;
+
+        int K0 = 64 * (rand() % 4 + 1);
+
+        // A[M0, M1, M2, K0]
+        std::vector<ck::index_t> a_ms_ks_lengths{M0, M1, M2, K0};
+        std::vector<ck::index_t> a_ms_ks_strides{M1 * M2 * K0, M2 * K0, K0, 1};
+        // B[N0, N1, K0]
+        std::vector<ck::index_t> b_ns_ks_lengths{N0, N1, K0};
+        std::vector<ck::index_t> b_ns_ks_strides{N1 * K0, K0, 1};
+#if 0
+        // D[M0, N0, M1, N1, M2]
+        std::vector<ck::index_t> d_ms_ns_lengths{M0, M1, M2, N0, N1};
+        std::vector<ck::index_t> d_ms_ns_strides{0, 0, 0, N1, 1};
+        // E[M0, N0, M1, N1, M2]
+        std::vector<ck::index_t> e_ms_ns_lengths{M0, M1, M2, N0, N1};
+        std::vector<ck::index_t> e_ms_ns_strides{N0 * M1 * N1 * M2, N1 * M2, 1, M1 * N1 * M2, M2};
+#else
+        // D[M0, N0, M1, N1, M2]
+        std::vector<ck::index_t> d_ms_ns_lengths{M0, M1, M2, N0, N1};
+        std::vector<ck::index_t> d_ms_ns_strides{0, 0, 0, N1, 1};
+        // E[M0, N0, M1, N1, M2]
+        std::vector<ck::index_t> e_ms_ns_lengths{M0, M1, M2, N0, N1};
+        std::vector<ck::index_t> e_ms_ns_strides{M1 * M2 * N0 * N1, M2 * N0 * N1, N0 * N1, N1, 1};
+#endif
+
+        contraction_descs.push_back(
+            ck::tensor_operation::device::ContractionDesc<1>{a_ms_ks_lengths,
+                                                             a_ms_ks_strides,
+                                                             b_ns_ks_lengths,
+                                                             b_ns_ks_strides,
+                                                             {d_ms_ns_lengths},
+                                                             {d_ms_ns_strides},
+                                                             e_ms_ns_lengths,
+                                                             e_ms_ns_strides});
+    }
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    std::vector<Tensor<BDataType>> b_tensors;
+    std::vector<Tensor<DDataType>> d_tensors;
+    std::vector<Tensor<EDataType>> e_device_tensors;
+
+    a_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    d_tensors.reserve(group_count);
+    e_device_tensors.reserve(group_count);
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+
+    std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, d_tensors_device,
+        e_tensors_device;
+
+    a_tensors_device.reserve(group_count);
+    b_tensors_device.reserve(group_count);
+    d_tensors_device.reserve(group_count);
+    e_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(std::size_t i = 0; i < contraction_descs.size(); i++)
+    {
+        const auto a_ms_ks_lengths = contraction_descs[i].a_ms_ks_lengths;
+        const auto a_ms_ks_strides = contraction_descs[i].a_ms_ks_strides;
+
+        const auto b_ns_ks_lengths = contraction_descs[i].b_ns_ks_lengths;
+        const auto b_ns_ks_strides = contraction_descs[i].b_ns_ks_strides;
+
+        const auto d_ms_ns_lengths = contraction_descs[i].ds_ms_ns_lengths[0];
+        const auto d_ms_ns_strides = contraction_descs[i].ds_ms_ns_strides[0];
+
+        const auto e_ms_ns_lengths = contraction_descs[i].e_ms_ns_lengths;
+        const auto e_ms_ns_strides = contraction_descs[i].e_ms_ns_strides;
+
+        Tensor<ADataType> a_ms_ks(a_ms_ks_lengths, a_ms_ks_strides);
+        Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
+        Tensor<DDataType> d_ms_ns(d_ms_ns_lengths, d_ms_ns_strides);
+        Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+        ck::index_t M_ =
+            ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+        ck::index_t N_ = ck::accumulate_n<ck::index_t>(
+            e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+        ck::index_t K_ = ck::accumulate_n<ck::index_t>(
+            a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+        a_tensors.push_back(a_ms_ks);
+        b_tensors.push_back(b_ns_ks);
+        d_tensors.push_back(d_ms_ns);
+
+        // e_host_tensors.push_back(e_ms_ns_host_result);
+        e_device_tensors.push_back(e_ms_ns_device_result);
+
+        flop += std::size_t(2) * M_ * K_ * N_;
+
+        num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
+                     sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * e_device_tensors[i].mDesc.GetElementSize();
+
+        std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
+                  << " b_n_k: " << b_tensors[i].mDesc << " c_m_n: " << e_device_tensors[i].mDesc
+                  << std::endl;
+
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            d_tensors[i].GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
+            break;
+        case 2:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+            d_tensors[i].GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
+            break;
+        default:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_1<ADataType>{});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_1<BDataType>{});
+            d_tensors[i].GenerateTensorValue(GeneratorTensor_1<DDataType>{});
+        }
+    }
+
+    for(std::size_t i = 0; i < contraction_descs.size(); i++)
+    {
+        a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpaceSize()));
+        b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(BDataType) * b_tensors[i].mDesc.GetElementSpaceSize()));
+        d_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(DDataType) * d_tensors[i].mDesc.GetElementSpaceSize()));
+        e_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(EDataType) * e_device_tensors[i].mDesc.GetElementSpaceSize()));
+
+        a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
+        b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
+        d_tensors_device[i]->ToDevice(d_tensors[i].mData.data());
+
+        p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
+        p_b.push_back(b_tensors_device[i]->GetDeviceBuffer());
+        p_ds.push_back({d_tensors_device[i]->GetDeviceBuffer()});
+        p_c.push_back(e_tensors_device[i]->GetDeviceBuffer());
+    }
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceOpInstanceKKNN{};
+    auto invoker = gemm.MakeInvoker();
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(
+        p_a, p_b, p_ds, p_c, contraction_descs, a_element_op, b_element_op, cde_element_op);
+
+    DeviceMem contraction_desc_workspace(gemm.GetWorkSpaceSize(&argument));
+
+    gemm.SetWorkSpacePointer(&argument, contraction_desc_workspace.GetDeviceBuffer());
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        for(std::size_t i = 0; i < group_count; i++)
+        {
+            const auto e_ms_ns_lengths = contraction_descs[i].e_ms_ns_lengths;
+            const auto e_ms_ns_strides = contraction_descs[i].e_ms_ns_strides;
+
+            Tensor<EDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+            Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+            e_tensors_device[i]->FromDevice(e_device_tensors[i].mData.data());
+
+            using ReferenceOpInstance = ReferenceContraction_M3_N2_K1<NumDimM,
+                                                                      NumDimN,
+                                                                      NumDimK,
+                                                                      ADataType,
+                                                                      BDataType,
+                                                                      CShuffleDataType,
+                                                                      AccDataType,
+                                                                      AElementOp,
+                                                                      BElementOp,
+                                                                      PassThrough>;
+
+            auto ref_gemm    = ReferenceOpInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
+                                                      b_tensors[i],
+                                                      c_ms_ns_host_result,
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      PassThrough{});
+
+            ref_invoker.Run(ref_argument);
+
+            for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
+            {
+                for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
+                {
+                    for(size_t m2 = 0; m2 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++m2)
+                    {
+                        for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n0)
+                        {
+                            for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[4]; ++n1)
+                            {
+                                cde_element_op(e_ms_ns_host_result(m0, m1, m2, n0, n1),
+                                               c_ms_ns_host_result(m0, m1, m2, n0, n1),
+                                               d_tensors[i](m0, m1, m2, n0, n1));
+                            }
+                        }
+                    }
+                }
+            }
+
+            pass &= ck::utils::check_err(e_device_tensors[i], e_ms_ns_host_result);
+        }
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/example/29_batched_gemm_bias_e_permute/CMakeLists.txt b/example/29_batched_gemm_bias_e_permute/CMakeLists.txt
new file mode 100644
index 0000000000..09c3e6c608
--- /dev/null
+++ b/example/29_batched_gemm_bias_e_permute/CMakeLists.txt
@@ -0,0 +1,7 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp)
+
+    if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
+        add_example_executable(example_batched_gemm_bias_e_permute_wmma_fp16 batched_gemm_bias_e_permute_wmma_fp16.cpp)
+    endif()
+endif()
diff --git a/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp b/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp
new file mode 100644
index 0000000000..62233e5351
--- /dev/null
+++ b/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp
@@ -0,0 +1,431 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 1;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceOpInstanceKKNN =
+    ck::tensor_operation::device::DeviceBatchedContractionMultipleD_Wmma_CShuffle<NumDimG,
+                                                                                  NumDimM,
+                                                                                  NumDimN,
+                                                                                  NumDimK,
+                                                                                  ADataType,
+                                                                                  BDataType,
+                                                                                  DsDataType,
+                                                                                  EDataType,
+                                                                                  AccDataType,
+                                                                                  CShuffleDataType,
+                                                                                  AElementOp,
+                                                                                  BElementOp,
+                                                                                  CDEElementOp,
+                                                                                  GemmSpec,
+                                                                                  ABSpec,
+                                                                                  ABSpec,
+                                                                                  DESpec,
+                                                                                  256,
+                                                                                  128,
+                                                                                  256,
+                                                                                  8,
+                                                                                  8,
+                                                                                  16,
+                                                                                  16,
+                                                                                  4,
+                                                                                  4,
+                                                                                  S<4, 64, 1>,
+                                                                                  S<1, 0, 2>,
+                                                                                  S<1, 0, 2>,
+                                                                                  2,
+                                                                                  8,
+                                                                                  8,
+                                                                                  true,
+                                                                                  S<4, 64, 1>,
+                                                                                  S<1, 0, 2>,
+                                                                                  S<1, 0, 2>,
+                                                                                  2,
+                                                                                  8,
+                                                                                  8,
+                                                                                  true,
+                                                                                  1,
+                                                                                  1,
+                                                                                  S<1, 32, 1, 8>,
+                                                                                  8>;
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimG,
+          ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimG == 2 && NumDimM == 2 && NumDimN == 2 && NumDimK == 1, bool> =
+              false>
+struct ReferenceContraction_G2_M2_N2_K1 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_gs_ms_ks,
+                 const Tensor<BDataType>& b_gs_ns_ks,
+                 Tensor<EDataType>& e_gs_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_gs_ms_ks_{a_gs_ms_ks},
+              b_gs_ns_ks_{b_gs_ns_ks},
+              e_gs_ms_ns_{e_gs_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_gs_ms_ks_;
+        const Tensor<BDataType>& b_gs_ns_ks_;
+        Tensor<EDataType>& e_gs_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_G2_M2_N2_K1::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto g0, auto g1, auto m0, auto m1, auto n0, auto n1) {
+                const int K0 = arg.a_gs_ms_ks_.mDesc.GetLengths()[4];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    AccDataType v_a;
+                    AccDataType v_b;
+
+                    arg.a_element_op_(
+                        v_a,
+                        ck::type_convert<const AccDataType>(arg.a_gs_ms_ks_(g0, g1, m0, m1, k0)));
+                    arg.b_element_op_(
+                        v_b,
+                        ck::type_convert<const AccDataType>(arg.b_gs_ns_ks_(g0, g1, n0, n1, k0)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_gs_ms_ns_(g0, g1, m0, m1, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[3],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[4],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[5])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_gs_ms_ks,
+                             const Tensor<BDataType>& b_gs_ns_ks,
+                             Tensor<EDataType>& e_gs_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{
+            a_gs_ms_ks, b_gs_ns_ks, e_gs_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_G2_M2_N2_K1"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    ck::index_t G0 = 1;
+    ck::index_t G1 = 2;
+
+    ck::index_t M0 = 4;
+    ck::index_t M1 = 128;
+
+    ck::index_t N0 = 16;
+    ck::index_t N1 = 256;
+
+    ck::index_t K0 = 2048;
+
+    // A[G0, G1, M0, M1, K0]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M0, M1, K0};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{G1 * M0 * M1 * K0, M0 * M1 * K0, M1 * K0, K0, 1};
+    // B[G0, G1, N0, N1, K0]
+    std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, G1, N0, N1, K0};
+    std::vector<ck::index_t> b_gs_ns_ks_strides{G1 * N0 * N1 * K0, N0 * N1 * K0, N1 * K0, K0, 1};
+
+    // D[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> d_gs_ms_ns_strides{G1 * N0 * N1, N0 * N1, 0, 0, N1, 1};
+    // E[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> e_gs_ms_ns_strides{
+        G1 * M0 * N0 * M1 * N1, M0 * N0 * M1 * N1, N0 * M1 * N1, N1, M1 * N1, 1};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<BDataType> b_gs_ns_ks(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
+    Tensor<DDataType> d_gs_ms_ns(d_gs_ms_ns_lengths, d_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_device_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
+    std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
+    std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
+        break;
+    }
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) *
+                           e_gs_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_gs_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_gs_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_gs_ms_ks_lengths,
+                                    a_gs_ms_ks_strides,
+                                    b_gs_ns_ks_lengths,
+                                    b_gs_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
+                                    e_gs_ms_ns_lengths,
+                                    e_gs_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t G =
+        ck::accumulate_n<ck::index_t>(e_gs_ms_ns_lengths.begin(), NumDimG, 1, std::multiplies<>{});
+
+    ck::index_t M = ck::accumulate_n<ck::index_t>(
+        e_gs_ms_ns_lengths.begin() + NumDimG, NumDimM, 1, std::multiplies<>{});
+
+    ck::index_t N = ck::accumulate_n<ck::index_t>(
+        e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    ck::index_t K = ck::accumulate_n<ck::index_t>(
+        a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM, NumDimK, 1, std::multiplies<>{});
+    std::cout << "GMNK=" << G << ", " << M << ", " << N << ", " << K << std::endl;
+    std::size_t flop      = std::size_t(2) * G * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * G * M * K + sizeof(BDataType) * G * K * N +
+                            sizeof(DDataType) * G * M * N + sizeof(EDataType) * G * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_gs_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+        using ReferenceOpInstance = ReferenceContraction_G2_M2_N2_K1<NumDimG,
+                                                                     NumDimM,
+                                                                     NumDimN,
+                                                                     NumDimK,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CShuffleDataType,
+                                                                     AccDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_gs_ms_ks, b_gs_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t g0 = 0; g0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[0]; ++g0)
+        {
+            for(size_t g1 = 0; g1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[1]; ++g1)
+            {
+                for(size_t m0 = 0; m0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[2]; ++m0)
+                {
+                    for(size_t m1 = 0; m1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[3]; ++m1)
+                    {
+                        for(size_t n0 = 0; n0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[4]; ++n0)
+                        {
+                            for(size_t n1 = 0; n1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[5];
+                                ++n1)
+                            {
+                                cde_element_op(e_gs_ms_ns_host_result(g0, g1, m0, m1, n0, n1),
+                                               c_ms_ns_host_result(g0, g1, m0, m1, n0, n1),
+                                               d_gs_ms_ns(g0, g1, m0, m1, n0, n1));
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_gs_ms_ns_device_result, e_gs_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp b/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp
new file mode 100644
index 0000000000..08158bfc25
--- /dev/null
+++ b/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp
@@ -0,0 +1,397 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 1;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //############################################| NumDimG| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           Gemm|              A|              B|             DE| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //############################################|        |        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Spacialization| Spacialization| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //############################################|        |        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |               |               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //############################################|        |        |        |        |      |      |        |         |           |      |             |            |             |               |               |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle< NumDimG, NumDimM, NumDimN, NumDimK,   F16,   F16,     F32,      F16, DsDataType,   F16,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1, 32, 1, 4>,               8>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimG,
+          ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimG == 2 && NumDimM == 2 && NumDimN == 2 && NumDimK == 1, bool> =
+              false>
+struct ReferenceContraction_G2_M2_N2_K1 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_gs_ms_ks,
+                 const Tensor<BDataType>& b_gs_ns_ks,
+                 Tensor<EDataType>& e_gs_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_gs_ms_ks_{a_gs_ms_ks},
+              b_gs_ns_ks_{b_gs_ns_ks},
+              e_gs_ms_ns_{e_gs_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_gs_ms_ks_;
+        const Tensor<BDataType>& b_gs_ns_ks_;
+        Tensor<EDataType>& e_gs_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_G2_M2_N2_K1::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto g0, auto g1, auto m0, auto m1, auto n0, auto n1) {
+                const int K0 = arg.a_gs_ms_ks_.mDesc.GetLengths()[4];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    AccDataType v_a;
+                    AccDataType v_b;
+
+                    arg.a_element_op_(
+                        v_a,
+                        ck::type_convert<const AccDataType>(arg.a_gs_ms_ks_(g0, g1, m0, m1, k0)));
+                    arg.b_element_op_(
+                        v_b,
+                        ck::type_convert<const AccDataType>(arg.b_gs_ns_ks_(g0, g1, n0, n1, k0)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_gs_ms_ns_(g0, g1, m0, m1, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[3],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[4],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[5])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_gs_ms_ks,
+                             const Tensor<BDataType>& b_gs_ns_ks,
+                             Tensor<EDataType>& e_gs_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{
+            a_gs_ms_ks, b_gs_ns_ks, e_gs_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_G2_M2_N2_K1"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::index_t G0 = 1;
+    ck::index_t G1 = 2;
+
+    ck::index_t M0 = 4;
+    ck::index_t M1 = 256;
+
+    ck::index_t N0 = 16;
+    ck::index_t N1 = 128;
+
+    ck::index_t K0 = 64;
+
+    // A[G0, G1, M0, M1, K0]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M0, M1, K0};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{G1 * M0 * M1 * K0, M0 * M1 * K0, M1 * K0, K0, 1};
+    // B[G0, G1, N0, N1, K0]
+    std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, G1, N0, N1, K0};
+    std::vector<ck::index_t> b_gs_ns_ks_strides{G1 * N0 * N1 * K0, N0 * N1 * K0, N1 * K0, K0, 1};
+
+    // D[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> d_gs_ms_ns_strides{G1 * N0 * N1, N0 * N1, 0, 0, N1, 1};
+    // E[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> e_gs_ms_ns_strides{
+        G1 * M0 * N0 * M1 * N1, M0 * N0 * M1 * N1, N0 * M1 * N1, N1, M1 * N1, 1};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<BDataType> b_gs_ns_ks(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
+    Tensor<DDataType> d_gs_ms_ns(d_gs_ms_ns_lengths, d_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+    Tensor<EDataType> e_gs_ms_ns_device_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
+    std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
+    std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) *
+                           e_gs_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_gs_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_gs_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_gs_ms_ks_lengths,
+                                    a_gs_ms_ks_strides,
+                                    b_gs_ns_ks_lengths,
+                                    b_gs_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
+                                    e_gs_ms_ns_lengths,
+                                    e_gs_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t G =
+        ck::accumulate_n<ck::index_t>(e_gs_ms_ns_lengths.begin(), NumDimG, 1, std::multiplies<>{});
+
+    ck::index_t M = ck::accumulate_n<ck::index_t>(
+        e_gs_ms_ns_lengths.begin() + NumDimG, NumDimM, 1, std::multiplies<>{});
+
+    ck::index_t N = ck::accumulate_n<ck::index_t>(
+        e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    ck::index_t K = ck::accumulate_n<ck::index_t>(
+        a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+    std::size_t flop      = std::size_t(2) * G * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * G * M * K + sizeof(BDataType) * G * K * N +
+                            sizeof(DDataType) * G * M * N + sizeof(EDataType) * G * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_gs_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+        using ReferenceOpInstance = ReferenceContraction_G2_M2_N2_K1<NumDimG,
+                                                                     NumDimM,
+                                                                     NumDimN,
+                                                                     NumDimK,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CShuffleDataType,
+                                                                     AccDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_gs_ms_ks, b_gs_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t g0 = 0; g0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[0]; ++g0)
+        {
+            for(size_t g1 = 0; g1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[1]; ++g1)
+            {
+                for(size_t m0 = 0; m0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[2]; ++m0)
+                {
+                    for(size_t m1 = 0; m1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[3]; ++m1)
+                    {
+                        for(size_t n0 = 0; n0 < e_gs_ms_ns_host_result.mDesc.GetLengths()[4]; ++n0)
+                        {
+                            for(size_t n1 = 0; n1 < e_gs_ms_ns_host_result.mDesc.GetLengths()[5];
+                                ++n1)
+                            {
+                                cde_element_op(e_gs_ms_ns_host_result(g0, g1, m0, m1, n0, n1),
+                                               c_ms_ns_host_result(g0, g1, m0, m1, n0, n1),
+                                               d_gs_ms_ns(g0, g1, m0, m1, n0, n1));
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        return ck::utils::check_err(e_gs_ms_ns_device_result, e_gs_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt b/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
new file mode 100644
index 0000000000..e37413c098
--- /dev/null
+++ b/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
@@ -0,0 +1,46 @@
+list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list2 gfx1100 gfx1101 gfx1102)
+
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
+   add_custom_target(example_grouped_conv_fwd_multiple_d)
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp16)
+      add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16)
+   endif()
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp32 grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp32)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_bf16 grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_bf16)
+   endif()
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int8 grouped_conv_fwd_bias_relu_add_xdl_int8.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int8)
+   endif()
+   if(USE_BITINT_EXTENSION_INT4)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int4 grouped_conv_fwd_bias_relu_add_xdl_int4.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int4)
+   endif() # USE_BITINT_EXTENSION_INT4
+
+   set(target 1)
+ endif()
+endforeach()
+
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
+   endif()
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp)
+   endif()
+   set(target 1)
+ endif()
+endforeach()
diff --git a/example/30_grouped_conv_fwd_multiple_d/README.md b/example/30_grouped_conv_fwd_multiple_d/README.md
new file mode 100644
index 0000000000..739a0425a8
--- /dev/null
+++ b/example/30_grouped_conv_fwd_multiple_d/README.md
@@ -0,0 +1,30 @@
+Command
+```bash
+arg1: verification (0=no, 1=yes)
+arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+arg3: time kernel (0=no, 1=yes)
+Following arguments (depending on number of spatial dims):
+ Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)
+ G, N, K, C,
+ <filter spatial dimensions>, (ie Y, X for 2D)
+ <input image spatial dimensions>, (ie Hi, Wi for 2D)
+ <strides>, (ie Sy, Sx for 2D)
+ <dilations>, (ie Dy, Dx for 2D)
+ <left padding>, (ie LeftPy, LeftPx for 2D)
+ <right padding>, (ie RightPy, RightPx for 2D)
+
+./bin/example_grouped_conv_fwd_bias_relu_add_xdl_fp16 1 1 1
+```
+
+Result (MI100)
+```
+in: dim 5, lengths {1, 128, 192, 71, 71}, strides {192, 967872, 1, 13632, 192}
+wei: dim 5, lengths {1, 256, 192, 3, 3}, strides {442368, 1728, 1, 576, 192}
+bias: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 0, 1, 0, 0}
+residual: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 0, 1, 0, 0}
+out: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 331776, 1, 9216, 256}
+launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 1.55981 ms, 94.0927 TFlops, 213.868 GB/s, DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<256, 128, 256, 16, Default>
+```
diff --git a/example/30_grouped_conv_fwd_multiple_d/common.hpp b/example/30_grouped_conv_fwd_multiple_d/common.hpp
new file mode 100644
index 0000000000..e60ebee6e4
--- /dev/null
+++ b/example/30_grouped_conv_fwd_multiple_d/common.hpp
@@ -0,0 +1,355 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <array>
+#include <iostream>
+#include <string>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using BF16 = ck::bhalf_t;
+using FP16 = ck::half_t;
+using FP32 = float;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+using I4 = ck::int4_t;
+#endif
+using I8  = std::int8_t;
+using I32 = std::int32_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <typename InputLay, typename WeightLay, typename OutputLay>
+struct CommonLayoutSetting
+{
+    using InputLayout  = InputLay;
+    using WeightLayout = WeightLay;
+    using OutputLayout = OutputLay;
+};
+
+template <ck::index_t NDimSpatial>
+struct CommonLayoutSettingSelector;
+
+namespace ctl = ck::tensor_layout::convolution;
+
+template <>
+struct CommonLayoutSettingSelector<1> final
+    : CommonLayoutSetting<ctl::G_NW_C, ctl::G_K_X_C, ctl::G_NW_K>
+{
+};
+
+template <>
+struct CommonLayoutSettingSelector<2> final
+    : CommonLayoutSetting<ctl::G_NHW_C, ctl::G_K_YX_C, ctl::G_NHW_K>
+{
+};
+
+template <>
+struct CommonLayoutSettingSelector<3> final
+    : CommonLayoutSetting<ctl::G_NDHW_C, ctl::G_K_ZYX_C, ctl::G_NDHW_K>
+{
+};
+
+template <ck::index_t NDimSpatial>
+using InputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::InputLayout;
+
+template <ck::index_t NDimSpatial>
+using WeightLayout = typename CommonLayoutSettingSelector<NDimSpatial>::WeightLayout;
+
+template <ck::index_t NDimSpatial>
+using OutputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::OutputLayout;
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+};
+
+#define DefaultConvParam                                                       \
+    ck::utils::conv::ConvParam                                                 \
+    {                                                                          \
+        2, 32, 2, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, { 1, 1 } \
+    }
+
+inline void print_help_msg()
+{
+    std::cerr << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+inline bool parse_cmd_args(int argc,
+                           char* argv[],
+                           ExecutionConfig& config,
+                           ck::utils::conv::ConvParam& conv_param)
+{
+    constexpr int num_execution_config_args =
+        3; // arguments for do_verification, init_method, time_kernel
+    constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
+
+    constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
+    constexpr int threshold_to_catch_all_args =
+        threshold_to_catch_partial_args + num_conv_param_leading_args;
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    // catch only ExecutionConfig arguments
+    else if(argc == threshold_to_catch_partial_args)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    // catch both ExecutionConfig & ConvParam arguments
+    else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+        conv_param                        = ck::utils::conv::parse_conv_param(
+            num_dim_spatial, threshold_to_catch_partial_args + 1, argv);
+    }
+    else
+    {
+        print_help_msg();
+        return false;
+    }
+
+    return true;
+}
+
+inline HostTensorDescriptor make_input_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.N_, conv_param.C_, conv_param.input_spatial_lengths_[0]},
+            {
+                conv_param.C_,                                                        // g
+                conv_param.input_spatial_lengths_[0] * conv_param.G_ * conv_param.C_, // n
+                1,                                                                    // c
+                conv_param.G_ * conv_param.C_                                         // wi
+            });
+
+    case 2:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.C_,
+             conv_param.input_spatial_lengths_[0],
+             conv_param.input_spatial_lengths_[1]},
+            {
+                conv_param.C_, // g
+                conv_param.input_spatial_lengths_[0] * conv_param.input_spatial_lengths_[1] *
+                    conv_param.G_ * conv_param.C_,                                    // n
+                1,                                                                    // c
+                conv_param.input_spatial_lengths_[1] * conv_param.G_ * conv_param.C_, // hi
+                conv_param.G_ * conv_param.C_                                         // wi
+            });
+
+    case 3:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.C_,
+             conv_param.input_spatial_lengths_[0],
+             conv_param.input_spatial_lengths_[1],
+             conv_param.input_spatial_lengths_[2]},
+            {
+                conv_param.C_, // g
+                conv_param.input_spatial_lengths_[0] * conv_param.input_spatial_lengths_[1] *
+                    conv_param.input_spatial_lengths_[2] * conv_param.G_ * conv_param.C_, // n
+                1,                                                                        // c
+                conv_param.input_spatial_lengths_[1] * conv_param.input_spatial_lengths_[2] *
+                    conv_param.G_ * conv_param.C_,                                    // di
+                conv_param.input_spatial_lengths_[2] * conv_param.G_ * conv_param.C_, // hi
+                conv_param.G_ * conv_param.C_                                         // wi
+            });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}
+
+inline HostTensorDescriptor make_weight_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.K_, conv_param.C_, conv_param.filter_spatial_lengths_[0]},
+            {
+                conv_param.K_ * conv_param.filter_spatial_lengths_[0] * conv_param.C_, // g
+                conv_param.filter_spatial_lengths_[0] * conv_param.C_,                 // k
+                1,                                                                     // c
+                conv_param.C_                                                          // x
+            });
+    case 2:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.K_,
+             conv_param.C_,
+             conv_param.filter_spatial_lengths_[0],
+             conv_param.filter_spatial_lengths_[1]},
+            {
+                conv_param.K_ * conv_param.filter_spatial_lengths_[0] *
+                    conv_param.filter_spatial_lengths_[1] * conv_param.C_, // g
+                conv_param.filter_spatial_lengths_[0] * conv_param.filter_spatial_lengths_[1] *
+                    conv_param.C_,                                     // k
+                1,                                                     // c
+                conv_param.filter_spatial_lengths_[1] * conv_param.C_, // y
+                conv_param.C_                                          // x
+            });
+    case 3:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.K_,
+             conv_param.C_,
+             conv_param.filter_spatial_lengths_[0],
+             conv_param.filter_spatial_lengths_[1],
+             conv_param.filter_spatial_lengths_[2]},
+            {
+                conv_param.K_ * conv_param.filter_spatial_lengths_[0] *
+                    conv_param.filter_spatial_lengths_[1] * conv_param.filter_spatial_lengths_[2] *
+                    conv_param.C_, // g
+                conv_param.filter_spatial_lengths_[0] * conv_param.filter_spatial_lengths_[1] *
+                    conv_param.filter_spatial_lengths_[2] * conv_param.C_, // k
+                1,                                                         // c
+                conv_param.filter_spatial_lengths_[1] * conv_param.filter_spatial_lengths_[2] *
+                    conv_param.C_,                                     // z
+                conv_param.filter_spatial_lengths_[2] * conv_param.C_, // y
+                conv_param.C_                                          // x
+            });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}
+
+inline HostTensorDescriptor make_bias_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.N_, conv_param.K_, conv_param.output_spatial_lengths_[0]},
+            {
+                conv_param.K_, // g
+                0,             // k
+                1,             // c
+                0              // x
+            });
+    case 2:
+        return HostTensorDescriptor({conv_param.G_,
+                                     conv_param.N_,
+                                     conv_param.K_,
+                                     conv_param.output_spatial_lengths_[0],
+                                     conv_param.output_spatial_lengths_[1]},
+                                    {
+                                        conv_param.K_, // g
+                                        0,             // n
+                                        1,             // k
+                                        0,             // ho
+                                        0              // wo
+                                    });
+    case 3:
+        return HostTensorDescriptor({conv_param.G_,
+                                     conv_param.N_,
+                                     conv_param.K_,
+                                     conv_param.output_spatial_lengths_[0],
+                                     conv_param.output_spatial_lengths_[1],
+                                     conv_param.output_spatial_lengths_[2]},
+                                    {
+                                        conv_param.K_, // g
+                                        0,             // n
+                                        1,             // k
+                                        0,             // z
+                                        0,             // y
+                                        0              // x
+                                    });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}
+
+inline HostTensorDescriptor make_output_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.N_, conv_param.K_, conv_param.output_spatial_lengths_[0]},
+            {
+                conv_param.K_,                                                         // g
+                conv_param.output_spatial_lengths_[0] * conv_param.G_ * conv_param.K_, // n
+                1,                                                                     // k
+                conv_param.G_ * conv_param.K_                                          // wo
+            });
+    case 2:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.K_,
+             conv_param.output_spatial_lengths_[0],
+             conv_param.output_spatial_lengths_[1]},
+            {
+                conv_param.K_, // g
+                conv_param.output_spatial_lengths_[0] * conv_param.output_spatial_lengths_[1] *
+                    conv_param.G_ * conv_param.K_,                                     // n
+                1,                                                                     // k
+                conv_param.output_spatial_lengths_[1] * conv_param.G_ * conv_param.K_, // ho
+                conv_param.G_ * conv_param.K_                                          // wo
+            });
+
+    case 3:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.K_,
+             conv_param.output_spatial_lengths_[0],
+             conv_param.output_spatial_lengths_[1],
+             conv_param.output_spatial_lengths_[2]},
+            {
+                conv_param.K_, // g
+                conv_param.output_spatial_lengths_[0] * conv_param.output_spatial_lengths_[1] *
+                    conv_param.output_spatial_lengths_[2] * conv_param.G_ * conv_param.K_, // n
+                1,                                                                         // k
+                conv_param.output_spatial_lengths_[1] * conv_param.output_spatial_lengths_[2] *
+                    conv_param.G_ * conv_param.K_,                                     // do
+                conv_param.output_spatial_lengths_[2] * conv_param.G_ * conv_param.K_, // ho
+                conv_param.G_ * conv_param.K_                                          // wo
+            });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}
diff --git a/example/30_grouped_conv_fwd_multiple_d/common_wmma.hpp b/example/30_grouped_conv_fwd_multiple_d/common_wmma.hpp
new file mode 100644
index 0000000000..ae769ff1d3
--- /dev/null
+++ b/example/30_grouped_conv_fwd_multiple_d/common_wmma.hpp
@@ -0,0 +1,355 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <array>
+#include <iostream>
+#include <string>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using BF16 = ck::bhalf_t;
+using FP16 = ck::half_t;
+using FP32 = float;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+using I4 = ck::int4_t;
+#endif
+using I8  = std::int8_t;
+using I32 = std::int32_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <typename InputLay, typename WeightLay, typename OutputLay>
+struct CommonLayoutSetting
+{
+    using InputLayout  = InputLay;
+    using WeightLayout = WeightLay;
+    using OutputLayout = OutputLay;
+};
+
+template <ck::index_t NDimSpatial>
+struct CommonLayoutSettingSelector;
+
+namespace ctl = ck::tensor_layout::convolution;
+
+template <>
+struct CommonLayoutSettingSelector<1> final
+    : CommonLayoutSetting<ctl::G_NW_C, ctl::G_K_X_C, ctl::G_NW_K>
+{
+};
+
+template <>
+struct CommonLayoutSettingSelector<2> final
+    : CommonLayoutSetting<ctl::G_NHW_C, ctl::G_K_YX_C, ctl::G_NHW_K>
+{
+};
+
+template <>
+struct CommonLayoutSettingSelector<3> final
+    : CommonLayoutSetting<ctl::G_NDHW_C, ctl::G_K_ZYX_C, ctl::G_NDHW_K>
+{
+};
+
+template <ck::index_t NDimSpatial>
+using InputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::InputLayout;
+
+template <ck::index_t NDimSpatial>
+using WeightLayout = typename CommonLayoutSettingSelector<NDimSpatial>::WeightLayout;
+
+template <ck::index_t NDimSpatial>
+using OutputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::OutputLayout;
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+};
+
+#define DefaultConvParam                                                       \
+    ck::utils::conv::ConvParam                                                 \
+    {                                                                          \
+        2, 32, 2, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, { 1, 1 } \
+    }
+
+inline void print_help_msg()
+{
+    std::cerr << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+inline bool parse_cmd_args(int argc,
+                           char* argv[],
+                           ExecutionConfig& config,
+                           ck::utils::conv::ConvParam& conv_param)
+{
+    constexpr int num_execution_config_args =
+        3; // arguments for do_verification, init_method, time_kernel
+    constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
+
+    constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
+    constexpr int threshold_to_catch_all_args =
+        threshold_to_catch_partial_args + num_conv_param_leading_args;
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    // catch only ExecutionConfig arguments
+    else if(argc == threshold_to_catch_partial_args)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    // catch both ExecutionConfig & ConvParam arguments
+    else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+        conv_param                        = ck::utils::conv::parse_conv_param(
+            num_dim_spatial, threshold_to_catch_partial_args + 1, argv);
+    }
+    else
+    {
+        print_help_msg();
+        return false;
+    }
+
+    return true;
+}
+
+inline HostTensorDescriptor make_input_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.N_, conv_param.C_, conv_param.input_spatial_lengths_[0]},
+            {
+                conv_param.C_,                                                        // g
+                conv_param.input_spatial_lengths_[0] * conv_param.G_ * conv_param.C_, // n
+                1,                                                                    // c
+                conv_param.G_ * conv_param.C_                                         // wi
+            });
+
+    case 2:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.C_,
+             conv_param.input_spatial_lengths_[0],
+             conv_param.input_spatial_lengths_[1]},
+            {
+                conv_param.C_, // g
+                conv_param.input_spatial_lengths_[0] * conv_param.input_spatial_lengths_[1] *
+                    conv_param.G_ * conv_param.C_,                                    // n
+                1,                                                                    // c
+                conv_param.input_spatial_lengths_[1] * conv_param.G_ * conv_param.C_, // hi
+                conv_param.G_ * conv_param.C_                                         // wi
+            });
+
+    case 3:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.C_,
+             conv_param.input_spatial_lengths_[0],
+             conv_param.input_spatial_lengths_[1],
+             conv_param.input_spatial_lengths_[2]},
+            {
+                conv_param.C_, // g
+                conv_param.input_spatial_lengths_[0] * conv_param.input_spatial_lengths_[1] *
+                    conv_param.input_spatial_lengths_[2] * conv_param.G_ * conv_param.C_, // n
+                1,                                                                        // c
+                conv_param.input_spatial_lengths_[1] * conv_param.input_spatial_lengths_[2] *
+                    conv_param.G_ * conv_param.C_,                                    // di
+                conv_param.input_spatial_lengths_[2] * conv_param.G_ * conv_param.C_, // hi
+                conv_param.G_ * conv_param.C_                                         // wi
+            });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}
+
+inline HostTensorDescriptor make_weight_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.K_, conv_param.C_, conv_param.filter_spatial_lengths_[0]},
+            {
+                conv_param.K_ * conv_param.filter_spatial_lengths_[0] * conv_param.C_, // g
+                conv_param.filter_spatial_lengths_[0] * conv_param.C_,                 // k
+                1,                                                                     // c
+                conv_param.C_                                                          // x
+            });
+    case 2:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.K_,
+             conv_param.C_,
+             conv_param.filter_spatial_lengths_[0],
+             conv_param.filter_spatial_lengths_[1]},
+            {
+                conv_param.K_ * conv_param.filter_spatial_lengths_[0] *
+                    conv_param.filter_spatial_lengths_[1] * conv_param.C_, // g
+                conv_param.filter_spatial_lengths_[0] * conv_param.filter_spatial_lengths_[1] *
+                    conv_param.C_,                                     // k
+                1,                                                     // c
+                conv_param.filter_spatial_lengths_[1] * conv_param.C_, // y
+                conv_param.C_                                          // x
+            });
+    case 3:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.K_,
+             conv_param.C_,
+             conv_param.filter_spatial_lengths_[0],
+             conv_param.filter_spatial_lengths_[1],
+             conv_param.filter_spatial_lengths_[2]},
+            {
+                conv_param.K_ * conv_param.filter_spatial_lengths_[0] *
+                    conv_param.filter_spatial_lengths_[1] * conv_param.filter_spatial_lengths_[2] *
+                    conv_param.C_, // g
+                conv_param.filter_spatial_lengths_[0] * conv_param.filter_spatial_lengths_[1] *
+                    conv_param.filter_spatial_lengths_[2] * conv_param.C_, // k
+                1,                                                         // c
+                conv_param.filter_spatial_lengths_[1] * conv_param.filter_spatial_lengths_[2] *
+                    conv_param.C_,                                     // z
+                conv_param.filter_spatial_lengths_[2] * conv_param.C_, // y
+                conv_param.C_                                          // x
+            });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}
+
+inline HostTensorDescriptor make_bias_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.N_, conv_param.K_, conv_param.output_spatial_lengths_[0]},
+            {
+                conv_param.K_, // g
+                0,             // k
+                1,             // c
+                0              // x
+            });
+    case 2:
+        return HostTensorDescriptor({conv_param.G_,
+                                     conv_param.N_,
+                                     conv_param.K_,
+                                     conv_param.output_spatial_lengths_[0],
+                                     conv_param.output_spatial_lengths_[1]},
+                                    {
+                                        conv_param.K_, // g
+                                        0,             // n
+                                        1,             // k
+                                        0,             // ho
+                                        0              // wo
+                                    });
+    case 3:
+        return HostTensorDescriptor({conv_param.G_,
+                                     conv_param.N_,
+                                     conv_param.K_,
+                                     conv_param.output_spatial_lengths_[0],
+                                     conv_param.output_spatial_lengths_[1],
+                                     conv_param.output_spatial_lengths_[2]},
+                                    {
+                                        conv_param.K_, // g
+                                        0,             // n
+                                        1,             // k
+                                        0,             // z
+                                        0,             // y
+                                        0              // x
+                                    });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}
+
+inline HostTensorDescriptor make_output_descriptor(const ck::utils::conv::ConvParam& conv_param)
+{
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1:
+        return HostTensorDescriptor(
+            {conv_param.G_, conv_param.N_, conv_param.K_, conv_param.output_spatial_lengths_[0]},
+            {
+                conv_param.K_,                                                         // g
+                conv_param.output_spatial_lengths_[0] * conv_param.G_ * conv_param.K_, // n
+                1,                                                                     // k
+                conv_param.G_ * conv_param.K_                                          // wo
+            });
+    case 2:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.K_,
+             conv_param.output_spatial_lengths_[0],
+             conv_param.output_spatial_lengths_[1]},
+            {
+                conv_param.K_, // g
+                conv_param.output_spatial_lengths_[0] * conv_param.output_spatial_lengths_[1] *
+                    conv_param.G_ * conv_param.K_,                                     // n
+                1,                                                                     // k
+                conv_param.output_spatial_lengths_[1] * conv_param.G_ * conv_param.K_, // ho
+                conv_param.G_ * conv_param.K_                                          // wo
+            });
+
+    case 3:
+        return HostTensorDescriptor(
+            {conv_param.G_,
+             conv_param.N_,
+             conv_param.K_,
+             conv_param.output_spatial_lengths_[0],
+             conv_param.output_spatial_lengths_[1],
+             conv_param.output_spatial_lengths_[2]},
+            {
+                conv_param.K_, // g
+                conv_param.output_spatial_lengths_[0] * conv_param.output_spatial_lengths_[1] *
+                    conv_param.output_spatial_lengths_[2] * conv_param.G_ * conv_param.K_, // n
+                1,                                                                         // k
+                conv_param.output_spatial_lengths_[1] * conv_param.output_spatial_lengths_[2] *
+                    conv_param.G_ * conv_param.K_,                                     // do
+                conv_param.output_spatial_lengths_[2] * conv_param.G_ * conv_param.K_, // ho
+                conv_param.G_ * conv_param.K_                                          // wo
+            });
+    }
+
+    throw std::runtime_error("unsuppored # dim spatial");
+}
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp
new file mode 100644
index 0000000000..039d250299
--- /dev/null
+++ b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common_wmma.hpp"
+
+// kernel data types
+using InKernelDataType       = FP16;
+using WeiKernelDataType      = FP16;
+using AccDataType            = FP32;
+using CShuffleDataType       = FP16;
+using BiasKernelDataType     = FP16;
+using ResidualKernelDataType = FP16;
+using OutKernelDataType      = FP16;
+
+// tensor data types
+using InUserDataType  = InKernelDataType;
+using WeiUserDataType = WeiKernelDataType;
+using OutUserDataType = OutKernelDataType;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
+
+#include "run_grouped_conv_fwd_bias_relu_add_wmma_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_int8.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_int8.cpp
new file mode 100644
index 0000000000..793324970e
--- /dev/null
+++ b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_int8.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common_wmma.hpp"
+
+// kernel data types
+using InKernelDataType       = I8;
+using WeiKernelDataType      = I8;
+using AccDataType            = I32;
+using CShuffleDataType       = I8;
+using BiasKernelDataType     = I8;
+using ResidualKernelDataType = I8;
+using OutKernelDataType      = I8;
+
+// tensor data types
+using InUserDataType  = InKernelDataType;
+using WeiUserDataType = WeiKernelDataType;
+using OutUserDataType = OutKernelDataType;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
+
+#include "run_grouped_conv_fwd_bias_relu_add_wmma_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp
new file mode 100644
index 0000000000..43c0d57dc2
--- /dev/null
+++ b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+// kernel data types
+using InKernelDataType       = BF16;
+using WeiKernelDataType      = BF16;
+using AccDataType            = FP32;
+using CShuffleDataType       = FP32;
+using BiasKernelDataType     = BF16;
+using ResidualKernelDataType = BF16;
+using OutKernelDataType      = BF16;
+
+// tensor data types
+using InUserDataType  = InKernelDataType;
+using WeiUserDataType = WeiKernelDataType;
+using OutUserDataType = OutKernelDataType;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
+
+#include "run_grouped_conv_fwd_bias_relu_add_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp
new file mode 100644
index 0000000000..40b4132b35
--- /dev/null
+++ b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+// kernel data types
+using InKernelDataType       = FP16;
+using WeiKernelDataType      = FP16;
+using AccDataType            = FP32;
+using CShuffleDataType       = FP16;
+using BiasKernelDataType     = FP16;
+using ResidualKernelDataType = FP16;
+using OutKernelDataType      = FP16;
+
+// tensor data types
+using InUserDataType  = InKernelDataType;
+using WeiUserDataType = WeiKernelDataType;
+using OutUserDataType = OutKernelDataType;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
+
+#include "run_grouped_conv_fwd_bias_relu_add_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp
new file mode 100644
index 0000000000..e05d384f26
--- /dev/null
+++ b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+// kernel data types
+using InKernelDataType       = FP32;
+using WeiKernelDataType      = FP32;
+using AccDataType            = FP32;
+using CShuffleDataType       = FP32;
+using BiasKernelDataType     = FP32;
+using ResidualKernelDataType = FP32;
+using OutKernelDataType      = FP32;
+
+// tensor data types
+using InUserDataType  = InKernelDataType;
+using WeiUserDataType = WeiKernelDataType;
+using OutUserDataType = OutKernelDataType;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
+
+#include "run_grouped_conv_fwd_bias_relu_add_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp
new file mode 100644
index 0000000000..5494563fdd
--- /dev/null
+++ b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+#error Should compile this file with ck::int4_t support
+#endif
+
+#include "common.hpp"
+
+// kernel data types
+using InKernelDataType       = I8;
+using WeiKernelDataType      = I8;
+using AccDataType            = I32;
+using CShuffleDataType       = I8;
+using BiasKernelDataType     = I8;
+using ResidualKernelDataType = I8;
+using OutKernelDataType      = I8;
+
+// tensor data types
+using InUserDataType  = I4;
+using WeiUserDataType = I4;
+using OutUserDataType = I4;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
+
+#define BUILD_INT4_EXAMPLE
+#include "run_grouped_conv_fwd_bias_relu_add_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp
new file mode 100644
index 0000000000..6bf2e8d963
--- /dev/null
+++ b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+// kernel data types
+using InKernelDataType       = I8;
+using WeiKernelDataType      = I8;
+using AccDataType            = I32;
+using CShuffleDataType       = I8;
+using BiasKernelDataType     = I8;
+using ResidualKernelDataType = I8;
+using OutKernelDataType      = I8;
+
+// tensor data types
+using InUserDataType  = InKernelDataType;
+using WeiUserDataType = WeiKernelDataType;
+using OutUserDataType = OutKernelDataType;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
+
+#include "run_grouped_conv_fwd_bias_relu_add_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp
new file mode 100644
index 0000000000..498eda2442
--- /dev/null
+++ b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+// kernel data types
+using InKernelDataType  = FP16;
+using WeiKernelDataType = FP16;
+using AccDataType       = FP32;
+using CShuffleDataType  = FP16;
+using OutKernelDataType = FP16;
+
+// tensor data types
+using InUserDataType  = InKernelDataType;
+using WeiUserDataType = WeiKernelDataType;
+using OutUserDataType = OutKernelDataType;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = PassThrough;
+
+#include "run_grouped_conv_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_example(argc, argv); }
diff --git a/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc b/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc
new file mode 100644
index 0000000000..eb242203ea
--- /dev/null
+++ b/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc
@@ -0,0 +1,288 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+template <typename BiasLay, typename ResidualLay>
+struct LayoutSetting
+{
+    using BiasLayout     = BiasLay;
+    using ResidualLayout = ResidualLay;
+};
+
+template <ck::index_t NDimSpatial>
+struct LayoutSettingSelector;
+
+template <>
+struct LayoutSettingSelector<1> final : LayoutSetting<ctl::G_K, ctl::G_NW_K>
+{
+};
+
+template <>
+struct LayoutSettingSelector<2> final : LayoutSetting<ctl::G_K, ctl::G_NHW_K>
+{
+};
+
+template <>
+struct LayoutSettingSelector<3> final : LayoutSetting<ctl::G_K, ctl::G_NDHW_K>
+{
+};
+
+template <ck::index_t NDimSpatial>
+using BiasLayout = typename LayoutSettingSelector<NDimSpatial>::BiasLayout;
+
+template <ck::index_t NDimSpatial>
+using ResidualLayout = typename LayoutSettingSelector<NDimSpatial>::ResidualLayout;
+
+template <ck::index_t NDimSpatial>
+using DeviceConvFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InputLayout<NDimSpatial>,
+        WeightLayout<NDimSpatial>,
+        ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>,
+        OutputLayout<NDimSpatial>,
+        InKernelDataType,
+        WeiKernelDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<BiasKernelDataType, ResidualKernelDataType>,
+        OutKernelDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        16,          // KPerBlock
+        4,           // AK1
+        4,           // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        4,           // ABlockTransferSrcScalarPerVector
+        4,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        4,           // BBlockTransferSrcScalarPerVector
+        4,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 16, 1, 16>,
+        4>;
+
+template <ck::index_t NDimSpatial>
+using HostConvFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                         InUserDataType,
+                                                                         WeiUserDataType,
+                                                                         CShuffleDataType,
+                                                                         InElementOp,
+                                                                         WeiElementOp,
+                                                                         PassThrough>;
+
+template <ck::index_t NDimSpatial>
+bool run_grouped_conv_fwd_bias_relu_add(const ExecutionConfig& config,
+                                        const ck::utils::conv::ConvParam& conv_param)
+{
+    static_assert(1 <= NDimSpatial && NDimSpatial <= 3, "Unsupported NDimSpatial");
+
+    const auto in_g_n_c_wis_desc   = make_input_descriptor(conv_param);
+    const auto wei_g_k_c_xs_desc   = make_weight_descriptor(conv_param);
+    const auto bias_g_n_k_wos_desc = make_bias_descriptor(conv_param);
+    const auto out_g_n_k_wos_desc  = make_output_descriptor(conv_param);
+
+    Tensor<InUserDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiUserDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<OutUserDataType> bias(bias_g_n_k_wos_desc);
+    Tensor<OutUserDataType> residual(bias_g_n_k_wos_desc);
+    Tensor<OutUserDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutKernelDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "bias: " << bias.mDesc << std::endl;
+    std::cout << "residual: " << residual.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InUserDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiUserDataType>{-5, 5});
+        bias.GenerateTensorValue(GeneratorTensor_2<OutUserDataType>{-5, 5});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InUserDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiUserDataType>{-0.5, 0.5});
+        bias.GenerateTensorValue(GeneratorTensor_3<OutUserDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InKernelDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiKernelDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(OutKernelDataType) * bias.mDesc.GetElementSpaceSize());
+    DeviceMem residual_device_buf(sizeof(OutKernelDataType) * residual.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutKernelDataType) * out_device.mDesc.GetElementSpaceSize());
+
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<InKernelDataType> in_converted(in);
+    const Tensor<WeiKernelDataType> wei_converted(wei);
+    const Tensor<OutKernelDataType> bias_converted(bias);
+    const Tensor<OutKernelDataType> residual_converted(residual);
+
+    in_device_buf.ToDevice(in_converted.mData.data());
+    wei_device_buf.ToDevice(wei_converted.mData.data());
+    bias_device_buf.ToDevice(bias_converted.mData.data());
+    residual_device_buf.ToDevice(residual_converted.mData.data());
+#else
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    bias_device_buf.ToDevice(bias.mData.data());
+    residual_device_buf.ToDevice(residual.mData.data());
+#endif
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d1_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d1_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(bias_g_n_k_wos_desc.GetLengths(), d0_g_n_k_wos_lengths);
+    copy(bias_g_n_k_wos_desc.GetStrides(), d0_g_n_k_wos_strides);
+    copy(bias_g_n_k_wos_desc.GetLengths(), d1_g_n_k_wos_lengths);
+    copy(bias_g_n_k_wos_desc.GetStrides(), d1_g_n_k_wos_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv    = DeviceConvFwdInstance<NDimSpatial>{};
+    auto invoker = conv.MakeInvoker();
+    auto argument =
+        conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                          wei_device_buf.GetDeviceBuffer(),
+                          std::array<const void*, 2>{bias_device_buf.GetDeviceBuffer(),
+                                                     residual_device_buf.GetDeviceBuffer()},
+                          out_device_buf.GetDeviceBuffer(),
+                          a_g_n_c_wis_lengths,
+                          a_g_n_c_wis_strides,
+                          b_g_k_c_xs_lengths,
+                          b_g_k_c_xs_strides,
+                          std::array<std::array<ck::index_t, NDimSpatial + 3>, 2>{
+                              {d0_g_n_k_wos_lengths, d1_g_n_k_wos_lengths}},
+                          std::array<std::array<ck::index_t, NDimSpatial + 3>, 2>{
+                              {d0_g_n_k_wos_strides, d1_g_n_k_wos_strides}},
+                          e_g_n_k_wos_lengths,
+                          e_g_n_k_wos_strides,
+                          conv_filter_strides,
+                          conv_filter_dilations,
+                          input_left_pads,
+                          input_right_pads,
+                          InElementOp{},
+                          WeiElementOp{},
+                          OutElementOp{});
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InUserDataType, WeiUserDataType, OutUserDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        Tensor<CShuffleDataType> c_host(out_g_n_k_wos_desc);
+
+        auto ref_conv     = HostConvFwdInstance<NDimSpatial>{};
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  c_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  InElementOp{},
+                                                  WeiElementOp{},
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        // TODO: implement elementwise operation for host
+        out_host.ForEach([&](auto&, auto idx) {
+            OutElementOp{}(out_host(idx), c_host(idx), bias(idx), residual(idx));
+        });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+#ifdef BUILD_INT4_EXAMPLE
+        const Tensor<OutUserDataType> out_device_converted(out_device);
+
+        return ck::utils::check_err(
+            out_device_converted, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+#else
+        return ck::utils::check_err(
+            out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+#endif
+    }
+
+    return true;
+}
+
+bool run_grouped_conv_fwd_bias_relu_add_example(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+
+    if(!parse_cmd_args(argc, argv, config, conv_param))
+    {
+        return false;
+    }
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1: return run_grouped_conv_fwd_bias_relu_add<1>(config, conv_param);
+    case 2: return run_grouped_conv_fwd_bias_relu_add<2>(config, conv_param);
+    case 3: return run_grouped_conv_fwd_bias_relu_add<3>(config, conv_param);
+    }
+
+    return false;
+}
diff --git a/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_wmma_example.inc b/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_wmma_example.inc
new file mode 100644
index 0000000000..360b2c8947
--- /dev/null
+++ b/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_wmma_example.inc
@@ -0,0 +1,286 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+template <typename BiasLay, typename ResidualLay>
+struct LayoutSetting
+{
+    using BiasLayout     = BiasLay;
+    using ResidualLayout = ResidualLay;
+};
+
+template <ck::index_t NDimSpatial>
+struct LayoutSettingSelector;
+
+template <>
+struct LayoutSettingSelector<1> final : LayoutSetting<ctl::G_K, ctl::G_NW_K>
+{
+};
+
+template <>
+struct LayoutSettingSelector<2> final : LayoutSetting<ctl::G_K, ctl::G_NHW_K>
+{
+};
+
+template <>
+struct LayoutSettingSelector<3> final : LayoutSetting<ctl::G_K, ctl::G_NDHW_K>
+{
+};
+
+template <ck::index_t NDimSpatial>
+using BiasLayout = typename LayoutSettingSelector<NDimSpatial>::BiasLayout;
+
+template <ck::index_t NDimSpatial>
+using ResidualLayout = typename LayoutSettingSelector<NDimSpatial>::ResidualLayout;
+
+template <ck::index_t NDimSpatial>
+using DeviceConvFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Wmma_CShuffle<
+        NDimSpatial,
+        InputLayout<NDimSpatial>,
+        WeightLayout<NDimSpatial>,
+        ck::Tuple<BiasLayout<NDimSpatial>, ResidualLayout<NDimSpatial>>,
+        OutputLayout<NDimSpatial>,
+        InKernelDataType,
+        WeiKernelDataType,
+        ck::Tuple<BiasKernelDataType, ResidualKernelDataType>,
+        OutKernelDataType,
+        AccDataType,
+        CShuffleDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        256,         // BlockSize
+        128,         // MPerBlock
+        128,         // NPerBlock
+        4,           // K0PerBlock
+        8,           // K1
+        16,          // MPerWMMA
+        16,          // NPerWMMA
+        4,           // MRepeat
+        2,           // NRepeat
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        8,           // ABlockTransferSrcScalarPerVector
+        8,           // ABlockTransferDstScalarPerVector_AK1
+        true,        // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        8,           // BBlockTransferSrcScalarPerVector
+        8,           // BBlockTransferDstScalarPerVector_BK1
+        true,        // BBlockLdsExtraN
+        4,
+        2,
+        S<1, 32, 1, 8>,
+        8>;
+
+template <ck::index_t NDimSpatial>
+using HostConvFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                         InUserDataType,
+                                                                         WeiUserDataType,
+                                                                         CShuffleDataType,
+                                                                         InElementOp,
+                                                                         WeiElementOp,
+                                                                         PassThrough>;
+
+template <ck::index_t NDimSpatial>
+bool run_grouped_conv_fwd_bias_relu_add(const ExecutionConfig& config,
+                                        const ck::utils::conv::ConvParam& conv_param)
+{
+    static_assert(1 <= NDimSpatial && NDimSpatial <= 3, "Unsupported NDimSpatial");
+
+    const auto in_g_n_c_wis_desc   = make_input_descriptor(conv_param);
+    const auto wei_g_k_c_xs_desc   = make_weight_descriptor(conv_param);
+    const auto bias_g_n_k_wos_desc = make_bias_descriptor(conv_param);
+    const auto out_g_n_k_wos_desc  = make_output_descriptor(conv_param);
+
+    Tensor<InUserDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiUserDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<OutUserDataType> bias(bias_g_n_k_wos_desc);
+    Tensor<OutUserDataType> residual(bias_g_n_k_wos_desc);
+    Tensor<OutUserDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutKernelDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "bias: " << bias.mDesc << std::endl;
+    std::cout << "residual: " << residual.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InUserDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiUserDataType>{-5, 5});
+        bias.GenerateTensorValue(GeneratorTensor_2<OutUserDataType>{-5, 5});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InUserDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiUserDataType>{-0.5, 0.5});
+        bias.GenerateTensorValue(GeneratorTensor_3<OutUserDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InKernelDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiKernelDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(OutKernelDataType) * bias.mDesc.GetElementSpaceSize());
+    DeviceMem residual_device_buf(sizeof(OutKernelDataType) * residual.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutKernelDataType) * out_device.mDesc.GetElementSpaceSize());
+
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<InKernelDataType> in_converted(in);
+    const Tensor<WeiKernelDataType> wei_converted(wei);
+    const Tensor<OutKernelDataType> bias_converted(bias);
+    const Tensor<OutKernelDataType> residual_converted(residual);
+
+    in_device_buf.ToDevice(in_converted.mData.data());
+    wei_device_buf.ToDevice(wei_converted.mData.data());
+    bias_device_buf.ToDevice(bias_converted.mData.data());
+    residual_device_buf.ToDevice(residual_converted.mData.data());
+#else
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    bias_device_buf.ToDevice(bias.mData.data());
+    residual_device_buf.ToDevice(residual.mData.data());
+#endif
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d1_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d1_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(bias_g_n_k_wos_desc.GetLengths(), d0_g_n_k_wos_lengths);
+    copy(bias_g_n_k_wos_desc.GetStrides(), d0_g_n_k_wos_strides);
+    copy(bias_g_n_k_wos_desc.GetLengths(), d1_g_n_k_wos_lengths);
+    copy(bias_g_n_k_wos_desc.GetStrides(), d1_g_n_k_wos_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv    = DeviceConvFwdInstance<NDimSpatial>{};
+    auto invoker = conv.MakeInvoker();
+    auto argument =
+        conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                          wei_device_buf.GetDeviceBuffer(),
+                          std::array<const void*, 2>{bias_device_buf.GetDeviceBuffer(),
+                                                     residual_device_buf.GetDeviceBuffer()},
+                          out_device_buf.GetDeviceBuffer(),
+                          a_g_n_c_wis_lengths,
+                          a_g_n_c_wis_strides,
+                          b_g_k_c_xs_lengths,
+                          b_g_k_c_xs_strides,
+                          std::array<std::array<ck::index_t, NDimSpatial + 3>, 2>{
+                              {d0_g_n_k_wos_lengths, d1_g_n_k_wos_lengths}},
+                          std::array<std::array<ck::index_t, NDimSpatial + 3>, 2>{
+                              {d0_g_n_k_wos_strides, d1_g_n_k_wos_strides}},
+                          e_g_n_k_wos_lengths,
+                          e_g_n_k_wos_strides,
+                          conv_filter_strides,
+                          conv_filter_dilations,
+                          input_left_pads,
+                          input_right_pads,
+                          InElementOp{},
+                          WeiElementOp{},
+                          OutElementOp{});
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InUserDataType, WeiUserDataType, OutUserDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        Tensor<CShuffleDataType> c_host(out_g_n_k_wos_desc);
+
+        auto ref_conv     = HostConvFwdInstance<NDimSpatial>{};
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  c_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  InElementOp{},
+                                                  WeiElementOp{},
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        // TODO: implement elementwise operation for host
+        out_host.ForEach([&](auto&, auto idx) {
+            OutElementOp{}(out_host(idx), c_host(idx), bias(idx), residual(idx));
+        });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+#ifdef BUILD_INT4_EXAMPLE
+        const Tensor<OutUserDataType> out_device_converted(out_device);
+
+        return ck::utils::check_err(
+            out_device_converted, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+#else
+        return ck::utils::check_err(
+            out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+#endif
+    }
+
+    return true;
+}
+
+bool run_grouped_conv_fwd_bias_relu_add_example(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+
+    if(!parse_cmd_args(argc, argv, config, conv_param))
+    {
+        return false;
+    }
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1: return run_grouped_conv_fwd_bias_relu_add<1>(config, conv_param);
+    case 2: return run_grouped_conv_fwd_bias_relu_add<2>(config, conv_param);
+    case 3: return run_grouped_conv_fwd_bias_relu_add<3>(config, conv_param);
+    }
+
+    return false;
+}
diff --git a/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc b/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc
new file mode 100644
index 0000000000..58ed69182e
--- /dev/null
+++ b/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc
@@ -0,0 +1,223 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+template <ck::index_t NDimSpatial>
+using DeviceConvFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InputLayout<NDimSpatial>,
+        WeightLayout<NDimSpatial>,
+        ck::Tuple<>,
+        OutputLayout<NDimSpatial>,
+        InKernelDataType,
+        WeiKernelDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutKernelDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        16,          // KPerBlock
+        4,           // AK1
+        4,           // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        4,           // ABlockTransferSrcScalarPerVector
+        4,           // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        4,           // BBlockTransferSrcScalarPerVector
+        4,           // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 16, 1, 16>,
+        4>;
+
+template <ck::index_t NDimSpatial>
+using HostConvFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                         InUserDataType,
+                                                                         WeiUserDataType,
+                                                                         CShuffleDataType,
+                                                                         InElementOp,
+                                                                         WeiElementOp,
+                                                                         PassThrough>;
+
+template <ck::index_t NDimSpatial>
+bool run_grouped_conv_fwd(const ExecutionConfig& config,
+                          const ck::utils::conv::ConvParam& conv_param)
+{
+    static_assert(1 <= NDimSpatial && NDimSpatial <= 3, "Unsupported NDimSpatial");
+
+    const auto in_g_n_c_wis_desc  = make_input_descriptor(conv_param);
+    const auto wei_g_k_c_xs_desc  = make_weight_descriptor(conv_param);
+    const auto out_g_n_k_wos_desc = make_output_descriptor(conv_param);
+
+    Tensor<InUserDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiUserDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<OutUserDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutKernelDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InUserDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiUserDataType>{-5, 5});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InUserDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiUserDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InKernelDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiKernelDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutKernelDataType) * out_device.mDesc.GetElementSpaceSize());
+
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<InKernelDataType> in_converted(in);
+    const Tensor<WeiKernelDataType> wei_converted(wei);
+
+    in_device_buf.ToDevice(in_converted.mData.data());
+    wei_device_buf.ToDevice(wei_converted.mData.data());
+#else
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+#endif
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvFwdInstance<NDimSpatial>{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                      wei_device_buf.GetDeviceBuffer(),
+                                      std::array<const void*, 0>{},
+                                      out_device_buf.GetDeviceBuffer(),
+                                      a_g_n_c_wis_lengths,
+                                      a_g_n_c_wis_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{},
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{},
+                                      e_g_n_k_wos_lengths,
+                                      e_g_n_k_wos_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      InElementOp{},
+                                      WeiElementOp{},
+                                      OutElementOp{});
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InUserDataType, WeiUserDataType, OutUserDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        auto ref_conv     = HostConvFwdInstance<NDimSpatial>{};
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  out_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  InElementOp{},
+                                                  WeiElementOp{},
+                                                  OutElementOp{});
+
+        ref_invoker.Run(ref_argument);
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+#ifdef BUILD_INT4_EXAMPLE
+        const Tensor<OutUserDataType> out_device_converted(out_device);
+
+        return ck::utils::check_err(
+            out_device_converted.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+#else
+        return ck::utils::check_err(
+            out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+#endif
+    }
+
+    return true;
+}
+
+bool run_grouped_conv_fwd_example(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+
+    if(!parse_cmd_args(argc, argv, config, conv_param))
+    {
+        return false;
+    }
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1: return run_grouped_conv_fwd<1>(config, conv_param);
+    case 2: return run_grouped_conv_fwd<2>(config, conv_param);
+    case 3: return run_grouped_conv_fwd<3>(config, conv_param);
+    }
+
+    return false;
+}
diff --git a/example/31_batched_gemm_gemm/CMakeLists.txt b/example/31_batched_gemm_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..2074520f8c
--- /dev/null
+++ b/example/31_batched_gemm_gemm/CMakeLists.txt
@@ -0,0 +1,26 @@
+list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list2 gfx908 gfx90a)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+      add_example_executable(example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp)
+   endif()
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp)
+   endif()
+   if(USE_BITINT_EXTENSION_INT4)
+      add_example_executable(example_batched_gemm_gemm_xdl_int4 batched_gemm_gemm_xdl_int4.cpp)
+   endif(USE_BITINT_EXTENSION_INT4)
+   set(target 1)
+ endif()
+endforeach()
+
+if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1")
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
+   endif()
+endif()
diff --git a/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_bf16.cpp b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_bf16.cpp
new file mode 100644
index 0000000000..7605d9c4f8
--- /dev/null
+++ b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_bf16.cpp
@@ -0,0 +1,136 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
+                                              |------------|
+                                                   Gemm0
+                                              |---------------------|
+                                                       Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = BF16;
+using B0DataType       = BF16;
+using B1DataType       = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = BF16;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = PassThrough;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmDefault,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    32,          // KPerBlock
+    128,         // Gemm1NPerBlock
+    32,          // Gemm1KPerBlock
+    8,           // AK1
+    8,           // BK1
+    2,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    4,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<8, 32, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    2,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                ADataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                CElementOp>;
+
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return run_batched_gemm_gemm_example(argc, argv) ? 0 : 1; }
diff --git a/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp
new file mode 100644
index 0000000000..92e746ca77
--- /dev/null
+++ b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp
@@ -0,0 +1,137 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
+                                              |------------|
+                                                   Gemm0
+                                              |---------------------|
+                                                       Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = PassThrough;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmDefault,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    32,          // KPerBlock
+    128,         // Gemm1NPerBlock
+    32,          // Gemm1KPerBlock
+    8,           // AK1
+    8,           // BK1
+    2,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    4,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<8, 32, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    2,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                ADataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                CElementOp>;
+
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return run_batched_gemm_gemm_example(argc, argv) ? 0 : 1; }
diff --git a/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp
new file mode 100644
index 0000000000..e0eb193ad0
--- /dev/null
+++ b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp
@@ -0,0 +1,135 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
+                                              |------------|
+                                                   Gemm0
+                                              |---------------------|
+                                                       Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F32;
+using B0DataType       = F32;
+using B1DataType       = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F32;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = PassThrough;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmDefault,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    16,          // KPerBlock
+    128,         // Gemm1NPerBlock
+    16,          // Gemm1KPerBlock
+    4,           // AK1
+    4,           // BK1
+    1,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    4,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    4,
+    4,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    4,
+    4,
+    true,
+    S<8, 32, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    1,
+    false,
+    1,               // CShuffleMXdlPerWavePerShuffle
+    2,               // CShuffleNXdlPerWavePerShuffle
+    S<1, 16, 1, 16>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    4>;              // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                ADataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                CElementOp>;
+
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return run_batched_gemm_gemm_example(argc, argv) ? 0 : 1; }
diff --git a/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp
new file mode 100644
index 0000000000..d166214c33
--- /dev/null
+++ b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp
@@ -0,0 +1,146 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
+                                              |------------|
+                                                   Gemm0
+                                              |---------------------|
+                                                       Gemm1
+*/
+
+#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+#error Should compile this file with ck::int4_t support
+#endif
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = ck::int4_t;
+using B0DataType       = ck::int4_t;
+using B1DataType       = ck::int4_t;
+using KernelADataType  = int8_t;
+using KernelB0DataType = int8_t;
+using KernelB1DataType = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int32_t;
+using CDataType        = ck::int4_t;
+using KernelCDataType  = int8_t;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = PassThrough;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    KernelADataType,
+    KernelB0DataType,
+    KernelB1DataType,
+    KernelCDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmDefault,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    64,          // KPerBlock
+    128,         // Gemm1NPerBlock
+    64,          // Gemm1KPerBlock
+    16,          // AK1
+    16,          // BK1
+    4,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    4,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    16,
+    16,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    16,
+    16,
+    true,
+    S<8, 32, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    4,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                ADataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                CElementOp>;
+
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#define BUILD_INT4_EXAMPLE
+#include "run_batched_gemm_gemm_example.inc"
+
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+#endif
+
+int main(int argc, char* argv[]) { return run_batched_gemm_gemm_example(argc, argv) ? 0 : 1; }
diff --git a/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp
new file mode 100644
index 0000000000..40f87d1f55
--- /dev/null
+++ b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp
@@ -0,0 +1,133 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Gemm fused operation. Computes C_m_o = A_m_k * B0_k_n * B1_n_o
+                                              |------------|
+                                                   Gemm0
+                                              |---------------------|
+                                                       Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = int8_t;
+using B0DataType       = int8_t;
+using B1DataType       = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = int32_t;
+using CDataType        = int8_t;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = PassThrough;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmDefault,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    64,          // KPerBlock
+    128,         // Gemm1NPerBlock
+    64,          // Gemm1KPerBlock
+    16,          // AK1
+    16,          // BK1
+    4,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    4,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    16,
+    16,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    16,
+    16,
+    true,
+    S<8, 32, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    4,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                ADataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                CElementOp>;
+
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return run_batched_gemm_gemm_example(argc, argv) ? 0 : 1; }
diff --git a/example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc b/example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc
new file mode 100644
index 0000000000..445fbc032e
--- /dev/null
+++ b/example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc
@@ -0,0 +1,277 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+bool run_batched_gemm_gemm_example(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M             = 1024;
+    ck::index_t N             = 1024;
+    ck::index_t K             = 64;
+    ck::index_t O             = 128;
+    ck::index_t BatchCount    = 4;
+    ck::index_t StrideA       = -1;
+    ck::index_t StrideB0      = -1;
+    ck::index_t StrideB1      = -1;
+    ck::index_t StrideC       = -1;
+    ck::index_t BatchStrideA  = -1;
+    ck::index_t BatchStrideB0 = -1;
+    ck::index_t BatchStrideB1 = -1;
+    ck::index_t BatchStrideC  = -1;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 9)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+    }
+    else if(argc == 17)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+
+        StrideA  = std::stoi(argv[9]);
+        StrideB0 = std::stoi(argv[10]);
+        StrideB1 = std::stoi(argv[11]);
+        StrideC  = std::stoi(argv[12]);
+
+        BatchStrideA  = std::stoi(argv[13]);
+        BatchStrideB0 = std::stoi(argv[14]);
+        BatchStrideB1 = std::stoi(argv[15]);
+        BatchStrideC  = std::stoi(argv[16]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 17: M, N, K, O, Batch, StrideA, StrideB0, StrideB1, StrideC, BatchStrideA, "
+               "BatchStrideB0, BatchStrideB1, BatchStrideC\n");
+        exit(0);
+    }
+
+    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideC  = ck::is_same_v<CLayout, Row> ? O : M;
+
+    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
+    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideC  = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
+    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideC  = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
+
+    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
+    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideC  = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
+        }
+    };
+
+    // C_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<CDataType> c_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    Tensor<CDataType> c_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-3.f, 3.f}(a_g_m_k);
+        ck::utils::FillUniformDistributionIntegerValue<B0DataType>{-3.f, 3.f}(b0_g_k_n);
+        ck::utils::FillUniformDistributionIntegerValue<B1DataType>{-3.f, 3.f}(b1_g_n_o);
+        break;
+    case 2:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+#ifdef BUILD_INT4_EXAMPLE
+    DeviceMem a_g_m_k_device_buf(sizeof(KernelADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(KernelB0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(KernelB1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
+    DeviceMem c_g_m_o_device_buf(sizeof(KernelCDataType) *
+                                 c_g_m_o_device_result.mDesc.GetElementSpaceSize());
+
+    const Tensor<KernelADataType> a_g_m_k_converted(a_g_m_k);
+    const Tensor<KernelB0DataType> b0_g_k_n_converted(b0_g_k_n);
+    const Tensor<KernelB1DataType> b1_g_n_o_converted(b1_g_n_o);
+
+    a_g_m_k_device_buf.ToDevice(a_g_m_k_converted.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n_converted.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o_converted.mData.data());
+#else
+    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
+    DeviceMem c_g_m_o_device_buf(sizeof(CDataType) *
+                                 c_g_m_o_device_result.mDesc.GetElementSpaceSize());
+
+    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+#endif
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(
+#ifdef BUILD_INT4_EXAMPLE
+        static_cast<KernelADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
+        static_cast<KernelB0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+        static_cast<KernelB1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+        static_cast<KernelCDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
+#else
+        static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
+        static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+        static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+        static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
+#endif
+        M,
+        N,
+        K,
+        O,
+        BatchCount,
+        StrideA,
+        StrideB0,
+        StrideB1,
+        StrideC,
+        BatchStrideA,
+        BatchStrideB0,
+        BatchStrideB1,
+        BatchStrideC,
+        a_element_op,
+        b0_element_op,
+        acc0_element_op,
+        b1_element_op,
+        c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                             sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                            BatchCount;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        // Output of Gemm0 is input A of Gemm1
+        Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, a1_g_m_n, a_element_op, b0_element_op, PassThrough{});
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+#ifdef BUILD_INT4_EXAMPLE
+        Tensor<KernelCDataType> c_g_m_o_device_result_converted(c_g_m_o_host_result.mDesc);
+
+        c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result_converted.mData.data());
+
+        c_g_m_o_device_result = c_g_m_o_device_result_converted.CopyAsType<CDataType>();
+#else
+        c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
+#endif
+
+        return ck::utils::check_err(c_g_m_o_device_result, c_g_m_o_host_result);
+    }
+
+    return true;
+}
diff --git a/example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt b/example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..0463bf6bd3
--- /dev/null
+++ b/example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
@@ -0,0 +1,24 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
+    add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
+    add_example_executable(example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
+    add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
+    add_example_executable(example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_bf16 batched_gemm_scale_softmax_gemm_xdl_bf16.cpp)
+    add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16 batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp)
+endif()
+
+add_custom_target(example_gemm_scale_softmax_gemm)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
+    add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
+    add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
+    add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
+    add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_bf16)
+    add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16)
+endif()
diff --git a/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
new file mode 100644
index 0000000000..1d1566d575
--- /dev/null
+++ b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
@@ -0,0 +1,160 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskOutUpperTriangle;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_scale_softmax_gemm_permute.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }
diff --git a/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp
new file mode 100644
index 0000000000..bae88d4b8e
--- /dev/null
+++ b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp
@@ -0,0 +1,159 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = BF16;
+using B0DataType       = BF16;
+using B1DataType       = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = BF16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: bf16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, bf16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: bf16 in, bf16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_scale_softmax_gemm_permute.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }
diff --git a/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
new file mode 100644
index 0000000000..a098ce6675
--- /dev/null
+++ b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
@@ -0,0 +1,160 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_scale_softmax_gemm_permute.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }
diff --git a/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp
new file mode 100644
index 0000000000..ce8caf7588
--- /dev/null
+++ b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp
@@ -0,0 +1,143 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = BF16;
+using B0DataType       = BF16;
+using B1DataType       = BF16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = BF16;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmSpec,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    32,          // KPerBlock
+    64,          // Gemm1NPerBlock
+    32,          // Gemm1KPerBlock
+    8,           // AK1
+    8,           // BK1
+    2,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    2,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<16, 16, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    2,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+    false>;
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_scale_softmax_gemm.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }
diff --git a/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp
new file mode 100644
index 0000000000..138db14963
--- /dev/null
+++ b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp
@@ -0,0 +1,144 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+
+using ALayout  = Row;
+using B0Layout = Col;
+using B1Layout = Row;
+using CLayout  = Row;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+    ALayout,
+    B0Layout,
+    B1Layout,
+    CLayout,
+    ADataType,
+    B0DataType,
+    B1DataType,
+    CDataType,
+    AccDataType,
+    CShuffleDataType,
+    AElementOp,
+    B0ElementOp,
+    Acc0ElementOp,
+    B1ElementOp,
+    CElementOp,
+    GemmSpec,
+    1,
+    256,
+    128,         // MPerBlock
+    128,         // NPerBlock
+    32,          // KPerBlock
+    64,          // Gemm1NPerBlock
+    32,          // Gemm1KPerBlock
+    8,           // AK1
+    8,           // BK1
+    2,           // B1K1
+    32,          // MPerXDL
+    32,          // NPerXDL
+    1,           // MXdlPerWave
+    4,           // NXdlPerWave
+    2,           // Gemm1NXdlPerWave
+    S<4, 64, 1>, // ABlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<4, 64, 1>, // BBlockTransfer
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    true,
+    S<16, 16, 1>, // B1BlockTransfer
+    S<0, 2, 1>,
+    S<0, 2, 1>,
+    1,
+    4,
+    2,
+    false,
+    1,              // CShuffleMXdlPerWavePerShuffle
+    2,              // CShuffleNXdlPerWavePerShuffle
+    S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+    8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+    false>;
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_batched_gemm_scale_softmax_gemm.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }
diff --git a/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
new file mode 100644
index 0000000000..5794924294
--- /dev/null
+++ b/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
@@ -0,0 +1,159 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskOutUpperTriangle;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_grouped_gemm_scale_softmax_gemm_permute.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }
diff --git a/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
new file mode 100644
index 0000000000..97caec6053
--- /dev/null
+++ b/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
@@ -0,0 +1,160 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
+                                                                  |-----------------|
+                                                                          Gemm0
+                                                          |-------------------------------------|
+                                                                          Gemm1
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using Acc0BiasDataType = ck::Tuple<>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using AElementOp    = PassThrough;
+using B0ElementOp   = PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
+using B1ElementOp   = PassThrough;
+using CElementOp    = PassThrough;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+static constexpr auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+#include "run_grouped_gemm_scale_softmax_gemm_permute.inc"
+
+int main(int argc, char* argv[]) { return run(argc, argv); }
diff --git a/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc b/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc
new file mode 100644
index 0000000000..27602e2313
--- /dev/null
+++ b/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc
@@ -0,0 +1,261 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+int run(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 2;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M             = 1020;
+    ck::index_t N             = 1020;
+    ck::index_t K             = 64;
+    ck::index_t O             = 128;
+    ck::index_t BatchCount    = 4;
+    ck::index_t StrideA       = -1;
+    ck::index_t StrideB0      = -1;
+    ck::index_t StrideB1      = -1;
+    ck::index_t StrideC       = -1;
+    ck::index_t BatchStrideA  = -1;
+    ck::index_t BatchStrideB0 = -1;
+    ck::index_t BatchStrideB1 = -1;
+    ck::index_t BatchStrideC  = -1;
+    float alpha               = 1;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 9)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+    }
+    else if(argc == 18)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+
+        StrideA  = std::stoi(argv[9]);
+        StrideB0 = std::stoi(argv[10]);
+        StrideB1 = std::stoi(argv[11]);
+        StrideC  = std::stoi(argv[12]);
+
+        BatchStrideA  = std::stoi(argv[13]);
+        BatchStrideB0 = std::stoi(argv[14]);
+        BatchStrideB1 = std::stoi(argv[15]);
+        BatchStrideC  = std::stoi(argv[16]);
+
+        alpha = std::stof(argv[17]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 16: M, N, K, O, Batch, StrideA, StrideB0, StrideB1, StrideC, BatchStrideA, "
+               "BatchStrideB0, BatchStrideB1, BatchStrideC\n");
+        printf("arg17: scale (alpha)\n");
+        exit(0);
+    }
+
+    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideC  = ck::is_same_v<CLayout, Row> ? O : M;
+
+    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
+    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideC  = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
+    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideC  = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
+
+    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
+    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideC  = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
+                                        std::vector<std::size_t>({batch_stride, stride, 1}));
+        }
+        else
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
+                                        std::vector<std::size_t>({batch_stride, 1, stride}));
+        }
+    };
+
+    // C_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<CDataType> c_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    Tensor<CDataType> c_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        break;
+    case 2:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSpaceSize());
+    DeviceMem c_g_m_o_device_buf(sizeof(CDataType) *
+                                 c_g_m_o_device_result.mDesc.GetElementSpaceSize());
+
+    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    // do GEMM
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+    auto argument =
+        gemm.MakeArgument(static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
+                          static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+                          static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+                          static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
+                          M,
+                          N,
+                          K,
+                          O,
+                          BatchCount,
+                          StrideA,
+                          StrideB0,
+                          StrideB1,
+                          StrideC,
+                          BatchStrideA,
+                          BatchStrideB0,
+                          BatchStrideB1,
+                          BatchStrideC,
+                          a_element_op,
+                          b0_element_op,
+                          acc0_element_op,
+                          b1_element_op,
+                          c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                             sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                            BatchCount;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
+
+    if(do_verification)
+    {
+        // Output of Gemm0 is input A of Gemm1
+        Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+        Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        return ck::utils::check_err(c_g_m_o_device_result.mData, c_g_m_o_host_result.mData) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc b/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc
new file mode 100644
index 0000000000..fa76faea84
--- /dev/null
+++ b/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc
@@ -0,0 +1,278 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+int run(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape for A/B0/B1/C
+    // C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
+    ck::index_t M = 120;
+    ck::index_t N = 1000;
+    ck::index_t K = 64;
+    ck::index_t O = 128;
+
+    // Output shape C[G0, M, G1, O]. Batch dim, outer dim, inner dim must match GEMM shape
+    // C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o])
+    // C_g0_m_g1_o = permute(C_g0_g1_m_o, [0, 2, 1, 3])
+    ck::index_t G0 = 7;
+    ck::index_t G1 = 13;
+
+    float alpha = 1;
+
+    bool input_permute  = false;
+    bool output_permute = true;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 13)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M  = std::stoi(argv[4]);
+        N  = std::stoi(argv[5]);
+        K  = std::stoi(argv[6]);
+        O  = std::stoi(argv[7]);
+        G0 = std::stoi(argv[8]);
+        G1 = std::stoi(argv[9]);
+
+        alpha = std::stof(argv[10]);
+
+        input_permute  = std::stoi(argv[11]);
+        output_permute = std::stoi(argv[12]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 11: M, N, K, O, G0, G1\n");
+        printf("arg10: scale (alpha)\n");
+        printf("arg11 to 12: input / output permute\n");
+        exit(0);
+    }
+
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides =
+        input_permute
+            ? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
+            : std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, K]
+
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides =
+        input_permute
+            ? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // B0 layout [G0, N, G1, K]
+            : std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
+
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides =
+        input_permute
+            ? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // B1 layout [G0, N, G1, O]
+            : std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, O]
+
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides =
+        output_permute
+            ? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
+            : std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+    Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+    Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+    Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
+    std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
+    std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) *
+                           c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    // do GEMM
+    // TODO ANT: replace array with vector?
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(
+        static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+        static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
+        static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
+        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+        {}, // std::array<void*, 1> p_acc0_biases;
+        {}, // std::array<void*, 1> p_acc1_biases;
+        a_gs_ms_ks_lengths,
+        a_gs_ms_ks_strides,
+        b0_gs_ns_ks_lengths,
+        b0_gs_ns_ks_strides,
+        b1_gs_os_ns_lengths,
+        b1_gs_os_ns_strides,
+        c_gs_ms_os_lengths,
+        c_gs_ms_os_strides,
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_lengths},
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_strides},
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
+        {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_strides},
+        a_element_op,
+        b0_element_op,
+        acc0_element_op,
+        b1_element_op,
+        c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    ck::index_t BatchCount = G0 * G1;
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                             sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                            BatchCount;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+        Tensor<ADataType> a_g_m_k({BatchCount, M, K});
+        Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
+        Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
+        Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N});        // scratch object after gemm0
+        Tensor<ADataType> a1_g_m_n({BatchCount, M, N});            // scratch object after softmax
+        Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
+
+        // permute
+        a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+            a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+        b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+            b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+            b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+
+        // gemm 0
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        // masking
+        const auto mask = DeviceGemmInstance::C0MatrixMask(N);
+        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+            if(mask.IsMaskedElement(idx[1], idx[2]))
+                self(idx) = -ck::NumericLimits<float>::Infinity();
+        });
+
+        // softmax
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        // gemm1
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        // permute
+        c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+            const size_t& g0 = idx[0];
+            const size_t& g1 = idx[1];
+
+            const size_t g = g0 * G1 + g1;
+
+            self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+        });
+
+        // default absolute error and relative error is 0.001
+        double rtol = 1e-3;
+        double atol = 1e-3;
+
+        // when BF16 is taken, set absolute error and relative error to 0.01
+        if(std::is_same_v<ADataType, ck::bhalf_t> && std::is_same_v<B0DataType, ck::bhalf_t> &&
+           std::is_same_v<B1DataType, ck::bhalf_t> && std::is_same_v<CDataType, ck::bhalf_t>)
+        {
+            rtol = 1e-2;
+            atol = 1e-2;
+        }
+
+        return ck::utils::check_err(c_gs_ms_os_device_result.mData,
+                                    c_gs_ms_os_host_result.mData,
+                                    "Error: Incorrect results!",
+                                    rtol,
+                                    atol)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}
diff --git a/example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc b/example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc
new file mode 100644
index 0000000000..ea1e2734a6
--- /dev/null
+++ b/example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc
@@ -0,0 +1,319 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+int run(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    bool input_permute  = false;
+    bool output_permute = true;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        input_permute  = std::stoi(argv[4]);
+        output_permute = std::stoi(argv[5]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 5: input / output permute\n");
+        exit(0);
+    }
+
+    float alpha = 1; // scaling after 1st gemm
+
+    std::size_t group_count = 7;
+
+    // Problem descs
+    std::vector<DeviceGemmInstance::ProblemDesc> problem_descs;
+    std::vector<const void*> p_a;
+    std::vector<const void*> p_b0;
+    std::vector<const void*> p_b1;
+    std::vector<void*> p_c;
+    std::vector<std::vector<int>> g0_g1_m_n_k_o;
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    std::vector<Tensor<B0DataType>> b0_tensors;
+    std::vector<Tensor<B1DataType>> b1_tensors;
+    std::vector<Tensor<CDataType>> c_tensors;
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+    std::vector<DeviceMemPtr> a_tensors_device;
+    std::vector<DeviceMemPtr> b0_tensors_device;
+    std::vector<DeviceMemPtr> b1_tensors_device;
+    std::vector<DeviceMemPtr> c_tensors_device;
+
+    std::size_t flop = 0, num_byte = 0;
+
+    std::cout << "group count " << group_count << ". printing first 4 groups\n";
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        int M  = 128 * (rand() % 8 + 1);
+        int N  = 128 * (rand() % 8 + 1);
+        int K  = 40;
+        int O  = 40 * (rand() % 2 + 1);
+        int G0 = rand() % 3 + 1;
+        int G1 = rand() % 5 + 1;
+
+        g0_g1_m_n_k_o.push_back({G0, G1, M, N, K, O});
+
+        std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+        std::vector<ck::index_t> a_gs_ms_ks_strides =
+            input_permute
+                ? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
+                : std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, K]
+
+        std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+        std::vector<ck::index_t> b0_gs_ns_ks_strides =
+            input_permute
+                ? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // B0 layout [G0, N, G1, K]
+                : std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
+
+        std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+        std::vector<ck::index_t> b1_gs_os_ns_strides =
+            input_permute
+                ? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // B1 layout [G0, N, G1, O]
+                : std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, O]
+
+        std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+        std::vector<ck::index_t> c_gs_ms_os_strides =
+            output_permute
+                ? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
+                : std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
+
+        problem_descs.push_back({a_gs_ms_ks_lengths,
+                                 a_gs_ms_ks_strides,
+                                 b0_gs_ns_ks_lengths,
+                                 b0_gs_ns_ks_strides,
+                                 b1_gs_os_ns_lengths,
+                                 b1_gs_os_ns_strides,
+                                 c_gs_ms_os_lengths,
+                                 c_gs_ms_os_strides,
+                                 {},   // acc0_biases_gs_ms_ns_lengths
+                                 {},   // acc0_biases_gs_ms_ns_strides
+                                 {},   // acc1_biases_gs_ms_os_lengths
+                                 {}}); // acc1_biases_gs_ms_os_strides
+
+        // C_m_o = A_m_k * B0_k_n * B1_n_o
+        Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+        Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+        Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+        Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+        int Batch = G0 * G1;
+        flop += (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * Batch;
+        num_byte += (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                    Batch;
+
+        if(i < 4)
+        {
+            std::cout << "a_gs_ms_ks[" << i << "]: " << a_gs_ms_ks.mDesc << ", "
+                      << "b0_gs_ns_ks[" << i << "]: " << b0_gs_ns_ks.mDesc << ", "
+                      << "b1_gs_os_ns[" << i << "]: " << b1_gs_os_ns.mDesc << ", "
+                      << "c_gs_ms_os[" << i << "]: " << c_gs_ms_os_device_result.mDesc << std::endl;
+        }
+
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+            break;
+        case 2:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+            break;
+        case 3:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+            break;
+        default:
+            a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+            b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+            b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        }
+
+        a_tensors.push_back(a_gs_ms_ks);
+        b0_tensors.push_back(b0_gs_ns_ks);
+        b1_tensors.push_back(b1_gs_os_ns);
+        c_tensors.push_back(c_gs_ms_os_device_result);
+
+        a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize()));
+        b0_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize()));
+        b1_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize()));
+        c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(CDataType) * c_gs_ms_os_device_result.mDesc.GetElementSpaceSize()));
+
+        a_tensors_device[i]->ToDevice(a_gs_ms_ks.mData.data());
+        b0_tensors_device[i]->ToDevice(b0_gs_ns_ks.mData.data());
+        b1_tensors_device[i]->ToDevice(b1_gs_os_ns.mData.data());
+
+        p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
+        p_b0.push_back(b0_tensors_device[i]->GetDeviceBuffer());
+        p_b1.push_back(b1_tensors_device[i]->GetDeviceBuffer());
+        p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
+    }
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(p_a,
+                                      p_b0,
+                                      p_b1,
+                                      p_c,
+                                      {}, // p_acc0_biases
+                                      {}, // p_acc1_biases
+                                      problem_descs,
+                                      a_element_op,
+                                      b0_element_op,
+                                      acc0_element_op,
+                                      b1_element_op,
+                                      c_element_op);
+
+    // specify workspace for problem_desc
+    DeviceMem problem_desc_workspace(gemm.GetWorkSpaceSize(&argument));
+
+    gemm.SetWorkSpacePointer(&argument, problem_desc_workspace.GetDeviceBuffer());
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        for(std::size_t i = 0; i < group_count; i++)
+        {
+            const int& G0 = g0_g1_m_n_k_o[i][0];
+            const int& G1 = g0_g1_m_n_k_o[i][1];
+            const int& M  = g0_g1_m_n_k_o[i][2];
+            const int& N  = g0_g1_m_n_k_o[i][3];
+            const int& K  = g0_g1_m_n_k_o[i][4];
+            const int& O  = g0_g1_m_n_k_o[i][5];
+
+            const auto& c_gs_ms_os_lengths = problem_descs[i].c_gs_ms_os_lengths;
+            const auto& c_gs_ms_os_strides = problem_descs[i].c_gs_ms_os_strides;
+
+            const auto& a_gs_ms_ks         = a_tensors[i];
+            const auto& b0_gs_ns_ks        = b0_tensors[i];
+            const auto& b1_gs_os_ns        = b1_tensors[i];
+            auto& c_gs_ms_os_device_result = c_tensors[i];
+            auto& c_gs_ms_os_device_buf    = *c_tensors_device[i];
+
+            c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+            Tensor<ADataType> a_g_m_k({G0 * G1, M, K});
+            Tensor<B0DataType> b0_g_k_n({G0 * G1, K, N});
+            Tensor<B1DataType> b1_g_n_o({G0 * G1, N, O});
+            Tensor<AccDataType> acc0_g_m_n({G0 * G1, M, N});        // scratch object after gemm0
+            Tensor<ADataType> a1_g_m_n({G0 * G1, M, N});            // scratch object after softmax
+            Tensor<CDataType> c_g_m_o_host_result({G0 * G1, M, O}); // scratch object after gemm1
+            Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+            // permute
+            a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+                a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+            });
+            b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+                b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+            });
+            b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+                b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+            });
+
+            // gemm 0
+            auto ref_gemm0          = ReferenceGemm0Instance{};
+            auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+            auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+                a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+            ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+            // masking
+            const auto mask = DeviceGemmInstance::C0MatrixMask(N);
+            acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+                if(mask.IsMaskedElement(idx[1], idx[2]))
+                    self(idx) = -ck::NumericLimits<float>::Infinity();
+            });
+
+            // softmax
+            auto ref_softmax          = ReferenceSoftmaxInstance{};
+            auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+            auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+            ref_softmax_invoker.Run(ref_softmax_argument);
+
+            // gemm 1
+            auto ref_gemm1          = ReferenceGemm1Instance{};
+            auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+            auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g_m_n,
+                                                             b1_g_n_o,
+                                                             c_g_m_o_host_result,
+                                                             PassThrough{},
+                                                             b1_element_op,
+                                                             c_element_op);
+
+            ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+            // permute
+            c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+                const size_t& g0 = idx[0];
+                const size_t& g1 = idx[1];
+
+                const size_t g = g0 * G1 + g1;
+
+                self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+            });
+
+            bool pass_ =
+                ck::utils::check_err(c_gs_ms_os_device_result.mData, c_gs_ms_os_host_result.mData);
+            pass &= pass_;
+        }
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/example/33_multiple_reduce/CMakeLists.txt b/example/33_multiple_reduce/CMakeLists.txt
new file mode 100644
index 0000000000..bc8c3eb04e
--- /dev/null
+++ b/example/33_multiple_reduce/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_example_executable(example_dual_reduce_multiblock dual_reduce_multiblock.cpp)
+add_example_executable(example_dual_reduce_threadwise dual_reduce_threadwise.cpp)
diff --git a/example/33_multiple_reduce/README.md b/example/33_multiple_reduce/README.md
new file mode 100644
index 0000000000..90762a692f
--- /dev/null
+++ b/example/33_multiple_reduce/README.md
@@ -0,0 +1,37 @@
+# Instructions for ```example_dual_reduce```
+
+## Run ```example_dual_reduce_multiblock```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+#arg1: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg2: time kernel (0=no, 1=yes) 
+./bin/example_dual_reduce_multiblock -D 600,28,28,256 -v 1 2 1
+```
+
+Result
+```
+./bin/example_dual_reduce_multiblock -D 600,28,28,256 -v 1 2 1                        
+launch_and_time_kernel: grid_dim {150, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 1.19529 ms, 201.499 GB/s, DeviceMultipleReduceBlockWise<256,M_C4_S1,K_C64_S1,InSrcVectorDim_1_InSrcVectorSize_1,OutDstVectorSize_1_1>
+```
+
+## Run ```example_dual_reduce_threadwise```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+#arg1: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg2: time kernel (0=no, 1=yes)
+./bin/example_dual_reduce_multiblock -D 8000,4,4,4 -v 1 2 1
+```
+
+Result
+```
+./bin/example_dual_reduce_threadwise -D 8000,4,4,4 -v 1 2 1
+launch_and_time_kernel: grid_dim {32, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 0.01512 ms, 71.9577 GB/s, DeviceMultipleReduceThreadwise<256,M_C256_S1,K_C1_S4,InSrcVectorDim_1_InSrcVectorSize_2,OutDstVectorSize_1_1>
+```
diff --git a/example/33_multiple_reduce/dual_reduce_common.hpp b/example/33_multiple_reduce/dual_reduce_common.hpp
new file mode 100644
index 0000000000..cd21790be6
--- /dev/null
+++ b/example/33_multiple_reduce/dual_reduce_common.hpp
@@ -0,0 +1,314 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths = {600, 28, 28, 256};
+    size_t n, h, w, c;
+
+    bool do_verification = true;
+    int init_method      = 2;
+    bool time_kernel     = true;
+
+    public:
+    SimpleAppArgs()
+    {
+        n = inLengths[0];
+        h = inLengths[1];
+        w = inLengths[2];
+        c = inLengths[3];
+    };
+
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
+                     "comparing with the host-based reduction"
+                  << std::endl;
+        std::cout << "Arg1 -- init method (0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg2 -- time kernel (0=no, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:l:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                if(inLengths.size() != 4)
+                    throw std::runtime_error(
+                        "Invalid option format! The number of integers is incorrect!");
+
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 2 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        n = inLengths[0];
+        h = inLengths[1];
+        w = inLengths[2];
+        c = inLengths[3];
+
+        return (0);
+    };
+};
+
+template <typename InDataType, typename OutDataType1, typename OutDataType2, typename AccDataType>
+static void mean_meansquare_host(const Tensor<InDataType>& in,
+                                 Tensor<OutDataType1>& mean_ref,
+                                 Tensor<OutDataType2>& meansquare_ref,
+                                 size_t n,
+                                 size_t h,
+                                 size_t w,
+                                 size_t c)
+
+{
+    auto thread_reduce_func = [&](auto iN) {
+        AccDataType mean       = ck::type_convert<AccDataType>(0.0f);
+        AccDataType meansquare = ck::type_convert<AccDataType>(0.0f);
+
+        // compute mean, meanquare, variance, invVariance
+        for(std::size_t iH = 0; iH < h; iH++)
+        {
+            for(std::size_t iW = 0; iW < w; iW++)
+            {
+                for(std::size_t iC = 0; iC < c; iC++)
+                {
+                    AccDataType curr_value = ck::type_convert<AccDataType>(in(iN, iH, iW, iC));
+
+                    mean += curr_value;
+                    meansquare += curr_value * curr_value;
+                };
+            }
+        };
+
+        mean       = mean / (h * w * c);
+        meansquare = meansquare / (h * w * c);
+
+        mean_ref(iN)       = ck::type_convert<OutDataType1>(mean);
+        meansquare_ref(iN) = ck::type_convert<OutDataType2>(meansquare);
+    };
+
+    std::size_t num_thread      = std::thread::hardware_concurrency();
+    std::size_t work_per_thread = (n + num_thread - 1) / num_thread;
+
+    std::vector<joinable_thread> threads(num_thread);
+
+    for(std::size_t it = 0; it < num_thread; it++)
+    {
+        std::size_t iN_begin = it * work_per_thread;
+        std::size_t iN_end   = std::min(static_cast<size_t>((it + 1) * work_per_thread), n);
+
+        auto f = [=] {
+            for(std::size_t iN = iN_begin; iN < iN_end; iN++)
+            {
+                thread_reduce_func(iN);
+            }
+        };
+
+        threads[it] = joinable_thread(f);
+    }
+};
+
+using ReduceOperation = ck::reduce::Add;
+
+using InElementwiseOperation_Mean  = ck::tensor_operation::element_wise::PassThrough;
+using AccElementwiseOperation_Mean = ck::tensor_operation::element_wise::UnaryDivide;
+
+using InElementwiseOperation_Meansquare  = ck::tensor_operation::element_wise::UnarySquare;
+using AccElementwiseOperation_Meansquare = ck::tensor_operation::element_wise::UnaryDivide;
+
+using InElementwiseOperationTuple =
+    ck::Tuple<InElementwiseOperation_Mean, InElementwiseOperation_Meansquare>;
+using AccElementwiseOperationTuple =
+    ck::Tuple<AccElementwiseOperation_Mean, AccElementwiseOperation_Meansquare>;
+
+template <typename DeviceDualReduce,
+          typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          int Rank,
+          int NumReduceDim>
+int mean_meansquare_dual_reduce_test(size_t n,
+                                     size_t h,
+                                     size_t w,
+                                     size_t c,
+                                     bool do_verification,
+                                     int init_method,
+                                     bool time_kernel,
+                                     const std::array<int, NumReduceDim> reduceDims)
+{
+    const std::vector<size_t> inLengths = {n, h, w, c};
+
+    Tensor<InDataType> in(inLengths);
+
+    std::vector<size_t> outLengths{n};
+
+    Tensor<OutDataType> mean_ref(outLengths);
+    Tensor<OutDataType> mean(outLengths);
+    Tensor<OutDataType> meansquare_ref(outLengths);
+    Tensor<OutDataType> meansquare(outLengths);
+
+    auto inStrides  = in.mDesc.GetStrides();
+    auto outStrides = mean.mDesc.GetStrides();
+
+    size_t invariant_total_length = n;
+    size_t reduce_total_length    = h * w * c;
+
+    const double alpha = 1.0f;
+    const double beta  = 0.0f;
+
+    std::size_t num_thread = 1;
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0: break;
+        case 1: in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread); break;
+        case 2: in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread); break;
+        default: in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0}, num_thread);
+        }
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem mean_dev(sizeof(OutDataType) * mean.mDesc.GetElementSpaceSize());
+    DeviceMem meansquare_dev(sizeof(OutDataType) * meansquare.mDesc.GetElementSpaceSize());
+
+    in_dev.ToDevice(in.mData.data());
+
+    if(do_verification)
+    {
+        mean_meansquare_host<InDataType, OutDataType, OutDataType, AccDataType>(
+            in, mean_ref, meansquare_ref, n, h, w, c);
+    };
+
+    constexpr ck::index_t NumInputDim  = Rank;
+    constexpr ck::index_t NumOutputDim = (Rank - NumReduceDim > 1) ? Rank - NumReduceDim : 1;
+
+    std::array<ck::index_t, NumInputDim> i_inLengths;
+    std::array<ck::index_t, NumInputDim> i_inStrides;
+    std::array<ck::index_t, NumOutputDim> i_outLengths;
+    std::array<ck::index_t, NumOutputDim> i_outStrides;
+
+    ck::ranges::copy(inLengths, i_inLengths.begin());
+    ck::ranges::copy(inStrides, i_inStrides.begin());
+    ck::ranges::copy(outLengths, i_outLengths.begin());
+    ck::ranges::copy(outStrides, i_outStrides.begin());
+
+    auto dual_reduce_op = DeviceDualReduce{};
+
+    auto argument_ptr = dual_reduce_op.MakeArgumentPointer(
+        i_inLengths,
+        i_inStrides,
+        i_outLengths,
+        {i_outStrides, i_outStrides},
+        reduceDims,
+        {alpha, alpha},
+        {beta, beta},
+        in_dev.GetDeviceBuffer(),
+        {mean_dev.GetDeviceBuffer(), meansquare_dev.GetDeviceBuffer()},
+        ck::make_tuple(InElementwiseOperation_Mean{}, InElementwiseOperation_Meansquare{}),
+        ck::make_tuple(
+            AccElementwiseOperation_Mean{static_cast<int32_t>(reduce_total_length)},
+            AccElementwiseOperation_Meansquare{static_cast<int32_t>(reduce_total_length)}));
+
+    if(!dual_reduce_op.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout
+            << "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
+            << std::endl;
+        return (-1);
+    };
+
+    std::string reduce_name = dual_reduce_op.GetTypeString();
+
+    auto invoker_ptr = dual_reduce_op.MakeInvokerPointer();
+
+    float avg_time = 0.0f;
+
+    avg_time += invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InDataType) +
+                            2 * invariant_total_length * sizeof(OutDataType);
+
+    float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+    std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        mean_dev.FromDevice(mean.mData.data());
+        meansquare_dev.FromDevice(meansquare.mData.data());
+        pass = pass && ck::utils::check_err(mean, mean_ref);
+        pass = pass && ck::utils::check_err(meansquare, meansquare_ref);
+    };
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/33_multiple_reduce/dual_reduce_multiblock.cpp b/example/33_multiple_reduce/dual_reduce_multiblock.cpp
new file mode 100644
index 0000000000..198931749b
--- /dev/null
+++ b/example/33_multiple_reduce/dual_reduce_multiblock.cpp
@@ -0,0 +1,98 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+
+#include "dual_reduce_common.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+using InDataType       = ck::half_t;
+using OutDataType      = float;
+using OutDataTypeTuple = Tuple<OutDataType, OutDataType>;
+using AccDataType      = float;
+
+// for NHWC layer-norm calculation of mean and meansquare
+constexpr int Rank         = 4;
+constexpr int NumReduceDim = 3;
+
+constexpr bool PropagateNan = false;
+
+constexpr InMemoryDataOperationEnum OutMemoryDataOperation = InMemoryDataOperationEnum::Set;
+
+using DeviceDualReduce = DeviceMultipleReduceMultiBlock<2,
+                                                        InDataType,
+                                                        AccDataType,
+                                                        OutDataTypeTuple,
+                                                        Rank,
+                                                        NumReduceDim,
+                                                        ReduceOperation,
+                                                        InElementwiseOperationTuple,
+                                                        AccElementwiseOperationTuple,
+                                                        OutMemoryDataOperation,
+                                                        PropagateNan,
+                                                        256,
+                                                        4,
+                                                        64,
+                                                        1,
+                                                        1,
+                                                        1, // InSrcVectorDim
+                                                        1,
+                                                        ck::Sequence<1, 1>>;
+
+int main(int argc, char* argv[])
+{
+    int retval = 0;
+
+    if(argc > 1)
+    {
+        SimpleAppArgs arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        std::array<int, NumReduceDim> reduceDims = {1, 2, 3};
+
+        retval = mean_meansquare_dual_reduce_test<DeviceDualReduce,
+                                                  InDataType,
+                                                  OutDataType,
+                                                  AccDataType,
+                                                  Rank,
+                                                  NumReduceDim>(arg.n,
+                                                                arg.h,
+                                                                arg.w,
+                                                                arg.c,
+                                                                arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                reduceDims);
+    }
+    else
+    {
+        std::array<int, NumReduceDim> reduceDims = {1, 2, 3};
+
+        retval = mean_meansquare_dual_reduce_test<DeviceDualReduce,
+                                                  InDataType,
+                                                  OutDataType,
+                                                  AccDataType,
+                                                  Rank,
+                                                  NumReduceDim>(
+            600, 28, 28, 256, true, 2, true, reduceDims);
+    };
+
+    return (retval);
+}
diff --git a/example/33_multiple_reduce/dual_reduce_threadwise.cpp b/example/33_multiple_reduce/dual_reduce_threadwise.cpp
new file mode 100644
index 0000000000..7609edad35
--- /dev/null
+++ b/example/33_multiple_reduce/dual_reduce_threadwise.cpp
@@ -0,0 +1,93 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+
+#include "dual_reduce_common.hpp"
+
+using namespace ck;
+using namespace ck::tensor_operation::device;
+
+using InDataType       = ck::half_t;
+using OutDataType      = float;
+using OutDataTypeTuple = Tuple<OutDataType, OutDataType>;
+using AccDataType      = float;
+
+// for NHWC layer-norm calculation of mean and meansquare
+constexpr int Rank         = 4;
+constexpr int NumReduceDim = 3;
+
+constexpr bool PropagateNan = false;
+
+using DeviceDualReduce = DeviceMultipleReduceThreadWise<2,
+                                                        InDataType,
+                                                        AccDataType,
+                                                        OutDataTypeTuple,
+                                                        Rank,
+                                                        NumReduceDim,
+                                                        ReduceOperation,
+                                                        InElementwiseOperationTuple,
+                                                        AccElementwiseOperationTuple,
+                                                        PropagateNan,
+                                                        256,
+                                                        1,
+                                                        4,
+                                                        1, // InSrcVectorDim
+                                                        2,
+                                                        ck::Sequence<1, 1>>;
+
+int main(int argc, char* argv[])
+{
+    int retval = 0;
+
+    if(argc > 1)
+    {
+        SimpleAppArgs arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        std::array<int, NumReduceDim> reduceDims = {1, 2, 3};
+
+        retval = mean_meansquare_dual_reduce_test<DeviceDualReduce,
+                                                  InDataType,
+                                                  OutDataType,
+                                                  AccDataType,
+                                                  Rank,
+                                                  NumReduceDim>(arg.n,
+                                                                arg.h,
+                                                                arg.w,
+                                                                arg.c,
+                                                                arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                reduceDims);
+    }
+    else
+    {
+        std::array<int, NumReduceDim> reduceDims = {1, 2, 3};
+
+        retval = mean_meansquare_dual_reduce_test<DeviceDualReduce,
+                                                  InDataType,
+                                                  OutDataType,
+                                                  AccDataType,
+                                                  Rank,
+                                                  NumReduceDim>(
+            8000, 4, 4, 4, true, 2, true, reduceDims);
+    };
+
+    return (retval);
+}
diff --git a/example/34_batchnorm/CMakeLists.txt b/example/34_batchnorm/CMakeLists.txt
new file mode 100644
index 0000000000..60824c5f4d
--- /dev/null
+++ b/example/34_batchnorm/CMakeLists.txt
@@ -0,0 +1,4 @@
+add_example_executable(example_batchnorm_forward_training batchnorm_forward_training_nhwc.cpp)
+add_example_executable(example_batchnorm_forward_training_obsolete batchnorm_forward_training_nhwc_obsolete.cpp)
+add_example_executable(example_batchnorm_forward_inferring batchnorm_forward_inferring_nhwc.cpp)
+add_example_executable(example_batchnorm_backward batchnorm_backward_nhwc.cpp)
diff --git a/example/34_batchnorm/README.md b/example/34_batchnorm/README.md
new file mode 100644
index 0000000000..294e32b998
--- /dev/null
+++ b/example/34_batchnorm/README.md
@@ -0,0 +1,81 @@
+# Instructions for ```batchnorm nhwc``` Example
+
+## Run ```batchnorm forward nhwc```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+#arg1:  data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)
+#arg2: 1/0 to indicate whether to update the moving average and variance (0=no, 1=yes)
+#arg3: 1/0 to indicate whether to save result mean/invVariance (0=no, 1=yes)
+#arg4: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg5: time kernel (0=no, 1=yes) 
+./bin/example_batchnorm_forward -D 128,16,16,1024 -v 1 0 0 1 2 1
+```
+
+Result 
+```
+./bin/example_batchnorm_forward -D 128,16,16,1024 -v 1 0 0 1 2 1
+launch_and_time_kernel: grid_dim {64, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {120, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {120, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 2.08231 ms, 354.519 GB/s
+```
+
+Result
+```
+./bin/example_batchnorm_forward -D 128,16,16,1024 -v 1 0 1 0 2 0
+echo $?
+0
+```
+
+## Run ```batchnorm infer nhwc```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+#arg1:  data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)
+#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+./bin/example_batchnorm_infer -D 128,16,16,1024 -v 1 0 2 1
+```
+
+Result
+```
+./bin/example_batchnorm_infer -D 128,16,16,1024 -v 1 0 2 1
+launch_and_time_kernel: grid_dim {120, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 1.28235 ms, 523.329 GB/s
+```
+
+## Run ```batchnorm backward nhwc```
+```bash
+# -D <xxx> : input 4-d tensor lengths
+# -v <x> :   verification (0=no, 1=yes)
+Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)
+Arg2 -- 1/0 to indicate whether to use saved mean and invVariance
+Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
+Arg4 -- time kernel (0=no, 1=yes)
+Arg5: use multi-block welford (0=n0, 1=yes)
+./bin/example_batchnorm_backward -D 128,16,3,1024 -v 1 0 0 3 1 1
+```
+
+Result 
+```
+./bin/example_batchnorm_backward -D 128,16,3,1024 -v 1 0 0 3 1 1
+launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+launch_and_time_kernel: grid_dim {6144, 1, 1}, block_dim {256, 1, 1} 
+Warm up 1 time
+Start running 10 times...
+Perf: 0.411026 ms, 91.8702 GB/s
+```
diff --git a/example/34_batchnorm/batchnorm_backward_nhwc.cpp b/example/34_batchnorm/batchnorm_backward_nhwc.cpp
new file mode 100644
index 0000000000..3756310fd7
--- /dev/null
+++ b/example/34_batchnorm/batchnorm_backward_nhwc.cpp
@@ -0,0 +1,506 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <limits>
+#include <iostream>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp"
+
+static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class BatchNormBwdArg
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inOutLengths;
+
+    bool do_verification = false;
+
+    bool haveSavedMeanInvVar;
+
+    int data_type               = 0;
+    int init_method             = 3;
+    bool time_kernel            = false;
+    bool use_multiblock_welford = false;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        // clang-format off
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension lengths, must have 4 integers for nhwc" << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the result by comparing with the host-based batch-normalization" << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2 -- 1/0 to indicate whether to use saved mean and invVariance" << std::endl;
+        std::cout << "Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)" << std::endl;
+        std::cout << "Arg4 -- time kernel (0=no, 1=yes)" << std::endl;
+        std::cout << "Arg5: use multi-block welford (0=n0, 1=yes)" << std::endl;
+        // clang-format on
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inOutLengths = getTypeValuesFromString<size_t>(optarg);
+
+                if(inOutLengths.size() != 4)
+                    throw std::runtime_error(
+                        "NHWC tensor layout should have 4 length values specified!");
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 5 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type              = std::atoi(argv[optind++]);
+        haveSavedMeanInvVar    = std::atoi(argv[optind++]);
+        init_method            = std::atoi(argv[optind++]);
+        time_kernel            = static_cast<bool>(std::atoi(argv[optind++]));
+        use_multiblock_welford = static_cast<bool>(std::atoi(argv[optind]));
+
+        return (0);
+    };
+};
+
+using namespace ck;
+
+template <typename XDataType, typename AccDataType, bool UseMultiblockInK>
+bool bnorm_bwd_nhwc_test(bool do_verification,
+                         int init_method,
+                         bool time_kernel,
+                         const std::vector<size_t> inOutLengths,
+                         bool haveSavedMeanInvVar,
+                         double epsilon)
+{
+    // for NHWC BatchNorm calculation of mean and meansquare
+    constexpr index_t Rank         = 4;
+    constexpr index_t NumReduceDim = 3;
+
+    using ScaleDataType = XDataType;
+
+    const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
+
+    // input data of the batchnorm backward algorithm
+    Tensor<XDataType> x(inOutLengths);
+    Tensor<AccDataType> dy(inOutLengths);
+
+    Tensor<ScaleDataType> bnScale(scaleBiasMeanVarLengths);
+
+    Tensor<AccDataType> savedMean(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> savedInvVar(scaleBiasMeanVarLengths);
+    // savedVariance is only used for initializing savedInvVar
+    Tensor<AccDataType> savedVariance(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm backward algorithm
+    Tensor<AccDataType> dx_ref(inOutLengths);
+    Tensor<AccDataType> dx(inOutLengths);
+
+    Tensor<AccDataType> dscale(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> dbias(scaleBiasMeanVarLengths);
+
+    Tensor<AccDataType> dscale_ref(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> dbias_ref(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = dy.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = dscale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if(haveSavedMeanInvVar)
+    {
+        const float x_mean       = 0.0f;
+        const float x_stddev     = 1.0f;
+        const float noise_stddev = 0.0001f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+
+        // initialize the savedMean to be values with tiny variation to the mean of the x values
+        savedMean.GenerateTensorValue(GeneratorTensor_4<AccDataType>{x_mean, noise_stddev},
+                                      num_thread);
+
+        // initialize the variance to be values with tiny variation to the variance of the x values
+        savedVariance.GenerateTensorValue(
+            GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+
+        auto it_src       = savedVariance.mData.begin();
+        auto it_dst       = savedInvVar.mData.begin();
+        float tmp_epsilon = std::numeric_limits<float>::epsilon();
+
+        while(it_src != savedVariance.mData.end())
+        {
+            *it_dst = type_convert<AccDataType>(
+                1.0f / std::sqrtf(type_convert<float>(*it_src) + tmp_epsilon));
+
+            it_src++;
+            it_dst++;
+        };
+    }
+    else
+    {
+        const float x_mean   = 0.0f;
+        const float x_stddev = 1.0f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            dy.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_0<ScaleDataType>{}, num_thread);
+            break;
+        case 1:
+            dy.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_1<ScaleDataType>{1}, num_thread);
+            break;
+        case 2:
+            dy.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-2, 2}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_2<ScaleDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            dy.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-0.2f, 0.2f}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_3<ScaleDataType>{-0.5f, 0.5f}, num_thread);
+        }
+    };
+
+    // input data of the batchnorm backward algorithm
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem dy_dev(sizeof(AccDataType) * dy.mDesc.GetElementSpaceSize());
+
+    DeviceMem bnScale_dev(sizeof(ScaleDataType) * bnScale.mDesc.GetElementSpaceSize());
+
+    DeviceMem savedMean_dev(sizeof(AccDataType) * savedMean.mDesc.GetElementSpaceSize());
+    DeviceMem savedInvVar_dev(sizeof(AccDataType) * savedInvVar.mDesc.GetElementSpaceSize());
+
+    // output data of the batchnorm backward algorithm
+    DeviceMem dx_dev(sizeof(AccDataType) * dx.mDesc.GetElementSpaceSize());
+
+    DeviceMem dscale_dev(sizeof(AccDataType) * dscale.mDesc.GetElementSpaceSize());
+    DeviceMem dbias_dev(sizeof(AccDataType) * dbias.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    dy_dev.ToDevice(dy.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+
+    if(haveSavedMeanInvVar)
+    {
+        savedMean_dev.ToDevice(savedMean.mData.data());
+        savedInvVar_dev.ToDevice(savedInvVar.mData.data());
+    };
+
+    std::array<index_t, Rank> i_inOutLengths;
+    std::array<index_t, Rank> i_inOutStrides;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
+
+    std::copy(inOutLengths.begin(), inOutLengths.end(), i_inOutLengths.begin());
+    std::copy(inOutStrides.begin(), inOutStrides.end(), i_inOutStrides.begin());
+    std::copy(scaleBiasMeanVarLengths.begin(),
+              scaleBiasMeanVarLengths.end(),
+              i_scaleBiasMeanVarLengths.begin());
+    std::copy(scaleBiasMeanVarStrides.begin(),
+              scaleBiasMeanVarStrides.end(),
+              i_scaleBiasMeanVarStrides.begin());
+
+    using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+    using DeviceBatchNormBwdInstance =
+        ck::tensor_operation::device::DeviceBatchNormBwdImpl<XDataType,
+                                                             AccDataType,
+                                                             AccDataType,
+                                                             AccDataType,
+                                                             ScaleDataType, // ScaleDataType
+                                                             AccDataType,   // DscaleDbiasDataType
+                                                             AccDataType,   // MeanVarDataType
+                                                             PassThroughOp,
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             UseMultiblockInK,
+                                                             256,
+                                                             16,
+                                                             16,
+                                                             1,
+                                                             2,
+                                                             0,
+                                                             1,  // XSrcVectorSize
+                                                             1,  // DySrcVectorSize
+                                                             1,  // DxDstVectorSize
+                                                             1,  // ScaleSrcVectorSize
+                                                             1,  // DscaleDbiasDstVectorSize
+                                                             1>; // MeanVarSrcVectorSize
+
+    auto batchnorm_bwd = DeviceBatchNormBwdInstance{};
+
+    auto argument_ptr = batchnorm_bwd.MakeArgumentPointer(
+        i_inOutLengths,
+        i_inOutStrides,
+        i_inOutStrides,
+        i_inOutStrides,
+        {0, 1, 2},
+        i_scaleBiasMeanVarLengths,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        x_dev.GetDeviceBuffer(),
+        dy_dev.GetDeviceBuffer(),
+        bnScale_dev.GetDeviceBuffer(),
+        haveSavedMeanInvVar ? savedMean_dev.GetDeviceBuffer() : nullptr,
+        haveSavedMeanInvVar ? savedInvVar_dev.GetDeviceBuffer() : nullptr,
+        epsilon,
+        PassThroughOp{},
+        dx_dev.GetDeviceBuffer(),
+        dscale_dev.GetDeviceBuffer(),
+        dbias_dev.GetDeviceBuffer());
+
+    if(!batchnorm_bwd.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters seems not supported by the BatchNorm device instance, "
+                     "exiting!"
+                  << std::endl;
+        return (false);
+    };
+
+    size_t workspace_sz = batchnorm_bwd.GetWorkSpaceSize(argument_ptr.get());
+
+    DeviceMem workspace_dev(workspace_sz);
+
+    batchnorm_bwd.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+    auto invoker_ptr = batchnorm_bwd.MakeInvokerPointer();
+
+    if(time_kernel)
+    {
+        float avg_time   = 0.0f;
+        size_t num_bytes = 0;
+
+        size_t total_length = inOutLengths[0] * inOutLengths[1] * inOutLengths[2] * inOutLengths[3];
+        size_t invariant_length = inOutLengths[3];
+
+        avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        // inputing of x, dy, scale, outputing of dx, dscale, dbias
+        num_bytes +=
+            total_length * sizeof(XDataType) * 3 + invariant_length * sizeof(AccDataType) * 3;
+
+        // outputing of mean, inv-variance
+        num_bytes += haveSavedMeanInvVar ? invariant_length * sizeof(AccDataType) * 2 : 0;
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+    }
+    else
+        (void)invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        using ReferenceBatchNormBwdInstance =
+            ck::tensor_operation::host::ReferenceBatchNormBwd<XDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              ScaleDataType, // ScaleDataType
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              PassThroughOp,
+                                                              Rank,
+                                                              NumReduceDim>;
+
+        auto batchNormBwd_ref = ReferenceBatchNormBwdInstance{};
+
+        auto argument_ptr_ref = batchNormBwd_ref.MakeArgumentPointer(
+            i_inOutLengths,
+            i_inOutStrides,
+            i_inOutStrides,
+            i_inOutStrides,
+            {0, 1, 2},
+            i_scaleBiasMeanVarLengths,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            x.mData.data(),
+            dy.mData.data(),
+            bnScale.mData.data(),
+            haveSavedMeanInvVar ? savedMean.mData.data() : nullptr,
+            haveSavedMeanInvVar ? savedInvVar.mData.data() : nullptr,
+            epsilon,
+            PassThroughOp{},
+            dx_ref.mData.data(),
+            dscale_ref.mData.data(),
+            dbias_ref.mData.data());
+
+        if(!batchNormBwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout
+                << "The runtime parameters seems not supported by the device instance, exiting!"
+                << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = batchNormBwd_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+
+        dx_dev.FromDevice(dx.mData.data());
+        dscale_dev.FromDevice(dscale.data());
+        dbias_dev.FromDevice(dbias.data());
+
+        // clang-format off
+        pass = pass && ck::utils::check_err(dbias.mData, dbias_ref.mData, "dBias result:", 2e-4, 2e-4);
+        pass = pass && ck::utils::check_err(dscale.mData, dscale_ref.mData, "dScale result:", 2e-4, 2e-4);
+        pass = pass && ck::utils::check_err(dx.mData, dx_ref.mData, "dx result:");
+        // clang-format on
+    };
+
+    return (pass);
+};
+
+static const double epsilon = std::numeric_limits<float>::epsilon();
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        BatchNormBwdArg arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<ck::half_t, float, true>(arg.do_verification,
+                                                                    arg.init_method,
+                                                                    arg.time_kernel,
+                                                                    arg.inOutLengths,
+                                                                    arg.haveSavedMeanInvVar,
+                                                                    epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<ck::half_t, float, false>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.haveSavedMeanInvVar,
+                                                                     epsilon);
+        }
+        else if(arg.data_type == 1)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<float, float, true>(arg.do_verification,
+                                                               arg.init_method,
+                                                               arg.time_kernel,
+                                                               arg.inOutLengths,
+                                                               arg.haveSavedMeanInvVar,
+                                                               epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<float, float, false>(arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                arg.inOutLengths,
+                                                                arg.haveSavedMeanInvVar,
+                                                                epsilon);
+        }
+        else if(arg.data_type == 5)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<ck::bhalf_t, float, true>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.haveSavedMeanInvVar,
+                                                                     epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<ck::bhalf_t, float, false>(arg.do_verification,
+                                                                      arg.init_method,
+                                                                      arg.time_kernel,
+                                                                      arg.inOutLengths,
+                                                                      arg.haveSavedMeanInvVar,
+                                                                      epsilon);
+        }
+        else if(arg.data_type == 6)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_bwd_nhwc_test<double, double, true>(arg.do_verification,
+                                                                 arg.init_method,
+                                                                 arg.time_kernel,
+                                                                 arg.inOutLengths,
+                                                                 arg.haveSavedMeanInvVar,
+                                                                 epsilon);
+            else
+                pass = bnorm_bwd_nhwc_test<double, double, false>(arg.do_verification,
+                                                                  arg.init_method,
+                                                                  arg.time_kernel,
+                                                                  arg.inOutLengths,
+                                                                  arg.haveSavedMeanInvVar,
+                                                                  epsilon);
+        }
+    }
+    else
+    {
+        pass = bnorm_bwd_nhwc_test<ck::half_t, float, true>(true,
+                                                            3,
+                                                            false, // don't time kernel
+                                                            {128, 16, 6, 512},
+                                                            false,
+                                                            epsilon);
+
+        pass = pass && bnorm_bwd_nhwc_test<ck::half_t, float, false>(true,
+                                                                     3,
+                                                                     false, // don't time kernel
+                                                                     {128, 16, 3, 1024},
+                                                                     false,
+                                                                     epsilon);
+    };
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/34_batchnorm/batchnorm_common.hpp b/example/34_batchnorm/batchnorm_common.hpp
new file mode 100644
index 0000000000..a1b8d253bf
--- /dev/null
+++ b/example/34_batchnorm/batchnorm_common.hpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cassert>
+#include <vector>
+#include <array>
+#include <type_traits>
+
+#include "ck/utility/data_type.hpp"
+
+struct NormalizeInInfer
+{
+    NormalizeInInfer(double epsilon = 1e-4) : epsilon_(epsilon) {}
+
+    template <typename T1, typename T2, typename T3, typename T4>
+    __host__ __device__ constexpr void operator()(T1& y,
+                                                  const T1& x,
+                                                  const T2& mean,
+                                                  const T2& variance,
+                                                  const T3& gamma,
+                                                  const T4& beta) const
+    {
+        static_assert(std::is_same<T2, float>::value || std::is_same<T2, double>::value,
+                      "Data type is not supported by this operation!");
+
+        using ck::type_convert;
+        using ck::math::sqrt;
+
+        T2 tmp_x, tmp_y;
+
+        tmp_x = type_convert<T2>(x);
+
+        tmp_y = ((tmp_x - mean) / sqrt(variance + type_convert<T2>(epsilon_))) *
+                    type_convert<T2>(gamma) +
+                type_convert<T2>(beta);
+        y = type_convert<T1>(tmp_y);
+    };
+
+    double epsilon_;
+};
+
+template <int Rank, int NumReduceDim>
+static inline std::array<int, Rank - NumReduceDim>
+get_invariant_dims(const std::array<int, NumReduceDim>& reduceDims)
+{
+    int reduceFlag = 0;
+
+    // flag the bits for the reduceDims
+    for(int i = 0; i < NumReduceDim; i++)
+    {
+        reduceFlag |= 1 << reduceDims[i];
+    };
+
+    std::array<int, Rank - NumReduceDim> invariantDims;
+
+    // collect invariant dimensions
+    int dim = 0;
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) == 0)
+        {
+            invariantDims[dim] = i;
+            dim++;
+        };
+
+    return invariantDims;
+};
diff --git a/example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp b/example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp
new file mode 100644
index 0000000000..6a8002025a
--- /dev/null
+++ b/example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp
@@ -0,0 +1,366 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <limits>
+#include <iostream>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp"
+
+#include "batchnorm_infer_impl.hpp"
+
+static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class BatchNormInferArg
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inOutLengths;
+
+    bool do_verification = false;
+
+    int data_type    = 0;
+    int init_method  = 2;
+    bool time_kernel = false;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension "
+                     "lengths, must have 4 integers for nhwc"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the batch-normalization "
+                     "result by "
+                     "comparing with the host-based batch-normalization"
+                  << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2: init method used for bnScale and bnBias (0=no init, 1=single integer "
+                     "value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg3: time kernel (0=no, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inOutLengths = getTypeValuesFromString<size_t>(optarg);
+
+                if(inOutLengths.size() != 4)
+                    throw std::runtime_error(
+                        "NHWC tensor layout should have 4 length values specified!");
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 3 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type   = std::atoi(argv[optind++]);
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return (-1);
+
+        return (0);
+    };
+};
+
+using namespace ck;
+
+template <typename InOutDataType, typename AccDataType>
+bool bnorm_infer_nhwc_test(bool do_verification,
+                           int init_method,
+                           bool time_kernel,
+                           const std::vector<size_t> inOutLengths,
+                           double epsilon)
+{
+    // for NHWC BatchNorm calculation of mean and meansquare
+    constexpr int Rank         = 4;
+    constexpr int NumReduceDim = 3;
+
+    // when using lengths[] to create a tensor, lengths[0] is the length of highest dimension
+    // eg. N of NHWC, so lengths[3] is the dimension C length of NHWC
+    const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
+
+    // input data of the batchnorm forward algorithm
+    Tensor<InOutDataType> x(inOutLengths);
+    Tensor<AccDataType> bnScale(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> bnBias(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm forward algorithm
+    Tensor<InOutDataType> y_ref(inOutLengths);
+    Tensor<InOutDataType> y(inOutLengths);
+
+    Tensor<AccDataType> estimatedMean(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> estimatedVariance(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = x.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if constexpr(std::is_same<InOutDataType, int8_t>::value)
+    {
+        x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+
+        const float x_mean       = 0.0f;
+        const float x_stddev     = 2.5f;
+        const float noise_stddev = 0.0001f;
+
+        estimatedMean.GenerateTensorValue(GeneratorTensor_4<AccDataType>{x_mean, noise_stddev},
+                                          num_thread);
+
+        estimatedVariance.GenerateTensorValue(
+            GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+    }
+    else
+    {
+        const float x_mean       = 0.0f;
+        const float x_stddev     = 1.0f;
+        const float noise_stddev = 0.0001f;
+
+        x.GenerateTensorValue(GeneratorTensor_4<InOutDataType>{x_mean, x_stddev}, num_thread);
+
+        // initialize the savedMean to be values with tiny variation to the mean of the x values
+        estimatedMean.GenerateTensorValue(GeneratorTensor_4<AccDataType>{x_mean, noise_stddev},
+                                          num_thread);
+
+        // initialize the variance to be values with tiny variation to the variance of the x values
+        estimatedVariance.GenerateTensorValue(
+            GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            bnScale.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            break;
+        case 1:
+            bnScale.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0}, num_thread);
+            break;
+        case 2:
+            bnScale.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            bnScale.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
+        }
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem x_dev(sizeof(InOutDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(InOutDataType) * y.mDesc.GetElementSpaceSize());
+    DeviceMem bnScale_dev(sizeof(AccDataType) * bnScale.mDesc.GetElementSpaceSize());
+    DeviceMem bnBias_dev(sizeof(AccDataType) * bnBias.mDesc.GetElementSpaceSize());
+
+    // mean_dev or resultSaveMean_dev
+    DeviceMem estimatedMean_dev(sizeof(AccDataType) * estimatedMean.mDesc.GetElementSpaceSize());
+    // meansquare_dev or resultSaveInvVariance_dev
+    DeviceMem estimatedVariance_dev(sizeof(AccDataType) *
+                                    estimatedVariance.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+    bnBias_dev.ToDevice(bnBias.mData.data());
+    estimatedMean_dev.ToDevice(estimatedMean.mData.data());
+    estimatedVariance_dev.ToDevice(estimatedVariance.mData.data());
+
+    using ck::index_t;
+
+    std::array<index_t, Rank> i_inOutLengths;
+    std::array<index_t, Rank> i_inOutStrides;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
+
+    ck::ranges::copy(inOutLengths, i_inOutLengths.begin());
+    ck::ranges::copy(inOutStrides, i_inOutStrides.begin());
+    ck::ranges::copy(scaleBiasMeanVarLengths, i_scaleBiasMeanVarLengths.begin());
+    ck::ranges::copy(scaleBiasMeanVarStrides, i_scaleBiasMeanVarStrides.begin());
+
+    int result = 0;
+
+    result = bnorm_infer<InOutDataType,
+                         InOutDataType,
+                         AccDataType,
+                         AccDataType,
+                         AccDataType,
+                         AccDataType,
+                         Rank,
+                         NumReduceDim,
+                         false>(time_kernel,
+                                {0, 1, 2},
+                                i_inOutLengths,
+                                i_inOutStrides,
+                                i_inOutStrides,
+                                i_scaleBiasMeanVarLengths,
+                                i_scaleBiasMeanVarStrides,
+                                i_scaleBiasMeanVarStrides,
+                                i_scaleBiasMeanVarStrides,
+                                x_dev.GetDeviceBuffer(),
+                                bnScale_dev.GetDeviceBuffer(),
+                                bnBias_dev.GetDeviceBuffer(),
+                                epsilon,
+                                estimatedMean_dev.GetDeviceBuffer(),
+                                estimatedVariance_dev.GetDeviceBuffer(),
+                                y_dev.GetDeviceBuffer());
+
+    if(result < 0)
+        return (false);
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+        using ReferenceBatchNormInferInstance =
+            ck::tensor_operation::host::ReferenceBatchNormInfer<InOutDataType,
+                                                                InOutDataType,
+                                                                AccDataType,
+                                                                AccDataType,
+                                                                AccDataType,
+                                                                AccDataType,
+                                                                PassThroughOp,
+                                                                Rank,
+                                                                NumReduceDim>;
+        auto batchNormInfer_ref = ReferenceBatchNormInferInstance{};
+
+        auto argument_ptr_ref =
+            batchNormInfer_ref.MakeArgumentPointer(i_inOutLengths,
+                                                   i_inOutStrides,
+                                                   i_inOutStrides,
+                                                   {0, 1, 2},
+                                                   i_scaleBiasMeanVarLengths,
+                                                   i_scaleBiasMeanVarStrides,
+                                                   i_scaleBiasMeanVarStrides,
+                                                   i_scaleBiasMeanVarStrides,
+                                                   x.mData.data(),
+                                                   bnScale.mData.data(),
+                                                   bnBias.mData.data(),
+                                                   epsilon,
+                                                   PassThroughOp{},
+                                                   estimatedMean.mData.data(),
+                                                   estimatedVariance.mData.data(),
+                                                   y_ref.mData.data());
+
+        if(!batchNormInfer_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout
+                << "The runtime parameters seems not supported by the BatchNorm instance, exiting!"
+                << std::endl;
+            return (-2);
+        };
+
+        auto invoker_ptr_ref = batchNormInfer_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+
+        y_dev.FromDevice(y.mData.data());
+        pass = pass && ck::utils::check_err(y, y_ref);
+    };
+
+    return (pass);
+};
+
+static const double epsilon = std::numeric_limits<float>::epsilon();
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        BatchNormInferArg arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            pass = bnorm_infer_nhwc_test<ck::half_t, float>(
+                arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
+        }
+        else if(arg.data_type == 1)
+        {
+            pass = bnorm_infer_nhwc_test<float, float>(
+                arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
+        }
+        else if(arg.data_type == 3)
+        {
+            pass = bnorm_infer_nhwc_test<int8_t, float>(
+                arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
+        }
+        else if(arg.data_type == 5)
+        {
+            pass = bnorm_infer_nhwc_test<ck::bhalf_t, float>(
+                arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
+        }
+        else if(arg.data_type == 6)
+        {
+            pass = bnorm_infer_nhwc_test<double, double>(
+                arg.do_verification, arg.init_method, arg.time_kernel, arg.inOutLengths, epsilon);
+        };
+    }
+    else
+    {
+        pass = bnorm_infer_nhwc_test<ck::half_t, float>(true,
+                                                        2,
+                                                        false, // don't time kernel
+                                                        {128, 16, 16, 1024},
+                                                        epsilon);
+    };
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/34_batchnorm/batchnorm_forward_training_nhwc.cpp b/example/34_batchnorm/batchnorm_forward_training_nhwc.cpp
new file mode 100644
index 0000000000..b27358fd9d
--- /dev/null
+++ b/example/34_batchnorm/batchnorm_forward_training_nhwc.cpp
@@ -0,0 +1,598 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <limits>
+#include <iostream>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class BatchNormFwdArg
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inOutLengths;
+
+    bool do_verification = false;
+
+    bool updateMovingAverage;
+    bool saveMeanAndInvVariance;
+
+    int data_type               = 0;
+    int init_method             = 2;
+    bool time_kernel            = false;
+    bool use_multiblock_welford = false;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension "
+                     "lengths, must have 4 integers for nhwc"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the batch-normalization "
+                     "result by "
+                     "comparing with the host-based batch-normalization"
+                  << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2: 1/0 to indicate whether to update the moving average and variance "
+                     "(0=no, 1=yes)"
+                  << std::endl;
+        std::cout << "Arg3: 1/0 to indicate whether to save the calculated mean and invVariance "
+                     "(0=no, 1=yes)"
+                  << std::endl;
+        std::cout << "Arg4: init method used for bnScale and bnBias (0=no init, 1=single integer "
+                     "value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg5: time kernel (0=no, 1=yes)" << std::endl;
+        std::cout << "Arg6: use multi-block welford (0=n0, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inOutLengths = getTypeValuesFromString<size_t>(optarg);
+
+                if(inOutLengths.size() != 4)
+                    throw std::runtime_error(
+                        "NHWC tensor layout should have 4 length values specified!");
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 6 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type              = std::atoi(argv[optind++]);
+        updateMovingAverage    = std::atoi(argv[optind++]);
+        saveMeanAndInvVariance = std::atoi(argv[optind++]);
+        init_method            = std::atoi(argv[optind++]);
+        time_kernel            = static_cast<bool>(std::atoi(argv[optind++]));
+        use_multiblock_welford = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return (-1);
+
+        return (0);
+    };
+};
+
+using namespace ck;
+
+template <typename InOutDataType, typename AccDataType, bool UseMultiblockInK>
+bool bnorm_fwd_nhwc_test(bool do_verification,
+                         int init_method,
+                         bool time_kernel,
+                         const std::vector<size_t> inOutLengths,
+                         bool updateMovingAverage,
+                         bool saveMeanAndInvVariance,
+                         double averageFactor,
+                         double epsilon)
+{
+    // for NHWC BatchNorm calculation of mean and meansquare
+    constexpr int Rank         = 4;
+    constexpr int NumReduceDim = 3;
+
+    // when using lengths[] to create a tensor, lengths[0] is the length of highest dimension
+    // eg. N of NHWC, so lengths[3] is the dimension C length of NHWC
+    const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
+
+    // input data of the batchnorm forward algorithm
+    Tensor<InOutDataType> x(inOutLengths);
+    Tensor<AccDataType> bnScale(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> bnBias(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm forward algorithm
+    Tensor<InOutDataType> y_ref(inOutLengths);
+    Tensor<InOutDataType> y(inOutLengths);
+
+    Tensor<AccDataType> resultSaveMean_ref(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> resultSaveInvVariance_ref(scaleBiasMeanVarLengths);
+
+    Tensor<AccDataType> resultRunningMean_ref(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> resultRunningVariance_ref(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = x.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if(updateMovingAverage)
+    {
+        if constexpr(std::is_same<InOutDataType, int8_t>::value)
+        {
+            x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+
+            const float x_mean       = 0.0f;
+            const float x_stddev     = 2.5f;
+            const float noise_stddev = 0.04f;
+
+            resultRunningMean_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_mean, noise_stddev}, num_thread);
+
+            resultRunningVariance_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+        }
+        else
+        {
+            const float x_mean       = 0.0f;
+            const float x_stddev     = 1.0f;
+            const float noise_stddev = 0.04f;
+
+            // input data in normal distribution
+            x.GenerateTensorValue(GeneratorTensor_4<InOutDataType>{x_mean, x_stddev}, num_thread);
+
+            // initialize the runningMean to be values with tiny variation to the mean of the x
+            // values
+            resultRunningMean_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_mean, noise_stddev}, num_thread);
+
+            // initialize the runningVariance to be values with tiny variation to the variance of
+            // the x values
+            resultRunningVariance_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+        };
+    }
+    else
+    {
+        if constexpr(std::is_same<InOutDataType, int8_t>::value)
+            x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+        else
+            x.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0f, 5.0f}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            bnScale.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            break;
+        case 1:
+            bnScale.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0}, num_thread);
+            break;
+        case 2:
+            bnScale.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            bnScale.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
+        }
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem x_dev(sizeof(InOutDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(InOutDataType) * y.mDesc.GetElementSpaceSize());
+    DeviceMem bnScale_dev(sizeof(AccDataType) * bnScale.mDesc.GetElementSpaceSize());
+    DeviceMem bnBias_dev(sizeof(AccDataType) * bnBias.mDesc.GetElementSpaceSize());
+
+    // mean_dev or resultSaveMean_dev
+    DeviceMem resultSaveMean_dev(sizeof(AccDataType) *
+                                 resultSaveMean_ref.mDesc.GetElementSpaceSize());
+    // meansquare_dev or resultSaveInvVariance_dev
+    DeviceMem resultSaveInvVariance_dev(sizeof(AccDataType) *
+                                        resultSaveInvVariance_ref.mDesc.GetElementSpaceSize());
+    // resultRunningMean_dev
+    DeviceMem resultRunningMean_dev(sizeof(AccDataType) *
+                                    resultRunningMean_ref.mDesc.GetElementSpaceSize());
+    // resultRunningVariance_dev
+    DeviceMem resultRunningVariance_dev(sizeof(AccDataType) *
+                                        resultRunningVariance_ref.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+    bnBias_dev.ToDevice(bnBias.mData.data());
+
+    if(updateMovingAverage)
+    {
+        resultRunningMean_dev.ToDevice(resultRunningMean_ref.mData.data());
+        resultRunningVariance_dev.ToDevice(resultRunningVariance_ref.mData.data());
+    };
+
+    std::array<index_t, Rank> i_inOutLengths;
+    std::array<index_t, Rank> i_inOutStrides;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
+
+    ck::ranges::copy(inOutLengths, i_inOutLengths.begin());
+    ck::ranges::copy(inOutStrides, i_inOutStrides.begin());
+    ck::ranges::copy(scaleBiasMeanVarLengths, i_scaleBiasMeanVarLengths.begin());
+    ck::ranges::copy(scaleBiasMeanVarStrides, i_scaleBiasMeanVarStrides.begin());
+
+    using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+    using DeviceBatchNormFwdInstance =
+        ck::tensor_operation::device::DeviceBatchNormFwdImpl<InOutDataType,
+                                                             InOutDataType,
+                                                             AccDataType,
+                                                             AccDataType,   // ScaleDataType
+                                                             AccDataType,   // BiasDataType
+                                                             AccDataType,   // MeanVarDataType
+                                                             PassThroughOp, // YElementwiseOp
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             UseMultiblockInK,
+                                                             256,
+                                                             16,
+                                                             16,
+                                                             1,
+                                                             2,
+                                                             0,
+                                                             1,
+                                                             1,
+                                                             1,
+                                                             1,
+                                                             1>;
+
+    auto batchnorm_fwd = DeviceBatchNormFwdInstance{};
+
+    auto argument_ptr = batchnorm_fwd.MakeArgumentPointer(
+        i_inOutLengths,
+        i_inOutStrides,
+        i_inOutStrides,
+        {0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
+        i_scaleBiasMeanVarLengths,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        x_dev.GetDeviceBuffer(),
+        bnScale_dev.GetDeviceBuffer(),
+        bnBias_dev.GetDeviceBuffer(),
+        epsilon,
+        PassThroughOp{},
+        y_dev.GetDeviceBuffer(),
+        saveMeanAndInvVariance ? resultSaveMean_dev.GetDeviceBuffer() : nullptr,
+        saveMeanAndInvVariance ? resultSaveInvVariance_dev.GetDeviceBuffer() : nullptr,
+        averageFactor,
+        updateMovingAverage ? resultRunningMean_dev.GetDeviceBuffer() : nullptr,
+        updateMovingAverage ? resultRunningVariance_dev.GetDeviceBuffer() : nullptr);
+
+    if(!batchnorm_fwd.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters seems not supported by the BatchNorm device instance, "
+                     "exiting!"
+                  << std::endl;
+        return (false);
+    };
+
+    size_t workspace_sz = batchnorm_fwd.GetWorkSpaceSize(argument_ptr.get());
+
+    DeviceMem workspace_dev(workspace_sz);
+
+    batchnorm_fwd.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+    auto invoker_ptr = batchnorm_fwd.MakeInvokerPointer();
+
+    if(time_kernel)
+    {
+        float avg_time   = 0.0f;
+        size_t num_bytes = 0;
+
+        size_t total_length = inOutLengths[0] * inOutLengths[1] * inOutLengths[2] * inOutLengths[3];
+        size_t invariant_length = inOutLengths[3];
+
+        avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        // inputing of x, scale, bias, outputing of y
+        num_bytes +=
+            total_length * sizeof(InOutDataType) * 2 + invariant_length * sizeof(AccDataType) * 2;
+
+        // outputing of mean, inv-variance
+        num_bytes += saveMeanAndInvVariance ? invariant_length * sizeof(AccDataType) * 2 : 0;
+
+        // updating of moving mean, variance
+        num_bytes += updateMovingAverage ? invariant_length * sizeof(AccDataType) * 4 : 0;
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+    }
+    else
+        (void)invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+
+        using ReferenceBatchNormFwdInstance =
+            ck::tensor_operation::host::ReferenceBatchNormFwd<InOutDataType,
+                                                              InOutDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              PassThroughOp,
+                                                              Rank,
+                                                              NumReduceDim>;
+
+        auto batchNormFwd_ref = ReferenceBatchNormFwdInstance{};
+
+        auto argument_ptr_ref = batchNormFwd_ref.MakeArgumentPointer(
+            i_inOutLengths,
+            i_inOutStrides,
+            i_inOutStrides,
+            {0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
+            i_scaleBiasMeanVarLengths,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            x.mData.data(),
+            bnScale.mData.data(),
+            bnBias.mData.data(),
+            epsilon,
+            PassThroughOp{},
+            y_ref.mData.data(),
+            saveMeanAndInvVariance ? resultSaveMean_ref.mData.data() : nullptr,
+            saveMeanAndInvVariance ? resultSaveInvVariance_ref.mData.data() : nullptr,
+            averageFactor,
+            updateMovingAverage ? resultRunningMean_ref.mData.data() : nullptr,
+            updateMovingAverage ? resultRunningVariance_ref.mData.data() : nullptr);
+
+        if(!batchNormFwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters seems not supported by the BatchNorm reference "
+                         "instance, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = batchNormFwd_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+
+        y_dev.FromDevice(y.mData.data());
+        pass = pass && ck::utils::check_err(y, y_ref, "Incorrect normalized output values");
+
+        if(updateMovingAverage)
+        {
+            Tensor<AccDataType> resultRunningMean(scaleBiasMeanVarLengths);
+            Tensor<AccDataType> resultRunningVariance(scaleBiasMeanVarLengths);
+
+            resultRunningMean_dev.FromDevice(resultRunningMean.mData.data());
+            resultRunningVariance_dev.FromDevice(resultRunningVariance.mData.data());
+
+            pass = pass && ck::utils::check_err(resultRunningMean,
+                                                resultRunningMean_ref,
+                                                "Incorrect running mean values");
+            pass = pass && ck::utils::check_err(resultRunningVariance,
+                                                resultRunningVariance_ref,
+                                                "Incorrect running variance values");
+        };
+
+        if(saveMeanAndInvVariance)
+        {
+            using ck::host_common::dumpBufferToFile;
+
+            Tensor<AccDataType> resultSaveMean(scaleBiasMeanVarLengths);
+            Tensor<AccDataType> resultSaveInvVariance(scaleBiasMeanVarLengths);
+
+            resultSaveMean_dev.FromDevice(resultSaveMean.mData.data());
+            resultSaveInvVariance_dev.FromDevice(resultSaveInvVariance.mData.data());
+
+            pass = pass && ck::utils::check_err(
+                               resultSaveMean, resultSaveMean_ref, "Incorrect saved mean values");
+            pass = pass && ck::utils::check_err(resultSaveInvVariance,
+                                                resultSaveInvVariance_ref,
+                                                "Incorrect saved invvariance values");
+        };
+    };
+
+    return (pass);
+};
+
+const double epsilon              = std::numeric_limits<float>::epsilon();
+static const double averageFactor = 0.1;
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        BatchNormFwdArg arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(arg.do_verification,
+                                                                    arg.init_method,
+                                                                    arg.time_kernel,
+                                                                    arg.inOutLengths,
+                                                                    arg.updateMovingAverage,
+                                                                    arg.saveMeanAndInvVariance,
+                                                                    averageFactor,
+                                                                    epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<ck::half_t, float, false>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.updateMovingAverage,
+                                                                     arg.saveMeanAndInvVariance,
+                                                                     averageFactor,
+                                                                     epsilon);
+        }
+        else if(arg.data_type == 1)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<float, float, true>(arg.do_verification,
+                                                               arg.init_method,
+                                                               arg.time_kernel,
+                                                               arg.inOutLengths,
+                                                               arg.updateMovingAverage,
+                                                               arg.saveMeanAndInvVariance,
+                                                               averageFactor,
+                                                               epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<float, float, false>(arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                arg.inOutLengths,
+                                                                arg.updateMovingAverage,
+                                                                arg.saveMeanAndInvVariance,
+                                                                averageFactor,
+                                                                epsilon);
+        }
+        else if(arg.data_type == 3)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<int8_t, float, true>(arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                arg.inOutLengths,
+                                                                arg.updateMovingAverage,
+                                                                arg.saveMeanAndInvVariance,
+                                                                averageFactor,
+                                                                epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<int8_t, float, false>(arg.do_verification,
+                                                                 arg.init_method,
+                                                                 arg.time_kernel,
+                                                                 arg.inOutLengths,
+                                                                 arg.updateMovingAverage,
+                                                                 arg.saveMeanAndInvVariance,
+                                                                 averageFactor,
+                                                                 epsilon);
+        }
+        else if(arg.data_type == 5)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, true>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.updateMovingAverage,
+                                                                     arg.saveMeanAndInvVariance,
+                                                                     averageFactor,
+                                                                     epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, false>(arg.do_verification,
+                                                                      arg.init_method,
+                                                                      arg.time_kernel,
+                                                                      arg.inOutLengths,
+                                                                      arg.updateMovingAverage,
+                                                                      arg.saveMeanAndInvVariance,
+                                                                      averageFactor,
+                                                                      epsilon);
+        }
+        else if(arg.data_type == 6)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<double, double, true>(arg.do_verification,
+                                                                 arg.init_method,
+                                                                 arg.time_kernel,
+                                                                 arg.inOutLengths,
+                                                                 arg.updateMovingAverage,
+                                                                 arg.saveMeanAndInvVariance,
+                                                                 averageFactor,
+                                                                 epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<double, double, false>(arg.do_verification,
+                                                                  arg.init_method,
+                                                                  arg.time_kernel,
+                                                                  arg.inOutLengths,
+                                                                  arg.updateMovingAverage,
+                                                                  arg.saveMeanAndInvVariance,
+                                                                  averageFactor,
+                                                                  epsilon);
+        }
+    }
+    else
+    {
+        pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(true,
+                                                            2,
+                                                            false, // don't time kernel
+                                                            {128, 16, 6, 512},
+                                                            true,
+                                                            true,
+                                                            averageFactor,
+                                                            epsilon);
+
+        pass = pass && bnorm_fwd_nhwc_test<ck::half_t, float, false>(true,
+                                                                     2,
+                                                                     false, // don't time kernel
+                                                                     {128, 16, 3, 1024},
+                                                                     true,
+                                                                     true,
+                                                                     averageFactor,
+                                                                     epsilon);
+    };
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/34_batchnorm/batchnorm_forward_training_nhwc_obsolete.cpp b/example/34_batchnorm/batchnorm_forward_training_nhwc_obsolete.cpp
new file mode 100644
index 0000000000..ffb9f4b584
--- /dev/null
+++ b/example/34_batchnorm/batchnorm_forward_training_nhwc_obsolete.cpp
@@ -0,0 +1,598 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <limits>
+#include <iostream>
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl_obsolete.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+static struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class BatchNormFwdArg
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inOutLengths;
+
+    bool do_verification = false;
+
+    bool updateMovingAverage;
+    bool saveMeanAndInvVariance;
+
+    int data_type               = 0;
+    int init_method             = 2;
+    bool time_kernel            = false;
+    bool use_multiblock_welford = false;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension "
+                     "lengths, must have 4 integers for nhwc"
+                  << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the batch-normalization "
+                     "result by "
+                     "comparing with the host-based batch-normalization"
+                  << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2: 1/0 to indicate whether to update the moving average and variance "
+                     "(0=no, 1=yes)"
+                  << std::endl;
+        std::cout << "Arg3: 1/0 to indicate whether to save the calculated mean and invVariance "
+                     "(0=no, 1=yes)"
+                  << std::endl;
+        std::cout << "Arg4: init method used for bnScale and bnBias (0=no init, 1=single integer "
+                     "value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+        std::cout << "Arg5: time kernel (0=no, 1=yes)" << std::endl;
+        std::cout << "Arg6: use multi-block welford (0=n0, 1=yes)" << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:v:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inOutLengths = getTypeValuesFromString<size_t>(optarg);
+
+                if(inOutLengths.size() != 4)
+                    throw std::runtime_error(
+                        "NHWC tensor layout should have 4 length values specified!");
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 6 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type              = std::atoi(argv[optind++]);
+        updateMovingAverage    = std::atoi(argv[optind++]);
+        saveMeanAndInvVariance = std::atoi(argv[optind++]);
+        init_method            = std::atoi(argv[optind++]);
+        time_kernel            = static_cast<bool>(std::atoi(argv[optind++]));
+        use_multiblock_welford = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return (-1);
+
+        return (0);
+    };
+};
+
+using namespace ck;
+
+template <typename InOutDataType, typename AccDataType, bool UseMultiblockInK>
+bool bnorm_fwd_nhwc_test(bool do_verification,
+                         int init_method,
+                         bool time_kernel,
+                         const std::vector<size_t> inOutLengths,
+                         bool updateMovingAverage,
+                         bool saveMeanAndInvVariance,
+                         double averageFactor,
+                         double epsilon)
+{
+    // for NHWC BatchNorm calculation of mean and meansquare
+    constexpr int Rank         = 4;
+    constexpr int NumReduceDim = 3;
+
+    // when using lengths[] to create a tensor, lengths[0] is the length of highest dimension
+    // eg. N of NHWC, so lengths[3] is the dimension C length of NHWC
+    const std::vector<size_t> scaleBiasMeanVarLengths = {inOutLengths[3]};
+
+    // input data of the batchnorm forward algorithm
+    Tensor<InOutDataType> x(inOutLengths);
+    Tensor<AccDataType> bnScale(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> bnBias(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm forward algorithm
+    Tensor<InOutDataType> y_ref(inOutLengths);
+    Tensor<InOutDataType> y(inOutLengths);
+
+    Tensor<AccDataType> resultSaveMean_ref(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> resultSaveInvVariance_ref(scaleBiasMeanVarLengths);
+
+    Tensor<AccDataType> resultRunningMean_ref(scaleBiasMeanVarLengths);
+    Tensor<AccDataType> resultRunningVariance_ref(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = x.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if(updateMovingAverage)
+    {
+        if constexpr(std::is_same<InOutDataType, int8_t>::value)
+        {
+            x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+
+            const float x_mean       = 0.0f;
+            const float x_stddev     = 2.5f;
+            const float noise_stddev = 0.04f;
+
+            resultRunningMean_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_mean, noise_stddev}, num_thread);
+
+            resultRunningVariance_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+        }
+        else
+        {
+            const float x_mean       = 0.0f;
+            const float x_stddev     = 1.0f;
+            const float noise_stddev = 0.04f;
+
+            // input data in normal distribution
+            x.GenerateTensorValue(GeneratorTensor_4<InOutDataType>{x_mean, x_stddev}, num_thread);
+
+            // initialize the runningMean to be values with tiny variation to the mean of the x
+            // values
+            resultRunningMean_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_mean, noise_stddev}, num_thread);
+
+            // initialize the runningVariance to be values with tiny variation to the variance of
+            // the x values
+            resultRunningVariance_ref.GenerateTensorValue(
+                GeneratorTensor_4<AccDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+        };
+    }
+    else
+    {
+        if constexpr(std::is_same<InOutDataType, int8_t>::value)
+            x.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
+        else
+            x.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0f, 5.0f}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            bnScale.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_0<AccDataType>{}, num_thread);
+            break;
+        case 1:
+            bnScale.GenerateTensorValue(GeneratorTensor_1<AccDataType>{1}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0}, num_thread);
+            break;
+        case 2:
+            bnScale.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_2<AccDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            bnScale.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_3<AccDataType>{-5.0f, 5.0f}, num_thread);
+        }
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem x_dev(sizeof(InOutDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(InOutDataType) * y.mDesc.GetElementSpaceSize());
+    DeviceMem bnScale_dev(sizeof(AccDataType) * bnScale.mDesc.GetElementSpaceSize());
+    DeviceMem bnBias_dev(sizeof(AccDataType) * bnBias.mDesc.GetElementSpaceSize());
+
+    // mean_dev or resultSaveMean_dev
+    DeviceMem resultSaveMean_dev(sizeof(AccDataType) *
+                                 resultSaveMean_ref.mDesc.GetElementSpaceSize());
+    // meansquare_dev or resultSaveInvVariance_dev
+    DeviceMem resultSaveInvVariance_dev(sizeof(AccDataType) *
+                                        resultSaveInvVariance_ref.mDesc.GetElementSpaceSize());
+    // resultRunningMean_dev
+    DeviceMem resultRunningMean_dev(sizeof(AccDataType) *
+                                    resultRunningMean_ref.mDesc.GetElementSpaceSize());
+    // resultRunningVariance_dev
+    DeviceMem resultRunningVariance_dev(sizeof(AccDataType) *
+                                        resultRunningVariance_ref.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+    bnBias_dev.ToDevice(bnBias.mData.data());
+
+    if(updateMovingAverage)
+    {
+        resultRunningMean_dev.ToDevice(resultRunningMean_ref.mData.data());
+        resultRunningVariance_dev.ToDevice(resultRunningVariance_ref.mData.data());
+    };
+
+    std::array<index_t, Rank> i_inOutLengths;
+    std::array<index_t, Rank> i_inOutStrides;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumReduceDim> i_scaleBiasMeanVarStrides;
+
+    ck::ranges::copy(inOutLengths, i_inOutLengths.begin());
+    ck::ranges::copy(inOutStrides, i_inOutStrides.begin());
+    ck::ranges::copy(scaleBiasMeanVarLengths, i_scaleBiasMeanVarLengths.begin());
+    ck::ranges::copy(scaleBiasMeanVarStrides, i_scaleBiasMeanVarStrides.begin());
+
+    using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+    using DeviceBatchNormFwdInstance =
+        ck::tensor_operation::device::DeviceBatchNormFwdImpl<InOutDataType,
+                                                             InOutDataType,
+                                                             AccDataType,
+                                                             AccDataType,   // ScaleDataType
+                                                             AccDataType,   // BiasDataType
+                                                             AccDataType,   // MeanVarDataType
+                                                             PassThroughOp, // YElementwiseOp
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             UseMultiblockInK,
+                                                             256,
+                                                             16,
+                                                             16,
+                                                             1,
+                                                             2,
+                                                             0,
+                                                             1,
+                                                             1,
+                                                             1,
+                                                             1,
+                                                             1>;
+
+    auto batchnorm_fwd = DeviceBatchNormFwdInstance{};
+
+    auto argument_ptr = batchnorm_fwd.MakeArgumentPointer(
+        i_inOutLengths,
+        i_inOutStrides,
+        i_inOutStrides,
+        {0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
+        i_scaleBiasMeanVarLengths,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        i_scaleBiasMeanVarStrides,
+        x_dev.GetDeviceBuffer(),
+        bnScale_dev.GetDeviceBuffer(),
+        bnBias_dev.GetDeviceBuffer(),
+        epsilon,
+        PassThroughOp{},
+        y_dev.GetDeviceBuffer(),
+        saveMeanAndInvVariance ? resultSaveMean_dev.GetDeviceBuffer() : nullptr,
+        saveMeanAndInvVariance ? resultSaveInvVariance_dev.GetDeviceBuffer() : nullptr,
+        averageFactor,
+        updateMovingAverage ? resultRunningMean_dev.GetDeviceBuffer() : nullptr,
+        updateMovingAverage ? resultRunningVariance_dev.GetDeviceBuffer() : nullptr);
+
+    if(!batchnorm_fwd.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters seems not supported by the BatchNorm device instance, "
+                     "exiting!"
+                  << std::endl;
+        return (false);
+    };
+
+    size_t workspace_sz = batchnorm_fwd.GetWorkSpaceSize(argument_ptr.get());
+
+    DeviceMem workspace_dev(workspace_sz);
+
+    batchnorm_fwd.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+    auto invoker_ptr = batchnorm_fwd.MakeInvokerPointer();
+
+    if(time_kernel)
+    {
+        float avg_time   = 0.0f;
+        size_t num_bytes = 0;
+
+        size_t total_length = inOutLengths[0] * inOutLengths[1] * inOutLengths[2] * inOutLengths[3];
+        size_t invariant_length = inOutLengths[3];
+
+        avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        // inputing of x, scale, bias, outputing of y
+        num_bytes +=
+            total_length * sizeof(InOutDataType) * 2 + invariant_length * sizeof(AccDataType) * 2;
+
+        // outputing of mean, inv-variance
+        num_bytes += saveMeanAndInvVariance ? invariant_length * sizeof(AccDataType) * 2 : 0;
+
+        // updating of moving mean, variance
+        num_bytes += updateMovingAverage ? invariant_length * sizeof(AccDataType) * 4 : 0;
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+    }
+    else
+        (void)invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+
+        using ReferenceBatchNormFwdInstance =
+            ck::tensor_operation::host::ReferenceBatchNormFwd<InOutDataType,
+                                                              InOutDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              AccDataType,
+                                                              PassThroughOp,
+                                                              Rank,
+                                                              NumReduceDim>;
+
+        auto batchNormFwd_ref = ReferenceBatchNormFwdInstance{};
+
+        auto argument_ptr_ref = batchNormFwd_ref.MakeArgumentPointer(
+            i_inOutLengths,
+            i_inOutStrides,
+            i_inOutStrides,
+            {0, 1, 2}, // indicates physical indices of reduce dimensions in lengths[] and strides[]
+            i_scaleBiasMeanVarLengths,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            i_scaleBiasMeanVarStrides,
+            x.mData.data(),
+            bnScale.mData.data(),
+            bnBias.mData.data(),
+            epsilon,
+            PassThroughOp{},
+            y_ref.mData.data(),
+            saveMeanAndInvVariance ? resultSaveMean_ref.mData.data() : nullptr,
+            saveMeanAndInvVariance ? resultSaveInvVariance_ref.mData.data() : nullptr,
+            averageFactor,
+            updateMovingAverage ? resultRunningMean_ref.mData.data() : nullptr,
+            updateMovingAverage ? resultRunningVariance_ref.mData.data() : nullptr);
+
+        if(!batchNormFwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters seems not supported by the BatchNorm reference "
+                         "instance, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = batchNormFwd_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+
+        y_dev.FromDevice(y.mData.data());
+        pass = pass && ck::utils::check_err(y, y_ref, "Incorrect normalized output values");
+
+        if(updateMovingAverage)
+        {
+            Tensor<AccDataType> resultRunningMean(scaleBiasMeanVarLengths);
+            Tensor<AccDataType> resultRunningVariance(scaleBiasMeanVarLengths);
+
+            resultRunningMean_dev.FromDevice(resultRunningMean.mData.data());
+            resultRunningVariance_dev.FromDevice(resultRunningVariance.mData.data());
+
+            pass = pass && ck::utils::check_err(resultRunningMean,
+                                                resultRunningMean_ref,
+                                                "Incorrect running mean values");
+            pass = pass && ck::utils::check_err(resultRunningVariance,
+                                                resultRunningVariance_ref,
+                                                "Incorrect running variance values");
+        };
+
+        if(saveMeanAndInvVariance)
+        {
+            using ck::host_common::dumpBufferToFile;
+
+            Tensor<AccDataType> resultSaveMean(scaleBiasMeanVarLengths);
+            Tensor<AccDataType> resultSaveInvVariance(scaleBiasMeanVarLengths);
+
+            resultSaveMean_dev.FromDevice(resultSaveMean.mData.data());
+            resultSaveInvVariance_dev.FromDevice(resultSaveInvVariance.mData.data());
+
+            pass = pass && ck::utils::check_err(
+                               resultSaveMean, resultSaveMean_ref, "Incorrect saved mean values");
+            pass = pass && ck::utils::check_err(resultSaveInvVariance,
+                                                resultSaveInvVariance_ref,
+                                                "Incorrect saved invvariance values");
+        };
+    };
+
+    return (pass);
+};
+
+const double epsilon              = std::numeric_limits<float>::epsilon();
+static const double averageFactor = 0.1;
+
+int main(int argc, char* argv[])
+{
+    bool pass = true;
+
+    if(argc > 1)
+    {
+        BatchNormFwdArg arg;
+
+        if(arg.processArgs(argc, argv) < 0)
+            return (-1);
+
+        if(arg.data_type == 0)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(arg.do_verification,
+                                                                    arg.init_method,
+                                                                    arg.time_kernel,
+                                                                    arg.inOutLengths,
+                                                                    arg.updateMovingAverage,
+                                                                    arg.saveMeanAndInvVariance,
+                                                                    averageFactor,
+                                                                    epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<ck::half_t, float, false>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.updateMovingAverage,
+                                                                     arg.saveMeanAndInvVariance,
+                                                                     averageFactor,
+                                                                     epsilon);
+        }
+        else if(arg.data_type == 1)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<float, float, true>(arg.do_verification,
+                                                               arg.init_method,
+                                                               arg.time_kernel,
+                                                               arg.inOutLengths,
+                                                               arg.updateMovingAverage,
+                                                               arg.saveMeanAndInvVariance,
+                                                               averageFactor,
+                                                               epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<float, float, false>(arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                arg.inOutLengths,
+                                                                arg.updateMovingAverage,
+                                                                arg.saveMeanAndInvVariance,
+                                                                averageFactor,
+                                                                epsilon);
+        }
+        else if(arg.data_type == 3)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<int8_t, float, true>(arg.do_verification,
+                                                                arg.init_method,
+                                                                arg.time_kernel,
+                                                                arg.inOutLengths,
+                                                                arg.updateMovingAverage,
+                                                                arg.saveMeanAndInvVariance,
+                                                                averageFactor,
+                                                                epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<int8_t, float, false>(arg.do_verification,
+                                                                 arg.init_method,
+                                                                 arg.time_kernel,
+                                                                 arg.inOutLengths,
+                                                                 arg.updateMovingAverage,
+                                                                 arg.saveMeanAndInvVariance,
+                                                                 averageFactor,
+                                                                 epsilon);
+        }
+        else if(arg.data_type == 5)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, true>(arg.do_verification,
+                                                                     arg.init_method,
+                                                                     arg.time_kernel,
+                                                                     arg.inOutLengths,
+                                                                     arg.updateMovingAverage,
+                                                                     arg.saveMeanAndInvVariance,
+                                                                     averageFactor,
+                                                                     epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<ck::bhalf_t, float, false>(arg.do_verification,
+                                                                      arg.init_method,
+                                                                      arg.time_kernel,
+                                                                      arg.inOutLengths,
+                                                                      arg.updateMovingAverage,
+                                                                      arg.saveMeanAndInvVariance,
+                                                                      averageFactor,
+                                                                      epsilon);
+        }
+        else if(arg.data_type == 6)
+        {
+            if(arg.use_multiblock_welford)
+                pass = bnorm_fwd_nhwc_test<double, double, true>(arg.do_verification,
+                                                                 arg.init_method,
+                                                                 arg.time_kernel,
+                                                                 arg.inOutLengths,
+                                                                 arg.updateMovingAverage,
+                                                                 arg.saveMeanAndInvVariance,
+                                                                 averageFactor,
+                                                                 epsilon);
+            else
+                pass = bnorm_fwd_nhwc_test<double, double, false>(arg.do_verification,
+                                                                  arg.init_method,
+                                                                  arg.time_kernel,
+                                                                  arg.inOutLengths,
+                                                                  arg.updateMovingAverage,
+                                                                  arg.saveMeanAndInvVariance,
+                                                                  averageFactor,
+                                                                  epsilon);
+        }
+    }
+    else
+    {
+        pass = bnorm_fwd_nhwc_test<ck::half_t, float, true>(true,
+                                                            2,
+                                                            false, // don't time kernel
+                                                            {128, 16, 6, 512},
+                                                            true,
+                                                            true,
+                                                            averageFactor,
+                                                            epsilon);
+
+        pass = pass && bnorm_fwd_nhwc_test<ck::half_t, float, false>(true,
+                                                                     2,
+                                                                     false, // don't time kernel
+                                                                     {128, 16, 3, 1024},
+                                                                     true,
+                                                                     true,
+                                                                     averageFactor,
+                                                                     epsilon);
+    };
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/34_batchnorm/batchnorm_infer_impl.hpp b/example/34_batchnorm/batchnorm_infer_impl.hpp
new file mode 100644
index 0000000000..d0b545b2a3
--- /dev/null
+++ b/example/34_batchnorm/batchnorm_infer_impl.hpp
@@ -0,0 +1,131 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cassert>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+
+#include "batchnorm_common.hpp"
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          ck::index_t Rank,
+          ck::index_t NumBatchNormReduceDim,
+          bool fastest_dim_is_reduced = false>
+int bnorm_infer(
+    bool time_kernel,
+    const std::array<int, NumBatchNormReduceDim> reduceDims,
+    const std::array<ck::index_t, Rank> xyLengths,
+    const std::array<ck::index_t, Rank> xStrides,
+    const std::array<ck::index_t, Rank> yStrides,
+    const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
+    const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+    const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+    const std::array<ck::index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
+    const void* p_x,
+    const void* p_scale,
+    const void* p_bias,
+    double epsilon,
+    const void* p_estimatedMean,
+    const void* p_estimatedVariance,
+    void* p_y)
+{
+    (void)bnScaleBiasMeanVarLengths;
+
+    static_assert(NumBatchNormReduceDim < Rank,
+                  "Invalid number of reduced dimensions for batchnorm!");
+
+    using DeviceNormalizeInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
+        ck::Tuple<XDataType, AccDataType, AccDataType, AccDataType, AccDataType>, // x, mean,
+                                                                                  // variance,
+                                                                                  // scale,
+                                                                                  // bias,
+        ck::Tuple<YDataType>,                                                     // y
+        NormalizeInInfer,
+        Rank,
+        2,                           // MPerthread
+        ck::Sequence<1, 1, 1, 1, 1>, // x, mean, variance, scale, bias
+        ck::Sequence<1>>;            // scalarPerVector: y
+
+    auto invariantDims = get_invariant_dims<Rank, NumBatchNormReduceDim>(reduceDims);
+    std::array<ck::index_t, Rank> aligned_bnScaleStrides{0};
+    std::array<ck::index_t, Rank> aligned_bnBiasStrides{0};
+    std::array<ck::index_t, Rank> aligned_bnMeanVarStrides{0};
+
+    int i = 0;
+    for(auto dim : invariantDims)
+    {
+        assert(xyLengths[dim] == bnScaleBiasMeanVarLengths[i]);
+
+        aligned_bnScaleStrides[dim]   = bnScaleStrides[i];
+        aligned_bnBiasStrides[dim]    = bnBiasStrides[i];
+        aligned_bnMeanVarStrides[dim] = bnMeanVarStrides[i];
+        i++;
+    };
+
+    int32_t reduceLength = 1;
+
+    for(auto dim : reduceDims)
+        reduceLength *= xyLengths[dim];
+
+    int32_t invariantLength = 1;
+
+    for(auto dim : invariantDims)
+        invariantLength *= xyLengths[dim];
+
+    size_t total_length = static_cast<size_t>(invariantLength) * reduceLength;
+
+    float avg_time        = 0.0f;
+    std::size_t num_bytes = 0;
+
+    auto dev_normalize = DeviceNormalizeInstance{};
+
+    auto argument_ptr1 = dev_normalize.MakeArgumentPointer(
+        xyLengths,
+        {xStrides,
+         aligned_bnMeanVarStrides,
+         aligned_bnMeanVarStrides,
+         aligned_bnScaleStrides,
+         aligned_bnBiasStrides},
+        {yStrides},
+        {p_x, p_estimatedMean, p_estimatedVariance, p_scale, p_bias},
+        {p_y},
+        NormalizeInInfer{epsilon});
+
+    if(!dev_normalize.IsSupportedArgument(argument_ptr1.get()))
+    {
+        std::cout << "The runtime parameters seems not supported by the Devic, exiting!"
+                  << std::endl;
+
+        return (-1);
+    };
+
+    auto invoker_ptr1 = dev_normalize.MakeInvokerPointer();
+
+    avg_time += invoker_ptr1->Run(argument_ptr1.get(), StreamConfig{nullptr, time_kernel});
+
+    num_bytes += total_length * sizeof(XDataType) +
+                 invariantLength *
+                     (sizeof(ScaleDataType) + sizeof(BiasDataType) + 2 * sizeof(MeanVarDataType)) +
+                 total_length * sizeof(YDataType);
+
+    if(time_kernel)
+    {
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+    };
+
+    return (0);
+};
diff --git a/example/35_splitK_gemm/CMakeLists.txt b/example/35_splitK_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..251a9b93c5
--- /dev/null
+++ b/example/35_splitK_gemm/CMakeLists.txt
@@ -0,0 +1,28 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_custom_target(example_splitK_gemm_xdl)
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
+    add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp32)
+   endif()
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
+    add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp16)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitK_gemm_xdl_bfp16 splitK_gemm_xdl_bfp16.cpp)
+    add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_bfp16)
+   endif()
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp)
+    add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int8)
+   endif()
+   if(USE_BITINT_EXTENSION_INT4)
+      add_example_executable(example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp)
+      add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int4)
+   endif()
+   set(target 1)
+ endif()
+endforeach()
diff --git a/example/35_splitK_gemm/run_splitK_gemm_example.inc b/example/35_splitK_gemm/run_splitK_gemm_example.inc
new file mode 100644
index 0000000000..e9bd5c552d
--- /dev/null
+++ b/example/35_splitK_gemm/run_splitK_gemm_example.inc
@@ -0,0 +1,217 @@
+#pragma once
+
+struct ProblemSize final
+{
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t stride_A = K;
+    ck::index_t stride_B = K;
+    ck::index_t stride_C = N;
+
+    ck::index_t k_batch = 4;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+bool run_splitK_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    using namespace ck::literals;
+
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+    static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+    static_assert(sizeof(ADataType) == sizeof(KernelADataType));
+    static_assert(sizeof(BDataType) == sizeof(KernelBDataType));
+#endif
+
+    auto& [M, N, K, StrideA, StrideB, StrideC, KBatch] = problem_size;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+        b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+    }
+
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+#ifdef BUILD_INT4_EXAMPLE
+    const Tensor<KernelADataType> a_m_k_converted(a_m_k);
+    const Tensor<KernelBDataType> b_k_n_converted(b_k_n);
+
+    a_m_k_device_buf.ToDevice(a_m_k_converted.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n_converted.mData.data());
+#else
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
+#endif
+    c_m_n_device_buf.SetZero();
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(
+#ifdef BUILD_INT4_EXAMPLE
+        static_cast<KernelADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+        static_cast<KernelBDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+#else
+        static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+        static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+#endif
+        static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+        M,
+        N,
+        K,
+        StrideA,
+        StrideB,
+        StrideC,
+        a_element_op,
+        b_element_op,
+        c_element_op,
+        KBatch);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+    bool pass = true;
+
+    if(config.do_verification)
+    {
+        c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        if(std::is_same<CDataType, ck::half_t>::value)
+        {
+            pass &= ck::utils::check_err(
+                c_m_n_device_result, c_m_n_host_result, "fp16 incorrect result", 3e-3, 1e-3);
+        }
+        else
+        {
+            pass &= ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+        }
+    }
+
+    if(config.time_kernel)
+    {
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+        std::size_t flop = std::size_t(2) * M * N * K;
+        std::size_t num_btype =
+            sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+        float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << gemm.GetTypeString() << std::endl;
+    }
+
+    return pass;
+}
+
+bool run_splitK_gemm_example(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 5)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+        problem_size.k_batch   = std::stoi(argv[4]);
+    }
+    else if(argc == 11)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+        problem_size.k_batch   = std::stoi(argv[4]);
+
+        problem_size.M = std::stoi(argv[5]);
+        problem_size.N = std::stoi(argv[6]);
+        problem_size.K = std::stoi(argv[7]);
+
+        problem_size.stride_A = std::stoi(argv[8]);
+        problem_size.stride_B = std::stoi(argv[9]);
+        problem_size.stride_C = std::stoi(argv[10]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4: KBatch\n");
+        printf("arg5 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
+        exit(0);
+    }
+
+    return run_splitK_gemm(problem_size, config);
+}
diff --git a/example/35_splitK_gemm/splitK_gemm_xdl_bfp16.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_bfp16.cpp
new file mode 100644
index 0000000000..fdf49a31b7
--- /dev/null
+++ b/example/35_splitK_gemm/splitK_gemm_xdl_bfp16.cpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType   = BF16;
+using BDataType   = BF16;
+using AccDataType = F32;
+using CDataType   = F32;
+using ComputeType = BF16;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
+    // clang-format off
+//######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer|  KPer|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|      Compute|
+//######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|         Type|
+//######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|             |
+//######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |             | 
+        < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   256,   128,     4,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               4, ComputeType>;
+// clang-format on
+
+#include "run_splitK_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }
diff --git a/example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp
new file mode 100644
index 0000000000..74fb16e15b
--- /dev/null
+++ b/example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType   = F16;
+using BDataType   = F16;
+using AccDataType = F32;
+using CDataType   = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
+    // clang-format off
+//######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer|  KPer|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+//######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+//######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   256,   128,     4,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+#include "run_splitK_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }
diff --git a/example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp
new file mode 100644
index 0000000000..7506f69420
--- /dev/null
+++ b/example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType   = F32;
+using BDataType   = F32;
+using AccDataType = F32;
+using CDataType   = F32;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
+    // clang-format off
+//######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer|  KPer|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+//######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+//######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   256,   128,     4,   4,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>;
+// clang-format on
+
+#include "run_splitK_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }
diff --git a/example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp
new file mode 100644
index 0000000000..7ebf914408
--- /dev/null
+++ b/example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp
@@ -0,0 +1,92 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType   = ck::int4_t;
+using BDataType   = ck::int4_t;
+using AccDataType = int32_t;
+using CDataType   = int32_t;
+
+using KernelADataType = int8_t;
+using KernelBDataType = int8_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
+    // clang-format off
+        <KernelADataType,      //ADataType    
+         KernelBDataType,      //BDataType   
+         CDataType,            //EDataType
+         AccDataType,          //AccDataType
+         ALayout,              //ALayout
+         BLayout,              //BLayout
+         CLayout,              //ELayout
+         AElementOp,           //AElementwiseOperation
+         BElementOp,           //BElementwiseOperation
+         CElementOp,           //CElementwiseOperation
+         GemmDefault,          //GEMMSpecialization
+         256,                  // BlockSize
+         256,                  // MPerBlock
+         128,                  // NPerBlock
+         4,                    // KPerBlock
+         16,                   // K1
+         32,                   // MPerXdl
+         32,                   // NPerXdl
+         4,                    // MXdlPerWave
+         2,                    // NXdlPerWave
+         S<1, 4, 64, 1>,       // ABlockTransfer ThreadCluster Lengths_K0_M_K1
+         S<0, 2, 1, 3>,        // ABlockTransfer ThreadCluster ArrangeOrder
+         S<0, 2, 1, 3>,        // ABlockTransfer SrcAccessOrder
+         3,                    // ABlockTransfer SrcVectorDim
+         16,                   // ABlockTransfer SrcScalarPerVector
+         16,                   // ABlockTransfer DstScalarPerVector_K1
+         true,                 // ABlockLdsExtraM
+         S<1, 4, 64, 1>,       // BBlockTransfer ThreadCluster Lengths_K0_N_K1
+         S<0, 1, 3, 2>,        // BBlockTransfer ThreadCluster ArrangeOrder
+         S<0, 1, 3, 2>,        // BBlockTransfer SrcAccessOrder
+         3,                    // BBlockTransfer SrcVectorDim
+         16,                   // BBlockTransfer SrcScalarPerVector
+         16,                   // BBlockTransfer DstScalarPerVector_K1
+         true,                 // BBlockLdsExtraN
+         1,                    // CShuffleMXdlPerWavePerShuffle
+         1,                    // CShuffleNXdlPerWavePerShuffle
+         S<1, 32, 1, 8>,       // CBlockTransferClusterLengths _MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+         4>;                   // CBlockTransferScalarPerVector_NWaveNPerXdl
+// clang-format on
+
+#define BUILD_INT4_EXAMPLE
+#include "run_splitK_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }
diff --git a/example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp
new file mode 100644
index 0000000000..6b0c1aa02d
--- /dev/null
+++ b/example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp
@@ -0,0 +1,56 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType   = int8_t;
+using BDataType   = int8_t;
+using AccDataType = int32_t;
+using CDataType   = int32_t;
+using ComputeType = int8_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
+    // clang-format off
+//######|     AData|     BData|     CData|     AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer|  KPer|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|     Compute|
+//######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise| Spacialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|        Type|
+//######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|            |
+//######|          |          |          |            |        |        |        |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |            |
+        < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,   256,   256,   128,     4,  16,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,             16,             16,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,             16,             16,      true,           1,           1,                   S<1, 32, 1, 8>,               4, ComputeType>;
+// clang-format on
+
+#include "run_splitK_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_splitK_gemm_example(argc, argv); }
diff --git a/example/36_sparse_embedding/CMakeLists.txt b/example/36_sparse_embedding/CMakeLists.txt
new file mode 100644
index 0000000000..9cbcf5540e
--- /dev/null
+++ b/example/36_sparse_embedding/CMakeLists.txt
@@ -0,0 +1 @@
+add_example_executable(example_sparse_embedding3_forward_layernorm sparse_embedding3_forward_layernorm.cpp)
diff --git a/example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp b/example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp
new file mode 100644
index 0000000000..d2337dcda5
--- /dev/null
+++ b/example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp
@@ -0,0 +1,194 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+#include <ctime>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_sparse_embeddings_forward_layernorm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp"
+
+// clang-format off
+using EmbType       = ck::half_t;
+using IndexType     = int64_t;
+using GammaDataType = ck::half_t;
+using BetaDataType  = ck::half_t;
+using AccDataType   = float;
+using OutType       = ck::half_t;
+using EmbElementwiseOperation = ck::tensor_operation::element_wise::AddAdd;
+
+using DeviceInstance_fp16_e256   = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  256,   1, 1, 3>;
+using DeviceInstance_fp16_e512   = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  512,   1, 2, 3>;
+using DeviceInstance_fp16_e768   = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  768,   1, 1, 3>;
+using DeviceInstance_fp16_e1024  = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  1024,  1, 2, 3>;
+using DeviceInstance_fp16_e1536  = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  1536,  1, 2, 3>;
+using DeviceInstance_fp16_e2048  = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  2048,  1, 2, 3>;
+using DeviceInstance_fp16_e4096  = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  4096,  1, 8, 3>;
+using DeviceInstance_fp16_e8192  = ck::tensor_operation::device::DeviceSparseEmbeddingsForwardLayernorm<EmbType, IndexType, GammaDataType, BetaDataType, AccDataType, OutType, EmbElementwiseOperation, 256,  1,  256, 1,  8192,  1, 8, 3>;
+
+template<typename emb_type, ck::index_t dim> struct emb_kernel{};
+
+template<> struct emb_kernel<ck::half_t, 256>  { using kernel_type = DeviceInstance_fp16_e256; };
+template<> struct emb_kernel<ck::half_t, 512>  { using kernel_type = DeviceInstance_fp16_e512; };
+template<> struct emb_kernel<ck::half_t, 768>  { using kernel_type = DeviceInstance_fp16_e768; };
+template<> struct emb_kernel<ck::half_t, 1024> { using kernel_type = DeviceInstance_fp16_e1024; };
+template<> struct emb_kernel<ck::half_t, 1536> { using kernel_type = DeviceInstance_fp16_e1536; };
+template<> struct emb_kernel<ck::half_t, 2048> { using kernel_type = DeviceInstance_fp16_e2048; };
+template<> struct emb_kernel<ck::half_t, 4096> { using kernel_type = DeviceInstance_fp16_e4096; };
+template<> struct emb_kernel<ck::half_t, 8192> { using kernel_type = DeviceInstance_fp16_e8192; };
+
+// clang-format on
+
+int main()
+{
+    bool time_kernel = true;
+
+    constexpr auto num_rows = 65536;
+    constexpr auto dims     = ck::Sequence<256, 512, 768, 1024, 1536, 2048, 4096, 8192>{};
+    // constexpr auto dims = ck::Sequence<256, 512>{};
+    constexpr auto index_length   = 2048;
+    constexpr AccDataType epsilon = 1e-4;
+
+    auto f_host_tensor_desc_1d = [](std::size_t len_) { return HostTensorDescriptor({len_}); };
+
+    auto f_host_tensor_desc_2d = [](std::size_t rows_, std::size_t cols_) {
+        return HostTensorDescriptor({rows_, cols_});
+    };
+
+    using ReferenceInstance =
+        ck::tensor_operation::host::ReferenceSparseEmbedding3ForwardLayernorm<EmbType,
+                                                                              IndexType,
+                                                                              GammaDataType,
+                                                                              BetaDataType,
+                                                                              AccDataType,
+                                                                              OutType>;
+
+    ck::static_for<0, dims.Size(), 1>{}([&](auto I) {
+        std::srand(std::time(nullptr));
+        constexpr auto current_dim = dims.At(I);
+        Tensor<EmbType> emb_a(f_host_tensor_desc_2d(num_rows, current_dim));
+        Tensor<EmbType> emb_b(f_host_tensor_desc_2d(num_rows, current_dim));
+        Tensor<EmbType> emb_c(f_host_tensor_desc_2d(num_rows, current_dim));
+
+        Tensor<IndexType> index_a(f_host_tensor_desc_1d(index_length));
+        Tensor<IndexType> index_b(f_host_tensor_desc_1d(index_length));
+        Tensor<IndexType> index_c(f_host_tensor_desc_1d(index_length));
+
+        Tensor<GammaDataType> gamma(f_host_tensor_desc_1d(current_dim));
+        Tensor<BetaDataType> beta(f_host_tensor_desc_1d(current_dim));
+
+        Tensor<OutType> out(f_host_tensor_desc_2d(index_length, current_dim));
+
+        emb_a.GenerateTensorValue(GeneratorTensor_3<EmbType>{0.0, 1.0});
+        emb_b.GenerateTensorValue(GeneratorTensor_3<EmbType>{0.0, 1.0});
+        emb_c.GenerateTensorValue(GeneratorTensor_3<EmbType>{0.0, 1.0});
+
+        index_a.GenerateTensorValue(GeneratorTensor_2<IndexType>{0, num_rows});
+        index_b.GenerateTensorValue(GeneratorTensor_2<IndexType>{0, num_rows});
+        index_c.GenerateTensorValue(GeneratorTensor_2<IndexType>{0, num_rows});
+
+        gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{0.0, 1.0});
+        beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{0.0, 1.0});
+
+        DeviceMem emb_a_dev(sizeof(EmbType) * emb_a.mDesc.GetElementSpaceSize());
+        DeviceMem emb_b_dev(sizeof(EmbType) * emb_b.mDesc.GetElementSpaceSize());
+        DeviceMem emb_c_dev(sizeof(EmbType) * emb_c.mDesc.GetElementSpaceSize());
+
+        DeviceMem index_a_dev(sizeof(IndexType) * index_a.mDesc.GetElementSpaceSize());
+        DeviceMem index_b_dev(sizeof(IndexType) * index_b.mDesc.GetElementSpaceSize());
+        DeviceMem index_c_dev(sizeof(IndexType) * index_c.mDesc.GetElementSpaceSize());
+
+        DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+        DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+
+        DeviceMem out_dev(sizeof(OutType) * out.mDesc.GetElementSpaceSize());
+
+        emb_a_dev.ToDevice(emb_a.mData.data());
+        emb_b_dev.ToDevice(emb_b.mData.data());
+        emb_c_dev.ToDevice(emb_c.mData.data());
+
+        index_a_dev.ToDevice(index_a.mData.data());
+        index_b_dev.ToDevice(index_b.mData.data());
+        index_c_dev.ToDevice(index_c.mData.data());
+
+        gamma_dev.ToDevice(gamma.mData.data());
+        beta_dev.ToDevice(beta.mData.data());
+
+        auto device_instance = typename emb_kernel<EmbType, current_dim>::kernel_type{};
+        auto argument_ptr    = device_instance.MakeArgumentPointer(
+            out_dev.GetDeviceBuffer(),
+            {ck::type_convert<EmbType*>(emb_a_dev.GetDeviceBuffer()),
+             ck::type_convert<EmbType*>(emb_b_dev.GetDeviceBuffer()),
+             ck::type_convert<EmbType*>(emb_c_dev.GetDeviceBuffer())},
+            {ck::type_convert<IndexType*>(index_a_dev.GetDeviceBuffer()),
+             ck::type_convert<IndexType*>(index_b_dev.GetDeviceBuffer()),
+             ck::type_convert<IndexType*>(index_c_dev.GetDeviceBuffer())},
+            gamma_dev.GetDeviceBuffer(),
+            beta_dev.GetDeviceBuffer(),
+            current_dim,
+            index_length,
+            epsilon,
+            EmbElementwiseOperation{});
+        std::cout << "Dim:" << current_dim << ", kernel:" << device_instance.GetTypeString()
+                  << std::endl
+                  << std::flush;
+
+        bool is_supported = device_instance.IsSupportedArgument(argument_ptr.get());
+
+        if(!is_supported)
+        {
+            std::cout << "Runtime parameters are not supported" << std::endl;
+            return;
+        }
+
+        auto invoker_ptr = device_instance.MakeInvokerPointer();
+        float time_ms    = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        bool pass = true;
+        {
+            Tensor<OutType> out_from_dev(f_host_tensor_desc_2d(index_length, current_dim));
+            ReferenceInstance ref;
+            auto ref_argument = ref.MakeArgument(out,
+                                                 emb_a,
+                                                 emb_b,
+                                                 emb_c,
+                                                 index_a,
+                                                 index_b,
+                                                 index_c,
+                                                 gamma,
+                                                 beta,
+                                                 num_rows,
+                                                 current_dim,
+                                                 index_length,
+                                                 epsilon);
+            auto ref_invoker  = ref.MakeInvoker();
+            ref_invoker.Run(ref_argument);
+
+            out_dev.FromDevice(out_from_dev.mData.data());
+            pass &= ck::utils::check_err(out_from_dev, out, "Error: Incorrect results", 1e-3, 1e-3);
+        }
+
+        double total_read = current_dim * index_length * 3 * sizeof(EmbType) +
+                            current_dim * sizeof(GammaDataType) +
+                            current_dim * sizeof(BetaDataType);
+        double total_write = current_dim * index_length * sizeof(OutType);
+        double gbps        = (total_read + total_write) / time_ms / 1e6;
+
+        std::cout << ", total bytes:" << (total_read + total_write) << ", time:" << time_ms
+                  << ", gbps:" << gbps << ", valid:" << (pass ? "y" : "n") << std::endl
+                  << std::flush;
+    });
+
+    return 0;
+}
diff --git a/example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt b/example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
new file mode 100644
index 0000000000..36bb5720d5
--- /dev/null
+++ b/example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
@@ -0,0 +1,3 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_batched_gemm_add_add_relu_gemm_add_xdl_fp16 batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp)
+endif()
diff --git a/example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp b/example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp
new file mode 100644
index 0000000000..36dcf58d70
--- /dev/null
+++ b/example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp
@@ -0,0 +1,521 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+/*
+Computes C_m_o = Relu(A0[m, k] * B0[n, k] + D00[m, n] + D01[mn]) * B1[n, o] + D1[m, o]
+*/
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using A0DataType        = F16;
+using B0DataType        = F16;
+using Acc0DataType      = F32;
+using D00DataType       = F16;
+using D01DataType       = F16;
+using B1DataType        = F16;
+using Acc1DataType      = F32;
+using C1ShuffleDataType = F32;
+using D1DataType        = F16;
+using E1DataType        = F16;
+
+using A0Layout  = Row;
+using B0Layout  = Col;
+using D00Layout = Row;
+using D01Layout = Row;
+using B1Layout  = Row;
+using D1Layout  = Row;
+using E1Layout  = Row;
+
+// E = Relu(C + D0 + D1)
+struct AddAddRelu
+{
+    __host__ __device__ void
+    operator()(ck::half_t& e, const ck::half_t& c, const ck::half_t& d0, const ck::half_t& d1) const
+    {
+        const ck::half_t x = c + d0 + d1;
+
+        ck::tensor_operation::element_wise::Relu{}.template operator()<ck::half_t>(e, x);
+    }
+    __host__ __device__ void
+    operator()(float& e, const float& c, const ck::half_t& d0, const ck::half_t& d1) const
+    {
+        const float x = c + (d0 + d1);
+
+        ck::tensor_operation::element_wise::Relu{}.template operator()<float>(e, x);
+    }
+};
+
+// E = Gelu(C + D0 + D1)
+struct AddAddGelu
+{
+    __host__ __device__ void
+    operator()(ck::half_t& e, const ck::half_t& c, const ck::half_t& d0, const ck::half_t& d1) const
+    {
+        const ck::half_t x = c + d0 + d1;
+
+        ck::tensor_operation::element_wise::Gelu{}.template operator()<ck::half_t, ck::half_t>(e,
+                                                                                               x);
+    }
+
+    __host__ __device__ void
+    operator()(float& e, const float& c, const ck::half_t& d0, const ck::half_t& d1) const
+    {
+        const float x = c + (d0 + d1);
+
+        ck::tensor_operation::element_wise::Gelu{}.template operator()<float, float>(e, x);
+    }
+};
+
+// E = FastGelu(C + D0 + D1)
+struct AddAddFastGelu
+{
+    __host__ __device__ void
+    operator()(float& e, const float& c, const ck::half_t& d0, const ck::half_t& d1) const
+    {
+        const float x = c + (d0 + d1);
+
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(e, x);
+    }
+};
+
+using A0ElementOp   = PassThrough;
+using B0ElementOp   = PassThrough;
+using CDE0ElementOp = AddAddRelu;
+using A1ElementOp   = PassThrough;
+using B1ElementOp   = PassThrough;
+using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr bool PadGemm0M = false;
+static constexpr bool PadGemm0N = false;
+static constexpr bool PadGemm0K = false;
+static constexpr bool PadGemm1N = false;
+static constexpr bool PadGemm1K = false;
+
+using DeviceGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<
+        A0Layout,
+        B0Layout,
+        ck::Tuple<D00Layout, D01Layout>,
+        B1Layout,
+        ck::Tuple<D1Layout>,
+        E1Layout,
+        A0DataType,
+        B0DataType,
+        Acc0DataType,
+        ck::Tuple<D00DataType, D01DataType>,
+        B1DataType,
+        Acc1DataType,
+        C1ShuffleDataType,
+        ck::Tuple<D1DataType>,
+        E1DataType,
+        A0ElementOp,
+        B0ElementOp,
+        CDE0ElementOp,
+        B1ElementOp,
+        CDE1ElementOp,
+        PadGemm0M,
+        PadGemm0N,
+        PadGemm0K,
+        PadGemm1N,
+        PadGemm1K,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        128,         // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        4,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        9,           // D0sTransferSrcVectorDim
+        4,           // D0sTransferSrcScalaerPerVector
+        S<8, 32, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // GEMM shape
+    ck::index_t M              = 1024;
+    ck::index_t N              = 1024;
+    ck::index_t K              = 64;
+    ck::index_t O              = 128;
+    ck::index_t BatchCount     = 4;
+    ck::index_t StrideA0       = -1;
+    ck::index_t StrideB0       = -1;
+    ck::index_t StrideD00      = -1;
+    ck::index_t StrideD01      = -1;
+    ck::index_t StrideB1       = -1;
+    ck::index_t StrideD1       = -1;
+    ck::index_t StrideE1       = -1;
+    ck::index_t BatchStrideA0  = -1;
+    ck::index_t BatchStrideB0  = -1;
+    ck::index_t BatchStrideD00 = -1;
+    ck::index_t BatchStrideD01 = -1;
+    ck::index_t BatchStrideB1  = -1;
+    ck::index_t BatchStrideD1  = -1;
+    ck::index_t BatchStrideE1  = -1;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 9)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+    }
+    else if(argc == 23)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+        O = std::stoi(argv[7]);
+
+        BatchCount = std::stoi(argv[8]);
+
+        StrideA0  = std::stoi(argv[9]);
+        StrideB0  = std::stoi(argv[10]);
+        StrideD00 = std::stoi(argv[11]);
+        StrideD01 = std::stoi(argv[12]);
+        StrideB1  = std::stoi(argv[13]);
+        StrideD1  = std::stoi(argv[14]);
+        StrideE1  = std::stoi(argv[15]);
+
+        BatchStrideA0  = std::stoi(argv[16]);
+        BatchStrideB0  = std::stoi(argv[17]);
+        BatchStrideD00 = std::stoi(argv[18]);
+        BatchStrideD01 = std::stoi(argv[19]);
+        BatchStrideB1  = std::stoi(argv[20]);
+        BatchStrideD1  = std::stoi(argv[21]);
+        BatchStrideE1  = std::stoi(argv[22]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 8: M, N, K, O, Batch\n");
+        printf(
+            "arg9 to 15: StrideA0, StrideB0, StrideD00, StrideD01, StrideB1, StrideD1, StrideE1\n");
+        printf("arg16 to 22: BatchStrideA0, BatchStrideB0, BatchStrideD00, BatchStrideD01, "
+               "BatchStrideB1, BatchStrideD1, BatchStrideE1 \n");
+        exit(0);
+    }
+
+    const int DefaultStrideA0  = ck::is_same_v<A0Layout, Row> ? K : M;
+    const int DefaultStrideB0  = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideD00 = ck::is_same_v<D00Layout, Row> ? N : M;
+    const int DefaultStrideD01 = ck::is_same_v<D01Layout, Row> ? N : M;
+    const int DefaultStrideB1  = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideD1  = ck::is_same_v<D1Layout, Row> ? O : M;
+    const int DefaultStrideE1  = ck::is_same_v<E1Layout, Row> ? O : M;
+
+    StrideA0  = (StrideA0 < 0) ? DefaultStrideA0 : StrideA0;
+    StrideB0  = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideD00 = (StrideD00 < 0) ? DefaultStrideD00 : StrideD00;
+    StrideD01 = (StrideD01 < 0) ? DefaultStrideD01 : StrideD01;
+    StrideB1  = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideD1  = (StrideD1 < 0) ? DefaultStrideD1 : StrideD1;
+    StrideE1  = (StrideE1 < 0) ? DefaultStrideE1 : StrideE1;
+
+    const int DefaultBatchStrideA0  = (ck::is_same_v<A0Layout, Col> ? K : M) * StrideA0;
+    const int DefaultBatchStrideB0  = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideD00 = (ck::is_same_v<D00Layout, Col> ? N : M) * StrideD00;
+    const int DefaultBatchStrideD01 = (ck::is_same_v<D01Layout, Col> ? N : M) * StrideD01;
+    const int DefaultBatchStrideB1  = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideD1  = (ck::is_same_v<D1Layout, Col> ? O : M) * StrideD1;
+    const int DefaultBatchStrideE1  = (ck::is_same_v<E1Layout, Col> ? O : M) * StrideE1;
+
+    BatchStrideA0  = BatchStrideA0 < 0 ? DefaultBatchStrideA0 : BatchStrideA0;
+    BatchStrideB0  = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideD00 = BatchStrideD00 < 0 ? DefaultBatchStrideD00 : BatchStrideD00;
+    BatchStrideD01 = BatchStrideD01 < 0 ? DefaultBatchStrideD01 : BatchStrideD01;
+    BatchStrideB1  = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideD1  = BatchStrideD1 < 0 ? DefaultBatchStrideD1 : BatchStrideD1;
+    BatchStrideE1  = BatchStrideE1 < 0 ? DefaultBatchStrideE1 : BatchStrideE1;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
+        }
+    };
+
+    // E_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<A0DataType> a0_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA0, BatchStrideA0, A0Layout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<D00DataType> d00_g_m_n(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideD00, BatchStrideD00, D00Layout{}));
+    Tensor<D01DataType> d01_g_m_n(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideD01, BatchStrideD01, D01Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<D1DataType> d1_g_m_o(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideD1, BatchStrideD1, D1Layout{}));
+    Tensor<E1DataType> e1_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideE1, BatchStrideE1, E1Layout{}));
+    Tensor<E1DataType> e1_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideE1, BatchStrideE1, E1Layout{}));
+
+    std::cout << "a0_g_m_k: " << a0_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "d00_g_m_n: " << d00_g_m_n.mDesc
+              << " size: " << d00_g_m_n.mDesc.GetElementSpaceSize() << std::endl;
+    std::cout << "d01_g_m_n: " << d01_g_m_n.mDesc
+              << " size: " << d01_g_m_n.mDesc.GetElementSpaceSize() << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "e1_g_m_o: " << e1_g_m_o_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_g_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 3});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 3});
+        d00_g_m_n.GenerateTensorValue(GeneratorTensor_2<D00DataType>{-2, 3});
+        d01_g_m_n.GenerateTensorValue(GeneratorTensor_2<D01DataType>{-2, 3});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 3});
+        d1_g_m_o.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-2, 3});
+        break;
+    case 2:
+        a0_g_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        d00_g_m_n.GenerateTensorValue(GeneratorTensor_3<D00DataType>{0.0, 1.0});
+        d01_g_m_n.GenerateTensorValue(GeneratorTensor_3<D01DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        d1_g_m_o.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+        break;
+    default:
+        a0_g_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{1});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        d00_g_m_n.GenerateTensorValue(GeneratorTensor_1<D00DataType>{1});
+        d01_g_m_n.GenerateTensorValue(GeneratorTensor_1<D01DataType>{1});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        d1_g_m_o.GenerateTensorValue(GeneratorTensor_1<D1DataType>{1});
+    }
+
+    DeviceMem a0_g_m_k_device_buf(sizeof(A0DataType) * a0_g_m_k.mDesc.GetElementSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSize());
+    DeviceMem d00_g_m_n_device_buf(sizeof(D00DataType) * d00_g_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d01_g_m_n_device_buf(sizeof(D01DataType) * d01_g_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSize());
+    DeviceMem e1_g_m_o_device_buf(sizeof(E1DataType) *
+                                  e1_g_m_o_device_result.mDesc.GetElementSize());
+    DeviceMem d1_g_m_o_device_buf(sizeof(D1DataType) * d1_g_m_o.mDesc.GetElementSpaceSize());
+
+    a0_g_m_k_device_buf.ToDevice(a0_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    d00_g_m_n_device_buf.ToDevice(d00_g_m_n.mData.data());
+    d01_g_m_n_device_buf.ToDevice(d01_g_m_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+    d1_g_m_o_device_buf.ToDevice(d1_g_m_o.mData.data());
+
+    auto a0_element_op   = A0ElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto cde0_element_op = CDE0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto cde1_element_op = CDE1ElementOp{};
+
+    // do GEMM
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+    auto argument =
+        gemm.MakeArgument(static_cast<A0DataType*>(a0_g_m_k_device_buf.GetDeviceBuffer()),
+                          static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+                          std::array<const void*, 2>{d00_g_m_n_device_buf.GetDeviceBuffer(),
+                                                     d01_g_m_n_device_buf.GetDeviceBuffer()},
+                          static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+                          std::array<const void*, 1>{d1_g_m_o_device_buf.GetDeviceBuffer()},
+                          static_cast<E1DataType*>(e1_g_m_o_device_buf.GetDeviceBuffer()),
+                          M,
+                          N,
+                          K,
+                          O,
+                          BatchCount,
+                          StrideA0,
+                          StrideB0,
+                          std::array<ck::index_t, 2>{StrideD00, StrideD01},
+                          StrideB1,
+                          std::array<ck::index_t, 1>{StrideD1},
+                          StrideE1,
+                          BatchStrideA0,
+                          BatchStrideB0,
+                          std::array<ck::index_t, 2>{BatchStrideD00, BatchStrideD01},
+                          BatchStrideB1,
+                          std::array<ck::index_t, 1>{BatchStrideD1},
+                          BatchStrideE1,
+                          a0_element_op,
+                          b0_element_op,
+                          cde0_element_op,
+                          b1_element_op,
+                          cde1_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype =
+        (sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(D00DataType) * N +
+         sizeof(D01DataType) * N + sizeof(B1DataType) * N * O + sizeof(E1DataType) * M * O +
+         sizeof(D1DataType) * O) *
+        BatchCount;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << gemm.GetTypeString() << std::endl;
+
+    e1_g_m_o_device_buf.FromDevice(e1_g_m_o_device_result.mData.data());
+
+    if(do_verification)
+    {
+        using ReferenceGemm0Instance =
+            ck::tensor_operation::host::ReferenceBatchedGemm<A0DataType,
+                                                             B0DataType,
+                                                             Acc0DataType,
+                                                             Acc0DataType,
+                                                             A0ElementOp,
+                                                             B0ElementOp,
+                                                             PassThrough>;
+
+        using ReferenceGemm1Instance =
+            ck::tensor_operation::host::ReferenceBatchedGemm<Acc0DataType,
+                                                             B1DataType,
+                                                             Acc1DataType,
+                                                             Acc1DataType,
+                                                             PassThrough,
+                                                             B1ElementOp,
+                                                             PassThrough>;
+
+        // Output of Gemm0 is input A of Gemm1
+        Tensor<Acc0DataType> c0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+        Tensor<Acc0DataType> e0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+        Tensor<Acc1DataType> c1_g_m_o(f_host_tensor_descriptor(BatchCount, M, O, O, M * O, Row{}));
+
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a0_g_m_k, b0_g_k_n, c0_g_m_n, a0_element_op, b0_element_op, PassThrough{});
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        // bias+bias+relu
+        e0_g_m_n.ForEach([&](auto&, auto idx) {
+            cde0_element_op(e0_g_m_n(idx), c0_g_m_n(idx), d00_g_m_n(idx), d01_g_m_n(idx));
+        });
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            e0_g_m_n, b1_g_n_o, c1_g_m_o, PassThrough{}, b1_element_op, PassThrough{});
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        // bias
+        e1_g_m_o_host_result.ForEach([&](auto&, auto idx) {
+            cde1_element_op(e1_g_m_o_host_result(idx), c1_g_m_o(idx), d1_g_m_o(idx));
+        });
+
+        return ck::utils::check_err(e1_g_m_o_device_result, e1_g_m_o_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}
diff --git a/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt b/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
new file mode 100644
index 0000000000..3821f8aaca
--- /dev/null
+++ b/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
@@ -0,0 +1,15 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_custom_target(example_grouped_conv_bwd_data)
+   add_example_executable(example_grouped_conv_bwd_data_fp16 grouped_conv_bwd_data_fp16.cpp)
+   add_example_executable(example_grouped_conv_bwd_data_bias_relu_fp16 grouped_conv_bwd_data_bias_relu_fp16.cpp)
+
+   add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_fp16)
+   add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_fp16)
+   set(target 1)
+ endif()
+endforeach()
+endif()
diff --git a/example/38_grouped_conv_bwd_data_multiple_d/common.hpp b/example/38_grouped_conv_bwd_data_multiple_d/common.hpp
new file mode 100644
index 0000000000..ca824b1075
--- /dev/null
+++ b/example/38_grouped_conv_bwd_data_multiple_d/common.hpp
@@ -0,0 +1,103 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <initializer_list>
+#include <iostream>
+#include <numeric>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static inline constexpr ck::index_t NDimSpatial = 2;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+using FP16 = ck::half_t;
+using FP32 = float;
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+};
+
+#define DefaultConvParams                                                                \
+    ck::utils::conv::ConvParam                                                           \
+    {                                                                                    \
+        NDimSpatial, 32, 4, 192, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, { 1, 1 } \
+    }
+
+inline void print_help_msg()
+{
+    std::cerr << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+inline bool parse_cmd_args(int argc,
+                           char* argv[],
+                           ExecutionConfig& config,
+                           ck::utils::conv::ConvParam& conv_params)
+{
+    constexpr int num_execution_config_args =
+        3; // arguments for do_verification, init_method, time_kernel
+    constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
+
+    constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
+    constexpr int threshold_to_catch_all_args =
+        threshold_to_catch_partial_args + num_conv_param_leading_args;
+
+    if(argc == 1)
+    {
+        // use default
+        config = ExecutionConfig{};
+    }
+    // catch only ExecutionConfig arguments
+    else if(argc == threshold_to_catch_partial_args)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    // catch both ExecutionConfig & ConvParam arguments
+    else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+        conv_params                       = ck::utils::conv::parse_conv_param(
+            num_dim_spatial, threshold_to_catch_partial_args, argv);
+    }
+    else
+    {
+        print_help_msg();
+        return false;
+    }
+
+    return true;
+}
diff --git a/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_fp16.cpp b/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_fp16.cpp
new file mode 100644
index 0000000000..a3533bb4cc
--- /dev/null
+++ b/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_fp16.cpp
@@ -0,0 +1,33 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using OutDataType      = FP16;
+using WeiDataType      = FP16;
+using AccDataType      = FP32;
+using CShuffleDataType = FP16;
+using BiasDataType     = FP16; // bias
+using InDataType       = FP16;
+
+using OutLayout  = ck::tensor_layout::convolution::GNHWK;
+using WeiLayout  = ck::tensor_layout::convolution::GKYXC;
+using BiasLayout = ck::Tuple<ck::tensor_layout::convolution::G_C>;
+using InLayout   = ck::tensor_layout::convolution::GNHWC;
+
+using OutElementOp = PassThrough;
+using WeiElementOp = PassThrough;
+using InElementOp  = ck::tensor_operation::element_wise::AddRelu;
+
+// clang-format off
+using DeviceConvInstance = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
+// ######| NDimSpatial|   ALayout|   BLayout|   DsLayout|  ELayout|       AData|       BData|     AccData|         CShuffle|                  DsData|      EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
+// ######|            |          |          |           |         |        Type|        Type|        Type|         DataType|                    Type|       Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
+// ######|            |          |          |           |         |            |            |            |                 |                        |           |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
+// ######|            |          |          |           |         |            |            |            |                 |                        |           |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
+         < NDimSpatial, OutLayout, WeiLayout, BiasLayout, InLayout, OutDataType, WeiDataType, AccDataType, CShuffleDataType, ck::Tuple<BiasDataType>, InDataType, OutElementOp, WeiElementOp,    InElementOp,  ConvBwdDataDefault,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>;
+// clang-format on
+
+#include "run_grouped_conv_bwd_data_bias_relu_example.inc"
+
+int main(int argc, char* argv[]) { return run_grouped_conv_bwd_data_bias_relu_example(argc, argv); }
diff --git a/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_fp16.cpp b/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_fp16.cpp
new file mode 100644
index 0000000000..fb688b6f3f
--- /dev/null
+++ b/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_fp16.cpp
@@ -0,0 +1,33 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using OutDataType      = FP16;
+using WeiDataType      = FP16;
+using AccDataType      = FP32;
+using CShuffleDataType = FP16;
+using DsDataType       = ck::Tuple<>;
+using InDataType       = FP16;
+
+using OutLayout = ck::tensor_layout::convolution::GNHWK;
+using WeiLayout = ck::tensor_layout::convolution::GKYXC;
+using DsLayout  = ck::Tuple<>;
+using InLayout  = ck::tensor_layout::convolution::GNHWC;
+
+using OutElementOp = PassThrough;
+using WeiElementOp = PassThrough;
+using InElementOp  = PassThrough;
+
+// clang-format off
+using DeviceConvInstance = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
+// ######| NDimSpatial|   ALayout|   BLayout|   DsLayout|  ELayout|       AData|       BData|     AccData|         CShuffle|       DsData|      EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
+// ######|            |          |          |           |         |        Type|        Type|        Type|         DataType|         Type|       Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
+// ######|            |          |          |           |         |            |            |            |                 |             |           |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
+// ######|            |          |          |           |         |            |            |            |                 |             |           |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
+         < NDimSpatial, OutLayout, WeiLayout,   DsLayout, InLayout, OutDataType, WeiDataType, AccDataType, CShuffleDataType,   DsDataType, InDataType, OutElementOp, WeiElementOp,    InElementOp,  ConvBwdDataDefault,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>;
+// clang-format on
+
+#include "run_grouped_conv_bwd_data_example.inc"
+
+int main(int argc, char* argv[]) { return run_grouped_conv_bwd_data_example(argc, argv); }
diff --git a/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc b/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc
new file mode 100644
index 0000000000..0f0b120cbc
--- /dev/null
+++ b/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc
@@ -0,0 +1,224 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+bool run_conv_bwd_data_bias_relu(const ExecutionConfig& config,
+                                 const ck::utils::conv::ConvParam& conv_params,
+                                 const HostTensorDescriptor& out_g_n_k_wos_desc,
+                                 const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                                 const HostTensorDescriptor& bias_g_n_c_wis_desc,
+                                 const HostTensorDescriptor& in_g_n_c_wis_desc,
+                                 const OutElementOp& out_element_op,
+                                 const WeiElementOp& wei_element_op,
+                                 const InElementOp& in_element_op)
+{
+    Tensor<OutDataType> out(out_g_n_k_wos_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<BiasDataType> bias(bias_g_n_c_wis_desc);
+    Tensor<InDataType> in_host(in_g_n_c_wis_desc);
+    Tensor<InDataType> in_device(in_g_n_c_wis_desc);
+
+    std::cout << "out: " << out.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "bias: " << bias.mDesc << std::endl;
+    std::cout << "in: " << in_host.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        bias.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-5, 5});
+        break;
+    default:
+        out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+        bias.GenerateTensorValue(GeneratorTensor_3<BiasDataType>{0.0, 1.0});
+    }
+
+    DeviceMem out_device_buf(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias.mDesc.GetElementSpaceSize());
+    DeviceMem in_device_buf(sizeof(InDataType) * in_device.mDesc.GetElementSpaceSize());
+
+    out_device_buf.ToDevice(out.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    bias_device_buf.ToDevice(bias.mData.data());
+
+    // reset input to zero
+    in_device_buf.SetZero();
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(out_g_n_k_wos_desc.GetLengths(), a_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), a_g_n_k_wos_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(bias_g_n_c_wis_desc.GetLengths(), d0_g_n_c_wis_lengths);
+    copy(bias_g_n_c_wis_desc.GetStrides(), d0_g_n_c_wis_strides);
+    copy(in_g_n_c_wis_desc.GetLengths(), e_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), e_g_n_c_wis_strides);
+    copy(conv_params.conv_filter_strides_, conv_filter_strides);
+    copy(conv_params.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_params.input_left_pads_, input_left_pads);
+    copy(conv_params.input_right_pads_, input_right_pads);
+
+    static_assert(std::is_default_constructible_v<DeviceConvInstance>);
+
+    // do conv
+    auto conv     = DeviceConvInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(
+        out_device_buf.GetDeviceBuffer(),
+        wei_device_buf.GetDeviceBuffer(),
+        std::array<const void*, 1>{bias_device_buf.GetDeviceBuffer()},
+        in_device_buf.GetDeviceBuffer(),
+        a_g_n_k_wos_lengths,
+        a_g_n_k_wos_strides,
+        b_g_k_c_xs_lengths,
+        b_g_k_c_xs_strides,
+        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{d0_g_n_c_wis_lengths},
+        std::array<std::array<ck::index_t, NDimSpatial + 3>, 1>{d0_g_n_c_wis_strides},
+        e_g_n_c_wis_lengths,
+        e_g_n_c_wis_strides,
+        conv_filter_strides,
+        conv_filter_dilations,
+        input_left_pads,
+        input_right_pads,
+        out_element_op,
+        wei_element_op,
+        in_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        std::cerr << "wrong! device_conv with the specified compilation parameters does "
+                     "not support this Conv problem"
+                  << std::endl;
+
+        return false;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = conv_params.GetFlops();
+    std::size_t num_btype = conv_params.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    if(config.do_verification)
+    {
+        // c doesn't physically exist, any layout is fine
+        Tensor<float> c_host(in_g_n_c_wis_desc);
+
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<NDimSpatial,
+                                                                         float,
+                                                                         WeiDataType,
+                                                                         OutDataType,
+                                                                         PassThrough,
+                                                                         WeiElementOp,
+                                                                         OutElementOp>();
+
+        auto ref_invoker = ref_conv.MakeInvoker();
+
+        auto ref_argument = ref_conv.MakeArgument(c_host,
+                                                  wei,
+                                                  out,
+                                                  conv_params.conv_filter_strides_,
+                                                  conv_params.conv_filter_dilations_,
+                                                  conv_params.input_left_pads_,
+                                                  conv_params.input_right_pads_,
+                                                  PassThrough{},
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        // TODO: implement elementwise operation for host
+        in_host.ForEach(
+            [&](auto&, auto idx) { in_element_op(in_host(idx), c_host(idx), bias(idx)); });
+
+        in_device_buf.FromDevice(in_device.mData.data());
+
+        return ck::utils::check_err(in_device, in_host);
+    }
+
+    return true;
+}
+
+int run_grouped_conv_bwd_data_bias_relu_example(int argc, char* argv[])
+{
+    namespace ctc = ck::tensor_layout::convolution;
+
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_params = DefaultConvParams;
+
+    if(!parse_cmd_args(argc, argv, config, conv_params))
+    {
+        return EXIT_FAILURE;
+    }
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    if(conv_params.num_dim_spatial_ != NDimSpatial)
+    {
+        std::cerr << "unsupported # of spatials dimensions" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    // output image: GNHWK
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+            conv_params);
+
+    // weight: GKYXC
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_params);
+
+    // input image bias: G_C
+    const auto bias_g_n_c_wis_desc = HostTensorDescriptor({conv_params.G_,
+                                                           conv_params.N_,
+                                                           conv_params.C_,
+                                                           conv_params.input_spatial_lengths_[0],
+                                                           conv_params.input_spatial_lengths_[1]},
+                                                          {
+                                                              conv_params.C_, // g
+                                                              0,              // n
+                                                              1,              // c
+                                                              0,              // hi
+                                                              0               // wi
+                                                          });
+
+    // input image: GNHWC
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_params);
+
+    return !run_conv_bwd_data_bias_relu(config,
+                                        conv_params,
+                                        out_g_n_k_wos_desc,
+                                        wei_g_k_c_xs_desc,
+                                        bias_g_n_c_wis_desc,
+                                        in_g_n_c_wis_desc,
+                                        wei_element_op,
+                                        out_element_op,
+                                        in_element_op);
+}
diff --git a/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc b/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc
new file mode 100644
index 0000000000..25678491ce
--- /dev/null
+++ b/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc
@@ -0,0 +1,190 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+bool run_conv_bwd_data(const ExecutionConfig& config,
+                       const ck::utils::conv::ConvParam& conv_params,
+                       const HostTensorDescriptor& out_g_n_k_wos_desc,
+                       const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                       const HostTensorDescriptor& in_g_n_c_wis_desc,
+                       const OutElementOp& out_element_op,
+                       const WeiElementOp& wei_element_op,
+                       const InElementOp& in_element_op)
+{
+    Tensor<OutDataType> out(out_g_n_k_wos_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<InDataType> in_host(in_g_n_c_wis_desc);
+    Tensor<InDataType> in_device(in_g_n_c_wis_desc);
+
+    std::cout << "out: " << out.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "in: " << in_host.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem out_device_buf(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem in_device_buf(sizeof(InDataType) * in_device.mDesc.GetElementSpaceSize());
+
+    out_device_buf.ToDevice(out.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+
+    // reset input to zero
+    in_device_buf.SetZero();
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(out_g_n_k_wos_desc.GetLengths(), a_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), a_g_n_k_wos_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(in_g_n_c_wis_desc.GetLengths(), e_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), e_g_n_c_wis_strides);
+    copy(conv_params.conv_filter_strides_, conv_filter_strides);
+    copy(conv_params.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_params.input_left_pads_, input_left_pads);
+    copy(conv_params.input_right_pads_, input_right_pads);
+
+    static_assert(std::is_default_constructible_v<DeviceConvInstance>);
+
+    // do conv
+    auto conv     = DeviceConvInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(out_device_buf.GetDeviceBuffer(),
+                                      wei_device_buf.GetDeviceBuffer(),
+                                      std::array<const void*, 0>{},
+                                      in_device_buf.GetDeviceBuffer(),
+                                      a_g_n_k_wos_lengths,
+                                      a_g_n_k_wos_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{},
+                                      std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{},
+                                      e_g_n_c_wis_lengths,
+                                      e_g_n_c_wis_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      out_element_op,
+                                      wei_element_op,
+                                      in_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        std::cerr << "wrong! device_conv with the specified compilation parameters does "
+                     "not support this Conv problem"
+                  << std::endl;
+
+        return false;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = conv_params.GetFlops();
+    std::size_t num_btype = conv_params.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    if(config.do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<NDimSpatial,
+                                                                         InDataType,
+                                                                         WeiDataType,
+                                                                         OutDataType,
+                                                                         PassThrough,
+                                                                         WeiElementOp,
+                                                                         OutElementOp>();
+
+        auto ref_invoker = ref_conv.MakeInvoker();
+
+        auto ref_argument = ref_conv.MakeArgument(in_host,
+                                                  wei,
+                                                  out,
+                                                  conv_params.conv_filter_strides_,
+                                                  conv_params.conv_filter_dilations_,
+                                                  conv_params.input_left_pads_,
+                                                  conv_params.input_right_pads_,
+                                                  PassThrough{},
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        in_device_buf.FromDevice(in_device.mData.data());
+
+        return ck::utils::check_err(in_device.mData, in_host.mData);
+    }
+
+    return true;
+}
+
+int run_grouped_conv_bwd_data_example(int argc, char* argv[])
+{
+    namespace ctc = ck::tensor_layout::convolution;
+
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_params = DefaultConvParams;
+
+    if(!parse_cmd_args(argc, argv, config, conv_params))
+    {
+        return EXIT_FAILURE;
+    }
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    if(conv_params.num_dim_spatial_ != NDimSpatial)
+    {
+        std::cerr << "unsupported # of spatials dimensions" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    // output image: GNHWK
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+            conv_params);
+
+    // weight: GKYXC
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_params);
+
+    // input image: GNHWC
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_params);
+
+    return !run_conv_bwd_data(config,
+                              conv_params,
+                              out_g_n_k_wos_desc,
+                              wei_g_k_c_xs_desc,
+                              in_g_n_c_wis_desc,
+                              wei_element_op,
+                              out_element_op,
+                              in_element_op);
+}
diff --git a/example/39_permute/CMakeLists.txt b/example/39_permute/CMakeLists.txt
new file mode 100644
index 0000000000..5b43de9725
--- /dev/null
+++ b/example/39_permute/CMakeLists.txt
@@ -0,0 +1,11 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_custom_target(example_permute)
+
+    add_example_executable(example_permute_1xHxW_fp16 permute_1xHxW_fp16.cpp)
+    add_example_executable(example_permute_NxHxW_fp16 permute_NxHxW_fp16.cpp)
+    add_example_executable(example_permute_HxWx4_fp16 permute_HxWx4_fp16.cpp)
+
+    add_dependencies(example_permute example_permute_1xHxW_fp16)
+    add_dependencies(example_permute example_permute_NxHxW_fp16)
+    add_dependencies(example_permute example_permute_HxWx4_fp16)
+endif()
diff --git a/example/39_permute/common.hpp b/example/39_permute/common.hpp
new file mode 100644
index 0000000000..54f3a78809
--- /dev/null
+++ b/example/39_permute/common.hpp
@@ -0,0 +1,456 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <cassert>
+#include <cstddef>
+#include <cstdlib>
+#include <cstring>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <type_traits>
+#include <utility>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/utility/type.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using F64 = double;
+
+struct Problem final
+{
+    static constexpr std::size_t NumDim = 3;
+
+    using Shape = std::array<std::size_t, NumDim>;
+    using Axes  = Shape;
+
+    Problem() = delete;
+
+    explicit Problem(const Shape& default_shape, const Axes& default_axes)
+        : shape(default_shape), axes(default_axes)
+    {
+    }
+
+    Shape shape;
+    Axes axes;
+};
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+namespace detail {
+
+template <typename Array, std::size_t Difference>
+struct enlarge_array_size;
+
+template <typename T, std::size_t Size, std::size_t Difference>
+struct enlarge_array_size<std::array<T, Size>, Difference>
+{
+    using type = std::array<T, Size + Difference>;
+};
+
+template <typename Array, std::size_t Difference>
+using enlarge_array_size_t = typename enlarge_array_size<Array, Difference>::type;
+
+template <typename Array>
+struct get_array_size;
+
+template <typename T, std::size_t Size>
+struct get_array_size<std::array<T, Size>> : std::integral_constant<std::size_t, Size>
+{
+};
+
+template <typename Array>
+inline constexpr std::size_t get_array_size_v = get_array_size<Array>::value;
+
+template <typename T, typename = void>
+struct is_iterator : std::false_type
+{
+};
+
+template <typename T>
+struct is_iterator<T,
+                   std::void_t<decltype(*std::declval<T>()),
+                               decltype(++std::declval<std::add_lvalue_reference_t<T>>()),
+                               decltype(std::declval<std::add_lvalue_reference_t<T>>()++)>>
+    : std::true_type
+{
+};
+
+template <typename T>
+inline constexpr bool is_iterator_v = is_iterator<T>::value;
+
+struct Placeholder final
+{
+    template <typename T>
+    constexpr inline operator T() const noexcept;
+};
+
+template <typename Iterator, typename = void>
+struct is_output_iterator : std::false_type
+{
+};
+
+template <typename Iterator>
+struct is_output_iterator<
+    Iterator,
+    std::void_t<decltype(*std::declval<Iterator>() = std::declval<Placeholder>())>>
+    : std::bool_constant<is_iterator_v<Iterator>>
+{
+};
+
+template <typename T>
+inline constexpr bool is_output_iterator_v = is_output_iterator<T>::value;
+
+template <typename Iterator, typename = void>
+struct is_bidirectional_iterator : std::false_type
+{
+};
+
+template <typename Iterator>
+struct is_bidirectional_iterator<
+    Iterator,
+    std::void_t<decltype(--std::declval<std::add_lvalue_reference_t<Iterator>>()),
+                decltype(std::declval<std::add_lvalue_reference_t<Iterator>>()--)>>
+    : std::bool_constant<is_iterator_v<Iterator>>
+{
+};
+
+template <typename Iterator>
+inline constexpr bool is_bidirectional_iterator_v = is_bidirectional_iterator<Iterator>::value;
+
+template <typename Iterator, typename = void>
+struct is_random_access_iterator : std::false_type
+{
+};
+
+template <typename Iterator>
+struct is_random_access_iterator<Iterator,
+                                 std::void_t<decltype(std::declval<Iterator>() + 1),
+                                             decltype(std::declval<Iterator>() - 1),
+                                             decltype(std::declval<Iterator>()[1])>>
+    : std::bool_constant<is_iterator_v<Iterator>>
+{
+};
+
+template <typename Iterator>
+inline constexpr bool is_random_access_iterator_v = is_random_access_iterator<Iterator>::value;
+
+template <typename T, typename = void>
+struct is_range : std::false_type
+{
+};
+
+template <typename T>
+struct is_range<T,
+                std::void_t<decltype(begin(std::declval<T>())),
+                            decltype(end(std::declval<T>())),
+                            decltype(begin(std::declval<T>()) != end(std::declval<T>()))>>
+    : std::bool_constant<is_iterator_v<ck::remove_cvref_t<decltype(begin(std::declval<T>()))>>>
+{
+};
+
+template <typename T>
+inline constexpr bool is_range_v = is_range<T>::value;
+
+template <typename Range, typename = void>
+struct is_sized_range : std::false_type
+{
+};
+
+template <typename Range>
+struct is_sized_range<Range, std::void_t<decltype(size(std::declval<Range>()))>>
+    : std::bool_constant<is_range_v<Range>>
+{
+};
+
+template <typename Range>
+inline constexpr bool is_sized_range_v = is_sized_range<Range>::value;
+
+template <typename Range, typename = void>
+struct is_bidirectional_range : std::false_type
+{
+};
+
+template <typename Range>
+struct is_bidirectional_range<Range, std::void_t<>>
+    : std::bool_constant<
+          is_range_v<Range> &&
+          is_bidirectional_iterator_v<ck::remove_cvref_t<decltype(begin(std::declval<Range>()))>>>
+{
+};
+
+template <typename Range>
+inline constexpr bool is_bidirectional_range_v = is_bidirectional_range<Range>::value;
+
+template <typename Range, typename = void>
+struct is_random_access_range : std::false_type
+{
+};
+
+template <typename Range>
+struct is_random_access_range<Range, std::void_t<>>
+    : std::bool_constant<
+          is_range_v<Range> &&
+          is_random_access_iterator_v<ck::remove_cvref_t<decltype(begin(std::declval<Range>()))>>>
+{
+};
+
+template <typename Range>
+inline constexpr bool is_random_access_range_v = is_random_access_range<Range>::value;
+
+template <typename Range>
+class to_array_proxy
+{
+    static_assert(is_range_v<Range>);
+
+    public:
+    explicit to_array_proxy(const Range& source) noexcept : source_(source) {}
+
+    template <typename T, std::size_t Size>
+    operator std::array<T, Size>() const
+    {
+        std::array<T, Size> destination;
+
+        std::copy_n(std::begin(source_),
+                    std::min<std::size_t>(Size, std::size(source_)),
+                    std::begin(destination));
+
+        return destination;
+    }
+
+    private:
+    const Range& source_;
+};
+
+} // namespace detail
+
+template <typename Range>
+inline auto to_array(Range& range) noexcept
+    -> std::enable_if_t<detail::is_range_v<Range>,
+                        detail::to_array_proxy<ck::remove_cvref_t<Range>>>
+{
+    return detail::to_array_proxy<ck::remove_cvref_t<Range>>{range};
+}
+
+template <typename Axes>
+inline auto is_valid_axes(const Axes& axes)
+    -> std::enable_if_t<detail::is_random_access_range_v<Axes>, bool>
+{
+    using std::empty;
+    if(empty(axes))
+    {
+        return false;
+    }
+
+    using std::begin, std::end;
+    std::vector<std::size_t> sorted_axes(begin(axes), end(axes));
+
+    std::sort(begin(sorted_axes), end(sorted_axes));
+    const auto last = std::unique(begin(sorted_axes), end(sorted_axes));
+
+    return (last == end(sorted_axes)) && (*begin(sorted_axes) == 0) &&
+           (*std::prev(last) == size(axes) - 1);
+}
+
+template <typename Shape>
+inline auto is_valid_shape(const Shape& shape) -> std::enable_if_t<detail::is_range_v<Shape>, bool>
+{
+    static_assert(std::is_unsigned_v<ck::remove_cvref_t<decltype(*std::begin(shape))>>);
+
+    using std::begin, std::end;
+    using std::empty;
+    return !empty(shape) && std::all_of(begin(shape), end(shape), [](auto dim) { return 0 < dim; });
+}
+
+template <typename Shape, typename Indices>
+inline auto is_valid_indices(const Shape& shape, const Indices& indices)
+    -> std::enable_if_t<detail::is_sized_range_v<Shape> && detail::is_sized_range_v<Indices>, bool>
+{
+    static_assert(std::is_unsigned_v<ck::remove_cvref_t<decltype(*std::begin(indices))>>);
+
+    if(!is_valid_shape(shape))
+    {
+        return false;
+    }
+
+    using std::empty;
+    if(empty(indices))
+    {
+        return false;
+    }
+
+    using std::size;
+    if(size(shape) != size(indices))
+    {
+        return false;
+    }
+
+    using std::begin, std::end;
+
+    auto dim = begin(shape);
+    auto idx = begin(indices);
+    for(; dim != end(shape) && idx != end(indices); ++dim, ++idx)
+    {
+        if(*dim <= *idx)
+        {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+template <std::size_t Size>
+std::array<std::size_t, Size> transpose(const std::array<std::size_t, Size>& shape,
+                                        const std::array<std::size_t, Size>& axes)
+{
+    assert(is_valid_shape(shape) && is_valid_axes(axes));
+
+    std::array<std::size_t, Size> transposed;
+    auto iter = std::begin(transposed);
+    for(const auto axis : axes)
+    {
+        *iter++ = shape[axis];
+    }
+
+    return transposed;
+}
+
+auto extend_shape(const Problem::Shape& shape, std::size_t new_dim)
+{
+    detail::enlarge_array_size_t<Problem::Shape, 1> extended_shape;
+
+    using std::begin, std::end;
+
+    ck::ranges::copy(shape, begin(extended_shape));
+    extended_shape.back() = new_dim;
+
+    return extended_shape;
+}
+
+auto extend_axes(const Problem::Axes& axes)
+{
+    detail::enlarge_array_size_t<Problem::Axes, 1> extended_axes;
+
+    using std::begin, std::end;
+
+    ck::ranges::copy(axes, begin(extended_axes));
+    extended_axes.back() = detail::get_array_size_v<Problem::Axes>;
+
+    return extended_axes;
+}
+
+template <typename Shape, typename Indices>
+auto advance_indices(const Shape& shape, Indices& indices) -> std::enable_if_t<
+    detail::is_bidirectional_range_v<Shape> && detail::is_sized_range_v<Shape> &&
+        detail::is_bidirectional_range_v<Indices> && detail::is_sized_range_v<Indices>,
+    bool>
+{
+    using std::size;
+    if(!(is_valid_shape(shape) && is_valid_indices(shape, indices) && size(shape) == size(indices)))
+    {
+        return false;
+    }
+
+    bool carry = true;
+
+    using std::rbegin, std::rend;
+    auto dim = rbegin(shape);
+    auto idx = rbegin(indices);
+    for(; carry && dim != rend(shape) && idx != rend(indices); ++dim, ++idx)
+    {
+        *idx  = (*idx + carry);
+        carry = ((*idx == *dim) ? (*idx = 0, true) : false);
+    }
+
+    return !carry;
+}
+
+template <typename Src, typename Axes, typename Functor, typename Dest>
+auto host_permute(const Tensor<Src>& src, const Axes& axes, Functor functor, Tensor<Dest>& dest)
+    -> std::enable_if_t<detail::is_random_access_range_v<Axes> && detail::is_sized_range_v<Axes> &&
+                            std::is_invocable_v<Functor,
+                                                std::add_lvalue_reference_t<Dest>,
+                                                std::add_lvalue_reference_t<Src>>,
+                        bool>
+{
+    const auto& shape            = src.mDesc.GetLengths();
+    const auto& transposed_shape = dest.mDesc.GetLengths();
+    if(!(is_valid_shape(shape) && is_valid_shape(transposed_shape)))
+    {
+        return false;
+    }
+
+    using std::size;
+    if(!is_valid_axes(axes))
+    {
+        return false;
+    }
+
+    static_assert(detail::is_sized_range_v<ck::remove_cvref_t<decltype(shape)>> &&
+                  detail::is_sized_range_v<ck::remove_cvref_t<decltype(transposed_shape)>>);
+
+    if(size(shape) != size(transposed_shape))
+    {
+        return false;
+    }
+
+    static_assert(detail::is_random_access_range_v<ck::remove_cvref_t<decltype(shape)>> &&
+                  detail::is_random_access_range_v<ck::remove_cvref_t<decltype(transposed_shape)>>);
+    {
+        for(std::size_t idx = 0; idx < size(shape); ++idx)
+        {
+            if(transposed_shape[idx] != shape[axes[idx]])
+            {
+                return false;
+            }
+        }
+    }
+
+    std::vector<std::size_t> indices(size(shape), 0);
+    if(!is_valid_indices(shape, indices))
+    {
+        return false;
+    }
+
+    switch(size(shape))
+    {
+    case 3: {
+        do
+        {
+            Dest output = 0;
+            functor(output, src(indices[0], indices[1], indices[2]));
+            dest(indices[axes[0]], indices[axes[1]], indices[axes[2]]) = output;
+        } while(advance_indices(shape, indices));
+    }
+    break;
+    case 4: {
+        do
+        {
+            Dest output = 0;
+            functor(output, src(indices[0], indices[1], indices[2], indices[3]));
+            dest(indices[axes[0]], indices[axes[1]], indices[axes[2]], indices[axes[3]]) = output;
+        } while(advance_indices(shape, indices));
+    }
+    break;
+    default: return false;
+    }
+
+    return true;
+}
diff --git a/example/39_permute/permute_1xHxW_fp16.cpp b/example/39_permute/permute_1xHxW_fp16.cpp
new file mode 100644
index 0000000000..7336c3b631
--- /dev/null
+++ b/example/39_permute/permute_1xHxW_fp16.cpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using InDataType  = F16;
+using OutDataType = F16;
+
+// clang-format off
+using DevicePermuteInstance = ck::tensor_operation::device::DevicePermuteImpl
+// ######| NumDim|     InData|     OutData| Elementwise| Block|  NPer|  HPer|  WPer|   InBlock|      InBlockTransfer|           InBlockTransfer|       Src|       Dst|             Src|             Dst|
+// ######|       |       Type|        Type|   Operation|  Size| Block| Block| Block| LdsExtraW| ThreadClusterLengths| ThreadClusterArrangeOrder| VectorDim| VectorDim| ScalarPerVector| ScalarPerVector|
+// ######|       |           |            |            |      |      |      |      |          |                     |                          |          |          |                |                |
+// ######|       |           |            |            |      |      |      |      |          |                     |                          |          |          |                |                |
+         <      3, InDataType, OutDataType, PassThrough,   256,     1,    32,    32,         3,         S<1, 32,  8>,                S<0, 1, 2>,         2,         1,               2,               1>;
+// clang-format on
+
+#include "run_permute_element_example.inc"
+
+int main() { return !run_permute_element_example({1, 32000, 80}, {0, 2, 1}); }
diff --git a/example/39_permute/permute_HxWx4_fp16.cpp b/example/39_permute/permute_HxWx4_fp16.cpp
new file mode 100644
index 0000000000..6c24919ded
--- /dev/null
+++ b/example/39_permute/permute_HxWx4_fp16.cpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using DataType   = F16;
+using BundleType = F64;
+
+static_assert(sizeof(BundleType) % sizeof(DataType) == 0);
+
+// clang-format off
+using DevicePermuteInstance = ck::tensor_operation::device::DevicePermuteImpl
+// ######| NumDim|     InData|     OutData| Elementwise| Block|  NPer|  HPer|  WPer|   InBlock|      InBlockTransfer|           InBlockTransfer|       Src|       Dst|             Src|             Dst|
+// ######|       |       Type|        Type|   Operation|  Size| Block| Block| Block| LdsExtraW| ThreadClusterLengths| ThreadClusterArrangeOrder| VectorDim| VectorDim| ScalarPerVector| ScalarPerVector|
+// ######|       |           |            |            |      |      |      |      |          |                     |                          |          |          |                |                |
+// ######|       |           |            |            |      |      |      |      |          |                     |                          |          |          |                |                |
+         <       3, BundleType, BundleType, PassThrough,   256,     1,    32,    32,         5,         S<1, 32,  8>,                S<0, 1, 2>,         2,         1,               4,               1>;
+// clang-format on
+
+#include "run_permute_bundle_example.inc"
+
+int main() { return !run_permute_bundle_example({1, 80, 32000}, {0, 2, 1}); }
diff --git a/example/39_permute/permute_NxHxW_fp16.cpp b/example/39_permute/permute_NxHxW_fp16.cpp
new file mode 100644
index 0000000000..3551d2a7c8
--- /dev/null
+++ b/example/39_permute/permute_NxHxW_fp16.cpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using InDataType  = F16;
+using OutDataType = F16;
+
+// clang-format off
+using DevicePermuteInstance = ck::tensor_operation::device::DevicePermuteImpl
+// ######| NumDim|     InData|     OutData| Elementwise| Block|  NPer|  HPer|  WPer|   InBlock|      InBlockTransfer|           InBlockTransfer|       Src|       Dst|             Src|             Dst|
+// ######|       |       Type|        Type|   Operation|  Size| Block| Block| Block| LdsExtraW| ThreadClusterLengths| ThreadClusterArrangeOrder| VectorDim| VectorDim| ScalarPerVector| ScalarPerVector|
+// ######|       |           |            |            |      |      |      |      |          |                     |                          |          |          |                |                |
+// ######|       |           |            |            |      |      |      |      |          |                     |                          |          |          |                |                |
+         <      3, InDataType, OutDataType, PassThrough,   128,     4,    16,     8,         6,          S<2, 16, 4>,                S<0, 1, 2>,         2,         1,               2,               1>;
+// clang-format on
+
+#include "run_permute_element_example.inc"
+
+int main() { return !run_permute_element_example({121, 768, 80}, {0, 2, 1}); }
diff --git a/example/39_permute/run_permute_bundle_example.inc b/example/39_permute/run_permute_bundle_example.inc
new file mode 100644
index 0000000000..2c19872922
--- /dev/null
+++ b/example/39_permute/run_permute_bundle_example.inc
@@ -0,0 +1,78 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+bool run_permute_bundle(const Problem& problem)
+{
+    const auto& input_bundle_shape = problem.shape;
+    const auto& input_bundle_axes  = problem.axes;
+
+    const auto output_bundle_shape = transpose(input_bundle_shape, input_bundle_axes);
+
+    Tensor<BundleType> input_bundle_tensor(input_bundle_shape);
+    Tensor<BundleType> output_bundle_tensor(output_bundle_shape);
+
+    // initialize tensor by assigning DataType values
+    ck::utils::FillUniformDistribution<DataType>{-1.f, 1.f}(input_bundle_tensor.AsSpan<DataType>());
+
+    DeviceMem input_device_buf(input_bundle_tensor.GetElementSpaceSizeInBytes());
+    DeviceMem output_device_buf(output_bundle_tensor.GetElementSpaceSizeInBytes());
+
+    using std::data;
+    input_device_buf.ToDevice(data(input_bundle_tensor));
+
+    static_assert(std::is_default_constructible_v<DevicePermuteInstance>);
+
+    auto permute  = DevicePermuteInstance{};
+    auto argument = permute.MakeArgument(to_array(input_bundle_shape),
+                                         to_array(input_bundle_tensor.GetStrides()),
+                                         to_array(output_bundle_shape),
+                                         to_array(output_bundle_tensor.GetStrides()),
+                                         input_device_buf.GetDeviceBuffer(),
+                                         output_device_buf.GetDeviceBuffer(),
+                                         PassThrough{});
+
+    if(!permute.IsSupportedArgument(argument))
+    {
+        std::cerr << "The runtime parameters seems not supported by the device instance, exiting!"
+                  << std::endl;
+        return false;
+    };
+
+    auto invoker   = permute.MakeInvoker();
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, true});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    output_device_buf.FromDevice(data(output_bundle_tensor));
+
+    constexpr std::size_t NumElemsInBundle = sizeof(BundleType) / sizeof(DataType);
+
+    // extend tensor shape from [N, H, W] to [N, H, W, NumElemsInBundle]
+    //               axes  from [0, 2, 1] to [0, 2, 1, 3]
+    const auto input_shape = extend_shape(input_bundle_shape, NumElemsInBundle);
+    const auto input_axes  = extend_axes(input_bundle_axes);
+
+    using std::begin;
+
+    Tensor<DataType> input_tensor(input_shape);
+    ck::ranges::copy(input_bundle_tensor.AsSpan<const DataType>(), begin(input_tensor));
+
+    Tensor<DataType> output_tensor(transpose(input_shape, input_axes));
+    if(!host_permute(input_tensor, input_axes, PassThrough{}, output_tensor))
+    {
+        return false;
+    }
+
+    return ck::utils::check_err(output_bundle_tensor.AsSpan<const DataType>(),
+                                output_tensor.AsSpan<const DataType>(),
+                                "Error: incorrect results in output tensor",
+                                1e-6,
+                                1e-6);
+}
+
+bool run_permute_bundle_example(const Problem::Shape& shape, const Problem::Axes& axes)
+{
+    return run_permute_bundle(Problem{shape, axes});
+}
diff --git a/example/39_permute/run_permute_element_example.inc b/example/39_permute/run_permute_element_example.inc
new file mode 100644
index 0000000000..3587134456
--- /dev/null
+++ b/example/39_permute/run_permute_element_example.inc
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+bool run_permute_element(const Problem& problem)
+{
+    const auto& input_shape = problem.shape;
+    const auto& input_axes  = problem.axes;
+
+    const auto output_shape = transpose(input_shape, input_axes);
+
+    Tensor<InDataType> input_tensor(input_shape);
+    Tensor<OutDataType> output_tensor(output_shape);
+
+    ck::utils::FillUniformDistribution<InDataType>{-1.f, 1.f}(input_tensor);
+
+    DeviceMem input_device_buf(input_tensor.GetElementSpaceSizeInBytes());
+    DeviceMem output_device_buf(output_tensor.GetElementSpaceSizeInBytes());
+
+    using std::data;
+    input_device_buf.ToDevice(data(input_tensor));
+
+    static_assert(std::is_default_constructible_v<DevicePermuteInstance>);
+
+    auto permute  = DevicePermuteInstance{};
+    auto argument = permute.MakeArgument(to_array(input_shape),
+                                         to_array(input_tensor.GetStrides()),
+                                         to_array(output_shape),
+                                         to_array(output_tensor.GetStrides()),
+                                         input_device_buf.GetDeviceBuffer(),
+                                         output_device_buf.GetDeviceBuffer(),
+                                         PassThrough{});
+
+    if(!permute.IsSupportedArgument(argument))
+    {
+        std::cerr << "The runtime parameters seems not supported by the device instance, exiting!"
+                  << std::endl;
+        return false;
+    };
+
+    auto invoker   = permute.MakeInvoker();
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, true});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    output_device_buf.FromDevice(data(output_tensor));
+
+    Tensor<OutDataType> output_tensor_host(output_shape);
+    if(!host_permute(input_tensor, input_axes, PassThrough{}, output_tensor_host))
+    {
+        return false;
+    }
+
+    return ck::utils::check_err(output_tensor.AsSpan<const OutDataType>(),
+                                output_tensor_host.AsSpan<const OutDataType>(),
+                                "Error: incorrect results in output tensor",
+                                1e-6,
+                                1e-6);
+}
+
+bool run_permute_element_example(const Problem::Shape& shape, const Problem::Axes& axes)
+{
+    return run_permute_element(Problem{shape, axes});
+}
diff --git a/example/40_conv2d_fwd_quantization/CMakeLists.txt b/example/40_conv2d_fwd_quantization/CMakeLists.txt
new file mode 100644
index 0000000000..55464957ac
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/CMakeLists.txt
@@ -0,0 +1,28 @@
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_example_executable(example_conv2d_fwd_xdl_perlayer_quantization_int8 conv2d_fwd_xdl_perlayer_quantization_int8.cpp)
+   add_example_executable(example_conv2d_fwd_xdl_perchannel_quantization_int8 conv2d_fwd_xdl_perchannel_quantization_int8.cpp)
+   add_example_executable(example_conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8 conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp)
+   add_example_executable(example_conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8 conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp)
+   set(target 1)
+ endif()
+endforeach()
+
+  if(DL_KERNELS)
+  # Conv perlayer quantization
+  add_example_executable(example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp)
+  # Conv perchannel quantization
+  add_example_executable(example_conv2d_fwd_dl_perchannel_quantization_int8 conv2d_fwd_dl_perchannel_quantization_int8.cpp)
+  # Conv + bias + relu perlayer quantization
+  add_example_executable(example_conv2d_fwd_dl_bias_relu_perlayer_quantization_int8 conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp)
+  # Conv + bias + relu perchannel quantization
+  add_example_executable(example_conv2d_fwd_dl_bias_relu_perchannel_quantization_int8 conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp)
+  # Conv + bias + tanh perlayer quantization
+  add_example_executable(example_conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8 conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp)
+  # Conv + bias + tanh perchannel quantization
+  add_example_executable(example_conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8 conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp)
+  endif()
+endif()
\ No newline at end of file
diff --git a/example/40_conv2d_fwd_quantization/common.hpp b/example/40_conv2d_fwd_quantization/common.hpp
new file mode 100644
index 0000000000..266b09145c
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/common.hpp
@@ -0,0 +1,18 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp
new file mode 100644
index 0000000000..4573c68658
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp
@@ -0,0 +1,85 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using BiasDataType         = int32_t;
+using RequantScaleDataType = float;
+using AccDataType          = int32_t;
+using OutDataType          = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename BiasLayout,
+          typename RequantScaleLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<
+        NDimSpatial,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<BiasDataType, RequantScaleDataType>,
+        OutDataType,
+        AccDataType,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout, RequantScaleLayout>,
+        OutLayout,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,            // ConvForwardSpecialization
+        GemmSpec,            // GemmSpecialization
+        256,                 // BlockSize
+        128,                 // MPerBlock
+        128,                 // NPerBlock
+        16,                  // K0PerBlock
+        4,                   // K1
+        4,                   // M1PerThread
+        4,                   // N1PerThread
+        1,                   // KPerThread
+        S<8, 2>,             // M1N1ThreadClusterM1Xs
+        S<8, 2>,             // M1N1ThreadClusterN1Xs
+        S<8, 1, 1, 4>,       // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
+        S<2, 1, 128, 1>,     // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // ABlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
+        S<8, 1, 1, 4>,       // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
+        S<2, 1, 128, 1>,     // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // BBlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
+        S<0, 1, 2, 3, 4, 5>, // CThreadTransferSrcDstAccessOrder
+        5,                   // CThreadTransferSrcDstVectorDim
+        4>;                  // CThreadTransferDstScalarPerVector
+
+#include "run_conv2d_fwd_bias_perchannel_quantization_example.inc"
+
+int main()
+{
+    const auto out_element_op = OutElementOp{ActivationOp{}};
+    run_conv2d_fwd_bias_perchannel_quantization_example(out_element_op);
+};
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp
new file mode 100644
index 0000000000..005f6263fd
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+
+using InDataType   = int8_t;
+using WeiDataType  = int8_t;
+using BiasDataType = int32_t;
+using AccDataType  = int32_t;
+using OutDataType  = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename BiasLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<
+        NDimSpatial,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<BiasDataType>,
+        OutDataType,
+        AccDataType,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout>,
+        OutLayout,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,            // ConvForwardSpecialization
+        GemmSpec,            // GemmSpecialization
+        256,                 // BlockSize
+        128,                 // MPerBlock
+        128,                 // NPerBlock
+        16,                  // K0PerBlock
+        4,                   // K1
+        4,                   // M1PerThread
+        4,                   // N1PerThread
+        1,                   // KPerThread
+        S<8, 2>,             // M1N1ThreadClusterM1Xs
+        S<8, 2>,             // M1N1ThreadClusterN1Xs
+        S<8, 1, 1, 4>,       // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
+        S<2, 1, 128, 1>,     // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // ABlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
+        S<8, 1, 1, 4>,       // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
+        S<2, 1, 128, 1>,     // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // BBlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
+        S<0, 1, 2, 3, 4, 5>, // CThreadTransferSrcDstAccessOrder
+        5,                   // CThreadTransferSrcDstVectorDim
+        4>;                  // CThreadTransferDstScalarPerVector
+
+#include "run_conv2d_fwd_bias_perlayer_quantization_example.inc"
+
+int main()
+{
+    float requant_scale       = 0.5f;
+    const auto out_element_op = OutElementOp{requant_scale, ActivationOp{}};
+    run_conv2d_fwd_bias_perlayer_quantization_example(out_element_op);
+}
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp
new file mode 100644
index 0000000000..62e5e583de
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp
@@ -0,0 +1,87 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using BiasDataType         = int32_t;
+using RequantScaleDataType = float;
+using AccDataType          = int32_t;
+using OutDataType          = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::TanH;
+using OutElementOp =
+    ck::tensor_operation::element_wise::Add_Mul2_Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename BiasLayout,
+          typename RequantScaleLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<
+        NDimSpatial,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<BiasDataType, RequantScaleDataType>,
+        OutDataType,
+        AccDataType,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout, RequantScaleLayout>,
+        OutLayout,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,            // ConvForwardSpecialization
+        GemmSpec,            // GemmSpecialization
+        256,                 // BlockSize
+        128,                 // MPerBlock
+        128,                 // NPerBlock
+        16,                  // K0PerBlock
+        4,                   // K1
+        4,                   // M1PerThread
+        4,                   // N1PerThread
+        1,                   // KPerThread
+        S<8, 2>,             // M1N1ThreadClusterM1Xs
+        S<8, 2>,             // M1N1ThreadClusterN1Xs
+        S<8, 1, 1, 4>,       // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
+        S<2, 1, 128, 1>,     // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // ABlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
+        S<8, 1, 1, 4>,       // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
+        S<2, 1, 128, 1>,     // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // BBlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
+        S<0, 1, 2, 3, 4, 5>, // CThreadTransferSrcDstAccessOrder
+        5,                   // CThreadTransferSrcDstVectorDim
+        4>;                  // CThreadTransferDstScalarPerVector
+
+#include "run_conv2d_fwd_bias_perchannel_quantization_example.inc"
+
+int main()
+{
+    float scale_z_inv         = 0.5f;
+    const auto out_element_op = OutElementOp{scale_z_inv, ActivationOp{}};
+    run_conv2d_fwd_bias_perchannel_quantization_example(out_element_op);
+};
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp
new file mode 100644
index 0000000000..ef98fe7e4f
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp
@@ -0,0 +1,85 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+
+using InDataType   = int8_t;
+using WeiDataType  = int8_t;
+using BiasDataType = int32_t;
+using AccDataType  = int32_t;
+using OutDataType  = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::TanH;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Mul_Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename BiasLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<
+        NDimSpatial,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<BiasDataType>,
+        OutDataType,
+        AccDataType,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout>,
+        OutLayout,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,            // ConvForwardSpecialization
+        GemmSpec,            // GemmSpecialization
+        256,                 // BlockSize
+        128,                 // MPerBlock
+        128,                 // NPerBlock
+        16,                  // K0PerBlock
+        4,                   // K1
+        4,                   // M1PerThread
+        4,                   // N1PerThread
+        1,                   // KPerThread
+        S<8, 2>,             // M1N1ThreadClusterM1Xs
+        S<8, 2>,             // M1N1ThreadClusterN1Xs
+        S<8, 1, 1, 4>,       // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
+        S<2, 1, 128, 1>,     // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // ABlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
+        S<8, 1, 1, 4>,       // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
+        S<2, 1, 128, 1>,     // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // BBlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
+        S<0, 1, 2, 3, 4, 5>, // CThreadTransferSrcDstAccessOrder
+        5,                   // CThreadTransferSrcDstVectorDim
+        4>;                  // CThreadTransferDstScalarPerVector
+
+#include "run_conv2d_fwd_bias_perlayer_quantization_example.inc"
+
+int main()
+{
+    float scale_acc           = 0.5f;
+    float scale_z_inv         = 0.5f;
+    const auto out_element_op = OutElementOp{scale_z_inv, scale_acc, ActivationOp{}};
+    run_conv2d_fwd_bias_perlayer_quantization_example(out_element_op);
+}
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp
new file mode 100644
index 0000000000..e524ddb2b2
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using RequantScaleDataType = float;
+using AccDataType          = int32_t;
+using OutDataType          = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename RequantScaleLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<
+        NDimSpatial,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<RequantScaleDataType>,
+        OutDataType,
+        AccDataType,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<RequantScaleLayout>,
+        OutLayout,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,            // ConvForwardSpecialization
+        GemmSpec,            // GemmSpecialization
+        256,                 // BlockSize
+        128,                 // MPerBlock
+        128,                 // NPerBlock
+        16,                  // K0PerBlock
+        4,                   // K1
+        4,                   // M1PerThread
+        4,                   // N1PerThread
+        1,                   // KPerThread
+        S<8, 2>,             // M1N1ThreadClusterM1Xs
+        S<8, 2>,             // M1N1ThreadClusterN1Xs
+        S<8, 1, 1, 4>,       // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
+        S<2, 1, 128, 1>,     // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // ABlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
+        S<8, 1, 1, 4>,       // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
+        S<2, 1, 128, 1>,     // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // BBlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
+        S<0, 1, 2, 3, 4, 5>, // CThreadTransferSrcDstAccessOrder
+        5,                   // CThreadTransferSrcDstVectorDim
+        4>;                  // CThreadTransferDstScalarPerVector
+
+#include "run_conv2d_fwd_perchannel_quantization_example.inc"
+
+int main()
+{
+    const auto out_element_op = OutElementOp{ActivationOp{}};
+    run_conv2d_fwd_perchannel_quantization_example(out_element_op);
+}
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perlayer_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perlayer_quantization_int8.cpp
new file mode 100644
index 0000000000..d29a3143c0
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perlayer_quantization_int8.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+
+using InDataType  = int8_t;
+using WeiDataType = int8_t;
+using AccDataType = int32_t;
+using OutDataType = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<
+        NDimSpatial,
+        InDataType,
+        WeiDataType,
+        ck::Tuple<>,
+        OutDataType,
+        AccDataType,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,            // ConvForwardSpecialization
+        GemmSpec,            // GemmSpecialization
+        256,                 // BlockSize
+        128,                 // MPerBlock
+        128,                 // NPerBlock
+        16,                  // K0PerBlock
+        4,                   // K1
+        4,                   // M1PerThread
+        4,                   // N1PerThread
+        1,                   // KPerThread
+        S<8, 2>,             // M1N1ThreadClusterM1Xs
+        S<8, 2>,             // M1N1ThreadClusterN1Xs
+        S<8, 1, 1, 4>,       // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
+        S<2, 1, 128, 1>,     // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // ABlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
+        S<1, 2, 0, 3>,       // ABlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
+        S<8, 1, 1, 4>,       // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
+        S<2, 1, 128, 1>,     // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferThreadClusterArrangeOrder
+        S<1, 2, 0, 3>,       // BBlockTransferSrcAccessOrder
+        S<4, 1, 1, 4>,       // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
+        S<1, 2, 0, 3>,       // BBlockTransferSrcVectorTensorContiguousDimOrder
+        S<1, 1, 1, 4>,       // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
+        S<0, 1, 2, 3, 4, 5>, // CThreadTransferSrcDstAccessOrder
+        5,                   // CThreadTransferSrcDstVectorDim
+        4>;                  // CThreadTransferDstScalarPerVector
+
+#include "run_conv2d_fwd_perlayer_quantization_example.inc"
+
+int main()
+{
+    float requant_scale       = 0.5f;
+    const auto out_element_op = OutElementOp{requant_scale, ActivationOp{}};
+    run_conv2d_fwd_perlayer_quantization_example(out_element_op);
+}
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp
new file mode 100644
index 0000000000..06c839e4e2
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp
@@ -0,0 +1,89 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using BiasDataType         = int32_t;
+using RequantScaleDataType = float;
+using AccDataType          = int32_t;
+using CShuffleDataType     = AccDataType;
+using OutDataType          = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename BiasLayout,
+          typename RequantScaleLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout, RequantScaleLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<BiasDataType, RequantScaleDataType>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        64,          // KPerBlock
+        16,          // AK1
+        16,          // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        16,          // ABlockTransferSrcScalarPerVector
+        16,          // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        16,          // BBlockTransferSrcScalarPerVector
+        16,          // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 64, 1, 4>,
+        8>;
+
+#include "run_conv2d_fwd_bias_perchannel_quantization_example.inc"
+
+int main()
+{
+    const auto out_element_op = OutElementOp{ActivationOp{}};
+    run_conv2d_fwd_bias_perchannel_quantization_example(out_element_op);
+};
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp
new file mode 100644
index 0000000000..7a9b42d39f
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp
@@ -0,0 +1,88 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+
+using InDataType       = int8_t;
+using WeiDataType      = int8_t;
+using BiasDataType     = int32_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = AccDataType;
+using OutDataType      = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename BiasLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<BiasLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<BiasDataType>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        64,          // KPerBlock
+        16,          // AK1
+        16,          // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        16,          // ABlockTransferSrcScalarPerVector
+        16,          // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        16,          // BBlockTransferSrcScalarPerVector
+        16,          // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 64, 1, 4>,
+        8>;
+
+#include "run_conv2d_fwd_bias_perlayer_quantization_example.inc"
+
+int main()
+{
+    float requant_scale       = 0.5f;
+    const auto out_element_op = OutElementOp{requant_scale, ActivationOp{}};
+    run_conv2d_fwd_bias_perlayer_quantization_example(out_element_op);
+}
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perchannel_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perchannel_quantization_int8.cpp
new file mode 100644
index 0000000000..3495636297
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perchannel_quantization_int8.cpp
@@ -0,0 +1,87 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+
+using InDataType           = int8_t;
+using WeiDataType          = int8_t;
+using RequantScaleDataType = float;
+using AccDataType          = int32_t;
+using CShuffleDataType     = AccDataType;
+using OutDataType          = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = ck::tensor_operation::element_wise::Relu;
+using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename RequantScaleLayout,
+          typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<RequantScaleLayout>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<RequantScaleDataType>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        64,          // KPerBlock
+        16,          // AK1
+        16,          // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        16,          // ABlockTransferSrcScalarPerVector
+        16,          // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        16,          // BBlockTransferSrcScalarPerVector
+        16,          // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 64, 1, 4>,
+        8>;
+
+#include "run_conv2d_fwd_perchannel_quantization_example.inc"
+
+int main()
+{
+    const auto out_element_op = OutElementOp{ActivationOp{}};
+    run_conv2d_fwd_perchannel_quantization_example(out_element_op);
+}
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp
new file mode 100644
index 0000000000..2611337254
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+
+using InDataType       = int8_t;
+using WeiDataType      = int8_t;
+using AccDataType      = int32_t;
+using CShuffleDataType = AccDataType;
+using OutDataType      = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using ActivationOp = PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<ActivationOp>;
+
+static constexpr auto ConvSpec =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
+using DeviceGroupedConvNDFwdInstance =
+    ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
+        NDimSpatial,
+        InLayout,
+        WeiLayout,
+        ck::Tuple<>,
+        OutLayout,
+        InDataType,
+        WeiDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        ConvSpec,    // ConvForwardSpecialization
+        GemmSpec,    // GemmSpecialization
+        1,           //
+        256,         // BlockSize
+        128,         // MPerBlock
+        256,         // NPerBlock
+        64,          // KPerBlock
+        16,          // AK1
+        16,          // BK1
+        32,          // MPerXdl
+        32,          // NPerXdl
+        2,           // MXdlPerWave
+        4,           // NXdlPerWave
+        S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
+        S<1, 0, 2>,  // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // ABlockTransferSrcAccessOrder
+        2,           // ABlockTransferSrcVectorDim
+        16,          // ABlockTransferSrcScalarPerVector
+        16,          // ABlockTransferDstScalarPerVector_AK1
+        1,           // ABlockLdsExtraM
+        S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
+        S<1, 0, 2>,  // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,  // BBlockTransferSrcAccessOrder
+        2,           // BBlockTransferSrcVectorDim
+        16,          // BBlockTransferSrcScalarPerVector
+        16,          // BBlockTransferDstScalarPerVector_BK1
+        1,           // BBlockLdsExtraN
+        1,
+        1,
+        S<1, 64, 1, 4>,
+        16>;
+
+#include "run_conv2d_fwd_perlayer_quantization_example.inc"
+
+int main()
+{
+    float requant_scale       = 0.5f;
+    const auto out_element_op = OutElementOp{requant_scale, ActivationOp{}};
+    run_conv2d_fwd_perlayer_quantization_example(out_element_op);
+}
diff --git a/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perchannel_quantization_example.inc b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perchannel_quantization_example.inc
new file mode 100644
index 0000000000..e5b924ad51
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perchannel_quantization_example.inc
@@ -0,0 +1,249 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          bool time_kernel,
+                          const ck::utils::conv::ConvParam& conv_param,
+                          const HostTensorDescriptor& in_g_n_c_wis_desc,
+                          const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                          const HostTensorDescriptor& bias_g_k_desc,
+                          const HostTensorDescriptor& requant_scale_g_k_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op,
+                          const OutElementOp& out_element_op)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<BiasDataType> bias(bias_g_k_desc);
+    Tensor<RequantScaleDataType> requant_scale(requant_scale_g_k_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "bias: " << bias.mDesc << std::endl;
+    std::cout << "requant_scale: " << requant_scale.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-128, 127});
+    wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-128, 127});
+    bias.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-128, 127});
+    requant_scale.GenerateTensorValue(GeneratorTensor_2<RequantScaleDataType>{0, 1});
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias.mDesc.GetElementSpaceSize());
+    DeviceMem requant_scale_device_buf(sizeof(RequantScaleDataType) *
+                                       requant_scale.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    bias_device_buf.ToDevice(bias.mData.data());
+    requant_scale_device_buf.ToDevice(requant_scale.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d1_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d1_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(bias_g_k_desc.GetLengths(), d0_g_n_k_wos_lengths);
+    copy(bias_g_k_desc.GetStrides(), d0_g_n_k_wos_strides);
+    copy(requant_scale_g_k_desc.GetLengths(), d1_g_n_k_wos_lengths);
+    copy(requant_scale_g_k_desc.GetStrides(), d1_g_n_k_wos_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(
+        in_device_buf.GetDeviceBuffer(),
+        wei_device_buf.GetDeviceBuffer(),
+        {bias_device_buf.GetDeviceBuffer(), requant_scale_device_buf.GetDeviceBuffer()},
+        out_device_buf.GetDeviceBuffer(),
+        a_g_n_c_wis_lengths,
+        a_g_n_c_wis_strides,
+        b_g_k_c_xs_lengths,
+        b_g_k_c_xs_strides,
+        {d0_g_n_k_wos_lengths, d1_g_n_k_wos_lengths},
+        {d0_g_n_k_wos_strides, d1_g_n_k_wos_strides},
+        e_g_n_k_wos_lengths,
+        e_g_n_k_wos_strides,
+        conv_filter_strides,
+        conv_filter_dilations,
+        input_left_pads,
+        input_right_pads,
+        in_element_op,
+        wei_element_op,
+        out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_host(out_g_n_k_wos_desc);
+
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     AccDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     PassThrough>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  c_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        // TODO: implement elementwise operation for host
+        out_host.ForEach([&](auto&, auto idx) {
+            out_element_op(out_host(idx), c_host(idx), bias(idx), requant_scale(idx));
+        });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        pass &=
+            ck::utils::check_err(out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return (pass ? 0 : 1);
+}
+
+int run_conv2d_fwd_bias_perchannel_quantization_example(const OutElementOp& out_element_op)
+{
+    bool do_verification           = true;
+    bool time_kernel               = true;
+    const ck::index_t ndim_spatial = 2;
+
+    ck::utils::conv::ConvParam conv_param{
+        ndim_spatial, // n_dim
+        1,            // group
+        4,            // batch
+        64,           // output channels
+        192,          // input chanels
+        {3, 3},       // weight HW
+        {71, 71},     // x HW
+        {2, 2},       // strides
+        {1, 1},       // dilations
+        {1, 1},       // left_pads
+        {1, 1}        // right_pads
+    };
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+
+    using InLayout           = ck::tensor_layout::convolution::NHWGC;
+    using WeiLayout          = ck::tensor_layout::convolution::KYXGC;
+    using BiasLayout         = ck::tensor_layout::convolution::G_K;
+    using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
+    using OutLayout          = ck::tensor_layout::convolution::NHWGK;
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    // TODO - make_bias_host_tensor_descriptor_g_n_k_wos_packed()
+    const auto bias_g_k_desc = HostTensorDescriptor({conv_param.G_,
+                                                     conv_param.N_,
+                                                     conv_param.K_,
+                                                     conv_param.output_spatial_lengths_[0],
+                                                     conv_param.output_spatial_lengths_[1]},
+                                                    {
+                                                        conv_param.K_, // g
+                                                        0,             // n
+                                                        1,             // k
+                                                        0,             // ho
+                                                        0              // wo
+                                                    });
+
+    const auto requant_scale_g_k_desc = bias_g_k_desc;
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    using deviceOp = DeviceGroupedConvNDFwdInstance<ndim_spatial,
+                                                    InLayout,
+                                                    WeiLayout,
+                                                    BiasLayout,
+                                                    RequantScaleLayout,
+                                                    OutLayout>;
+
+    return run_grouped_conv_fwd<ndim_spatial,
+                                InDataType,
+                                WeiDataType,
+                                OutDataType,
+                                InElementOp,
+                                WeiElementOp,
+                                OutElementOp,
+                                deviceOp>(do_verification,
+                                          time_kernel,
+                                          conv_param,
+                                          in_g_n_c_wis_desc,
+                                          wei_g_k_c_xs_desc,
+                                          bias_g_k_desc,
+                                          requant_scale_g_k_desc,
+                                          out_g_n_k_wos_desc,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+}
diff --git a/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perlayer_quantization_example.inc b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perlayer_quantization_example.inc
new file mode 100644
index 0000000000..9f3a769dcf
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perlayer_quantization_example.inc
@@ -0,0 +1,228 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          bool time_kernel,
+                          const ck::utils::conv::ConvParam& conv_param,
+                          const HostTensorDescriptor& in_g_n_c_wis_desc,
+                          const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                          const HostTensorDescriptor& bias_g_k_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op,
+                          const OutElementOp& out_element_op)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<BiasDataType> bias(bias_g_k_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "bias: " << bias.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+    wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+    bias.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-5, 5});
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    bias_device_buf.ToDevice(bias.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(bias_g_k_desc.GetLengths(), d0_g_n_k_wos_lengths);
+    copy(bias_g_k_desc.GetStrides(), d0_g_n_k_wos_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                      wei_device_buf.GetDeviceBuffer(),
+                                      {bias_device_buf.GetDeviceBuffer()},
+                                      out_device_buf.GetDeviceBuffer(),
+                                      a_g_n_c_wis_lengths,
+                                      a_g_n_c_wis_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      {d0_g_n_k_wos_lengths},
+                                      {d0_g_n_k_wos_strides},
+                                      e_g_n_k_wos_lengths,
+                                      e_g_n_k_wos_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      in_element_op,
+                                      wei_element_op,
+                                      out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_host(out_g_n_k_wos_desc);
+
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     AccDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     PassThrough>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  c_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        // TODO: implement elementwise operation for host
+        out_host.ForEach(
+            [&](auto&, auto idx) { out_element_op(out_host(idx), c_host(idx), bias(idx)); });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        pass &=
+            ck::utils::check_err(out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return (pass ? 0 : 1);
+}
+
+int run_conv2d_fwd_bias_perlayer_quantization_example(const OutElementOp& out_element_op)
+{
+    bool do_verification           = true;
+    bool time_kernel               = true;
+    const ck::index_t ndim_spatial = 2;
+
+    ck::utils::conv::ConvParam conv_param{
+        ndim_spatial, // n_dim
+        1,            // group
+        4,            // batch
+        64,           // output channels
+        192,          // input chanels
+        {3, 3},       // weight HW
+        {71, 71},     // x HW
+        {2, 2},       // strides
+        {1, 1},       // dilations
+        {1, 1},       // left_pads
+        {1, 1}        // right_pads
+    };
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+
+    using InLayout   = ck::tensor_layout::convolution::NHWGC;
+    using WeiLayout  = ck::tensor_layout::convolution::KYXGC;
+    using BiasLayout = ck::tensor_layout::convolution::G_K;
+    using OutLayout  = ck::tensor_layout::convolution::NHWGK;
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    // TODO - make_bias_host_tensor_descriptor_g_n_k_wos_packed()
+    const auto bias_g_k_desc = HostTensorDescriptor({conv_param.G_,
+                                                     conv_param.N_,
+                                                     conv_param.K_,
+                                                     conv_param.output_spatial_lengths_[0],
+                                                     conv_param.output_spatial_lengths_[1]},
+                                                    {
+                                                        conv_param.K_, // g
+                                                        0,             // n
+                                                        1,             // k
+                                                        0,             // ho
+                                                        0              // wo
+                                                    });
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    return run_grouped_conv_fwd<
+        ndim_spatial,
+        InDataType,
+        WeiDataType,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        DeviceGroupedConvNDFwdInstance<ndim_spatial, InLayout, WeiLayout, BiasLayout, OutLayout>>(
+        do_verification,
+        time_kernel,
+        conv_param,
+        in_g_n_c_wis_desc,
+        wei_g_k_c_xs_desc,
+        bias_g_k_desc,
+        out_g_n_k_wos_desc,
+        in_element_op,
+        wei_element_op,
+        out_element_op);
+}
diff --git a/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perchannel_quantization_example.inc b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perchannel_quantization_example.inc
new file mode 100644
index 0000000000..9b08fc690d
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perchannel_quantization_example.inc
@@ -0,0 +1,234 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          bool time_kernel,
+                          const ck::utils::conv::ConvParam& conv_param,
+                          const HostTensorDescriptor& in_g_n_c_wis_desc,
+                          const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                          const HostTensorDescriptor& requant_scale_g_k_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op,
+                          const OutElementOp& out_element_op)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<RequantScaleDataType> requant_scale(requant_scale_g_k_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "requant_scale: " << requant_scale.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-128, 127});
+    wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-128, 127});
+    requant_scale.GenerateTensorValue(GeneratorTensor_2<RequantScaleDataType>{0, 1});
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem requant_scale_device_buf(sizeof(RequantScaleDataType) *
+                                       requant_scale.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+    requant_scale_device_buf.ToDevice(requant_scale.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> d0_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(requant_scale_g_k_desc.GetLengths(), d0_g_n_k_wos_lengths);
+    copy(requant_scale_g_k_desc.GetStrides(), d0_g_n_k_wos_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                      wei_device_buf.GetDeviceBuffer(),
+                                      {requant_scale_device_buf.GetDeviceBuffer()},
+                                      out_device_buf.GetDeviceBuffer(),
+                                      a_g_n_c_wis_lengths,
+                                      a_g_n_c_wis_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      {d0_g_n_k_wos_lengths},
+                                      {d0_g_n_k_wos_strides},
+                                      e_g_n_k_wos_lengths,
+                                      e_g_n_k_wos_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      in_element_op,
+                                      wei_element_op,
+                                      out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_host(out_g_n_k_wos_desc);
+
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     AccDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     PassThrough>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  c_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        // TODO: implement elementwise operation for host
+        out_host.ForEach([&](auto&, auto idx) {
+            out_element_op(out_host(idx), c_host(idx), requant_scale(idx));
+        });
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        pass &=
+            ck::utils::check_err(out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return (pass ? 0 : 1);
+}
+
+int run_conv2d_fwd_perchannel_quantization_example(const OutElementOp& out_element_op)
+{
+    bool do_verification           = true;
+    bool time_kernel               = true;
+    const ck::index_t ndim_spatial = 2;
+
+    ck::utils::conv::ConvParam conv_param{
+        ndim_spatial, // n_dim
+        1,            // group
+        4,            // batch
+        64,           // output channels
+        192,          // input chanels
+        {3, 3},       // weight HW
+        {71, 71},     // x HW
+        {2, 2},       // strides
+        {1, 1},       // dilations
+        {1, 1},       // left_pads
+        {1, 1}        // right_pads
+    };
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+
+    using InLayout           = ck::tensor_layout::convolution::NHWGC;
+    using WeiLayout          = ck::tensor_layout::convolution::KYXGC;
+    using RequantScaleLayout = ck::tensor_layout::convolution::G_K;
+    using OutLayout          = ck::tensor_layout::convolution::NHWGK;
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto requant_scale_g_k_desc =
+        HostTensorDescriptor({conv_param.G_,
+                              conv_param.N_,
+                              conv_param.K_,
+                              conv_param.output_spatial_lengths_[0],
+                              conv_param.output_spatial_lengths_[1]},
+                             {
+                                 conv_param.K_, // g
+                                 0,             // n
+                                 1,             // k
+                                 0,             // ho
+                                 0              // wo
+                             });
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    using deviceOp = DeviceGroupedConvNDFwdInstance<ndim_spatial,
+                                                    InLayout,
+                                                    WeiLayout,
+                                                    RequantScaleLayout,
+                                                    OutLayout>;
+
+    return run_grouped_conv_fwd<ndim_spatial,
+                                InDataType,
+                                WeiDataType,
+                                OutDataType,
+                                InElementOp,
+                                WeiElementOp,
+                                OutElementOp,
+                                deviceOp>(do_verification,
+                                          time_kernel,
+                                          conv_param,
+                                          in_g_n_c_wis_desc,
+                                          wei_g_k_c_xs_desc,
+                                          requant_scale_g_k_desc,
+                                          out_g_n_k_wos_desc,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+}
diff --git a/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perlayer_quantization_example.inc b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perlayer_quantization_example.inc
new file mode 100644
index 0000000000..267c737e00
--- /dev/null
+++ b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perlayer_quantization_example.inc
@@ -0,0 +1,196 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementOp,
+          typename WeiElementOp,
+          typename OutElementOp,
+          typename DeviceConvNDFwdInstance>
+bool run_grouped_conv_fwd(bool do_verification,
+                          bool time_kernel,
+                          const ck::utils::conv::ConvParam& conv_param,
+                          const HostTensorDescriptor& in_g_n_c_wis_desc,
+                          const HostTensorDescriptor& wei_g_k_c_xs_desc,
+                          const HostTensorDescriptor& out_g_n_k_wos_desc,
+                          const InElementOp& in_element_op,
+                          const WeiElementOp& wei_element_op,
+                          const OutElementOp& out_element_op)
+{
+    Tensor<InDataType> in(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> out_host(out_g_n_k_wos_desc);
+    Tensor<OutDataType> out_device(out_g_n_k_wos_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "out: " << out_host.mDesc << std::endl;
+
+    in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+    wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    // do Conv
+    auto conv     = DeviceConvNDFwdInstance{};
+    auto invoker  = conv.MakeInvoker();
+    auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                      wei_device_buf.GetDeviceBuffer(),
+                                      {},
+                                      out_device_buf.GetDeviceBuffer(),
+                                      a_g_n_c_wis_lengths,
+                                      a_g_n_c_wis_strides,
+                                      b_g_k_c_xs_lengths,
+                                      b_g_k_c_xs_strides,
+                                      {},
+                                      {},
+                                      e_g_n_k_wos_lengths,
+                                      e_g_n_k_wos_strides,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads,
+                                      in_element_op,
+                                      wei_element_op,
+                                      out_element_op);
+
+    if(!conv.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv_param.GetFlops();
+    std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << conv.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp>();
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(in,
+                                                  wei,
+                                                  out_host,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        pass &=
+            ck::utils::check_err(out_device, out_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return (pass ? 0 : 1);
+}
+
+int run_conv2d_fwd_perlayer_quantization_example(const OutElementOp& out_element_op)
+{
+    bool do_verification           = true;
+    bool time_kernel               = false;
+    const ck::index_t ndim_spatial = 2;
+
+    ck::utils::conv::ConvParam conv_param{
+        ndim_spatial, // n_dim
+        1,            // group
+        4,            // batch
+        64,           // output channels
+        192,          // input chanels
+        {3, 3},       // weight HW
+        {71, 71},     // x HW
+        {2, 2},       // strides
+        {1, 1},       // dilations
+        {1, 1},       // left_pads
+        {1, 1}        // right_pads
+    };
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+
+    using InLayout  = ck::tensor_layout::convolution::NHWGC;
+    using WeiLayout = ck::tensor_layout::convolution::KYXGC;
+    using OutLayout = ck::tensor_layout::convolution::NHWGK;
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    return run_grouped_conv_fwd<
+        ndim_spatial,
+        InDataType,
+        WeiDataType,
+        OutDataType,
+        InElementOp,
+        WeiElementOp,
+        OutElementOp,
+        DeviceGroupedConvNDFwdInstance<ndim_spatial, InLayout, WeiLayout, OutLayout>>(
+        do_verification,
+        time_kernel,
+        conv_param,
+        in_g_n_c_wis_desc,
+        wei_g_k_c_xs_desc,
+        out_g_n_k_wos_desc,
+        in_element_op,
+        wei_element_op,
+        out_element_op);
+}
diff --git a/example/41_grouped_conv_conv_fwd/CMakeLists.txt b/example/41_grouped_conv_conv_fwd/CMakeLists.txt
new file mode 100644
index 0000000000..085d90a9e9
--- /dev/null
+++ b/example/41_grouped_conv_conv_fwd/CMakeLists.txt
@@ -0,0 +1,26 @@
+list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list2 gfx908 gfx90a)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp)
+   endif()
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp)
+   endif()
+   if(USE_BITINT_EXTENSION_INT4)
+      add_example_executable(example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_xdl_int4.cpp)
+   endif(USE_BITINT_EXTENSION_INT4)
+   set(target 1)
+ endif()
+endforeach()
+
+if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1")
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
+   endif()
+endif()
diff --git a/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp
new file mode 100644
index 0000000000..e37d413695
--- /dev/null
+++ b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp
@@ -0,0 +1,109 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using In0DataType       = ck::bhalf_t;
+using Wei0DataType      = ck::bhalf_t;
+using Acc0DataType      = float;
+using Wei1DataType      = ck::bhalf_t;
+using Acc1DataType      = float;
+using C1ShuffleDataType = float;
+using Out1DataType      = ck::bhalf_t;
+
+// This is used for reference code
+using Out0DataType = ck::bhalf_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using In0ElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using Wei0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Wei1ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out1ElementOp = ck::tensor_operation::element_wise::UnaryConvert;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceBatchedGemmGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+        Row,               // ALayout
+        Col,               // B0Layout
+        Col,               // B1Layout
+        Row,               // CLayout
+        In0DataType,       // ADataType,
+        Wei0DataType,      // B0DataType,
+        Wei1DataType,      // B1DataType,
+        Out1DataType,      // CDataType,
+        Acc0DataType,      // AccDataType,
+        C1ShuffleDataType, // CShuffleDataType,
+        In0ElementOp,      // AElementOp,
+        Wei0ElementOp,     // B0ElementOp,
+        Out0ElementOp,     // Acc0ElementOp,
+        Wei1ElementOp,     // B1ElementOp,
+        Out1ElementOp,     // CElementOp,
+        GemmDefault,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        128,         // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        4,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        4,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // B1BlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        4,
+        4,
+        true,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+#include "run_grouped_conv_conv_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_grouped_conv_conv_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp
new file mode 100644
index 0000000000..496e676a40
--- /dev/null
+++ b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp
@@ -0,0 +1,109 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using In0DataType       = ck::half_t;
+using Wei0DataType      = ck::half_t;
+using Acc0DataType      = float;
+using Wei1DataType      = ck::half_t;
+using Acc1DataType      = float;
+using C1ShuffleDataType = float;
+using Out1DataType      = ck::half_t;
+
+// This is used for reference code
+using Out0DataType = ck::half_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using In0ElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using Wei0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Wei1ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out1ElementOp = ck::tensor_operation::element_wise::UnaryConvert;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceBatchedGemmGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+        Row,               // ALayout
+        Col,               // B0Layout
+        Col,               // B1Layout
+        Row,               // CLayout
+        In0DataType,       // ADataType,
+        Wei0DataType,      // B0DataType,
+        Wei1DataType,      // B1DataType,
+        Out1DataType,      // CDataType,
+        Acc0DataType,      // AccDataType,
+        C1ShuffleDataType, // CShuffleDataType,
+        In0ElementOp,      // AElementOp,
+        Wei0ElementOp,     // B0ElementOp,
+        Out0ElementOp,     // Acc0ElementOp,
+        Wei1ElementOp,     // B1ElementOp,
+        Out1ElementOp,     // CElementOp,
+        GemmDefault,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        128,         // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        4,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        4,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // B1BlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        4,
+        4,
+        true,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+#include "run_grouped_conv_conv_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_grouped_conv_conv_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp
new file mode 100644
index 0000000000..35d50721dc
--- /dev/null
+++ b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp
@@ -0,0 +1,109 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using In0DataType       = float;
+using Wei0DataType      = float;
+using Acc0DataType      = float;
+using Wei1DataType      = float;
+using Acc1DataType      = float;
+using C1ShuffleDataType = float;
+using Out1DataType      = float;
+
+// This is used for reference code
+using Out0DataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using In0ElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using Wei0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Wei1ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out1ElementOp = ck::tensor_operation::element_wise::UnaryConvert;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceBatchedGemmGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+        Row,               // ALayout
+        Col,               // B0Layout
+        Col,               // B1Layout
+        Row,               // CLayout
+        In0DataType,       // ADataType,
+        Wei0DataType,      // B0DataType,
+        Wei1DataType,      // B1DataType,
+        Out1DataType,      // CDataType,
+        Acc0DataType,      // AccDataType,
+        C1ShuffleDataType, // CShuffleDataType,
+        In0ElementOp,      // AElementOp,
+        Wei0ElementOp,     // B0ElementOp,
+        Out0ElementOp,     // Acc0ElementOp,
+        Wei1ElementOp,     // B1ElementOp,
+        Out1ElementOp,     // CElementOp,
+        GemmDefault,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        16,          // KPerBlock
+        128,         // Gemm1NPerBlock
+        16,          // Gemm1KPerBlock
+        4,           // AK1
+        4,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        4,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        4,
+        4,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        4,
+        4,
+        true,
+        S<4, 64, 1>, // B1BlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        2,
+        2,
+        true,
+        1,               // CShuffleMXdlPerWavePerShuffle
+        2,               // CShuffleNXdlPerWavePerShuffle
+        S<1, 16, 1, 16>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        4>;              // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+#include "run_grouped_conv_conv_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_grouped_conv_conv_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp
new file mode 100644
index 0000000000..80f6e9ae05
--- /dev/null
+++ b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+#error Should compile this file with ck::int4_t support
+#endif
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using In0DataType        = ck::int4_t;
+using Wei0DataType       = ck::int4_t;
+using KernelIn0DataType  = int8_t;
+using KernelWei0DataType = int8_t;
+using Acc0DataType       = int32_t;
+using Wei1DataType       = ck::int4_t;
+using KernelWei1DataType = int8_t;
+using Acc1DataType       = int32_t;
+using C1ShuffleDataType  = int32_t;
+using Out1DataType       = ck::int4_t;
+using KernelOut1DataType = int8_t;
+
+// This is used for reference code
+using Out0DataType = ck::int4_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using In0ElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using Wei0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Wei1ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out1ElementOp = ck::tensor_operation::element_wise::UnaryConvert;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceBatchedGemmGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+        Row,                // ALayout
+        Col,                // B0Layout
+        Col,                // B1Layout
+        Row,                // CLayout
+        KernelIn0DataType,  // ADataType,
+        KernelWei0DataType, // B0DataType,
+        KernelWei1DataType, // B1DataType,
+        KernelOut1DataType, // CDataType,
+        Acc0DataType,       // AccDataType,
+        C1ShuffleDataType,  // CShuffleDataType,
+        In0ElementOp,       // AElementOp,
+        Wei0ElementOp,      // B0ElementOp,
+        Out0ElementOp,      // Acc0ElementOp,
+        Wei1ElementOp,      // B1ElementOp,
+        Out1ElementOp,      // CElementOp,
+        GemmDefault,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        64,          // KPerBlock
+        128,         // Gemm1NPerBlock
+        64,          // Gemm1KPerBlock
+        16,          // AK1
+        16,          // BK1
+        4,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        4,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        16,
+        16,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        16,
+        16,
+        true,
+        S<4, 64, 1>, // B1BlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        4,
+        4,
+        true,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+#define BUILD_INT4_EXAMPLE
+#include "run_grouped_conv_conv_fwd_example.inc"
+
+#if defined(BUILD_INT4_EXAMPLE) && defined(CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4)
+static_assert(sizeof(ck::int4_t) == sizeof(int8_t));
+#endif
+
+int main(int argc, char* argv[]) { return run_grouped_conv_conv_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp
new file mode 100644
index 0000000000..3ade6c811a
--- /dev/null
+++ b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp
@@ -0,0 +1,109 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <type_traits>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+using In0DataType       = int8_t;
+using Wei0DataType      = int8_t;
+using Acc0DataType      = int32_t;
+using Wei1DataType      = int8_t;
+using Acc1DataType      = int32_t;
+using C1ShuffleDataType = int32_t;
+using Out1DataType      = int8_t;
+
+// This is used for reference code
+using Out0DataType = int8_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using In0ElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using Wei0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Wei1ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using Out1ElementOp = ck::tensor_operation::element_wise::UnaryConvert;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceBatchedGemmGemmInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+        Row,               // ALayout
+        Col,               // B0Layout
+        Col,               // B1Layout
+        Row,               // CLayout
+        In0DataType,       // ADataType,
+        Wei0DataType,      // B0DataType,
+        Wei1DataType,      // B1DataType,
+        Out1DataType,      // CDataType,
+        Acc0DataType,      // AccDataType,
+        C1ShuffleDataType, // CShuffleDataType,
+        In0ElementOp,      // AElementOp,
+        Wei0ElementOp,     // B0ElementOp,
+        Out0ElementOp,     // Acc0ElementOp,
+        Wei1ElementOp,     // B1ElementOp,
+        Out1ElementOp,     // CElementOp,
+        GemmDefault,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        64,          // KPerBlock
+        128,         // Gemm1NPerBlock
+        64,          // Gemm1KPerBlock
+        16,          // AK1
+        16,          // BK1
+        4,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        4,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        16,
+        16,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        16,
+        16,
+        true,
+        S<4, 64, 1>, // B1BlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        4,
+        4,
+        true,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+#include "run_grouped_conv_conv_fwd_example.inc"
+
+int main(int argc, char* argv[]) { return run_grouped_conv_conv_fwd_example(argc, argv) ? 0 : 1; }
diff --git a/example/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc b/example/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc
new file mode 100644
index 0000000000..0722d497d8
--- /dev/null
+++ b/example/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc
@@ -0,0 +1,379 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+template <ck::index_t NDimSpatial,
+          typename In0DataType,
+          typename Wei0DataType,
+          typename Out0DataType,
+          typename Wei1DataType,
+          typename Out1DataType,
+          typename In0ElementOp,
+          typename Wei0ElementOp,
+          typename Out0ElementOp,
+          typename Wei1ElementOp,
+          typename Out1ElementOp,
+          typename DeviceOpInstance>
+bool run_grouped_conv_conv_fwd(bool do_verification,
+                               int init_method,
+                               bool time_kernel,
+                               const ck::utils::conv::ConvParam& conv0_param,
+                               const ck::utils::conv::ConvParam& conv1_param,
+                               const HostTensorDescriptor& in0_g_n_c_wis_desc,
+                               const HostTensorDescriptor& wei0_g_k_c_xs_desc,
+                               const HostTensorDescriptor& out0_g_n_k_wos_desc,
+                               const HostTensorDescriptor& wei1_g_k_c_xs_desc,
+                               const HostTensorDescriptor& out1_g_n_k_wos_desc,
+                               const In0ElementOp& in0_element_op,
+                               const Wei0ElementOp& wei0_element_op,
+                               const Wei1ElementOp& wei1_element_op,
+                               const Out0ElementOp& out0_element_op,
+                               const Out1ElementOp& out1_element_op)
+{
+    Tensor<In0DataType> in0(in0_g_n_c_wis_desc);
+    Tensor<Wei0DataType> wei0(wei0_g_k_c_xs_desc);
+    Tensor<Wei1DataType> wei1(wei1_g_k_c_xs_desc);
+    Tensor<Out1DataType> out1_host(out1_g_n_k_wos_desc);
+    Tensor<Out1DataType> out1_device(out1_g_n_k_wos_desc);
+
+    std::cout << "in0: " << in0.mDesc << std::endl;
+    std::cout << "wei0: " << wei0.mDesc << std::endl;
+    std::cout << "wei1: " << wei1.mDesc << std::endl;
+    std::cout << "out1: " << out1_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in0.GenerateTensorValue(GeneratorTensor_2<In0DataType>{-5, 5});
+        wei0.GenerateTensorValue(GeneratorTensor_2<Wei0DataType>{-5, 5});
+        wei1.GenerateTensorValue(GeneratorTensor_2<Wei1DataType>{-5, 5});
+        break;
+    default:
+        in0.GenerateTensorValue(GeneratorTensor_3<In0DataType>{0.0, 1.0});
+        wei0.GenerateTensorValue(GeneratorTensor_3<Wei0DataType>{-0.5, 0.5});
+        wei1.GenerateTensorValue(GeneratorTensor_3<Wei1DataType>{-0.5, 0.5});
+    }
+
+#ifdef BUILD_INT4_EXAMPLE
+    DeviceMem in0_device_buf(sizeof(KernelIn0DataType) * in0.mDesc.GetElementSpaceSize());
+    DeviceMem wei0_device_buf(sizeof(KernelWei0DataType) * wei0.mDesc.GetElementSpaceSize());
+    DeviceMem wei1_device_buf(sizeof(KernelWei1DataType) * wei1.mDesc.GetElementSpaceSize());
+    DeviceMem out1_device_buf(sizeof(KernelOut1DataType) * out1_device.mDesc.GetElementSpaceSize());
+
+    const Tensor<KernelIn0DataType> in0_converted(in0);
+    const Tensor<KernelWei0DataType> wei0_converted(wei0);
+    const Tensor<KernelWei1DataType> wei1_converted(wei1);
+
+    in0_device_buf.ToDevice(in0_converted.mData.data());
+    wei0_device_buf.ToDevice(wei0_converted.mData.data());
+    wei1_device_buf.ToDevice(wei1_converted.mData.data());
+#else
+    DeviceMem in0_device_buf(sizeof(In0DataType) * in0.mDesc.GetElementSpaceSize());
+    DeviceMem wei0_device_buf(sizeof(Wei0DataType) * wei0.mDesc.GetElementSpaceSize());
+    DeviceMem wei1_device_buf(sizeof(Wei1DataType) * wei1.mDesc.GetElementSpaceSize());
+    DeviceMem out1_device_buf(sizeof(Out1DataType) * out1_device.mDesc.GetElementSpaceSize());
+
+    in0_device_buf.ToDevice(in0.mData.data());
+    wei0_device_buf.ToDevice(wei0.mData.data());
+    wei1_device_buf.ToDevice(wei1.mData.data());
+#endif
+
+    std::array<ck::index_t, NDimSpatial + 3> a0_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a0_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b0_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b0_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b1_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b1_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e1_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e1_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv0_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv0_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input0_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input0_right_pads{};
+    std::array<ck::index_t, NDimSpatial> conv1_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv1_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input1_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input1_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in0_g_n_c_wis_desc.GetLengths(), a0_g_n_c_wis_lengths);
+    copy(in0_g_n_c_wis_desc.GetStrides(), a0_g_n_c_wis_strides);
+    copy(wei0_g_k_c_xs_desc.GetLengths(), b0_g_k_c_xs_lengths);
+    copy(wei0_g_k_c_xs_desc.GetStrides(), b0_g_k_c_xs_strides);
+    copy(wei1_g_k_c_xs_desc.GetLengths(), b1_g_k_c_xs_lengths);
+    copy(wei1_g_k_c_xs_desc.GetStrides(), b1_g_k_c_xs_strides);
+    copy(out1_g_n_k_wos_desc.GetLengths(), e1_g_n_k_wos_lengths);
+    copy(out1_g_n_k_wos_desc.GetStrides(), e1_g_n_k_wos_strides);
+    copy(conv0_param.conv_filter_strides_, conv0_filter_strides);
+    copy(conv0_param.conv_filter_dilations_, conv0_filter_dilations);
+    copy(conv0_param.input_left_pads_, input0_left_pads);
+    copy(conv0_param.input_right_pads_, input0_right_pads);
+    copy(conv1_param.conv_filter_strides_, conv1_filter_strides);
+    copy(conv1_param.conv_filter_dilations_, conv1_filter_dilations);
+    copy(conv1_param.input_left_pads_, input1_left_pads);
+    copy(conv1_param.input_right_pads_, input1_right_pads);
+
+    // do Conv using GEMM, only works for 1x1 conv for now
+    const ck::index_t gemm_batch = a0_g_n_c_wis_lengths[0];
+
+    const ck::index_t gemm0_m_length =
+        e1_g_n_k_wos_lengths[1] *
+        ck::accumulate_n<ck::index_t>(
+            e1_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>{});
+
+    const ck::index_t gemm0_n_length = b0_g_k_c_xs_lengths[1];
+
+    const ck::index_t gemm0_k_length = ck::accumulate_n<ck::index_t>(
+        b0_g_k_c_xs_lengths.begin() + 2, NDimSpatial + 1, 1, std::multiplies<>{});
+
+    const ck::index_t gemm1_n_length = b1_g_k_c_xs_lengths[1];
+
+    //
+    const ck::index_t a0_stride = a0_g_n_c_wis_strides[2 + NDimSpatial];
+    const ck::index_t b0_stride = b0_g_k_c_xs_strides[2 + NDimSpatial];
+    const ck::index_t b1_stride = b1_g_k_c_xs_strides[2 + NDimSpatial];
+    const ck::index_t e1_stride = e1_g_n_k_wos_strides[2 + NDimSpatial];
+
+    //
+    const ck::index_t a0_batch_stride = a0_g_n_c_wis_strides[0];
+    const ck::index_t b0_batch_stride = b0_g_k_c_xs_strides[0];
+    const ck::index_t b1_batch_stride = b1_g_k_c_xs_strides[0];
+    const ck::index_t e1_batch_stride = e1_g_n_k_wos_strides[0];
+
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument  = device_op.MakeArgument(
+#ifdef BUILD_INT4_EXAMPLE
+        static_cast<KernelIn0DataType*>(in0_device_buf.GetDeviceBuffer()),
+        static_cast<KernelWei0DataType*>(wei0_device_buf.GetDeviceBuffer()),
+        static_cast<KernelWei1DataType*>(wei1_device_buf.GetDeviceBuffer()),
+        static_cast<KernelOut1DataType*>(out1_device_buf.GetDeviceBuffer()),
+#else
+        static_cast<In0DataType*>(in0_device_buf.GetDeviceBuffer()),
+        static_cast<Wei0DataType*>(wei0_device_buf.GetDeviceBuffer()),
+        static_cast<Wei1DataType*>(wei1_device_buf.GetDeviceBuffer()),
+        static_cast<Out1DataType*>(out1_device_buf.GetDeviceBuffer()),
+#endif
+        gemm0_m_length,
+        gemm0_n_length,
+        gemm0_k_length,
+        gemm1_n_length,
+        gemm_batch,
+        a0_stride,
+        b0_stride,
+        b1_stride,
+        e1_stride,
+        a0_batch_stride,
+        b0_batch_stride,
+        b1_batch_stride,
+        e1_batch_stride,
+        in0_element_op,
+        wei0_element_op,
+        out0_element_op,
+        wei1_element_op,
+        out1_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_conv with the specified compilation parameters does "
+            "not support this Conv problem");
+    }
+
+    float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = conv0_param.GetFlops() + conv1_param.GetFlops();
+    std::size_t num_btype = conv0_param.template GetInputByte<In0DataType>() +
+                            conv0_param.template GetWeightByte<Wei0DataType>() +
+                            conv1_param.template GetWeightByte<Wei1DataType>() +
+                            conv1_param.template GetOutputByte<Out1DataType>();
+
+    float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+    float gb_per_sec = num_btype / 1.E6 / avg_time;
+    std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << device_op.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+        Tensor<Out0DataType> out0_host(out0_g_n_k_wos_desc);
+
+        auto ref_conv0 = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                      In0DataType,
+                                                                      Wei0DataType,
+                                                                      Out0DataType,
+                                                                      In0ElementOp,
+                                                                      Wei0ElementOp,
+                                                                      Out0ElementOp>();
+
+        auto ref_conv1 = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                      Out0DataType,
+                                                                      Wei1DataType,
+                                                                      Out1DataType,
+                                                                      PassThrough,
+                                                                      Wei1ElementOp,
+                                                                      Out1ElementOp>();
+
+        auto ref_conv0_invoker = ref_conv0.MakeInvoker();
+        auto ref_conv1_invoker = ref_conv1.MakeInvoker();
+
+        auto ref_conv0_argument = ref_conv0.MakeArgument(in0,
+                                                         wei0,
+                                                         out0_host,
+                                                         conv0_param.conv_filter_strides_,
+                                                         conv0_param.conv_filter_dilations_,
+                                                         conv0_param.input_left_pads_,
+                                                         conv0_param.input_right_pads_,
+                                                         in0_element_op,
+                                                         wei0_element_op,
+                                                         out0_element_op);
+
+        auto ref_conv1_argument = ref_conv1.MakeArgument(out0_host,
+                                                         wei1,
+                                                         out1_host,
+                                                         conv1_param.conv_filter_strides_,
+                                                         conv1_param.conv_filter_dilations_,
+                                                         conv1_param.input_left_pads_,
+                                                         conv1_param.input_right_pads_,
+                                                         out0_element_op,
+                                                         wei1_element_op,
+                                                         out1_element_op);
+
+        ref_conv0_invoker.Run(ref_conv0_argument);
+        ref_conv1_invoker.Run(ref_conv1_argument);
+
+#ifdef BUILD_INT4_EXAMPLE
+        Tensor<KernelOut1DataType> out1_device_converted(out1_host.mDesc);
+
+        out1_device_buf.FromDevice(out1_device_converted.mData.data());
+
+        out1_device = out1_device_converted.CopyAsType<Out1DataType>();
+#else
+        out1_device_buf.FromDevice(out1_device.mData.data());
+#endif
+
+        return ck::utils::check_err(
+            out1_device, out1_host, "Error: incorrect results!", 1e-5f, 1e-4f);
+    }
+
+    return true;
+}
+
+bool run_grouped_conv_conv_fwd_example(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    ck::utils::conv::ConvParam conv0_param{
+        2, 1, 128, 512, 128, {1, 1}, {28, 28}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
+
+    ck::utils::conv::ConvParam conv1_param{
+        2, 1, 128, 128, 512, {1, 1}, {28, 28}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    const auto in0_element_op  = In0ElementOp{};
+    const auto wei0_element_op = Wei0ElementOp{};
+    const auto wei1_element_op = Wei1ElementOp{};
+    const auto out0_element_op = Out0ElementOp{};
+    const auto out1_element_op = Out1ElementOp{};
+
+    const auto run = [&](auto ndim_spatial,
+                         auto in0_layout,
+                         auto wei0_layout,
+                         auto wei1_layout,
+                         auto out1_layout) {
+        constexpr ck::index_t ndim_spatial_value = ndim_spatial.value;
+
+        using In0Layout  = decltype(in0_layout);
+        using Wei0Layout = decltype(wei0_layout);
+        using Wei1Layout = decltype(wei1_layout);
+        using Out1Layout = decltype(out1_layout);
+
+        const auto in0_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<In0Layout>(
+                conv0_param);
+
+        const auto wei0_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<Wei0Layout>(
+                conv0_param);
+
+        // out0 doesn't physical exist, any layout for host verification is OK
+        const auto out0_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<Out1Layout>(
+                conv0_param);
+
+        const auto wei1_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<Wei1Layout>(
+                conv1_param);
+
+        const auto out1_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<Out1Layout>(
+                conv1_param);
+
+        return run_grouped_conv_conv_fwd<ndim_spatial_value,
+                                         In0DataType,
+                                         Wei0DataType,
+                                         Out0DataType,
+                                         Wei1DataType,
+                                         Out1DataType,
+                                         In0ElementOp,
+                                         Wei0ElementOp,
+                                         Out0ElementOp,
+                                         Wei1ElementOp,
+                                         Out1ElementOp,
+                                         DeviceBatchedGemmGemmInstance>(do_verification,
+                                                                        init_method,
+                                                                        time_kernel,
+                                                                        conv0_param,
+                                                                        conv1_param,
+                                                                        in0_g_n_c_wis_desc,
+                                                                        wei0_g_k_c_xs_desc,
+                                                                        out0_g_n_k_wos_desc,
+                                                                        wei1_g_k_c_xs_desc,
+                                                                        out1_g_n_k_wos_desc,
+                                                                        in0_element_op,
+                                                                        wei0_element_op,
+                                                                        wei1_element_op,
+                                                                        out0_element_op,
+                                                                        out1_element_op);
+    };
+
+    namespace ctc = ck::tensor_layout::convolution;
+
+    if(conv0_param.num_dim_spatial_ == 1)
+    {
+        return run(ck::Number<1>{}, ctc::GNWC{}, ctc::GKXC{}, ctc::GKXC{}, ctc::GNWK{});
+    }
+    else if(conv0_param.num_dim_spatial_ == 2)
+    {
+        return run(ck::Number<2>{}, ctc::GNHWC{}, ctc::GKYXC{}, ctc::GKYXC{}, ctc::GNHWK{});
+    }
+    else if(conv0_param.num_dim_spatial_ == 3)
+    {
+        return run(ck::Number<3>{}, ctc::GNDHWC{}, ctc::GKZYXC{}, ctc::GKZYXC{}, ctc::GNDHWK{});
+    }
+
+    return true;
+}
diff --git a/example/42_groupnorm/CMakeLists.txt b/example/42_groupnorm/CMakeLists.txt
new file mode 100644
index 0000000000..bc2246a2bf
--- /dev/null
+++ b/example/42_groupnorm/CMakeLists.txt
@@ -0,0 +1,5 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
+    add_example_executable(example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp)
+    add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)
+endif()
diff --git a/example/42_groupnorm/common.hpp b/example/42_groupnorm/common.hpp
new file mode 100644
index 0000000000..c8f91eb53b
--- /dev/null
+++ b/example/42_groupnorm/common.hpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp"
diff --git a/example/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp b/example/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp
new file mode 100644
index 0000000000..cc107b63dc
--- /dev/null
+++ b/example/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp
@@ -0,0 +1,56 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+constexpr int Rank         = 5;
+constexpr int NumReduceDim = 3;
+
+using XDataType       = ck::half_t;
+using GammaDataType   = ck::half_t;
+using BetaDataType    = ck::half_t;
+using YDataType       = ck::half_t;
+using ComputeDataType = float;
+
+struct YElementOp
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(ck::is_same<T, float>::value || ck::is_same<T, double>::value ||
+                          ck::is_same<T, ck::half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        T a;
+
+        ck::tensor_operation::element_wise::Sigmoid{}(a, x);
+
+        y = x * a;
+    };
+};
+
+using DeviceInstance =
+    ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
+                                                          GammaDataType,
+                                                          BetaDataType,
+                                                          ComputeDataType,
+                                                          YDataType,
+                                                          YElementOp,
+                                                          Rank,
+                                                          NumReduceDim,
+                                                          1024, // BlockSize
+                                                          1,    // ClusterM
+                                                          1024, // ClusterK
+                                                          1,    // SliceM
+                                                          32,   // SliceK
+                                                          1,    // SrcVecDim (0=M, 1=K)
+                                                          2,    // SrcScalarPerVector
+                                                          1,    // GammaVecDim (0=M, 1=K)
+                                                          2,    // GammaScalarPerVector
+                                                          1,    // BetaVecDim (0=M, 1=K)
+                                                          2,    // BetaScalarPerVector
+                                                          2>;   // OutScalarPerVector
+
+#include "run_groupnorm_example.inc"
+
+int main(int argc, char* argv[]) { run_groupnorm_example(argc, argv); }
diff --git a/example/42_groupnorm/groupnorm_splitk_fp16.cpp b/example/42_groupnorm/groupnorm_splitk_fp16.cpp
new file mode 100644
index 0000000000..057b240a63
--- /dev/null
+++ b/example/42_groupnorm/groupnorm_splitk_fp16.cpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+constexpr int Rank         = 5;
+constexpr int NumReduceDim = 3;
+
+using XDataType       = ck::half_t;
+using GammaDataType   = ck::half_t;
+using BetaDataType    = ck::half_t;
+using YDataType       = ck::half_t;
+using ComputeDataType = float;
+using YElementOp      = ck::tensor_operation::element_wise::Swish;
+
+using DeviceInstance =
+    ck::tensor_operation::device::DeviceNormalizationSplitKImpl<XDataType,
+                                                                GammaDataType,
+                                                                BetaDataType,
+                                                                ComputeDataType,
+                                                                YDataType,
+                                                                YElementOp,
+                                                                Rank,
+                                                                NumReduceDim,
+                                                                256, // BlockSize
+                                                                1,   // ClusterM
+                                                                256, // ClusterK
+                                                                1,   // SliceM
+                                                                16,  // SliceK
+                                                                1,   // SrcVecDim (0=M, 1=K)
+                                                                2,   // SrcScalarPerVector
+                                                                1,   // GammaVecDim (0=M, 1=K)
+                                                                2,   // GammaScalarPerVector
+                                                                1,   // BetaVecDim (0=M, 1=K)
+                                                                2,   // BetaScalarPerVector
+                                                                2>;  // OutScalarPerVector
+
+#include "run_groupnorm_example.inc"
+
+int main(int argc, char* argv[]) { run_groupnorm_example(argc, argv); }
diff --git a/example/42_groupnorm/groupnorm_swish_fp16.cpp b/example/42_groupnorm/groupnorm_swish_fp16.cpp
new file mode 100644
index 0000000000..363f22ed4c
--- /dev/null
+++ b/example/42_groupnorm/groupnorm_swish_fp16.cpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+constexpr int Rank         = 5;
+constexpr int NumReduceDim = 3;
+
+using XDataType       = ck::half_t;
+using GammaDataType   = ck::half_t;
+using BetaDataType    = ck::half_t;
+using YDataType       = ck::half_t;
+using ComputeDataType = float;
+using YElementOp      = ck::tensor_operation::element_wise::Swish;
+
+using DeviceInstance =
+    ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
+                                                          GammaDataType,
+                                                          BetaDataType,
+                                                          ComputeDataType,
+                                                          YDataType,
+                                                          YElementOp,
+                                                          Rank,
+                                                          NumReduceDim,
+                                                          1024, // BlockSize
+                                                          1,    // ClusterM
+                                                          1024, // ClusterK
+                                                          1,    // SliceM
+                                                          32,   // SliceK
+                                                          1,    // SrcVecDim (0=M, 1=K)
+                                                          2,    // SrcScalarPerVector
+                                                          1,    // GammaVecDim (0=M, 1=K)
+                                                          2,    // GammaScalarPerVector
+                                                          1,    // BetaVecDim (0=M, 1=K)
+                                                          2,    // BetaScalarPerVector
+                                                          2>;   // OutScalarPerVector
+
+#include "run_groupnorm_example.inc"
+
+int main(int argc, char* argv[]) { run_groupnorm_example(argc, argv); }
diff --git a/example/42_groupnorm/run_groupnorm_example.inc b/example/42_groupnorm/run_groupnorm_example.inc
new file mode 100644
index 0000000000..16065c8d46
--- /dev/null
+++ b/example/42_groupnorm/run_groupnorm_example.inc
@@ -0,0 +1,113 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+int run_groupnorm_example(int argc, char* argv[])
+{
+    ck::index_t N = 32;
+    ck::index_t H = 16;
+    ck::index_t W = 16;
+    ck::index_t G = 64;
+    ck::index_t C = 128;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 6)
+    {
+        N = std::stoi(argv[1]);
+        H = std::stoi(argv[2]);
+        W = std::stoi(argv[3]);
+        G = std::stoi(argv[4]);
+        C = std::stoi(argv[5]);
+    }
+    else
+    {
+        std::cerr << "arg1 to 5: N, H, W, G, C" << std::endl;
+
+        return 1;
+    }
+
+    Tensor<XDataType> x({N, H, W, G, C});
+    Tensor<YDataType> y({N, H, W, G, C});
+    Tensor<GammaDataType> gamma({G, C});
+    Tensor<BetaDataType> beta({G, C});
+
+    ck::utils::FillUniformDistribution<XDataType>{0.f, 1.f}(x);
+    ck::utils::FillUniformDistribution<GammaDataType>{0.f, 1.f}(gamma);
+    ck::utils::FillUniformDistribution<BetaDataType>{0.f, 1.f}(beta);
+
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    beta_dev.ToDevice(beta.mData.data());
+
+    const auto y_element_op = YElementOp{};
+
+    auto device_instance = DeviceInstance{};
+    auto argument_ptr    = device_instance.MakeArgumentPointer(
+        {N, H, W, G, C},
+        std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
+        {0, 0, 0, C, 1},
+        {0, 0, 0, C, 1},
+        std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
+        {1, 2, 4}, // reduction dimension: [H, W, C]
+        1e-6,
+        x_dev.GetDeviceBuffer(),
+        gamma_dev.GetDeviceBuffer(),
+        beta_dev.GetDeviceBuffer(),
+        y_dev.GetDeviceBuffer(),
+        nullptr,
+        nullptr,
+        y_element_op);
+
+    if(!device_instance.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters are not supported" << std::endl;
+        return 1;
+    };
+
+    size_t workspace_sz = device_instance.GetWorkSpaceSize(argument_ptr.get());
+    DeviceMem workspace_dev(workspace_sz);
+    device_instance.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+    auto invoker_ptr = device_instance.MakeInvokerPointer();
+    float ave_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true, true});
+
+    std::size_t num_btype = sizeof(XDataType) * N * H * W * G * C +
+                            sizeof(YDataType) * N * H * W * G * C + sizeof(GammaDataType) * G * C +
+                            sizeof(BetaDataType) * G * C;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s, "
+              << device_instance.GetTypeString() << std::endl;
+
+    bool pass = true;
+    {
+        Tensor<YDataType> host_y({N, H, W, G, C});
+        using ReferenceInstance = ck::tensor_operation::host::ReferenceGroupnorm<XDataType,
+                                                                                 GammaDataType,
+                                                                                 BetaDataType,
+                                                                                 YDataType,
+                                                                                 ComputeDataType,
+                                                                                 YElementOp>;
+
+        ReferenceInstance ref;
+        auto ref_argument =
+            ref.MakeArgument(x, gamma, beta, host_y, y_element_op, {N, H, W, G, C}, 1e-6);
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+
+        y_dev.FromDevice(y.mData.data());
+        pass &= ck::utils::check_err(y, host_y, "Error: Incorrect results", 1e-3, 1e-3);
+    }
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/43_splitk_gemm_bias_e_permute/CMakeLists.txt b/example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
new file mode 100644
index 0000000000..7e070f5357
--- /dev/null
+++ b/example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
@@ -0,0 +1,6 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)
+endif()
diff --git a/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp b/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp
new file mode 100644
index 0000000000..ebba88cf41
--- /dev/null
+++ b/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp
@@ -0,0 +1,407 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 1;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //############################################| NumDimG| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           Gemm|              A|              B|             DE| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|CBlockTransferClusterLengths|   CBlockTransfer|
+        //############################################|        |        |        |        |  Type|  Type|    Type| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Spacialization| Spacialization| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|        _MBlock_MWaveMPerXdl|  ScalarPerVector|
+        //############################################|        |        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |               |               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|        _NBlock_NWaveNPerXdl|    _NWaveNPerXdl|
+        //############################################|        |        |        |        |      |      |        |         |           |      |             |            |             |               |               |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                            |                 |
+        DeviceSplitKContractionMultipleD_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK,   F16,   F16,     F32,      F16, DsDataType,   F16,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,              S<1, 32, 1, 4>,               8>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimG,
+          ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimG == 2 && NumDimM == 2 && NumDimN == 2 && NumDimK == 1, bool> =
+              false>
+struct ReferenceContraction_G2_M2_N2_K1 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_gs_ms_ks,
+                 const Tensor<BDataType>& b_gs_ns_ks,
+                 Tensor<EDataType>& e_gs_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_gs_ms_ks_{a_gs_ms_ks},
+              b_gs_ns_ks_{b_gs_ns_ks},
+              e_gs_ms_ns_{e_gs_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_gs_ms_ks_;
+        const Tensor<BDataType>& b_gs_ns_ks_;
+        Tensor<EDataType>& e_gs_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_G2_M2_N2_K1::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto g0, auto g1, auto m0, auto m1, auto n0, auto n1) {
+                const int K0 = arg.a_gs_ms_ks_.mDesc.GetLengths()[4];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    AccDataType v_a;
+                    AccDataType v_b;
+
+                    arg.a_element_op_(
+                        v_a,
+                        ck::type_convert<const AccDataType>(arg.a_gs_ms_ks_(g0, g1, m0, m1, k0)));
+                    arg.b_element_op_(
+                        v_b,
+                        ck::type_convert<const AccDataType>(arg.b_gs_ns_ks_(g0, g1, n0, n1, k0)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_gs_ms_ns_(g0, g1, m0, m1, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[3],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[4],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[5])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_gs_ms_ks,
+                             const Tensor<BDataType>& b_gs_ns_ks,
+                             Tensor<EDataType>& e_gs_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{
+            a_gs_ms_ks, b_gs_ns_ks, e_gs_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_G2_M2_N2_K1"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    int split_k          = 1;
+
+    ck::index_t G0 = 1;
+    ck::index_t G1 = 2;
+
+    ck::index_t M0 = 4;
+    ck::index_t M1 = 256;
+
+    ck::index_t N0 = 16;
+    ck::index_t N1 = 128;
+
+    ck::index_t K0 = 64 * 2;
+
+    // A[G0, G1, M0, M1, K0]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M0, M1, K0};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{G1 * M0 * M1 * K0, M0 * M1 * K0, M1 * K0, K0, 1};
+    // B[G0, G1, N0, N1, K0]
+    std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, G1, N0, N1, K0};
+    std::vector<ck::index_t> b_gs_ns_ks_strides{G1 * N0 * N1 * K0, N0 * N1 * K0, N1 * K0, K0, 1};
+
+    // D[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> d_gs_ms_ns_strides{G1 * N0 * N1, N0 * N1, 0, 0, N1, 1};
+    // E[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> e_gs_ms_ns_strides{
+        G1 * M0 * N0 * M1 * N1, M0 * N0 * M1 * N1, N0 * M1 * N1, N1, M1 * N1, 1};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 5)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        split_k         = std::stoi(argv[4]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_gs_ms_ks(
+        std::vector<std::size_t>(a_gs_ms_ks_lengths.begin(), a_gs_ms_ks_lengths.end()),
+        std::vector<std::size_t>(a_gs_ms_ks_strides.begin(), a_gs_ms_ks_strides.end()));
+    Tensor<BDataType> b_gs_ns_ks(
+        std::vector<std::size_t>(b_gs_ns_ks_lengths.begin(), b_gs_ns_ks_lengths.end()),
+        std::vector<std::size_t>(b_gs_ns_ks_strides.begin(), b_gs_ns_ks_strides.end()));
+    Tensor<DDataType> d_gs_ms_ns(
+        std::vector<std::size_t>(d_gs_ms_ns_lengths.begin(), d_gs_ms_ns_lengths.end()),
+        std::vector<std::size_t>(d_gs_ms_ns_strides.begin(), d_gs_ms_ns_strides.end()));
+    Tensor<EDataType> e_gs_ms_ns_host_result(
+        std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
+        std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
+    Tensor<EDataType> e_gs_ms_ns_device_result(
+        std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
+        std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
+    std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
+    std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) *
+                           e_gs_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_gs_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_gs_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_gs_ms_ks_lengths,
+                                    a_gs_ms_ks_strides,
+                                    b_gs_ns_ks_lengths,
+                                    b_gs_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
+                                    e_gs_ms_ns_lengths,
+                                    e_gs_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op,
+                                    split_k);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t G = std::accumulate(e_gs_ms_ns_lengths.begin(),
+                                    e_gs_ms_ns_lengths.begin() + NumDimG,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    ck::index_t M = std::accumulate(e_gs_ms_ns_lengths.begin() + NumDimG,
+                                    e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    ck::index_t N = std::accumulate(e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM,
+                                    e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM + NumDimN,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    ck::index_t K = std::accumulate(a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM,
+                                    a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM + NumDimK,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    std::size_t flop      = std::size_t(2) * G * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * G * M * K + sizeof(BDataType) * G * K * N +
+                            sizeof(DDataType) * G * M * N + sizeof(EDataType) * G * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_gs_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(
+            std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
+            std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
+
+        using ReferenceOpInstance = ReferenceContraction_G2_M2_N2_K1<NumDimG,
+                                                                     NumDimM,
+                                                                     NumDimN,
+                                                                     NumDimK,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CShuffleDataType,
+                                                                     AccDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_gs_ms_ks, b_gs_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        e_gs_ms_ns_host_result.ForEach([&](auto&, auto idx) {
+            cde_element_op(e_gs_ms_ns_host_result(idx), c_ms_ns_host_result(idx), d_gs_ms_ns(idx));
+        });
+
+        return ck::utils::check_err(e_gs_ms_ns_device_result.mData, e_gs_ms_ns_host_result.mData)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}
diff --git a/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp b/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp
new file mode 100644
index 0000000000..4ab26293cc
--- /dev/null
+++ b/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp
@@ -0,0 +1,407 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F32;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F32;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 1;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+static constexpr auto ABSpec = ck::tensor_operation::device::TensorSpecialization::Packed;
+static constexpr auto DESpec = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstanceKKNN = ck::tensor_operation::device::
+        //############################################| NumDimG| NumDimM| NumDimN| NumDimK|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|            A|           B|          CDE|           Gemm|              A|              B|             DE| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|CBlockTransferClusterLengths|   CBlockTransfer|
+        //############################################|        |        |        |        |      Type|      Type|        Type|         DataType|       Type|      Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Spacialization| Spacialization| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|        _MBlock_MWaveMPerXdl|  ScalarPerVector|
+        //############################################|        |        |        |        |          |          |            |                 |           |          |    Operation|   Operation|    Operation|               |               |               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|        _NBlock_NWaveNPerXdl|    _NWaveNPerXdl|
+        //############################################|        |        |        |        |          |          |            |                 |           |          |             |            |             |               |               |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                            |                 |
+        DeviceSplitKContractionMultipleD_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   4,   4,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,         1,           1,           1,              S<1, 32, 1, 4>,               4>;
+// clang-format on
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimG,
+          ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ck::enable_if_t<NumDimG == 2 && NumDimM == 2 && NumDimN == 2 && NumDimK == 1, bool> =
+              false>
+struct ReferenceContraction_G2_M2_N2_K1 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_gs_ms_ks,
+                 const Tensor<BDataType>& b_gs_ns_ks,
+                 Tensor<EDataType>& e_gs_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_gs_ms_ks_{a_gs_ms_ks},
+              b_gs_ns_ks_{b_gs_ns_ks},
+              e_gs_ms_ns_{e_gs_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_gs_ms_ks_;
+        const Tensor<BDataType>& b_gs_ns_ks_;
+        Tensor<EDataType>& e_gs_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_G2_M2_N2_K1::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto g0, auto g1, auto m0, auto m1, auto n0, auto n1) {
+                const int K0 = arg.a_gs_ms_ks_.mDesc.GetLengths()[4];
+
+                AccDataType v_acc = 0;
+
+                for(int k0 = 0; k0 < K0; ++k0)
+                {
+                    AccDataType v_a;
+                    AccDataType v_b;
+
+                    arg.a_element_op_(
+                        v_a,
+                        ck::type_convert<const AccDataType>(arg.a_gs_ms_ks_(g0, g1, m0, m1, k0)));
+                    arg.b_element_op_(
+                        v_b,
+                        ck::type_convert<const AccDataType>(arg.b_gs_ns_ks_(g0, g1, n0, n1, k0)));
+
+                    v_acc += v_a * v_b;
+                }
+
+                AccDataType v_c;
+
+                arg.cde_element_op_(v_c, v_acc);
+
+                arg.e_gs_ms_ns_(g0, g1, m0, m1, n0, n1) = v_c;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[3],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[4],
+                                       arg.e_gs_ms_ns_.mDesc.GetLengths()[5])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_gs_ms_ks,
+                             const Tensor<BDataType>& b_gs_ns_ks,
+                             Tensor<EDataType>& e_gs_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{
+            a_gs_ms_ks, b_gs_ns_ks, e_gs_ms_ns, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_G2_M2_N2_K1"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    int split_k          = 1;
+
+    ck::index_t G0 = 1;
+    ck::index_t G1 = 2;
+
+    ck::index_t M0 = 4;
+    ck::index_t M1 = 256;
+
+    ck::index_t N0 = 16;
+    ck::index_t N1 = 128;
+
+    ck::index_t K0 = 64 * 2;
+
+    // A[G0, G1, M0, M1, K0]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M0, M1, K0};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{G1 * M0 * M1 * K0, M0 * M1 * K0, M1 * K0, K0, 1};
+    // B[G0, G1, N0, N1, K0]
+    std::vector<ck::index_t> b_gs_ns_ks_lengths{G0, G1, N0, N1, K0};
+    std::vector<ck::index_t> b_gs_ns_ks_strides{G1 * N0 * N1 * K0, N0 * N1 * K0, N1 * K0, K0, 1};
+
+    // D[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> d_gs_ms_ns_strides{G1 * N0 * N1, N0 * N1, 0, 0, N1, 1};
+    // E[G0, G1, M0, N0, M1, N1]
+    std::vector<ck::index_t> e_gs_ms_ns_lengths{G0, G1, M0, M1, N0, N1};
+    std::vector<ck::index_t> e_gs_ms_ns_strides{
+        G1 * M0 * N0 * M1 * N1, M0 * N0 * M1 * N1, N0 * M1 * N1, N1, M1 * N1, 1};
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 5)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        split_k         = std::stoi(argv[4]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    Tensor<ADataType> a_gs_ms_ks(
+        std::vector<std::size_t>(a_gs_ms_ks_lengths.begin(), a_gs_ms_ks_lengths.end()),
+        std::vector<std::size_t>(a_gs_ms_ks_strides.begin(), a_gs_ms_ks_strides.end()));
+    Tensor<BDataType> b_gs_ns_ks(
+        std::vector<std::size_t>(b_gs_ns_ks_lengths.begin(), b_gs_ns_ks_lengths.end()),
+        std::vector<std::size_t>(b_gs_ns_ks_strides.begin(), b_gs_ns_ks_strides.end()));
+    Tensor<DDataType> d_gs_ms_ns(
+        std::vector<std::size_t>(d_gs_ms_ns_lengths.begin(), d_gs_ms_ns_lengths.end()),
+        std::vector<std::size_t>(d_gs_ms_ns_strides.begin(), d_gs_ms_ns_strides.end()));
+    Tensor<EDataType> e_gs_ms_ns_host_result(
+        std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
+        std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
+    Tensor<EDataType> e_gs_ms_ns_device_result(
+        std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
+        std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
+    std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
+    std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) *
+                           e_gs_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_gs_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_gs_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // device operation
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                    b_device_buf.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                                    e_device_buf.GetDeviceBuffer(),
+                                    a_gs_ms_ks_lengths,
+                                    a_gs_ms_ks_strides,
+                                    b_gs_ns_ks_lengths,
+                                    b_gs_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_gs_ms_ns_strides},
+                                    e_gs_ms_ns_lengths,
+                                    e_gs_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op,
+                                    split_k);
+
+    if(!op.IsSupportedArgument(argument))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t G = std::accumulate(e_gs_ms_ns_lengths.begin(),
+                                    e_gs_ms_ns_lengths.begin() + NumDimG,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    ck::index_t M = std::accumulate(e_gs_ms_ns_lengths.begin() + NumDimG,
+                                    e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    ck::index_t N = std::accumulate(e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM,
+                                    e_gs_ms_ns_lengths.begin() + NumDimG + NumDimM + NumDimN,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    ck::index_t K = std::accumulate(a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM,
+                                    a_gs_ms_ks_lengths.begin() + NumDimG + NumDimM + NumDimK,
+                                    ck::index_t{1},
+                                    std::multiplies<ck::index_t>{});
+
+    std::size_t flop      = std::size_t(2) * G * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * G * M * K + sizeof(BDataType) * G * K * N +
+                            sizeof(DDataType) * G * M * N + sizeof(EDataType) * G * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf.FromDevice(e_gs_ms_ns_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_ms_ns_host_result(
+            std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
+            std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
+
+        using ReferenceOpInstance = ReferenceContraction_G2_M2_N2_K1<NumDimG,
+                                                                     NumDimM,
+                                                                     NumDimN,
+                                                                     NumDimK,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CShuffleDataType,
+                                                                     AccDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     PassThrough>;
+
+        auto ref_gemm    = ReferenceOpInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_gs_ms_ks, b_gs_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        e_gs_ms_ns_host_result.ForEach([&](auto&, auto idx) {
+            cde_element_op(e_gs_ms_ns_host_result(idx), c_ms_ns_host_result(idx), d_gs_ms_ns(idx));
+        });
+
+        return ck::utils::check_err(e_gs_ms_ns_device_result.mData, e_gs_ms_ns_host_result.mData)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}
diff --git a/example/44_elementwise_permute/CMakeLists.txt b/example/44_elementwise_permute/CMakeLists.txt
new file mode 100644
index 0000000000..877a820315
--- /dev/null
+++ b/example/44_elementwise_permute/CMakeLists.txt
@@ -0,0 +1,4 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
+    add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
+endif()
diff --git a/example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp b/example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp
new file mode 100644
index 0000000000..2ceda86839
--- /dev/null
+++ b/example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp
@@ -0,0 +1,116 @@
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ADataType = F16;
+using BDataType = F16;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using DeviceElementwisePermuteInstance =
+    ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>,
+                                                        ck::Tuple<BDataType>,
+                                                        PassThrough,
+                                                        4,
+                                                        8,
+                                                        ck::Sequence<8>,
+                                                        ck::Sequence<1>>;
+
+template <typename HostTensorA, typename HostTensorB, typename Functor>
+void host_elementwise4D(HostTensorB& B_nhwc, const HostTensorA& A_nchw, Functor functor)
+{
+    for(std::size_t n = 0; n < A_nchw.mDesc.GetLengths()[0]; ++n)
+        for(std::size_t c = 0; c < A_nchw.mDesc.GetLengths()[1]; ++c)
+            for(std::size_t h = 0; h < A_nchw.mDesc.GetLengths()[2]; ++h)
+                for(std::size_t w = 0; w < A_nchw.mDesc.GetLengths()[3]; ++w)
+                {
+                    auto a_val = A_nchw(n, c, h, w);
+                    functor(B_nhwc(n, h, w, c), a_val);
+                }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = true;
+
+    std::vector<std::size_t> nchw = {16, 128, 32, 64};
+    std::vector<std::size_t> nhwc = {16, 32, 64, 128};
+    Tensor<ADataType> a(nchw);
+    Tensor<BDataType> b(nhwc);
+
+    a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+
+    std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 4> ab_lengths;
+    std::array<ck::index_t, 4> a_strides = {static_cast<int>(nchw[1] * nchw[2] * nchw[3]),
+                                            static_cast<int>(nchw[2] * nchw[3]),
+                                            static_cast<int>(nchw[3]),
+                                            1};
+    std::array<ck::index_t, 4> b_strides = {static_cast<int>(nhwc[1] * nhwc[2] * nhwc[3]),
+                                            1,
+                                            static_cast<int>(nhwc[2] * nhwc[3]),
+                                            static_cast<int>(nhwc[3])};
+
+    ck::ranges::copy(nchw, ab_lengths.begin());
+
+    auto broadcastPermute = DeviceElementwisePermuteInstance{};
+    auto argument         = broadcastPermute.MakeArgumentPointer(
+        ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
+
+    if(!broadcastPermute.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    std::cout << "A (nchw): " << a.mDesc << std::endl;
+    std::cout << "B (nhwc): " << b.mDesc << std::endl;
+
+    auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
+    float ave_time =
+        broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+    std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
+
+    std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
+                            sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        b_device_buf.FromDevice(b.mData.data());
+        Tensor<BDataType> host_b(nhwc);
+        host_elementwise4D(host_b, a, PassThrough{});
+
+        pass &=
+            ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp b/example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp
new file mode 100644
index 0000000000..6b94a5d46f
--- /dev/null
+++ b/example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp
@@ -0,0 +1,130 @@
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+using F16 = ck::half_t;
+
+using ADataType = F16;
+using BDataType = F16;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using DeviceElementwisePermuteInstance =
+    ck::tensor_operation::device::DeviceElementwise2dImpl<ck::Tuple<ADataType>,
+                                                          ck::Tuple<BDataType>,
+                                                          PassThrough,
+                                                          3, // NumDim_M
+                                                          1, // NumDim_N
+                                                          8,
+                                                          8,
+                                                          ck::Sequence<8>,
+                                                          ck::Sequence<8>>;
+
+template <typename HostTensorA, typename HostTensorB, typename Functor>
+void host_elementwise4D(HostTensorB& B_nhwc,
+                        const HostTensorA& A_nchw,
+                        const std::vector<std::size_t>& shape_nchw,
+                        Functor functor)
+{
+    for(std::size_t n = 0; n < shape_nchw[0]; ++n)
+        for(std::size_t c = 0; c < shape_nchw[1]; ++c)
+            for(std::size_t h = 0; h < shape_nchw[2]; ++h)
+                for(std::size_t w = 0; w < shape_nchw[3]; ++w)
+                {
+                    auto a_val = A_nchw(n, c, h, w);
+                    functor(B_nhwc(n, h, w, c), a_val);
+                }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = true;
+
+    const int N = 120;
+    const int C = 128;
+    const int H = 32;
+    const int W = 1024;
+
+    /**const int N = 120;
+    const int H = 32;
+    const int W = 64;
+
+    const int C = 128;**/
+
+    std::vector<std::size_t> nchw = {N, C, H, W};
+    std::vector<std::size_t> nhwc = {N, H, W, C};
+
+    Tensor<ADataType> a(nchw);
+    Tensor<BDataType> b(nhwc);
+
+    a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+    // LogRangeAsType<float>(std::cout << "Tensor a  : ", a.mData, ",") << std::endl;
+
+    std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
+
+    std::array<ck::index_t, 4> ab_lengths{N, H, W, C};
+
+    std::array<ck::index_t, 4> a_strides = {C * H * W, W, 1, H * W};
+    std::array<ck::index_t, 4> b_strides = {H * W * C, W * C, C, 1};
+
+    auto broadcastPermute = DeviceElementwisePermuteInstance{};
+    auto argument         = broadcastPermute.MakeArgumentPointer(
+        ab_lengths, {a_strides}, {b_strides}, input, output, PassThrough{});
+
+    if(!broadcastPermute.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error(
+            "The runtime parameters seems not supported by the device instance, exiting!");
+    };
+
+    std::cout << "A (nchw): " << a.mDesc << std::endl;
+    std::cout << "B (nhwc): " << b.mDesc << std::endl;
+
+    auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
+    float ave_time =
+        broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
+
+    std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
+                            sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        b_device_buf.FromDevice(b.mData.data());
+        // LogRangeAsType<float>(std::cout << "Tensor b  : ", b.mData, ",") << std::endl;
+
+        Tensor<BDataType> host_b(nhwc);
+        host_elementwise4D<Tensor<ADataType>, Tensor<BDataType>, PassThrough>(
+            host_b, a, nchw, PassThrough{});
+
+        // LogRangeAsType<float>(std::cout << "Host b  : ", host_b.mData, ",") << std::endl;
+        pass &=
+            ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}
diff --git a/example/45_elementwise_normalization/CMakeLists.txt b/example/45_elementwise_normalization/CMakeLists.txt
new file mode 100644
index 0000000000..8f5b9d4d87
--- /dev/null
+++ b/example/45_elementwise_normalization/CMakeLists.txt
@@ -0,0 +1 @@
+add_example_executable(example_elementwise_layernorm_blockwise elementwise_layernorm_blockwise.cpp)
diff --git a/example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp b/example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp
new file mode 100644
index 0000000000..76361f87a5
--- /dev/null
+++ b/example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp
@@ -0,0 +1,195 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <getopt.h>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
+
+using ADataType             = ck::half_t; // Input 1
+using BDataType             = ck::half_t; // Input 2
+using XDataType             = ck::half_t;
+using GammaDataType         = ck::half_t;
+using BetaDataType          = ck::half_t;
+using YDataType             = ck::half_t;
+using AccDataType           = float;
+using XElementwiseOperation = ck::tensor_operation::element_wise::Add;
+using YElementwiseOperation = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+// X = Elementwise(input1, input2, input3, ...)
+// Y = Layernorm(X, beta, gamma)
+using DeviceInstance = ck::tensor_operation::device::DeviceElementwiseNormalizationImpl<
+    ck::Tuple<ADataType, BDataType>,
+    GammaDataType,
+    BetaDataType,
+    AccDataType,
+    YDataType,
+    XElementwiseOperation,
+    YElementwiseOperation,
+    Rank,
+    NumReduceDim,
+    256, // BlockSize
+    8,   // ClusterM
+    32,  // ClusterK
+    1,   // SliceM
+    32,  // SliceK
+    1,   // SrcVecDim (0=M, 1=K)
+    8,   // SrcScalarPerVector
+    1,   // GammaVecDim (0=M, 1=K)
+    8,   // GammaScalarPerVector
+    1,   // BetaVecDim (0=M, 1=K)
+    8,   // BetaScalarPerVector
+    8>;  // OutScalarPerVector
+
+template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
+void host_elementwise2D(HostTensorC& C,
+                        const HostTensorA& A,
+                        const HostTensorB& B,
+                        const std::vector<std::size_t>& shape,
+                        Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0))>;
+
+    for(std::size_t m = 0; m < shape[0]; ++m)
+        for(std::size_t n = 0; n < shape[1]; ++n)
+        {
+            auto a_val  = A(m, n);
+            auto b_val  = B(m, n);
+            ctype c_val = 0;
+            functor(c_val, a_val, b_val);
+            C(m, n) = c_val;
+        }
+}
+
+int main()
+{
+    bool time_kernel = true;
+
+    ck::index_t M      = 48 * 256;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = N;
+
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({len}),
+                                    std::vector<std::size_t>({stride}));
+    };
+
+    auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                    std::vector<std::size_t>({stride, 1}));
+    };
+
+    Tensor<ADataType> a(f_host_tensor_descriptor2d(M, N, Stride));
+    Tensor<BDataType> b(f_host_tensor_descriptor2d(M, N, Stride));
+    Tensor<GammaDataType> gamma(f_host_tensor_descriptor1d(N, 1));
+    Tensor<BetaDataType> beta(f_host_tensor_descriptor1d(N, 1));
+    Tensor<YDataType> y(f_host_tensor_descriptor2d(M, N, Stride));
+
+    a.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+    b.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+    gamma.GenerateTensorValue(GeneratorTensor_2<GammaDataType>{-5, 5});
+    beta.GenerateTensorValue(GeneratorTensor_2<BetaDataType>{-5, 5});
+
+    DeviceMem a_dev(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_dev(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+
+    a_dev.ToDevice(a.mData.data());
+    b_dev.ToDevice(b.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    beta_dev.ToDevice(beta.mData.data());
+
+    std::array<const void*, 2> input = {a_dev.GetDeviceBuffer(), b_dev.GetDeviceBuffer()};
+
+    auto device_instance = DeviceInstance{};
+    auto argument_ptr    = device_instance.MakeArgumentPointer(
+        {M, N},
+        {
+            std::vector<ck::index_t>{a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end()},
+            std::vector<ck::index_t>{b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end()},
+        },
+        {0, 1},
+        {0, 1},
+        std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
+        {1},
+        1e-4,
+        input,
+        gamma_dev.GetDeviceBuffer(),
+        beta_dev.GetDeviceBuffer(),
+        y_dev.GetDeviceBuffer(),
+        XElementwiseOperation{},
+        YElementwiseOperation{});
+
+    if(!device_instance.IsSupportedArgument(argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters are not supported" << std::endl;
+        return 1;
+    };
+
+    auto invoker_ptr = device_instance.MakeInvokerPointer();
+    float ela_time   = 0;
+    ela_time         = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    float data_mem_size = M * N * sizeof(ADataType) + M * N * sizeof(BDataType) +
+                          M * N * sizeof(YDataType) + N * sizeof(GammaDataType) +
+                          N * sizeof(BetaDataType);
+    float bandwidth = data_mem_size * 1000 / ela_time / 1024 / 1024 / 1024;
+
+    std::cout << "Bandwidth is : " << bandwidth << "GB/s . " << std::endl;
+    std::cout << "Time elapase is : " << ela_time << " ms . " << std::endl;
+
+    bool pass = true;
+    {
+        std::vector<std::size_t> mn = {static_cast<unsigned long>(M),
+                                       static_cast<unsigned long>(N)};
+        Tensor<XDataType> x(f_host_tensor_descriptor2d(M, N, Stride));
+        host_elementwise2D<Tensor<ADataType>,
+                           Tensor<BDataType>,
+                           Tensor<XDataType>,
+                           XElementwiseOperation>(x, a, b, mn, XElementwiseOperation{});
+
+        Tensor<YDataType> host_y(f_host_tensor_descriptor2d(M, N, Stride));
+        using ReferenceInstance =
+            ck::tensor_operation::host::ReferenceLayernorm<XDataType,
+                                                           GammaDataType,
+                                                           BetaDataType,
+                                                           YDataType,
+                                                           AccDataType,
+                                                           YElementwiseOperation,
+                                                           Rank,
+                                                           NumReduceDim>;
+
+        ReferenceInstance ref;
+        auto ref_argument =
+            ref.MakeArgument(x, gamma, beta, host_y, YElementwiseOperation{}, {M, N}, {1}, 1e-4);
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+
+        y_dev.FromDevice(y.mData.data());
+        pass &=
+            ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
+        if(!(pass))
+        {
+            std::cout << "layernorm wrong" << std::endl;
+        }
+    }
+    return (pass ? 0 : 1);
+}
diff --git a/example/46_gemm_add_multiply/CMakeLists.txt b/example/46_gemm_add_multiply/CMakeLists.txt
new file mode 100644
index 0000000000..cf7d81f895
--- /dev/null
+++ b/example/46_gemm_add_multiply/CMakeLists.txt
@@ -0,0 +1,6 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    if(DL_KERNELS)
+        add_example_executable(example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp)
+    endif()
+    add_example_executable(example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp)
+endif()
diff --git a/example/46_gemm_add_multiply/README.md b/example/46_gemm_add_multiply/README.md
new file mode 100644
index 0000000000..ee5cdee365
--- /dev/null
+++ b/example/46_gemm_add_multiply/README.md
@@ -0,0 +1,26 @@
+# Instructions for ```example_gemm_add_multiply_dl_fp16```
+
+## Run ```example_gemm_add_multiply_dl_fp16```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+#arg4 to 11: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, StrideE"
+./bin/example_gemm_add_multiply_dl_fp16 1 1 1
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {4096, 1}
+d0_m_n: dim 2, lengths {3840, 4096}, strides {0, 1}
+d1_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+e_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+arg.a_grid_desc_k0_m0_m1_k1_{2048, 3840, 2}
+arg.b_grid_desc_k0_n0_n1_k1_{2048, 4096, 2}
+arg.e_grid_desc_m_n_{ 3840, 4096}
+launch_and_time_kernel: grid_dim {960, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 3.99904 ms, 32.22 TFlops, 31.9913 GB/s, DeviceGemmMultipleD_Dl<256, 128, 128, 16, 2, 4, 4, 1>
+```
diff --git a/example/46_gemm_add_multiply/common.hpp b/example/46_gemm_add_multiply/common.hpp
new file mode 100644
index 0000000000..2c656cf441
--- /dev/null
+++ b/example/46_gemm_add_multiply/common.hpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <cstddef>
+#include <iostream>
+#include <stdexcept>
+#include <string>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddMultiply = ck::tensor_operation::element_wise::AddMultiply;
+
+using BF16 = ck::bhalf_t;
+using F16  = ck::half_t;
+using F32  = float;
+using I8   = int8_t;
+using I32  = int32_t;
+
+struct ProblemSize final
+{
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA  = 4096;
+    ck::index_t StrideB  = 4096;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = 4096;
+    ck::index_t StrideE  = 4096;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+inline bool
+parse_cmd_args(int argc, char* argv[], ProblemSize& problem_size, ExecutionConfig& config)
+{
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 12)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        problem_size.M = std::stoi(argv[4]);
+        problem_size.N = std::stoi(argv[5]);
+        problem_size.K = std::stoi(argv[6]);
+
+        problem_size.StrideA  = std::stoi(argv[7]);
+        problem_size.StrideB  = std::stoi(argv[8]);
+        problem_size.StrideD0 = std::stoi(argv[9]);
+        problem_size.StrideD1 = std::stoi(argv[10]);
+        problem_size.StrideE  = std::stoi(argv[11]);
+    }
+    else
+    {
+        std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
+                  << std::endl
+                  << "arg3: time kernel (0=no, 1=yes)" << std::endl
+                  << "arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD0, StrideD1, "
+                     "StrideE"
+                  << std::endl;
+        return false;
+    }
+
+    return true;
+}
diff --git a/example/46_gemm_add_multiply/gemm_add_multiply_dl_fp16.cpp b/example/46_gemm_add_multiply/gemm_add_multiply_dl_fp16.cpp
new file mode 100644
index 0000000000..58a399f226
--- /dev/null
+++ b/example/46_gemm_add_multiply/gemm_add_multiply_dl_fp16.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp"
+
+using ADataType   = F16;
+using BDataType   = F16;
+using AccDataType = F32;
+using D0DataType  = F16;
+using D1DataType  = F16;
+using DsDataType  = ck::Tuple<D0DataType, D1DataType>;
+using EDataType   = F16;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddMultiply;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::
+        //  ##################| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|     DsData|     EData|           A|           B|          CDE|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ##################|        |        |         |        |      Type|      Type|        Type|       Type|      Type| Elementwise| Elementwise|  Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ##################|        |        |         |        |          |          |            |           |          |   Operation|   Operation|    Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ##################|        |        |         |        |          |          |            |           |          |            |            |             |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmMultipleD_Dl< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        AccDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_add_multiply_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_multiply_example(argc, argv); }
diff --git a/example/46_gemm_add_multiply/gemm_add_multiply_xdl_fp16.cpp b/example/46_gemm_add_multiply/gemm_add_multiply_xdl_fp16.cpp
new file mode 100644
index 0000000000..56417b101d
--- /dev/null
+++ b/example/46_gemm_add_multiply/gemm_add_multiply_xdl_fp16.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+
+using ADataType   = F16;
+using BDataType   = F16;
+using AccDataType = F32;
+using D0DataType  = F16;
+using D1DataType  = F16;
+using DsDataType  = ck::Tuple<D0DataType, D1DataType>;
+using EDataType   = F16;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using D0Layout = Row;
+using D1Layout = Row;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddMultiply;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::
+        //##############################|      A|      B|       Ds|      E| AData| BData| AccData| CShuffle|     DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|   Layout| Layout|  Type|  Type|    Type| DataType|       Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |         |       |      |      |        |         |           |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |         |       |      |      |        |         |           |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, DsLayout,    Row,   F16,   F16,     F32,      F16, DsDataType,   F16, PassThrough, PassThrough, CDEElementOp,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        AccDataType,
+                                                                        AccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        PassThrough>;
+
+#include "run_gemm_add_multiply_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_add_multiply_example(argc, argv); }
diff --git a/example/46_gemm_add_multiply/run_gemm_add_multiply_example.inc b/example/46_gemm_add_multiply/run_gemm_add_multiply_example.inc
new file mode 100644
index 0000000000..e1b2bccfe1
--- /dev/null
+++ b/example/46_gemm_add_multiply/run_gemm_add_multiply_example.inc
@@ -0,0 +1,139 @@
+#pragma once
+
+bool run_gemm_add_multiply(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    using namespace ck::literals;
+
+    auto& [M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE] = problem_size;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-1, 1});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                               b_device_buf.GetDeviceBuffer(),
+                               {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               {StrideD0, StrideD1},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        std::cout << "wrong! this device_op instance does not support this problem" << std::endl;
+        return true;
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+
+    std::size_t flop      = 2_uz * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(D0DataType) * N + sizeof(D1DataType) * M * N +
+                            sizeof(EDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << device_op.GetTypeString() << std::endl;
+
+    if(config.do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+    }
+
+    return true;
+}
+
+bool run_gemm_add_multiply_example(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    return !parse_cmd_args(argc, argv, problem_size, config) ||
+           run_gemm_add_multiply(problem_size, config);
+}
diff --git a/example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt b/example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt
new file mode 100644
index 0000000000..14432f6e23
--- /dev/null
+++ b/example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt
@@ -0,0 +1,8 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_example_executable(example_gemm_bias_softmax_gemm_permute gemm_bias_softmax_gemm_permute.cpp)
+   set(target 1)
+ endif()
+endforeach()
diff --git a/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp b/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp
new file mode 100644
index 0000000000..a90a6340a4
--- /dev/null
+++ b/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp
@@ -0,0 +1,409 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp    = ck::tensor_operation::element_wise::PassThrough;
+using B0ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using C0DEElementOp = ck::tensor_operation::element_wise::ScaleAdd;
+using Acc0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using B1ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp    = ck::tensor_operation::element_wise::PassThrough;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+constexpr static auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using F16              = ck::half_t;
+using F32              = float;
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using D0DataType       = F16;
+using Acc0BiasDataType = ck::Tuple<D0DataType>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using DeviceOpInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        C0DEElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec,    // MaskingSpecialization
+        1>;
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    int G0      = 3;
+    int G1      = 2;
+    int M       = 1024;
+    int N       = 1024;
+    int K       = 64;
+    int O       = 64;
+    float alpha = 1;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 11)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M  = std::stoi(argv[4]);
+        N  = std::stoi(argv[5]);
+        K  = std::stoi(argv[6]);
+        O  = std::stoi(argv[7]);
+        G0 = std::stoi(argv[8]);
+        G1 = std::stoi(argv[9]);
+
+        alpha = std::stof(argv[10]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 11: M, N, K, O, G0, G1\n");
+        printf("arg10: scale (alpha)\n");
+        exit(0);
+    }
+
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{
+        M * G1 * K, K, G1 * K, 1}; // A layout [G0, M, G1, K]
+
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{
+        N * G1 * K, K, G1 * K, 1}; // B0 layout [G0, N, G1, K]
+
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{
+        N * G1 * O, O, 1, G1 * O}; // B1 layout [G0, N, G1, O]
+
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{
+        M * G1 * O, O, G1 * O, 1}; // C layout [G0, M, G1, O]
+
+    // D layout [G0, M, G1, N]
+    std::vector<ck::index_t> d0_gs_ms_ns_lengths{G0, G1, M, N};
+    std::vector<ck::index_t> d0_gs_ms_ns_strides{M * G1 * N, N, G1 * N, 1};
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+    Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+    Tensor<D0DataType> d0_gs_ms_ns(d0_gs_ms_ns_lengths, d0_gs_ms_ns_strides);
+    Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+    Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
+    std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
+    std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-2, 2});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-1, 1});
+        break;
+    case 3:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * G0 * G1 * M * K);
+    DeviceMem b0_device_buf(sizeof(B0DataType) * G0 * G1 * N * K);
+    DeviceMem d0_device_buf(sizeof(D0DataType) * G0 * G1 * M * N);
+    DeviceMem b1_device_buf(sizeof(B1DataType) * G0 * G1 * O * N);
+    DeviceMem c_device_buf(sizeof(CDataType) * G0 * G1 * M * O);
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
+    d0_device_buf.ToDevice(d0_gs_ms_ns.mData.data());
+
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto c0de_element_op = C0DEElementOp{alpha};
+    auto acc0_element_op = Acc0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    auto argument = device_op.MakeArgument(
+        static_cast<const ADataType*>(a_device_buf.GetDeviceBuffer()),
+        static_cast<const B0DataType*>(b0_device_buf.GetDeviceBuffer()),
+        static_cast<const B1DataType*>(b1_device_buf.GetDeviceBuffer()),
+        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+        std::array<void*, 1>{d0_device_buf.GetDeviceBuffer()}, // p_acc0_biases
+        {},                                                    // p_acc1_biases
+        a_gs_ms_ks_lengths,
+        a_gs_ms_ks_strides,
+        b0_gs_ns_ks_lengths,
+        b0_gs_ns_ks_strides,
+        b1_gs_os_ns_lengths,
+        b1_gs_os_ns_strides,
+        c_gs_ms_os_lengths,
+        c_gs_ms_os_strides,
+        std::array<std::vector<ck::index_t>, 1>{
+            d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+        std::array<std::vector<ck::index_t>, 1>{
+            d0_gs_ms_ns_strides}, // acc0_biases_gs_ms_ns_strides
+        {},                       // acc1_biases_gs_ms_os_lengths
+        {},                       // acc1_biases_gs_ms_os_strides
+        a_element_op,
+        b0_element_op,
+        c0de_element_op,
+        b1_element_op,
+        c_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t BatchCount = G0 * G1;
+    std::size_t flop       = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype =
+        (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N + sizeof(B1DataType) * N * O +
+         sizeof(CDataType) * M * O + sizeof(D0DataType) * M * N) *
+        BatchCount;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << std::endl;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+        Tensor<ADataType> a_g_m_k({BatchCount, M, K});
+        Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
+        Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
+        Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N});        // scratch object after gemm0
+        Tensor<ADataType> a1_g_m_n({BatchCount, M, N});            // scratch object after softmax
+        Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
+        Tensor<D0DataType> d0_g_m_n({BatchCount, M, N});
+
+        // permute
+        a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+            a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+        b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+            b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+            b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        d0_gs_ms_ns.ForEach([&](auto& self, auto idx) {
+            d0_g_m_n(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+
+        // gemm 0
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        acc0_g_m_n.ForEach([&](auto&, auto idx) {
+            c0de_element_op(acc0_g_m_n(idx), acc0_g_m_n(idx), d0_g_m_n(idx));
+        });
+        // masking
+        const auto mask = DeviceOpInstance::C0MatrixMask(N);
+        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+            if(mask.IsMaskedElement(idx[1], idx[2]))
+                self(idx) = -ck::NumericLimits<float>::Infinity();
+        });
+
+        // softmax
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        // gemm1
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        // permute
+        c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+            const size_t& g0 = idx[0];
+            const size_t& g1 = idx[1];
+
+            const size_t g = g0 * G1 + g1;
+
+            self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+        });
+
+        // default absolute error and relative error is 0.001
+        double rtol = 1e-3;
+        double atol = 1e-3;
+
+        return ck::utils::check_err(c_gs_ms_os_device_result.mData,
+                                    c_gs_ms_os_host_result.mData,
+                                    "Error: Incorrect results!",
+                                    rtol,
+                                    atol)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}
diff --git a/example/48_pool3d_fwd/CMakeLists.txt b/example/48_pool3d_fwd/CMakeLists.txt
new file mode 100644
index 0000000000..f821f2c97b
--- /dev/null
+++ b/example/48_pool3d_fwd/CMakeLists.txt
@@ -0,0 +1,3 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp)
+endif()
diff --git a/example/48_pool3d_fwd/pool3d_fwd_common.hpp b/example/48_pool3d_fwd/pool3d_fwd_common.hpp
new file mode 100644
index 0000000000..39032fa123
--- /dev/null
+++ b/example/48_pool3d_fwd/pool3d_fwd_common.hpp
@@ -0,0 +1,198 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+#pragma once
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp"
+
+template <typename TensorLayout>
+std::vector<ck::index_t> f_tensor_strides_ncdhw(ck::index_t N_,
+                                                ck::index_t C_,
+                                                ck::index_t D,
+                                                ck::index_t H,
+                                                ck::index_t W,
+                                                TensorLayout layout)
+{
+    using namespace ck::literals;
+    (void)N_;
+    if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCDHW>::value)
+        return {C_ * D * H * W, D * H * W, H * W, W, 1_uz};
+    else if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NDHWC>::value)
+        return {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_};
+};
+
+template <typename TensorLayout>
+HostTensorDescriptor f_host_tensor_descriptor(std::size_t N_,
+                                              std::size_t C_,
+                                              std::size_t D,
+                                              std::size_t H,
+                                              std::size_t W,
+                                              TensorLayout layout)
+{
+    using namespace ck::literals;
+
+    if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCDHW>::value)
+    {
+        return HostTensorDescriptor({N_, C_, D, H, W}, {C_ * D * H * W, D * H * W, H * W, W, 1_uz});
+    }
+    else if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NDHWC>::value)
+    {
+        return HostTensorDescriptor({N_, C_, D, H, W},
+                                    {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
+    }
+};
+
+template <typename DevicePoolFwdInstance,
+          typename InDataType,
+          typename OutDataType,
+          typename ComputeDataType,
+          typename IndexDataType,
+          typename InLayout,
+          typename OutLayout,
+          ck::ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool OutputIndex>
+bool pool3d_test(bool do_verification,
+                 bool time_kernel,
+                 ck::index_t N,
+                 ck::index_t C,
+                 ck::index_t Z,
+                 ck::index_t Y,
+                 ck::index_t X,
+                 ck::index_t Di,
+                 ck::index_t Hi,
+                 ck::index_t Wi,
+                 ck::index_t window_stride_d,
+                 ck::index_t window_stride_h,
+                 ck::index_t window_stride_w,
+                 ck::index_t window_dilation_d,
+                 ck::index_t window_dilation_h,
+                 ck::index_t window_dilation_w,
+                 ck::index_t in_left_pad_d,
+                 ck::index_t in_left_pad_h,
+                 ck::index_t in_left_pad_w,
+                 ck::index_t in_right_pad_d,
+                 ck::index_t in_right_pad_h,
+                 ck::index_t in_right_pad_w)
+{
+    const ck::index_t Zs = (Z - 1) * window_dilation_d + 1;
+    const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
+    const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
+    const ck::index_t Do = (Di + in_left_pad_d + in_right_pad_d - Zs) / window_stride_d + 1;
+    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
+    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
+
+    const std::vector<ck::index_t> window_spatial_lengths{Z, Y, X};
+    const std::vector<ck::index_t> window_strides{
+        window_stride_d, window_stride_h, window_stride_w};
+    const std::vector<ck::index_t> window_dilations{
+        window_dilation_d, window_dilation_h, window_dilation_w};
+    const std::vector<ck::index_t> input_left_pads{in_left_pad_d, in_left_pad_h, in_left_pad_w};
+    const std::vector<ck::index_t> input_right_pads{in_right_pad_d, in_right_pad_h, in_right_pad_w};
+
+    Tensor<InDataType> in_n_c_di_hi_wi(f_host_tensor_descriptor(N, C, Di, Hi, Wi, InLayout{}));
+    Tensor<OutDataType> out_n_c_do_ho_wo_host(
+        f_host_tensor_descriptor(N, C, Do, Ho, Wo, OutLayout{}));
+    Tensor<IndexDataType> out_indices_n_c_do_ho_wo_host(
+        f_host_tensor_descriptor(N, C, Do, Ho, Wo, OutLayout{}));
+    Tensor<OutDataType> out_n_c_do_ho_wo_device(
+        f_host_tensor_descriptor(N, C, Do, Ho, Wo, OutLayout{}));
+    Tensor<IndexDataType> out_indices_n_c_do_ho_wo_device(
+        f_host_tensor_descriptor(N, C, Do, Ho, Wo, OutLayout{}));
+
+    std::cout << "in_n_c_di_hi_wi: " << in_n_c_di_hi_wi.mDesc << std::endl;
+    std::cout << "out_n_c_do_ho_wo: " << out_n_c_do_ho_wo_host.mDesc << std::endl;
+
+    in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-1.0, 1.0});
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_di_hi_wi.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) *
+                             out_n_c_do_ho_wo_device.mDesc.GetElementSpaceSize());
+    DeviceMem out_indices_device_buf(sizeof(IndexDataType) *
+                                     out_indices_n_c_do_ho_wo_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in_n_c_di_hi_wi.mData.data());
+
+    auto pool         = DevicePoolFwdInstance{};
+    auto invoker_ptr  = pool.MakeInvokerPointer();
+    auto argument_ptr = pool.MakeArgumentPointer(
+        static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+        static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
+        {N, C, Di, Hi, Wi},
+        {Z, Y, X},
+        {N, C, Do, Ho, Wo},
+        f_tensor_strides_ncdhw(N, C, Di, Hi, Wi, InLayout{}),
+        f_tensor_strides_ncdhw(N, C, Do, Ho, Wo, OutLayout{}),
+        f_tensor_strides_ncdhw(N, C, Do, Ho, Wo, OutLayout{}),
+        window_strides,
+        window_dilations,
+        input_left_pads,
+        input_right_pads,
+        {2, 3, 4});
+
+    if(!pool.IsSupportedArgument(argument_ptr.get()))
+    {
+        throw std::runtime_error("wrong! device_op with the specified compilation parameters does "
+                                 "not support this problem");
+    }
+
+    float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+    std::cout << "Perf: " << ave_time << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        using ReferencePoolingFwdInstance =
+            ck::tensor_operation::host::ReferencePoolingFwd<5,
+                                                            3,
+                                                            InDataType,
+                                                            OutDataType,
+                                                            ComputeDataType,
+                                                            IndexDataType,
+                                                            ReduceOpId,
+                                                            PropagateNan,
+                                                            OutputIndex>;
+
+        auto ref_pooling          = ReferencePoolingFwdInstance{};
+        auto ref_pooling_invoker  = ref_pooling.MakeInvoker();
+        auto ref_pooling_argument = ref_pooling.MakeArgument(in_n_c_di_hi_wi,
+                                                             out_n_c_do_ho_wo_host,
+                                                             out_indices_n_c_do_ho_wo_host,
+                                                             window_spatial_lengths,
+                                                             window_strides,
+                                                             window_dilations,
+                                                             input_left_pads,
+                                                             input_right_pads);
+
+        ref_pooling_invoker.Run(ref_pooling_argument);
+
+        out_device_buf.FromDevice(out_n_c_do_ho_wo_device.mData.data());
+
+        pass = pass && ck::utils::check_err(out_n_c_do_ho_wo_device, out_n_c_do_ho_wo_host);
+
+        if constexpr(OutputIndex)
+        {
+            out_indices_device_buf.FromDevice(out_indices_n_c_do_ho_wo_device.mData.data());
+
+            pass = pass && ck::utils::check_err(out_indices_n_c_do_ho_wo_device,
+                                                out_indices_n_c_do_ho_wo_host);
+        };
+    }
+
+    return (pass);
+};
diff --git a/example/48_pool3d_fwd/pool3d_fwd_fp16.cpp b/example/48_pool3d_fwd/pool3d_fwd_fp16.cpp
new file mode 100644
index 0000000000..b9ac61033d
--- /dev/null
+++ b/example/48_pool3d_fwd/pool3d_fwd_fp16.cpp
@@ -0,0 +1,104 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "pool3d_fwd_common.hpp"
+
+using InDataType      = ck::half_t;
+using OutDataType     = ck::half_t;
+using ComputeDataType = float;
+
+using IndexDataType = int32_t;
+
+using InLayout  = ck::tensor_layout::convolution::NDHWC;
+using OutLayout = ck::tensor_layout::convolution::NDHWC;
+
+#if 1
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+#else
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+#endif
+
+static constexpr bool OutputIndex  = false;
+static constexpr bool PropagateNan = false;
+
+using DevicePoolFwdInstance =
+    ck::tensor_operation::device::DevicePool3dFwd_NDHWC_NDHWC<InDataType,
+                                                              OutDataType,
+                                                              IndexDataType,
+                                                              ComputeDataType,
+                                                              ReduceOpId,
+                                                              OutputIndex,
+                                                              64, // BlockSize
+                                                              64, // ReduceMThreadClusterSize
+                                                              1,  // ReduceKThreadClusterSize
+                                                              1,  // ReduceMThreadSliceSize
+                                                              1,  // ReduceKThreadSliceSize
+                                                              1>; // InSrcOutDstVectorSize
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    // Pool shape
+    ck::index_t N                 = 2;
+    ck::index_t C                 = 32;
+    ck::index_t Z                 = 2;
+    ck::index_t Y                 = 2;
+    ck::index_t X                 = 2;
+    ck::index_t Di                = 30;
+    ck::index_t Hi                = 30;
+    ck::index_t Wi                = 30;
+    ck::index_t window_stride_d   = 2;
+    ck::index_t window_stride_h   = 2;
+    ck::index_t window_stride_w   = 2;
+    ck::index_t window_dilation_d = 1;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
+    ck::index_t in_left_pad_d     = 1;
+    ck::index_t in_left_pad_h     = 1;
+    ck::index_t in_left_pad_w     = 1;
+    ck::index_t in_right_pad_d    = 1;
+    ck::index_t in_right_pad_h    = 1;
+    ck::index_t in_right_pad_w    = 1;
+
+    bool pass = pool3d_test<DevicePoolFwdInstance,
+                            InDataType,
+                            OutDataType,
+                            ComputeDataType,
+                            IndexDataType,
+                            InLayout,
+                            OutLayout,
+                            ReduceOpId,
+                            PropagateNan,
+                            OutputIndex>(do_verification,
+                                         time_kernel,
+                                         N,
+                                         C,
+                                         Z,
+                                         Y,
+                                         X,
+                                         Di,
+                                         Hi,
+                                         Wi,
+                                         window_stride_d,
+                                         window_stride_h,
+                                         window_stride_w,
+                                         window_dilation_d,
+                                         window_dilation_h,
+                                         window_dilation_w,
+                                         in_left_pad_d,
+                                         in_left_pad_h,
+                                         in_left_pad_w,
+                                         in_right_pad_d,
+                                         in_right_pad_h,
+                                         in_right_pad_w);
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/49_maxpool2d_bwd/CMakeLists.txt b/example/49_maxpool2d_bwd/CMakeLists.txt
new file mode 100644
index 0000000000..fe98b7c997
--- /dev/null
+++ b/example/49_maxpool2d_bwd/CMakeLists.txt
@@ -0,0 +1,9 @@
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_maxpool2d_bwd_fp16 maxpool2d_bwd_fp16.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_maxpool2d_bwd_fp32 maxpool2d_bwd_fp32.cpp)
+endif()
diff --git a/example/49_maxpool2d_bwd/maxpool2d_bwd_bf16.cpp b/example/49_maxpool2d_bwd/maxpool2d_bwd_bf16.cpp
new file mode 100644
index 0000000000..a6c8f3b884
--- /dev/null
+++ b/example/49_maxpool2d_bwd/maxpool2d_bwd_bf16.cpp
@@ -0,0 +1,66 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "maxpool2d_bwd_common.hpp"
+
+using InDataType      = ck::bhalf_t;
+using OutDataType     = ck::bhalf_t;
+using IndexDataType   = int32_t;
+using ComputeDataType = float;
+using DInDataType     = ck::bhalf_t;
+using DOutDataType    = ck::bhalf_t;
+
+static constexpr bool PropagateNan = false;
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    // Pool shape
+    ck::index_t N                 = 1;
+    ck::index_t C                 = 1;
+    ck::index_t Y                 = 3;
+    ck::index_t X                 = 3;
+    ck::index_t Hi                = 32;
+    ck::index_t Wi                = 32;
+    ck::index_t window_stride_h   = 1;
+    ck::index_t window_stride_w   = 1;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
+    ck::index_t in_left_pad_h     = 0;
+    ck::index_t in_left_pad_w     = 0;
+    ck::index_t in_right_pad_h    = 0;
+    ck::index_t in_right_pad_w    = 0;
+
+    bool pass = maxpool_bwd_test<InDataType,
+                                 OutDataType,
+                                 IndexDataType,
+                                 ComputeDataType,
+                                 DInDataType,
+                                 DOutDataType,
+                                 PropagateNan>(do_verification,
+                                               time_kernel,
+                                               N,
+                                               C,
+                                               Y,
+                                               X,
+                                               Hi,
+                                               Wi,
+                                               window_stride_h,
+                                               window_stride_w,
+                                               window_dilation_h,
+                                               window_dilation_w,
+                                               in_left_pad_h,
+                                               in_left_pad_w,
+                                               in_right_pad_h,
+                                               in_right_pad_w);
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/49_maxpool2d_bwd/maxpool2d_bwd_common.hpp b/example/49_maxpool2d_bwd/maxpool2d_bwd_common.hpp
new file mode 100644
index 0000000000..2c1e669375
--- /dev/null
+++ b/example/49_maxpool2d_bwd/maxpool2d_bwd_common.hpp
@@ -0,0 +1,229 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp"
+
+template <typename InDataType,
+          typename OutDataType,
+          typename IndexDataType,
+          typename ComputeDataType,
+          typename DInDataType,
+          typename DOutDataType,
+          bool PropagateNan>
+bool maxpool_bwd_test(bool do_verification,
+                      bool time_kernel,
+                      ck::index_t N,
+                      ck::index_t C,
+                      ck::index_t Y,
+                      ck::index_t X,
+                      ck::index_t Hi,
+                      ck::index_t Wi,
+                      ck::index_t window_stride_h,
+                      ck::index_t window_stride_w,
+                      ck::index_t window_dilation_h,
+                      ck::index_t window_dilation_w,
+                      ck::index_t in_left_pad_h,
+                      ck::index_t in_left_pad_w,
+                      ck::index_t in_right_pad_h,
+                      ck::index_t in_right_pad_w)
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    using DevicePoolFwdInstance =
+        ck::tensor_operation::device::DevicePool2dFwd_NHWC_NHWC<InDataType,      // InDataType
+                                                                OutDataType,     // OutDataType
+                                                                IndexDataType,   // IndexDataType
+                                                                ComputeDataType, // ComputeDataType
+                                                                ck::ReduceTensorOp::MAX,
+                                                                true,
+                                                                64, // BlockSize
+                                                                64, // ReduceMThreadClusterSize
+                                                                1,  // ReduceKThreadClusterSize
+                                                                4,  // ReduceMThreadSliceSize
+                                                                1,  // ReduceKThreadSliceSize
+                                                                1>; // InSrcOutDstVectorSize
+
+    using DeviceMaxPoolBwdInstance = ck::tensor_operation::device::
+        DeviceMaxPoolBwdImpl<DOutDataType, IndexDataType, DInDataType, 4>;
+
+    const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
+    const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
+    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Ys) / window_stride_h + 1;
+    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - Xs) / window_stride_w + 1;
+
+    const std::vector<ck::index_t> window_spatial_lengths{Y, X};
+    const std::vector<ck::index_t> window_strides{window_stride_h, window_stride_w};
+    const std::vector<ck::index_t> window_dilations{window_dilation_h, window_dilation_w};
+    const std::vector<ck::index_t> input_left_pads{in_left_pad_h, in_left_pad_w};
+    const std::vector<ck::index_t> input_right_pads{in_right_pad_h, in_right_pad_w};
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W) {
+            using namespace ck::literals;
+            // reference need Tensor with NCHW order
+            return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, 1_uz, W * C_, C_});
+        };
+
+    // in
+    Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi));
+
+    // out
+    Tensor<OutDataType> out_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo));
+    Tensor<OutDataType> out_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo));
+
+    // indices
+    Tensor<IndexDataType> indices_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo));
+    Tensor<IndexDataType> indices_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo));
+
+    // dout
+    Tensor<DOutDataType> dout_n_c_ho_wo(f_host_tensor_descriptor(N, C, Ho, Wo));
+
+    // din
+    Tensor<DInDataType> din_n_c_hi_wi_host(f_host_tensor_descriptor(N, C, Hi, Wi));
+    Tensor<DInDataType> din_n_c_hi_wi_device(f_host_tensor_descriptor(N, C, Hi, Wi));
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "out_n_c_ho_wo: " << out_n_c_ho_wo_host.mDesc << std::endl;
+    std::cout << "indices_n_c_ho_wo: " << indices_n_c_ho_wo_host.mDesc << std::endl;
+    std::cout << "dout_n_c_ho_wo: " << dout_n_c_ho_wo.mDesc << std::endl;
+    std::cout << "din_n_c_hi_wi: " << din_n_c_hi_wi_host.mDesc << std::endl;
+
+    in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-1.0, 1.0});
+    dout_n_c_ho_wo.GenerateTensorValue(GeneratorTensor_3<DOutDataType>{-1.0, 1.0});
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) *
+                             out_n_c_ho_wo_device.mDesc.GetElementSpaceSize());
+    DeviceMem indices_device_buf(sizeof(IndexDataType) *
+                                 indices_n_c_ho_wo_device.mDesc.GetElementSpaceSize());
+    DeviceMem dout_device_buf(sizeof(DOutDataType) * dout_n_c_ho_wo.mDesc.GetElementSpaceSize());
+    DeviceMem din_device_buf(sizeof(DInDataType) *
+                             din_n_c_hi_wi_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+    dout_device_buf.ToDevice(dout_n_c_ho_wo.mData.data());
+
+    auto pool_fwd              = DevicePoolFwdInstance{};
+    auto pool_fwd_invoker_ptr  = pool_fwd.MakeInvokerPointer();
+    auto pool_fwd_argument_ptr = pool_fwd.MakeArgumentPointer(
+        static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+        static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
+        {N, C, Hi, Wi},
+        window_spatial_lengths,
+        {N, C, Ho, Wo},
+        {C * Hi * Wi, 1, Wi * C, C},
+        {C * Ho * Wo, 1, Wo * C, C},
+        {C * Ho * Wo, 1, Wo * C, C},
+        window_strides,
+        window_dilations,
+        input_left_pads,
+        input_right_pads,
+        {2, 3});
+
+    if(!pool_fwd.IsSupportedArgument(pool_fwd_argument_ptr.get()))
+    {
+        throw std::runtime_error("wrong! pool_fwd with the specified compilation parameters does "
+                                 "not support this problem");
+    }
+
+    float ave_time_fwd =
+        pool_fwd_invoker_ptr->Run(pool_fwd_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    auto pool_bwd              = DeviceMaxPoolBwdInstance{};
+    auto pool_bwd_invoker_ptr  = pool_bwd.MakeInvokerPointer();
+    auto pool_bwd_argument_ptr = pool_bwd.MakeArgumentPointer(
+        static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
+        static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
+        static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+        dout_n_c_ho_wo.mDesc.GetElementSpaceSize(),
+        din_n_c_hi_wi_device.mDesc.GetElementSpaceSize(),
+        window_spatial_lengths,
+        window_strides,
+        window_dilations);
+
+    if(!pool_bwd.IsSupportedArgument(pool_bwd_argument_ptr.get()))
+    {
+        throw std::runtime_error("wrong! pool_bwd with the specified compilation parameters does "
+                                 "not support this problem");
+    }
+
+    size_t pool_bwd_workspace_sz = pool_bwd.GetWorkSpaceSize(pool_bwd_argument_ptr.get());
+    DeviceMem pool_bwd_workspace_device_buf(pool_bwd_workspace_sz);
+    pool_bwd.SetWorkSpacePointer(pool_bwd_argument_ptr.get(),
+                                 pool_bwd_workspace_device_buf.GetDeviceBuffer());
+
+    float ave_time_bwd =
+        pool_bwd_invoker_ptr->Run(pool_bwd_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Pool fwd perf: " << ave_time_fwd << " ms" << std::endl;
+    std::cout << "Pool bwd perf: " << ave_time_bwd << " ms" << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        using ReferencePoolingFwdInstance =
+            ck::tensor_operation::host::ReferencePoolingFwd<4,
+                                                            2,
+                                                            InDataType,
+                                                            OutDataType,
+                                                            ComputeDataType,
+                                                            IndexDataType,
+                                                            ck::ReduceTensorOp::MAX,
+                                                            PropagateNan,
+                                                            true>;
+
+        auto ref_pooling_fwd          = ReferencePoolingFwdInstance{};
+        auto ref_pooling_fwd_invoker  = ref_pooling_fwd.MakeInvoker();
+        auto ref_pooling_fwd_argument = ref_pooling_fwd.MakeArgument(in_n_c_hi_wi,
+                                                                     out_n_c_ho_wo_host,
+                                                                     indices_n_c_ho_wo_host,
+                                                                     window_spatial_lengths,
+                                                                     window_strides,
+                                                                     window_dilations,
+                                                                     input_left_pads,
+                                                                     input_right_pads);
+        ref_pooling_fwd_invoker.Run(ref_pooling_fwd_argument);
+
+        using ReferencePoolingBwdInstance =
+            ck::tensor_operation::host::ReferenceMaxPoolBwd<DOutDataType,
+                                                            IndexDataType,
+                                                            ComputeDataType,
+                                                            DInDataType,
+                                                            PassThrough>;
+
+        auto ref_pooling_bwd          = ReferencePoolingBwdInstance{};
+        auto ref_pooling_bwd_invoker  = ref_pooling_bwd.MakeInvoker();
+        auto ref_pooling_bwd_argument = ref_pooling_bwd.MakeArgument(
+            dout_n_c_ho_wo, indices_n_c_ho_wo_host, din_n_c_hi_wi_host, PassThrough{});
+
+        ref_pooling_bwd_invoker.Run(ref_pooling_bwd_argument);
+
+        out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
+        indices_device_buf.FromDevice(indices_n_c_ho_wo_device.mData.data());
+        din_device_buf.FromDevice(din_n_c_hi_wi_device.mData.data());
+
+        pass = pass && ck::utils::check_err(out_n_c_ho_wo_device, out_n_c_ho_wo_host);
+        pass = pass && ck::utils::check_err(indices_n_c_ho_wo_device, indices_n_c_ho_wo_host);
+        pass = pass && ck::utils::check_err(din_n_c_hi_wi_device, din_n_c_hi_wi_host);
+    }
+
+    return (pass);
+};
diff --git a/example/49_maxpool2d_bwd/maxpool2d_bwd_fp16.cpp b/example/49_maxpool2d_bwd/maxpool2d_bwd_fp16.cpp
new file mode 100644
index 0000000000..a4f982d855
--- /dev/null
+++ b/example/49_maxpool2d_bwd/maxpool2d_bwd_fp16.cpp
@@ -0,0 +1,66 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "maxpool2d_bwd_common.hpp"
+
+using InDataType      = ck::half_t;
+using OutDataType     = ck::half_t;
+using IndexDataType   = int32_t;
+using ComputeDataType = float;
+using DInDataType     = ck::half_t;
+using DOutDataType    = ck::half_t;
+
+static constexpr bool PropagateNan = false;
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    // Pool shape
+    ck::index_t N                 = 1;
+    ck::index_t C                 = 1;
+    ck::index_t Y                 = 3;
+    ck::index_t X                 = 3;
+    ck::index_t Hi                = 32;
+    ck::index_t Wi                = 32;
+    ck::index_t window_stride_h   = 1;
+    ck::index_t window_stride_w   = 1;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
+    ck::index_t in_left_pad_h     = 0;
+    ck::index_t in_left_pad_w     = 0;
+    ck::index_t in_right_pad_h    = 0;
+    ck::index_t in_right_pad_w    = 0;
+
+    bool pass = maxpool_bwd_test<InDataType,
+                                 OutDataType,
+                                 IndexDataType,
+                                 ComputeDataType,
+                                 DInDataType,
+                                 DOutDataType,
+                                 PropagateNan>(do_verification,
+                                               time_kernel,
+                                               N,
+                                               C,
+                                               Y,
+                                               X,
+                                               Hi,
+                                               Wi,
+                                               window_stride_h,
+                                               window_stride_w,
+                                               window_dilation_h,
+                                               window_dilation_w,
+                                               in_left_pad_h,
+                                               in_left_pad_w,
+                                               in_right_pad_h,
+                                               in_right_pad_w);
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/49_maxpool2d_bwd/maxpool2d_bwd_fp32.cpp b/example/49_maxpool2d_bwd/maxpool2d_bwd_fp32.cpp
new file mode 100644
index 0000000000..c14928a936
--- /dev/null
+++ b/example/49_maxpool2d_bwd/maxpool2d_bwd_fp32.cpp
@@ -0,0 +1,66 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "maxpool2d_bwd_common.hpp"
+
+using InDataType      = float;
+using OutDataType     = float;
+using IndexDataType   = int32_t;
+using ComputeDataType = float;
+using DInDataType     = float;
+using DOutDataType    = float;
+
+static constexpr bool PropagateNan = false;
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    // Pool shape
+    ck::index_t N                 = 1;
+    ck::index_t C                 = 1;
+    ck::index_t Y                 = 2;
+    ck::index_t X                 = 2;
+    ck::index_t Hi                = 32;
+    ck::index_t Wi                = 32;
+    ck::index_t window_stride_h   = 2;
+    ck::index_t window_stride_w   = 2;
+    ck::index_t window_dilation_h = 1;
+    ck::index_t window_dilation_w = 1;
+    ck::index_t in_left_pad_h     = 0;
+    ck::index_t in_left_pad_w     = 0;
+    ck::index_t in_right_pad_h    = 0;
+    ck::index_t in_right_pad_w    = 0;
+
+    bool pass = maxpool_bwd_test<InDataType,
+                                 OutDataType,
+                                 IndexDataType,
+                                 ComputeDataType,
+                                 DInDataType,
+                                 DOutDataType,
+                                 PropagateNan>(do_verification,
+                                               time_kernel,
+                                               N,
+                                               C,
+                                               Y,
+                                               X,
+                                               Hi,
+                                               Wi,
+                                               window_stride_h,
+                                               window_stride_w,
+                                               window_dilation_h,
+                                               window_dilation_w,
+                                               in_left_pad_h,
+                                               in_left_pad_w,
+                                               in_right_pad_h,
+                                               in_right_pad_w);
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/50_put_element/CMakeLists.txt b/example/50_put_element/CMakeLists.txt
new file mode 100644
index 0000000000..eca410008f
--- /dev/null
+++ b/example/50_put_element/CMakeLists.txt
@@ -0,0 +1,3 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_put_element_fp16 put_element_fp16.cpp)
+endif()
diff --git a/example/50_put_element/put_element_fp16.cpp b/example/50_put_element/put_element_fp16.cpp
new file mode 100644
index 0000000000..791747d875
--- /dev/null
+++ b/example/50_put_element/put_element_fp16.cpp
@@ -0,0 +1,88 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_put_element_impl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+using XDataType     = ck::half_t;
+using YDataType     = ck::half_t;
+using IndexDataType = int32_t;
+
+using YElementwiseOp = ck::tensor_operation::element_wise::PassThrough;
+
+using DeviceInstance =
+    ck::tensor_operation::device::DevicePutElementImpl<XDataType,     // XDataType
+                                                       IndexDataType, // IndexDataType
+                                                       YDataType,     // YDataType
+                                                       YElementwiseOp,
+                                                       ck::InMemoryDataOperationEnum::Set,
+                                                       1>;
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    int N = 1024;
+
+    Tensor<XDataType> x(HostTensorDescriptor{N});
+    Tensor<IndexDataType> indices(HostTensorDescriptor{N});
+    Tensor<YDataType> y(HostTensorDescriptor{N});
+
+    x.GenerateTensorValue(GeneratorTensor_3<XDataType>{-1.0, 1.0});
+    for(int i = 0; i < N; ++i)
+        indices(i) = i;
+
+    DeviceMem x_device_buf(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem y_device_buf(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+    DeviceMem indices_device_buf(sizeof(IndexDataType) * indices.mDesc.GetElementSpaceSize());
+
+    x_device_buf.ToDevice(x.mData.data());
+    indices_device_buf.ToDevice(indices.mData.data());
+
+    auto put_instance     = DeviceInstance{};
+    auto put_invoker_ptr  = put_instance.MakeInvokerPointer();
+    auto put_argument_ptr = put_instance.MakeArgumentPointer(
+        static_cast<XDataType*>(x_device_buf.GetDeviceBuffer()),
+        static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
+        static_cast<YDataType*>(y_device_buf.GetDeviceBuffer()),
+        N,
+        N,
+        YElementwiseOp{});
+
+    if(!put_instance.IsSupportedArgument(put_argument_ptr.get()))
+    {
+        throw std::runtime_error("argument is not supported!");
+    }
+
+    float ave_time =
+        put_invoker_ptr->Run(put_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        Tensor<YDataType> y_host(HostTensorDescriptor{N});
+
+        for(int i = 0; i < N; ++i)
+        {
+            IndexDataType idx = indices(i);
+            y_host(idx)       = x(i);
+        }
+
+        y_device_buf.FromDevice(y.mData.data());
+        pass = ck::utils::check_err(y, y_host);
+    }
+
+    return (pass ? 0 : 1);
+}
diff --git a/example/51_avgpool3d_bwd/CMakeLists.txt b/example/51_avgpool3d_bwd/CMakeLists.txt
new file mode 100644
index 0000000000..fef0c66835
--- /dev/null
+++ b/example/51_avgpool3d_bwd/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_example_executable(example_avgpool3d_bwd_bf16 avgpool3d_bwd_bf16.cpp)
+add_example_executable(example_avgpool3d_bwd_fp16 avgpool3d_bwd_fp16.cpp)
+add_example_executable(example_avgpool3d_bwd_fp32 avgpool3d_bwd_fp32.cpp)
diff --git a/example/51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp b/example/51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp
new file mode 100644
index 0000000000..a5ab6ef24a
--- /dev/null
+++ b/example/51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp"
+
+#include "avgpool3d_bwd_common.hpp"
+
+using DOutDataType    = ck::bhalf_t;
+using DInDataType     = ck::bhalf_t;
+using ComputeDataType = float;
+
+#if 1
+using DOutLayout = ck::tensor_layout::convolution::NDHWC;
+using DInLayout  = ck::tensor_layout::convolution::NDHWC;
+#else
+using DOutLayout = ck::tensor_layout::convolution::NCDHW;
+using DInLayout  = ck::tensor_layout::convolution::NCDHW;
+#endif
+
+using DevicePoolBwdInstance =
+    ck::tensor_operation::device::DeviceAvgPool3dBwd_NDHWC_NDHWC<DOutDataType,
+                                                                 DInDataType,
+                                                                 ComputeDataType,
+                                                                 64, // BlockSize
+                                                                 64, // ReduceMThreadClusterSize
+                                                                 1,  // ReduceKThreadClusterSize
+                                                                 1,  // ReduceMThreadSliceSize
+                                                                 1,  // ReduceKThreadSliceSize
+                                                                 1>; // InSrcOutDstVectorSize
+
+int main()
+{
+    std::vector<ck::index_t> window_lengths    = {5, 5, 5};
+    std::vector<ck::index_t> window_strides    = {2, 2, 2};
+    std::vector<ck::index_t> window_dilations  = {2, 2, 2};
+    std::vector<ck::index_t> dinput_left_pads  = {0, 0, 0};
+    std::vector<ck::index_t> dinput_right_pads = {0, 0, 0};
+
+    ck::index_t N  = 1;
+    ck::index_t C  = 16;
+    ck::index_t Di = 40;
+    ck::index_t Hi = 40;
+    ck::index_t Wi = 40;
+
+    pool3d_bwd_test<DevicePoolBwdInstance, DOutDataType, DInDataType, DOutLayout, DInLayout>(
+        true,
+        false,
+        N,
+        C,
+        Di,
+        Hi,
+        Wi,
+        window_lengths,
+        window_strides,
+        window_dilations,
+        dinput_left_pads,
+        dinput_right_pads);
+}
diff --git a/example/51_avgpool3d_bwd/avgpool3d_bwd_common.hpp b/example/51_avgpool3d_bwd/avgpool3d_bwd_common.hpp
new file mode 100644
index 0000000000..394f046b1e
--- /dev/null
+++ b/example/51_avgpool3d_bwd/avgpool3d_bwd_common.hpp
@@ -0,0 +1,147 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp"
+
+template <typename TensorLayout>
+std::vector<ck::index_t> f_tensor_strides_ncdhw(ck::index_t N_,
+                                                ck::index_t C_,
+                                                ck::index_t D,
+                                                ck::index_t H,
+                                                ck::index_t W,
+                                                TensorLayout layout)
+{
+    using namespace ck::literals;
+    (void)N_;
+    if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCDHW>::value)
+        return {C_ * D * H * W, D * H * W, H * W, W, 1_uz};
+    else if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NDHWC>::value)
+        return {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_};
+};
+
+template <typename TensorLayout>
+HostTensorDescriptor f_host_tensor_descriptor(std::size_t N_,
+                                              std::size_t C_,
+                                              std::size_t D,
+                                              std::size_t H,
+                                              std::size_t W,
+                                              TensorLayout layout)
+{
+    using namespace ck::literals;
+
+    if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCDHW>::value)
+    {
+        return HostTensorDescriptor({N_, C_, D, H, W}, {C_ * D * H * W, D * H * W, H * W, W, 1_uz});
+    }
+    else if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NDHWC>::value)
+    {
+        return HostTensorDescriptor({N_, C_, D, H, W},
+                                    {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
+    }
+};
+
+template <typename DevicePoolBwdInstance,
+          typename DOutDataType,
+          typename DInDataType,
+          typename DOutLayout,
+          typename DInLayout>
+bool pool3d_bwd_test(bool do_verification,
+                     bool time_kernel,
+                     ck::index_t N,
+                     ck::index_t C,
+                     ck::index_t Di,
+                     ck::index_t Hi,
+                     ck::index_t Wi,
+                     std::vector<ck::index_t> window_lengths,
+                     std::vector<ck::index_t> window_strides,
+                     std::vector<ck::index_t> window_dilations,
+                     std::vector<ck::index_t> dinput_left_pads,
+                     std::vector<ck::index_t> dinput_right_pads)
+{
+    auto OutSpatialLength = [&](auto InSpatialLength, int index) {
+        ck::index_t left_pad   = dinput_left_pads[index];
+        ck::index_t right_pad  = dinput_right_pads[index];
+        ck::index_t window_len = window_lengths[index];
+        ck::index_t stride     = window_strides[index];
+        ck::index_t dilation   = window_dilations[index];
+        ck::index_t eff        = (window_len - 1) * dilation + 1;
+        return (InSpatialLength + left_pad + right_pad - eff) / stride + 1;
+    };
+
+    ck::index_t Do = OutSpatialLength(Di, 0);
+    ck::index_t Ho = OutSpatialLength(Hi, 1);
+    ck::index_t Wo = OutSpatialLength(Wi, 2);
+
+    Tensor<DOutDataType> dout(f_host_tensor_descriptor(N, C, Do, Ho, Wo, DOutLayout{}));
+    Tensor<DInDataType> din_dev(f_host_tensor_descriptor(N, C, Di, Hi, Wi, DInLayout{}));
+    Tensor<DInDataType> din_host(f_host_tensor_descriptor(N, C, Di, Hi, Wi, DInLayout{}));
+
+    std::cout << "dout: " << dout.mDesc << std::endl;
+    std::cout << "din_host: " << din_host.mDesc << std::endl;
+
+    dout.GenerateTensorValue(GeneratorTensor_3<DOutDataType>{0.0, 1.0});
+
+    DeviceMem dout_device_buf(sizeof(DOutDataType) * dout.mDesc.GetElementSpaceSize());
+    DeviceMem din_device_buf(sizeof(DInDataType) * din_dev.mDesc.GetElementSpaceSize());
+
+    dout_device_buf.ToDevice(dout.mData.data());
+    din_device_buf.SetZero();
+
+    auto pool        = DevicePoolBwdInstance{};
+    auto invoker_ptr = pool.MakeInvokerPointer();
+    auto argument_ptr =
+        pool.MakeArgumentPointer(static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
+                                 static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+                                 {N, C, Do, Ho, Wo},
+                                 {N, C, Di, Hi, Wi},
+                                 f_tensor_strides_ncdhw(N, C, Do, Ho, Wo, DOutLayout{}),
+                                 f_tensor_strides_ncdhw(N, C, Di, Hi, Wi, DInLayout{}),
+                                 window_lengths,
+                                 window_strides,
+                                 window_dilations,
+                                 dinput_left_pads,
+                                 dinput_right_pads);
+
+    if(!pool.IsSupportedArgument(argument_ptr.get()))
+    {
+        throw std::runtime_error("wrong! device_op with the specified compilation parameters does "
+                                 "not support this problem");
+    }
+
+    float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+    std::cout << "Perf: " << ave_time << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        auto ref_pool =
+            ck::tensor_operation::host::ReferenceAvgPoolBwd<3, DInDataType, DOutDataType>();
+
+        auto ref_invoker = ref_pool.MakeInvoker();
+
+        auto ref_argument = ref_pool.MakeArgument(din_host,
+                                                  dout,
+                                                  window_lengths,
+                                                  window_strides,
+                                                  window_dilations,
+                                                  dinput_left_pads,
+                                                  dinput_right_pads);
+
+        ref_invoker.Run(ref_argument);
+
+        din_device_buf.FromDevice(din_dev.mData.data());
+        pass = ck::utils::check_err(din_dev, din_host);
+    }
+
+    return pass;
+}
diff --git a/example/51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp b/example/51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp
new file mode 100644
index 0000000000..578f563d5c
--- /dev/null
+++ b/example/51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp"
+
+#include "avgpool3d_bwd_common.hpp"
+
+using DOutDataType    = ck::half_t;
+using DInDataType     = ck::half_t;
+using ComputeDataType = float;
+
+#if 1
+using DOutLayout = ck::tensor_layout::convolution::NDHWC;
+using DInLayout  = ck::tensor_layout::convolution::NDHWC;
+#else
+using DOutLayout = ck::tensor_layout::convolution::NCDHW;
+using DInLayout  = ck::tensor_layout::convolution::NCDHW;
+#endif
+
+using DevicePoolBwdInstance =
+    ck::tensor_operation::device::DeviceAvgPool3dBwd_NDHWC_NDHWC<DOutDataType,
+                                                                 DInDataType,
+                                                                 ComputeDataType,
+                                                                 64, // BlockSize
+                                                                 64, // ReduceMThreadClusterSize
+                                                                 1,  // ReduceKThreadClusterSize
+                                                                 1,  // ReduceMThreadSliceSize
+                                                                 1,  // ReduceKThreadSliceSize
+                                                                 1>; // InSrcOutDstVectorSize
+
+int main()
+{
+    std::vector<ck::index_t> window_lengths    = {5, 5, 5};
+    std::vector<ck::index_t> window_strides    = {2, 2, 2};
+    std::vector<ck::index_t> window_dilations  = {2, 2, 2};
+    std::vector<ck::index_t> dinput_left_pads  = {0, 0, 0};
+    std::vector<ck::index_t> dinput_right_pads = {0, 0, 0};
+
+    ck::index_t N  = 1;
+    ck::index_t C  = 16;
+    ck::index_t Di = 40;
+    ck::index_t Hi = 40;
+    ck::index_t Wi = 40;
+
+    pool3d_bwd_test<DevicePoolBwdInstance, DOutDataType, DInDataType, DOutLayout, DInLayout>(
+        true,
+        false,
+        N,
+        C,
+        Di,
+        Hi,
+        Wi,
+        window_lengths,
+        window_strides,
+        window_dilations,
+        dinput_left_pads,
+        dinput_right_pads);
+}
diff --git a/example/51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp b/example/51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp
new file mode 100644
index 0000000000..c2910c5567
--- /dev/null
+++ b/example/51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp"
+
+#include "avgpool3d_bwd_common.hpp"
+
+using DOutDataType    = float;
+using DInDataType     = float;
+using ComputeDataType = float;
+
+#if 1
+using DOutLayout = ck::tensor_layout::convolution::NDHWC;
+using DInLayout  = ck::tensor_layout::convolution::NDHWC;
+#else
+using DOutLayout = ck::tensor_layout::convolution::NCDHW;
+using DInLayout  = ck::tensor_layout::convolution::NCDHW;
+#endif
+
+using DevicePoolBwdInstance =
+    ck::tensor_operation::device::DeviceAvgPool3dBwd_NDHWC_NDHWC<DOutDataType,
+                                                                 DInDataType,
+                                                                 ComputeDataType,
+                                                                 64, // BlockSize
+                                                                 64, // ReduceMThreadClusterSize
+                                                                 1,  // ReduceKThreadClusterSize
+                                                                 1,  // ReduceMThreadSliceSize
+                                                                 1,  // ReduceKThreadSliceSize
+                                                                 1>; // InSrcOutDstVectorSize
+
+int main()
+{
+    std::vector<ck::index_t> window_lengths    = {5, 5, 5};
+    std::vector<ck::index_t> window_strides    = {2, 2, 2};
+    std::vector<ck::index_t> window_dilations  = {2, 2, 2};
+    std::vector<ck::index_t> dinput_left_pads  = {0, 0, 0};
+    std::vector<ck::index_t> dinput_right_pads = {0, 0, 0};
+
+    ck::index_t N  = 1;
+    ck::index_t C  = 16;
+    ck::index_t Di = 40;
+    ck::index_t Hi = 40;
+    ck::index_t Wi = 40;
+
+    pool3d_bwd_test<DevicePoolBwdInstance, DOutDataType, DInDataType, DOutLayout, DInLayout>(
+        true,
+        false,
+        N,
+        C,
+        Di,
+        Hi,
+        Wi,
+        window_lengths,
+        window_strides,
+        window_dilations,
+        dinput_left_pads,
+        dinput_right_pads);
+}
diff --git a/example/52_image_to_column/CMakeLists.txt b/example/52_image_to_column/CMakeLists.txt
new file mode 100644
index 0000000000..226e1fc5ae
--- /dev/null
+++ b/example/52_image_to_column/CMakeLists.txt
@@ -0,0 +1,10 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_custom_target(example_image_to_column)
+   add_example_executable(example_image_to_column_f32 image_to_column_f32.cpp)
+   add_dependencies(example_image_to_column example_image_to_column_f32)
+   set(target 1)
+ endif()
+endforeach()
diff --git a/example/52_image_to_column/common.hpp b/example/52_image_to_column/common.hpp
new file mode 100644
index 0000000000..8510fa1e6d
--- /dev/null
+++ b/example/52_image_to_column/common.hpp
@@ -0,0 +1,95 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <initializer_list>
+#include <iostream>
+#include <numeric>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static inline constexpr ck::index_t NDimSpatial = 2;
+
+using FP32 = float;
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+};
+
+#define DefaultConvParams                                                            \
+    ck::utils::conv::ConvParam                                                       \
+    {                                                                                \
+        NDimSpatial, 1, 32, 1, 1, {4, 4}, {64, 64}, {1, 1}, {1, 1}, {0, 0}, { 0, 0 } \
+    }
+
+inline void print_help_msg()
+{
+    std::cerr << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+inline bool parse_cmd_args(int argc,
+                           char* argv[],
+                           ExecutionConfig& config,
+                           ck::utils::conv::ConvParam& conv_params)
+{
+    constexpr int num_execution_config_args =
+        3; // arguments for do_verification, init_method, time_kernel
+    constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
+
+    constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
+    constexpr int threshold_to_catch_all_args =
+        threshold_to_catch_partial_args + num_conv_param_leading_args;
+
+    if(argc == 1)
+    {
+        // use default
+        config = ExecutionConfig{};
+    }
+    // catch only ExecutionConfig arguments
+    else if(argc == threshold_to_catch_partial_args)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    // catch both ExecutionConfig & ConvParam arguments
+    else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+        conv_params                       = ck::utils::conv::parse_conv_param(
+            num_dim_spatial, threshold_to_catch_partial_args, argv);
+    }
+    else
+    {
+        print_help_msg();
+        return false;
+    }
+
+    return true;
+}
diff --git a/example/52_image_to_column/image_to_column_f32.cpp b/example/52_image_to_column/image_to_column_f32.cpp
new file mode 100644
index 0000000000..c8a7e5f221
--- /dev/null
+++ b/example/52_image_to_column/image_to_column_f32.cpp
@@ -0,0 +1,166 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using InDataType  = FP32;
+using OutDataType = FP32;
+
+using InLayout = ck::tensor_layout::convolution::GNHWC;
+
+// clang-format off
+using DeviceImgToColInstance = ck::tensor_operation::device::DeviceImageToColumnImpl
+        //#####################|        Num| InLayout| InDataType| OutDataType| Block|  MPer|  KPer|    Thread| Scalar|
+        //#####################|        Dim|         |           |            |  Size| Block| Block|   Cluster|    Per|
+        //#####################|    Spatial|         |           |            |      |      |      |   Lengths| Vector|
+        //#####################|           |         |           |            |      |      |      |          |       |
+                              < NDimSpatial, InLayout, InDataType, OutDataType,   256,   128,   128, S<16, 16>,     1>;
+// clang-format on
+
+bool RunImageToColumn(const ExecutionConfig& config, const ck::utils::conv::ConvParam& conv_params)
+{
+
+    const auto N = conv_params.N_;
+    const auto C = conv_params.C_;
+
+    const ck::index_t NDoHoWo =
+        N * ck::accumulate_n<ck::index_t>(
+                conv_params.output_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+    const ck::index_t CZYX =
+        C * ck::accumulate_n<ck::index_t>(
+                conv_params.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+
+    const auto in_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_params);
+    const auto out_desc = HostTensorDescriptor({NDoHoWo, CZYX});
+
+    std::array<ck::index_t, NDimSpatial> input_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> input_g_n_c_wis_strides{};
+    std::array<ck::index_t, 2> output_m_k_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
+
+    copy(conv_params.input_spatial_lengths_, input_spatial_lengths);
+    copy(conv_params.filter_spatial_lengths_, filter_spatial_lengths);
+    copy(conv_params.output_spatial_lengths_, output_spatial_lengths);
+    copy(in_desc.GetStrides(), input_g_n_c_wis_strides);
+    copy(out_desc.GetStrides(), output_m_k_strides);
+    copy(conv_params.conv_filter_strides_, conv_filter_strides);
+    copy(conv_params.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_params.input_left_pads_, input_left_pads);
+    copy(conv_params.input_right_pads_, input_right_pads);
+
+    Tensor<InDataType> in(in_desc);
+    Tensor<OutDataType> out_device(out_desc);
+    Tensor<OutDataType> out_host(out_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "out: " << out_device.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1: in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
+    default: in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+
+    // reset input to zero
+    out_device_buf.SetZero();
+
+    static_assert(std::is_default_constructible_v<DeviceImgToColInstance>);
+
+    // do conv
+    auto img2col  = DeviceImgToColInstance{};
+    auto invoker  = img2col.MakeInvoker();
+    auto argument = img2col.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                         out_device_buf.GetDeviceBuffer(),
+                                         N,
+                                         C,
+                                         input_spatial_lengths,
+                                         filter_spatial_lengths,
+                                         output_spatial_lengths,
+                                         input_g_n_c_wis_strides,
+                                         output_m_k_strides,
+                                         conv_filter_strides,
+                                         conv_filter_dilations,
+                                         input_left_pads,
+                                         input_right_pads);
+
+    if(!img2col.IsSupportedArgument(argument))
+    {
+        std::cerr << "wrong! device_img2col with the specified compilation parameters does "
+                     "not support this img2col problem"
+                  << std::endl;
+
+        return false;
+    }
+
+    float ave_time        = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+    std::size_t num_btype = NDoHoWo * CZYX * (sizeof(OutDataType) + sizeof(InDataType));
+    float gb_per_sec      = num_btype / 1.E6 / ave_time;
+    std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+
+    if(config.do_verification)
+    {
+        auto ref_image_to_column = ck::tensor_operation::host::
+            ReferenceImageToColumn<NDimSpatial, InLayout, InDataType, OutDataType>();
+
+        auto ref_invoker = ref_image_to_column.MakeInvoker();
+
+        auto ref_argument = ref_image_to_column.MakeArgument(in,
+                                                             out_host,
+                                                             conv_params.filter_spatial_lengths_,
+                                                             conv_params.conv_filter_strides_,
+                                                             conv_params.conv_filter_dilations_,
+                                                             conv_params.input_left_pads_,
+                                                             conv_params.input_right_pads_);
+
+        if(!ref_image_to_column.IsSupportedArgument(&ref_argument))
+        {
+            std::cerr << "wrong! ref_img2col with the specified compilation parameters does "
+                         "not support this img2col problem"
+                      << std::endl;
+            return false;
+        }
+
+        ref_invoker.Run(ref_argument);
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        return ck::utils::check_err(out_device.mData, out_host.mData);
+    }
+
+    return true;
+}
+
+int RunImageToColumnExample(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_params = DefaultConvParams;
+
+    if(!parse_cmd_args(argc, argv, config, conv_params))
+    {
+        return EXIT_FAILURE;
+    }
+
+    if(conv_params.num_dim_spatial_ != NDimSpatial)
+    {
+        std::cerr << "unsupported # of spatial dimensions" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    return !RunImageToColumn(config, conv_params);
+}
+
+int main(int argc, char* argv[]) { return RunImageToColumnExample(argc, argv); }
diff --git a/example/91_tile_program/CMakeLists.txt b/example/91_tile_program/CMakeLists.txt
new file mode 100644
index 0000000000..7a352b482c
--- /dev/null
+++ b/example/91_tile_program/CMakeLists.txt
@@ -0,0 +1,9 @@
+add_example_executable(example_im2col im2col.cpp)
+add_example_executable(example_gemm gemm.cpp)
+add_example_executable(example_gemm_gemm gemm_gemm.cpp)
+add_example_executable(example_reduce reduce.cpp)
+add_example_executable(example_softmax softmax.cpp)
+add_example_executable(example_gemm_softmax_gemm gemm_softmax_gemm.cpp)
+add_example_executable(example_batched_gemm_softmax_gemm batched_gemm_softmax_gemm.cpp)
+add_example_executable(example_fmha_fwd fmha_fwd.cpp)
+add_example_executable(example_flash_attention_fwd flash_attention_fwd.cpp)
diff --git a/example/91_tile_program/batched_gemm_softmax_gemm.cpp b/example/91_tile_program/batched_gemm_softmax_gemm.cpp
new file mode 100644
index 0000000000..6e6dc335bd
--- /dev/null
+++ b/example/91_tile_program/batched_gemm_softmax_gemm.cpp
@@ -0,0 +1,166 @@
+#include <cstring>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor/tensor_view.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+#include "reference_batched_gemm.hpp"
+#include "reference_batched_softmax.hpp"
+#include "batched_gemm_softmax_gemm.hpp"
+
+int main(int argc, char* argv[])
+{
+    using QDataType           = ck::half_t;
+    using KDataType           = ck::half_t;
+    using VDataType           = ck::half_t;
+    using SaccDataType        = float;
+    using SMPLComputeDataType = float;
+    using PDataType           = ck::half_t;
+    using OaccDataType        = float;
+    using ODataType           = ck::half_t;
+
+    ck::index_t Batch = 16;
+    ck::index_t M0    = 3328;
+    ck::index_t N0    = 4096;
+    ck::index_t K0    = 128;
+    ck::index_t N1    = 128;
+
+    if(argc == 6)
+    {
+        Batch = std::stoi(argv[1]);
+        M0    = std::stoi(argv[2]);
+        N0    = std::stoi(argv[3]);
+        K0    = std::stoi(argv[4]);
+        N1    = std::stoi(argv[5]);
+    }
+
+    std::array<ck::index_t, 3> q_lengths{Batch, M0, K0};
+    std::array<ck::index_t, 3> q_strides{M0 * K0, K0, 1};
+
+    std::array<ck::index_t, 3> k_lengths{Batch, N0, K0};
+    std::array<ck::index_t, 3> k_strides{N0 * K0, K0, 1};
+
+    std::array<ck::index_t, 3> v_lengths{Batch, N1, N0};
+    std::array<ck::index_t, 3> v_strides{N1 * N0, N0, 1};
+
+    std::array<ck::index_t, 3> s_lengths{Batch, M0, N0};
+    std::array<ck::index_t, 3> s_strides{M0 * N0, N0, 1};
+
+    std::array<ck::index_t, 3> p_lengths{Batch, M0, N0};
+    std::array<ck::index_t, 3> p_strides{M0 * N0, N0, 1};
+
+    std::array<ck::index_t, 3> o_lengths{Batch, M0, N1};
+    std::array<ck::index_t, 3> o_strides{M0 * N1, N1, 1};
+
+    // host verify
+    Tensor<QDataType> q_host(q_lengths, q_strides);
+    Tensor<KDataType> k_host(k_lengths, k_strides);
+    Tensor<VDataType> v_host(v_lengths, v_strides);
+    Tensor<SMPLComputeDataType> s_host_ref(s_lengths, s_strides);
+    Tensor<PDataType> p_host_ref(p_lengths, p_strides);
+    Tensor<ODataType> o_host_ref(o_lengths, o_strides);
+    Tensor<ODataType> o_host_dev(o_lengths, o_strides);
+
+#if 0
+    ck::utils::FillUniformDistributionIntegerValue<QDataType>{-3.f, 3.f}(q_host);
+    ck::utils::FillUniformDistributionIntegerValue<KDataType>{-3.f, 3.f}(k_host);
+    ck::utils::FillUniformDistributionIntegerValue<VDataType>{-3.f, 3.f}(v_host);
+#else
+    ck::utils::FillUniformDistribution<QDataType>{-3.f, 3.f}(q_host);
+    ck::utils::FillUniformDistribution<KDataType>{-3.f, 3.f}(k_host);
+    ck::utils::FillUniformDistribution<VDataType>{-3.f, 3.f}(v_host);
+#endif
+
+    // reference
+    reference_batched_gemm<QDataType, KDataType, SaccDataType, SMPLComputeDataType>(
+        q_host, k_host, s_host_ref);
+    reference_batched_softmax<SMPLComputeDataType, SMPLComputeDataType, PDataType>(s_host_ref,
+                                                                                   p_host_ref);
+    reference_batched_gemm<PDataType, VDataType, OaccDataType, ODataType>(
+        p_host_ref, v_host, o_host_ref);
+
+    DeviceMem q_buf(sizeof(QDataType) * q_host.GetElementSpaceSize());
+    DeviceMem k_buf(sizeof(KDataType) * k_host.GetElementSpaceSize());
+    DeviceMem v_buf(sizeof(VDataType) * v_host.GetElementSpaceSize());
+    DeviceMem o_buf(sizeof(ODataType) * o_host_ref.GetElementSpaceSize());
+
+    q_buf.ToDevice(q_host.mData.data());
+    k_buf.ToDevice(k_host.mData.data());
+    v_buf.ToDevice(v_host.mData.data());
+
+    constexpr ck::index_t kM0PerBlock = 128;
+    constexpr ck::index_t kN0PerBlock = 128;
+    constexpr ck::index_t kK0PerBlock = 32;
+    constexpr ck::index_t kN1PerBlock = 128;
+
+    constexpr ck::index_t kBlockSize = 256;
+    ck::index_t kGridSize            = Batch * (M0 / kM0PerBlock) * (N1 / kN1PerBlock);
+
+    std::cout << "grid size " << kGridSize << std::endl;
+
+    constexpr ck::index_t kWarpPerCu    = 8; // 2 warps per SIMD
+    constexpr ck::index_t kWarpPerBlock = kBlockSize / warpSize;
+    constexpr ck::index_t kBlockPerCu   = kWarpPerCu / kWarpPerBlock;
+
+    float ave_time =
+        launch_kernel<kBlockSize, kBlockPerCu>(StreamConfig{nullptr, true},
+                                               BatchedGemmSoftmaxGemm<QDataType,
+                                                                      KDataType,
+                                                                      VDataType,
+                                                                      SaccDataType,
+                                                                      SMPLComputeDataType,
+                                                                      PDataType,
+                                                                      OaccDataType,
+                                                                      ODataType,
+                                                                      kBlockSize,
+                                                                      kM0PerBlock,
+                                                                      kN0PerBlock,
+                                                                      kK0PerBlock,
+                                                                      kN1PerBlock>{},
+                                               kGridSize,
+                                               kBlockSize,
+                                               0,
+                                               static_cast<QDataType*>(q_buf.GetDeviceBuffer()),
+                                               static_cast<KDataType*>(k_buf.GetDeviceBuffer()),
+                                               static_cast<VDataType*>(v_buf.GetDeviceBuffer()),
+                                               static_cast<ODataType*>(o_buf.GetDeviceBuffer()),
+                                               M0,
+                                               N0,
+                                               K0,
+                                               N1,
+                                               Batch,
+                                               K0,       // StrideQ
+                                               K0,       // StrideK
+                                               N0,       // StrideV
+                                               N1,       // StrideO
+                                               M0 * K0,  // BatchStrideQ
+                                               N0 * K0,  // BatchStrideK
+                                               N1 * N0,  // BatchStrideV
+                                               M0 * N1); // BatchStrideO
+
+    o_buf.FromDevice(o_host_dev.mData.data());
+
+    std::size_t flop =
+        std::size_t(2) * Batch * M0 * N0 * K0 + std::size_t(2) * Batch * M0 * N1 * N0;
+    std::size_t num_btype =
+        sizeof(QDataType) * Batch * M0 * K0 + sizeof(KDataType) * Batch * N0 * K0 +
+        sizeof(VDataType) * Batch * N1 * N0 + sizeof(ODataType) * Batch * M0 * N1;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    return !ck::utils::check_err(o_host_dev, o_host_ref);
+}
diff --git a/example/91_tile_program/batched_gemm_softmax_gemm.hpp b/example/91_tile_program/batched_gemm_softmax_gemm.hpp
new file mode 100644
index 0000000000..f396222fef
--- /dev/null
+++ b/example/91_tile_program/batched_gemm_softmax_gemm.hpp
@@ -0,0 +1,109 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_problem.hpp"
+#include "ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1.hpp"
+#include "ck/tile_program/block_tile/block_reduce.hpp"
+
+#include "gemm_softmax_gemm_impl.hpp"
+
+// S[M0, N0] = Q[M0, K0] * K[N0, K0]
+// P[M0, N0] = Softmax(S[M0, N0])
+// O[M0, N1] = P[M0, N0] * V[N1, N0]
+template <typename QDataType,
+          typename KDataType,
+          typename VDataType,
+          typename SaccDataType,
+          typename SMPLComputeDataType,
+          typename PDataType,
+          typename OaccDataType,
+          typename ODataType,
+          ck::index_t kBlockSize,
+          ck::index_t kM0PerBlock,
+          ck::index_t kN0PerBlock,
+          ck::index_t kK0PerBlock,
+          ck::index_t kN1PerBlock>
+struct BatchedGemmSoftmaxGemm
+{
+    __device__ void operator()(const QDataType* q_ptr,
+                               const KDataType* k_ptr,
+                               const VDataType* v_ptr,
+                               ODataType* o_ptr,
+                               const ck::index_t M0,
+                               const ck::index_t N0,
+                               const ck::index_t K0,
+                               const ck::index_t N1,
+                               const ck::index_t /* Batch */,
+                               const ck::index_t StrideQ,
+                               const ck::index_t StrideK,
+                               const ck::index_t StrideV,
+                               const ck::index_t StrideO,
+                               const ck::index_t BatchStrideQ,
+                               const ck::index_t BatchStrideK,
+                               const ck::index_t BatchStrideV,
+                               const ck::index_t BatchStrideO) const
+    {
+        using namespace ck;
+
+        // divide problem
+        const index_t num_tile_m0 = M0 / kM0PerBlock;
+        const index_t num_tile_n1 = N1 / kN1PerBlock;
+
+        const index_t id_block = get_block_id();
+
+        const auto f = [](index_t dividend, index_t divisor) {
+            index_t quotient = dividend / divisor;
+            index_t modulus  = dividend - quotient * divisor;
+
+            return ck::make_tuple(quotient, modulus);
+        };
+
+        const auto [itmp, id_tile_n]          = f(id_block, num_tile_n1);
+        const auto [id_tile_batch, id_tile_m] = f(itmp, num_tile_m0);
+
+        const index_t iBatch = __builtin_amdgcn_readfirstlane(id_tile_batch);
+        const index_t iM0    = __builtin_amdgcn_readfirstlane(id_tile_m * kM0PerBlock);
+        const index_t iN1    = __builtin_amdgcn_readfirstlane(id_tile_n * kN1PerBlock);
+
+        const auto kernel_impl = GemmSoftmaxGemmImpl<QDataType,
+                                                     KDataType,
+                                                     VDataType,
+                                                     SaccDataType,
+                                                     SMPLComputeDataType,
+                                                     PDataType,
+                                                     OaccDataType,
+                                                     ODataType,
+                                                     kBlockSize,
+                                                     kM0PerBlock,
+                                                     kN0PerBlock,
+                                                     kK0PerBlock,
+                                                     kN1PerBlock>{};
+
+        kernel_impl(q_ptr + iBatch * BatchStrideQ,
+                    k_ptr + iBatch * BatchStrideK,
+                    v_ptr + iBatch * BatchStrideV,
+                    o_ptr + iBatch * BatchStrideO,
+                    M0,
+                    N0,
+                    K0,
+                    N1,
+                    StrideQ,
+                    StrideK,
+                    StrideV,
+                    StrideO,
+                    iM0,
+                    iN1);
+    }
+};
diff --git a/example/91_tile_program/flash_attention_fwd.cpp b/example/91_tile_program/flash_attention_fwd.cpp
new file mode 100644
index 0000000000..ecfb4a5b5d
--- /dev/null
+++ b/example/91_tile_program/flash_attention_fwd.cpp
@@ -0,0 +1,209 @@
+#include <cstring>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor/tensor_view.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+#include "reference_batched_gemm.hpp"
+#include "reference_batched_softmax.hpp"
+#include "flash_attention_fwd.hpp"
+
+/*
+ * Toy code of flash attention forward pass
+ * Assume simplest case.
+ * Q [Batch, HeadNum, SeqenceLengthQ, HeadDim]
+ * K [Batch, HeadNum, SeqenceLengthK, HeadDim]
+ * V [Batch, HeadNum, HeadDim, SeqenceLengthK]
+ * O [Batch, HeadNum, SeqenceLengthQ, HeadDim]
+ */
+
+int main(int argc, char* argv[])
+{
+    using QDataType           = ck::half_t;
+    using KDataType           = ck::half_t;
+    using VDataType           = ck::half_t;
+    using SaccDataType        = float;
+    using SMPLComputeDataType = float;
+    using PDataType           = ck::half_t;
+    using OaccDataType        = float;
+    using ODataType           = ck::half_t;
+
+    ck::index_t Batch        = 64;   // Batch Number * Head Number
+    ck::index_t M0           = 4096; // SequenceLengthQ
+    ck::index_t N0           = 4096; // SequencelengthK
+    ck::index_t K0           = 128;  // HeadDim
+    ck::index_t N1           = 128;  // HeadDim
+    ck::index_t verification = 0;
+    ck::index_t init_method  = 1;
+    ck::index_t time_kernel  = 0;
+
+    if(argc == 4)
+    {
+        init_method  = std::stoi(argv[1]);
+        time_kernel  = std::stoi(argv[2]);
+        verification = std::stoi(argv[3]);
+    }
+
+    if(argc == 9)
+    {
+        init_method  = std::stoi(argv[1]);
+        time_kernel  = std::stoi(argv[2]);
+        verification = std::stoi(argv[3]);
+        Batch        = std::stoi(argv[4]);
+        M0           = std::stoi(argv[5]);
+        N0           = std::stoi(argv[6]);
+        K0           = std::stoi(argv[7]);
+        N1           = std::stoi(argv[8]);
+    }
+
+    std::array<ck::index_t, 3> q_lengths{Batch, M0, K0};
+    std::array<ck::index_t, 3> q_strides{M0 * K0, K0, 1};
+
+    std::array<ck::index_t, 3> k_lengths{Batch, N0, K0};
+    std::array<ck::index_t, 3> k_strides{N0 * K0, K0, 1};
+
+    std::array<ck::index_t, 3> v_lengths{Batch, N1, N0};
+    std::array<ck::index_t, 3> v_strides{N1 * N0, N0, 1};
+
+    std::array<ck::index_t, 3> s_lengths{Batch, M0, N0};
+    std::array<ck::index_t, 3> s_strides{M0 * N0, N0, 1};
+
+    std::array<ck::index_t, 3> p_lengths{Batch, M0, N0};
+    std::array<ck::index_t, 3> p_strides{M0 * N0, N0, 1};
+
+    std::array<ck::index_t, 3> o_lengths{Batch, M0, N1};
+    std::array<ck::index_t, 3> o_strides{M0 * N1, N1, 1};
+
+    // host verify
+    Tensor<QDataType> q_host(q_lengths, q_strides);
+    Tensor<KDataType> k_host(k_lengths, k_strides);
+    Tensor<VDataType> v_host(v_lengths, v_strides);
+    Tensor<ODataType> o_host_dev(o_lengths, o_strides);
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        ck::utils::FillUniformDistributionIntegerValue<QDataType>{-3.f, 3.f}(q_host);
+        ck::utils::FillUniformDistributionIntegerValue<KDataType>{-3.f, 3.f}(k_host);
+        ck::utils::FillUniformDistributionIntegerValue<VDataType>{-3.f, 3.f}(v_host);
+        break;
+    case 2:
+        ck::utils::FillUniformDistribution<QDataType>{-3.f, 3.f}(q_host);
+        ck::utils::FillUniformDistribution<KDataType>{-3.f, 3.f}(k_host);
+        ck::utils::FillUniformDistribution<VDataType>{-3.f, 3.f}(v_host);
+        break;
+    default:
+        ck::utils::FillUniformDistributionIntegerValue<QDataType>{-2.f, 2.f}(q_host);
+        ck::utils::FillUniformDistributionIntegerValue<KDataType>{-2.f, 2.f}(k_host);
+        ck::utils::FillUniformDistributionIntegerValue<VDataType>{-2.f, 2.f}(v_host);
+    }
+
+    DeviceMem q_buf(sizeof(QDataType) * q_host.GetElementSpaceSize());
+    DeviceMem k_buf(sizeof(KDataType) * k_host.GetElementSpaceSize());
+    DeviceMem v_buf(sizeof(VDataType) * v_host.GetElementSpaceSize());
+    DeviceMem o_buf(sizeof(ODataType) * o_host_dev.GetElementSpaceSize());
+
+    q_buf.ToDevice(q_host.mData.data());
+    k_buf.ToDevice(k_host.mData.data());
+    v_buf.ToDevice(v_host.mData.data());
+
+    constexpr ck::index_t kM0PerBlock = 128;
+    constexpr ck::index_t kN0PerBlock = 128;
+    constexpr ck::index_t kK0PerBlock = 32;
+    constexpr ck::index_t kN1PerBlock = 128;
+    constexpr ck::index_t kK1PerBlock = 32;
+
+    constexpr ck::index_t kBlockSize = 256;
+    constexpr ck::index_t kHeadDim   = 128;
+
+    ck::index_t kGridSize = Batch * (M0 / kM0PerBlock) * (N1 / kN1PerBlock);
+
+    std::cout << "grid size " << kGridSize << std::endl;
+
+    constexpr ck::index_t kWarpPerCu    = 8; // 2 warps per SIMD
+    constexpr ck::index_t kWarpPerBlock = kBlockSize / warpSize;
+    constexpr ck::index_t kBlockPerCu   = kWarpPerCu / kWarpPerBlock;
+
+    float ave_time = launch_kernel<kBlockSize, kBlockPerCu>(
+        StreamConfig{nullptr, static_cast<bool>(time_kernel)},
+        FlashAttentionFwd<QDataType,
+                          KDataType,
+                          VDataType,
+                          SaccDataType,
+                          SMPLComputeDataType,
+                          PDataType,
+                          OaccDataType,
+                          ODataType,
+                          kBlockSize,
+                          kHeadDim,
+                          kM0PerBlock,
+                          kN0PerBlock,
+                          kK0PerBlock,
+                          kN1PerBlock,
+                          kK1PerBlock>{},
+        kGridSize,
+        kBlockSize,
+        0,
+        static_cast<QDataType*>(q_buf.GetDeviceBuffer()),
+        static_cast<KDataType*>(k_buf.GetDeviceBuffer()),
+        static_cast<VDataType*>(v_buf.GetDeviceBuffer()),
+        static_cast<ODataType*>(o_buf.GetDeviceBuffer()),
+        M0,
+        N0,
+        K0,
+        N1,
+        Batch,
+        K0,       // StrideQ
+        K0,       // StrideK
+        N0,       // StrideV
+        N1,       // StrideO
+        M0 * K0,  // BatchStrideQ
+        N0 * K0,  // BatchStrideK
+        N1 * N0,  // BatchStrideV
+        M0 * N1); // BatchStrideO
+
+    // reference
+    auto pass = true;
+    if(verification)
+    {
+        o_buf.FromDevice(o_host_dev.mData.data());
+
+        Tensor<SMPLComputeDataType> s_host_ref(s_lengths, s_strides);
+        Tensor<PDataType> p_host_ref(p_lengths, p_strides);
+        Tensor<ODataType> o_host_ref(o_lengths, o_strides);
+
+        reference_batched_gemm<QDataType, KDataType, SaccDataType, SMPLComputeDataType>(
+            q_host, k_host, s_host_ref);
+        reference_batched_softmax<SMPLComputeDataType, SMPLComputeDataType, PDataType>(s_host_ref,
+                                                                                       p_host_ref);
+        reference_batched_gemm<PDataType, VDataType, OaccDataType, ODataType>(
+            p_host_ref, v_host, o_host_ref);
+
+        pass &= ck::utils::check_err(o_host_dev, o_host_ref);
+    }
+
+    std::size_t flop =
+        std::size_t(2) * Batch * M0 * N0 * K0 + std::size_t(2) * Batch * M0 * N1 * N0;
+    std::size_t num_btype =
+        sizeof(QDataType) * Batch * M0 * K0 + sizeof(KDataType) * Batch * N0 * K0 +
+        sizeof(VDataType) * Batch * N1 * N0 + sizeof(ODataType) * Batch * M0 * N1;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    return !pass;
+}
diff --git a/example/91_tile_program/flash_attention_fwd.hpp b/example/91_tile_program/flash_attention_fwd.hpp
new file mode 100644
index 0000000000..5e0d36319a
--- /dev/null
+++ b/example/91_tile_program/flash_attention_fwd.hpp
@@ -0,0 +1,113 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_problem.hpp"
+#include "ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1.hpp"
+#include "ck/tile_program/block_tile/block_reduce.hpp"
+
+#include "flash_attention_fwd_impl.hpp"
+
+// S[M0, N0] = Q[M0, K0] * K[N0, K0]
+// P[M0, N0] = Softmax(S[M0, N0])
+// O[M0, N1] = P[M0, N0] * V[N1, N0]
+template <typename QDataType,
+          typename KDataType,
+          typename VDataType,
+          typename SaccDataType,
+          typename SMPLComputeDataType,
+          typename PDataType,
+          typename OaccDataType,
+          typename ODataType,
+          ck::index_t kBlockSize,
+          ck::index_t kHeadDim,
+          ck::index_t kM0PerBlock,
+          ck::index_t kN0PerBlock,
+          ck::index_t kK0PerBlock,
+          ck::index_t kN1PerBlock,
+          ck::index_t kK1PerBlock>
+struct FlashAttentionFwd
+{
+    __device__ void operator()(const QDataType* q_ptr,
+                               const KDataType* k_ptr,
+                               const VDataType* v_ptr,
+                               ODataType* o_ptr,
+                               const ck::index_t M0,
+                               const ck::index_t N0,
+                               const ck::index_t K0,
+                               const ck::index_t N1,
+                               const ck::index_t /* Batch */,
+                               const ck::index_t StrideQ,
+                               const ck::index_t StrideK,
+                               const ck::index_t StrideV,
+                               const ck::index_t StrideO,
+                               const ck::index_t BatchStrideQ,
+                               const ck::index_t BatchStrideK,
+                               const ck::index_t BatchStrideV,
+                               const ck::index_t BatchStrideO) const
+    {
+        using namespace ck;
+
+        // divide problem
+        const index_t num_tile_m0 = M0 / kM0PerBlock;
+        const index_t num_tile_n1 = N1 / kN1PerBlock;
+
+        const index_t id_block = get_block_id();
+
+        const auto f = [](index_t dividend, index_t divisor) {
+            index_t quotient = dividend / divisor;
+            index_t modulus  = dividend - quotient * divisor;
+
+            return ck::make_tuple(quotient, modulus);
+        };
+
+        const auto [itmp, id_tile_n]          = f(id_block, num_tile_n1);
+        const auto [id_tile_batch, id_tile_m] = f(itmp, num_tile_m0);
+
+        const index_t iBatch = __builtin_amdgcn_readfirstlane(id_tile_batch);
+        const index_t iM0    = __builtin_amdgcn_readfirstlane(id_tile_m * kM0PerBlock);
+        const index_t iN1    = __builtin_amdgcn_readfirstlane(id_tile_n * kN1PerBlock);
+
+        const auto kernel_impl = FlashAttentionFwdImpl<QDataType,
+                                                       KDataType,
+                                                       VDataType,
+                                                       SaccDataType,
+                                                       SMPLComputeDataType,
+                                                       PDataType,
+                                                       OaccDataType,
+                                                       ODataType,
+                                                       kBlockSize,
+                                                       kHeadDim,
+                                                       kM0PerBlock,
+                                                       kN0PerBlock,
+                                                       kK0PerBlock,
+                                                       kN1PerBlock,
+                                                       kK1PerBlock>{};
+
+        kernel_impl(q_ptr + iBatch * BatchStrideQ,
+                    k_ptr + iBatch * BatchStrideK,
+                    v_ptr + iBatch * BatchStrideV,
+                    o_ptr + iBatch * BatchStrideO,
+                    M0,
+                    N0,
+                    K0,
+                    N1,
+                    StrideQ,
+                    StrideK,
+                    StrideV,
+                    StrideO,
+                    iM0,
+                    iN1);
+    }
+};
diff --git a/example/91_tile_program/flash_attention_fwd_impl.hpp b/example/91_tile_program/flash_attention_fwd_impl.hpp
new file mode 100644
index 0000000000..40e4519744
--- /dev/null
+++ b/example/91_tile_program/flash_attention_fwd_impl.hpp
@@ -0,0 +1,368 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/tile/slice_tile.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2_askiplds.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_problem.hpp"
+#include "ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1.hpp"
+#include "ck/tile_program/block_tile/block_reduce.hpp"
+
+// S[M0, N0] = Q[M0, K0] * K[N0, K0]
+// P[M0, N0] = Softmax(S[M0, N0])
+// O[M0, N1] = P[M0, N0] * V[N1, N0]
+template <typename QDataType,
+          typename KDataType,
+          typename VDataType,
+          typename SaccDataType,
+          typename SMPLComputeDataType,
+          typename PDataType,
+          typename OaccDataType,
+          typename ODataType,
+          ck::index_t kBlockSize,
+          ck::index_t kHeadDim,
+          ck::index_t kM0PerBlock,
+          ck::index_t kN0PerBlock,
+          ck::index_t kK0PerBlock,
+          ck::index_t kN1PerBlock,
+          ck::index_t kK1PerBlock>
+struct FlashAttentionFwdImpl
+{
+    // block gemm0 pipeline
+    using BlockGemm0Problem = ck::tile_program::block::BlockGemmPipelineProblem<
+        QDataType,
+        KDataType,
+        SaccDataType,
+        kBlockSize,
+        ck::tile_program::TileGemmShape<kM0PerBlock, kN0PerBlock, kK0PerBlock>>;
+
+    using BlockGemm0Policy =
+        ck::tile_program::block::BlockGemmPipelineAGmemBGmemCRegV2SkipALdsPersistentQRegCachePolicy<
+            kHeadDim>;
+
+    using BlockGemm0Pipeline =
+        ck::tile_program::block::BlockGemmPipelineAGmemBGmemCRegV2<BlockGemm0Problem,
+                                                                   BlockGemm0Policy>;
+
+    // block gemm1
+    using BlockGemm1 = ck::tile_program::block::BlockGemmARegBSmemCRegV1<
+        ck::tile_program::block::BlockGemmARegBSmemCRegProblem<
+            PDataType,
+            VDataType,
+            OaccDataType,
+            kBlockSize,
+            ck::tile_program::TileGemmShape<kM0PerBlock, kN1PerBlock, kK1PerBlock>>,
+        ck::tile_program::block::BlockGemmARegBSmemCRegV1DefaultPolicy>;
+
+    // 3d, with padding
+    __device__ static constexpr auto MakeVLdsBlockDescriptor()
+    {
+        using namespace ck;
+
+        // using BDataType = B1DataType;
+
+        constexpr index_t kNPerBlock = kN1PerBlock;
+        constexpr index_t kKPerBlock = kK1PerBlock;
+        constexpr index_t kPad       = 1;
+        // 2% faster than use kK1 = 8
+        constexpr index_t kK1 = 4;
+
+        constexpr auto b_lds_block_desc_0 = make_naive_tensor_descriptor(
+            make_tuple(Number<kKPerBlock / kK1>{}, Number<kNPerBlock>{}, Number<kK1>{}),
+            make_tuple(Number<(kNPerBlock + kPad) * kK1>{}, Number<kK1>{}, Number<1>{}),
+            Number<kK1>{},
+            Number<1>{});
+
+        constexpr auto b_lds_block_desc = transform_tensor_descriptor(
+            b_lds_block_desc_0,
+            make_tuple(make_pass_through_transform(kNPerBlock),
+                       make_merge_transform(make_tuple(Number<kKPerBlock / kK1>{}, Number<kK1>{}))),
+            make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return b_lds_block_desc;
+    }
+
+    __device__ static constexpr auto MakeVDramTileDistribution()
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        using BDataType = VDataType;
+
+        constexpr index_t kNPerBlock = kN1PerBlock;
+        constexpr index_t kKPerBlock = kK1PerBlock;
+
+        constexpr index_t K1 = 16 / sizeof(BDataType);
+        constexpr index_t K0 = kKPerBlock / K1;
+        constexpr index_t N2 = get_warp_size() / K0;
+        constexpr index_t N1 = kBlockSize / get_warp_size();
+        constexpr index_t N0 = kNPerBlock / (N2 * N1);
+
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<1>,
+                                           Tuple<Sequence<N0, N1, N2>, Sequence<K0, K1>>,
+                                           Tuple<Sequence<1>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1>, Sequence<2, 0>>,
+                                           Sequence<1, 2>,
+                                           Sequence<0, 1>>{});
+    }
+
+    __device__ static constexpr ck::index_t GetStaticLdsSize()
+    {
+        using namespace ck;
+
+        return math::max(BlockGemm0Pipeline::GetStaticLdsSize(),
+                         static_cast<index_t>(MakeVLdsBlockDescriptor().GetElementSpaceSize() *
+                                              sizeof(VDataType)));
+    }
+
+    __device__ void operator()(const QDataType* q_ptr,
+                               const KDataType* k_ptr,
+                               const VDataType* v_ptr,
+                               ODataType* o_ptr,
+                               const ck::index_t M0,
+                               const ck::index_t N0,
+                               const ck::index_t K0,
+                               const ck::index_t N1,
+                               const ck::index_t StrideQ,
+                               const ck::index_t StrideK,
+                               const ck::index_t StrideV,
+                               const ck::index_t StrideO,
+                               const ck::index_t iM0,
+                               const ck::index_t iN1) const
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+        using namespace ck::tile_program::block;
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        // allocate LDS
+        __shared__ char smem_ptr[GetStaticLdsSize()];
+
+        // Q/K/V DRAM and DRAM window
+        const auto q_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            q_ptr, make_tuple(M0, K0), make_tuple(StrideQ, 1), Number<32>{}, Number<1>{});
+
+        const auto k_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            k_ptr, make_tuple(N0, K0), make_tuple(StrideK, 1), Number<32>{}, Number<1>{});
+
+        const auto v_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            v_ptr, make_tuple(N1, N0), make_tuple(StrideV, 1), Number<32>{}, Number<1>{});
+
+        auto q_dram_window = make_tile_window(
+            q_dram, make_tuple(Number<kM0PerBlock>{}, Number<kK0PerBlock>{}), {iM0, 0});
+
+        auto k_dram_window = make_tile_window(
+            k_dram, make_tuple(Number<kN0PerBlock>{}, Number<kK0PerBlock>{}), {0, 0});
+
+        auto v_dram_window =
+            make_tile_window(v_dram,
+                             make_tuple(Number<kN1PerBlock>{}, Number<kK1PerBlock>{}),
+                             {iN1, 0},
+                             MakeVDramTileDistribution());
+
+        // Q in Register
+        auto q_reg_tensor = make_static_distributed_tensor<QDataType>(
+            BlockGemm0Policy::template MakeARegBlockDescriptor<BlockGemm0Problem>());
+
+        // V LDS and LDS window
+        // V LDS occupies the same LDS allocation Q/K LDS
+        auto v_lds = make_tensor_view<AddressSpaceEnum::Lds>(reinterpret_cast<VDataType*>(smem_ptr),
+                                                             MakeVLdsBlockDescriptor());
+
+        auto v_lds_window = make_tile_window(
+            v_lds, make_tuple(Number<kN1PerBlock>{}, Number<kK1PerBlock>{}), {0, 0});
+
+        // Block GEMM0 pipeline and Block GEMM1
+        constexpr auto gemm0_pipeline = BlockGemm0Pipeline{};
+        constexpr auto gemm1          = BlockGemm1{};
+
+        // reduction function for softmax
+        const auto f_max = [](auto e0, auto e1) { return max(e0, e1); };
+        const auto f_sum = [](auto e0, auto e1) { return e0 + e1; };
+
+        // infer Sacc, S, P, M, L, Oacc type
+        using SaccBlockTileType =
+            decltype(gemm0_pipeline(q_dram_window, k_dram_window, q_reg_tensor, nullptr));
+
+        using SBlockTileType = decltype(tile_elementwise_in(
+            type_convert<SMPLComputeDataType, SaccDataType>, SaccBlockTileType{}));
+
+        using PBlockTileType = decltype(tile_elementwise_in(type_convert<PDataType, SaccDataType>,
+                                                            SaccBlockTileType{}));
+
+        using MLBlockTileType = decltype(block_tile_reduce<SMPLComputeDataType>(
+            SBlockTileType{}, Sequence<1>{}, f_max, SMPLComputeDataType{0}));
+
+        using OaccBlockTileType = decltype(gemm1(
+            get_slice_tile(
+                PBlockTileType{}, Sequence<0, 0>{}, Sequence<kM0PerBlock, kK1PerBlock>{}),
+            v_dram_window));
+
+        // init Sacc, Oacc, M, L
+        auto s_acc = SaccBlockTileType{};
+        auto o_acc = OaccBlockTileType{};
+        auto m     = MLBlockTileType{};
+        auto l     = MLBlockTileType{};
+
+        tile_elementwise_inout([](auto& e) { e = 0; }, o_acc);
+        tile_elementwise_inout([](auto& e) { e = NumericLimits<SMPLComputeDataType>::Lowest(); },
+                               m);
+        tile_elementwise_inout([](auto& e) { e = 0; }, l);
+
+        // loop over Column of S (J loop)
+        index_t iN0 = 0;
+
+        // Cold Q_Reg_Cache
+        s_acc = gemm0_pipeline(q_dram_window, k_dram_window, q_reg_tensor, smem_ptr);
+        do
+        {
+            // Hot Q_Reg_Cache
+            if(iN0 > 0)
+            {
+                s_acc = gemm0_pipeline(k_dram_window, q_reg_tensor, smem_ptr);
+            }
+            // S{j}
+            const auto s =
+                tile_elementwise_in(type_convert<SMPLComputeDataType, SaccDataType>, s_acc);
+
+            // prefetch load v tile
+            const auto v_prefetch = load_tile(v_dram_window);
+
+            // m_local = rowmax(S{j})
+            auto m_local = block_tile_reduce<SMPLComputeDataType>(
+                s, Sequence<1>{}, f_max, NumericLimits<SMPLComputeDataType>::Lowest());
+
+            block_tile_reduce_sync(m_local, f_max);
+
+            // m{j-1}
+            const auto m_old = m;
+
+            // m{j}
+            tile_elementwise_inout(
+                [](auto& e0, auto e1, auto e2) { e0 = max(e1, e2); }, m, m_old, m_local);
+
+            // Pcompute{j}
+            auto p_compute =
+                make_static_distributed_tensor<SMPLComputeDataType>(s.GetTileDistribution());
+
+            constexpr auto p_spans = decltype(p_compute)::GetDistributedSpans();
+
+            sweep_tile_span(p_spans[I0], [&](auto idx0) {
+                constexpr auto i_idx = make_tuple(idx0);
+
+                sweep_tile_span(p_spans[I1], [&](auto idx1) {
+                    constexpr auto i_j_idx = make_tuple(idx0, idx1);
+
+                    p_compute(i_j_idx) = math::exp(s[i_j_idx] - m[i_idx]);
+                });
+            });
+
+            // rowsum(Pcompute{j})
+            auto rowsum_p = block_tile_reduce<SMPLComputeDataType>(
+                p_compute, Sequence<1>{}, f_sum, SMPLComputeDataType{0});
+
+            block_tile_reduce_sync(rowsum_p, f_sum);
+
+            // l{j}, Oacc{j}
+            sweep_tile_span(p_spans[I0], [&](auto idx0) {
+                constexpr auto i_idx = make_tuple(idx0);
+
+                const auto tmp = math::exp(m_old[i_idx] - m[i_idx]);
+
+                l(i_idx) = tmp * l[i_idx] + rowsum_p[i_idx];
+
+                sweep_tile_span(p_spans[I1], [&](auto idx1) {
+                    constexpr auto i_j_idx = make_tuple(idx0, idx1);
+
+                    o_acc(i_j_idx) *= tmp;
+                });
+            });
+
+            block_sync_lds();
+            store_tile(v_lds_window, v_prefetch);
+            move_tile_window(v_dram_window, {0, kK1PerBlock});
+
+            // type cast Pcompute{j} into P{j}
+            const auto p =
+                tile_elementwise_in(type_convert<PDataType, SMPLComputeDataType>, p_compute);
+
+            // Oacc{j}
+            constexpr index_t k1_loops = kN0PerBlock / kK1PerBlock;
+
+            if constexpr(k1_loops > 1)
+            {
+                static_for<0, k1_loops - 1, 1>{}([&](auto i_k1) {
+                    const auto v = load_tile(v_dram_window); // load next v
+                    block_sync_lds();
+                    gemm1(o_acc,
+                          get_slice_tile(p,
+                                         Sequence<0, i_k1 * kK1PerBlock>{},
+                                         Sequence<kM0PerBlock, (i_k1 + 1) * kK1PerBlock>{}),
+                          v_lds_window);
+                    block_sync_lds();
+                    store_tile(v_lds_window, v);
+                    move_tile_window(v_dram_window, {0, kK1PerBlock});
+                });
+            }
+            // tail
+            {
+                block_sync_lds();
+                gemm1(o_acc,
+                      get_slice_tile(p,
+                                     Sequence<0, (k1_loops - 1) * kK1PerBlock>{},
+                                     Sequence<kM0PerBlock, kN0PerBlock>{}),
+                      v_lds_window);
+                block_sync_lds();
+            }
+            // move tile windows
+            move_tile_window(k_dram_window, {kN0PerBlock, 0});
+            iN0 += kN0PerBlock;
+        } while(iN0 < N0);
+
+        // Oacc
+        constexpr auto o_spans = decltype(o_acc)::GetDistributedSpans();
+
+        sweep_tile_span(o_spans[I0], [&](auto idx0) {
+            constexpr auto i_idx = make_tuple(idx0);
+
+            const auto tmp = 1 / l[i_idx];
+
+            sweep_tile_span(o_spans[I1], [&](auto idx1) {
+                constexpr auto i_j_idx = make_tuple(idx0, idx1);
+
+                o_acc(i_j_idx) *= tmp;
+            });
+        });
+
+        // type cast Oacc into O
+        const auto o = tile_elementwise_in(type_convert<ODataType, OaccDataType>, o_acc);
+
+        // O DRAM and O DRAM window
+        auto o_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            o_ptr, make_tuple(M0, N1), make_tuple(StrideO, 1), Number<32>{}, Number<1>{});
+
+        auto o_dram_window =
+            make_tile_window(o_dram,
+                             make_tuple(Number<kM0PerBlock>{}, Number<kN1PerBlock>{}),
+                             {iM0, iN1},
+                             o.GetTileDistribution());
+
+        // store O
+        store_tile(o_dram_window, o);
+    }
+};
diff --git a/example/91_tile_program/fmha_fwd.cpp b/example/91_tile_program/fmha_fwd.cpp
new file mode 100644
index 0000000000..89e5541735
--- /dev/null
+++ b/example/91_tile_program/fmha_fwd.cpp
@@ -0,0 +1,176 @@
+#include <cstring>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor/tensor_view.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+#include "ck/tile_program/block_tile_pipeline/block_fmha_pipeline_qkvs.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_fmha_pipeline_qkvs_default_policy.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_fmha_pipeline_problem.hpp"
+#include "ck/tile_program/tile/tile_fmha_shape.hpp"
+
+#include "reference_batched_gemm.hpp"
+#include "reference_batched_softmax.hpp"
+#include "fmha_fwd_kernel.hpp"
+#include "fmha_fwd_epilogue.hpp"
+
+using QDataType           = ck::half_t;
+using KDataType           = ck::half_t;
+using VDataType           = ck::half_t;
+using SaccDataType        = float;      // data type for first gemm accumulation
+using SMPLComputeDataType = float;      // data type for reduction, softmax
+using PDataType           = ck::half_t; // data type for A matrix of second gemm
+using OaccDataType        = float;      // data type for second gemm accumulation
+using ODataType           = ck::half_t;
+
+using FmhaShape =
+    ck::tile_program::TileFmhaShape<128 /*M0*/, 128 /*N0*/, 32 /*K0*/, 128 /*N1*/, 32 /*K1*/>;
+
+using FmhaPipelineProblem = ck::tile_program::block::BlockFmhaPipelineProblem<QDataType,
+                                                                              KDataType,
+                                                                              VDataType,
+                                                                              SaccDataType,
+                                                                              SMPLComputeDataType,
+                                                                              PDataType,
+                                                                              OaccDataType,
+                                                                              ODataType,
+                                                                              256, // BlockSize
+                                                                              FmhaShape>;
+using FmhaPipeline        = ck::tile_program::block::BlockFmhaPipelineQKVS<FmhaPipelineProblem>;
+
+using FmhaEpilogue = FmhaFwdEpilogue<FmhaFwdEpilogueProblem<OaccDataType, ODataType>>;
+using FmhaKernel   = FmhaFwdKernel<FmhaPipeline, FmhaEpilogue>;
+
+int main(int argc, char* argv[])
+{
+    ck::index_t Batch = 16;   // batch * nheads
+    ck::index_t M0    = 3328; // seqlen_q
+    ck::index_t N0    = 4096; // seqlen_k
+    ck::index_t K0    = 128;  // hdim_q
+    ck::index_t N1    = 128;  // hdim_v
+
+    if(argc == 6)
+    {
+        Batch = std::stoi(argv[1]);
+        M0    = std::stoi(argv[2]);
+        N0    = std::stoi(argv[3]);
+        K0    = std::stoi(argv[4]);
+        N1    = std::stoi(argv[5]);
+    }
+
+    std::array<ck::index_t, 3> q_lengths{Batch, M0, K0};
+    std::array<ck::index_t, 3> q_strides{M0 * K0, K0, 1};
+
+    std::array<ck::index_t, 3> k_lengths{Batch, N0, K0};
+    std::array<ck::index_t, 3> k_strides{N0 * K0, K0, 1};
+
+    std::array<ck::index_t, 3> v_lengths{Batch, N1, N0};
+    std::array<ck::index_t, 3> v_strides{N1 * N0, N0, 1};
+
+    std::array<ck::index_t, 3> s_lengths{Batch, M0, N0};
+    std::array<ck::index_t, 3> s_strides{M0 * N0, N0, 1};
+
+    std::array<ck::index_t, 3> p_lengths{Batch, M0, N0};
+    std::array<ck::index_t, 3> p_strides{M0 * N0, N0, 1};
+
+    std::array<ck::index_t, 3> o_lengths{Batch, M0, N1};
+    std::array<ck::index_t, 3> o_strides{M0 * N1, N1, 1};
+
+    // host verify
+    Tensor<QDataType> q_host(q_lengths, q_strides);
+    Tensor<KDataType> k_host(k_lengths, k_strides);
+    Tensor<VDataType> v_host(v_lengths, v_strides);
+    Tensor<SMPLComputeDataType> s_host_ref(s_lengths, s_strides);
+    Tensor<PDataType> p_host_ref(p_lengths, p_strides);
+    Tensor<ODataType> o_host_ref(o_lengths, o_strides);
+    Tensor<ODataType> o_host_dev(o_lengths, o_strides);
+
+#if 0
+    ck::utils::FillUniformDistributionIntegerValue<QDataType>{-3.f, 3.f}(q_host);
+    ck::utils::FillUniformDistributionIntegerValue<KDataType>{-3.f, 3.f}(k_host);
+    ck::utils::FillUniformDistributionIntegerValue<VDataType>{-3.f, 3.f}(v_host);
+#else
+    ck::utils::FillUniformDistribution<QDataType>{-3.f, 3.f}(q_host);
+    ck::utils::FillUniformDistribution<KDataType>{-3.f, 3.f}(k_host);
+    ck::utils::FillUniformDistribution<VDataType>{-3.f, 3.f}(v_host);
+#endif
+
+    // reference
+    reference_batched_gemm<QDataType, KDataType, SaccDataType, SMPLComputeDataType>(
+        q_host, k_host, s_host_ref);
+    reference_batched_softmax<SMPLComputeDataType, SMPLComputeDataType, PDataType>(s_host_ref,
+                                                                                   p_host_ref);
+    reference_batched_gemm<PDataType, VDataType, OaccDataType, ODataType>(
+        p_host_ref, v_host, o_host_ref);
+
+    DeviceMem q_buf(sizeof(QDataType) * q_host.GetElementSpaceSize());
+    DeviceMem k_buf(sizeof(KDataType) * k_host.GetElementSpaceSize());
+    DeviceMem v_buf(sizeof(VDataType) * v_host.GetElementSpaceSize());
+    DeviceMem o_buf(sizeof(ODataType) * o_host_ref.GetElementSpaceSize());
+
+    q_buf.ToDevice(q_host.mData.data());
+    k_buf.ToDevice(k_host.mData.data());
+    v_buf.ToDevice(v_host.mData.data());
+
+    dim3 kGridSize            = FmhaKernel::GridSize(Batch, M0, N1);
+    constexpr dim3 kBlockSize = FmhaKernel::BlockSize();
+
+    std::cout << "batch:" << Batch << ", seqlen_q:" << M0 << ", seqlen_k:" << N0
+              << ", hdim_q:" << K0 << ", hdim_v:" << N1 << ", grid_size " << kGridSize.x
+              << std::endl;
+
+    constexpr ck::index_t kWarpPerCu    = 8; // 2 warps per SIMD
+    constexpr ck::index_t kWarpPerBlock = kBlockSize.x / warpSize;
+    constexpr ck::index_t kBlockPerCu   = kWarpPerCu / kWarpPerBlock;
+
+    auto kargs = FmhaKernel::MakeKargs(q_buf.GetDeviceBuffer(),
+                                       k_buf.GetDeviceBuffer(),
+                                       v_buf.GetDeviceBuffer(),
+                                       o_buf.GetDeviceBuffer(),
+                                       M0,       // seqlen_q
+                                       N0,       // seqlen_k
+                                       K0,       // hdim_q
+                                       N1,       // hdim_v
+                                       K0,       // stride_q
+                                       K0,       // stride_k
+                                       N0,       // stride_v
+                                       N1,       // stride_o
+                                       M0 * K0,  // batch_stride_q
+                                       N0 * K0,  // batch_stride_k
+                                       N1 * N0,  // batch_stride_v
+                                       M0 * N1); // batch_stride_o
+
+    float ave_time = launch_kernel<kBlockSize.x, kBlockPerCu>(StreamConfig{nullptr, true},
+                                                              FmhaKernel{},
+                                                              kGridSize,
+                                                              kBlockSize,
+                                                              0,
+                                                              kargs); // BatchStrideO
+
+    o_buf.FromDevice(o_host_dev.mData.data());
+
+    std::size_t flop =
+        std::size_t(2) * Batch * M0 * N0 * K0 + std::size_t(2) * Batch * M0 * N1 * N0;
+
+    std::size_t num_btype =
+        sizeof(QDataType) * Batch * M0 * K0 + sizeof(KDataType) * Batch * N0 * K0 +
+        sizeof(VDataType) * Batch * N1 * N0 + sizeof(ODataType) * Batch * M0 * N1;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    return !ck::utils::check_err(o_host_dev, o_host_ref);
+}
diff --git a/example/91_tile_program/fmha_fwd_epilogue.hpp b/example/91_tile_program/fmha_fwd_epilogue.hpp
new file mode 100644
index 0000000000..94f4b17e5f
--- /dev/null
+++ b/example/91_tile_program/fmha_fwd_epilogue.hpp
@@ -0,0 +1,35 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tile_program/tile/store_tile.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+
+template <typename OaccDataType_, typename ODataType_>
+struct FmhaFwdEpilogueProblem
+{
+    using OaccDataType = ck::remove_cvref_t<OaccDataType_>;
+    using ODataType    = ck::remove_cvref_t<ODataType_>;
+};
+
+template <typename Problem_, typename Policy_ = void>
+struct FmhaFwdEpilogue
+{
+    using Problem      = ck::remove_cvref_t<Problem_>;
+    using OaccDataType = ck::remove_cvref_t<typename Problem::OaccDataType>;
+    using ODataType    = ck::remove_cvref_t<typename Problem::ODataType>;
+
+    __host__ __device__ static constexpr ck::index_t GetSmemSize() { return 0; }
+
+    template <typename ODramWindowTmp, typename OAccTile>
+    __device__ auto operator()(ODramWindowTmp& o_dram_window_tmp, const OAccTile& o_acc_tile)
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        const auto o = tile_elementwise_in(type_convert<ODataType, OaccDataType>, o_acc_tile);
+        store_tile(o_dram_window_tmp, o);
+    }
+};
diff --git a/example/91_tile_program/fmha_fwd_kernel.hpp b/example/91_tile_program/fmha_fwd_kernel.hpp
new file mode 100644
index 0000000000..70ac09757e
--- /dev/null
+++ b/example/91_tile_program/fmha_fwd_kernel.hpp
@@ -0,0 +1,191 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor/tensor_view.hpp"
+#include "ck/tile_program/tile/tile_window.hpp"
+
+// S[seqlen_q, seqlen_k] = Q[seqlen_q, hdim_q] * K[seqlen_k, hdim_q]
+// P[seqlen_q, seqlen_k] = Softmax(S[seqlen_q, seqlen_k])
+// O[seqlen_q, hdim_v] = P[seqlen_q, seqlen_k] * V[hdim_v, seqlen_k]
+
+template <typename FmhaPipeline_, typename EpiloguePipeline_>
+struct FmhaFwdKernel
+{
+    using FmhaPipeline                      = ck::remove_cvref_t<FmhaPipeline_>;
+    using EpiloguePipeline                  = ck::remove_cvref_t<EpiloguePipeline_>;
+    static constexpr ck::index_t kBlockSize = FmhaPipeline::kBlockSize;
+
+    using QDataType = ck::remove_cvref_t<typename FmhaPipeline::QDataType>;
+    using KDataType = ck::remove_cvref_t<typename FmhaPipeline::KDataType>;
+    using VDataType = ck::remove_cvref_t<typename FmhaPipeline::VDataType>;
+    using ODataType = ck::remove_cvref_t<typename FmhaPipeline::ODataType>;
+
+    struct Kargs
+    {
+        const void* q_ptr;
+        const void* k_ptr;
+        const void* v_ptr;
+        void* o_ptr;
+        ck::index_t seqlen_q;
+        ck::index_t seqlen_k;
+        ck::index_t hdim_q;
+        ck::index_t hdim_v;
+        ck::index_t stride_q;
+        ck::index_t stride_k;
+        ck::index_t stride_v;
+        ck::index_t stride_o;
+        ck::index_t batch_stride_q;
+        ck::index_t batch_stride_k;
+        ck::index_t batch_stride_v;
+        ck::index_t batch_stride_o;
+    };
+
+    __host__ static constexpr Kargs MakeKargs(const void* q_ptr,
+                                              const void* k_ptr,
+                                              const void* v_ptr,
+                                              void* o_ptr,
+                                              ck::index_t seqlen_q,
+                                              ck::index_t seqlen_k,
+                                              ck::index_t hdim_q,
+                                              ck::index_t hdim_v,
+                                              ck::index_t stride_q,
+                                              ck::index_t stride_k,
+                                              ck::index_t stride_v,
+                                              ck::index_t stride_o,
+                                              ck::index_t batch_stride_q,
+                                              ck::index_t batch_stride_k,
+                                              ck::index_t batch_stride_v,
+                                              ck::index_t batch_stride_o)
+    {
+        return Kargs{q_ptr,
+                     k_ptr,
+                     v_ptr,
+                     o_ptr,
+                     seqlen_q,
+                     seqlen_k,
+                     hdim_q,
+                     hdim_v,
+                     stride_q,
+                     stride_k,
+                     stride_v,
+                     stride_o,
+                     batch_stride_q,
+                     batch_stride_k,
+                     batch_stride_v,
+                     batch_stride_o};
+    }
+
+    __host__ static constexpr auto
+    GridSize(ck::index_t batch_size_, ck::index_t seqlen_q_, ck::index_t hdim_v_)
+    {
+        return dim3(batch_size_ * (seqlen_q_ / FmhaPipeline::kM0) * (hdim_v_ / FmhaPipeline::kN1));
+    }
+
+    __host__ static constexpr auto BlockSize() { return dim3(kBlockSize); }
+
+    __host__ __device__ static constexpr ck::index_t GetSmemSize()
+    {
+        return ck::math::max(FmhaPipeline::GetSmemSize(), EpiloguePipeline::GetSmemSize());
+    }
+
+    __device__ void operator()(Kargs kargs) const
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+        using namespace ck::tile_program::block;
+
+        // allocate LDS
+        __shared__ char smem_ptr[GetSmemSize()];
+
+        // divide problem
+        const index_t num_tile_m0 = kargs.seqlen_q / FmhaPipeline::kM0;
+        const index_t num_tile_n1 = kargs.hdim_v / FmhaPipeline::kN1;
+
+        const index_t id_block = ck::get_block_id();
+
+        const auto f = [](index_t dividend, index_t divisor) {
+            index_t quotient = dividend / divisor;
+            index_t modulus  = dividend - quotient * divisor;
+
+            return ck::make_tuple(quotient, modulus);
+        };
+
+        const auto [itmp, id_tile_n]          = f(id_block, num_tile_n1);
+        const auto [id_tile_batch, id_tile_m] = f(itmp, num_tile_m0);
+
+        const index_t i_batch = __builtin_amdgcn_readfirstlane(id_tile_batch);
+        const index_t i_m0    = __builtin_amdgcn_readfirstlane(id_tile_m * FmhaPipeline::kM0);
+        const index_t i_n1    = __builtin_amdgcn_readfirstlane(id_tile_n * FmhaPipeline::kN1);
+
+        // for simplicity, batch stride we just modify the pointer
+        const QDataType* q_ptr =
+            reinterpret_cast<const QDataType*>(kargs.q_ptr) + i_batch * kargs.batch_stride_q;
+        const KDataType* k_ptr =
+            reinterpret_cast<const KDataType*>(kargs.k_ptr) + i_batch * kargs.batch_stride_k;
+        const VDataType* v_ptr =
+            reinterpret_cast<const VDataType*>(kargs.v_ptr) + i_batch * kargs.batch_stride_v;
+        ODataType* o_ptr =
+            reinterpret_cast<ODataType*>(kargs.o_ptr) + i_batch * kargs.batch_stride_o;
+
+        // Q/K/V DRAM and DRAM window
+        // FIXME: assume layout Q[seqlen_q, hdim_q], K[seqlen_k, hdim_q], V[hdim_v, seqlen_k],
+        const auto q_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            q_ptr,
+            make_tuple(kargs.seqlen_q, kargs.hdim_q),
+            make_tuple(kargs.stride_q, 1),
+            Number<32>{},
+            Number<1>{});
+
+        const auto k_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            k_ptr,
+            make_tuple(kargs.seqlen_k, kargs.hdim_q),
+            make_tuple(kargs.stride_k, 1),
+            Number<32>{},
+            Number<1>{});
+
+        const auto v_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            v_ptr,
+            make_tuple(kargs.hdim_v, kargs.seqlen_k),
+            make_tuple(kargs.stride_v, 1),
+            Number<32>{},
+            Number<1>{});
+
+        auto q_dram_window =
+            make_tile_window(q_dram,
+                             make_tuple(Number<FmhaPipeline::kM0>{}, Number<FmhaPipeline::kK0>{}),
+                             {i_m0, 0});
+
+        auto k_dram_window = make_tile_window(
+            k_dram, make_tuple(Number<FmhaPipeline::kN0>{}, Number<FmhaPipeline::kK0>{}), {0, 0});
+
+        auto v_dram_window =
+            make_tile_window(v_dram,
+                             make_tuple(Number<FmhaPipeline::kN1>{}, Number<FmhaPipeline::kK1>{}),
+                             {i_n1, 0});
+
+        auto o_acc_tile = FmhaPipeline{}(q_dram_window,
+                                         k_dram_window,
+                                         v_dram_window,
+                                         kargs.seqlen_k / FmhaPipeline::kN0,
+                                         kargs.hdim_q / FmhaPipeline::kK0,
+                                         smem_ptr);
+
+        // O DRAM and O DRAM window
+        auto o_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            o_ptr,
+            make_tuple(kargs.seqlen_q, kargs.hdim_v),
+            make_tuple(kargs.stride_o, 1),
+            Number<32>{},
+            Number<1>{});
+
+        auto o_dram_window =
+            make_tile_window(o_dram,
+                             make_tuple(Number<FmhaPipeline::kM0>{}, Number<FmhaPipeline::kN1>{}),
+                             {i_m0, i_n1});
+
+        EpiloguePipeline{}(o_dram_window, o_acc_tile);
+    }
+};
diff --git a/example/91_tile_program/gemm.cpp b/example/91_tile_program/gemm.cpp
new file mode 100644
index 0000000000..bb57464f40
--- /dev/null
+++ b/example/91_tile_program/gemm.cpp
@@ -0,0 +1,195 @@
+#include <cstring>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor/tensor_view.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+#include "reference_gemm.hpp"
+#include "gemm.hpp"
+
+/*
+ * Toy code of GEMM
+ * Assume simplest case.
+ * A [M, K]
+ * B [N, K]
+ * C [M, N]
+ */
+
+// elementwise lambda
+struct AElementFunction
+{
+    template <typename X>
+    __host__ __device__ auto operator()(const X& x) const
+    {
+        return x;
+    }
+};
+
+struct BElementFunction
+{
+    template <typename X>
+    __host__ __device__ auto operator()(const X& x) const
+    {
+        return x;
+    }
+};
+
+struct CElementFunction
+{
+    template <typename X>
+    __host__ __device__ auto operator()(const X& x) const
+    {
+        return x;
+    }
+};
+
+int main(int argc, char* argv[])
+{
+    using ADataType   = ck::half_t;
+    using BDataType   = ck::half_t;
+    using AccDataType = float;
+    using CDataType   = ck::half_t;
+
+    using ALayout = ck::tensor_layout::gemm::RowMajor;
+    using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+    using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+    ck::index_t verification = 0;
+    ck::index_t M            = 3328;
+    ck::index_t N            = 4096;
+    ck::index_t K            = 4096;
+
+    if(argc == 2)
+    {
+        verification = std::stoi(argv[1]);
+    }
+    if(argc == 5)
+    {
+        verification = std::stoi(argv[1]);
+        M            = std::stoi(argv[1]);
+        N            = std::stoi(argv[2]);
+        K            = std::stoi(argv[3]);
+    }
+
+    const ck::index_t Lda = std::is_same_v<ALayout, ck::tensor_layout::gemm::RowMajor> ? K : M;
+    const ck::index_t Ldb = std::is_same_v<BLayout, ck::tensor_layout::gemm::ColumnMajor> ? K : N;
+    const ck::index_t Ldc = std::is_same_v<CLayout, ck::tensor_layout::gemm::RowMajor> ? N : M;
+
+    const auto a_lengths = std::array<ck::index_t, 2>{M, K};
+    const auto a_strides = std::is_same_v<ALayout, ck::tensor_layout::gemm::RowMajor>
+                               ? std::array<ck::index_t, 2>{Lda, 1}
+                               : std::array<ck::index_t, 2>{1, Lda};
+
+    const auto b_lengths = std::array<ck::index_t, 2>{N, K};
+    const auto b_strides = std::is_same_v<BLayout, ck::tensor_layout::gemm::ColumnMajor>
+                               ? std::array<ck::index_t, 2>{Ldb, 1}
+                               : std::array<ck::index_t, 2>{1, Ldb};
+
+    const auto c_lengths = std::array<ck::index_t, 2>{M, N};
+    const auto c_strides = std::is_same_v<CLayout, ck::tensor_layout::gemm::RowMajor>
+                               ? std::array<ck::index_t, 2>{Ldc, 1}
+                               : std::array<ck::index_t, 2>{1, Ldc};
+
+    // host verify
+    Tensor<ADataType> a_host(a_lengths, a_strides);
+    Tensor<BDataType> b_host(b_lengths, b_strides);
+    Tensor<CDataType> c_host_dev(c_lengths, c_strides);
+
+    ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_host);
+    ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_host);
+
+    DeviceMem a_buf(sizeof(ADataType) * a_host.GetElementSpaceSize());
+    DeviceMem b_buf(sizeof(BDataType) * b_host.GetElementSpaceSize());
+    DeviceMem c_buf(sizeof(CDataType) * c_host_dev.GetElementSpaceSize());
+
+    a_buf.ToDevice(a_host.mData.data());
+    b_buf.ToDevice(b_host.mData.data());
+
+    // Alignment
+    constexpr ck::index_t kAAlignment = 32;
+    constexpr ck::index_t kBAlignment = 32;
+    constexpr ck::index_t kCAlignment = 32;
+
+    constexpr ck::index_t kBlockSize = 256;
+
+    constexpr ck::index_t kGemmMPerBlock = 256;
+    constexpr ck::index_t kGemmNPerBlock = 128;
+    constexpr ck::index_t kGemmKPerBlock = 32;
+
+    ck::index_t kGridSize = (M / kGemmMPerBlock) * (N / kGemmNPerBlock);
+
+    std::cout << "grid size " << kGridSize << std::endl;
+
+    constexpr ck::index_t kWarpPerCu    = 8; // 2 warps per SIMD
+    constexpr ck::index_t kWarpPerBlock = kBlockSize / warpSize;
+    constexpr ck::index_t kBlockPerCu   = kWarpPerCu / kWarpPerBlock;
+
+    const auto gemm_kernel = Gemm<ADataType,
+                                  BDataType,
+                                  AccDataType,
+                                  CDataType,
+                                  ALayout,
+                                  BLayout,
+                                  CLayout,
+                                  AElementFunction,
+                                  BElementFunction,
+                                  CElementFunction,
+                                  kAAlignment,
+                                  kBAlignment,
+                                  kCAlignment,
+                                  kBlockSize,
+                                  kGemmMPerBlock,
+                                  kGemmNPerBlock,
+                                  kGemmKPerBlock>{};
+
+    float ave_time =
+        launch_kernel<kBlockSize, kBlockPerCu>(StreamConfig{nullptr, true},
+                                               gemm_kernel,
+                                               kGridSize,
+                                               kBlockSize,
+                                               0,
+                                               static_cast<ADataType*>(a_buf.GetDeviceBuffer()),
+                                               static_cast<BDataType*>(b_buf.GetDeviceBuffer()),
+                                               static_cast<CDataType*>(c_buf.GetDeviceBuffer()),
+                                               M,
+                                               N,
+                                               K,
+                                               Lda,
+                                               Ldb,
+                                               Ldc,
+                                               AElementFunction{},
+                                               BElementFunction{},
+                                               CElementFunction{});
+    auto pass = true;
+
+    if(verification)
+    {
+        // reference gemm
+        Tensor<CDataType> c_host_ref(c_lengths, c_strides);
+        reference_gemm<ADataType, ADataType, AccDataType, CDataType>(a_host, b_host, c_host_ref);
+        c_buf.FromDevice(c_host_dev.mData.data());
+        pass &= ck::utils::check_err(c_host_dev, c_host_ref);
+    }
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    return !pass;
+}
diff --git a/example/91_tile_program/gemm.hpp b/example/91_tile_program/gemm.hpp
new file mode 100644
index 0000000000..d161cbf8ed
--- /dev/null
+++ b/example/91_tile_program/gemm.hpp
@@ -0,0 +1,171 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v1.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2.hpp"
+#include "ck/tile_program/grid/grid_gemm_problem.hpp"
+#include "ck/tile_program/grid/grid_gemm_v1.hpp"
+#include "ck/tile_program/grid/grid_gemm_v1_default_policy.hpp"
+
+// C = A * B
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementFunction,
+          typename BElementFunction,
+          typename CElementFunction,
+          ck::index_t kAAlignment,
+          ck::index_t kBAlignment,
+          ck::index_t kCAlignment,
+          ck::index_t kBlockSize_,
+          ck::index_t kMPerBlock_,
+          ck::index_t kNPerBlock_,
+          ck::index_t kKPerBlock_>
+struct Gemm
+{
+    using GridGemmProblem = ck::tile_program::grid::GridGemmProblem<ADataType,
+                                                                    BDataType,
+                                                                    AccDataType,
+                                                                    CDataType,
+                                                                    AElementFunction,
+                                                                    BElementFunction,
+                                                                    CElementFunction>;
+
+    struct GridGemmPolicy
+    {
+        static constexpr ck::index_t kBlockSize = kBlockSize_;
+        static constexpr ck::index_t kMPerBlock = kMPerBlock_;
+        static constexpr ck::index_t kNPerBlock = kNPerBlock_;
+        static constexpr ck::index_t kKPerBlock = kKPerBlock_;
+
+        template <typename Problem>
+        __host__ __device__ static constexpr auto MakeBlock2TileMap(ck::index_t NumTilesM,
+                                                                    ck::index_t NumTilesN)
+        {
+            using namespace ck;
+
+            const auto unmerge = make_merge_transform(make_tuple(NumTilesN, NumTilesM));
+
+            return [unmerge](index_t block_id) {
+                MultiIndex<2> unmerged;
+                unmerge.CalculateLowerIndex(unmerged, make_multi_index(block_id));
+
+                return make_multi_index(unmerged.At<1>(), unmerged.At<0>());
+            };
+        }
+
+        template <typename Problem>
+        __host__ __device__ static constexpr auto GetBlockGemmPipeline()
+        {
+            using namespace ck;
+            using namespace ck::tile_program;
+            using namespace ck::tile_program::block;
+
+            using BlockGemmPipelineProblem_ =
+                BlockGemmPipelineProblem<ADataType,
+                                         BDataType,
+                                         AccDataType,
+                                         kBlockSize,
+                                         TileGemmShape<kMPerBlock, kNPerBlock, kKPerBlock>>;
+
+            return BlockGemmPipelineAGmemBGmemCRegV2<
+                BlockGemmPipelineProblem_,
+                BlockGemmPipelineAGmemBGmemCRegV2DefaultPolicy>{};
+        }
+    };
+
+    using GridGemm = ck::GridGemmV1<GridGemmProblem, GridGemmPolicy>;
+
+    __device__ void operator()(const ADataType* p_a,
+                               const BDataType* p_b,
+                               CDataType* p_c,
+                               const ck::index_t M,
+                               const ck::index_t N,
+                               const ck::index_t K,
+                               const ck::index_t Lda,
+                               const ck::index_t Ldb,
+                               const ck::index_t Ldc,
+                               const AElementFunction& a_element_func,
+                               const BElementFunction& b_element_func,
+                               const CElementFunction& c_element_func) const
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+        using namespace ck::tile_program::block;
+
+        const auto a_dram = [&] {
+            if constexpr(is_same_v<ALayout, ck::tensor_layout::gemm::RowMajor>)
+            {
+                return make_naive_tensor_view<AddressSpaceEnum::Global>(
+                    p_a, make_tuple(M, K), make_tuple(Lda, 1), Number<kAAlignment>{}, Number<1>{});
+            }
+            else
+            {
+                const auto a_k_m_desc = make_naive_tensor_view<AddressSpaceEnum::Global>(
+                    p_a, make_tuple(K, M), make_tuple(Lda, 1), Number<kAAlignment>{}, Number<1>{});
+
+                return transform_tensor_view(
+                    a_k_m_desc,
+                    make_tuple(make_pass_through_transform(M), make_pass_through_transform(K)),
+                    make_tuple(Sequence<1>{}, Sequence<0>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto b_dram = [&] {
+            if constexpr(is_same_v<BLayout, ck::tensor_layout::gemm::ColumnMajor>)
+            {
+                return make_naive_tensor_view<AddressSpaceEnum::Global>(
+                    p_b, make_tuple(N, K), make_tuple(Ldb, 1), Number<kBAlignment>{}, Number<1>{});
+            }
+            else
+            {
+                const auto b_k_n_desc = make_naive_tensor_view<AddressSpaceEnum::Global>(
+                    p_b, make_tuple(K, N), make_tuple(Ldb, 1), Number<kBAlignment>{}, Number<1>{});
+
+                return transform_tensor_view(
+                    b_k_n_desc,
+                    make_tuple(make_pass_through_transform(N), make_pass_through_transform(K)),
+                    make_tuple(Sequence<1>{}, Sequence<0>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto c_dram = [&] {
+            if constexpr(is_same_v<CLayout, ck::tensor_layout::gemm::RowMajor>)
+            {
+                return make_naive_tensor_view<AddressSpaceEnum::Global>(
+                    p_c, make_tuple(M, N), make_tuple(Ldc, 1), Number<kCAlignment>{}, Number<1>{});
+            }
+            else
+            {
+                const auto c_n_m_desc = make_naive_tensor_view<AddressSpaceEnum::Global>(
+                    p_c, make_tuple(N, M), make_tuple(Ldc, 1), Number<kCAlignment>{}, Number<1>{});
+
+                return transform_tensor_view(
+                    c_n_m_desc,
+                    make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                    make_tuple(Sequence<1>{}, Sequence<0>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        GridGemm{}(a_dram, b_dram, c_dram, a_element_func, b_element_func, c_element_func);
+    }
+};
diff --git a/example/91_tile_program/gemm_gemm.cpp b/example/91_tile_program/gemm_gemm.cpp
new file mode 100644
index 0000000000..0e14576ace
--- /dev/null
+++ b/example/91_tile_program/gemm_gemm.cpp
@@ -0,0 +1,143 @@
+#include <cstring>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor/tensor_view.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+#include "reference_gemm.hpp"
+#include "gemm_gemm.hpp"
+
+int main(int argc, char* argv[])
+{
+    using A0DataType   = ck::half_t;
+    using B0DataType   = ck::half_t;
+    using B1DataType   = ck::half_t;
+    using Acc0DataType = float;
+    using C0DataType   = ck::half_t;
+    using Acc1DataType = float;
+    using C1DataType   = ck::half_t;
+
+    ck::index_t M0 = 13312;
+    ck::index_t N0 = 4096;
+    ck::index_t K0 = 128;
+    ck::index_t N1 = 128;
+
+    if(argc == 5)
+    {
+        M0 = std::stoi(argv[1]);
+        N0 = std::stoi(argv[2]);
+        K0 = std::stoi(argv[3]);
+        N1 = std::stoi(argv[4]);
+    }
+
+    std::array<ck::index_t, 2> a0_lengths{M0, K0};
+    std::array<ck::index_t, 2> a0_strides{K0, 1};
+
+    std::array<ck::index_t, 2> b0_lengths{N0, K0};
+    std::array<ck::index_t, 2> b0_strides{K0, 1};
+
+    std::array<ck::index_t, 2> c0_lengths{M0, N0};
+    std::array<ck::index_t, 2> c0_strides{N0, 1};
+
+    std::array<ck::index_t, 2> b1_lengths{N1, N0};
+    std::array<ck::index_t, 2> b1_strides{N0, 1};
+
+    std::array<ck::index_t, 2> c1_lengths{M0, N1};
+    std::array<ck::index_t, 2> c1_strides{N1, 1};
+
+    // host verify
+    Tensor<A0DataType> a0_host(a0_lengths, a0_strides);
+    Tensor<B0DataType> b0_host(b0_lengths, b0_strides);
+    Tensor<B1DataType> b1_host(b1_lengths, b1_strides);
+    Tensor<C0DataType> c0_host_ref(c0_lengths, c0_strides);
+    Tensor<C1DataType> c1_host_ref(c1_lengths, c1_strides);
+    Tensor<C1DataType> c1_host_dev(c1_lengths, c1_strides);
+
+    ck::utils::FillUniformDistributionIntegerValue<A0DataType>{-3.f, 3.f}(a0_host);
+    ck::utils::FillUniformDistributionIntegerValue<B0DataType>{-3.f, 3.f}(b0_host);
+    ck::utils::FillUniformDistributionIntegerValue<B1DataType>{-3.f, 3.f}(b1_host);
+
+    // reference gemm
+    reference_gemm<A0DataType, B0DataType, Acc0DataType, C0DataType>(a0_host, b0_host, c0_host_ref);
+    reference_gemm<C0DataType, B1DataType, Acc1DataType, C1DataType>(
+        c0_host_ref, b1_host, c1_host_ref);
+
+    DeviceMem a0_buf(sizeof(A0DataType) * a0_host.GetElementSpaceSize());
+    DeviceMem b0_buf(sizeof(B0DataType) * b0_host.GetElementSpaceSize());
+    DeviceMem b1_buf(sizeof(B1DataType) * b1_host.GetElementSpaceSize());
+    DeviceMem c1_buf(sizeof(C1DataType) * c1_host_ref.GetElementSpaceSize());
+
+    a0_buf.ToDevice(a0_host.mData.data());
+    b0_buf.ToDevice(b0_host.mData.data());
+    b1_buf.ToDevice(b1_host.mData.data());
+
+    constexpr ck::index_t kM0PerBlock = 128;
+    constexpr ck::index_t kN0PerBlock = 128;
+    constexpr ck::index_t kK0PerBlock = 32;
+    constexpr ck::index_t kN1PerBlock = 128;
+    constexpr ck::index_t kK1PerBlock = 32;
+
+    constexpr ck::index_t kBlockSize = 256;
+    ck::index_t kGridSize            = (M0 / kM0PerBlock) * (N1 / kN1PerBlock);
+
+    std::cout << "grid size " << kGridSize << std::endl;
+
+    constexpr ck::index_t kWarpPerCu    = 8; // 2 warps per SIMD
+    constexpr ck::index_t kWarpPerBlock = kBlockSize / warpSize;
+    constexpr ck::index_t kBlockPerCu   = kWarpPerCu / kWarpPerBlock;
+
+    float ave_time =
+        launch_kernel<kBlockSize, kBlockPerCu>(StreamConfig{nullptr, true},
+                                               GemmGemm<A0DataType,
+                                                        B0DataType,
+                                                        B1DataType,
+                                                        Acc0DataType,
+                                                        C0DataType,
+                                                        Acc1DataType,
+                                                        C1DataType,
+                                                        kBlockSize,
+                                                        kM0PerBlock,
+                                                        kN0PerBlock,
+                                                        kK0PerBlock,
+                                                        kN1PerBlock,
+                                                        kK1PerBlock>{},
+                                               kGridSize,
+                                               kBlockSize,
+                                               0,
+                                               static_cast<A0DataType*>(a0_buf.GetDeviceBuffer()),
+                                               static_cast<B0DataType*>(b0_buf.GetDeviceBuffer()),
+                                               static_cast<B1DataType*>(b1_buf.GetDeviceBuffer()),
+                                               static_cast<C1DataType*>(c1_buf.GetDeviceBuffer()),
+                                               M0,
+                                               N0,
+                                               K0,
+                                               N1,
+                                               K0,  // Lda0
+                                               K0,  // Ldb0
+                                               N0,  // Ldb1
+                                               N1); // Ldc1
+
+    c1_buf.FromDevice(c1_host_dev.mData.data());
+
+    std::size_t flop      = std::size_t(2) * M0 * N0 * K0 + std::size_t(2) * M0 * N1 * N0;
+    std::size_t num_btype = sizeof(A0DataType) * M0 * K0 + sizeof(B0DataType) * N0 * K0 +
+                            sizeof(B1DataType) * N1 * N0 + sizeof(C1DataType) * M0 * N1;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    return !ck::utils::check_err(c1_host_dev, c1_host_ref);
+}
diff --git a/example/91_tile_program/gemm_gemm.hpp b/example/91_tile_program/gemm_gemm.hpp
new file mode 100644
index 0000000000..5f189a2f3b
--- /dev/null
+++ b/example/91_tile_program/gemm_gemm.hpp
@@ -0,0 +1,326 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/tile/slice_tile.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_problem.hpp"
+#include "ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1.hpp"
+
+// C0 = A0 * B0
+// C1 = C0 * B1
+template <typename A0DataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename Acc0DataType,
+          typename C0DataType,
+          typename Acc1DataType,
+          typename C1DataType,
+          ck::index_t kBlockSize,
+          ck::index_t kM0PerBlock,
+          ck::index_t kN0PerBlock,
+          ck::index_t kK0PerBlock,
+          ck::index_t kN1PerBlock,
+          ck::index_t kK1PerBlock>
+struct GemmGemm
+{
+    static constexpr auto I0         = ck::Number<0>{};
+    static constexpr auto BlockSize  = ck::Number<kBlockSize>{};
+    static constexpr auto M0PerBlock = ck::Number<kM0PerBlock>{};
+    static constexpr auto N0PerBlock = ck::Number<kN0PerBlock>{};
+    static constexpr auto K0PerBlock = ck::Number<kK0PerBlock>{};
+    static constexpr auto N1PerBlock = ck::Number<kN1PerBlock>{};
+    static constexpr auto K1PerBlock = ck::Number<kK1PerBlock>{};
+
+    // block gemm0 pipeline
+    using BlockGemm0Pipeline = ck::tile_program::block::BlockGemmPipelineAGmemBGmemCRegV2<
+        ck::tile_program::block::BlockGemmPipelineProblem<
+            A0DataType,
+            B0DataType,
+            Acc0DataType,
+            kBlockSize,
+            ck::tile_program::TileGemmShape<kM0PerBlock, kN0PerBlock, kK0PerBlock>>,
+        ck::tile_program::block::BlockGemmPipelineAGmemBGmemCRegV2DefaultPolicy>;
+
+    // block gemm1
+    using BlockGemm1 = ck::tile_program::block::BlockGemmARegBSmemCRegV1<
+        ck::tile_program::block::BlockGemmARegBSmemCRegV1Problem<
+            C0DataType,
+            B1DataType,
+            Acc1DataType,
+            kBlockSize,
+            ck::tile_program::TileGemmShape<kM0PerBlock, kN1PerBlock, kK1PerBlock>>,
+        ck::tile_program::block::BlockGemmARegBSmemCRegV1DefaultPolicy>;
+
+#if 0
+    // 2d
+    __device__ static constexpr auto MakeB1LdsBlockDescriptor()
+    {
+        using namespace ck;
+
+        constexpr index_t kNPerBlock = kN1PerBlock;
+        constexpr index_t kKPerBlock = kK1PerBlock;
+
+        constexpr auto b_lds_block_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(kNPerBlock, kKPerBlock), Number<32>{});
+
+        return b_lds_block_desc;
+    }
+#elif 1
+    // 3d, with padding
+    __device__ static constexpr auto MakeB1LdsBlockDescriptor()
+    {
+        using namespace ck;
+
+        // using BDataType = B1DataType;
+
+        constexpr index_t kNPerBlock = kN1PerBlock;
+        constexpr index_t kKPerBlock = kK1PerBlock;
+        constexpr index_t kPad       = 1;
+        constexpr index_t kK1        = 8;
+
+        constexpr auto b_lds_block_desc_0 = make_naive_tensor_descriptor(
+            make_tuple(Number<kKPerBlock / kK1>{}, Number<kNPerBlock>{}, Number<kK1>{}),
+            make_tuple(Number<(kNPerBlock + kPad) * kK1>{}, Number<kK1>{}, Number<1>{}),
+            Number<kK1>{},
+            Number<1>{});
+
+        constexpr auto b_lds_block_desc = transform_tensor_descriptor(
+            b_lds_block_desc_0,
+            make_tuple(make_pass_through_transform(kNPerBlock),
+                       make_merge_transform(make_tuple(Number<kKPerBlock / kK1>{}, Number<kK1>{}))),
+            make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return b_lds_block_desc;
+    }
+#else
+    // fake XOR
+    __host__ __device__ static constexpr auto MakeB1LdsBlockDescriptor()
+    {
+        using namespace ck;
+
+        using BDataType = B1DataType;
+
+        constexpr index_t kNPerBlock = kN1PerBlock;
+        constexpr index_t kKPerBlock = kK1PerBlock;
+
+        constexpr auto b_lds_block_desc_d1_d2_d3 = make_naive_tensor_descriptor_packed(
+            make_tuple(kNPerBlock / 2, 2, kKPerBlock), Number<kKPerBlock>{});
+
+        constexpr index_t kK1 = 16 / sizeof(BDataType);
+
+        constexpr auto b_lds_block_desc_d4_d5_d6 = transform_tensor_descriptor(
+            b_lds_block_desc_d1_d2_d3,
+            make_tuple(make_xor_transform(make_tuple(kNPerBlock / 2, kKPerBlock), kK1),
+                       make_pass_through_transform(2)),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+        constexpr auto b_lds_block_desc_n_k = transform_tensor_descriptor(
+            b_lds_block_desc_d4_d5_d6,
+            make_tuple(make_merge_transform(make_tuple(kNPerBlock / 2, 2)),
+                       make_pass_through_transform(kKPerBlock)),
+            make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return b_lds_block_desc_n_k;
+    }
+#endif
+
+    __device__ static constexpr auto MakeB1DramTileDistribution()
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        using BDataType = B1DataType;
+
+        constexpr index_t kNPerBlock = kN1PerBlock;
+        constexpr index_t kKPerBlock = kK1PerBlock;
+
+        constexpr index_t K1 = 16 / sizeof(BDataType);
+        constexpr index_t K0 = kKPerBlock / K1;
+        constexpr index_t N2 = get_warp_size() / K0;
+        constexpr index_t N1 = kBlockSize / get_warp_size();
+        constexpr index_t N0 = kNPerBlock / (N2 * N1);
+
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<1>,
+                                           Tuple<Sequence<N0, N1, N2>, Sequence<K0, K1>>,
+                                           Tuple<Sequence<1>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1>, Sequence<2, 0>>,
+                                           Sequence<1, 2>,
+                                           Sequence<0, 1>>{});
+    }
+
+    __device__ static constexpr ck::index_t GetStaticLdsSize()
+    {
+        using namespace ck;
+
+        return math::max(BlockGemm0Pipeline::GetStaticLdsSize(),
+                         static_cast<index_t>(MakeB1LdsBlockDescriptor().GetElementSpaceSize() *
+                                              sizeof(B1DataType)));
+    }
+
+    __device__ void operator()(const A0DataType* p_a0,
+                               const B0DataType* p_b0,
+                               const B1DataType* p_b1,
+                               C1DataType* p_c1,
+                               const ck::index_t M0,
+                               const ck::index_t N0,
+                               const ck::index_t K0,
+                               const ck::index_t N1,
+                               const ck::index_t Lda0,
+                               const ck::index_t Ldb0,
+                               const ck::index_t Ldb1,
+                               const ck::index_t Ldc1)
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+        using namespace ck::tile_program::block;
+
+        // FIXME: assume layout A0[M0, K0], B0[N0, K0], B1[N1, N0], C1[M0, N1]
+        const auto a0_dram_grid = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_a0, make_tuple(M0, K0), make_tuple(Lda0, 1), Number<32>{}, Number<1>{});
+
+        const auto b0_dram_grid = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_b0, make_tuple(N0, K0), make_tuple(Ldb0, 1), Number<32>{}, Number<1>{});
+
+        const auto b1_dram_grid = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_b1, make_tuple(N1, N0), make_tuple(Ldb1, 1), Number<32>{}, Number<1>{});
+
+        // divide problem
+        const auto id_block = get_block_id();
+
+        const auto num_tile_m0 = M0 / kM0PerBlock;
+        const auto num_tile_n1 = N1 / kN1PerBlock;
+
+        const auto block2tile = make_cluster_descriptor(make_tuple(num_tile_m0, num_tile_n1));
+
+        const auto id_tile = block2tile.CalculateBottomIndex(make_tuple(id_block));
+
+        const auto iM0 = __builtin_amdgcn_readfirstlane(id_tile.At<0>() * kM0PerBlock);
+        const auto iN1 = __builtin_amdgcn_readfirstlane(id_tile.At<1>() * kN1PerBlock);
+
+        __shared__ char p_smem_char[GetStaticLdsSize()];
+
+        // A0 DRAM block window
+        auto a0_dram_block_window =
+            make_tile_window(a0_dram_grid, make_tuple(M0PerBlock, K0PerBlock), {iM0, 0});
+
+        // B0 DRAM block window
+        auto b0_dram_block_window =
+            make_tile_window(b0_dram_grid, make_tuple(N0PerBlock, K0PerBlock), {0, 0});
+
+        // Block GEMM0 pipeline
+        constexpr auto block_gemm0_pipeline = BlockGemm0Pipeline{};
+
+        // B1 DRAM window
+        auto b1_dram_block_window = make_tile_window(b1_dram_grid,
+                                                     make_tuple(N1PerBlock, K1PerBlock),
+                                                     {iN1, 0},
+                                                     MakeB1DramTileDistribution());
+
+        // B1 LDS tensor view: occupies the same LDS allocation as block_gemm0_pipeline
+        auto b1_lds_block = make_tensor_view<AddressSpaceEnum::Lds>(
+            reinterpret_cast<B1DataType*>(p_smem_char), MakeB1LdsBlockDescriptor());
+
+        auto b1_lds_block_window =
+            make_tile_window(b1_lds_block, make_tuple(N1PerBlock, K1PerBlock), {0, 0});
+
+        // Bock GEMM1
+        constexpr auto block_gemm1 = BlockGemm1{};
+
+        // Acc1 tile
+        auto acc1_block_tile = decltype(block_gemm1(
+            get_slice_tile(
+                tile_elementwise_in(
+                    type_convert<C0DataType, Acc0DataType>,
+                    block_gemm0_pipeline(a0_dram_block_window, b0_dram_block_window, 0, nullptr)),
+                Sequence<0, 0>{},
+                Sequence<kM0PerBlock, kK1PerBlock>{}),
+            b1_dram_block_window)){};
+
+        // init Acc1
+        tile_elementwise_inout([](auto& acc1) { acc1 = 0; }, acc1_block_tile);
+
+        index_t iN0 = 0;
+
+        do
+        {
+            // Block GEMM0 pipeline: acc0 = a0 * b0
+            const auto acc0_block_tile = block_gemm0_pipeline(
+                a0_dram_block_window, b0_dram_block_window, K0 / kK0PerBlock, p_smem_char);
+
+            // type cast acc0 into c0
+            const auto c0_block_tile =
+                tile_elementwise_in(type_convert<C0DataType, Acc0DataType>, acc0_block_tile);
+
+            // prefetch load b1
+            const auto b1_block_tile = load_tile(b1_dram_block_window);
+            move_tile_window(b1_dram_block_window, {0, kK1PerBlock});
+
+            block_sync_lds();
+
+            store_tile(b1_lds_block_window, b1_block_tile);
+
+            constexpr index_t k1_loops = kN0PerBlock / kK1PerBlock;
+
+            if constexpr(k1_loops > 1)
+            {
+                static_for<0, k1_loops - 1, 1>{}([&](auto i) {
+                    // acc1 += c0 * b1
+                    const auto b1_block_tile_1 = load_tile(b1_dram_block_window);
+                    block_sync_lds();
+                    block_gemm1(acc1_block_tile,
+                                get_slice_tile(c0_block_tile,
+                                               Sequence<0, i * kK1PerBlock>{},
+                                               Sequence<kM0PerBlock, (i + 1) * kK1PerBlock>{}),
+                                b1_lds_block_window);
+                    block_sync_lds();
+                    move_tile_window(b1_dram_block_window, {0, kK1PerBlock});
+                    store_tile(b1_lds_block_window, b1_block_tile_1);
+                });
+            }
+            // tail
+            {
+                block_sync_lds();
+                block_gemm1(acc1_block_tile,
+                            get_slice_tile(c0_block_tile,
+                                           Sequence<0, (k1_loops - 1) * kK1PerBlock>{},
+                                           Sequence<kM0PerBlock, kN0PerBlock>{}),
+                            b1_lds_block_window);
+            }
+
+            move_tile_window(b0_dram_block_window, {kN0PerBlock, 0});
+            block_sync_lds();
+            iN0 += kN0PerBlock;
+
+        } while(iN0 < N0);
+
+        // type cast acc1 into c1
+        const auto c1_block_tile =
+            tile_elementwise_in(type_convert<C1DataType, Acc1DataType>, acc1_block_tile);
+
+        // store c1
+        auto c1_dram_grid = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_c1, make_tuple(M0, N1), make_tuple(Ldc1, 1), Number<32>{}, Number<1>{});
+
+        auto c1_dram_window = make_tile_window(c1_dram_grid,
+                                               make_tuple(M0PerBlock, N1PerBlock),
+                                               {iM0, iN1},
+                                               c1_block_tile.GetTileDistribution());
+
+        store_tile(c1_dram_window, c1_block_tile);
+    }
+};
diff --git a/example/91_tile_program/gemm_softmax_gemm.cpp b/example/91_tile_program/gemm_softmax_gemm.cpp
new file mode 100644
index 0000000000..9d2c962ade
--- /dev/null
+++ b/example/91_tile_program/gemm_softmax_gemm.cpp
@@ -0,0 +1,155 @@
+#include <cstring>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor/tensor_view.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+#include "reference_gemm.hpp"
+#include "reference_softmax.hpp"
+#include "gemm_softmax_gemm.hpp"
+
+int main(int argc, char* argv[])
+{
+    using QDataType           = ck::half_t;
+    using KDataType           = ck::half_t;
+    using VDataType           = ck::half_t;
+    using SaccDataType        = float;
+    using SMPLComputeDataType = float;
+    using PDataType           = ck::half_t;
+    using OaccDataType        = float;
+    using ODataType           = ck::half_t;
+
+    ck::index_t M0 = 13312;
+    ck::index_t N0 = 4096;
+    ck::index_t K0 = 128;
+    ck::index_t N1 = 128;
+
+    if(argc == 5)
+    {
+        M0 = std::stoi(argv[1]);
+        N0 = std::stoi(argv[2]);
+        K0 = std::stoi(argv[3]);
+        N1 = std::stoi(argv[4]);
+    }
+
+    std::array<ck::index_t, 2> q_lengths{M0, K0};
+    std::array<ck::index_t, 2> q_strides{K0, 1};
+
+    std::array<ck::index_t, 2> k_lengths{N0, K0};
+    std::array<ck::index_t, 2> k_strides{K0, 1};
+
+    std::array<ck::index_t, 2> v_lengths{N1, N0};
+    std::array<ck::index_t, 2> v_strides{N0, 1};
+
+    std::array<ck::index_t, 2> s_lengths{M0, N0};
+    std::array<ck::index_t, 2> s_strides{N0, 1};
+
+    std::array<ck::index_t, 2> p_lengths{M0, N0};
+    std::array<ck::index_t, 2> p_strides{N0, 1};
+
+    std::array<ck::index_t, 2> o_lengths{M0, N1};
+    std::array<ck::index_t, 2> o_strides{N1, 1};
+
+    // host verify
+    Tensor<QDataType> q_host(q_lengths, q_strides);
+    Tensor<KDataType> k_host(k_lengths, k_strides);
+    Tensor<VDataType> v_host(v_lengths, v_strides);
+    Tensor<SMPLComputeDataType> s_host_ref(s_lengths, s_strides);
+    Tensor<PDataType> p_host_ref(p_lengths, p_strides);
+    Tensor<ODataType> o_host_ref(o_lengths, o_strides);
+    Tensor<ODataType> o_host_dev(o_lengths, o_strides);
+
+#if 0
+    ck::utils::FillUniformDistributionIntegerValue<QDataType>{-3.f, 3.f}(q_host);
+    ck::utils::FillUniformDistributionIntegerValue<KDataType>{-3.f, 3.f}(k_host);
+    ck::utils::FillUniformDistributionIntegerValue<VDataType>{-3.f, 3.f}(v_host);
+#else
+    ck::utils::FillUniformDistribution<QDataType>{-3.f, 3.f}(q_host);
+    ck::utils::FillUniformDistribution<KDataType>{-3.f, 3.f}(k_host);
+    ck::utils::FillUniformDistribution<VDataType>{-3.f, 3.f}(v_host);
+#endif
+
+    // reference
+    reference_gemm<QDataType, KDataType, SaccDataType, SMPLComputeDataType>(
+        q_host, k_host, s_host_ref);
+    reference_softmax<SMPLComputeDataType, SMPLComputeDataType, PDataType>(s_host_ref, p_host_ref);
+    reference_gemm<PDataType, VDataType, OaccDataType, ODataType>(p_host_ref, v_host, o_host_ref);
+
+    DeviceMem q_buf(sizeof(QDataType) * q_host.GetElementSpaceSize());
+    DeviceMem k_buf(sizeof(KDataType) * k_host.GetElementSpaceSize());
+    DeviceMem v_buf(sizeof(VDataType) * v_host.GetElementSpaceSize());
+    DeviceMem o_buf(sizeof(ODataType) * o_host_ref.GetElementSpaceSize());
+
+    q_buf.ToDevice(q_host.mData.data());
+    k_buf.ToDevice(k_host.mData.data());
+    v_buf.ToDevice(v_host.mData.data());
+
+    constexpr ck::index_t kM0PerBlock = 128;
+    constexpr ck::index_t kN0PerBlock = 128;
+    constexpr ck::index_t kK0PerBlock = 32;
+    constexpr ck::index_t kN1PerBlock = 128;
+
+    constexpr ck::index_t kBlockSize = 256;
+    ck::index_t kGridSize            = (M0 / kM0PerBlock) * (N1 / kN1PerBlock);
+
+    std::cout << "grid size " << kGridSize << std::endl;
+
+    constexpr ck::index_t kWarpPerCu    = 8; // 2 warps per SIMD
+    constexpr ck::index_t kWarpPerBlock = kBlockSize / warpSize;
+    constexpr ck::index_t kBlockPerCu   = kWarpPerCu / kWarpPerBlock;
+
+    float ave_time =
+        launch_kernel<kBlockSize, kBlockPerCu>(StreamConfig{nullptr, true},
+                                               GemmSoftmaxGemm<QDataType,
+                                                               KDataType,
+                                                               VDataType,
+                                                               SaccDataType,
+                                                               SMPLComputeDataType,
+                                                               PDataType,
+                                                               OaccDataType,
+                                                               ODataType,
+                                                               kBlockSize,
+                                                               kM0PerBlock,
+                                                               kN0PerBlock,
+                                                               kK0PerBlock,
+                                                               kN1PerBlock>{},
+                                               kGridSize,
+                                               kBlockSize,
+                                               0,
+                                               static_cast<QDataType*>(q_buf.GetDeviceBuffer()),
+                                               static_cast<KDataType*>(k_buf.GetDeviceBuffer()),
+                                               static_cast<VDataType*>(v_buf.GetDeviceBuffer()),
+                                               static_cast<ODataType*>(o_buf.GetDeviceBuffer()),
+                                               M0,
+                                               N0,
+                                               K0,
+                                               N1,
+                                               K0,  // StrideQ
+                                               K0,  // StrideK
+                                               N0,  // StrideV
+                                               N1); // StrideO
+
+    o_buf.FromDevice(o_host_dev.mData.data());
+
+    std::size_t flop      = std::size_t(2) * M0 * N0 * K0 + std::size_t(2) * M0 * N1 * N0;
+    std::size_t num_btype = sizeof(QDataType) * M0 * K0 + sizeof(KDataType) * N0 * K0 +
+                            sizeof(VDataType) * N1 * N0 + sizeof(ODataType) * M0 * N1;
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
+              << std::endl;
+
+    return !ck::utils::check_err(o_host_dev, o_host_ref);
+}
diff --git a/example/91_tile_program/gemm_softmax_gemm.hpp b/example/91_tile_program/gemm_softmax_gemm.hpp
new file mode 100644
index 0000000000..657e7bff53
--- /dev/null
+++ b/example/91_tile_program/gemm_softmax_gemm.hpp
@@ -0,0 +1,95 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_problem.hpp"
+#include "ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1.hpp"
+#include "ck/tile_program/block_tile/block_reduce.hpp"
+
+#include "gemm_softmax_gemm_impl.hpp"
+
+// S[M0, N0] = Q[M0, K0] * K[N0, K0]
+// P[M0, N0] = Softmax(S[M0, N0])
+// O[M0, N1] = P[M0, N0] * V[N1, N0]
+template <typename QDataType,
+          typename KDataType,
+          typename VDataType,
+          typename SaccDataType,
+          typename SMPLComputeDataType,
+          typename PDataType,
+          typename OaccDataType,
+          typename ODataType,
+          ck::index_t kBlockSize,
+          ck::index_t kM0PerBlock,
+          ck::index_t kN0PerBlock,
+          ck::index_t kK0PerBlock,
+          ck::index_t kN1PerBlock>
+struct GemmSoftmaxGemm
+{
+    __device__ void operator()(const QDataType* q_ptr,
+                               const KDataType* k_ptr,
+                               const VDataType* v_ptr,
+                               ODataType* o_ptr,
+                               const ck::index_t M0,
+                               const ck::index_t N0,
+                               const ck::index_t K0,
+                               const ck::index_t N1,
+                               const ck::index_t StrideQ,
+                               const ck::index_t StrideK,
+                               const ck::index_t StrideV,
+                               const ck::index_t StrideO) const
+    {
+        using namespace ck;
+
+        // divide problem
+        const auto num_tile_n1 = N1 / kN1PerBlock;
+
+        const auto id_block = get_block_id();
+
+        const auto id_tile_m = id_block / num_tile_n1;
+        const auto id_tile_n = id_block - id_tile_m * num_tile_n1;
+
+        const auto iM0 = __builtin_amdgcn_readfirstlane(id_tile_m * kM0PerBlock);
+        const auto iN1 = __builtin_amdgcn_readfirstlane(id_tile_n * kN1PerBlock);
+
+        const auto kernel_impl = GemmSoftmaxGemmImpl<QDataType,
+                                                     KDataType,
+                                                     VDataType,
+                                                     SaccDataType,
+                                                     SMPLComputeDataType,
+                                                     PDataType,
+                                                     OaccDataType,
+                                                     ODataType,
+                                                     kBlockSize,
+                                                     kM0PerBlock,
+                                                     kN0PerBlock,
+                                                     kK0PerBlock,
+                                                     kN1PerBlock>{};
+
+        kernel_impl(q_ptr,
+                    k_ptr,
+                    v_ptr,
+                    o_ptr,
+                    M0,
+                    N0,
+                    K0,
+                    N1,
+                    StrideQ,
+                    StrideK,
+                    StrideV,
+                    StrideO,
+                    iM0,
+                    iN1);
+    }
+};
diff --git a/example/91_tile_program/gemm_softmax_gemm_impl.hpp b/example/91_tile_program/gemm_softmax_gemm_impl.hpp
new file mode 100644
index 0000000000..021a68eeb6
--- /dev/null
+++ b/example/91_tile_program/gemm_softmax_gemm_impl.hpp
@@ -0,0 +1,263 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_problem.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2_default_policy.hpp"
+#include "ck/tile_program/block_tile/block_gemm_areg_bgmem_creg_problem.hpp"
+#include "ck/tile_program/block_tile/block_gemm_areg_bgmem_creg_v1.hpp"
+#include "ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1_default_policy.hpp"
+#include "ck/tile_program/block_tile/block_reduce.hpp"
+
+// S[M0, N0] = Q[M0, K0] * K[N0, K0]
+// P[M0, N0] = Softmax(S[M0, N0])
+// O[M0, N1] = P[M0, N0] * V[N1, N0]
+template <typename QDataType,
+          typename KDataType,
+          typename VDataType,
+          typename SaccDataType,
+          typename SMPLComputeDataType,
+          typename PDataType,
+          typename OaccDataType,
+          typename ODataType,
+          ck::index_t kBlockSize,
+          ck::index_t kM0PerBlock,
+          ck::index_t kN0PerBlock,
+          ck::index_t kK0PerBlock,
+          ck::index_t kN1PerBlock>
+struct GemmSoftmaxGemmImpl
+{
+    // block gemm0 pipeline
+    using BlockGemm0Pipeline = ck::tile_program::block::BlockGemmPipelineAGmemBGmemCRegV2<
+        ck::tile_program::block::BlockGemmPipelineProblem<
+            QDataType,
+            KDataType,
+            SaccDataType,
+            kBlockSize,
+            ck::tile_program::TileGemmShape<kM0PerBlock, kN0PerBlock, kK0PerBlock>>,
+        ck::tile_program::block::BlockGemmPipelineAGmemBGmemCRegV2DefaultPolicy>;
+
+    // block gemm1
+    using BlockGemm1 = ck::tile_program::block::BlockGemmARegBGmemCRegV1<
+        ck::tile_program::block::BlockGemmARegBGmemCRegProblem<
+            PDataType,
+            VDataType,
+            OaccDataType,
+            kBlockSize,
+            ck::tile_program::TileGemmShape<kM0PerBlock, kN1PerBlock, kN0PerBlock>>,
+        ck::tile_program::block::BlockGemmARegBGmemCRegV1DefaultPolicy>;
+
+    __device__ static constexpr ck::index_t GetStaticLdsSize()
+    {
+        return ck::math::max(BlockGemm0Pipeline::GetStaticLdsSize(),
+                             BlockGemm1::GetStaticLdsSize());
+    }
+
+    __device__ void operator()(const QDataType* q_ptr,
+                               const KDataType* k_ptr,
+                               const VDataType* v_ptr,
+                               ODataType* o_ptr,
+                               const ck::index_t M0,
+                               const ck::index_t N0,
+                               const ck::index_t K0,
+                               const ck::index_t N1,
+                               const ck::index_t StrideQ,
+                               const ck::index_t StrideK,
+                               const ck::index_t StrideV,
+                               const ck::index_t StrideO,
+                               const ck::index_t iM0,
+                               const ck::index_t iN1) const
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+        using namespace ck::tile_program::block;
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        // allocate LDS
+        __shared__ char smem_ptr[GetStaticLdsSize()];
+
+        // Q/K/V DRAM
+        // FIXME: assume layout Q[M0, K0], K[N0, K0], V[N1, N0], O[M0, N1]
+        const auto q_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            q_ptr, make_tuple(M0, K0), make_tuple(StrideQ, 1), Number<32>{}, Number<1>{});
+
+        const auto k_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            k_ptr, make_tuple(N0, K0), make_tuple(StrideK, 1), Number<32>{}, Number<1>{});
+
+        const auto v_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            v_ptr, make_tuple(N1, N0), make_tuple(StrideV, 1), Number<32>{}, Number<1>{});
+
+        // Q/K/V DRAM window
+        auto q_dram_window = make_tile_window(
+            q_dram, make_tuple(Number<kM0PerBlock>{}, Number<kK0PerBlock>{}), {iM0, 0});
+
+        auto k_dram_window = make_tile_window(
+            k_dram, make_tuple(Number<kN0PerBlock>{}, Number<kK0PerBlock>{}), {0, 0});
+
+        auto v_dram_window = make_tile_window(
+            v_dram, make_tuple(Number<kN1PerBlock>{}, Number<kN0PerBlock>{}), {iN1, 0});
+
+        // Block GEMM0 pipeline and Block GEMM1
+        constexpr auto gemm0_pipeline = BlockGemm0Pipeline{};
+        constexpr auto gemm1          = BlockGemm1{};
+
+        // reduction function for softmax
+        const auto f_max = [](auto e0, auto e1) { return max(e0, e1); };
+        const auto f_sum = [](auto e0, auto e1) { return e0 + e1; };
+
+        // infer Sacc, S, P, M, L, Oacc type
+        using SaccBlockTileType =
+            decltype(gemm0_pipeline(q_dram_window, k_dram_window, 0, nullptr));
+
+        using SBlockTileType = decltype(tile_elementwise_in(
+            type_convert<SMPLComputeDataType, SaccDataType>, SaccBlockTileType{}));
+
+        using PBlockTileType = decltype(tile_elementwise_in(type_convert<PDataType, SaccDataType>,
+                                                            SaccBlockTileType{}));
+
+        using MLBlockTileType = decltype(block_tile_reduce<SMPLComputeDataType>(
+            SBlockTileType{}, Sequence<1>{}, f_max, SMPLComputeDataType{0}));
+
+        using OaccBlockTileType = decltype(gemm1(PBlockTileType{}, v_dram_window, smem_ptr));
+
+        // init Oacc, M, L
+        auto o_acc = OaccBlockTileType{};
+        auto m     = MLBlockTileType{};
+        auto l     = MLBlockTileType{};
+
+        tile_elementwise_inout([](auto& e) { e = 0; }, o_acc);
+        tile_elementwise_inout([](auto& e) { e = NumericLimits<SMPLComputeDataType>::Lowest(); },
+                               m);
+        tile_elementwise_inout([](auto& e) { e = 0; }, l);
+
+        // loop over Column of S (J loop)
+        index_t iN0 = 0;
+
+        do
+        {
+            // Sacc{j} = Q * K{j}
+            const auto s_acc =
+                gemm0_pipeline(q_dram_window, k_dram_window, K0 / kK0PerBlock, smem_ptr);
+
+            // S{j}
+            const auto s =
+                tile_elementwise_in(type_convert<SMPLComputeDataType, SaccDataType>, s_acc);
+
+            // m_local = rowmax(S{j})
+            auto m_local = block_tile_reduce<SMPLComputeDataType>(
+                s, Sequence<1>{}, f_max, NumericLimits<SMPLComputeDataType>::Lowest());
+
+            block_tile_reduce_sync(m_local, f_max);
+
+            // m{j-1}
+            const auto m_old = m;
+
+            // m{j}
+            tile_elementwise_inout(
+                [](auto& e0, auto e1, auto e2) { e0 = max(e1, e2); }, m, m_old, m_local);
+
+            // Pcompute{j}
+            auto p_compute =
+                make_static_distributed_tensor<SMPLComputeDataType>(s.GetTileDistribution());
+
+            constexpr auto p_spans = decltype(p_compute)::GetDistributedSpans();
+
+            sweep_tile_span(p_spans[I0], [&](auto idx0) {
+                constexpr auto i_idx = make_tuple(idx0);
+
+                sweep_tile_span(p_spans[I1], [&](auto idx1) {
+                    constexpr auto i_j_idx = make_tuple(idx0, idx1);
+
+                    p_compute(i_j_idx) = math::exp(s[i_j_idx] - m[i_idx]);
+                });
+            });
+
+            // rowsum(Pcompute{j})
+            auto rowsum_p = block_tile_reduce<SMPLComputeDataType>(
+                p_compute, Sequence<1>{}, f_sum, SMPLComputeDataType{0});
+
+            block_tile_reduce_sync(rowsum_p, f_sum);
+
+            // l{j}, Oacc{j}
+            sweep_tile_span(p_spans[I0], [&](auto idx0) {
+                constexpr auto i_idx = make_tuple(idx0);
+
+                const auto tmp = math::exp(m_old[i_idx] - m[i_idx]);
+
+                l(i_idx) = tmp * l[i_idx] + rowsum_p[i_idx];
+
+                sweep_tile_span(p_spans[I1], [&](auto idx1) {
+                    constexpr auto i_j_idx = make_tuple(idx0, idx1);
+
+                    // FIXME: this use different equation from FA v2 paper,
+                    // but produce correct result.
+                    // Is the equation wrong?
+                    o_acc(i_j_idx) *= tmp;
+                });
+            });
+
+            // type cast Pcompute{j} into P{j}
+            const auto p =
+                tile_elementwise_in(type_convert<PDataType, SMPLComputeDataType>, p_compute);
+
+            // wait for gemm0 pipeline to finish reading Lds
+            block_sync_lds();
+
+            // Block GEMM1: Oacc{j} += P{j} * V{j}
+            gemm1(o_acc, p, v_dram_window, smem_ptr);
+
+            // move K/V tile windows for next iteration (J loop)
+            move_tile_window(k_dram_window, {kN0PerBlock, 0});
+            move_tile_window(v_dram_window, {0, kN0PerBlock});
+
+            // wait for gemm1 to finish reading Lds, before next iteration (J loop)
+            block_sync_lds();
+
+            iN0 += kN0PerBlock;
+
+        } while(iN0 < N0);
+
+        // O
+        constexpr auto o_spans = decltype(o_acc)::GetDistributedSpans();
+
+        sweep_tile_span(o_spans[I0], [&](auto idx0) {
+            constexpr auto i_idx = make_tuple(idx0);
+
+            const auto tmp = 1 / l[i_idx];
+
+            sweep_tile_span(o_spans[I1], [&](auto idx1) {
+                constexpr auto i_j_idx = make_tuple(idx0, idx1);
+
+                o_acc(i_j_idx) *= tmp;
+            });
+        });
+
+        // type cast Oacc into O
+        const auto o = tile_elementwise_in(type_convert<ODataType, OaccDataType>, o_acc);
+
+        // O DRAM and O DRAM window
+        auto o_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            o_ptr, make_tuple(M0, N1), make_tuple(StrideO, 1), Number<32>{}, Number<1>{});
+
+        auto o_dram_window =
+            make_tile_window(o_dram,
+                             make_tuple(Number<kM0PerBlock>{}, Number<kN1PerBlock>{}),
+                             {iM0, iN1},
+                             o.GetTileDistribution());
+
+        // store O
+        store_tile(o_dram_window, o);
+    }
+};
diff --git a/example/91_tile_program/im2col.cpp b/example/91_tile_program/im2col.cpp
new file mode 100644
index 0000000000..83a8ba55fd
--- /dev/null
+++ b/example/91_tile_program/im2col.cpp
@@ -0,0 +1,356 @@
+#include <string_view>
+#include <tuple>
+#include <array>
+#include <utility>
+#include <type_traits>
+#include <cstring>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor/tensor_view.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_window.hpp"
+#include "ck/tile_program/tile/load_tile.hpp"
+#include "ck/tile_program/tile/store_tile.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+
+template <typename T>
+void reference_im2col(Tensor<T>& in_mtx_host_ref,
+                      const Tensor<T>& in_host,
+                      int /*N*/,
+                      int /*K*/,
+                      int C,
+                      int /*Y*/,
+                      int X,
+                      int Hi,
+                      int Wi,
+                      int Ho,
+                      int Wo,
+                      int ConvStrideH,
+                      int ConvStrideW,
+                      int ConvDilationH,
+                      int ConvDilationW,
+                      int InLeftPadH,
+                      int InLeftPadW,
+                      int /*InRightPadH*/,
+                      int /*InRightPadW*/)
+{
+    int GemmM = in_mtx_host_ref.GetLengths()[0];
+    int GemmK = in_mtx_host_ref.GetLengths()[1];
+
+    for(int gemm_m = 0; gemm_m < GemmM; ++gemm_m)
+    {
+        int mtmp = gemm_m;
+        int n    = mtmp / (Ho * Wo);
+        mtmp -= n * Ho * Wo;
+        int ho = mtmp / Wo;
+        int wo = mtmp - ho * Wo;
+
+        for(int gemm_k = 0; gemm_k < GemmK; ++gemm_k)
+        {
+            int ktmp = gemm_k;
+            int y    = ktmp / (X * C);
+            ktmp -= y * X * C;
+            int x = ktmp / C;
+            int c = ktmp - x * C;
+
+            int hi = y * ConvDilationH + ho * ConvStrideH - InLeftPadH;
+            int wi = x * ConvDilationW + wo * ConvStrideW - InLeftPadW;
+
+            bool inbound = (hi >= 0 && hi < Hi && wi >= 0 && wi < Wi);
+
+            in_mtx_host_ref(gemm_m, gemm_k) = inbound ? in_host(n, hi, wi, c) : 0;
+        }
+    }
+}
+
+template <ck::index_t NDimSpatial,
+          typename T,
+          ck::index_t kBlockSize,
+          ck::index_t kMPerBlock,
+          ck::index_t kKPerBlock>
+struct Im2Col
+{
+    __host__ __device__ static constexpr auto MakeBlockCopyTileDistribution()
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        constexpr index_t NumWarp = kBlockSize / get_warp_size();
+
+        constexpr index_t K1 = 16 / sizeof(T);
+        constexpr index_t K0 = kKPerBlock / K1;
+
+        constexpr index_t M2 = get_warp_size() / K0;
+        constexpr index_t M1 = NumWarp;
+        constexpr index_t M0 = kMPerBlock / (M1 * M2);
+
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<>,
+                                           Tuple<Sequence<M0, M1, M2>, Sequence<K0, K1>>,
+                                           Tuple<Sequence<1>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1>, Sequence<2, 0>>,
+                                           Sequence<1, 2>,
+                                           Sequence<0, 1>>{});
+    }
+
+    __host__ __device__ void
+    operator()(const std::array<ck::index_t, NDimSpatial + 2>& a_n_wis_c_lengths,
+               const std::array<ck::index_t, NDimSpatial + 2>& /* a_n_wis_c_strides */,
+               const std::array<ck::index_t, NDimSpatial + 2>& b_k_xs_c_lengths,
+               const std::array<ck::index_t, NDimSpatial + 2>& /* b_k_xs_c_strides */,
+               const std::array<ck::index_t, NDimSpatial + 2>& c_n_wos_k_lengths,
+               const std::array<ck::index_t, NDimSpatial + 2>& /* c_n_wos_k_strides */,
+               const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+               const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+               const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+               const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+               const std::array<ck::index_t, 2> a_gemmm_gemmk_lengths,
+               const std::array<ck::index_t, 2> a_gemmm_gemmk_strides,
+               const T* p_a_img,
+               T* p_a_mtx)
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        const index_t N = a_n_wis_c_lengths[0];
+        const index_t C = a_n_wis_c_lengths[3];
+
+        const index_t Hi = a_n_wis_c_lengths[1];
+        const index_t Wi = a_n_wis_c_lengths[2];
+
+        const index_t Ho = c_n_wos_k_lengths[1];
+        const index_t Wo = c_n_wos_k_lengths[2];
+
+        const index_t Y = b_k_xs_c_lengths[1];
+        const index_t X = b_k_xs_c_lengths[2];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const auto a_n_hi_wi_c = make_naive_tensor_view_packed<AddressSpaceEnum::Global>(
+            p_a_img, make_tuple(N, Hi, Wi, C), Number<32>{});
+
+        const auto a_n_hip_wip_c = transform_tensor_view(
+            a_n_hi_wi_c,
+            make_tuple(make_pass_through_transform(N),
+                       make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                       make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+        const auto a_n_y_ho_x_wo_c = transform_tensor_view(
+            a_n_hip_wip_c,
+            make_tuple(
+                make_pass_through_transform(N),
+                make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+        const auto src_gemmm_gemmk =
+            transform_tensor_view(a_n_y_ho_x_wo_c,
+                                  make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
+                                             make_merge_transform(make_tuple(Y, X, C))),
+                                  make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}),
+                                  make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        auto dst_gemmm_gemmk = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_a_mtx,
+            make_tuple(a_gemmm_gemmk_lengths[0], a_gemmm_gemmk_lengths[1]),
+            make_tuple(a_gemmm_gemmk_strides[0], a_gemmm_gemmk_strides[1]),
+            Number<32>{},
+            Number<1>{});
+
+        const auto numGemmM = a_gemmm_gemmk_lengths[0];
+        const auto numGemmK = a_gemmm_gemmk_lengths[1];
+
+        const auto id_block = get_block_id();
+
+        const auto num_tile_m = __builtin_amdgcn_readfirstlane(numGemmM / kMPerBlock);
+
+        const auto block2tile = make_cluster_descriptor(make_tuple(num_tile_m));
+
+        const auto i_gemmm_gemmk = block2tile.CalculateBottomIndex(make_multi_index(id_block));
+
+        const auto iGemmM = __builtin_amdgcn_readfirstlane(i_gemmm_gemmk[0]) * kMPerBlock;
+
+        // src window
+        auto src_block_window =
+            make_tile_window(src_gemmm_gemmk,
+                             make_tuple(Number<kMPerBlock>{}, Number<kKPerBlock>{}),
+                             {iGemmM, 0},
+                             MakeBlockCopyTileDistribution());
+
+        // dst window
+        auto dst_block_window = make_tile_window(
+            dst_gemmm_gemmk, make_tuple(Number<kMPerBlock>{}, Number<kKPerBlock>{}), {iGemmM, 0});
+
+        index_t iGemmK = 0;
+
+        do
+        {
+            const auto block_tile = load_tile(src_block_window);
+
+            store_tile(dst_block_window, block_tile);
+
+            move_tile_window(src_block_window, {0, kKPerBlock});
+            move_tile_window(dst_block_window, {0, kKPerBlock});
+
+            iGemmK += kKPerBlock;
+
+        } while(iGemmK < numGemmK);
+    }
+};
+
+int main()
+{
+    using DataType = ck::half_t;
+
+    constexpr ck::index_t NumDimSpatial = 2;
+
+    ck::index_t N  = 128;
+    ck::index_t K  = 1;
+    ck::index_t C  = 256;
+    ck::index_t Y  = 3;
+    ck::index_t X  = 3;
+    ck::index_t Hi = 28;
+    ck::index_t Wi = 28;
+    ck::index_t Ho = 14;
+    ck::index_t Wo = 14;
+
+    std::array<ck::index_t, NumDimSpatial + 2> in_lengths{N, Hi, Wi, C};
+    std::array<ck::index_t, NumDimSpatial + 2> in_strides{0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 2> wei_lengths{K, Y, X, C};
+    std::array<ck::index_t, NumDimSpatial + 2> wei_strides{0, 0, 0, 1};
+
+    std::array<ck::index_t, NumDimSpatial + 2> out_lengths{N, Ho, Wo, K};
+    std::array<ck::index_t, NumDimSpatial + 2> out_strides{0, 0, 0, 1};
+
+    std::partial_sum(rbegin(in_lengths),
+                     std::prev(rend(in_lengths)),
+                     std::next(rbegin(in_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(wei_lengths),
+                     std::prev(rend(wei_lengths)),
+                     std::next(rbegin(wei_strides)),
+                     std::multiplies<>{});
+    std::partial_sum(rbegin(out_lengths),
+                     std::prev(rend(out_lengths)),
+                     std::next(rbegin(out_strides)),
+                     std::multiplies<>{});
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{2, 2};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
+
+    // matrix
+    std::array<ck::index_t, 2> in_mtx_lengths{N * Ho * Wo, Y * X * C};
+    std::array<ck::index_t, 2> in_mtx_strides{0, 1};
+
+    std::partial_sum(rbegin(in_mtx_lengths),
+                     std::prev(rend(in_mtx_lengths)),
+                     std::next(rbegin(in_mtx_strides)),
+                     std::multiplies<>{});
+
+    // host verify
+    Tensor<DataType> in_host(in_lengths, in_strides);
+    Tensor<DataType> in_mtx_host_ref(in_mtx_lengths, in_mtx_strides);
+    Tensor<DataType> in_mtx_host_dev(in_mtx_lengths, in_mtx_strides);
+
+    std::cout << " image tensor element size: " << in_host.GetElementSize() << std::endl;
+    std::cout << "matrix tensor element size: " << in_mtx_host_ref.GetElementSize() << std::endl;
+
+    std::cout << " image tensor element space size: " << in_host.GetElementSpaceSize() << std::endl;
+    std::cout << "matrix tensor element sapce size: " << in_mtx_host_ref.GetElementSpaceSize()
+              << std::endl;
+
+    ck::utils::FillUniformDistributionIntegerValue<DataType>{-5.f, 5.f}(in_host);
+
+    reference_im2col(in_mtx_host_ref,
+                     in_host,
+                     N,
+                     K,
+                     C,
+                     Y,
+                     X,
+                     Hi,
+                     Wi,
+                     Ho,
+                     Wo,
+                     filter_strides[0],
+                     filter_strides[1],
+                     filter_dilations[0],
+                     filter_dilations[1],
+                     input_left_pads[0],
+                     input_left_pads[1],
+                     input_right_pads[0],
+                     input_right_pads[1]);
+
+    DeviceMem in_buf(sizeof(DataType) * in_host.GetElementSpaceSize());
+    DeviceMem in_mtx_buf(sizeof(DataType) * in_mtx_host_ref.GetElementSpaceSize());
+
+    in_buf.ToDevice(in_host.mData.data());
+
+    constexpr ck::index_t kBlockSize = 256;
+
+    constexpr ck::index_t kGemmMPerBlock = 256;
+    constexpr ck::index_t kGemmKPerBlock = 128;
+
+    ck::index_t kGridSize = (N * Ho * Wo) / kGemmMPerBlock;
+
+    float ave_time =
+        launch_kernel(StreamConfig{nullptr, true},
+                      Im2Col<2, DataType, kBlockSize, kGemmMPerBlock, kGemmKPerBlock>{},
+                      kGridSize,
+                      kBlockSize,
+                      0,
+                      in_lengths,
+                      in_strides,
+                      wei_lengths,
+                      wei_strides,
+                      out_lengths,
+                      out_strides,
+                      filter_strides,
+                      filter_dilations,
+                      input_left_pads,
+                      input_right_pads,
+                      in_mtx_lengths,
+                      in_mtx_strides,
+                      static_cast<DataType*>(in_buf.GetDeviceBuffer()),
+                      static_cast<DataType*>(in_mtx_buf.GetDeviceBuffer()));
+
+    std::size_t num_btype = sizeof(DataType) * in_host.GetElementSize() +
+                            sizeof(DataType) * in_mtx_host_ref.GetElementSize();
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+
+    in_mtx_buf.FromDevice(in_mtx_host_dev.mData.data());
+
+    return !ck::utils::check_err(in_mtx_host_dev, in_mtx_host_ref);
+}
diff --git a/example/91_tile_program/reduce.cpp b/example/91_tile_program/reduce.cpp
new file mode 100644
index 0000000000..5f94e35f0e
--- /dev/null
+++ b/example/91_tile_program/reduce.cpp
@@ -0,0 +1,105 @@
+#include <cstring>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor/tensor_view.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+#include "reduce.hpp"
+
+template <typename ADataType, typename AccDataType, typename BDataType>
+void reference_reduce(const Tensor<ADataType>& a_m_n, Tensor<BDataType>& b_m)
+{
+    auto f = [&](auto m) {
+        const int N = a_m_n.mDesc.GetLengths()[1];
+
+        AccDataType v_acc = 0;
+
+        for(int n = 0; n < N; ++n)
+        {
+            const ADataType v_a = a_m_n(m, n);
+
+            v_acc += v_a;
+        }
+
+        b_m(m) = ck::type_convert<BDataType>(v_acc);
+    };
+
+    make_ParallelTensorFunctor(f, b_m.mDesc.GetLengths()[0])(std::thread::hardware_concurrency());
+}
+
+int main(int argc, char* argv[])
+{
+    using ADataType   = ck::half_t;
+    using AccDataType = float;
+    using BDataType   = ck::half_t;
+
+    ck::index_t M = 3328;
+    ck::index_t N = 4096;
+
+    if(argc == 3)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+    }
+
+    std::array<ck::index_t, 2> a_lengths{M, N};
+    std::array<ck::index_t, 2> a_strides{N, 1};
+
+    std::array<ck::index_t, 1> b_lengths{M};
+    std::array<ck::index_t, 1> b_strides{1};
+
+    // host verify
+    Tensor<ADataType> a_host(a_lengths, a_strides);
+    Tensor<BDataType> b_host_ref(b_lengths, b_strides);
+    Tensor<BDataType> b_host_dev(b_lengths, b_strides);
+
+    ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_host);
+
+    // reference
+    reference_reduce<ADataType, AccDataType, BDataType>(a_host, b_host_ref);
+
+    DeviceMem a_buf(sizeof(ADataType) * a_host.GetElementSpaceSize());
+    DeviceMem b_buf(sizeof(BDataType) * b_host_ref.GetElementSpaceSize());
+
+    a_buf.ToDevice(a_host.mData.data());
+
+    constexpr ck::index_t kMPerBlock = 128;
+    constexpr ck::index_t kNPerBlock = 128;
+
+    constexpr ck::index_t kBlockSize = 256;
+    ck::index_t kGridSize            = (M / kMPerBlock);
+
+    std::cout << "grid size " << kGridSize << std::endl;
+
+    const auto kernel =
+        Reduce<ADataType, AccDataType, BDataType, kBlockSize, kMPerBlock, kNPerBlock>{};
+
+    float ave_time = launch_kernel(StreamConfig{nullptr, true},
+                                   kernel,
+                                   kGridSize,
+                                   kBlockSize,
+                                   0,
+                                   static_cast<ADataType*>(a_buf.GetDeviceBuffer()),
+                                   static_cast<BDataType*>(b_buf.GetDeviceBuffer()),
+                                   M,
+                                   N);
+
+    b_buf.FromDevice(b_host_dev.mData.data());
+
+    std::size_t num_btype = sizeof(ADataType) * M * N + sizeof(BDataType) * M;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+
+    return !ck::utils::check_err(b_host_dev, b_host_ref);
+}
diff --git a/example/91_tile_program/reduce.hpp b/example/91_tile_program/reduce.hpp
new file mode 100644
index 0000000000..a2bc6ccd23
--- /dev/null
+++ b/example/91_tile_program/reduce.hpp
@@ -0,0 +1,141 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_window.hpp"
+#include "ck/tile_program/tile/load_tile.hpp"
+#include "ck/tile_program/tile/store_tile.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/block_tile/block_reduce.hpp"
+
+template <typename ADataType,
+          typename AccDataType,
+          typename BDataType,
+          ck::index_t kBlockSize,
+          ck::index_t kMPerBlock,
+          ck::index_t kNPerBlock>
+struct Reduce
+{
+#if 0
+     __device__ static constexpr auto MakeABlockTileDistribution()
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        // 2x2 wave
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<>,
+                                           Tuple<Sequence<2, 2, 4, 2, 4>, Sequence<2, 2, 32>>,
+                                           Tuple<Sequence<1, 2>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1, 1>, Sequence<3, 2>>,
+                                           Sequence<1, 2, 1, 1>,
+                                           Sequence<0, 0, 2, 4>>{});
+    }
+#elif 0
+    __device__ static constexpr auto MakeABlockTileDistribution()
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        // 2x2 wave
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<>,
+                                           Tuple<Sequence<2, 2, 32>, Sequence<2, 2, 4, 2, 4>>,
+                                           Tuple<Sequence<2, 1>, Sequence<2, 1>>,
+                                           Tuple<Sequence<1, 1>, Sequence<3, 2>>,
+                                           Sequence<2, 1, 2, 2>,
+                                           Sequence<0, 0, 2, 4>>{});
+    }
+#elif 1
+    __device__ static constexpr auto MakeABlockTileDistribution()
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        // 4x1 wave
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<>,
+                                           Tuple<Sequence<1, 4, 4, 2, 4>, Sequence<4, 1, 32>>,
+                                           Tuple<Sequence<1, 2>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1, 1>, Sequence<3, 2>>,
+                                           Sequence<1, 2, 1, 1>,
+                                           Sequence<0, 0, 2, 4>>{});
+    }
+#endif
+
+    __device__ void
+    operator()(const ADataType* p_a, BDataType* p_b, ck::index_t M, ck::index_t N) const
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+        using namespace ck::tile_program::block;
+
+        const auto a_m_n = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_a, make_tuple(M, N), make_tuple(N, 1), Number<32>{}, Number<1>{});
+
+        const auto iM = get_block_id() * kMPerBlock;
+
+        // A window
+        auto a_block_window =
+            make_tile_window(a_m_n,
+                             make_tuple(Number<kMPerBlock>{}, Number<kNPerBlock>{}),
+                             {iM, 0},
+                             MakeABlockTileDistribution());
+
+        const auto f_reduce = [](const auto& v0, const auto& v1) { return v0 + v1; };
+
+        const ADataType reduce_init_value = 0;
+
+        constexpr auto reduce_dims = Sequence<1>{};
+
+        // Acc tile
+        // FIXME: support cross warp reduction
+        auto acc_block_tensor = decltype(block_tile_reduce<AccDataType>(
+            load_tile(a_block_window), reduce_dims, f_reduce, reduce_init_value)){};
+
+        // init Acc tile
+        tile_elementwise_inout(
+            [&](auto& acc) { acc = type_convert<AccDataType>(reduce_init_value); },
+            acc_block_tensor);
+
+        // loop
+        index_t iN = 0;
+
+        do
+        {
+            const auto a_block_tensor = load_tile(a_block_window);
+
+            // FIXME: support cross warp reduction
+            block_tile_reduce(acc_block_tensor, a_block_tensor, reduce_dims, f_reduce);
+
+            move_tile_window(a_block_window, {0, kNPerBlock});
+
+            iN += kNPerBlock;
+
+        } while(iN < N);
+
+        // FIXME: support cross warp reduction
+        block_tile_reduce_sync(acc_block_tensor, f_reduce);
+
+        // convert acc_block_tensor to b_block_tensor
+        const auto b_block_tensor = tile_elementwise_in(
+            [](const auto& acc) { return type_convert<BDataType>(acc); }, acc_block_tensor);
+
+        // B
+        const auto b_m = make_naive_tensor_view_packed<AddressSpaceEnum::Global>(
+            p_b, make_tuple(M), Number<32>{});
+
+        // B window
+        auto b_block_window = make_tile_window(b_m, make_tuple(Number<kMPerBlock>{}), {iM});
+
+        // store B tile
+        store_tile(b_block_window, b_block_tensor);
+    }
+};
diff --git a/example/91_tile_program/reference_batched_gemm.hpp b/example/91_tile_program/reference_batched_gemm.hpp
new file mode 100644
index 0000000000..3898950400
--- /dev/null
+++ b/example/91_tile_program/reference_batched_gemm.hpp
@@ -0,0 +1,36 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+template <typename ADataType, typename BDataType, typename AccDataType, typename CDataType>
+void reference_batched_gemm(const Tensor<ADataType>& a_b_m_k,
+                            const Tensor<BDataType>& b_b_n_k,
+                            Tensor<CDataType>& c_b_m_n)
+{
+    const int N = b_b_n_k.mDesc.GetLengths()[1];
+    const int K = b_b_n_k.mDesc.GetLengths()[2];
+
+    auto f = [&](auto batch, auto m) {
+        for(int n = 0; n < N; ++n)
+        {
+            AccDataType v_acc = 0;
+
+            for(int k = 0; k < K; ++k)
+            {
+                ADataType v_a = a_b_m_k(batch, m, k);
+                BDataType v_b = b_b_n_k(batch, n, k);
+
+                v_acc += ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
+            }
+
+            c_b_m_n(batch, m, n) = ck::type_convert<CDataType>(v_acc);
+        }
+    };
+
+    make_ParallelTensorFunctor(f, c_b_m_n.mDesc.GetLengths()[0], c_b_m_n.mDesc.GetLengths()[1])(
+        std::thread::hardware_concurrency());
+}
diff --git a/example/91_tile_program/reference_batched_softmax.hpp b/example/91_tile_program/reference_batched_softmax.hpp
new file mode 100644
index 0000000000..344acfee73
--- /dev/null
+++ b/example/91_tile_program/reference_batched_softmax.hpp
@@ -0,0 +1,46 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+template <typename ADataType, typename AccDataType, typename BDataType>
+void reference_batched_softmax(const Tensor<ADataType>& a_b_m_n, Tensor<BDataType>& b_b_m_n)
+{
+    const int N = a_b_m_n.mDesc.GetLengths()[2];
+
+    auto f = [&](auto batch, auto m) {
+        AccDataType v_max = ck::NumericLimits<ADataType>::Lowest();
+
+        // max
+        for(int n = 0; n < N; ++n)
+        {
+            const ADataType v_a = a_b_m_n(batch, m, n);
+
+            v_max = v_max < v_a ? v_a : v_max;
+        }
+
+        AccDataType v_exp_sum = 0;
+
+        // sum
+        for(int n = 0; n < N; ++n)
+        {
+            const ADataType v_a = a_b_m_n(batch, m, n);
+
+            v_exp_sum += ck::math::exp(v_a - v_max);
+        }
+
+        // elementwise
+        for(int n = 0; n < N; ++n)
+        {
+            const ADataType v_a = a_b_m_n(batch, m, n);
+
+            b_b_m_n(batch, m, n) = ck::math::exp(v_a - v_max) / v_exp_sum;
+        }
+    };
+
+    make_ParallelTensorFunctor(f, b_b_m_n.mDesc.GetLengths()[0], b_b_m_n.mDesc.GetLengths()[1])(
+        std::thread::hardware_concurrency());
+}
diff --git a/example/91_tile_program/reference_gemm.hpp b/example/91_tile_program/reference_gemm.hpp
new file mode 100644
index 0000000000..a558e5719d
--- /dev/null
+++ b/example/91_tile_program/reference_gemm.hpp
@@ -0,0 +1,35 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+template <typename ADataType, typename BDataType, typename AccDataType, typename CDataType>
+void reference_gemm(const Tensor<ADataType>& a_m_k,
+                    const Tensor<BDataType>& b_n_k,
+                    Tensor<CDataType>& c_m_n)
+{
+    const int N = b_n_k.mDesc.GetLengths()[0];
+    const int K = b_n_k.mDesc.GetLengths()[1];
+
+    auto f = [&](auto m) {
+        for(int n = 0; n < N; ++n)
+        {
+            AccDataType v_acc = 0;
+
+            for(int k = 0; k < K; ++k)
+            {
+                ADataType v_a = a_m_k(m, k);
+                BDataType v_b = b_n_k(n, k);
+
+                v_acc += ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
+            }
+
+            c_m_n(m, n) = ck::type_convert<CDataType>(v_acc);
+        }
+    };
+
+    make_ParallelTensorFunctor(f, c_m_n.mDesc.GetLengths()[0])(std::thread::hardware_concurrency());
+}
diff --git a/example/91_tile_program/reference_softmax.hpp b/example/91_tile_program/reference_softmax.hpp
new file mode 100644
index 0000000000..41ad5b2726
--- /dev/null
+++ b/example/91_tile_program/reference_softmax.hpp
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+template <typename ADataType, typename AccDataType, typename BDataType>
+void reference_softmax(const Tensor<ADataType>& a_m_n, Tensor<BDataType>& b_m_n)
+{
+    auto f = [&](auto m) {
+        const int N = a_m_n.mDesc.GetLengths()[1];
+
+        AccDataType v_max = ck::NumericLimits<ADataType>::Lowest();
+
+        // max
+        for(int n = 0; n < N; ++n)
+        {
+            const ADataType v_a = a_m_n(m, n);
+
+            v_max = v_max < v_a ? v_a : v_max;
+        }
+
+        AccDataType v_exp_sum = 0;
+
+        // sum
+        for(int n = 0; n < N; ++n)
+        {
+            const ADataType v_a = a_m_n(m, n);
+
+            v_exp_sum += ck::math::exp(v_a - v_max);
+        }
+
+        // elementwise
+        for(int n = 0; n < N; ++n)
+        {
+            const ADataType v_a = a_m_n(m, n);
+
+            b_m_n(m, n) = ck::math::exp(v_a - v_max) / v_exp_sum;
+        }
+    };
+
+    make_ParallelTensorFunctor(f, b_m_n.mDesc.GetLengths()[0])(std::thread::hardware_concurrency());
+}
diff --git a/example/91_tile_program/softmax.cpp b/example/91_tile_program/softmax.cpp
new file mode 100644
index 0000000000..f78d609f20
--- /dev/null
+++ b/example/91_tile_program/softmax.cpp
@@ -0,0 +1,85 @@
+#include <cstring>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor/tensor_view.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+#include "reference_softmax.hpp"
+#include "softmax.hpp"
+
+int main(int argc, char* argv[])
+{
+    using ADataType   = ck::half_t;
+    using AccDataType = float;
+    using BDataType   = ck::half_t;
+
+    ck::index_t M = 13312;
+    ck::index_t N = 4096;
+
+    if(argc == 3)
+    {
+        M = std::stoi(argv[1]);
+        N = std::stoi(argv[2]);
+    }
+
+    std::array<ck::index_t, 2> a_lengths{M, N};
+    std::array<ck::index_t, 2> a_strides{N, 1};
+
+    std::array<ck::index_t, 2> b_lengths{M, N};
+    std::array<ck::index_t, 2> b_strides{N, 1};
+
+    // host verify
+    Tensor<ADataType> a_host(a_lengths, a_strides);
+    Tensor<BDataType> b_host_ref(b_lengths, b_strides);
+    Tensor<BDataType> b_host_dev(b_lengths, b_strides);
+
+    ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_host);
+
+    // reference
+    reference_softmax<ADataType, AccDataType, BDataType>(a_host, b_host_ref);
+
+    DeviceMem a_buf(sizeof(ADataType) * a_host.GetElementSpaceSize());
+    DeviceMem b_buf(sizeof(BDataType) * b_host_ref.GetElementSpaceSize());
+
+    a_buf.ToDevice(a_host.mData.data());
+
+    constexpr ck::index_t kMPerBlock = 128;
+    constexpr ck::index_t kNPerBlock = 128;
+
+    constexpr ck::index_t kBlockSize = 256;
+    ck::index_t kGridSize            = (M / kMPerBlock);
+
+    std::cout << "grid size " << kGridSize << std::endl;
+
+    const auto kernel =
+        Softmax<ADataType, AccDataType, BDataType, kBlockSize, kMPerBlock, kNPerBlock>{};
+
+    float ave_time = launch_kernel(StreamConfig{nullptr, true},
+                                   kernel,
+                                   kGridSize,
+                                   kBlockSize,
+                                   0,
+                                   static_cast<ADataType*>(a_buf.GetDeviceBuffer()),
+                                   static_cast<BDataType*>(b_buf.GetDeviceBuffer()),
+                                   M,
+                                   N);
+
+    b_buf.FromDevice(b_host_dev.mData.data());
+
+    std::size_t num_btype = sizeof(ADataType) * M * N + sizeof(BDataType) * M * N;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+
+    return !ck::utils::check_err(b_host_dev, b_host_ref);
+}
diff --git a/example/91_tile_program/softmax.hpp b/example/91_tile_program/softmax.hpp
new file mode 100644
index 0000000000..b69c5a10d8
--- /dev/null
+++ b/example/91_tile_program/softmax.hpp
@@ -0,0 +1,273 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_window.hpp"
+#include "ck/tile_program/tile/load_tile.hpp"
+#include "ck/tile_program/tile/store_tile.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/block_tile/block_reduce.hpp"
+
+template <typename ADataType,
+          typename AccDataType,
+          typename BDataType,
+          ck::index_t kBlockSize,
+          ck::index_t kMPerBlock,
+          ck::index_t kNPerBlock>
+struct Softmax
+{
+    __device__ static constexpr auto MakeABlockTileDistribution()
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        // 4x1 wave
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<>,
+                                           Tuple<Sequence<1, 4, 4, 2, 4>, Sequence<4, 1, 32>>,
+                                           Tuple<Sequence<1, 2>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1, 1>, Sequence<3, 2>>,
+                                           Sequence<1, 2, 1, 1>,
+                                           Sequence<0, 0, 2, 4>>{});
+    }
+
+    __device__ void
+    MultiPassSoftmax(const ADataType* p_a, BDataType* p_b, ck::index_t M, ck::index_t N) const
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+        using namespace ck::tile_program::block;
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        // A DRAM tensor view
+        const auto a_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_a, make_tuple(M, N), make_tuple(N, 1), Number<32>{}, Number<1>{});
+
+        const auto iM = get_block_id() * kMPerBlock;
+
+        // A DRAM window
+        auto a_dram_window =
+            make_tile_window(a_dram,
+                             make_tuple(Number<kMPerBlock>{}, Number<kNPerBlock>{}),
+                             {iM, 0},
+                             MakeABlockTileDistribution());
+
+        // m = rowmax(A)
+        const auto f_max = [](auto e0, auto e1) { return max(e0, e1); };
+
+        auto m = decltype(block_tile_reduce<AccDataType>(
+            load_tile(a_dram_window), Sequence<1>{}, f_max, ADataType{})){};
+
+        tile_elementwise_inout(
+            [&](auto& e) { e = type_convert<AccDataType>(NumericLimits<ADataType>::Lowest()); }, m);
+
+        index_t iN = 0;
+
+        do
+        {
+            // load A tile from DRAM
+            const auto a = load_tile(a_dram_window);
+
+            // m = rowmax(A)
+            block_tile_reduce(m, a, Sequence<1>{}, f_max);
+
+            move_tile_window(a_dram_window, {0, kNPerBlock});
+
+            iN += kNPerBlock;
+
+        } while(iN < N);
+
+        // cross lane reduce: max
+        block_tile_reduce_sync(m, f_max);
+
+        // reset window location
+        iN = 0;
+        move_tile_window(a_dram_window, {0, -N});
+
+        // l = rowsum(exp(A - m))
+        auto l = make_static_distributed_tensor<AccDataType>(m.GetTileDistribution());
+
+        tile_elementwise_inout([&](auto& e) { e = 0; }, l);
+
+        do
+        {
+            // load A tile from DRAM
+            const auto a = load_tile(a_dram_window);
+
+            constexpr auto a_spans = decltype(a)::GetDistributedSpans();
+
+            sweep_tile_span(a_spans[I0], [&](auto idx0) {
+                constexpr auto i_idx = make_tuple(idx0);
+
+                sweep_tile_span(a_spans[I1], [&](auto idx1) {
+                    constexpr auto i_j_idx = make_tuple(idx0, idx1);
+
+                    // l = rowsum(exp(A - m))
+                    l(i_idx) += math::exp(a[i_j_idx] - m[i_idx]);
+                });
+            });
+
+            move_tile_window(a_dram_window, {0, kNPerBlock});
+
+            iN += kNPerBlock;
+
+        } while(iN < N);
+
+        // cross lane reduce: sum
+        block_tile_reduce_sync(l, [](auto e0, auto e1) { return e0 + e1; });
+
+        // reset window location
+        iN = 0;
+        move_tile_window(a_dram_window, {0, -N});
+
+        // B DRAM tensor view
+        const auto b_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_b, make_tuple(M, N), make_tuple(N, 1), Number<32>{}, Number<1>{});
+
+        // B DRAM window
+        auto b_dram_window = make_tile_window(
+            b_dram, make_tuple(Number<kMPerBlock>{}, Number<kNPerBlock>{}), {iM, 0});
+
+        // B = exp(A - m) / l
+        do
+        {
+            // load A tile from DRAM
+            const auto a = load_tile(a_dram_window);
+
+            constexpr auto a_spans = decltype(a)::GetDistributedSpans();
+
+            auto b = make_static_distributed_tensor<BDataType>(a.GetTileDistribution());
+
+            sweep_tile_span(a_spans[I0], [&](auto idx0) {
+                constexpr auto i_idx = make_tuple(idx0);
+
+                sweep_tile_span(a_spans[I1], [&](auto idx1) {
+                    constexpr auto i_j_idx = make_tuple(idx0, idx1);
+
+                    // B = exp(A - m) / l
+                    b(i_j_idx) =
+                        type_convert<BDataType>(math::exp(a[i_j_idx] - m[i_idx]) / l[i_idx]);
+                });
+            });
+
+            // store B tile
+            store_tile(b_dram_window, b);
+
+            move_tile_window(a_dram_window, {0, kNPerBlock});
+            move_tile_window(b_dram_window, {0, kNPerBlock});
+
+            iN += kNPerBlock;
+
+        } while(iN < N);
+    }
+
+    __device__ void
+    SinglePassSoftmax(const ADataType* p_a, BDataType* p_b, ck::index_t M, ck::index_t N) const
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+        using namespace ck::tile_program::block;
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        // A DRAM tensor view
+        const auto a_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_a, make_tuple(M, N), make_tuple(N, 1), Number<32>{}, Number<1>{});
+
+        const auto iM = get_block_id() * kMPerBlock;
+
+        // A DRAM window
+        auto a_dram_window =
+            make_tile_window(a_dram,
+                             make_tuple(Number<kMPerBlock>{}, Number<kNPerBlock>{}),
+                             {iM, 0},
+                             MakeABlockTileDistribution());
+
+        // f_max
+        const auto f_max = [](auto e0, auto e1) { return max(e0, e1); };
+
+        // m = rowmax(A)
+        auto m = decltype(block_tile_reduce<AccDataType>(
+            load_tile(a_dram_window), Sequence<1>{}, f_max, ADataType{})){};
+
+        tile_elementwise_inout(
+            [&](auto& e) { e = type_convert<AccDataType>(NumericLimits<ADataType>::Lowest()); }, m);
+
+        // l = rowsum(exp(A - m))
+        auto l = make_static_distributed_tensor<AccDataType>(m.GetTileDistribution());
+
+        tile_elementwise_inout([&](auto& e) { e = 0; }, l);
+
+        // load A tile from DRAM
+        const auto a = load_tile(a_dram_window);
+
+        constexpr auto a_spans = decltype(a)::GetDistributedSpans();
+
+        // m = rowmax(A)
+        block_tile_reduce(m, a, Sequence<1>{}, f_max);
+
+        // cross lane reduce: max
+        block_tile_reduce_sync(m, f_max);
+
+        // l = rowsum(exp(A - m))
+        sweep_tile_span(a_spans[I0], [&](auto idx0) {
+            constexpr auto i_idx = make_tuple(idx0);
+
+            sweep_tile_span(a_spans[I1], [&](auto idx1) {
+                constexpr auto i_j_idx = make_tuple(idx0, idx1);
+
+                l(i_idx) += math::exp(a[i_j_idx] - m[i_idx]);
+            });
+        });
+
+        // cross lane reduce: sum
+        block_tile_reduce_sync(l, [](auto e0, auto e1) { return e0 + e1; });
+
+        auto b = make_static_distributed_tensor<BDataType>(a.GetTileDistribution());
+
+        // B = exp(A - m) / l
+        sweep_tile_span(a_spans[I0], [&](auto idx0) {
+            constexpr auto i_idx = make_tuple(idx0);
+
+            sweep_tile_span(a_spans[I1], [&](auto idx1) {
+                constexpr auto i_j_idx = make_tuple(idx0, idx1);
+
+                b(i_j_idx) = type_convert<BDataType>(math::exp(a[i_j_idx] - m[i_idx]) / l[i_idx]);
+            });
+        });
+
+        // B DRAM tensor view
+        const auto b_dram = make_naive_tensor_view<AddressSpaceEnum::Global>(
+            p_b, make_tuple(M, N), make_tuple(N, 1), Number<32>{}, Number<1>{});
+
+        // B DRAM window
+        auto b_dram_window = make_tile_window(
+            b_dram, make_tuple(Number<kMPerBlock>{}, Number<kNPerBlock>{}), {iM, 0});
+
+        // store B tile
+        store_tile(b_dram_window, b);
+    }
+
+    __device__ void
+    operator()(const ADataType* p_a, BDataType* p_b, ck::index_t M, ck::index_t N) const
+    {
+        if(N > kNPerBlock)
+        {
+            MultiPassSoftmax(p_a, p_b, M, N);
+        }
+        else
+        {
+            SinglePassSoftmax(p_a, p_b, M, N);
+        }
+    }
+};
diff --git a/example/CMakeLists.txt b/example/CMakeLists.txt
new file mode 100644
index 0000000000..1fdd2f6d14
--- /dev/null
+++ b/example/CMakeLists.txt
@@ -0,0 +1,32 @@
+include_directories(BEFORE
+    ${PROJECT_SOURCE_DIR}/include
+    ${PROJECT_SOURCE_DIR}/library/include
+)
+
+add_custom_target(examples)
+
+function(add_example_executable EXAMPLE_NAME FILE_NAME)
+    message("adding example ${EXAMPLE_NAME}")
+    add_executable(${EXAMPLE_NAME} ${FILE_NAME})
+    target_link_libraries(${EXAMPLE_NAME} PRIVATE utility)
+    add_test(NAME ${EXAMPLE_NAME} COMMAND $<TARGET_FILE:${EXAMPLE_NAME}> ${ARGN})
+    add_dependencies(examples ${EXAMPLE_NAME})
+    add_dependencies(check ${EXAMPLE_NAME})
+    rocm_install(TARGETS ${EXAMPLE_NAME} COMPONENT examples)
+endfunction(add_example_executable EXAMPLE_NAME)
+
+function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
+    message("adding example ${EXAMPLE_NAME}")
+    add_executable(${EXAMPLE_NAME} ${FILE_NAME})
+    target_link_libraries(${EXAMPLE_NAME} PRIVATE utility)
+    add_dependencies(examples ${EXAMPLE_NAME})
+    rocm_install(TARGETS ${EXAMPLE_NAME} COMPONENT examples)
+endfunction(add_example_executable_no_testing EXAMPLE_NAME)
+
+# add all example subdir
+file(GLOB dir_list LIST_DIRECTORIES true *)
+FOREACH(subdir ${dir_list})
+    IF(IS_DIRECTORY "${subdir}")
+        add_subdirectory(${subdir})
+    ENDIF()
+ENDFOREACH()
diff --git a/include/ck/ck.hpp b/include/ck/ck.hpp
new file mode 100644
index 0000000000..8eaa8356ea
--- /dev/null
+++ b/include/ck/ck.hpp
@@ -0,0 +1,243 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/config.h"
+
+#ifndef CK_DONT_USE_HIP_RUNTIME_HEADERS
+#include "hip/hip_runtime.h"
+#include "hip/hip_fp16.h"
+#endif
+
+#define CK_TIME_KERNEL 1
+
+// constant address space for kernel parameter
+// https://llvm.org/docs/AMDGPUUsage.html#address-spaces
+#define CK_CONSTANT_ADDRESS_SPACE __attribute__((address_space(4)))
+
+// launch bounds
+#define CK_USE_LAUNCH_BOUNDS 1
+
+#ifdef CK_USE_LAUNCH_BOUNDS
+// for most kernels
+#define CK_MAX_THREAD_PER_BLOCK 256
+#define CK_MIN_BLOCK_PER_CU 2
+
+// for wavelet GEMM kernel
+#define CK_WAVELET_MAX_THREAD_PER_BLOCK 512
+#define CK_WAVELET_MIN_BLOCK_PER_CU 2
+#endif
+
+// kernel attribute: amdgpu_waves_per_eu()
+#ifdef CK_USE_WAVES_PER_EU
+// for 1-wave kernels, control arguments of amdgpu_waves_per_eu() attribute
+#ifndef CK_MIN_WAVES_PER_EU
+#define CK_MIN_WAVES_PER_EU 0
+#endif
+
+#ifndef CK_MAX_WAVES_PER_EU
+#define CK_MAX_WAVES_PER_EU 0
+#endif
+
+#else
+#define CK_USE_WAVES_PER_EU 0
+#endif
+
+// buffer resource
+#ifndef __HIP_DEVICE_COMPILE__ // for host code
+#define CK_BUFFER_RESOURCE_3RD_DWORD -1
+#elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \
+    defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) ||                          \
+    defined(__gfx942__) // for GPU code
+#define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
+#elif defined(__gfx1030__) // for GPU code
+#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
+#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
+#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31004000
+#endif
+
+// FMA instruction
+#ifndef __HIP_DEVICE_COMPILE__                   // for host code, define nothing
+#elif defined(__gfx803__) || defined(__gfx900__) // for GPU code
+#define CK_USE_AMD_V_MAC_F32
+#elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) // for GPU code
+#define CK_USE_AMD_V_FMAC_F32
+#define CK_USE_AMD_V_DOT2_F32_F16
+#define CK_USE_AMD_V_DOT4_I32_I8
+#endif
+
+// MFMA instruction
+#ifndef __HIP_DEVICE_COMPILE__ // for host code
+#define CK_USE_AMD_MFMA
+#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || \
+    defined(__gfx942__) // for GPU code
+#define CK_USE_AMD_MFMA
+#endif
+
+#if(defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+#define CK_USE_AMD_MFMA_BF16_1K_OP
+#endif
+
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+#define CK_USE_AMD_MFMA_GFX940
+#endif
+
+// WMMA instruction
+#ifndef __HIP_DEVICE_COMPILE__ // for host code
+#define CK_USE_AMD_WMMA
+#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) // for GPU code
+#define CK_USE_AMD_WMMA
+#endif
+
+// buffer load
+#define CK_USE_AMD_BUFFER_LOAD 1
+
+// buffer store
+#define CK_USE_AMD_BUFFER_STORE 1
+
+// buffer atomic add: integer
+#define CK_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER 1
+
+// buffer atomic add: floating point
+#ifndef __HIP_DEVICE_COMPILE__ // for host code
+#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
+#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || \
+    defined(__gfx942__) // for GPU code
+#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
+#else // for GPU code
+#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 0
+#endif
+
+#if(defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || \
+    defined(__gfx942__)) // for GPU code
+#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 1
+#else
+#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 0
+#endif
+
+// inline asm
+#define CK_USE_AMD_INLINE_ASM 1
+
+// inner product (V_MAC/V_FMAC)
+#define CK_USE_AMD_V_MAC_INLINE_ASM 1
+
+// V_DOT inline instructions, less efficient since they require adding
+// `s_nop`s to avoid hazard
+#define CK_USE_AMD_V_DOT_INLINE_ASM 0
+
+// inner product using V_DOT with DPP8 modifiers
+#define CK_USE_AMD_V_DOT_DPP8_INLINE_ASM 1
+
+// block synchronization only s_wait lgkmcnt(0), not vmcnt(0)
+#define CK_EXPERIMENTAL_BLOCK_SYNC_LDS_WITHOUT_SYNC_VMEM 1
+
+// experimental feature: multi index implemented as array
+#define CK_EXPERIMENTAL_USE_DYNAMICALLY_INDEXED_MULTI_INDEX 1
+
+// experimental feature: static tensor descriptor
+#define CK_EXPERIMENTAL_STATIC_TENSOR_DESCRIPTOR 0
+
+// experimental feature: buffer load/store/atomic-add/ OOB trick
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting. Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter for each usage
+#ifndef CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 0
+#endif
+#define CK_EXPERIMENTAL_USE_BUFFER_STORE_OOB_CHECK_OFFSET_TRICK 1
+#define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_ADD_OOB_CHECK_OFFSET_TRICK 1
+#define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_MAX_OOB_CHECK_OFFSET_TRICK 1
+
+// experimental feature: merge transformation use magic number division
+#define CK_EXPERIMENTAL_MERGE_USE_MAGIC_DIVISION 1
+
+// experimental feature: use __builtin_memcpy instead of pointer cast to access a vector from
+// pointer of scalar
+#define CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS 0
+
+// experimental feature: optimize for inter-wave scheduling policy
+#define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING 1
+#define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS 1
+// this will let make_default_loop_scheduler() return interwave scheduling flag by default
+#define CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING 0
+// experimental feature: add instances using interwave scheduling
+#define CK_EXPERIMENTAL_INTER_WAVE_INSTANCES 1
+// experimental feature: add instances using pipeline v2
+#define CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES 1
+// experimental feature: optimize pipeline v2 by IGLP strategy (value=ID of strategy)
+#ifndef CK_EXPERIMENTAL_PIPELINE_V2_IGLP_OPT
+#define CK_EXPERIMENTAL_PIPELINE_V2_IGLP_OPT 0
+#endif
+
+// hack: have underlying assumption that need to be satsified, otherwise it's a bug
+// hack for forcing register to keep idx_diff_low_const in SGPR. idx_diff_low_const must be
+// thread-invariant, otherwise it's a bug
+// TODO: separate index calculation into "compile-time", "global", "block", "wave", "thread"
+#define CK_HACK_MERGE_CALCULATE_IDX_DIFF_LOW_CONST_USE_AMD_GCN_READ_FIRST_LANE 0
+
+// workaround: compiler crash when compiling recursive lambda
+#define CK_WORKAROUND_SWDEV_275126 1
+
+// workaround: compiler crash when using buffer load/store for i8
+#define CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE 1
+
+// workaround: compiler gnerating inefficient ds_write instructions
+#define CK_WORKAROUND_SWDEV_XXXXXX_INT8_DS_WRITE_ISSUE 1
+
+// workaround: verifaction failure, due to compiler regression, for conv bwd-data fp16 using some
+// tuning parameter
+#define CK_WORKAROUND_SWDEV_325164 0
+
+// workaround: compiler not emiting reciprocal instruction frm __frcp_rn()
+#define CK_WORKAROUND_SWDEV_383542 1
+
+// workaround: compiler issue on gfx908
+#define CK_WORKAROUND_SWDEV_388832 1
+
+// flag to enable (1) or disable (0) the debugging output in some kernels
+#define DEBUG_LOG 0
+
+// denorm test fix, required to work around dissue
+#ifndef CK_WORKAROUND_DENORM_FIX
+#define CK_WORKAROUND_DENORM_FIX 0
+#elif
+// enable only on MI200
+#define CK_WORKAROUND_DENORM_FIX = CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__)
+#endif // CK_WORKAROUND_DENORM_FIX
+
+namespace ck {
+
+enum struct ArchEnum
+{
+    Gfx90a,
+};
+
+enum struct InMemoryDataOperationEnum
+{
+    Set,
+    AtomicAdd,
+    AtomicMax,
+    Add
+};
+
+// FIXME: use regular Sequence and remove this
+template <InMemoryDataOperationEnum... Is>
+struct InMemoryDataOperationEnumSequence
+{
+    static constexpr int mSize = sizeof...(Is);
+
+    __host__ __device__ static constexpr InMemoryDataOperationEnum At(int I)
+    {
+        // the last dummy element is to prevent compiler complain about empty array, when mSize = 0
+        const InMemoryDataOperationEnum mData[mSize + 1] = {Is..., InMemoryDataOperationEnum::Set};
+        return mData[I];
+    }
+};
+
+// index type
+using index_t      = int32_t;
+using long_index_t = int64_t;
+
+} // namespace ck
diff --git a/include/ck/config.h.in b/include/ck/config.h.in
new file mode 100644
index 0000000000..13dc5da5d1
--- /dev/null
+++ b/include/ck/config.h.in
@@ -0,0 +1,102 @@
+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+#ifndef CK_CONFIG_H_IN
+#define CK_CONFIG_H_IN
+
+// clang-format off
+//
+// DataType supports in the current CK build
+//
+#ifndef DTYPES
+#cmakedefine DTYPES "@DTYPES@"
+#endif
+// if DTYPES is not defined, enable all datatypes in headerfiles
+#ifndef CK_ENABLE_ALL_DTYPES
+#cmakedefine CK_ENABLE_ALL_DTYPES @CK_ENABLE_ALL_DTYPES@
+#if defined(CK_ENABLE_ALL_DTYPES)
+#ifndef CK_ENABLE_INT8
+#define CK_ENABLE_INT8 "ON"
+#endif
+#ifndef CK_ENABLE_FP8
+#define CK_ENABLE_FP8 "ON"
+#endif
+#ifndef CK_ENABLE_FP16
+#define CK_ENABLE_FP16 "ON"
+#endif
+#ifndef CK_ENABLE_BF16
+#define CK_ENABLE_BF16 "ON"
+#endif
+#ifndef CK_ENABLE_FP32
+#define CK_ENABLE_FP32 "ON"
+#endif
+#ifndef CK_ENABLE_FP64
+#define CK_ENABLE_FP64 "ON"
+#endif
+#endif
+#endif
+// if DTYPES are selectively enabled
+#ifndef CK_ENABLE_INT8
+#cmakedefine CK_ENABLE_INT8 @CK_ENABLE_INT8@
+#endif
+
+#ifndef CK_ENABLE_FP8
+#cmakedefine CK_ENABLE_FP8 @CK_ENABLE_FP8@
+#endif
+
+#ifndef CK_ENABLE_FP16
+#cmakedefine CK_ENABLE_FP16 @CK_ENABLE_FP16@
+#endif
+
+#ifndef CK_ENABLE_BF16
+#cmakedefine CK_ENABLE_BF16 @CK_ENABLE_BF16@
+#endif
+
+#ifndef CK_ENABLE_FP32
+#cmakedefine CK_ENABLE_FP32 @CK_ENABLE_FP32@
+#endif
+
+#ifndef CK_ENABLE_FP64
+#cmakedefine CK_ENABLE_FP64 @CK_ENABLE_FP64@
+#endif
+
+//
+// Legacy DL kernel supports in the current CK build
+// by default DL kernels are turned OFF
+//
+#ifndef CK_ENABLE_DL_KERNELS
+#cmakedefine CK_ENABLE_DL_KERNELS @CK_ENABLE_DL_KERNELS@
+#endif
+
+//
+// Instances supports in the current CK build
+//
+#ifndef CK_ENABLE_INSTANCES_ONLY
+#cmakedefine CK_ENABLE_INSTANCES_ONLY @CK_ENABLE_INSTANCES_ONLY@
+#endif
+
+// clang-format on
+
+#endif // CK_CONFIG_H_IN
diff --git a/include/ck/host_utility/device_prop.hpp b/include/ck/host_utility/device_prop.hpp
new file mode 100644
index 0000000000..be1dbc1657
--- /dev/null
+++ b/include/ck/host_utility/device_prop.hpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <string>
+#include <map>
+#include <hip/hip_runtime.h>
+
+namespace ck {
+
+inline std::string get_device_name()
+{
+    hipDeviceProp_t props{};
+    int device;
+    auto status = hipGetDevice(&device);
+    if(status != hipSuccess)
+    {
+        return std::string();
+    }
+
+    status = hipGetDeviceProperties(&props, device);
+    if(status != hipSuccess)
+    {
+        return std::string();
+    }
+    const std::string raw_name(props.gcnArchName);
+
+    // https://github.com/ROCmSoftwarePlatform/MIOpen/blob/8498875aef84878e04c1eabefdf6571514891086/src/target_properties.cpp#L40
+    static std::map<std::string, std::string> device_name_map = {
+        {"Ellesmere", "gfx803"},
+        {"Baffin", "gfx803"},
+        {"RacerX", "gfx803"},
+        {"Polaris10", "gfx803"},
+        {"Polaris11", "gfx803"},
+        {"Tonga", "gfx803"},
+        {"Fiji", "gfx803"},
+        {"gfx800", "gfx803"},
+        {"gfx802", "gfx803"},
+        {"gfx804", "gfx803"},
+        {"Vega10", "gfx900"},
+        {"gfx901", "gfx900"},
+        {"10.3.0 Sienna_Cichlid 18", "gfx1030"},
+    };
+
+    const auto name = raw_name.substr(0, raw_name.find(':')); // str.substr(0, npos) returns str.
+
+    auto match = device_name_map.find(name);
+    if(match != device_name_map.end())
+        return match->second;
+    return name;
+}
+
+inline bool is_xdl_supported()
+{
+    return ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+           ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
+           ck::get_device_name() == "gfx942";
+}
+
+} // namespace ck
diff --git a/include/ck/host_utility/hip_check_error.hpp b/include/ck/host_utility/hip_check_error.hpp
new file mode 100644
index 0000000000..af7bebd9d6
--- /dev/null
+++ b/include/ck/host_utility/hip_check_error.hpp
@@ -0,0 +1,17 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <hip/hip_runtime.h>
+
+inline void hip_check_error(hipError_t x)
+{
+    if(x != hipSuccess)
+    {
+        std::ostringstream ss;
+        ss << "HIP runtime error: " << hipGetErrorString(x) << ". " << __FILE__ << ": " << __LINE__
+           << "in function: " << __func__;
+        throw std::runtime_error(ss.str());
+    }
+}
diff --git a/include/ck/host_utility/io.hpp b/include/ck/host_utility/io.hpp
new file mode 100644
index 0000000000..55734bab2e
--- /dev/null
+++ b/include/ck/host_utility/io.hpp
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <iostream>
+#include <vector>
+#include <iterator>
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+
+template <typename T>
+std::ostream& operator<<(std::ostream& os, const std::vector<T>& v)
+{
+    std::copy(std::begin(v), std::end(v), std::ostream_iterator<T>(os, " "));
+    return os;
+}
+
+template <typename T, std::size_t N>
+std::ostream& operator<<(std::ostream& os, const std::array<T, N>& v)
+{
+    std::copy(std::begin(v), std::end(v), std::ostream_iterator<T>(os, " "));
+    return os;
+}
+
+template <typename... Ts>
+std::ostream& operator<<(std::ostream& os, const ck::TensorDescriptor<Ts...>& desc)
+{
+    constexpr ck::index_t nDim = ck::remove_cvref_t<decltype(desc)>::GetNumOfDimension();
+
+    os << "{";
+
+    ck::static_for<0, nDim - 1, 1>{}([&](auto i) { os << desc.GetLength(i) << ", "; });
+
+    os << desc.GetLength(ck::Number<nDim - 1>{});
+
+    os << "}";
+
+    return os;
+}
diff --git a/include/ck/host_utility/kernel_launch.hpp b/include/ck/host_utility/kernel_launch.hpp
new file mode 100644
index 0000000000..8e82afb351
--- /dev/null
+++ b/include/ck/host_utility/kernel_launch.hpp
@@ -0,0 +1,170 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <hip/hip_runtime.h>
+
+#include "ck/ck.hpp"
+#include "ck/stream_config.hpp"
+#include "ck/host_utility/hip_check_error.hpp"
+
+template <int MaxThreadPerBlock, int MinBlockPerCu, typename Kernel, typename... Args>
+#if CK_USE_LAUNCH_BOUNDS
+__launch_bounds__(MaxThreadPerBlock, MinBlockPerCu)
+#endif
+    __global__ void kernel_wrapper(Kernel f, Args... args)
+{
+    f(args...);
+}
+
+template <typename... Args, typename F>
+float launch_and_time_kernel(const StreamConfig& stream_config,
+                             F kernel,
+                             dim3 grid_dim,
+                             dim3 block_dim,
+                             std::size_t lds_byte,
+                             Args... args)
+{
+#if CK_TIME_KERNEL
+    if(stream_config.time_kernel_)
+    {
+#if DEBUG_LOG
+        printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n",
+               __func__,
+               grid_dim.x,
+               grid_dim.y,
+               grid_dim.z,
+               block_dim.x,
+               block_dim.y,
+               block_dim.z);
+
+        printf("Warm up 1 time\n");
+#endif
+        // warm up
+        kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
+
+        const int nrepeat = 10;
+#if DEBUG_LOG
+        printf("Start running %d times...\n", nrepeat);
+#endif
+        hipEvent_t start, stop;
+
+        hip_check_error(hipEventCreate(&start));
+        hip_check_error(hipEventCreate(&stop));
+
+        hip_check_error(hipDeviceSynchronize());
+        hip_check_error(hipEventRecord(start, stream_config.stream_id_));
+
+        for(int i = 0; i < nrepeat; ++i)
+        {
+            kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
+        }
+
+        hip_check_error(hipEventRecord(stop, stream_config.stream_id_));
+        hip_check_error(hipEventSynchronize(stop));
+
+        float total_time = 0;
+
+        hip_check_error(hipEventElapsedTime(&total_time, start, stop));
+
+        return total_time / nrepeat;
+    }
+    else
+    {
+        kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
+
+        return 0;
+    }
+#else
+    kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
+
+    return 0;
+#endif
+}
+
+template <typename... Args, typename F, typename PreProcessFunc>
+float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
+                                             PreProcessFunc preprocess,
+                                             F kernel,
+                                             dim3 grid_dim,
+                                             dim3 block_dim,
+                                             std::size_t lds_byte,
+                                             Args... args)
+{
+#if CK_TIME_KERNEL
+    if(stream_config.time_kernel_)
+    {
+#if DEBUG_LOG
+        printf("%s: grid_dim {%d, %d, %d}, block_dim {%d, %d, %d} \n",
+               __func__,
+               grid_dim.x,
+               grid_dim.y,
+               grid_dim.z,
+               block_dim.x,
+               block_dim.y,
+               block_dim.z);
+
+        printf("Warm up 1 time\n");
+#endif
+        // warm up
+        preprocess();
+        kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
+
+        const int nrepeat = 10;
+#if DEBUG_LOG
+        printf("Start running %d times...\n", nrepeat);
+#endif
+        hipEvent_t start, stop;
+
+        hip_check_error(hipEventCreate(&start));
+        hip_check_error(hipEventCreate(&stop));
+
+        hip_check_error(hipDeviceSynchronize());
+        hip_check_error(hipEventRecord(start, stream_config.stream_id_));
+
+        for(int i = 0; i < nrepeat; ++i)
+        {
+            preprocess();
+            kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
+        }
+
+        hip_check_error(hipEventRecord(stop, stream_config.stream_id_));
+        hip_check_error(hipEventSynchronize(stop));
+
+        float total_time = 0;
+
+        hip_check_error(hipEventElapsedTime(&total_time, start, stop));
+
+        return total_time / nrepeat;
+    }
+    else
+    {
+        preprocess();
+        kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
+
+        return 0;
+    }
+#else
+    kernel<<<grid_dim, block_dim, lds_byte, stream_config.stream_id_>>>(args...);
+
+    return 0;
+#endif
+}
+
+template <int MaxThreadPerBlock = CK_MAX_THREAD_PER_BLOCK,
+          int MinBlockPerCu     = CK_MIN_BLOCK_PER_CU,
+          typename KernelImpl,
+          typename... Args>
+float launch_kernel(const StreamConfig& stream_config,
+                    KernelImpl kernel_impl,
+                    dim3 grid_dim,
+                    dim3 block_dim,
+                    std::size_t dynamic_smem_byte,
+                    Args... args)
+{
+    const auto kernel = kernel_wrapper<MaxThreadPerBlock, MinBlockPerCu, KernelImpl, Args...>;
+
+    return launch_and_time_kernel(
+        stream_config, kernel, grid_dim, block_dim, dynamic_smem_byte, kernel_impl, args...);
+}
diff --git a/include/ck/host_utility/stream_utility.hpp b/include/ck/host_utility/stream_utility.hpp
new file mode 100644
index 0000000000..9ab49489bb
--- /dev/null
+++ b/include/ck/host_utility/stream_utility.hpp
@@ -0,0 +1,43 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <hip/hip_runtime.h>
+
+#include "ck/stream_config.hpp"
+#include "ck/host_utility/hip_check_error.hpp"
+
+static inline int getAvailableComputeUnitCount(const StreamConfig& stream_config)
+{
+    constexpr int MAX_MASK_DWORDS = 64;
+
+    // assume at most 64*32 = 2048 CUs
+    uint32_t cuMask[MAX_MASK_DWORDS];
+
+    for(int i = 0; i < MAX_MASK_DWORDS; i++)
+        cuMask[i] = 0;
+
+    auto countSetBits = [](uint32_t dword) {
+        int count = 0;
+
+        while(dword != 0)
+        {
+            if(dword & 0x1)
+                count++;
+
+            dword = dword >> 1;
+        };
+
+        return (count);
+    };
+
+    hip_check_error(hipExtStreamGetCUMask(stream_config.stream_id_, MAX_MASK_DWORDS, &cuMask[0]));
+
+    int ret = 0;
+
+    for(int i = 0; i < MAX_MASK_DWORDS; i++)
+        ret += countSetBits(cuMask[i]);
+
+    return (ret);
+};
diff --git a/include/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp b/include/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp
new file mode 100644
index 0000000000..6b118e972e
--- /dev/null
+++ b/include/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp
@@ -0,0 +1,153 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_TRANSFORM_FORWARD_CONVOLUTION3D_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_FORWARD_CONVOLUTION3D_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// A: in
+// B: wei
+// C: out
+// GemmM = N * Do * Ho * Wo
+// GemmN = K
+// GemmK = Z * Y * X * C
+template <typename... In,
+          typename... Wei,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value>
+__host__ __device__ constexpr auto
+transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk_pad(
+    const TensorDescriptor<In...>& in_grid_desc_n_di_hi_wi_c,
+    const TensorDescriptor<Wei...>& wei_k_z_y_x_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_do_ho_wo_k_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+    constexpr auto I4 = Number<4>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_grid_desc_n_di_hi_wi_c.GetLength(I0);
+    const auto K = out_n_do_ho_wo_k_grid_desc.GetLength(I4);
+    const auto C = in_grid_desc_n_di_hi_wi_c.GetLength(I4);
+
+    const auto Di = in_grid_desc_n_di_hi_wi_c.GetLength(I1);
+    const auto Hi = in_grid_desc_n_di_hi_wi_c.GetLength(I2);
+    const auto Wi = in_grid_desc_n_di_hi_wi_c.GetLength(I3);
+
+    const auto Do = out_n_do_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Ho = out_n_do_ho_wo_k_grid_desc.GetLength(I2);
+    const auto Wo = out_n_do_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Z = wei_k_z_y_x_c_grid_desc.GetLength(I1);
+    const auto Y = wei_k_z_y_x_c_grid_desc.GetLength(I2);
+    const auto X = wei_k_z_y_x_c_grid_desc.GetLength(I3);
+
+    const auto ConvStrideD = conv_strides[I0];
+    const auto ConvStrideH = conv_strides[I1];
+    const auto ConvStrideW = conv_strides[I2];
+
+    const auto ConvDilationD = conv_dilations[I0];
+    const auto ConvDilationH = conv_dilations[I1];
+    const auto ConvDilationW = conv_dilations[I2];
+
+    const auto InLeftPadD = in_left_pads[I0];
+    const auto InLeftPadH = in_left_pads[I1];
+    const auto InLeftPadW = in_left_pads[I2];
+
+    const auto InRightPadD = in_right_pads[I0];
+    const auto InRightPadH = in_right_pads[I1];
+    const auto InRightPadW = in_right_pads[I2];
+
+    const auto GemmM  = N * Do * Ho * Wo;
+    const auto GemmN  = K;
+    const auto GemmK  = Z * Y * X * C;
+    const auto GemmK0 = GemmK / GemmK1;
+
+    // A: input tensor
+    const auto in_grid_desc_n_dip_hip_wip_c = transform_tensor_descriptor(
+        in_grid_desc_n_di_hi_wi_c,
+        make_tuple(make_pass_through_transform(N),
+                   make_pad_transform(Di, InLeftPadD, InRightPadD),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+    const auto in_grid_desc_n_z_do_y_ho_x_wo_c = transform_tensor_descriptor(
+        in_grid_desc_n_dip_hip_wip_c,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+        make_tuple(
+            Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5, 6>{}, Sequence<7>{}));
+
+    const auto in_grid_desc_gemmk_gemmm =
+        transform_tensor_descriptor(in_grid_desc_n_z_do_y_ho_x_wo_c,
+                                    make_tuple(make_merge_transform(make_tuple(Z, Y, X, C)),
+                                               make_merge_transform(make_tuple(N, Do, Ho, Wo))),
+                                    make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto in_grid_desc_gemmk0_gemmm_gemmk1 =
+        transform_tensor_descriptor(in_grid_desc_gemmk_gemmm,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmM)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // B: weight tensor
+    const auto wei_grid_desc_gemmk_gemmn = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, Z * Y * X * C)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(Z * Y * X * C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto wei_grid_desc_gemmk0_gemmn_gemmk1 =
+        transform_tensor_descriptor(wei_grid_desc_gemmk_gemmn,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmN)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // C: output tensor
+    const auto out_grid_desc_gemmm_gemmn = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K)),
+        make_tuple(make_pass_through_transform(N * Do * Ho * Wo), make_pass_through_transform(K)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    // const auto out_grid_desc_gemmm_gemmn = transform_tensor_descriptor(
+    //     out_n_do_ho_wo_k_grid_desc,
+    //     make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+    //                make_pass_through_transform(K)),
+    //     make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<3>{}),
+    //     make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    return make_tuple(in_grid_desc_gemmk0_gemmm_gemmk1,
+                      wei_grid_desc_gemmk0_gemmn_gemmk1,
+                      out_grid_desc_gemmm_gemmn);
+}
+
+} // namespace ck
+#endif
diff --git a/include/ck/stream_config.hpp b/include/ck/stream_config.hpp
new file mode 100644
index 0000000000..505a602b24
--- /dev/null
+++ b/include/ck/stream_config.hpp
@@ -0,0 +1,14 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <hip/hip_runtime.h>
+#include <hip/hip_fp16.h>
+
+struct StreamConfig
+{
+    hipStream_t stream_id_ = nullptr;
+    bool time_kernel_      = false;
+    int log_level_         = 0;
+};
diff --git a/include/ck/tensor/static_tensor.hpp b/include/ck/tensor/static_tensor.hpp
new file mode 100644
index 0000000000..f4bcb6b3ad
--- /dev/null
+++ b/include/ck/tensor/static_tensor.hpp
@@ -0,0 +1,277 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_STATIC_TENSOR_HPP
+#define CK_STATIC_TENSOR_HPP
+
+namespace ck {
+
+// StaticTensor for Scalar
+template <AddressSpaceEnum AddressSpace,
+          typename T,
+          typename TensorDesc,
+          bool InvalidElementUseNumericalZeroValue,
+          typename enable_if<TensorDesc::IsKnownAtCompileTime(), bool>::type = false>
+struct StaticTensor
+{
+    static constexpr auto desc_                  = TensorDesc{};
+    static constexpr index_t ndim_               = TensorDesc::GetNumOfDimension();
+    static constexpr index_t element_space_size_ = desc_.GetElementSpaceSize();
+
+    __host__ __device__ constexpr StaticTensor() : invalid_element_scalar_value_{0} {}
+
+    __host__ __device__ constexpr StaticTensor(T invalid_element_value)
+        : invalid_element_scalar_value_{invalid_element_value}
+    {
+    }
+
+    // read access
+    template <typename Idx,
+              typename enable_if<is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
+                                 bool>::type = false>
+    __host__ __device__ constexpr const T& operator[](Idx) const
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        if constexpr(is_valid)
+        {
+            return data_[Number<offset>{}];
+        }
+        else
+        {
+            if constexpr(InvalidElementUseNumericalZeroValue)
+            {
+                return zero_scalar_value_;
+            }
+            else
+            {
+                return invalid_element_scalar_value_;
+            }
+        }
+    }
+
+    // write access
+    template <typename Idx,
+              typename enable_if<is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
+                                 bool>::type = false>
+    __host__ __device__ constexpr T& operator()(Idx)
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        if constexpr(is_valid)
+        {
+            return data_(Number<offset>{});
+        }
+        else
+        {
+            return ignored_element_scalar_;
+        }
+    }
+
+    StaticBuffer<AddressSpace, T, element_space_size_, true> data_;
+    static constexpr T zero_scalar_value_ = T{0};
+    // for read access of invalid element
+    const T invalid_element_scalar_value_;
+    // for write access of invalid element
+    T ignored_element_scalar_;
+};
+
+// StaticTensor for vector
+template <AddressSpaceEnum AddressSpace,
+          typename S,
+          index_t ScalarPerVector,
+          typename TensorDesc,
+          bool InvalidElementUseNumericalZeroValue,
+          typename enable_if<TensorDesc::IsKnownAtCompileTime(), bool>::type = false>
+struct StaticTensorTupleOfVectorBuffer
+{
+    static constexpr auto desc_                  = TensorDesc{};
+    static constexpr index_t ndim_               = TensorDesc::GetNumOfDimension();
+    static constexpr index_t element_space_size_ = desc_.GetElementSpaceSize();
+
+    static constexpr index_t num_of_vector_ =
+        math::integer_divide_ceil(element_space_size_, ScalarPerVector);
+
+    using V = vector_type<S, ScalarPerVector>;
+
+    __host__ __device__ constexpr StaticTensorTupleOfVectorBuffer()
+        : invalid_element_scalar_value_{0}, ignored_element_scalar_{0}
+    {
+    }
+
+    __host__ __device__ constexpr StaticTensorTupleOfVectorBuffer(S invalid_element_value)
+        : invalid_element_scalar_value_{invalid_element_value}, ignored_element_scalar_{0}
+    {
+    }
+
+    // Get S
+    // Idx is for S, not V
+    template <typename Idx,
+              typename enable_if<is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
+                                 bool>::type = false>
+    __host__ __device__ constexpr const S& operator[](Idx) const
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        if constexpr(is_valid)
+        {
+            return data_[Number<offset>{}];
+        }
+        else
+        {
+            if constexpr(InvalidElementUseNumericalZeroValue)
+            {
+                return zero_scalar_value_;
+            }
+            else
+            {
+                return invalid_element_scalar_value_;
+            }
+        }
+    }
+
+    // Set S
+    // Idx is for S, not V
+    template <typename Idx,
+              typename enable_if<is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
+                                 bool>::type = false>
+    __host__ __device__ constexpr S& operator()(Idx)
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        if constexpr(is_valid)
+        {
+            return data_(Number<offset>{});
+        }
+        else
+        {
+            return ignored_element_scalar_;
+        }
+    }
+
+    // Get X
+    // Idx is for S, not X. Idx should be aligned with X
+    template <typename X,
+              typename Idx,
+              typename enable_if<has_same_scalar_type<S, X>::value &&
+                                     is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
+                                 bool>::type = false>
+    __host__ __device__ constexpr X GetAsType(Idx) const
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        if constexpr(is_valid)
+        {
+            return data_.template GetAsType<X>(Number<offset>{});
+        }
+        else
+        {
+            if constexpr(InvalidElementUseNumericalZeroValue)
+            {
+                // TODO: is this right way to initialize a vector?
+                return X{0};
+            }
+            else
+            {
+                // TODO: is this right way to initialize a vector?
+                return X{invalid_element_scalar_value_};
+            }
+        }
+    }
+
+    // Set X
+    // Idx is for S, not X. Idx should be aligned with X
+    template <typename X,
+              typename Idx,
+              typename enable_if<has_same_scalar_type<S, X>::value &&
+                                     is_known_at_compile_time<Idx>::value && Idx::Size() == ndim_,
+                                 bool>::type = false>
+    __host__ __device__ constexpr void SetAsType(Idx, X x)
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        if constexpr(is_valid)
+        {
+            data_.template SetAsType<X>(Number<offset>{}, x);
+        }
+    }
+
+    // Get read access to V. No is_valid check
+    // Idx is for S, not V. Idx should be aligned with V
+    template <typename Idx>
+    __host__ __device__ constexpr const V& GetVectorTypeReference(Idx) const
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        return data_.GetVectorTypeReference(Number<offset>{});
+    }
+
+    // Get read access to V. No is_valid check
+    // Idx is for S, not V. Idx should be aligned with V
+    template <typename Idx>
+    __host__ __device__ constexpr V& GetVectorTypeReference(Idx)
+    {
+        constexpr auto coord = make_tensor_coordinate(desc_, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        return data_.GetVectorTypeReference(Number<offset>{});
+    }
+
+    StaticBufferTupleOfVector<AddressSpace, S, num_of_vector_, ScalarPerVector, true> data_;
+    static constexpr S zero_scalar_value_ = S{0};
+    // for read access of invalid element
+    const S invalid_element_scalar_value_ = S{0};
+    // for write access of invalid element
+    S ignored_element_scalar_;
+};
+
+template <AddressSpaceEnum AddressSpace,
+          typename T,
+          typename TensorDesc,
+          typename enable_if<TensorDesc::IsKnownAtCompileTime(), bool>::type = false>
+__host__ __device__ constexpr auto make_static_tensor(TensorDesc)
+{
+    return StaticTensor<AddressSpace, T, TensorDesc, true>{};
+}
+
+template <
+    AddressSpaceEnum AddressSpace,
+    typename T,
+    typename TensorDesc,
+    typename X,
+    typename enable_if<TensorDesc::IsKnownAtCompileTime(), bool>::type                   = false,
+    typename enable_if<is_same<remove_cvref_t<T>, remove_cvref_t<X>>::value, bool>::type = false>
+__host__ __device__ constexpr auto make_static_tensor(TensorDesc, X invalid_element_value)
+{
+    return StaticTensor<AddressSpace, T, TensorDesc, true>{invalid_element_value};
+}
+
+} // namespace ck
+#endif
diff --git a/include/ck/tensor/tensor_view.hpp b/include/ck/tensor/tensor_view.hpp
new file mode 100644
index 0000000000..b13ad19285
--- /dev/null
+++ b/include/ck/tensor/tensor_view.hpp
@@ -0,0 +1,170 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_description/tensor_coordinate.hpp"
+
+namespace ck {
+
+template <typename BufferView_, typename TensorDesc_>
+struct TensorView
+{
+    using BufferView  = remove_reference_t<BufferView_>;
+    using DataType    = typename BufferView::type;
+    using TensorDesc  = remove_cvref_t<TensorDesc_>;
+    using TensorIndex = Array<index_t, TensorDesc::GetNumOfTopDimension()>;
+    using TensorCoord = decltype(make_tensor_coordinate(TensorDesc{}, TensorIndex{}));
+
+    __host__ __device__ constexpr TensorView() = default;
+
+    __host__ __device__ constexpr TensorView(const BufferView& buffer_view, const TensorDesc& desc)
+        : buf_{buffer_view}, desc_{desc}
+    {
+    }
+
+    __host__ __device__ constexpr auto& GetTensorDescriptor() const { return desc_; }
+
+    __host__ __device__ static constexpr index_t GetNumOfDimension()
+    {
+        return TensorDesc::GetNumOfTopDimension();
+    }
+
+    __host__ __device__ constexpr const auto& GetBufferView() const { return buf_; }
+
+    __host__ __device__ constexpr auto& GetBufferView() { return buf_; }
+
+    __host__ __device__ constexpr DataType GetElement(const TensorCoord& coord) const
+    {
+        return buf_.template Get<DataType>(
+            coord.GetOffset(),
+            coordinate_has_valid_offset_assuming_top_index_is_valid(desc_, coord));
+    }
+
+    __host__ __device__ constexpr void SetElement(const TensorCoord& coord, const DataType& x)
+    {
+        buf_.template Set<DataType>(
+            coord.GetOffset(),
+            coordinate_has_valid_offset_assuming_top_index_is_valid(desc_, coord),
+            x);
+    }
+
+    // X is vector of DataType.
+    // "coord" is coordinate of DataType, not X. "coord" should be aligned to X
+    template <typename X,
+              typename enable_if<is_same_v<typename scalar_type<remove_cvref_t<X>>::type,
+                                           typename scalar_type<remove_cvref_t<DataType>>::type>,
+                                 bool>::type = false>
+    __host__ __device__ constexpr remove_cvref_t<X>
+    GetVectorizedElements(const TensorCoord& coord) const
+    {
+        return buf_.template Get<X>(
+            coord.GetOffset(),
+            coordinate_has_valid_offset_assuming_top_index_is_valid(desc_, coord));
+    }
+
+    // X is vector of DataType.
+    // "coord" is coordinate of DataType, not X. "coord" should be aligned to X
+    template <typename X,
+              typename enable_if<is_same_v<typename scalar_type<remove_cvref_t<X>>::type,
+                                           typename scalar_type<remove_cvref_t<DataType>>::type>,
+                                 bool>::type = false>
+    __host__ __device__ constexpr void SetVectorizedElements(const TensorCoord& coord, const X& x)
+    {
+        buf_.template Set<X>(coord.GetOffset(),
+                             coordinate_has_valid_offset_assuming_top_index_is_valid(desc_, coord),
+                             x);
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("TensorView{");
+
+        // buf_
+        printf("buf_: ");
+        print(buf_);
+        printf(", ");
+
+        // desc_
+        printf("desc_: ");
+        print(desc_);
+
+        printf("}");
+    }
+
+    // member
+    BufferView buf_;
+    TensorDesc desc_;
+};
+
+template <AddressSpaceEnum BufferAddressSpace = AddressSpaceEnum::Generic,
+          typename DataType,
+          typename... Ts>
+__host__ __device__ constexpr auto make_tensor_view(DataType* p,
+                                                    const TensorDescriptor<Ts...>& desc)
+{
+    auto buffer_view = make_buffer_view<BufferAddressSpace>(p, desc.GetElementSpaceSize());
+
+    return TensorView<decltype(buffer_view), decltype(desc)>{buffer_view, desc};
+}
+
+template <AddressSpaceEnum BufferAddressSpace = AddressSpaceEnum::Generic,
+          typename DataType,
+          typename... Lengths,
+          typename... Strides,
+          index_t GuaranteedLastDimensionVectorLength                              = -1,
+          index_t GuaranteedLastDimensionVectorStride                              = -1,
+          typename enable_if<sizeof...(Lengths) == sizeof...(Strides), bool>::type = false>
+__host__ __device__ constexpr auto
+make_naive_tensor_view(DataType* p,
+                       const Tuple<Lengths...>& lengths,
+                       const Tuple<Strides...>& strides,
+                       Number<GuaranteedLastDimensionVectorLength> = Number<-1>{},
+                       Number<GuaranteedLastDimensionVectorStride> = Number<-1>{})
+{
+    auto desc = make_naive_tensor_descriptor(lengths,
+                                             strides,
+                                             Number<GuaranteedLastDimensionVectorLength>{},
+                                             Number<GuaranteedLastDimensionVectorStride>{});
+
+    auto buffer_view = make_buffer_view<BufferAddressSpace>(p, desc.GetElementSpaceSize());
+
+    return TensorView<decltype(buffer_view), decltype(desc)>{buffer_view, desc};
+}
+
+template <AddressSpaceEnum BufferAddressSpace = AddressSpaceEnum::Generic,
+          typename DataType,
+          typename... Lengths,
+          index_t GuaranteedLastDimensionVectorLength = -1>
+__host__ __device__ constexpr auto
+make_naive_tensor_view_packed(DataType* p,
+                              const Tuple<Lengths...>& lengths,
+                              Number<GuaranteedLastDimensionVectorLength> = Number<-1>{})
+{
+    auto desc =
+        make_naive_tensor_descriptor_packed(lengths, Number<GuaranteedLastDimensionVectorLength>{});
+
+    auto buffer_view = make_buffer_view<BufferAddressSpace>(p, desc.GetElementSpaceSize());
+
+    return TensorView<decltype(buffer_view), decltype(desc)>{buffer_view, desc};
+}
+
+template <typename OldTensorView,
+          typename NewTransforms,
+          typename NewLowerDimensionOldVisibleIdss,
+          typename NewUpperDimensionNewVisibleIdss>
+__host__ __device__ constexpr auto transform_tensor_view(const OldTensorView& old_tensor_view,
+                                                         const NewTransforms& new_transforms,
+                                                         NewLowerDimensionOldVisibleIdss,
+                                                         NewUpperDimensionNewVisibleIdss)
+{
+    auto new_desc = transform_tensor_descriptor(old_tensor_view.desc_,
+                                                new_transforms,
+                                                NewLowerDimensionOldVisibleIdss{},
+                                                NewUpperDimensionNewVisibleIdss{});
+
+    return TensorView<typename OldTensorView::BufferView, remove_cvref_t<decltype(new_desc)>>{
+        old_tensor_view.buf_, new_desc};
+}
+
+} // namespace ck
diff --git a/include/ck/tensor/thread_private_tensor.hpp b/include/ck/tensor/thread_private_tensor.hpp
new file mode 100644
index 0000000000..9a85635ba8
--- /dev/null
+++ b/include/ck/tensor/thread_private_tensor.hpp
@@ -0,0 +1,128 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/buffer_view.hpp"
+
+namespace ck {
+
+template <typename DataType, typename TensorDescriptor>
+struct ThreadPrivateTensor
+{
+    using DataType_         = DataType;
+    using T                 = DataType_;
+    using TensorDescriptor_ = remove_cvref_t<TensorDescriptor>;
+
+    __host__ __device__ constexpr ThreadPrivateTensor()
+        // FIXME: remove static_cast
+        : desc_{}, buf_view_{p_data_, static_cast<index_t>(desc_.GetElementSpaceSize())}
+    {
+    }
+
+    __host__ __device__ constexpr ThreadPrivateTensor(const TensorDescriptor_& desc)
+        // FIXME: remove static_cast
+        : desc_{desc}, buf_view_{p_data_, static_cast<index_t>(desc_.GetElementSpaceSize())}
+    {
+    }
+
+    // member
+    TensorDescriptor_ desc_;
+
+    //
+    static constexpr index_t kMaxBufferSize_ = 32;
+    DataType_ p_data_[kMaxBufferSize_];
+
+    // FIXME: remove assumption that type of BufferSize should be index_t
+    BufferView<AddressSpaceEnum::Vgpr, DataType_, index_t, true> buf_view_;
+
+    // function
+    // FIXME: doesn't do is_valid check
+    template <typename Idx>
+    __device__ constexpr const T& operator[](const Idx& idx) const
+    {
+#if 1 // debug
+      // FIXME: remove to_multi_index
+        const auto coord = make_tensor_coordinate(desc_, to_multi_index(idx));
+
+        const index_t offset = coord.GetOffset();
+
+        return buf_view_[offset];
+#else
+        constexpr auto coord = make_tensor_coordinate(TensorDescriptor_{}, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        return buf_view_[offset];
+#endif
+    }
+
+    // FIXME: doesn't do is_valid check
+    template <typename Idx>
+    __device__ constexpr T& operator()(const Idx& idx)
+    {
+#if 1 // debug
+      // FIXME: remove to_multi_index
+        const auto coord = make_tensor_coordinate(desc_, to_multi_index(idx));
+
+        const index_t offset = coord.GetOffset();
+
+        return buf_view_(offset);
+#else
+        constexpr auto coord = make_tensor_coordinate(TensorDescriptor_{}, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        return buf_view_(offset);
+#endif
+    }
+
+    // idx is the index of T, not X. idx should be aligned to X
+    template <typename X, typename Idx>
+    __device__ constexpr X Get(const Idx& idx) const
+    {
+#if 1 // debug
+        const auto coord = make_tensor_coordinate(desc_, idx);
+
+        const index_t offset = coord.GetOffset();
+
+        const bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        return buf_view_.template Get<X>(offset, is_valid);
+#else
+        constexpr auto coord = make_tensor_coordinate(TensorDescriptor_{}, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(TensorDescriptor_{}, coord);
+
+        return buf_view_.template Get<X>(offset, is_valid);
+#endif
+    }
+
+    // idx is the index of T, not X. idx should be aligned to X
+    template <typename X, typename Idx>
+    __device__ void Set(const Idx& idx, const X& x)
+    {
+
+#if 1 // debug
+        const auto coord = make_tensor_coordinate(desc_, idx);
+
+        const index_t offset = coord.GetOffset();
+
+        const bool is_valid = coordinate_has_valid_offset(desc_, coord);
+
+        buf_view_.template Set<X>(offset, is_valid, x);
+#else
+        constexpr auto coord = make_tensor_coordinate(TensorDescriptor_{}, to_multi_index(Idx{}));
+
+        constexpr index_t offset = coord.GetOffset();
+
+        constexpr bool is_valid = coordinate_has_valid_offset(TensorDescriptor_{}, coord);
+
+        buf_view_.template Set<X>(offset, is_valid, x);
+#endif
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_description/cluster_descriptor.hpp b/include/ck/tensor_description/cluster_descriptor.hpp
new file mode 100644
index 0000000000..2dfcad8e04
--- /dev/null
+++ b/include/ck/tensor_description/cluster_descriptor.hpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+namespace ck {
+
+template <typename Lengths,
+          typename ArrangeOrder = typename arithmetic_sequence_gen<0, Lengths::Size(), 1>::type>
+__host__ __device__ constexpr auto make_cluster_descriptor(
+    const Lengths& lengths,
+    ArrangeOrder order = typename arithmetic_sequence_gen<0, Lengths::Size(), 1>::type{})
+{
+    constexpr index_t ndim_low = Lengths::Size();
+
+    const auto reordered_lengths = container_reorder_given_new2old(lengths, order);
+
+    const auto low_lengths = generate_tuple(
+        [&](auto idim_low) { return reordered_lengths[idim_low]; }, Number<ndim_low>{});
+
+    const auto transform = make_merge_transform(low_lengths);
+
+    constexpr auto low_dim_old_top_ids = ArrangeOrder{};
+
+    constexpr auto up_dim_new_top_ids = Sequence<0>{};
+
+    return make_single_stage_tensor_adaptor(
+        make_tuple(transform), make_tuple(low_dim_old_top_ids), make_tuple(up_dim_new_top_ids));
+}
+
+} // namespace ck
diff --git a/include/ck/tensor_description/macro_func_tensor_adaptor_from_encoding.hpp b/include/ck/tensor_description/macro_func_tensor_adaptor_from_encoding.hpp
new file mode 100644
index 0000000000..7a92932a05
--- /dev/null
+++ b/include/ck/tensor_description/macro_func_tensor_adaptor_from_encoding.hpp
@@ -0,0 +1,255 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+// Macro function
+// construct constexpr TensorAdaptor from constexpr encoding
+// encoded_tensor_adaptor are Tuple of following objects:
+//    1. encoded transforms (Array of fixed size). Each encoded transform is a Tuple of following:
+//           1.1 name (IndexTransformEnum)
+//           1.2 meta data for constructor of the transform
+//           1.3 num of lower dimension (index_t)
+//           1.4 lower dimension Ids (Array of fixed size)
+//           1.5 num of up dimension (index_t)
+//           1.6 upper dimension Ids (Array of fixed size)
+//    2. num of transforms (index_t)
+//    3. encoded bottom dimension Ids (Array of fixed size)
+//    4. num of bottom dimension (index_t)
+//    5. encoded top dimension Ids (Array of fixed size)
+//    6. num of top dimension (index_t)
+#define CONSTRUCT_TENSOR_ADAPTOR_FROM_ENCODING(encoded_tensor_adaptor)                            \
+    [encoded_tensor_adaptor]() {                                                                  \
+        using namespace ck;                                                                       \
+                                                                                                  \
+        constexpr auto encoded_transforms  = encoded_tensor_adaptor.template At<0>();             \
+        constexpr index_t num_transform    = encoded_tensor_adaptor.template At<1>();             \
+        constexpr auto encoded_bottom_dims = encoded_tensor_adaptor.template At<2>();             \
+        constexpr index_t num_bottom_dim   = encoded_tensor_adaptor.template At<3>();             \
+        constexpr auto encoded_top_dims    = encoded_tensor_adaptor.template At<4>();             \
+        constexpr index_t num_top_dim      = encoded_tensor_adaptor.template At<5>();             \
+                                                                                                  \
+        constexpr auto trans = [&encoded_transforms, &num_transform]() {                          \
+            return generate_tuple(                                                                \
+                [&encoded_transforms](auto i) constexpr {                                         \
+                    constexpr auto name        = encoded_transforms[i].template At<0>();          \
+                    constexpr auto meta_data   = encoded_transforms[i].template At<1>();          \
+                    constexpr auto num_low_dim = encoded_transforms[i].template At<2>();          \
+                    constexpr auto num_up_dim  = encoded_transforms[i].template At<4>();          \
+                                                                                                  \
+                    STATIC_ASSERT(name == IndexTransformEnum::PassThrough ||                      \
+                                      name == IndexTransformEnum::Pad ||                          \
+                                      name == IndexTransformEnum::Embed ||                        \
+                                      name == IndexTransformEnum::Merge ||                        \
+                                      name == IndexTransformEnum::UnMerge,                        \
+                                  "");                                                            \
+                                                                                                  \
+                    if constexpr(name == IndexTransformEnum::PassThrough)                         \
+                    {                                                                             \
+                        index_t pos  = 0;                                                         \
+                        auto low_len = meta_data.template Pop<index_t>(pos);                      \
+                                                                                                  \
+                        return make_pass_through_transform(low_len);                              \
+                    }                                                                             \
+                    else if constexpr(name == IndexTransformEnum::Pad)                            \
+                    {                                                                             \
+                        index_t pos    = 0;                                                       \
+                        auto low_len   = meta_data.template Pop<index_t>(pos);                    \
+                        auto left_pad  = meta_data.template Pop<index_t>(pos);                    \
+                        auto right_pad = meta_data.template Pop<index_t>(pos);                    \
+                                                                                                  \
+                        return make_pad_transform(low_len, left_pad, right_pad);                  \
+                    }                                                                             \
+                    else if constexpr(name == IndexTransformEnum::Embed)                          \
+                    {                                                                             \
+                        index_t pos  = 0;                                                         \
+                        auto up_lens = meta_data.template Pop<Array<index_t, num_up_dim>>(pos);   \
+                        auto coefficients =                                                       \
+                            meta_data.template Pop<Array<index_t, num_up_dim>>(pos);              \
+                                                                                                  \
+                        return make_embed_transform(up_lens, coefficients);                       \
+                    }                                                                             \
+                    else if constexpr(name == IndexTransformEnum::Merge)                          \
+                    {                                                                             \
+                        index_t pos   = 0;                                                        \
+                        auto low_lens = meta_data.template Pop<Array<index_t, num_low_dim>>(pos); \
+                                                                                                  \
+                        return make_merge_transform(low_lens);                                    \
+                    }                                                                             \
+                    else if constexpr(name == IndexTransformEnum::UnMerge)                        \
+                    {                                                                             \
+                        index_t pos  = 0;                                                         \
+                        auto up_lens = meta_data.template Pop<Array<index_t, num_up_dim>>(pos);   \
+                                                                                                  \
+                        return make_unmerge_transform(up_lens);                                   \
+                    }                                                                             \
+                    else if constexpr(name == IndexTransformEnum::Replicate)                      \
+                    {                                                                             \
+                        index_t pos  = 0;                                                         \
+                        auto up_lens = meta_data.template Pop<Array<index_t, num_up_dim>>(pos);   \
+                                                                                                  \
+                        return make_replicate_transform(up_lens);                                 \
+                    }                                                                             \
+                },                                                                                \
+                Number<num_transform>{});                                                         \
+        }();                                                                                      \
+                                                                                                  \
+        constexpr auto low_dim_idss = [&encoded_transforms, &num_transform]() {                   \
+            return generate_tuple(                                                                \
+                [&encoded_transforms](auto i) {                                                   \
+                    constexpr auto num_low_dim = encoded_transforms[i].template At<2>();          \
+                    constexpr auto low_dims    = encoded_transforms[i].template At<3>();          \
+                                                                                                  \
+                    return TO_SEQUENCE(low_dims, num_low_dim);                                    \
+                },                                                                                \
+                Number<num_transform>());                                                         \
+        }();                                                                                      \
+                                                                                                  \
+        constexpr auto up_dim_idss = [&encoded_transforms, &num_transform] {                      \
+            return generate_tuple(                                                                \
+                [&encoded_transforms](auto i) {                                                   \
+                    constexpr auto num_up_dim = encoded_transforms[i].template At<4>();           \
+                    constexpr auto up_dims    = encoded_transforms[i].template At<5>();           \
+                                                                                                  \
+                    return TO_SEQUENCE(up_dims, num_up_dim);                                      \
+                },                                                                                \
+                Number<num_transform>());                                                         \
+        }();                                                                                      \
+                                                                                                  \
+        constexpr auto bottom_dim_ids = TO_SEQUENCE(encoded_bottom_dims, num_bottom_dim);         \
+        constexpr auto top_dim_ids    = TO_SEQUENCE(encoded_top_dims, num_top_dim);               \
+                                                                                                  \
+        return TensorAdaptor<remove_cvref_t<decltype(trans)>,                                     \
+                             remove_cvref_t<decltype(low_dim_idss)>,                              \
+                             remove_cvref_t<decltype(up_dim_idss)>,                               \
+                             remove_cvref_t<decltype(bottom_dim_ids)>,                            \
+                             remove_cvref_t<decltype(top_dim_ids)>>{trans};                       \
+    }()
+
+// Macro function
+// construct static TensorAdaptor from constexpr encoding
+// encoded_tensor_adaptor are Tuple of following objects:
+//    1. encoded transforms (Array of fixed size). Each encoded transform is a Tuple of following:
+//           1.1 name (IndexTransformEnum)
+//           1.2 meta data for constructor of the transform
+//           1.3 num of lower dimension (index_t)
+//           1.4 lower dimension Ids (Array of fixed size)
+//           1.5 num of up dimension (index_t)
+//           1.6 upper dimension Ids (Array of fixed size)
+//    2. num of transforms (index_t)
+//    3. encoded bottom dimension Ids (Array of fixed size)
+//    4. num of bottom dimension (index_t)
+//    5. encoded top dimension Ids (Array of fixed size)
+//    6. num of top dimension (index_t)
+#define CONSTRUCT_STATIC_TENSOR_ADAPTOR_FROM_ENCODING(encoded_tensor_adaptor)                      \
+    [encoded_tensor_adaptor]() {                                                                   \
+        using namespace ck;                                                                        \
+                                                                                                   \
+        constexpr auto encoded_transforms  = encoded_tensor_adaptor.template At<0>();              \
+        constexpr index_t num_transform    = encoded_tensor_adaptor.template At<1>();              \
+        constexpr auto encoded_bottom_dims = encoded_tensor_adaptor.template At<2>();              \
+        constexpr index_t num_bottom_dim   = encoded_tensor_adaptor.template At<3>();              \
+        constexpr auto encoded_top_dims    = encoded_tensor_adaptor.template At<4>();              \
+        constexpr index_t num_top_dim      = encoded_tensor_adaptor.template At<5>();              \
+                                                                                                   \
+        constexpr auto trans = [&encoded_transforms, &num_transform]() {                           \
+            return generate_tuple(                                                                 \
+                [&encoded_transforms](auto i) constexpr {                                          \
+                    constexpr auto name        = encoded_transforms[i].template At<0>();           \
+                    constexpr auto meta_data   = encoded_transforms[i].template At<1>();           \
+                    constexpr auto num_low_dim = encoded_transforms[i].template At<2>();           \
+                    constexpr auto num_up_dim  = encoded_transforms[i].template At<4>();           \
+                                                                                                   \
+                    STATIC_ASSERT(name == IndexTransformEnum::PassThrough ||                       \
+                                      name == IndexTransformEnum::Pad ||                           \
+                                      name == IndexTransformEnum::Embed ||                         \
+                                      name == IndexTransformEnum::Merge ||                         \
+                                      name == IndexTransformEnum::UnMerge,                         \
+                                  "");                                                             \
+                                                                                                   \
+                    if constexpr(name == IndexTransformEnum::PassThrough)                          \
+                    {                                                                              \
+                        constexpr index_t low_len = meta_data.template Get<index_t>(0);            \
+                                                                                                   \
+                        return make_pass_through_transform(Number<low_len>{});                     \
+                    }                                                                              \
+                    else if constexpr(name == IndexTransformEnum::Pad)                             \
+                    {                                                                              \
+                        constexpr index_t low_len = meta_data.template Get<index_t>(0);            \
+                                                                                                   \
+                        constexpr index_t left_pad =                                               \
+                            meta_data.template Get<index_t>(sizeof(low_len));                      \
+                                                                                                   \
+                        constexpr index_t right_pad =                                              \
+                            meta_data.template Pop<index_t>(sizeof(low_len) + sizeof(left_pad));   \
+                                                                                                   \
+                        return make_pad_transform(                                                 \
+                            Number<low_len>{}, Number<left_pad>{}, Number<right_pad>{});           \
+                    }                                                                              \
+                    else if constexpr(name == IndexTransformEnum::Embed)                           \
+                    {                                                                              \
+                        constexpr auto up_lens =                                                   \
+                            meta_data.template Get<Array<index_t, num_up_dim>>(0);                 \
+                                                                                                   \
+                        constexpr auto coefficients =                                              \
+                            meta_data.template Get<Array<index_t, num_up_dim>>(sizeof(up_lens));   \
+                                                                                                   \
+                        return make_embed_transform(TO_TUPLE_OF_NUMBER(up_lens, num_up_dim),       \
+                                                    TO_TUPLE_OF_NUMBER(coefficients, num_up_dim)); \
+                    }                                                                              \
+                    else if constexpr(name == IndexTransformEnum::Merge)                           \
+                    {                                                                              \
+                        constexpr auto low_lens =                                                  \
+                            meta_data.template Get<Array<index_t, num_low_dim>>(0);                \
+                                                                                                   \
+                        return make_merge_transform(TO_TUPLE_OF_NUMBER(low_lens, num_low_dim));    \
+                    }                                                                              \
+                    else if constexpr(name == IndexTransformEnum::UnMerge)                         \
+                    {                                                                              \
+                        constexpr auto up_lens =                                                   \
+                            meta_data.template Get<Array<index_t, num_up_dim>>(0);                 \
+                                                                                                   \
+                        return make_unmerge_transform(TO_TUPLE_OF_NUMBER(up_lens, num_up_dim));    \
+                    }                                                                              \
+                    else if constexpr(name == IndexTransformEnum::Replicate)                       \
+                    {                                                                              \
+                        constexpr auto up_lens =                                                   \
+                            meta_data.template Get<Array<index_t, num_up_dim>>(0);                 \
+                                                                                                   \
+                        return make_replicate_transform(TO_TUPLE_OF_NUMBER(up_lens, num_up_dim));  \
+                    }                                                                              \
+                },                                                                                 \
+                Number<num_transform>{});                                                          \
+        }();                                                                                       \
+                                                                                                   \
+        constexpr auto low_dim_idss = [&encoded_transforms, &num_transform]() {                    \
+            return generate_tuple(                                                                 \
+                [&encoded_transforms](auto i) {                                                    \
+                    constexpr auto num_low_dim = encoded_transforms[i].template At<2>();           \
+                    constexpr auto low_dims    = encoded_transforms[i].template At<3>();           \
+                                                                                                   \
+                    return TO_SEQUENCE(low_dims, num_low_dim);                                     \
+                },                                                                                 \
+                Number<num_transform>());                                                          \
+        }();                                                                                       \
+                                                                                                   \
+        constexpr auto up_dim_idss = [&encoded_transforms, &num_transform] {                       \
+            return generate_tuple(                                                                 \
+                [&encoded_transforms](auto i) {                                                    \
+                    constexpr auto num_up_dim = encoded_transforms[i].template At<4>();            \
+                    constexpr auto up_dims    = encoded_transforms[i].template At<5>();            \
+                                                                                                   \
+                    return TO_SEQUENCE(up_dims, num_up_dim);                                       \
+                },                                                                                 \
+                Number<num_transform>());                                                          \
+        }();                                                                                       \
+                                                                                                   \
+        constexpr auto bottom_dim_ids = TO_SEQUENCE(encoded_bottom_dims, num_bottom_dim);          \
+        constexpr auto top_dim_ids    = TO_SEQUENCE(encoded_top_dims, num_top_dim);                \
+                                                                                                   \
+        return TensorAdaptor<remove_cvref_t<decltype(trans)>,                                      \
+                             remove_cvref_t<decltype(low_dim_idss)>,                               \
+                             remove_cvref_t<decltype(up_dim_idss)>,                                \
+                             remove_cvref_t<decltype(bottom_dim_ids)>,                             \
+                             remove_cvref_t<decltype(top_dim_ids)>>{trans};                        \
+    }()
diff --git a/include/ck/tensor_description/multi_index_transform.hpp b/include/ck/tensor_description/multi_index_transform.hpp
new file mode 100644
index 0000000000..aa3e92f51e
--- /dev/null
+++ b/include/ck/tensor_description/multi_index_transform.hpp
@@ -0,0 +1,1337 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/multi_index.hpp"
+
+namespace ck {
+
+enum struct IndexTransformEnum
+{
+    Undefined,
+    PassThrough,
+    Pad,
+    Embed,
+    Merge,
+    UnMerge,
+    Replicate,
+    Xor,
+};
+
+template <index_t NDimLow, index_t NDimUp>
+struct BaseTransform
+{
+    __host__ __device__ static constexpr auto GetTypeEnum()
+    {
+        return IndexTransformEnum::Undefined;
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return NDimLow; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return NDimUp; }
+
+    // return safe value for vector length/stride, based on compile-time known only
+    // variables
+    // MUST be static function
+    template <typename LowVectorLengths, typename LowVectorStrides>
+    __host__ __device__ static constexpr auto
+    CalculateUpperDimensionSafeVectorLengthStrides(const LowVectorLengths&, const LowVectorStrides&)
+    {
+        if constexpr(NDimUp > 0)
+        {
+            Array<index_t, NDimUp> up_vector_lengths{-1};
+            Array<index_t, NDimUp> up_vector_strides{-1};
+
+            return make_tuple(up_vector_lengths, up_vector_strides);
+        }
+        else
+        {
+            return make_tuple(Array<index_t, 0>{}, Array<index_t, 0>{});
+        }
+    }
+};
+
+template <typename LowLength>
+struct PassThrough : public BaseTransform<1, 1>
+{
+    static constexpr auto type_enum = IndexTransformEnum::PassThrough;
+
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<1>;
+
+    using UpLengths = decltype(make_tuple(LowLength{}));
+
+    UpLengths up_lengths_;
+
+    __host__ __device__ constexpr PassThrough() = default;
+
+    __host__ __device__ constexpr PassThrough(const LowLength& low_length)
+        : up_lengths_{make_tuple(low_length)}
+    {
+    }
+
+    __host__ __device__ static constexpr auto GetTypeEnum()
+    {
+        return IndexTransformEnum::PassThrough;
+    }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ static constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                                  const UpIdx& idx_up)
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = idx_up[Number<0>{}];
+    }
+
+    template <typename LowIdxDiff, typename UpIdxDiff, typename LowIdx, typename UpIdx>
+    __host__ __device__ static void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                                     const UpIdxDiff& idx_diff_up,
+                                                     LowIdx& idx_low,
+                                                     const UpIdx&)
+    {
+        static_assert(LowIdxDiff::Size() == 1 && UpIdxDiff::Size() == 1 && LowIdx::Size() == 1 &&
+                          UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0 = Number<0>{};
+
+        idx_diff_low(I0) = idx_diff_up[I0];
+
+        idx_low += idx_diff_low;
+    }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value;
+    }
+
+    // MUST be static function
+    template <typename LowVectorLengths, typename LowVectorStrides>
+    __host__ __device__ static constexpr auto
+    CalculateUpperDimensionSafeVectorLengthStrides(const LowVectorLengths& low_vector_lengths,
+                                                   const LowVectorStrides& low_vector_strides)
+    {
+        return make_tuple(low_vector_lengths, low_vector_strides);
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("PassThrough{");
+
+        //
+        printf("up_lengths_:");
+        print(up_lengths_);
+
+        //
+        printf("}");
+    }
+};
+
+template <typename LowLength,
+          typename LeftPadLength,
+          typename RightPadLength,
+          bool SkipIsValidCheck = false>
+struct Pad : public BaseTransform<1, 1>
+{
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<1>;
+
+    using UpLengths = decltype(make_tuple(LowLength{} + LeftPadLength{} + RightPadLength{}));
+
+    UpLengths up_lengths_;
+    LeftPadLength left_pad_length_;
+    RightPadLength right_pad_length_;
+
+    __host__ __device__ constexpr Pad() : up_lengths_{}, left_pad_length_{}, right_pad_length_{} {}
+
+    __host__ __device__ constexpr Pad(const LowLength& low_length,
+                                      const LeftPadLength& left_pad_length,
+                                      const RightPadLength& right_pad_length)
+        : up_lengths_{make_tuple(low_length + left_pad_length + right_pad_length)},
+          left_pad_length_{left_pad_length},
+          right_pad_length_{right_pad_length}
+    {
+    }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = idx_up[Number<0>{}] - left_pad_length_;
+    }
+
+    template <typename LowIdxDiff, typename UpIdxDiff, typename LowIdx, typename UpIdx>
+    __host__ __device__ static void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                                     const UpIdxDiff& idx_diff_up,
+                                                     LowIdx& idx_low,
+                                                     const UpIdx&)
+    {
+        static_assert(LowIdxDiff::Size() == 1 && UpIdxDiff::Size() == 1 && LowIdx::Size() == 1 &&
+                          UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0 = Number<0>{};
+
+        idx_diff_low(I0) = idx_diff_up[I0];
+
+        idx_low += idx_diff_low;
+    }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return SkipIsValidCheck;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& idx_up) const
+    {
+        return SkipIsValidCheck ||
+               ((idx_up[Number<0>{}] >= left_pad_length_) &&
+                (idx_up[Number<0>{}] < up_lengths_[Number<0>{}] - right_pad_length_));
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value &&
+               is_known_at_compile_time<LeftPadLength>::value &&
+               is_known_at_compile_time<RightPadLength>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("Pad{");
+
+        //
+        printf("up_lengths_: ");
+        print(up_lengths_);
+        printf(", ");
+
+        //
+        printf("left_pad_length_: ");
+        print(left_pad_length_);
+        printf(", ");
+
+        //
+        printf("right_pad_length_: ");
+        print(right_pad_length_);
+
+        printf("}");
+    }
+};
+
+template <typename LowLength, typename LeftPadLength, bool SkipIsValidCheck = false>
+struct LeftPad
+{
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<1>;
+
+    using UpLengths = decltype(make_tuple(LowLength{} + LeftPadLength{}));
+
+    UpLengths up_lengths_;
+    LeftPadLength left_pad_length_;
+
+    __host__ __device__ constexpr LeftPad() = default;
+
+    __host__ __device__ constexpr LeftPad(const LowLength& low_length,
+                                          const LeftPadLength& left_pad_length)
+        : up_lengths_{make_tuple(low_length + left_pad_length)}, left_pad_length_{left_pad_length}
+    {
+    }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = idx_up[Number<0>{}] - left_pad_length_;
+    }
+
+    template <typename LowIdxDiff, typename UpIdxDiff, typename LowIdx, typename UpIdx>
+    __host__ __device__ static void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                                     const UpIdxDiff& idx_diff_up,
+                                                     LowIdx& idx_low,
+                                                     const UpIdx&)
+    {
+        static_assert(LowIdxDiff::Size() == 1 && UpIdxDiff::Size() == 1 && LowIdx::Size() == 1 &&
+                          UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0 = Number<0>{};
+
+        idx_diff_low(I0) = idx_diff_up[I0];
+
+        idx_low += idx_diff_low;
+    }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return SkipIsValidCheck;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& idx_up) const
+    {
+        return SkipIsValidCheck || (idx_up[Number<0>{}] >= left_pad_length_);
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value &&
+               is_known_at_compile_time<LeftPadLength>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("LeftPad{");
+
+        //
+        printf("up_lengths_: ");
+        print(up_lengths_);
+        printf(", ");
+
+        //
+        printf("left_pad_length_: ");
+        print(left_pad_length_);
+
+        printf("}");
+    }
+};
+
+template <typename LowLength, typename RightPadLength, bool SkipIsValidCheck = false>
+struct RightPad : public BaseTransform<1, 1>
+{
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<1>;
+
+    using UpLengths = decltype(make_tuple(LowLength{} + RightPadLength{}));
+
+    UpLengths up_lengths_;
+    LowLength low_length_;
+    RightPadLength right_pad_length_;
+
+    __host__ __device__ constexpr RightPad() = default;
+
+    __host__ __device__ constexpr RightPad(const LowLength& low_length,
+                                           const RightPadLength& right_pad_length)
+        : up_lengths_{make_tuple(low_length + right_pad_length)},
+          low_length_{low_length},
+          right_pad_length_{right_pad_length}
+    {
+    }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ static constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                                  const UpIdx& idx_up)
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = idx_up[Number<0>{}];
+    }
+
+    template <typename LowIdxDiff, typename UpIdxDiff, typename LowIdx, typename UpIdx>
+    __host__ __device__ static void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                                     const UpIdxDiff& idx_diff_up,
+                                                     LowIdx& idx_low,
+                                                     const UpIdx&)
+    {
+        static_assert(LowIdxDiff::Size() == 1 && UpIdxDiff::Size() == 1 && LowIdx::Size() == 1 &&
+                          UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0 = Number<0>{};
+
+        idx_diff_low(I0) = idx_diff_up[I0];
+
+        idx_low += idx_diff_low;
+    }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return SkipIsValidCheck;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& idx_up) const
+    {
+        return SkipIsValidCheck || (idx_up[Number<0>{}] < low_length_);
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value &&
+               is_known_at_compile_time<LowLength>::value &&
+               is_known_at_compile_time<RightPadLength>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("LeftPad{");
+
+        //
+        printf("up_lengths_: ");
+        print(up_lengths_);
+        printf(", ");
+
+        //
+        printf("right_pad_length_: ");
+        print(right_pad_length_);
+
+        printf("}");
+    }
+};
+
+// idx_low = coefficients[0, ...nDimUp-1] * idx_up[0, ...nDimUp-1]
+// UpLengths and Coefficients can be either of the followings:
+//   1) Tuple of index_t, which is known at run-time, or
+//   2) Tuple of Number, which is known at compile-time, or
+//   3) Tuple of mixture of index_t and Number, which is known partially at run-time and partially
+//   at compile-time
+template <typename UpLengths,
+          typename Coefficients,
+          typename enable_if<UpLengths::Size() == Coefficients::Size(), bool>::type = false>
+struct Embed : public BaseTransform<1, UpLengths::Size()>
+{
+    static constexpr index_t NDimUp = UpLengths::Size();
+
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<NDimUp>;
+
+    UpLengths up_lengths_;
+    Coefficients coefficients_;
+
+    __host__ __device__ constexpr Embed() = default;
+
+    __host__ __device__ constexpr Embed(const UpLengths& up_lengths,
+                                        const Coefficients& coefficients)
+        : up_lengths_{up_lengths}, coefficients_{coefficients}
+    {
+    }
+
+    __host__ __device__ static constexpr auto GetTypeEnum() { return IndexTransformEnum::Embed; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == NDimUp,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = 0;
+
+        static_for<0, NDimUp, 1>{}([&idx_low, &idx_up, this](auto i) {
+            idx_low(Number<0>{}) += idx_up[i] * this->coefficients_[i];
+        });
+    }
+
+    template <typename LowIdxDiff, typename UpIdxDiff, typename LowIdx, typename UpIdx>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff& idx_diff_up,
+                                              LowIdx& idx_low,
+                                              const UpIdx&) const
+    {
+        static_assert(LowIdxDiff::Size() == 1 && UpIdxDiff::Size() == NDimUp &&
+                          LowIdx::Size() == 1 && UpIdx::Size() == NDimUp,
+                      "wrong! inconsistent # of dimension");
+
+        idx_diff_low(Number<0>{}) = 0;
+
+        static_for<0, NDimUp, 1>{}(
+            [&](auto i) { idx_diff_low(Number<0>{}) += idx_diff_up[i] * coefficients_[i]; });
+
+        idx_low += idx_diff_low;
+    }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value &&
+               is_known_at_compile_time<Coefficients>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("Embed{");
+
+        //
+        printf("up_lengths_: ");
+        print(up_lengths_);
+        printf(", ");
+
+        //
+        printf("coefficients_: ");
+        print(coefficients_);
+
+        printf("}");
+    }
+};
+
+template <typename LowLengths>
+struct lambda_merge_generate_MagicDivision_calculate_magic_divisor
+{
+    template <index_t I>
+    __host__ __device__ constexpr auto operator()(Number<I> i) const
+    {
+        return MagicDivision::CalculateMagicNumbers(LowLengths{}[i]);
+    }
+};
+
+// Implementation of "Merge" transformation primitive that uses magic-number-division to do lowering
+// of both multi-index and delta of multi-index
+// Caution:
+//   1. The magic number division implementation being used would produce correct result if the
+//   dividended is uint32_t and its value is with in 31-bit value range of uint32_t.
+//   2. The magic number division for int32_t dividened has not been implemented, the int32_t
+//   dividend would be bit-wise interpreted as uint32_t and magic number division implementation for
+//   uint32_t is then used.
+//   3. For Merge primitive, upper-index is the dividend.
+//   4. When upper-index is uint32_t, its value need to be within 31-bit range.
+//   5. When upper-index is int32_t type (when index_t is int32_t), its value need to be
+//   non-negative.
+template <typename LowLengths>
+struct Merge_v2_magic_division : public BaseTransform<LowLengths::Size(), 1>
+{
+    static constexpr index_t NDimLow = LowLengths::Size();
+
+    using LowerIndex = MultiIndex<NDimLow>;
+    using UpperIndex = MultiIndex<1>;
+
+    using UpLengths =
+        decltype(make_tuple(container_reduce(LowLengths{}, math::multiplies{}, Number<1>{})));
+
+    using LowLengthsMagicDivisor = decltype(generate_tuple(
+        lambda_merge_generate_MagicDivision_calculate_magic_divisor<LowLengths>{},
+        Number<NDimLow>{}));
+
+    LowLengths low_lengths_;
+    LowLengthsMagicDivisor low_lengths_magic_divisor_;
+    UpLengths up_lengths_;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    __host__ __device__ constexpr Merge_v2_magic_division() = default;
+
+    __host__ __device__ constexpr Merge_v2_magic_division(const LowLengths& low_lengths)
+        : low_lengths_{low_lengths},
+          low_lengths_magic_divisor_{generate_tuple(
+              [&](auto i) { return MagicDivision::CalculateMagicNumbers(low_lengths[i]); },
+              Number<NDimLow>{})},
+          up_lengths_{make_tuple(container_reduce(low_lengths, math::multiplies{}, I1))}
+    {
+        static_assert(LowerIndex::Size() == NDimLow, "wrong!");
+    }
+
+    __host__ __device__ static constexpr auto GetTypeEnum() { return IndexTransformEnum::Merge; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        index_t tmp = idx_up[I0];
+
+        static_for<NDimLow - 1, 0, -1>{}([&, this](auto i) {
+            index_t tmp2 = MagicDivision::DoMagicDivision(tmp,
+                                                          this->low_lengths_magic_divisor_[i][I0],
+                                                          this->low_lengths_magic_divisor_[i][I1]);
+            idx_low(i)   = tmp - tmp2 * this->low_lengths_[i];
+            tmp          = tmp2;
+        });
+
+        idx_low(Number<0>{}) = tmp;
+    }
+
+    template <typename LowIdxDiff, typename UpIdxDiff, typename LowIdx, typename UpIdx>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff&,
+                                              LowIdx& idx_low,
+                                              const UpIdx& idx_up_new) const
+    {
+        static_assert(LowIdxDiff::Size() == NDimLow && UpIdxDiff::Size() == 1 &&
+                          LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        index_t tmp = idx_up_new[Number<0>{}];
+
+        static_for<NDimLow - 1, 0, -1>{}([&, this](auto i) {
+            index_t tmp2 = MagicDivision::DoMagicDivision(tmp,
+                                                          this->low_lengths_magic_divisor_[i][I0],
+                                                          this->low_lengths_magic_divisor_[i][I1]);
+
+            index_t idx_low_old = idx_low[i];
+
+            idx_low(i) = tmp - tmp2 * this->low_lengths_[i];
+            tmp        = tmp2;
+
+            idx_diff_low(i) = idx_low[i] - idx_low_old;
+        });
+
+        idx_diff_low(Number<0>{}) = tmp - idx_low(Number<0>{});
+
+        idx_low(Number<0>{}) = tmp;
+    }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<LowLengths>::value &&
+               is_known_at_compile_time<LowLengthsMagicDivisor>::value &&
+               is_known_at_compile_time<UpLengths>::value;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    // MUST be static function
+    template <typename LowVectorLengths, typename LowVectorStrides>
+    __host__ __device__ static constexpr auto
+    CalculateUpperDimensionSafeVectorLengthStrides(const LowVectorLengths& low_vector_lengths,
+                                                   const LowVectorStrides& low_vector_strides)
+    {
+        Array<index_t, 1> up_vector_lengths{-1};
+        Array<index_t, 1> up_vector_strides{-1};
+
+        up_vector_lengths(0) = low_vector_lengths[Number<NDimLow - 1>{}];
+        up_vector_strides(0) = low_vector_strides[Number<NDimLow - 1>{}];
+
+        return make_tuple(up_vector_lengths, up_vector_strides);
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("Merge_v2_magic_division{");
+
+        //
+        printf("low_lengths_ ");
+        print(low_lengths_);
+        printf(", ");
+
+        //
+        printf("up_lengths_ ");
+        print(up_lengths_);
+
+        printf("}");
+    }
+};
+
+// Implementation of "Merge" transformation primitive that uses division and mod. It is supposed to
+// be used for low_lengths that are known at compile time and are power of 2, otherwise performance
+// will be very bad
+template <typename LowLengths>
+struct Merge_v3_division_mod : public BaseTransform<LowLengths::Size(), 1>
+{
+    static constexpr index_t NDimLow = LowLengths::Size();
+
+    using LowerIndex = MultiIndex<NDimLow>;
+    using UpperIndex = MultiIndex<1>;
+
+    using LowLengthsScan =
+        decltype(container_reverse_exclusive_scan(LowLengths{}, math::multiplies{}, Number<1>{}));
+
+    using UpLengths =
+        decltype(make_tuple(container_reduce(LowLengths{}, math::multiplies{}, Number<1>{})));
+
+    LowLengths low_lengths_;
+    LowLengthsScan low_lengths_scan_;
+    UpLengths up_lengths_;
+
+    __host__ __device__ constexpr Merge_v3_division_mod() = default;
+
+    __host__ __device__ constexpr Merge_v3_division_mod(const LowLengths& low_lengths)
+        : low_lengths_{low_lengths},
+          low_lengths_scan_{
+              container_reverse_exclusive_scan(low_lengths, math::multiplies{}, Number<1>{})},
+          up_lengths_{make_tuple(container_reduce(low_lengths, math::multiplies{}, Number<1>{}))}
+    {
+        static_assert(LowerIndex::Size() == NDimLow, "wrong!");
+    }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        index_t tmp = idx_up[Number<0>{}];
+
+        // division and mod
+        static_for<0, NDimLow - 1, 1>{}([&](auto i) {
+            idx_low(i) = tmp / this->low_lengths_scan_[i];
+            tmp %= this->low_lengths_scan_[i];
+        });
+
+        idx_low(Number<NDimLow - 1>{}) = tmp;
+    }
+
+    template <typename LowIdxDiff, typename UpIdxDiff, typename LowIdx, typename UpIdx>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff&,
+                                              LowIdx& idx_low,
+                                              const UpIdx& idx_up_new) const
+    {
+        static_assert(LowIdxDiff::Size() == NDimLow && UpIdxDiff::Size() == 1 &&
+                          LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0   = Number<0>{};
+        constexpr auto INm1 = Number<NDimLow - 1>{};
+
+        index_t tmp = idx_up_new[I0];
+
+        static_for<0, NDimLow - 1, 1>{}([&](auto i) {
+            const index_t tmp2 = idx_low[i];
+            idx_low(i)         = tmp / this->low_lengths_scan_[i];
+            idx_diff_low(i)    = idx_low[i] - tmp2;
+            tmp %= this->low_lengths_scan_[i];
+        });
+
+        const index_t tmp2 = idx_low[INm1];
+        idx_low(INm1)      = tmp;
+        idx_diff_low(INm1) = idx_low[INm1] - tmp2;
+    }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<LowLengths>::value &&
+               is_known_at_compile_time<LowLengthsScan>::value &&
+               is_known_at_compile_time<UpLengths>::value;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    // MUST be static function
+    template <typename LowVectorLengths, typename LowVectorStrides>
+    __host__ __device__ static constexpr auto
+    CalculateUpperDimensionSafeVectorLengthStrides(const LowVectorLengths& low_vector_lengths,
+                                                   const LowVectorStrides& low_vector_strides)
+    {
+        Array<index_t, 1> up_vector_lengths{-1};
+        Array<index_t, 1> up_vector_strides{-1};
+
+        up_vector_lengths(0) = low_vector_lengths[Number<NDimLow - 1>{}];
+        up_vector_strides(0) = low_vector_strides[Number<NDimLow - 1>{}];
+
+        return make_tuple(up_vector_lengths, up_vector_strides);
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("Merge_v3_direct_division_mod{");
+
+        //
+        printf("low_lengths_ ");
+        print(low_lengths_);
+        printf(", ");
+
+        //
+        printf("low_lengths_scan_ ");
+        print(low_lengths_scan_);
+        printf(", ");
+
+        //
+        printf("up_lengths_ ");
+        print(up_lengths_);
+
+        printf("}");
+    }
+};
+
+template <typename UpLengths, bool Use24BitIntegerCalculation>
+struct UnMerge : public BaseTransform<1, UpLengths::Size()>
+{
+    static constexpr index_t NDimUp = UpLengths::Size();
+
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<NDimUp>;
+
+    using UpLengthsScan =
+        decltype(container_reverse_exclusive_scan(UpLengths{}, math::multiplies{}, Number<1>{}));
+
+    UpLengths up_lengths_;
+    UpLengthsScan up_lengths_scan_;
+
+    __host__ __device__ constexpr UnMerge() = default;
+
+    __host__ __device__ constexpr UnMerge(const UpLengths& up_lengths)
+        : up_lengths_{up_lengths},
+          up_lengths_scan_{
+              container_reverse_exclusive_scan(up_lengths, math::multiplies{}, Number<1>{})}
+    {
+    }
+
+    __host__ __device__ static constexpr auto GetTypeEnum() { return IndexTransformEnum::UnMerge; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        if constexpr(!Use24BitIntegerCalculation)
+        {
+            idx_low(Number<0>{}) = idx_up[Number<NDimUp - 1>{}];
+
+            static_for<0, NDimUp - 1, 1>{}(
+                [&](auto i) { idx_low(Number<0>{}) += idx_up[i] * up_lengths_scan_[i]; });
+        }
+        else
+        {
+            idx_low(Number<0>{}) = idx_up[Number<NDimUp - 1>{}];
+
+            static_for<0, NDimUp - 1, 1>{}([&](auto i) {
+                idx_low(Number<0>{}) =
+                    (0x00ffffff & idx_low[Number<0>{}]) +
+                    (0x00ffffff & idx_up[i]) * (0x00ffffff & up_lengths_scan_[i]);
+            });
+        }
+    }
+
+    template <typename LowIdxDiff, typename UpIdxDiff, typename LowIdx, typename UpIdx>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff& idx_diff_up,
+                                              LowIdx& idx_low,
+                                              const UpIdx&) const
+    {
+        CalculateLowerIndex(idx_diff_low, idx_diff_up);
+
+        idx_low += idx_diff_low;
+    }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value &&
+               is_known_at_compile_time<UpLengthsScan>::value;
+    }
+
+    // MUST be static function
+    template <typename LowVectorLengths, typename LowVectorStrides>
+    __host__ __device__ static constexpr auto
+    CalculateUpperDimensionSafeVectorLengthStrides(const LowVectorLengths& low_vector_lengths,
+                                                   const LowVectorStrides& low_vector_strides)
+    {
+        Array<index_t, NDimUp> up_vector_lengths{-1};
+        Array<index_t, NDimUp> up_vector_strides{-1};
+
+        constexpr auto up_length_last = UpLengths{}[Number<NDimUp - 1>{}];
+
+        if constexpr(is_known_at_compile_time<decltype(up_length_last)>::value)
+        {
+            if(low_vector_lengths[0] != -1)
+            {
+                up_vector_lengths(NDimUp - 1) = math::gcd(low_vector_lengths[0], up_length_last);
+            }
+        }
+
+        up_vector_strides(NDimUp - 1) = low_vector_strides[0];
+
+        return make_tuple(up_vector_lengths, up_vector_strides);
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("UnMerge{");
+
+        //
+        printf("up_lengths_");
+        print(up_lengths_);
+        printf(", ");
+
+        //
+        printf("up_lengths_scan_");
+        print(up_lengths_scan_);
+
+        printf("}");
+    }
+};
+
+template <typename LowerIndex>
+struct Freeze : public BaseTransform<1, 0>
+{
+    LowerIndex low_idx_;
+
+    __host__ __device__ constexpr Freeze() = default;
+
+    __host__ __device__ constexpr Freeze(const LowerIndex& low_idx) : low_idx_{low_idx} {}
+
+    __host__ __device__ static constexpr auto GetUpperLengths() { return Tuple<>{}; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& /* idx_up */) const
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == 0,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = low_idx_;
+    }
+
+    template <typename LowIdxDiff, typename UpIdxDiff, typename LowIdx, typename UpIdx>
+    __host__ __device__ static void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                                     const UpIdxDiff& /* idx_diff_up */,
+                                                     LowIdx& /* idx_low */,
+                                                     const UpIdx& /* idx_up_new */)
+    {
+        idx_diff_low(Number<0>{}) = 0;
+    }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<LowerIndex>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("Freeze{");
+
+        //
+        printf("low_idx_: ");
+        print(low_idx_);
+
+        printf("}");
+    }
+};
+
+// Replicate the original tensor and create a higher dimensional tensor
+template <typename UpLengths>
+struct Replicate : public BaseTransform<0, UpLengths::Size()>
+{
+    static constexpr index_t NDimUp = UpLengths::Size();
+
+    __host__ __device__ constexpr Replicate() = default;
+
+    __host__ __device__ constexpr Replicate(const UpLengths& up_lengths) : up_lengths_{up_lengths}
+    {
+    }
+
+    __host__ __device__ constexpr auto GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx&, const UpIdx&) const
+    {
+        static_assert(LowIdx::Size() == 0 && UpIdx::Size() == NDimUp,
+                      "wrong! inconsistent # of dimension");
+    }
+
+    template <typename LowIdxDiff, typename UpIdxDiff, typename LowIdx, typename UpIdx>
+    __host__ __device__ static void
+    UpdateLowerIndex(LowIdxDiff&, const UpIdxDiff&, LowIdx&, const UpIdx&)
+    {
+        static_assert(LowIdxDiff::Size() == 0 && UpIdxDiff::Size() == NDimUp &&
+                          LowIdx::Size() == 0 && UpIdx::Size() == NDimUp,
+                      "wrong! inconsistent # of dimension");
+    }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("Replicate{");
+
+        //
+        printf("up_lengths_: ");
+        print(up_lengths_);
+
+        printf("}");
+    }
+
+    //
+    UpLengths up_lengths_;
+};
+
+template <typename LowLength, typename SliceBegin, typename SliceEnd>
+struct Slice : public BaseTransform<1, 1>
+{
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<1>;
+
+    using UpLengths = decltype(make_tuple(SliceEnd{} - SliceBegin{}));
+
+    UpLengths up_lengths_;
+    SliceBegin slice_begin_;
+    SliceEnd slice_end_;
+
+    __host__ __device__ constexpr Slice() = default;
+
+    __host__ __device__ constexpr Slice(const LowLength&,
+                                        const SliceBegin& slice_begin,
+                                        const SliceEnd& slice_end)
+        : up_lengths_{make_tuple(slice_end - slice_begin)},
+          slice_begin_{slice_begin},
+          slice_end_{slice_end}
+    {
+    }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = idx_up[Number<0>{}] + slice_begin_;
+    }
+
+    template <typename LowIdxDiff, typename UpIdxDiff, typename LowIdx, typename UpIdx>
+    __host__ __device__ static void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                                     const UpIdxDiff& idx_diff_up,
+                                                     LowIdx& idx_low,
+                                                     const UpIdx&)
+    {
+        static_assert(LowIdxDiff::Size() == 1 && UpIdxDiff::Size() == 1 && LowIdx::Size() == 1 &&
+                          UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0 = Number<0>{};
+
+        idx_diff_low(I0) = idx_diff_up[I0];
+
+        idx_low += idx_diff_low;
+    }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ constexpr bool IsValidUpperIndexMappedToValidLowerIndex(const UpIdx&) const
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value &&
+               is_known_at_compile_time<SliceBegin>::value &&
+               is_known_at_compile_time<SliceEnd>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("Slice{");
+
+        //
+        printf("up_lengths_: ");
+        print(up_lengths_);
+        printf(", ");
+
+        //
+        printf("slice_begin_: ");
+        print(slice_begin_);
+        printf(", ");
+
+        //
+        printf("slice_end_: ");
+        print(slice_end_);
+
+        printf("}");
+    } // namespace ck
+};    // namespace ck
+
+/*
+ * \brief lower_idx = upper_idx % modulus.
+ * TODO: Need an improved implementation since the modulo operation is expensive.
+ */
+template <typename Modulus, typename UpLength>
+struct Modulo : public BaseTransform<1, 1>
+{
+    using LowerIndex = MultiIndex<1>;
+    using UpperIndex = MultiIndex<1>;
+    using UpLengths  = decltype(make_tuple(UpLength{}));
+
+    Modulus modulus_;
+    UpLengths up_lengths_;
+
+    __host__ __device__ constexpr Modulo() = default;
+
+    __host__ __device__ constexpr Modulo(const Modulus& modulus, const UpLength& up_length)
+        : modulus_{modulus}, up_lengths_{make_tuple(up_length)}
+    {
+    }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == 1 && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = idx_up[Number<0>{}] % modulus_;
+    }
+
+    template <typename LowIdxDiff, typename UpIdxDiff, typename LowIdx, typename UpIdx>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff& idx_diff_up,
+                                              LowIdx& idx_low,
+                                              const UpIdx& up_idx) const
+    {
+        static_assert(LowIdxDiff::Size() == 1 && UpIdxDiff::Size() == 1 && LowIdx::Size() == 1 &&
+                          UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0 = Number<0>{};
+
+        const auto idx_low_old = idx_low;
+        idx_low(I0)            = (up_idx(I0) + idx_diff_up(I0)) % modulus_;
+        idx_diff_low(I0)       = idx_low - idx_low_old;
+    }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("Modulus{");
+
+        //
+        printf("up_lengths_: ");
+        print(up_lengths_);
+
+        printf("}");
+    }
+};
+
+// 2D XOR
+template <typename LowLengths, typename RightShift>
+struct Xor : public BaseTransform<2, 2>
+{
+    static constexpr auto type_enum = IndexTransformEnum::Xor;
+
+    using LowerIndex = MultiIndex<2>;
+    using UpperIndex = MultiIndex<2>;
+
+    using UpLengths = LowLengths;
+
+    UpLengths up_lengths_;
+    RightShift right_shift_;
+
+    __host__ __device__ constexpr Xor() : up_lengths_{}, right_shift_{} {}
+
+    __host__ __device__ constexpr Xor(const LowLengths& low_lengths, const RightShift& right_shift)
+        : up_lengths_{low_lengths}, right_shift_{right_shift}
+    {
+    }
+
+    __host__ __device__ static constexpr auto GetTypeEnum() { return IndexTransformEnum::Xor; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == 2 && UpIdx::Size() == 2,
+                      "wrong! inconsistent # of dimension");
+
+        idx_low(Number<0>{}) = idx_up[Number<0>{}];
+
+        const auto idx_low_1_tmp =
+            (idx_up[Number<1>{}] - idx_up[Number<0>{}] * right_shift_) % up_lengths_[Number<1>{}];
+
+        const auto idx_low_1 =
+            (idx_low_1_tmp >= 0) ? idx_low_1_tmp : up_lengths_[Number<1>{}] + idx_low_1_tmp;
+
+        idx_low(Number<1>{}) = idx_low_1;
+    }
+
+    template <typename LowIdxDiff, typename UpIdxDiff, typename LowIdx, typename UpIdx>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff&,
+                                              LowIdx& idx_low,
+                                              const UpIdx& idx_up) const
+    {
+        static_assert(LowIdxDiff::Size() == 2 && UpIdxDiff::Size() == 2 && LowIdx::Size() == 2 &&
+                          UpIdx::Size() == 2,
+                      "wrong! inconsistent # of dimension");
+
+        const auto idx_low_old = idx_low;
+
+        CalculateLowerIndex(idx_low, idx_up);
+
+        idx_diff_low = idx_low - idx_low_old;
+    }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<UpLengths>::value &&
+               is_known_at_compile_time<RightShift>::value;
+    }
+
+    // MUST be static function
+    template <typename LowVectorLengths, typename LowVectorStrides>
+    __host__ __device__ constexpr auto
+    CalculateUpperDimensionSafeVectorLengthStrides(const LowVectorLengths& low_vector_lengths,
+                                                   const LowVectorStrides& low_vector_strides) const
+    {
+        Array<index_t, 2> up_vector_lengths = low_vector_lengths;
+        Array<index_t, 2> up_vector_strides = low_vector_strides;
+
+        if constexpr(is_known_at_compile_time<RightShift>::value)
+        {
+            if(low_vector_lengths[1] != -1)
+            {
+                up_vector_lengths(1) = math::gcd(low_vector_lengths[1], math::abs(right_shift_));
+            }
+        }
+
+        return make_tuple(up_vector_lengths, up_vector_strides);
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("Xor{");
+
+        //
+        printf("up_lengths_: ");
+        print(up_lengths_);
+        printf(", ");
+
+        //
+        printf("right_shift_: ");
+        print(right_shift_);
+
+        printf("}");
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_description/multi_index_transform_helper.hpp b/include/ck/tensor_description/multi_index_transform_helper.hpp
new file mode 100644
index 0000000000..9bb7bdb2c5
--- /dev/null
+++ b/include/ck/tensor_description/multi_index_transform_helper.hpp
@@ -0,0 +1,116 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform.hpp"
+
+namespace ck {
+
+template <typename LowLength>
+__host__ __device__ constexpr auto make_pass_through_transform(const LowLength& low_length)
+{
+    return PassThrough<LowLength>{low_length};
+}
+
+template <typename LowLength, typename LeftPad, typename RightPad, bool SkipIsValidCheck = false>
+__host__ __device__ constexpr auto
+make_pad_transform(const LowLength& low_length,
+                   const LeftPad& left_pad,
+                   const RightPad& right_pad,
+                   integral_constant<bool, SkipIsValidCheck> = integral_constant<bool, false>{})
+{
+    return Pad<LowLength, LeftPad, RightPad, SkipIsValidCheck>{low_length, left_pad, right_pad};
+}
+
+template <typename LowLength, typename LeftPadLength, bool SkipIsValidCheck = false>
+__host__ __device__ constexpr auto make_left_pad_transform(
+    const LowLength& low_length,
+    const LeftPadLength& left_pad,
+    integral_constant<bool, SkipIsValidCheck> = integral_constant<bool, false>{})
+{
+    return LeftPad<LowLength, LeftPadLength, SkipIsValidCheck>{low_length, left_pad};
+}
+
+template <typename LowLength, typename RightPadLength, bool SkipIsValidCheck = false>
+__host__ __device__ constexpr auto make_right_pad_transform(
+    const LowLength& low_length,
+    const RightPadLength& right_pad,
+    integral_constant<bool, SkipIsValidCheck> = integral_constant<bool, false>{})
+{
+    return RightPad<LowLength, RightPadLength, SkipIsValidCheck>{low_length, right_pad};
+}
+
+template <typename UpLengths,
+          typename Coefficients,
+          typename enable_if<UpLengths::Size() == Coefficients::Size(), bool>::type = false>
+__host__ __device__ constexpr auto make_embed_transform(const UpLengths& up_lengths,
+                                                        const Coefficients& coefficients)
+{
+    return Embed<UpLengths, Coefficients>{up_lengths, coefficients};
+}
+
+template <typename LowLengths>
+__host__ __device__ constexpr auto
+make_merge_transform_v2_magic_division(const LowLengths& low_lengths)
+{
+    return Merge_v2_magic_division<LowLengths>{low_lengths};
+}
+
+template <typename LowLengths>
+__host__ __device__ constexpr auto
+make_merge_transform_v3_division_mod(const LowLengths& low_lengths)
+{
+    return Merge_v3_division_mod<LowLengths>{low_lengths};
+}
+
+template <typename LowLengths>
+__host__ __device__ constexpr auto make_merge_transform(const LowLengths& low_lengths)
+{
+    return make_merge_transform_v2_magic_division(low_lengths);
+}
+
+template <typename UpLengths, bool Use24BitIntegerCalculation = false>
+__host__ __device__ constexpr auto make_unmerge_transform(
+    const UpLengths& up_lengths,
+    integral_constant<bool, Use24BitIntegerCalculation> = integral_constant<bool, false>{})
+{
+    return UnMerge<UpLengths, Use24BitIntegerCalculation>{up_lengths};
+}
+
+template <typename LowerIndex>
+__host__ __device__ constexpr auto make_freeze_transform(const LowerIndex& low_idx)
+{
+    return Freeze<LowerIndex>{low_idx};
+}
+
+template <typename UpLengths>
+__host__ __device__ constexpr auto make_replicate_transform(const UpLengths& up_lengths)
+{
+    return Replicate<UpLengths>{up_lengths};
+}
+
+template <typename LowLength, typename SliceBegin, typename SliceEnd>
+__host__ __device__ constexpr auto make_slice_transform(const LowLength& low_length,
+                                                        const SliceBegin& slice_begin,
+                                                        const SliceEnd& slice_end)
+{
+    return Slice<LowLength, SliceBegin, SliceEnd>{low_length, slice_begin, slice_end};
+}
+
+template <typename Modulus, typename UpLength>
+__host__ __device__ constexpr auto make_modulo_transform(const Modulus& modulus,
+                                                         const UpLength& up_length)
+{
+    return Modulo<Modulus, UpLength>{modulus, up_length};
+}
+
+template <typename LowLengths, typename RightShift>
+__host__ __device__ constexpr auto make_xor_transform(const LowLengths& low_lengths,
+                                                      const RightShift& right_shift)
+{
+    return Xor<LowLengths, RightShift>{low_lengths, right_shift};
+}
+
+} // namespace ck
diff --git a/include/ck/tensor_description/tensor_adaptor.hpp b/include/ck/tensor_description/tensor_adaptor.hpp
new file mode 100644
index 0000000000..bf1170d074
--- /dev/null
+++ b/include/ck/tensor_description/tensor_adaptor.hpp
@@ -0,0 +1,675 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+
+namespace ck {
+
+// Transforms: Tuple<transforms...>
+// LowerDimensionHiddenIdss : Tuple<Sequence<...>, ...>
+// UpperDimensionHiddenIdss : Tuple<Sequence<...>, ...>
+// BottomDimensionHiddenIds : Sequence<...>
+// TopDimensionHiddenIds : Sequence<...>
+template <typename Transforms,
+          typename LowerDimensionHiddenIdss,
+          typename UpperDimensionHiddenIdss,
+          typename BottomDimensionHiddenIds,
+          typename TopDimensionHiddenIds>
+struct TensorAdaptor
+{
+    __host__ __device__ static constexpr index_t GetNumOfTransform() { return Transforms::Size(); }
+
+    __host__ __device__ constexpr const auto& GetTransforms() const { return transforms_; }
+
+    __host__ __device__ static constexpr auto GetLowerDimensionHiddenIdss()
+    {
+        return LowerDimensionHiddenIdss{};
+    }
+
+    __host__ __device__ static constexpr auto GetUpperDimensionHiddenIdss()
+    {
+        return UpperDimensionHiddenIdss{};
+    }
+
+    __host__ __device__ static constexpr auto GetBottomDimensionHiddenIds()
+    {
+        return BottomDimensionHiddenIds{};
+    }
+
+    __host__ __device__ static constexpr auto GetTopDimensionHiddenIds()
+    {
+        return TopDimensionHiddenIds{};
+    }
+
+    __host__ __device__ static constexpr auto InitializeElementSize(const Transforms& transforms)
+    {
+        const auto lengths = generate_tuple(
+            [&](auto idim_top) {
+                constexpr index_t idim_hidden = TopDimensionHiddenIds::At(idim_top);
+
+                constexpr auto tmp = GetTransformAndItsUpperDimension(Number<idim_hidden>{});
+
+                constexpr index_t itran   = tmp[Number<0>{}];
+                constexpr index_t idim_up = tmp[Number<1>{}];
+                constexpr bool found      = tmp[Number<2>{}];
+
+                static_assert(found == true,
+                              "wrong! not found matching transformation and upper-dimension");
+
+                const auto length =
+                    transforms[Number<itran>{}].GetUpperLengths()[Number<idim_up>{}];
+
+                return length;
+            },
+            Number<ndim_top_>{});
+
+        // TODO: make container_reduce support tuple of Number and index_t
+        return container_reduce(lengths, math::multiplies{}, Number<1>{});
+    }
+
+    template <index_t IDimHidden>
+    __host__ __device__ static constexpr auto GetTransformAndItsUpperDimension(Number<IDimHidden>)
+    {
+        // FIXME: length of bottom dimension is not known, since info about lower dim length are not
+        // saved in transformation
+        static_assert(IDimHidden >= ndim_bottom_, "wrong! not implemented");
+
+        index_t itran_found   = 0;
+        index_t idim_up_found = 0;
+        bool found            = false;
+
+        static_for<0, ntransform_, 1>{}([&](auto itran) {
+            constexpr auto up_dim_ids = UpperDimensionHiddenIdss{}[itran];
+
+            static_for<0, up_dim_ids.Size(), 1>{}([&](auto idim_up) {
+                if constexpr(up_dim_ids[idim_up] == IDimHidden)
+                {
+                    itran_found   = itran;
+                    idim_up_found = idim_up;
+                    found         = true;
+                }
+            });
+        });
+
+        return make_tuple(itran_found, idim_up_found, found);
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfBottomDimension()
+    {
+        return BottomDimensionHiddenIds::Size();
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfTopDimension()
+    {
+        return TopDimensionHiddenIds::Size();
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfHiddenDimension()
+    {
+        constexpr auto all_low_dim_ids = unpack(
+            [](auto&&... xs) constexpr { return merge_sequences(xs...); },
+            LowerDimensionHiddenIdss{});
+
+        constexpr auto all_up_dim_ids = unpack(
+            [](auto&&... xs) constexpr { return merge_sequences(xs...); },
+            UpperDimensionHiddenIdss{});
+
+        constexpr auto all_dim_ids = merge_sequences(all_low_dim_ids, all_up_dim_ids);
+
+        using unique_sort_all_dim_ids = typename sequence_unique_sort<decltype(all_dim_ids),
+                                                                      math::less<index_t>,
+                                                                      math::equal<index_t>>::type;
+
+        return unique_sort_all_dim_ids::Size();
+    }
+
+    constexpr static index_t ntransform_  = GetNumOfTransform();
+    constexpr static index_t ndim_hidden_ = GetNumOfHiddenDimension();
+    constexpr static index_t ndim_bottom_ = GetNumOfBottomDimension();
+    constexpr static index_t ndim_top_    = GetNumOfTopDimension();
+
+    using HiddenIndex = MultiIndex<ndim_hidden_>;
+    using BottomIndex = MultiIndex<ndim_bottom_>;
+    using TopIndex    = MultiIndex<ndim_top_>;
+
+    // may be index_t or Number<>
+    using ElementSize = remove_cv_t<decltype(InitializeElementSize(Transforms{}))>;
+
+    public:
+    __host__ __device__ constexpr TensorAdaptor() = default;
+
+    __host__ __device__ constexpr TensorAdaptor(const Transforms& transforms)
+        : transforms_{transforms}, element_size_{InitializeElementSize(transforms)}
+    {
+        static_assert(Transforms::Size() == ntransform_ &&
+                          LowerDimensionHiddenIdss::Size() == ntransform_ &&
+                          UpperDimensionHiddenIdss::Size() == ntransform_,
+                      "wrong! inconsistent # of transformations");
+
+        // TODO check dependency of dimensions is valid
+    }
+
+    __host__ __device__ constexpr auto GetElementSize() const { return element_size_; }
+
+    // FIXME: this logic is wrong when getting bottome dimension lengths
+    template <index_t IDimHidden>
+    __host__ __device__ constexpr auto GetHiddenDimensionLength(Number<IDimHidden>) const
+    {
+        static_assert(IDimHidden >= 0 && IDimHidden < ndim_hidden_, "wrong! out of range");
+
+        constexpr auto tmp = GetTransformAndItsUpperDimension(Number<IDimHidden>{});
+
+        constexpr index_t itran   = tmp[Number<0>{}];
+        constexpr index_t idim_up = tmp[Number<1>{}];
+        constexpr bool found      = tmp[Number<2>{}];
+
+        static_assert(found == true,
+                      "wrong! not found matching transformation and upper-dimension");
+
+        return transforms_[Number<itran>{}].GetUpperLengths()[Number<idim_up>{}];
+    }
+
+    template <index_t IDimTop>
+    __host__ __device__ constexpr auto GetTopDimensionLength(Number<IDimTop> idim_top) const
+    {
+        return GetHiddenDimensionLength(TopDimensionHiddenIds::At(idim_top));
+    }
+
+#if 0
+    // FIXME: GetHiddenDimensionLength is wrong when getting bottome dimension lengths
+    template <index_t IDimBottom>
+    __host__ __device__ constexpr index_t
+    GetBottomDimensionLength(Number<IDimBottom> idim_bottom) const
+    {
+        return GetHiddenDimensionLength(TopDimensionHiddenIds::At(idim_bottom));
+    }
+#endif
+
+    __host__ __device__ constexpr auto GetTopDimensionLengths() const
+    {
+        return generate_tuple([&](auto i) { return GetTopDimensionLength(i); },
+                              Number<ndim_top_>{});
+    }
+
+#if 0
+    // FIXME: GetHiddenDimensionLength is wrong when getting bottome dimension lengths
+    __host__ __device__ constexpr auto GetBottomDimensionLengths() const
+    {
+        return generate_tuple([&](auto i) { return GetBottomDimensionLength(i); },
+                              Number<ndim_bottom_>{});
+    }
+#endif
+
+    template <typename TopIdx>
+    __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+    {
+        static_assert(TopIdx::Size() == TopDimensionHiddenIds::Size(),
+                      "wrong! # of dimension inconsistent");
+
+        constexpr index_t ntransform  = GetNumOfTransform();
+        constexpr index_t ndim_hidden = GetNumOfHiddenDimension();
+
+        MultiIndex<ndim_hidden> idx_hidden;
+
+        // initialize uppest index
+        set_container_subset(idx_hidden, GetTopDimensionHiddenIds(), idx_top);
+
+        // calculate hidden index
+        static_for<ntransform, 0, -1>{}([&](auto itran_p1) {
+            auto itran              = itran_p1 - Number<1>{};
+            const auto& tran        = GetTransforms().At(itran);
+            constexpr auto dims_low = GetLowerDimensionHiddenIdss().At(itran);
+            constexpr auto dims_up  = GetUpperDimensionHiddenIdss().At(itran);
+
+            const auto idx_up = get_container_subset(idx_hidden, dims_up);
+
+            MultiIndex<dims_low.Size()> idx_low;
+
+            tran.CalculateLowerIndex(idx_low, idx_up);
+
+            set_container_subset(idx_hidden, dims_low, idx_low);
+        });
+
+        return get_container_subset(idx_hidden, BottomDimensionHiddenIds{});
+    }
+
+    __host__ __device__ static constexpr bool IsStatic()
+    {
+        bool is_known = true;
+
+        static_for<0, Transforms::Size(), 1>{}([&](auto i) {
+            is_known &= remove_cvref_t<decltype(Transforms{}[i])>::IsKnownAtCompileTime();
+        });
+
+        return is_known && is_known_at_compile_time<ElementSize>::value;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime() { return IsStatic(); }
+
+    __host__ __device__ static constexpr auto GetTopDimensionSafeVectorLengthStrides(
+        const Array<index_t, ndim_hidden_>& guaranteed_vector_lengths,
+        const Array<index_t, ndim_hidden_>& guaranteed_vector_strides)
+    {
+        auto vector_lengths = guaranteed_vector_lengths;
+        auto vector_strides = guaranteed_vector_strides;
+
+        static_for<0, GetNumOfTransform(), 1>{}([&](auto itran) {
+            constexpr auto low_dims = GetLowerDimensionHiddenIdss().At(itran);
+            constexpr auto up_dims  = GetUpperDimensionHiddenIdss().At(itran);
+
+            const auto up_guaranteed_vector_lengths =
+                get_container_subset(guaranteed_vector_lengths, up_dims);
+            const auto up_guaranteed_vector_strides =
+                get_container_subset(guaranteed_vector_strides, up_dims);
+
+            // only need type of transform
+            auto [up_vector_lengths, up_vector_strides] =
+                Transforms{}.At(itran).CalculateUpperDimensionSafeVectorLengthStrides(
+                    get_container_subset(vector_lengths, low_dims),
+                    get_container_subset(vector_strides, low_dims));
+
+            if constexpr(up_dims.Size() > 0)
+            {
+                for(index_t i = 0; i < up_dims.Size(); ++i)
+                {
+                    up_vector_lengths(i) = (up_guaranteed_vector_lengths[i] != -1)
+                                               ? up_guaranteed_vector_lengths[i]
+                                               : up_vector_lengths[i];
+
+                    up_vector_strides(i) = (up_guaranteed_vector_strides[i] != -1)
+                                               ? up_guaranteed_vector_strides[i]
+                                               : up_vector_strides[i];
+                }
+            }
+
+            set_container_subset(vector_lengths, up_dims, up_vector_lengths);
+            set_container_subset(vector_strides, up_dims, up_vector_strides);
+        });
+
+        constexpr auto top_dims = TopDimensionHiddenIds{};
+
+        return make_tuple(get_container_subset(vector_lengths, top_dims),
+                          get_container_subset(vector_strides, top_dims));
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("TensorAdaptor{");
+
+        //
+        printf("transforms: ");
+        print(transforms_);
+        printf(", ");
+
+        //
+        printf("LowerDimensionHiddenIds: ");
+        print(LowerDimensionHiddenIdss{});
+        printf(", ");
+
+        //
+        printf("UpperDimensionHiddenIds: ");
+        print(UpperDimensionHiddenIdss{});
+        printf(", ");
+
+        //
+        printf("BottomDimensionHiddenIds: ");
+        print(BottomDimensionHiddenIds{});
+        printf(", ");
+
+        //
+        printf("TopDimensionHiddenIds: ");
+        print(TopDimensionHiddenIds{});
+
+        printf("}");
+    }
+
+    private:
+    Transforms transforms_;
+    ElementSize element_size_;
+};
+
+// Transforms: Tuple<transforms...>
+// LowerDimensionOldTopIdss: Tuple<Sequence<...>, ...>
+// UpperDimensionNewTopIdss: Tuple<Sequence<...>, ...>
+template <typename Transforms, typename LowerDimensionOldTopIdss, typename UpperDimensionNewTopIdss>
+__host__ __device__ constexpr auto make_single_stage_tensor_adaptor(const Transforms& transforms,
+                                                                    LowerDimensionOldTopIdss,
+                                                                    UpperDimensionNewTopIdss)
+{
+    constexpr index_t ntransform = Transforms::Size();
+
+    static_assert(LowerDimensionOldTopIdss::Size() == ntransform &&
+                      UpperDimensionNewTopIdss::Size() == ntransform,
+                  "wrong!");
+
+    // sanity check on LowerDimensionOldTopIdss and UpperDimensionNewTopIdss
+    constexpr auto all_low_dim_old_top_ids = unpack(
+        [](auto&&... xs) constexpr { return merge_sequences(xs...); }, LowerDimensionOldTopIdss{});
+
+    constexpr auto all_up_dim_new_top_ids = unpack(
+        [](auto&&... xs) constexpr { return merge_sequences(xs...); }, UpperDimensionNewTopIdss{});
+
+    static_assert(is_valid_sequence_map<decltype(all_low_dim_old_top_ids)>::value &&
+                      is_valid_sequence_map<decltype(all_up_dim_new_top_ids)>::value,
+                  "wrong!");
+
+    constexpr index_t ndim_old_top = all_low_dim_old_top_ids.Size();
+    constexpr index_t ndim_new_top = all_up_dim_new_top_ids.Size();
+
+    // low_dim_hidden_idss
+    constexpr auto low_dim_hidden_idss = LowerDimensionOldTopIdss{};
+
+    // up_dim_hidden_idss: shift UpperDimensionNewTopIdss by ndim_bottom
+    constexpr auto up_dim_hidden_idss = generate_tuple(
+        [](auto itran) { return UpperDimensionNewTopIdss{}[itran] + Number<ndim_old_top>{}; },
+        Number<ntransform>{});
+
+    // bottom_dim_hidden_ids
+    constexpr auto bottom_dim_hidden_ids =
+        typename arithmetic_sequence_gen<0, ndim_old_top, 1>::type{};
+
+    // top_dim_hidden_ids
+    constexpr auto top_dim_hidden_ids =
+        typename arithmetic_sequence_gen<0, ndim_new_top, 1>::type{} + Number<ndim_old_top>{};
+
+    return TensorAdaptor<remove_cvref_t<Transforms>,
+                         remove_cvref_t<decltype(low_dim_hidden_idss)>,
+                         remove_cvref_t<decltype(up_dim_hidden_idss)>,
+                         remove_cvref_t<decltype(bottom_dim_hidden_ids)>,
+                         remove_cvref_t<decltype(top_dim_hidden_ids)>>{transforms};
+}
+
+// TODO: How to fix this? It uses an struct instead of lambda because lambda
+// doesn't have constructor, and to put it outside the scope where it is used
+// (transform_tensor_adaptor) because template cannot be defined inside a function
+// template
+template <typename NewTransforms>
+struct lambda_get_up_dim_num
+{
+    template <typename I>
+    __host__ __device__ constexpr auto operator()(I) const
+    {
+        using Tran = remove_reference_t<decltype(NewTransforms{}.At(I{}))>;
+        return Number<Tran::GetNumOfUpperDimension()>{};
+    }
+};
+
+template <typename OldTensorAdaptor,
+          typename NewTransforms,
+          typename NewLowerDimensionOldTopIdss,
+          typename NewUpperDimensionNewTopIdss>
+__host__ __device__ constexpr auto
+transform_tensor_adaptor(const OldTensorAdaptor& old_tensor_adaptor,
+                         const NewTransforms& new_transforms,
+                         NewLowerDimensionOldTopIdss,
+                         NewUpperDimensionNewTopIdss)
+{
+    // sanity check
+    {
+        static_assert(NewTransforms::Size() == NewLowerDimensionOldTopIdss::Size() &&
+                          NewTransforms::Size() == NewUpperDimensionNewTopIdss::Size(),
+                      "wrong! inconsitent number of transform");
+
+        constexpr auto all_old_top_ids = unpack([](auto... xs) { return merge_sequences(xs...); },
+                                                NewLowerDimensionOldTopIdss{});
+
+        constexpr auto all_new_top_ids = unpack([](auto... xs) { return merge_sequences(xs...); },
+                                                NewUpperDimensionNewTopIdss{});
+
+        static_assert(is_valid_sequence_map<decltype(all_old_top_ids)>::value &&
+                          is_valid_sequence_map<decltype(all_new_top_ids)>::value,
+                      "wrong!");
+    }
+
+    // lower dimension's hidden idss
+    // convert lower dimension top idss (tuple of sequences) to hidden idss (tuple of
+    // sequences)
+    constexpr auto low_dim_hidden_idss = transform_tuples(
+        // convert lower dimension top ids (a sequence) to hidden ids (a sequence)
+        [](auto low_dim_top_ids) constexpr {
+            return transform_sequences(
+                // convert lower dimension top id to hidden id
+                [](auto low_dim_top_id) constexpr {
+                    return OldTensorAdaptor::GetTopDimensionHiddenIds()[low_dim_top_id];
+                },
+                low_dim_top_ids);
+        },
+        NewLowerDimensionOldTopIdss{});
+
+    constexpr index_t num_new_transform = NewTransforms::Size();
+
+    // upper dimension's hidden idss
+    constexpr index_t old_hidden_dim_number = OldTensorAdaptor::GetNumOfHiddenDimension();
+
+    constexpr auto up_dim_numbers =
+        generate_sequence(lambda_get_up_dim_num<NewTransforms>{}, Number<num_new_transform>{});
+
+    constexpr auto up_dim_numbers_scan = merge_sequences(
+        Sequence<0>{}, inclusive_scan_sequence(up_dim_numbers, math::plus<index_t>{}, Number<0>{}));
+
+    constexpr auto up_dim_hidden_idss = generate_tuple(
+        [ old_hidden_dim_number, up_dim_numbers_scan ](auto i) constexpr {
+            return
+                typename arithmetic_sequence_gen<old_hidden_dim_number + up_dim_numbers_scan[i],
+                                                 old_hidden_dim_number + up_dim_numbers_scan[i + 1],
+                                                 1>::type{};
+        },
+        Number<num_new_transform>{});
+
+    // new top dimension's hidden ids
+    constexpr auto unordered_new_top_dim_hidden_ids = unpack(
+        [](auto... xs) constexpr { return merge_sequences(xs...); }, up_dim_hidden_idss);
+
+    constexpr auto new_top_dim_unordered2ordered = unpack(
+        [](auto... xs) constexpr { return merge_sequences(xs...); }, NewUpperDimensionNewTopIdss{});
+
+    constexpr auto new_top_dim_hidden_ids =
+        unordered_new_top_dim_hidden_ids.ReorderGivenOld2New(new_top_dim_unordered2ordered);
+
+    // put everything together
+    const auto all_transforms =
+        container_concat(old_tensor_adaptor.GetTransforms(), new_transforms);
+
+    constexpr auto all_low_dim_hidden_idss =
+        container_concat(OldTensorAdaptor::GetLowerDimensionHiddenIdss(), low_dim_hidden_idss);
+
+    constexpr auto all_up_dim_hidden_idss =
+        container_concat(OldTensorAdaptor::GetUpperDimensionHiddenIdss(), up_dim_hidden_idss);
+
+    return TensorAdaptor<remove_cvref_t<decltype(all_transforms)>,
+                         remove_cvref_t<decltype(all_low_dim_hidden_idss)>,
+                         remove_cvref_t<decltype(all_up_dim_hidden_idss)>,
+                         remove_cvref_t<decltype(OldTensorAdaptor::GetBottomDimensionHiddenIds())>,
+                         remove_cvref_t<decltype(new_top_dim_hidden_ids)>>{all_transforms};
+}
+
+template <typename TensorAdaptor0, typename TensorAdaptor1>
+__host__ __device__ constexpr auto chain_tensor_adaptors(const TensorAdaptor0& adaptor0,
+                                                         const TensorAdaptor1& adaptor1)
+{
+    static_assert(TensorAdaptor0::GetNumOfTopDimension() ==
+                      TensorAdaptor1::GetNumOfBottomDimension(),
+                  "wrong!");
+
+    // all_transforms = transform0 + transform1
+    const auto all_transforms =
+        container_concat(adaptor0.GetTransforms(), adaptor1.GetTransforms());
+
+    // shift
+    constexpr index_t adaptor0_max_hidden_id = [&]() {
+        index_t adaptor0_max_hidden_id_ = NumericLimits<index_t>::Min();
+
+        static_for<0, TensorAdaptor0::GetNumOfTransform(), 1>{}([&](auto itran) {
+            constexpr index_t ndim_low =
+                TensorAdaptor0{}.GetTransforms()[itran].GetNumOfLowerDimension();
+
+            static_for<0, ndim_low, 1>{}([&](auto idim_low) {
+                adaptor0_max_hidden_id_ =
+                    math::max(adaptor0_max_hidden_id_,
+                              TensorAdaptor0::GetLowerDimensionHiddenIdss()[itran][idim_low].value);
+            });
+
+            constexpr index_t ndim_up =
+                TensorAdaptor0{}.GetTransforms()[itran].GetNumOfUpperDimension();
+
+            static_for<0, ndim_up, 1>{}([&](auto idim_up) {
+                adaptor0_max_hidden_id_ =
+                    math::max(adaptor0_max_hidden_id_,
+                              TensorAdaptor0::GetUpperDimensionHiddenIdss()[itran][idim_up].value);
+            });
+        });
+
+        return adaptor0_max_hidden_id_;
+    }();
+
+    constexpr index_t adaptor1_min_hidden_id = [&]() {
+        index_t adaptor1_min_hidden_id_ = NumericLimits<index_t>::Max();
+
+        static_for<0, TensorAdaptor1::GetNumOfTransform(), 1>{}([&](auto itran) {
+            constexpr index_t ndim_low =
+                TensorAdaptor1{}.GetTransforms()[itran].GetNumOfLowerDimension();
+
+            // get the min of all lower dimenions, but not bottom dimension (because their id will
+            // be matched with top id from adaptor0)
+            static_for<0, ndim_low, 1>{}([&](auto idim_low) {
+                constexpr index_t low_dim_hidden_id =
+                    TensorAdaptor1::GetLowerDimensionHiddenIdss()[itran][idim_low].value;
+
+                bool is_bottom_dim = false;
+                static_for<0, TensorAdaptor1::GetNumOfBottomDimension(), 1>{}([&](auto i) {
+                    if constexpr(low_dim_hidden_id ==
+                                 TensorAdaptor1::GetBottomDimensionHiddenIds()[i])
+                    {
+                        is_bottom_dim = true;
+                    }
+                });
+
+                if(!is_bottom_dim)
+                {
+                    adaptor1_min_hidden_id_ = math::min(adaptor1_min_hidden_id_, low_dim_hidden_id);
+                }
+            });
+
+            constexpr index_t ndim_up =
+                TensorAdaptor1{}.GetTransforms()[itran].GetNumOfUpperDimension();
+
+            // get the min of all upper dimensions
+            static_for<0, ndim_up, 1>{}([&](auto idim_up) {
+                adaptor1_min_hidden_id_ =
+                    math::min(adaptor1_min_hidden_id_,
+                              TensorAdaptor1::GetUpperDimensionHiddenIdss()[itran][idim_up].value);
+            });
+        });
+
+        return adaptor1_min_hidden_id_;
+    }();
+
+    constexpr index_t adaptor1_hidden_id_shift =
+        adaptor0_max_hidden_id + 1 - adaptor1_min_hidden_id;
+
+    constexpr index_t ndim_bottom_1 = TensorAdaptor1::GetNumOfBottomDimension();
+
+    // all_low_dim_hidden_idss =
+    // low_dim_hidden_idss_0 + match_hidden_id_for_1(shift_hidden_id_for_1(low_dim_hiden_idss_1))
+    constexpr auto low_dim_hidden_idss_1 = generate_tuple(
+        // generate sequence of ids for a transform
+        [&](auto itran) {
+            constexpr auto ndim_low_1 = TensorAdaptor1::GetLowerDimensionHiddenIdss()[itran].Size();
+
+            constexpr auto low_dim_hidden_ids_1 =
+                TensorAdaptor1::GetLowerDimensionHiddenIdss()[itran];
+
+            // sequence in, sequence out
+            constexpr auto low_dim_hidden_ids_1_mod = [&]() constexpr
+            {
+                auto low_dim_hidden_ids_1_mod_ = to_multi_index(low_dim_hidden_ids_1);
+
+                // shift hidden id so every dim id is unique
+                static_for<0, ndim_low_1, 1>{}([&](auto idim_low_1) {
+                    low_dim_hidden_ids_1_mod_(idim_low_1) += adaptor1_hidden_id_shift;
+                });
+
+                // match hidden id
+                static_for<0, ndim_low_1, 1>{}([&](auto idim_low_1) {
+                    static_for<0, ndim_bottom_1, 1>{}([&](auto idim_bottom_1) {
+                        // if this low dim is bottom dim, then do id matching
+                        if constexpr(low_dim_hidden_ids_1[idim_low_1] ==
+                                     TensorAdaptor1::GetBottomDimensionHiddenIds()[idim_bottom_1])
+                        {
+                            low_dim_hidden_ids_1_mod_(idim_low_1) =
+                                TensorAdaptor0::GetTopDimensionHiddenIds()[idim_bottom_1];
+                        }
+                    });
+                });
+
+                return low_dim_hidden_ids_1_mod_;
+            }
+            ();
+
+            return generate_sequence_v2(
+                [&](auto i) constexpr { return Number<low_dim_hidden_ids_1_mod[i]>{}; },
+                Number<ndim_low_1>{});
+        },
+        Number<TensorAdaptor1::GetNumOfTransform()>{});
+
+    constexpr auto all_low_dim_hidden_idss =
+        container_concat(TensorAdaptor0::GetLowerDimensionHiddenIdss(), low_dim_hidden_idss_1);
+
+    // all_up_dim_hidden_idss =
+    // up_dim_hidden_idss_0 + shift_hidden_id_for_1(up_dim_hiden_idss_1)
+    constexpr auto up_dim_hidden_idss_1 = generate_tuple(
+        // generate sequence of ids for a transform
+        [&](auto itran) {
+            constexpr auto ndim_up_1 = TensorAdaptor1::GetUpperDimensionHiddenIdss()[itran].Size();
+
+            constexpr auto up_dim_hidden_ids_1 =
+                TensorAdaptor1::GetUpperDimensionHiddenIdss()[itran];
+
+            // sequence in, constexpr tuple out
+            constexpr auto up_dim_hidden_ids_1_mod = [&]() constexpr
+            {
+                auto up_dim_hidden_ids_1_mod_ = to_multi_index(up_dim_hidden_ids_1);
+
+                // shift hidden id
+                static_for<0, ndim_up_1, 1>{}([&](auto idim_up_1) {
+                    up_dim_hidden_ids_1_mod_(idim_up_1) += adaptor1_hidden_id_shift;
+                });
+
+                return up_dim_hidden_ids_1_mod_;
+            }
+            ();
+
+            // constexpr tuple to sequence
+            return generate_sequence_v2(
+                [&](auto i) constexpr { return Number<up_dim_hidden_ids_1_mod[i]>{}; },
+                Number<ndim_up_1>{});
+        },
+        Number<TensorAdaptor1::GetNumOfTransform()>{});
+
+    constexpr auto all_up_dim_hidden_idss =
+        container_concat(TensorAdaptor0::GetUpperDimensionHiddenIdss(), up_dim_hidden_idss_1);
+
+    // bottom_dim_hidden_ids = bottom_dim_hidden_ids_0
+    constexpr auto bottom_dim_hidden_ids = TensorAdaptor0::GetBottomDimensionHiddenIds();
+
+    // top_dim_hidden_ids = shift_hidden_id(top_dim_hidden_ids_1)
+    constexpr auto top_dim_hidden_ids =
+        TensorAdaptor1::GetTopDimensionHiddenIds() + Number<adaptor1_hidden_id_shift>{};
+
+    // put everything together
+    return TensorAdaptor<remove_cvref_t<decltype(all_transforms)>,
+                         remove_cvref_t<decltype(all_low_dim_hidden_idss)>,
+                         remove_cvref_t<decltype(all_up_dim_hidden_idss)>,
+                         remove_cvref_t<decltype(bottom_dim_hidden_ids)>,
+                         remove_cvref_t<decltype(top_dim_hidden_ids)>>{all_transforms};
+}
+
+template <typename X, typename... Xs, typename enable_if<sizeof...(Xs) >= 2, bool>::type = false>
+__host__ __device__ constexpr auto chain_tensor_adaptors(const X& x, const Xs&... xs)
+{
+    return chain_tensor_adaptors(x, chain_tensor_adaptors(xs...));
+}
+
+} // namespace ck
diff --git a/include/ck/tensor_description/tensor_adaptor_coordinate.hpp b/include/ck/tensor_description/tensor_adaptor_coordinate.hpp
new file mode 100644
index 0000000000..e5d14f04d9
--- /dev/null
+++ b/include/ck/tensor_description/tensor_adaptor_coordinate.hpp
@@ -0,0 +1,249 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+namespace ck {
+
+template <index_t NDimHidden, typename BottomDimensionHiddenIds, typename TopDimensionHiddenIds>
+struct TensorAdaptorCoordinate
+{
+    static constexpr index_t ndim_bottom_ = BottomDimensionHiddenIds::Size();
+    static constexpr index_t ndim_top_    = TopDimensionHiddenIds::Size();
+
+    using HiddenIndex = MultiIndex<NDimHidden>;
+    using BottomIndex = MultiIndex<ndim_bottom_>;
+    using TopIndex    = MultiIndex<ndim_top_>;
+
+    public:
+    __host__ __device__ constexpr TensorAdaptorCoordinate() = default;
+
+    __host__ __device__ constexpr TensorAdaptorCoordinate(const HiddenIndex& idx_hidden)
+        : idx_hidden_{idx_hidden}
+    {
+    }
+
+    __host__ __device__ constexpr auto GetTopIndex() const
+    {
+        return get_container_subset(idx_hidden_, TopDimensionHiddenIds{});
+    }
+
+    __host__ __device__ constexpr auto GetBottomIndex() const
+    {
+        return get_container_subset(idx_hidden_, BottomDimensionHiddenIds{});
+    }
+
+    __host__ __device__ constexpr const auto& GetHiddenIndex() const { return idx_hidden_; }
+
+    __host__ __device__ constexpr auto& GetHiddenIndex() { return idx_hidden_; }
+
+    //
+    HiddenIndex idx_hidden_;
+};
+
+template <typename Adaptor, typename TopIndex>
+__host__ __device__ constexpr auto make_tensor_adaptor_coordinate(const Adaptor& adaptor,
+                                                                  const TopIndex& idx_top)
+{
+    static_assert(Adaptor::GetNumOfTopDimension() == TopIndex::Size(),
+                  "wrong! # of dimension inconsistent");
+
+    constexpr index_t ntransform  = Adaptor::GetNumOfTransform();
+    constexpr index_t ndim_hidden = Adaptor::GetNumOfHiddenDimension();
+    constexpr auto bottom_dim_ids = Adaptor::GetBottomDimensionHiddenIds();
+    constexpr auto top_dim_ids    = Adaptor::GetTopDimensionHiddenIds();
+
+    MultiIndex<ndim_hidden> idx_hidden;
+
+    // initialize visible index
+    set_container_subset(idx_hidden, top_dim_ids, idx_top);
+
+    // calculate hidden index
+    static_for<ntransform, 0, -1>{}([&adaptor, &idx_hidden](auto itran_p1) {
+        auto itran              = itran_p1 - Number<1>{};
+        const auto& tran        = adaptor.GetTransforms().At(itran);
+        constexpr auto dims_low = Adaptor::GetLowerDimensionHiddenIdss().At(itran);
+        constexpr auto dims_up  = Adaptor::GetUpperDimensionHiddenIdss().At(itran);
+
+        const auto idx_up = get_container_subset(idx_hidden, dims_up);
+
+        MultiIndex<dims_low.Size()> idx_low;
+
+        tran.CalculateLowerIndex(idx_low, idx_up);
+
+        set_container_subset(idx_hidden, dims_low, idx_low);
+    });
+
+    return TensorAdaptorCoordinate<ndim_hidden,
+                                   remove_cvref_t<decltype(bottom_dim_ids)>,
+                                   remove_cvref_t<decltype(top_dim_ids)>>{idx_hidden};
+}
+
+template <bool JudgeDoTransforms = true,
+          typename Adaptor,
+          typename AdaptorCoord,
+          typename TopIndex,
+          typename BottomIndex>
+__host__ __device__ constexpr void move_tensor_adaptor_coordinate(const Adaptor& adaptor,
+                                                                  AdaptorCoord& coord,
+                                                                  const TopIndex& idx_diff_top,
+                                                                  BottomIndex& idx_diff_bottom)
+{
+    constexpr index_t ndim_hidden = Adaptor::GetNumOfHiddenDimension();
+    constexpr index_t ndim_top    = Adaptor::GetNumOfTopDimension();
+    //  constexpr index_t ndim_bottom = Adaptor::GetNumOfBottomDimension();
+    constexpr index_t ntransform = Adaptor::GetNumOfTransform();
+
+    //  STATIC_ASSERT(TopIndex::Size() == ndim_top && BottomIndex::Size() == ndim_bottom, "");
+
+    // judge whether calculation of lower diff is needed for each transform
+    // use index_t for boolean type
+    auto do_transforms = make_zero_multi_index<ntransform>();
+
+    if constexpr(JudgeDoTransforms)
+    {
+        auto is_non_zero_diff = make_zero_multi_index<ndim_hidden>();
+
+        // decide do_transform by checkout non-zero index diff components
+        MultiIndex<ndim_top> non_zero_diff_pick_top;
+
+        static_for<0, ndim_top, 1>{}(
+            [&](auto i) { non_zero_diff_pick_top(i) = (idx_diff_top[i] != 0); });
+
+        set_container_subset(
+            is_non_zero_diff, Adaptor::GetTopDimensionHiddenIds(), non_zero_diff_pick_top);
+
+        static_for<ntransform - 1, -1, -1>{}([&](auto itran) {
+            constexpr auto dims_low = Adaptor::GetLowerDimensionHiddenIdss().At(itran);
+            constexpr auto dims_up  = Adaptor::GetUpperDimensionHiddenIdss().At(itran);
+
+            const auto non_zero_diff_pick_up = get_container_subset(is_non_zero_diff, dims_up);
+
+            MultiIndex<dims_low.Size()> non_zero_diff_pick_low;
+
+            // if any of upper index diff components is non-zero, then
+            //   1) Need to do this transform
+            //   2) all components of lower index diff will assume to be non-zero and need to be
+            //   computed
+            const bool idx_diff_up_has_non_zero = container_reduce(
+                non_zero_diff_pick_up, [](auto a, auto b) constexpr { return a or b; }, false);
+
+            do_transforms(itran) = idx_diff_up_has_non_zero;
+
+            static_for<0, dims_low.Size(), 1>{}(
+                [&](auto i) { non_zero_diff_pick_low(i) = idx_diff_up_has_non_zero; });
+
+            set_container_subset(is_non_zero_diff, dims_low, non_zero_diff_pick_low);
+        });
+    }
+    else
+    {
+        static_for<ntransform - 1, -1, -1>{}([&](auto itran) { do_transforms(itran) = 1; });
+    }
+
+    // this is what needs to be calculated
+    auto idx_diff_hidden = make_zero_multi_index<ndim_hidden>();
+
+    // initialize top index diff
+    set_container_subset(idx_diff_hidden, Adaptor::GetTopDimensionHiddenIds(), idx_diff_top);
+
+    // this is what needs to be updated
+    auto& idx_hidden = coord.GetHiddenIndex();
+
+    // update top index
+    auto idx_hidden_pick_top =
+        get_container_subset(idx_hidden, Adaptor::GetTopDimensionHiddenIds());
+
+    idx_hidden_pick_top += idx_diff_top;
+
+    set_container_subset(idx_hidden, Adaptor::GetTopDimensionHiddenIds(), idx_hidden_pick_top);
+
+    // update rest of hidden index
+    static_for<ntransform - 1, -1, -1>{}([&](auto itran) {
+        if(do_transforms[itran])
+        {
+            const auto& tran        = adaptor.GetTransforms().At(itran);
+            constexpr auto dims_low = Adaptor::GetLowerDimensionHiddenIdss().At(itran);
+            constexpr auto dims_up  = Adaptor::GetUpperDimensionHiddenIdss().At(itran);
+
+            const auto idx_up_new  = get_container_subset(idx_hidden, dims_up);
+            auto idx_low           = get_container_subset(idx_hidden, dims_low);
+            const auto idx_diff_up = get_container_subset(idx_diff_hidden, dims_up);
+
+            MultiIndex<dims_low.Size()> idx_diff_low;
+
+            tran.UpdateLowerIndex(idx_diff_low, idx_diff_up, idx_low, idx_up_new);
+
+            set_container_subset(idx_diff_hidden, dims_low, idx_diff_low);
+            set_container_subset(idx_hidden, dims_low, idx_low);
+        }
+    });
+
+    // set bottom index diff
+    idx_diff_bottom = get_container_subset(idx_diff_hidden, Adaptor::GetBottomDimensionHiddenIds());
+}
+
+template <bool JudgeDoTransforms = true, typename Adaptor, typename AdaptorCoord, typename TopIndex>
+__host__ __device__ constexpr void move_tensor_adaptor_coordinate(const Adaptor& adaptor,
+                                                                  AdaptorCoord& coord,
+                                                                  const TopIndex& idx_diff_top)
+{
+    constexpr index_t ndim_bottom = Adaptor::GetNumOfBottomDimension();
+
+    MultiIndex<ndim_bottom> tmp;
+
+    move_tensor_adaptor_coordinate<JudgeDoTransforms>(adaptor, coord, idx_diff_top, tmp);
+}
+
+template <typename Adaptor, typename AdaptorCoord>
+__host__ __device__ constexpr bool
+adaptor_coordinate_is_valid_assuming_top_index_is_valid(const Adaptor& adaptor,
+                                                        const AdaptorCoord& coord)
+{
+    bool valid = true;
+
+    constexpr index_t ntransform = Adaptor::GetNumOfTransform();
+
+    const auto& idx_hidden = coord.GetHiddenIndex();
+
+    static_for<ntransform - 1, -1, -1>{}([&adaptor, &idx_hidden, &valid](auto itran) {
+        const auto tran = adaptor.GetTransforms().At(itran);
+
+        // check validity, only if current transformation does not always has a valid mapping
+        if constexpr(!decltype(tran)::IsValidUpperIndexAlwaysMappedToValidLowerIndex())
+        {
+            const auto idx_up =
+                get_container_subset(idx_hidden, Adaptor::GetUpperDimensionHiddenIdss().At(itran));
+
+            // Comment: using valid = valid && .. will result in weird control flow in ISA
+            valid &= tran.IsValidUpperIndexMappedToValidLowerIndex(idx_up);
+        }
+    });
+
+    return valid;
+}
+
+template <typename Adaptor, typename AdpatorCoord>
+__host__ __device__ constexpr bool adaptor_coordinate_is_valid(const Adaptor& adaptor,
+                                                               const AdpatorCoord& coord)
+{
+    // check top index
+    const auto& idx_top = coord.GetTopIndex();
+
+    bool is_top_index_valid = true;
+
+    static_for<0, Adaptor::GetNumOfDimension(), 1>{}(
+        [&is_top_index_valid, &idx_top, &adaptor](auto i) {
+            is_top_index_valid =
+                is_top_index_valid && (idx_top[i] >= 0 && idx_top[i] < adaptor.GetLength(i));
+        });
+
+    // check other hidden index
+    return is_top_index_valid &&
+           adaptor_coordinate_is_valid_assuming_top_index_is_valid(adaptor, coord);
+}
+
+} // namespace ck
diff --git a/include/ck/tensor_description/tensor_coordinate.hpp b/include/ck/tensor_description/tensor_coordinate.hpp
new file mode 100644
index 0000000000..cb0c986165
--- /dev/null
+++ b/include/ck/tensor_description/tensor_coordinate.hpp
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_adaptor_coordinate.hpp"
+
+namespace ck {
+
+template <index_t NDimHidden, typename TopDimensionHiddenIds>
+struct TensorCoordinate
+    : public TensorAdaptorCoordinate<NDimHidden, Sequence<0>, TopDimensionHiddenIds>
+{
+    using Base = TensorAdaptorCoordinate<NDimHidden, Sequence<0>, TopDimensionHiddenIds>;
+
+    // TODO make these private
+    static constexpr index_t ndim_top_ = TopDimensionHiddenIds::Size();
+
+    using HiddenIndex = MultiIndex<NDimHidden>;
+    using TopIndex    = MultiIndex<ndim_top_>;
+
+    public:
+    __host__ __device__ constexpr TensorCoordinate() = default;
+
+    __host__ __device__ constexpr TensorCoordinate(const HiddenIndex& idx_hidden) : Base{idx_hidden}
+    {
+    }
+
+    // construct from TensorAdaptorCoordinte base class
+    __host__ __device__ constexpr TensorCoordinate(const Base& adaptor_coord) : Base{adaptor_coord}
+    {
+    }
+
+    __host__ __device__ constexpr auto GetIndex() const { return Base::GetTopIndex(); }
+
+    __host__ __device__ constexpr index_t GetOffset() const
+    {
+        return Base::GetBottomIndex()[Number<0>{}];
+    }
+
+    __host__ __device__ constexpr const auto& GetHiddenIndex() const
+    {
+        return Base::GetHiddenIndex();
+    }
+
+    __host__ __device__ auto& GetHiddenIndex() { return Base::GetHiddenIndex(); }
+};
+
+template <typename TensorDesc, typename TopIndex>
+__host__ __device__ constexpr auto make_tensor_coordinate(const TensorDesc& tensor_desc,
+                                                          const TopIndex& idx_top)
+{
+    const auto adaptor_coord = make_tensor_adaptor_coordinate(tensor_desc, idx_top);
+
+    return TensorCoordinate<TensorDesc::GetNumOfHiddenDimension(),
+                            remove_cvref_t<decltype(TensorDesc::GetTopDimensionHiddenIds())>>{
+        adaptor_coord};
+}
+
+template <bool JudgeDoTransforms = true, typename TensorDesc, typename TensorCoord, typename Index>
+__host__ __device__ constexpr void
+move_tensor_coordinate(const TensorDesc& tensor_desc, TensorCoord& coord, const Index& coord_step)
+{
+    move_tensor_adaptor_coordinate(tensor_desc, coord, coord_step);
+}
+
+template <typename TensorDesc, typename TensorCoord>
+__host__ __device__ constexpr bool
+coordinate_has_valid_offset_assuming_top_index_is_valid(const TensorDesc& tensor_desc,
+                                                        const TensorCoord& coord)
+{
+    return adaptor_coordinate_is_valid_assuming_top_index_is_valid(tensor_desc, coord);
+}
+
+template <typename TensorDesc, typename TensorCoord>
+__host__ __device__ constexpr bool coordinate_has_valid_offset(const TensorDesc& tensor_desc,
+                                                               const TensorCoord& coord)
+{
+    return adaptor_coordinate_is_valid(tensor_desc, coord);
+}
+
+} // namespace ck
diff --git a/include/ck/tensor_description/tensor_descriptor.hpp b/include/ck/tensor_description/tensor_descriptor.hpp
new file mode 100644
index 0000000000..bad61678e8
--- /dev/null
+++ b/include/ck/tensor_description/tensor_descriptor.hpp
@@ -0,0 +1,203 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+namespace ck {
+
+// Transforms: Tuple<transforms...>
+// LowerDimensionHiddenIdss : Tuple<Sequence<...>, ...>
+// UpperDimensionHiddenIdss : Tuple<Sequence<...>, ...>
+// TopDimensionHiddenIds> : Sequence<...>
+template <typename Transforms,
+          typename LowerDimensionHiddenIdss,
+          typename UpperDimensionHiddenIdss,
+          typename TopDimensionHiddenIds,
+          typename ElementSpaceSize,
+          typename GuaranteedVectorLengths_,
+          typename GuaranteedVectorSrides_>
+struct TensorDescriptor : public TensorAdaptor<Transforms,
+                                               LowerDimensionHiddenIdss,
+                                               UpperDimensionHiddenIdss,
+                                               Sequence<0>,
+                                               TopDimensionHiddenIds>
+{
+    using Base = TensorAdaptor<Transforms,
+                               LowerDimensionHiddenIdss,
+                               UpperDimensionHiddenIdss,
+                               Sequence<0>,
+                               TopDimensionHiddenIds>;
+
+    using ElementSpaceSizeType = ElementSpaceSize;
+
+    constexpr static index_t ntransform_  = Base::GetNumOfTransform();
+    constexpr static index_t ndim_hidden_ = Base::GetNumOfHiddenDimension();
+    constexpr static index_t ndim_top_    = Base::GetNumOfTopDimension();
+
+    using GuaranteedVectorLengths = GuaranteedVectorLengths_;
+    using GuaranteedVectorStrides = GuaranteedVectorSrides_;
+
+    static_assert(GuaranteedVectorLengths::Size() == ndim_hidden_ &&
+                      GuaranteedVectorStrides::Size() == ndim_hidden_,
+                  "wrong! inconsistent # of hidden dimensions");
+
+    using TopIndex    = MultiIndex<ndim_top_>;
+    using HiddenIndex = MultiIndex<ndim_hidden_>;
+
+    public:
+    __host__ __device__ constexpr TensorDescriptor() = default;
+
+    __host__ __device__ constexpr TensorDescriptor(const Transforms& transforms,
+                                                   ElementSpaceSize element_space_size)
+        : Base{transforms}, element_space_size_{element_space_size}
+
+    {
+        static_assert(Transforms::Size() == ntransform_ &&
+                          LowerDimensionHiddenIdss::Size() == ntransform_ &&
+                          UpperDimensionHiddenIdss::Size() == ntransform_,
+                      "wrong! inconsistent # of transformations");
+
+        // TODO check dependency of dimensions is valid
+    }
+
+    // construct from TensorAdaptor base class
+    __host__ __device__ constexpr TensorDescriptor(const Base& adaptor,
+                                                   ElementSpaceSize element_space_size)
+        : Base{adaptor}, element_space_size_{element_space_size}
+    {
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfDimension()
+    {
+        return Base::GetNumOfTopDimension();
+    }
+
+    template <index_t IDim>
+    __host__ __device__ constexpr auto GetLength(Number<IDim> idim) const
+    {
+        return Base::GetTopDimensionLength(idim);
+    }
+
+    __host__ __device__ constexpr auto GetLengths() const { return Base::GetTopDimensionLengths(); }
+
+    __host__ __device__ constexpr auto GetElementSpaceSize() const { return element_space_size_; }
+
+    template <typename Idx>
+    __host__ __device__ constexpr index_t CalculateOffset(const Idx& idx) const
+    {
+        return Base::CalculateBottomIndex(idx)[Number<0>{}];
+    }
+
+    // TODO make these private
+    __host__ __device__ constexpr const auto& GetTransforms() const
+    {
+        return Base::GetTransforms();
+    }
+
+    __host__ __device__ static constexpr auto GetLowerDimensionHiddenIdss()
+    {
+        return Base::GetLowerDimensionHiddenIdss();
+    }
+
+    __host__ __device__ static constexpr auto GetUpperDimensionHiddenIdss()
+    {
+        return Base::GetUpperDimensionHiddenIdss();
+    }
+
+    __host__ __device__ static constexpr auto GetTopDimensionHiddenIds()
+    {
+        return Base::GetTopDimensionHiddenIds();
+    }
+
+    __host__ __device__ static constexpr bool IsStatic()
+    {
+        return Base::IsKnownAtCompileTime() && is_known_at_compile_time<ElementSpaceSize>::value;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime() { return IsStatic(); }
+
+    __host__ __device__ static constexpr auto GetTopDimensionSafeVectorLengthStrides()
+    {
+        return Base::GetTopDimensionSafeVectorLengthStrides(
+            to_array<index_t, ndim_hidden_>(GuaranteedVectorLengths{}),
+            to_array<index_t, ndim_hidden_>(GuaranteedVectorStrides{}));
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("TensorDescriptor{");
+
+        // TensorAdaptor
+        Base::Print();
+        printf(", ");
+
+        // element_space_size_
+        printf("element_space_size_: ");
+        print(element_space_size_);
+
+        printf("}");
+    }
+
+    // TODO make these private
+    ElementSpaceSize element_space_size_;
+};
+
+template <typename Adaptor, typename ElementSpaceSize>
+__host__ __device__ constexpr auto
+make_tensor_descriptor_from_adaptor(const Adaptor& adaptor,
+                                    const ElementSpaceSize& element_space_size)
+{
+    constexpr index_t NDimHidden = Adaptor::GetNumOfHiddenDimension();
+
+    return TensorDescriptor<remove_cvref_t<decltype(adaptor.GetTransforms())>,
+                            remove_cvref_t<decltype(adaptor.GetLowerDimensionHiddenIdss())>,
+                            remove_cvref_t<decltype(adaptor.GetUpperDimensionHiddenIdss())>,
+                            remove_cvref_t<decltype(adaptor.GetTopDimensionHiddenIds())>,
+                            remove_cvref_t<decltype(element_space_size)>,
+                            typename uniform_sequence_gen<NDimHidden, -1>::type,
+                            typename uniform_sequence_gen<NDimHidden, -1>::type>{
+        adaptor, element_space_size};
+}
+
+template <typename OldTensorDescriptor,
+          typename NewTransforms,
+          typename NewLowerDimensionOldTopIdss,
+          typename NewUpperDimensionNewTopIdss>
+__host__ __device__ constexpr auto
+transform_tensor_descriptor(const OldTensorDescriptor& old_tensor_desc,
+                            const NewTransforms& new_transforms,
+                            NewLowerDimensionOldTopIdss,
+                            NewUpperDimensionNewTopIdss)
+{
+    const auto element_space_size = old_tensor_desc.GetElementSpaceSize();
+
+    const auto new_tensor_adaptor = transform_tensor_adaptor(old_tensor_desc,
+                                                             new_transforms,
+                                                             NewLowerDimensionOldTopIdss{},
+                                                             NewUpperDimensionNewTopIdss{});
+
+    constexpr index_t NDimHiddenOld = OldTensorDescriptor::GetNumOfHiddenDimension();
+    constexpr index_t NDimHiddenNew = decltype(new_tensor_adaptor)::GetNumOfHiddenDimension();
+
+    using NewGuaranteedVectorLengths = typename sequence_merge<
+        typename OldTensorDescriptor::GuaranteedVectorLengths,
+        typename uniform_sequence_gen<NDimHiddenNew - NDimHiddenOld, -1>::type>::type;
+
+    using NewGuaranteedVectorStrides = typename sequence_merge<
+        typename OldTensorDescriptor::GuaranteedVectorStrides,
+        typename uniform_sequence_gen<NDimHiddenNew - NDimHiddenOld, -1>::type>::type;
+
+    return TensorDescriptor<
+        remove_cvref_t<decltype(new_tensor_adaptor.GetTransforms())>,
+        remove_cvref_t<decltype(new_tensor_adaptor.GetLowerDimensionHiddenIdss())>,
+        remove_cvref_t<decltype(new_tensor_adaptor.GetUpperDimensionHiddenIdss())>,
+        remove_cvref_t<decltype(new_tensor_adaptor.GetTopDimensionHiddenIds())>,
+        remove_cvref_t<decltype(element_space_size)>,
+        NewGuaranteedVectorLengths,
+        NewGuaranteedVectorStrides>{new_tensor_adaptor, element_space_size};
+}
+
+} // namespace ck
diff --git a/include/ck/tensor_description/tensor_descriptor_helper.hpp b/include/ck/tensor_description/tensor_descriptor_helper.hpp
new file mode 100644
index 0000000000..c49cc91c09
--- /dev/null
+++ b/include/ck/tensor_description/tensor_descriptor_helper.hpp
@@ -0,0 +1,167 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+
+namespace ck {
+
+namespace detail {
+
+template <typename Lengths, typename Strides, index_t I, typename AccOld>
+__host__ __device__ constexpr auto calculate_element_space_size_impl(const Lengths& lengths,
+                                                                     const Strides& strides,
+                                                                     Number<I> i,
+                                                                     AccOld acc_old)
+{
+    auto acc_new = acc_old + (lengths[i] - Number<1>{}) * strides[i];
+
+    if constexpr(i.value < Lengths::Size() - 1)
+    {
+        return calculate_element_space_size_impl(lengths, strides, i + Number<1>{}, acc_new);
+    }
+    else
+    {
+        return acc_new;
+    }
+}
+
+} // namespace detail
+
+/*
+ * These functions create naive tensor descriptor
+ */
+
+// Lengths..., Strides... could be:
+//   1) index_t, which is known at run-time, or
+//   2) Number<>, which is known at compile-time
+// element_space_size could be:
+//   1) long_index_t, or
+//   2) LongNumber<>
+template <typename... Lengths,
+          typename... Strides,
+          index_t GuaranteedLastDimensionVectorLength                              = -1,
+          index_t GuaranteedLastDimensionVectorStride                              = -1,
+          typename enable_if<sizeof...(Lengths) == sizeof...(Strides), bool>::type = false>
+__host__ __device__ constexpr auto
+make_naive_tensor_descriptor(const Tuple<Lengths...>& lengths,
+                             const Tuple<Strides...>& strides,
+                             Number<GuaranteedLastDimensionVectorLength> = Number<-1>{},
+                             Number<GuaranteedLastDimensionVectorStride> = Number<-1>{})
+{
+    constexpr index_t N = sizeof...(Lengths);
+
+    const auto transforms = make_tuple(make_embed_transform(lengths, strides));
+
+    constexpr auto low_dim_hidden_idss = make_tuple(Sequence<0>{});
+
+    constexpr auto up_dim_hidden_idss =
+        make_tuple(typename arithmetic_sequence_gen<1, N + 1, 1>::type{});
+
+    constexpr auto visible_dim_hidden_ids = typename arithmetic_sequence_gen<1, N + 1, 1>::type{};
+
+    const auto element_space_size =
+        detail::calculate_element_space_size_impl(lengths, strides, Number<0>{}, LongNumber<1>{});
+
+    using GuaranteedVectorLengths =
+        typename sequence_merge<typename uniform_sequence_gen<N, -1>::type,
+                                Sequence<GuaranteedLastDimensionVectorLength>>::type;
+
+    using GuaranteedVectorStrides =
+        typename sequence_merge<typename uniform_sequence_gen<N, -1>::type,
+                                Sequence<GuaranteedLastDimensionVectorStride>>::type;
+
+    return TensorDescriptor<remove_cv_t<decltype(transforms)>,
+                            remove_cv_t<decltype(low_dim_hidden_idss)>,
+                            remove_cv_t<decltype(up_dim_hidden_idss)>,
+                            remove_cv_t<decltype(visible_dim_hidden_ids)>,
+                            remove_cv_t<decltype(element_space_size)>,
+                            GuaranteedVectorLengths,
+                            GuaranteedVectorStrides>{transforms, element_space_size};
+}
+
+// Lengths... could be:
+//   1) index_t, which is known at run-time, or
+//   2) Number<>, which is known at compile-time
+// element_space_size could be:
+//   1) long_index_t, or
+//   2) LongNumber<>
+template <typename... Lengths, index_t GuaranteedLastDimensionVectorLength = -1>
+__host__ __device__ constexpr auto
+make_naive_tensor_descriptor_packed(const Tuple<Lengths...>& lengths,
+                                    Number<GuaranteedLastDimensionVectorLength> = Number<-1>{})
+{
+    constexpr index_t N = sizeof...(Lengths);
+
+    const auto transforms = make_tuple(make_unmerge_transform(lengths));
+
+    constexpr auto low_dim_hidden_idss = make_tuple(Sequence<0>{});
+
+    constexpr auto up_dim_hidden_idss =
+        make_tuple(typename arithmetic_sequence_gen<1, N + 1, 1>::type{});
+
+    constexpr auto visible_dim_hidden_ids = typename arithmetic_sequence_gen<1, N + 1, 1>::type{};
+
+    const auto element_space_size = container_reduce(lengths, math::multiplies{}, LongNumber<1>{});
+
+    using GuaranteedVectorLengths =
+        typename sequence_merge<typename uniform_sequence_gen<N, -1>::type,
+                                Sequence<GuaranteedLastDimensionVectorLength>>::type;
+
+    using GuaranteedVectorStrides =
+        typename sequence_merge<typename uniform_sequence_gen<N, -1>::type, Sequence<1>>::type;
+
+    return TensorDescriptor<remove_cv_t<decltype(transforms)>,
+                            remove_cv_t<decltype(low_dim_hidden_idss)>,
+                            remove_cv_t<decltype(up_dim_hidden_idss)>,
+                            remove_cv_t<decltype(visible_dim_hidden_ids)>,
+                            remove_cv_t<decltype(element_space_size)>,
+                            GuaranteedVectorLengths,
+                            GuaranteedVectorStrides>{transforms, element_space_size};
+}
+
+// Lengths... could be:
+//   1) index_t, which is known at run-time, or
+//   2) Number<>, which is known at compile-time
+// align could be:
+//   1) index_t, or
+//   2) Number<>
+template <typename... Lengths, typename Align>
+__host__ __device__ constexpr auto
+make_naive_tensor_descriptor_aligned(const Tuple<Lengths...>& lengths, Align align)
+{
+    constexpr auto I1 = Number<1>{};
+
+    constexpr index_t N = sizeof...(Lengths);
+
+    const auto stride_n_minus_2 = math::integer_least_multiple(lengths[Number<N - 1>{}], align);
+
+    auto strides = generate_tuple(
+        [&](auto i) {
+            if constexpr(i.value == N - 1)
+            {
+                return I1;
+            }
+            else if constexpr(i.value == N - 2)
+            {
+                return Number<stride_n_minus_2>{};
+            }
+            else
+            {
+                return container_reduce(lengths,
+                                        math::multiplies{},
+                                        Number<stride_n_minus_2>{},
+                                        i + I1,
+                                        Number<N - 1>{},
+                                        I1);
+            }
+        },
+        Number<N>{});
+
+    return make_naive_tensor_descriptor(lengths, strides);
+}
+
+} // namespace ck
diff --git a/include/ck/tensor_description/tensor_space_filling_curve.hpp b/include/ck/tensor_description/tensor_space_filling_curve.hpp
new file mode 100644
index 0000000000..c7adf69a51
--- /dev/null
+++ b/include/ck/tensor_description/tensor_space_filling_curve.hpp
@@ -0,0 +1,162 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/math.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/sequence_helper.hpp"
+#include "ck/utility/multi_index.hpp"
+#include "ck/utility/tuple_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+namespace ck {
+
+template <typename TensorLengths,
+          typename DimAccessOrder,
+          typename ScalarsPerAccess,
+          bool SnakeCurved = true> // # of scalars per access in each dimension
+struct SpaceFillingCurve
+{
+    static constexpr index_t nDim = TensorLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    static constexpr index_t ScalarPerVector =
+        reduce_on_sequence(ScalarsPerAccess{}, math::multiplies{}, Number<1>{});
+
+    static constexpr auto access_lengths   = TensorLengths{} / ScalarsPerAccess{};
+    static constexpr auto dim_access_order = DimAccessOrder{};
+    static constexpr auto ordered_access_lengths =
+        container_reorder_given_new2old(access_lengths, dim_access_order);
+
+    static constexpr auto to_index_adaptor = make_single_stage_tensor_adaptor(
+        make_tuple(make_merge_transform(ordered_access_lengths)),
+        make_tuple(typename arithmetic_sequence_gen<0, nDim, 1>::type{}),
+        make_tuple(Sequence<0>{}));
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    __host__ __device__ static constexpr index_t GetNumOfAccess()
+    {
+        static_assert(TensorLengths::Size() == ScalarsPerAccess::Size());
+        static_assert(TensorLengths{} % ScalarsPerAccess{} ==
+                      typename uniform_sequence_gen<TensorLengths::Size(), 0>::type{});
+
+        return reduce_on_sequence(TensorLengths{}, math::multiplies{}, Number<1>{}) /
+               ScalarPerVector;
+    }
+
+    template <index_t AccessIdx1dHead, index_t AccessIdx1dTail>
+    static __device__ __host__ constexpr auto GetStepBetween(Number<AccessIdx1dHead>,
+                                                             Number<AccessIdx1dTail>)
+    {
+        static_assert(AccessIdx1dHead >= 0 && AccessIdx1dHead < GetNumOfAccess(),
+                      "1D index out of range");
+        static_assert(AccessIdx1dTail >= 0 && AccessIdx1dTail < GetNumOfAccess(),
+                      "1D index out of range");
+
+        constexpr auto idx_head = GetIndex(Number<AccessIdx1dHead>{});
+        constexpr auto idx_tail = GetIndex(Number<AccessIdx1dTail>{});
+        return idx_tail - idx_head;
+    }
+
+    template <index_t AccessIdx1d>
+    static __device__ __host__ constexpr auto GetForwardStep(Number<AccessIdx1d>)
+    {
+        static_assert(AccessIdx1d < GetNumOfAccess(), "1D index should be larger than 0");
+        return GetStepBetween(Number<AccessIdx1d>{}, Number<AccessIdx1d + 1>{});
+    }
+
+    template <index_t AccessIdx1d>
+    static __device__ __host__ constexpr auto GetBackwardStep(Number<AccessIdx1d>)
+    {
+        static_assert(AccessIdx1d > 0, "1D index should be larger than 0");
+
+        return GetStepBetween(Number<AccessIdx1d>{}, Number<AccessIdx1d - 1>{});
+    }
+
+    template <index_t AccessIdx1d>
+    static __device__ __host__ constexpr Index GetIndex(Number<AccessIdx1d>)
+    {
+#if 0
+        /*
+         * \todo: TensorAdaptor::CalculateBottomIndex does NOT return constexpr as expected.
+         */
+        constexpr auto ordered_access_idx = to_index_adaptor.CalculateBottomIndex(make_multi_index(Number<AccessIdx1d>{}));
+#else
+
+        constexpr auto access_strides = container_reverse_exclusive_scan(
+            ordered_access_lengths, math::multiplies{}, Number<1>{});
+
+        constexpr auto idx_1d = Number<AccessIdx1d>{};
+        // Given tensor strides \p access_lengths, and 1D index of space-filling-curve, compute the
+        // idim-th element of multidimensional index.
+        // All constexpr variables have to be captured by VALUE.
+        constexpr auto compute_index = [ idx_1d, access_strides ](auto idim) constexpr
+        {
+            constexpr auto compute_index_impl = [ idx_1d, access_strides ](auto jdim) constexpr
+            {
+                auto res = idx_1d.value;
+                auto id  = 0;
+
+                static_for<0, jdim.value + 1, 1>{}([&](auto kdim) {
+                    id = res / access_strides[kdim].value;
+                    res -= id * access_strides[kdim].value;
+                });
+
+                return id;
+            };
+
+            constexpr auto id = compute_index_impl(idim);
+            return Number<id>{};
+        };
+
+        constexpr auto ordered_access_idx = generate_tuple(compute_index, Number<nDim>{});
+#endif
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto idim) {
+                index_t tmp = ordered_access_idx[I0];
+
+                static_for<1, idim, 1>{}(
+                    [&](auto j) { tmp = tmp * ordered_access_lengths[j] + ordered_access_idx[j]; });
+
+                forward_sweep_(idim) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate multi-dim tensor index
+        auto idx_md = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto idim) {
+                ordered_idx(idim) =
+                    !SnakeCurved || forward_sweep[idim]
+                        ? ordered_access_idx[idim]
+                        : ordered_access_lengths[idim] - 1 - ordered_access_idx[idim];
+            });
+
+            return container_reorder_given_old2new(ordered_idx, dim_access_order) *
+                   ScalarsPerAccess{};
+        }();
+        return idx_md;
+    }
+
+    // FIXME: rename this function
+    template <index_t AccessIdx1d>
+    static __device__ __host__ constexpr auto GetIndexTupleOfNumber(Number<AccessIdx1d>)
+    {
+        constexpr auto idx = GetIndex(Number<AccessIdx1d>{});
+
+        return generate_tuple([&](auto i) { return Number<idx[i]>{}; }, Number<nDim>{});
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_dpp8.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_dpp8.hpp
new file mode 100644
index 0000000000..e527509f57
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_dpp8.hpp
@@ -0,0 +1,370 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/amd_gemm_dpp.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_contraction_dl_dpp8.hpp"
+
+namespace ck {
+
+/**
+ * DPP8 version of blockwise GEMM algorithm. It uses DPP8 instruction modifier to limit
+ * the data loaded from LDS to registers.
+ *
+ * The algorithm groups threads into groups of size `dpp8::lane_group_size` and splits the matrix C
+ * between them in such a way that threads from the same group need the same chunk of either
+ * matrix A (or B, respectively). Without the usage of DPP8, each thread would need to load the
+ * whole chunk from LDS to its own register space.
+ * Usage of DPP8 modifiers allow each thread to load less data, exactly `1 / dpp8::lane_group_size`
+ * of the chunk, and then share that data with other threads from the same lane group.
+ *
+ * Assumptions coming from the usage of DPP8:
+ *   1. `BM10BN10ThreadClusterBM10Xs[1] == dpp8::lane_group_size` or
+ *      `BM10BN10ThreadClusterBN10Xs[1] == dpp8::lane_group_size` -
+ *        - it makes consecutive `dpp8::lane_group_size` threads use the same chunk of either
+ *          matrix A or B;
+ *        - based on these values we determine which matrix to share.
+ *   2. `BM1PerThreadBM11 % dpp8::lane_group_size == 0` (if sharing A) or
+ *      `BN1PerThreadBN11 % dpp8::lane_group_size == 0` (if sharing B) -
+ *        - we have to make sure that the data to split is divisible by the number of
+ *          threads in the group.
+ *
+ * General algorithm:
+ * C[BM0, BM1, BN0, BN1] += transpose(A[K, BM0, BM1]) * B[K, BN0, BN1]
+ * A and B are visible to the whole block, C is distributed among each thread
+ * Assume:
+ *   1. A:
+ *     1. ABlockDesc_BK0_BM_BK1 is known at compile-time
+ *     2. ABlockBuffer is DynamicBuffer
+ *   2. B:
+ *     1. BBlockDesc_BK0_BN_BK1 is known at compile-time
+ *     2. BBlockBuffer is DynamicBuffer
+ *   3. C:
+ *     1. CThreadDesc_BM0_BM11_BN0_BN11 is known at compile-time
+ *     2. CThreadBuffer is StaticBuffer
+ *   4. BM10BN10ThreadClusterBM10Xs::Size() = BM10BN10ThreadClusterBN10Xs::Size() == 2
+ */
+template <index_t BlockSize,
+          typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename ABlockDesc_BK0_BM_BK1,
+          typename BBlockDesc_BK0_BN_BK1,
+          index_t BM1PerThreadBM11,
+          index_t BN1PerThreadBN11,
+          index_t BK0PerThread,
+          typename BM10BN10ThreadClusterBM10Xs, // Sequence<BM10BN10ThreadClusterBM100,
+                                                //          BM10BN10ThreadClusterBM101, ...>
+          typename BM10BN10ThreadClusterBN10Xs, // Sequence<BM10BN10ThreadClusterBN100,
+                                                //          BM10BN10ThreadClusterBN101, ...>
+          index_t AThreadCopyScalarPerVector_BM11,
+          index_t BThreadCopyScalarPerVector_BN11,
+          typename enable_if<ABlockDesc_BK0_BM_BK1::IsKnownAtCompileTime() &&
+                                 BBlockDesc_BK0_BN_BK1::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct BlockwiseGemmDlDpp8_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_loop_BM0_BN0
+{
+    using AIndex = MultiIndex<4>;
+    using BIndex = MultiIndex<4>;
+    using CIndex = MultiIndex<4>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr index_t BK0 = ABlockDesc_BK0_BM_BK1{}.GetLength(I0);
+    static constexpr index_t BK1 = ABlockDesc_BK0_BM_BK1{}.GetLength(I2);
+    static constexpr index_t BM  = ABlockDesc_BK0_BM_BK1{}.GetLength(I1);
+    static constexpr index_t BN  = BBlockDesc_BK0_BN_BK1{}.GetLength(I1);
+
+    static constexpr index_t BM100 = BM10BN10ThreadClusterBM10Xs{}[I0];
+    static constexpr index_t BN100 = BM10BN10ThreadClusterBN10Xs{}[I0];
+
+    static constexpr index_t BM101 = BM10BN10ThreadClusterBM10Xs{}[I1];
+    static constexpr index_t BN101 = BM10BN10ThreadClusterBN10Xs{}[I1];
+
+    static constexpr index_t BM11 = BM1PerThreadBM11;
+    static constexpr index_t BN11 = BN1PerThreadBN11;
+
+    static constexpr index_t BM1 = BM100 * BM101 * BM11;
+    static constexpr index_t BN1 = BN100 * BN101 * BN11;
+
+    static constexpr index_t BM0 = BM / BM1;
+    static constexpr index_t BN0 = BN / BN1;
+
+    // We assume that either `BM101` or `BN101` is equal to `dpp8::lane_group_size`. It makes all
+    // threads in a lane group need the same chunk of B or A matrices and we can share them using
+    // DPP.
+    static_assert(BM101 == dpp8::lane_group_size || BN101 == dpp8::lane_group_size);
+    static constexpr bool ShareB = BM101 == dpp8::lane_group_size ? true : false;
+    static constexpr bool ShareA = !ShareB;
+
+    // If DPP shares A (B, respectively), lane group gets `BM1PerThreadBM11` (`BN1PerThreadBN11`,
+    // respectively) elements, so we split them between threads in lane group so each thread loads
+    // less data from LDS.
+    static constexpr index_t BM1PerThread =
+        ShareA ? BM1PerThreadBM11 / dpp8::lane_group_size : BM1PerThreadBM11;
+    static constexpr index_t BN1PerThread =
+        ShareB ? BN1PerThreadBN11 / dpp8::lane_group_size : BN1PerThreadBN11;
+
+    __host__ __device__ static constexpr auto
+    MakeABlockDescriptor_BK0_BM0_BM1_BK1(const ABlockDesc_BK0_BM_BK1& a_block_desc_bk0_bm_bk1)
+    {
+        const auto a_block_bk0_bm0_bm1_bk1 = transform_tensor_descriptor(
+            a_block_desc_bk0_bm_bk1,
+            make_tuple(make_pass_through_transform(Number<BK0>{}),
+                       make_unmerge_transform(make_tuple(Number<BM0>{}, Number<BM1>{})),
+                       make_pass_through_transform(Number<BK1>{})),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+        return a_block_bk0_bm0_bm1_bk1;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeBBlockDescriptor_BK0_BN0_BN1_BK1(const BBlockDesc_BK0_BN_BK1& b_block_desc_bk0_bn_bk1)
+    {
+        const auto b_block_desc_bk0_bn0_bn1_bk1 = transform_tensor_descriptor(
+            b_block_desc_bk0_bn_bk1,
+            make_tuple(make_pass_through_transform(Number<BK0>{}),
+                       make_unmerge_transform(make_tuple(Number<BN0>{}, Number<BN1>{})),
+                       make_pass_through_transform(Number<BK1>{})),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+        return b_block_desc_bk0_bn0_bn1_bk1;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCBlockAdaptor_BM0_BM100_BM101_BM11_BN0_BN100_BN101_BN11_To_BM_BN()
+    {
+        // upper: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
+        // lower: [BM, BN]
+        constexpr auto c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m_n =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_unmerge_transform(make_tuple(
+                               Number<BM0>{}, Number<BM100>{}, Number<BM101>{}, Number<BM11>{})),
+                           make_unmerge_transform(make_tuple(
+                               Number<BN0>{}, Number<BN100>{}, Number<BN101>{}, Number<BN11>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4, 5, 6, 7>{}));
+
+        return c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m_n;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCBlockAdaptor_BM0_BM100_BM101_BM11_BN0_BN100_BN101_BN11_To_BM0_BM1_BN0_BN1()
+    {
+        // upper: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
+        // lower: [BM0, BM1, BN0, BN1]
+        constexpr auto c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m0_m1_n0_n1 =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_pass_through_transform(Number<BM0>{}),
+                           make_unmerge_transform(
+                               make_tuple(Number<BM100>{}, Number<BM101>{}, Number<BM11>{})),
+                           make_pass_through_transform(Number<BN0>{}),
+                           make_unmerge_transform(
+                               make_tuple(Number<BN100>{}, Number<BN101>{}, Number<BN11>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}, Sequence<5, 6, 7>{}));
+
+        return c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m0_m1_n0_n1;
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadTensorLengths_BM0_BM1_BN0_BN1()
+    {
+        return Sequence<BM0, BM11, BN0, BN11>{};
+    }
+
+    static constexpr auto a_block_desc_bk0_bm0_bm1_bk1_ =
+        MakeABlockDescriptor_BK0_BM0_BM1_BK1(ABlockDesc_BK0_BM_BK1{});
+
+    static constexpr auto b_block_desc_bk0_bn0_bn1_bk1_ =
+        MakeBBlockDescriptor_BK0_BN0_BN1_BK1(BBlockDesc_BK0_BN_BK1{});
+
+    public:
+    __device__ BlockwiseGemmDlDpp8_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_loop_BM0_BN0()
+        : c_thread_origin_data_idx_{CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
+              get_thread_local_1d_id())},
+          a_thread_copy_{CalculateAThreadOriginOnBlock_BK0_BM0_BM1_BK1()},
+          b_thread_copy_{CalculateBThreadOriginOnBlock_BK0_BN0_BN1_BK1()}
+    {
+        static_assert(ABlockDesc_BK0_BM_BK1::IsKnownAtCompileTime() &&
+                          BBlockDesc_BK0_BN_BK1::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(BM % BM1 == 0 && BN % BN1 == 0, "wrong!");
+
+        static_assert(ABlockDesc_BK0_BM_BK1{}.GetLength(I0) ==
+                          BBlockDesc_BK0_BN_BK1{}.GetLength(I0),
+                      "wrong! K dimension not consistent");
+
+        static_assert(BM10BN10ThreadClusterBM10Xs::Size() == 2 &&
+                          BM10BN10ThreadClusterBN10Xs::Size() == 2,
+                      "wrong!");
+    }
+
+    __device__ static CIndex CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(index_t thread_id)
+    {
+        // lower: [BM0, BM1, BN0, BN1]
+        // upper: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
+        constexpr auto adaptor0 =
+            MakeCBlockAdaptor_BM0_BM100_BM101_BM11_BN0_BN100_BN101_BN11_To_BM0_BM1_BN0_BN1();
+
+        // lower: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
+        // upper: [Tid, BM0, BM11, BN0, BN11]
+        constexpr auto adaptor1 = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(BM100, BN100, BM101, BN101)),
+                       make_pass_through_transform(BM0),
+                       make_pass_through_transform(BM11),
+                       make_pass_through_transform(BN0),
+                       make_pass_through_transform(BN11)),
+            make_tuple(
+                Sequence<1, 5, 2, 6>{}, Sequence<0>{}, Sequence<3>{}, Sequence<4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+        constexpr auto adaptor = chain_tensor_adaptors(adaptor0, adaptor1);
+
+        return adaptor.CalculateBottomIndex(make_multi_index(thread_id, 0, 0, 0, 0));
+    }
+
+    __device__ AIndex CalculateAThreadOriginOnBlock_BK0_BM0_BM1_BK1()
+    {
+        const auto offsetBM0 = c_thread_origin_data_idx_[I0];
+        // If sharing matrix A, we need a separate BM1 offset for each thread in lane group.
+        const auto offsetBM1 = ShareA ? c_thread_origin_data_idx_[I1] +
+                                            dpp8::get_thread_idx_in_lane_group() * BM1PerThread
+                                      : c_thread_origin_data_idx_[I1];
+        return make_tuple(0, offsetBM0, offsetBM1, 0);
+    }
+
+    __device__ BIndex CalculateBThreadOriginOnBlock_BK0_BN0_BN1_BK1()
+    {
+        const auto offsetBN0 = c_thread_origin_data_idx_[I2];
+        // If sharing matrix B, we need a separate BN1 offset for each thread in lane group.
+        const auto offsetBN1 = ShareB ? c_thread_origin_data_idx_[I3] +
+                                            dpp8::get_thread_idx_in_lane_group() * BN1PerThread
+                                      : c_thread_origin_data_idx_[I3];
+        return make_tuple(0, offsetBN0, offsetBN1, 0);
+    }
+
+    template <typename CThreadDesc_BM0_BM11_BN0_BN11,
+              typename ABlockBuffer,
+              typename BBlockBuffer,
+              typename CThreadBuffer>
+    __device__ void Run(const CThreadDesc_BM0_BM11_BN0_BN11&,
+                        const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_block_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        static_assert(CThreadDesc_BM0_BM11_BN0_BN11::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
+            a_thread_desc_bk0_bm0_bm1_bk1_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
+            b_thread_desc_bk0_bn0_bn1_bk1_.GetElementSpaceSize());
+
+        constexpr auto threadwise_contraction =
+            ThreadwiseContractionDlDpp8_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1<
+                FloatA,
+                FloatB,
+                FloatC,
+                decltype(a_thread_desc_bk0_bm0_bm1_bk1_),
+                decltype(b_thread_desc_bk0_bn0_bn1_bk1_),
+                CThreadDesc_BM0_BM11_BN0_BN11,
+                Sequence<BK0PerThread, BK1>,
+                Sequence<1, BM1PerThreadBM11>,
+                Sequence<1, BN1PerThreadBN11>,
+                ShareA>{};
+
+        static_for<0, BN0, 1>{}([&](auto bn0) {
+            static_for<0, BM0, 1>{}([&](auto bm0) {
+                a_thread_copy_.Run(a_block_desc_bk0_bm0_bm1_bk1_,
+                                   make_tuple(I0, bm0, I0, I0),
+                                   a_block_buf,
+                                   a_thread_desc_bk0_bm0_bm1_bk1_,
+                                   make_tuple(I0, I0, I0, I0),
+                                   a_thread_buf);
+
+                b_thread_copy_.Run(b_block_desc_bk0_bn0_bn1_bk1_,
+                                   make_tuple(I0, bn0, I0, I0),
+                                   b_block_buf,
+                                   b_thread_desc_bk0_bn0_bn1_bk1_,
+                                   make_tuple(I0, I0, I0, I0),
+                                   b_thread_buf);
+
+                threadwise_contraction.Run(a_thread_buf,
+                                           make_tuple(I0, I0, I0, I0),
+                                           b_thread_buf,
+                                           make_tuple(I0, I0, I0, I0),
+                                           c_thread_buf,
+                                           make_tuple(bm0, I0, bn0, I0));
+
+                static_for<BK0PerThread, BK0, BK0PerThread>{}([&](auto bk0) {
+                    a_thread_copy_.Run(a_block_desc_bk0_bm0_bm1_bk1_,
+                                       make_tuple(bk0, bm0, I0, I0),
+                                       a_block_buf,
+                                       a_thread_desc_bk0_bm0_bm1_bk1_,
+                                       make_tuple(I0, I0, I0, I0),
+                                       a_thread_buf);
+
+                    b_thread_copy_.Run(b_block_desc_bk0_bn0_bn1_bk1_,
+                                       make_tuple(bk0, bn0, I0, I0),
+                                       b_block_buf,
+                                       b_thread_desc_bk0_bn0_bn1_bk1_,
+                                       make_tuple(I0, I0, I0, I0),
+                                       b_thread_buf);
+
+                    threadwise_contraction.Run(a_thread_buf,
+                                               make_tuple(I0, I0, I0, I0),
+                                               b_thread_buf,
+                                               make_tuple(I0, I0, I0, I0),
+                                               c_thread_buf,
+                                               make_tuple(bm0, I0, bn0, I0));
+                });
+            });
+        });
+    }
+
+    private:
+    // A[BK0, BM0, BM1, BK1]
+    static constexpr auto a_thread_desc_bk0_bm0_bm1_bk1_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<BK0PerThread>{}, Number<BM0>{}, Number<BM1PerThread>{}, Number<BK1>{}));
+
+    // B[BK0, BN0, BN1, BK1]
+    static constexpr auto b_thread_desc_bk0_bn0_bn1_bk1_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<BK0PerThread>{}, Number<BN0>{}, Number<BN1PerThread>{}, Number<BK1>{}));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4r1<
+        FloatA,
+        FloatA,
+        decltype(a_block_desc_bk0_bm0_bm1_bk1_),
+        decltype(a_thread_desc_bk0_bm0_bm1_bk1_),
+        Sequence<BK0PerThread, 1, BM1PerThread, BK1>, // SliceLengths
+        Sequence<0, 1, 2, 3>,                         // DimAccessOrder
+        Sequence<1, 1, BM1PerThread, BK1>,            // SrcVectorTensorLengths
+        Sequence<0, 1, 2, 3>>;                        // SrcVectorTensorContiguousDimOrder
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4r1<
+        FloatB,
+        FloatB,
+        decltype(b_block_desc_bk0_bn0_bn1_bk1_),
+        decltype(b_thread_desc_bk0_bn0_bn1_bk1_),
+        Sequence<BK0PerThread, 1, BN1PerThread, BK1>, // SliceLengths
+        Sequence<0, 1, 2, 3>,                         // DimAccessOrder
+        Sequence<1, 1, BN1PerThread, BK1>,            // SrcVectorTensorLengths
+        Sequence<0, 1, 2, 3>>;                        // SrcVectorTensorContiguousDimOrder
+
+    CIndex c_thread_origin_data_idx_;
+
+    AThreadCopy a_thread_copy_;
+    BThreadCopy b_thread_copy_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp
new file mode 100644
index 0000000000..f23404a1d7
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp
@@ -0,0 +1,412 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp"
+
+namespace ck {
+
+// C[BM0, BM1, BN0, BN1] += transpose(A[K, BM0, BM1]) * B[K, BN0, BN1]
+// A and B are visible to the whole block, C is distributed among each thread
+// Assume:
+//   1. A:
+//     1. ABlockDesc_BK0_BM_BK1 is known at compile-time
+//     2. ABlockBuffer is DynamicBuffer
+//   2. B:
+//     1. BBlockDesc_BK0_BN_BK1 is known at compile-time
+//     2. BBlockBuffer is DynamicBuffer
+//   3. C:
+//     1. CThreadDesc_BM0_BM11_BN0_BN11 is known at compile-time
+//     2. CThreadBuffer is StaticBuffer
+// Also assume:
+//   BM10BN10ThreadClusterBM10Xs::Size() = BM10BN10ThreadClusterBN10Xs::Size() == 2
+//   BM0 = BN0 = 2. It will do 2x2 pipelined read and fma (ABBA optimization)
+template <index_t BlockSize,
+          typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename ABlockDesc_BK0_BM_BK1,
+          typename BBlockDesc_BK0_BN_BK1,
+          index_t BM1PerThreadBM11,
+          index_t BN1PerThreadBN11,
+          index_t BK0PerThread,
+          typename BM10BN10ThreadClusterBM10Xs, // Sequence<BM10BN10ThreadClusterBM100,
+                                                //          BM10BN10ThreadClusterBM101, ...>
+          typename BM10BN10ThreadClusterBN10Xs, // Sequence<BM10BN10ThreadClusterBN100,
+                                                //          BM10BN10ThreadClusterBN101, ...>
+          index_t AThreadCopyScalarPerVector_BM11,
+          index_t BThreadCopyScalarPerVector_BN11,
+          typename enable_if<ABlockDesc_BK0_BM_BK1::IsKnownAtCompileTime() &&
+                                 BBlockDesc_BK0_BN_BK1::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2
+{
+    using AIndex = MultiIndex<3>;
+    using BIndex = MultiIndex<3>;
+    using CIndex = MultiIndex<4>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr index_t BK0 = ABlockDesc_BK0_BM_BK1{}.GetLength(I0);
+    static constexpr index_t BK1 = ABlockDesc_BK0_BM_BK1{}.GetLength(I2);
+    static constexpr index_t BM  = ABlockDesc_BK0_BM_BK1{}.GetLength(I1);
+    static constexpr index_t BN  = BBlockDesc_BK0_BN_BK1{}.GetLength(I1);
+
+    static constexpr index_t BM100 = BM10BN10ThreadClusterBM10Xs{}[I0];
+    static constexpr index_t BN100 = BM10BN10ThreadClusterBN10Xs{}[I0];
+
+    static constexpr index_t BM101 = BM10BN10ThreadClusterBM10Xs{}[I1];
+    static constexpr index_t BN101 = BM10BN10ThreadClusterBN10Xs{}[I1];
+
+    static constexpr index_t BM11 = BM1PerThreadBM11;
+    static constexpr index_t BN11 = BN1PerThreadBN11;
+
+    static constexpr index_t BM1 = BM100 * BM101 * BM11;
+    static constexpr index_t BN1 = BN100 * BN101 * BN11;
+
+    static constexpr index_t BM0 = BM / BM1;
+    static constexpr index_t BN0 = BN / BN1;
+
+    __host__ __device__ static constexpr auto
+    MakeABlockDescriptor_BK0_BM0_BM1_BK1(const ABlockDesc_BK0_BM_BK1& a_block_desc_bk0_bm_bk1)
+    {
+        const auto a_block_bk0_bm0_bm1_bk1 = transform_tensor_descriptor(
+            a_block_desc_bk0_bm_bk1,
+            make_tuple(make_pass_through_transform(Number<BK0>{}),
+                       make_unmerge_transform(make_tuple(Number<BM0>{}, Number<BM1>{})),
+                       make_pass_through_transform(Number<BK1>{})),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+        return a_block_bk0_bm0_bm1_bk1;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeBBlockDescriptor_BK0_BN0_BN1_BK1(const BBlockDesc_BK0_BN_BK1& b_block_desc_bk0_bn_bk1)
+    {
+        const auto b_block_desc_bk0_bn0_bn1_bk1 = transform_tensor_descriptor(
+            b_block_desc_bk0_bn_bk1,
+            make_tuple(make_pass_through_transform(Number<BK0>{}),
+                       make_unmerge_transform(make_tuple(Number<BN0>{}, Number<BN1>{})),
+                       make_pass_through_transform(Number<BK1>{})),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+        return b_block_desc_bk0_bn0_bn1_bk1;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCBlockAdaptor_BM0_BM100_BM101_BM11_BN0_BN100_BN101_BN11_To_BM_BN()
+    {
+        // upper: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
+        // lower: [BM, BN]
+        constexpr auto c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m_n =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_unmerge_transform(make_tuple(
+                               Number<BM0>{}, Number<BM100>{}, Number<BM101>{}, Number<BM11>{})),
+                           make_unmerge_transform(make_tuple(
+                               Number<BN0>{}, Number<BN100>{}, Number<BN101>{}, Number<BN11>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4, 5, 6, 7>{}));
+
+        return c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m_n;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCBlockAdaptor_BM0_BM100_BM101_BM11_BN0_BN100_BN101_BN11_To_BM0_BM1_BN0_BN1()
+    {
+        // upper: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
+        // lower: [BM0, BM1, BN0, BN1]
+        constexpr auto c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m0_m1_n0_n1 =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_pass_through_transform(Number<BM0>{}),
+                           make_unmerge_transform(
+                               make_tuple(Number<BM100>{}, Number<BM101>{}, Number<BM11>{})),
+                           make_pass_through_transform(Number<BN0>{}),
+                           make_unmerge_transform(
+                               make_tuple(Number<BN100>{}, Number<BN101>{}, Number<BN11>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}, Sequence<5, 6, 7>{}));
+
+        return c_block_adaptor_m0_m100_m101_m11_n0_n100_n101_n11_to_m0_m1_n0_n1;
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadTensorLengths_BM0_BM1_BN0_BN1()
+    {
+        return Sequence<BM0, BM11, BN0, BN11>{};
+    }
+
+    static constexpr auto a_block_desc_bk0_bm0_bm1_bk1_ =
+        MakeABlockDescriptor_BK0_BM0_BM1_BK1(ABlockDesc_BK0_BM_BK1{});
+
+    static constexpr auto b_block_desc_bk0_bn0_bn1_bk1_ =
+        MakeBBlockDescriptor_BK0_BN0_BN1_BK1(BBlockDesc_BK0_BN_BK1{});
+
+    public:
+    __device__ BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2()
+        : c_thread_origin_data_idx_{CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
+              get_thread_local_1d_id())},
+          a_thread_copy_{
+              make_tuple(0, c_thread_origin_data_idx_[I0], c_thread_origin_data_idx_[I1], 0)},
+          b_thread_copy_{
+              make_tuple(0, c_thread_origin_data_idx_[I2], c_thread_origin_data_idx_[I3], 0)}
+    {
+        static_assert(ABlockDesc_BK0_BM_BK1::IsKnownAtCompileTime() &&
+                          BBlockDesc_BK0_BN_BK1::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(BlockSize == BM101 * BM100 * BN101 * BN100,
+                      "wrong! blocksize and cluster size not consistent");
+
+        static_assert(BM % BM1 == 0 && BN % BN1 == 0, "wrong!");
+
+        static_assert(ABlockDesc_BK0_BM_BK1{}.GetLength(I0) ==
+                          BBlockDesc_BK0_BN_BK1{}.GetLength(I0),
+                      "wrong! K dimension not consistent");
+
+        // TODO remove this restriction
+        static_assert(BM10BN10ThreadClusterBM10Xs::Size() == 2 &&
+                          BM10BN10ThreadClusterBN10Xs::Size() == 2,
+                      "wrong!");
+
+        // TODO: remove this restriction
+        static_assert(BM0 == 2, "wrong");
+        static_assert(BM0 == 2 && BN0 == 2, "wrong");
+    }
+
+    __device__ static CIndex CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(index_t thread_id)
+    {
+        // lower: [BM0, BM1, BN0, BN1]
+        // upper: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
+        constexpr auto adaptor0 =
+            MakeCBlockAdaptor_BM0_BM100_BM101_BM11_BN0_BN100_BN101_BN11_To_BM0_BM1_BN0_BN1();
+
+        // lower: [BM0, BM100, BM101, BM11, BN0, BN100, BN101, BN11]
+        // upper: [Tid, BM0, BM11, BN0, BN11]
+        constexpr auto adaptor1 = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(BM100, BN100, BM101, BN101)),
+                       make_pass_through_transform(BM0),
+                       make_pass_through_transform(BM11),
+                       make_pass_through_transform(BN0),
+                       make_pass_through_transform(BN11)),
+            make_tuple(
+                Sequence<1, 5, 2, 6>{}, Sequence<0>{}, Sequence<3>{}, Sequence<4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+        constexpr auto adaptor = chain_tensor_adaptors(adaptor0, adaptor1);
+
+        return adaptor.CalculateBottomIndex(make_multi_index(thread_id, 0, 0, 0, 0));
+    }
+
+    template <typename CThreadDesc_BM0_BM11_BN0_BN11,
+              typename ABlockBuffer,
+              typename BBlockBuffer,
+              typename CThreadBuffer>
+    __device__ void Run(const CThreadDesc_BM0_BM11_BN0_BN11&,
+                        const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_block_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        static_assert(CThreadDesc_BM0_BM11_BN0_BN11::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        // TODO: remove this restriction
+        static_assert(BM0 == 2 && BN0 == 2 &&
+                          CThreadDesc_BM0_BM11_BN0_BN11{}.GetLength(I0) == BM0 &&
+                          CThreadDesc_BM0_BM11_BN0_BN11{}.GetLength(I2) == BN0,
+                      "wrong");
+
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
+            a_thread_desc_bk0_bm0_bm1_bk1_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
+            b_thread_desc_bk0_bn0_bn1_bk1_.GetElementSpaceSize());
+
+        constexpr auto threadwise_contraction =
+            ThreadwiseContractionDl_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1<
+                FloatA,
+                FloatB,
+                FloatC,
+                decltype(a_thread_desc_bk0_bm0_bm1_bk1_),
+                decltype(b_thread_desc_bk0_bn0_bn1_bk1_),
+                CThreadDesc_BM0_BM11_BN0_BN11,
+                Sequence<BK0PerThread, BK1>,
+                Sequence<1, BM1PerThreadBM11>,
+                Sequence<1, BN1PerThreadBN11>>{};
+
+        // read A_sub_0
+        a_thread_copy_.Run(a_block_desc_bk0_bm0_bm1_bk1_,
+                           make_tuple(I0, I0, I0, I0),
+                           a_block_buf,
+                           a_thread_desc_bk0_bm0_bm1_bk1_,
+                           make_tuple(I0, I0, I0, I0),
+                           a_thread_buf);
+
+        // read B_sub_0
+        b_thread_copy_.Run(b_block_desc_bk0_bn0_bn1_bk1_,
+                           make_tuple(I0, I0, I0, I0),
+                           b_block_buf,
+                           b_thread_desc_bk0_bn0_bn1_bk1_,
+                           make_tuple(I0, I0, I0, I0),
+                           b_thread_buf);
+
+        // read B_sub_1
+        b_thread_copy_.Run(b_block_desc_bk0_bn0_bn1_bk1_,
+                           make_tuple(I0, I1, I0, I0),
+                           b_block_buf,
+                           b_thread_desc_bk0_bn0_bn1_bk1_,
+                           make_tuple(I0, I1, I0, I0),
+                           b_thread_buf);
+
+        // read A_sub_1
+        a_thread_copy_.Run(a_block_desc_bk0_bm0_bm1_bk1_,
+                           make_tuple(I0, I1, I0, I0),
+                           a_block_buf,
+                           a_thread_desc_bk0_bm0_bm1_bk1_,
+                           make_tuple(I0, I1, I0, I0),
+                           a_thread_buf);
+
+        // C_sub_00 += transpose(A_sub_0) * B_sub_0
+        threadwise_contraction.Run(a_thread_buf,
+                                   make_tuple(I0, I0, I0, I0),
+                                   b_thread_buf,
+                                   make_tuple(I0, I0, I0, I0),
+                                   c_thread_buf,
+                                   make_tuple(I0, I0, I0, I0));
+
+        // C_sub_01 += transpose(A_sub_0) * B_sub_1
+        threadwise_contraction.Run(a_thread_buf,
+                                   make_tuple(I0, I0, I0, I0),
+                                   b_thread_buf,
+                                   make_tuple(I0, I1, I0, I0),
+                                   c_thread_buf,
+                                   make_tuple(I0, I0, I1, I0));
+
+        // loop over rest of bk0
+        static_for<BK0PerThread, BK0, BK0PerThread>{}([&](auto bk0) {
+            // read A_sub_0
+            a_thread_copy_.Run(a_block_desc_bk0_bm0_bm1_bk1_,
+                               make_tuple(bk0, I0, I0, I0),
+                               a_block_buf,
+                               a_thread_desc_bk0_bm0_bm1_bk1_,
+                               make_tuple(I0, I0, I0, I0),
+                               a_thread_buf);
+
+            // C_sub_10 += transpose(A_sub_1) * B_sub_0
+            threadwise_contraction.Run(a_thread_buf,
+                                       make_tuple(I0, I1, I0, I0),
+                                       b_thread_buf,
+                                       make_tuple(I0, I0, I0, I0),
+                                       c_thread_buf,
+                                       make_tuple(I1, I0, I0, I0));
+
+            // read B_sub_0
+            b_thread_copy_.Run(b_block_desc_bk0_bn0_bn1_bk1_,
+                               make_tuple(bk0, I0, I0, I0),
+                               b_block_buf,
+                               b_thread_desc_bk0_bn0_bn1_bk1_,
+                               make_tuple(I0, I0, I0, I0),
+                               b_thread_buf);
+
+            // C_sub_11 += transpose(A_sub_1) * B_sub_1
+            threadwise_contraction.Run(a_thread_buf,
+                                       make_tuple(I0, I1, I0, I0),
+                                       b_thread_buf,
+                                       make_tuple(I0, I1, I0, I0),
+                                       c_thread_buf,
+                                       make_tuple(I1, I0, I1, I0));
+
+            // read B_sub_1
+            b_thread_copy_.Run(b_block_desc_bk0_bn0_bn1_bk1_,
+                               make_tuple(bk0, I1, I0, I0),
+                               b_block_buf,
+                               b_thread_desc_bk0_bn0_bn1_bk1_,
+                               make_tuple(I0, I1, I0, I0),
+                               b_thread_buf);
+
+            // read A_sub_1
+            a_thread_copy_.Run(a_block_desc_bk0_bm0_bm1_bk1_,
+                               make_tuple(bk0, I1, I0, I0),
+                               a_block_buf,
+                               a_thread_desc_bk0_bm0_bm1_bk1_,
+                               make_tuple(I0, I1, I0, I0),
+                               a_thread_buf);
+
+            // C_sub_00 += transpose(A_sub_0) * B_sub_0
+            threadwise_contraction.Run(a_thread_buf,
+                                       make_tuple(I0, I0, I0, I0),
+                                       b_thread_buf,
+                                       make_tuple(I0, I0, I0, I0),
+                                       c_thread_buf,
+                                       make_tuple(I0, I0, I0, I0));
+
+            // C_sub_01 += transpose(A_sub_0) * B_sub_1
+            threadwise_contraction.Run(a_thread_buf,
+                                       make_tuple(I0, I0, I0, I0),
+                                       b_thread_buf,
+                                       make_tuple(I0, I1, I0, I0),
+                                       c_thread_buf,
+                                       make_tuple(I0, I0, I1, I0));
+        });
+
+        // C_sub_10 += transpose(A_sub_1) * B_sub_0
+        threadwise_contraction.Run(a_thread_buf,
+                                   make_tuple(I0, I1, I0, I0),
+                                   b_thread_buf,
+                                   make_tuple(I0, I0, I0, I0),
+                                   c_thread_buf,
+                                   make_tuple(I1, I0, I0, I0));
+
+        // C_sub_11 += transpose(A_sub_1) * B_sub_1
+        threadwise_contraction.Run(a_thread_buf,
+                                   make_tuple(I0, I1, I0, I0),
+                                   b_thread_buf,
+                                   make_tuple(I0, I1, I0, I0),
+                                   c_thread_buf,
+                                   make_tuple(I1, I0, I1, I0));
+    }
+
+    private:
+    // A[BK0, BM0, BM1, BK1]
+    static constexpr auto a_thread_desc_bk0_bm0_bm1_bk1_ =
+        make_naive_tensor_descriptor_packed(make_tuple(
+            Number<BK0PerThread>{}, Number<BM0>{}, Number<BM1PerThreadBM11>{}, Number<BK1>{}));
+
+    // B[BK0, BN0, BN1, BK1]
+    static constexpr auto b_thread_desc_bk0_bn0_bn1_bk1_ =
+        make_naive_tensor_descriptor_packed(make_tuple(
+            Number<BK0PerThread>{}, Number<BN0>{}, Number<BN1PerThreadBN11>{}, Number<BK1>{}));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4r1<
+        FloatA,
+        FloatA,
+        decltype(a_block_desc_bk0_bm0_bm1_bk1_),
+        decltype(a_thread_desc_bk0_bm0_bm1_bk1_),
+        Sequence<BK0PerThread, 1, BM1PerThreadBM11, BK1>, // SliceLengths
+        Sequence<0, 1, 2, 3>,                             // DimAccessOrder
+        Sequence<1, 1, BM1PerThreadBM11, BK1>,            // SrcVectorTensorLengths
+        Sequence<0, 1, 2, 3>>;                            // SrcVectorTensorContiguousDimOrder
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4r1<
+        FloatB,
+        FloatB,
+        decltype(b_block_desc_bk0_bn0_bn1_bk1_),
+        decltype(b_thread_desc_bk0_bn0_bn1_bk1_),
+        Sequence<BK0PerThread, 1, BN1PerThreadBN11, BK1>, // SliceLengths
+        Sequence<0, 1, 2, 3>,                             // DimAccessOrder
+        Sequence<1, 1, BN1PerThreadBN11, BK1>,            // SrcVectorTensorLengths
+        Sequence<0, 1, 2, 3>>;                            // SrcVectorTensorContiguousDimOrder
+
+    CIndex c_thread_origin_data_idx_;
+
+    AThreadCopy a_thread_copy_;
+    BThreadCopy b_thread_copy_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp
new file mode 100644
index 0000000000..b0143366c1
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp
@@ -0,0 +1,397 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_BLOCKWISE_GEMM_DLOPS_V2R2_HPP
+#define CK_BLOCKWISE_GEMM_DLOPS_V2R2_HPP
+
+#include "common_header.hpp"
+#include "tensor_adaptor.hpp"
+#include "threadwise_tensor_slice_transfer.hpp"
+#include "threadwise_contraction_dlops.hpp"
+
+namespace ck {
+
+// C[M0, M1, N0, N1] += transpose(A[K, M0, M1]) * B[K, N0, N1]
+// A and B are visable to the whole block, C is distributed among each thread
+// Assume:
+//   1. A:
+//     1. AKMBlockDesc is known at compile-time
+//     2. ABlockBuffer is DynamicBuffer
+//   2. B:
+//     1. BKNBlockDesc is known at compile-time
+//     2. BBlockBuffer is DynamicBuffer
+//   3. C:
+//     1. CM0M1N0N1ThreadDesc is known at compile-time
+//     2. CThreadBuffer is StaticBuffer
+// Also assume:
+//   M0 = N0 = 2. It will do 2x2 pipelined read and fma (ABBA optimization)
+template <
+    index_t BlockSize,
+    typename FloatA,
+    typename FloatB,
+    typename FloatC,
+    typename AKMBlockDesc,
+    typename BKNBlockDesc,
+    index_t M1PerThreadM11,
+    index_t N1PerThreadN11,
+    index_t KPerThread,
+    index_t M1N1ThreadClusterM100,
+    index_t M1N1ThreadClusterN100,
+    index_t M1N1ThreadClusterM101,
+    index_t M1N1ThreadClusterN101,
+    index_t AThreadCopyScalarPerVector_M11,
+    index_t BThreadCopyScalarPerVector_N11,
+    typename enable_if<AKMBlockDesc::IsKnownAtCompileTime() && BKNBlockDesc::IsKnownAtCompileTime(),
+                       bool>::type = false>
+struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v2r2_pipeline_2x2
+{
+    using AIndex = MultiIndex<3>;
+    using BIndex = MultiIndex<3>;
+    using CIndex = MultiIndex<4>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr index_t K = AKMBlockDesc{}.GetLength(I0);
+    static constexpr index_t M = AKMBlockDesc{}.GetLength(I1);
+    static constexpr index_t N = BKNBlockDesc{}.GetLength(I1);
+
+    static constexpr index_t M100 = M1N1ThreadClusterM100;
+    static constexpr index_t N100 = M1N1ThreadClusterN100;
+
+    static constexpr index_t M101 = M1N1ThreadClusterM101;
+    static constexpr index_t N101 = M1N1ThreadClusterN101;
+
+    static constexpr index_t M11 = M1PerThreadM11;
+    static constexpr index_t N11 = N1PerThreadN11;
+
+    static constexpr index_t M1 = M1N1ThreadClusterM100 * M1N1ThreadClusterM101 * M1PerThreadM11;
+    static constexpr index_t N1 = M1N1ThreadClusterN100 * M1N1ThreadClusterN101 * N1PerThreadN11;
+
+    static constexpr index_t M0 = M / M1;
+    static constexpr index_t N0 = N / N1;
+
+    __host__ __device__ static constexpr auto
+    MakeAKM0M1BlockDescriptor(const AKMBlockDesc& /* a_k_m_block_desc */)
+    {
+        const auto a_k_m0_m1_block_desc = transform_tensor_descriptor(
+            AKMBlockDesc{},
+            make_tuple(make_pass_through_transform(Number<K>{}),
+                       make_unmerge_transform(make_tuple(Number<M0>{}, Number<M1>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}));
+
+        return a_k_m0_m1_block_desc;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeBKN0N1BlockDescriptor(const BKNBlockDesc& /* b_k_n_block_desc */)
+    {
+        const auto b_k_n0_n1_block_desc = transform_tensor_descriptor(
+            BKNBlockDesc{},
+            make_tuple(make_pass_through_transform(Number<K>{}),
+                       make_unmerge_transform(make_tuple(Number<N0>{}, Number<N1>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}));
+
+        return b_k_n0_n1_block_desc;
+    }
+
+    __host__ __device__ static constexpr auto MakeCM0M100M101M11N0N100N101N11ToMNBlockAdaptor()
+    {
+        // upper: [M0, M100, M101, M11, N0, N100, N101, N11]
+        // lower: [M, N]
+        constexpr auto c_m0_m100_m101_m11_n0_n100_n101_n11_to_m_n_block_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_unmerge_transform(make_tuple(
+                               Number<M0>{}, Number<M100>{}, Number<M101>{}, Number<M11>{})),
+                           make_unmerge_transform(make_tuple(
+                               Number<N0>{}, Number<N100>{}, Number<N101>{}, Number<N11>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4, 5, 6, 7>{}));
+
+        return c_m0_m100_m101_m11_n0_n100_n101_n11_to_m_n_block_adaptor;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCM0M100M101M11N0N100N101N11ToM0M1N0N1BlockAdaptor()
+    {
+        // upper: [M0, M100, M101, M11, N0, N100, N101, N11]
+        // lower: [M0, M1, N0, N1]
+        constexpr auto c_m0_m100_m101_m11_n0_n100_n101_n11_to_m0_m1_n0_n1_block_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_pass_through_transform(Number<M0>{}),
+                           make_unmerge_transform(
+                               make_tuple(Number<M100>{}, Number<M101>{}, Number<M11>{})),
+                           make_pass_through_transform(Number<N0>{}),
+                           make_unmerge_transform(
+                               make_tuple(Number<N100>{}, Number<N101>{}, Number<N11>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}, Sequence<5, 6, 7>{}));
+
+        return c_m0_m100_m101_m11_n0_n100_n101_n11_to_m0_m1_n0_n1_block_adaptor;
+    }
+
+    __host__ __device__ static constexpr auto GetCM0M1N0N1ThreadTensorLengths()
+    {
+        return Sequence<M0, M11, N0, N11>{};
+    }
+
+    static constexpr auto a_k_m0_m1_block_desc_ = MakeAKM0M1BlockDescriptor(AKMBlockDesc{});
+    static constexpr auto b_k_n0_n1_block_desc_ = MakeBKN0N1BlockDescriptor(BKNBlockDesc{});
+
+    public:
+    __device__ BlockwiseGemmDlops_km_kn_m0m1n0n1_v2r2_pipeline_2x2()
+        : c_thread_origin_data_idx_{CalculateCM0M1N0N1ThreadOriginOnBlock(
+              get_thread_local_1d_id())},
+          a_thread_copy_{
+              make_tuple(0, c_thread_origin_data_idx_[I0], c_thread_origin_data_idx_[I1])},
+          b_thread_copy_{
+              make_tuple(0, c_thread_origin_data_idx_[I2], c_thread_origin_data_idx_[I3])}
+    {
+        static_assert(AKMBlockDesc::IsKnownAtCompileTime() && BKNBlockDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(BlockSize == M101 * M100 * N101 * N100,
+                      "wrong! blocksize and cluster size not consistent");
+
+        static_assert(M % M1 == 0 && N % N1 == 0, "wrong!");
+
+        static_assert(AKMBlockDesc{}.GetLength(I0) == BKNBlockDesc{}.GetLength(I0),
+                      "wrong! K dimension not consistent");
+
+        // TODO: remove this restriction
+        static_assert(M0 == 2 && N0 == 2, "wrong");
+    }
+
+    __device__ static CIndex CalculateCM0M1N0N1ThreadOriginOnBlock(index_t thread_id)
+    {
+        // lower: [M0, M1, N0, N1]
+        // upper: [M0, M100, M101, M11, N0, N100, N101, N11]
+        constexpr auto adaptor0 = MakeCM0M100M101M11N0N100N101N11ToM0M1N0N1BlockAdaptor();
+
+        // lower: [M0, M100, M101, M11, N0, N100, N101, N11]
+        // upper: [Tid, M0, M11, N0, N11]
+        constexpr auto adaptor1 = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(M100, N100, M101, N101)),
+                       make_pass_through_transform(M0),
+                       make_pass_through_transform(M11),
+                       make_pass_through_transform(N0),
+                       make_pass_through_transform(N11)),
+            make_tuple(
+                Sequence<1, 5, 2, 6>{}, Sequence<0>{}, Sequence<3>{}, Sequence<4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+        constexpr auto adaptor = chain_tensor_adaptors(adaptor0, adaptor1);
+
+        return adaptor.CalculateBottomIndex(make_multi_index(thread_id, 0, 0, 0, 0));
+    }
+
+    __host__ __device__ static constexpr index_t GetABlockAlignment() { return M1PerThreadM11; }
+
+    __host__ __device__ static constexpr auto GetBBlockAlignment() { return N1PerThreadN11; }
+
+    template <typename CM0M1N0N1ThreadDesc,
+              typename ABlockBuffer,
+              typename BBlockBuffer,
+              typename CThreadBuffer>
+    __device__ void Run(const CM0M1N0N1ThreadDesc& /* c_m0_m1_n0_n1_thread_desc */,
+                        const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_block_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        static_assert(CM0M1N0N1ThreadDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        // TODO: remove this restriction
+        static_assert(M0 == 2 && N0 == 2 && CM0M1N0N1ThreadDesc{}.GetLength(I0) == M0 &&
+                          CM0M1N0N1ThreadDesc{}.GetLength(I2) == N0,
+                      "wrong");
+
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
+            a_k_m0_m1_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
+            b_k_n0_n1_thread_desc_.GetElementSpaceSize());
+
+        constexpr auto threadwise_gemm =
+            ThreadwiseGemmDlops_km0m1_kn0n1_m0m1n0n1<FloatA,
+                                                     FloatB,
+                                                     FloatC,
+                                                     decltype(a_k_m0_m1_thread_desc_),
+                                                     decltype(b_k_n0_n1_thread_desc_),
+                                                     CM0M1N0N1ThreadDesc,
+                                                     Sequence<KPerThread>,
+                                                     Sequence<1, M1PerThreadM11>,
+                                                     Sequence<1, N1PerThreadN11>>{};
+
+        // read A_sub_0
+        a_thread_copy_.Run(a_k_m0_m1_block_desc_,
+                           make_tuple(I0, I0, I0),
+                           a_block_buf,
+                           a_k_m0_m1_thread_desc_,
+                           make_tuple(I0, I0, I0),
+                           a_thread_buf);
+
+        // read B_sub_0
+        b_thread_copy_.Run(b_k_n0_n1_block_desc_,
+                           make_tuple(I0, I0, I0),
+                           b_block_buf,
+                           b_k_n0_n1_thread_desc_,
+                           make_tuple(I0, I0, I0),
+                           b_thread_buf);
+
+        // read B_sub_1
+        b_thread_copy_.Run(b_k_n0_n1_block_desc_,
+                           make_tuple(I0, I1, I0),
+                           b_block_buf,
+                           b_k_n0_n1_thread_desc_,
+                           make_tuple(I0, I1, I0),
+                           b_thread_buf);
+
+        // read A_sub_1
+        a_thread_copy_.Run(a_k_m0_m1_block_desc_,
+                           make_tuple(I0, I1, I0),
+                           a_block_buf,
+                           a_k_m0_m1_thread_desc_,
+                           make_tuple(I0, I1, I0),
+                           a_thread_buf);
+
+        // C_sub_00 += transpose(A_sub_0) * B_sub_0
+        threadwise_gemm.Run(a_thread_buf,
+                            make_tuple(I0, I0, I0),
+                            b_thread_buf,
+                            make_tuple(I0, I0, I0),
+                            c_thread_buf,
+                            make_tuple(I0, I0, I0, I0));
+
+        // C_sub_01 += transpose(A_sub_0) * B_sub_1
+        threadwise_gemm.Run(a_thread_buf,
+                            make_tuple(I0, I0, I0),
+                            b_thread_buf,
+                            make_tuple(I0, I1, I0),
+                            c_thread_buf,
+                            make_tuple(I0, I0, I1, I0));
+
+        // loop over rest of k
+        static_for<KPerThread, K, KPerThread>{}([&](auto k) {
+            // read A_sub_0
+            a_thread_copy_.Run(a_k_m0_m1_block_desc_,
+                               make_tuple(k, I0, I0),
+                               a_block_buf,
+                               a_k_m0_m1_thread_desc_,
+                               make_tuple(I0, I0, I0),
+                               a_thread_buf);
+
+            // C_sub_10 += transpose(A_sub_1) * B_sub_0
+            threadwise_gemm.Run(a_thread_buf,
+                                make_tuple(I0, I1, I0),
+                                b_thread_buf,
+                                make_tuple(I0, I0, I0),
+                                c_thread_buf,
+                                make_tuple(I1, I0, I0, I0));
+
+            // read B_sub_0
+            b_thread_copy_.Run(b_k_n0_n1_block_desc_,
+                               make_tuple(k, I0, I0),
+                               b_block_buf,
+                               b_k_n0_n1_thread_desc_,
+                               make_tuple(I0, I0, I0),
+                               b_thread_buf);
+
+            // C_sub_11 += transpose(A_sub_1) * B_sub_1
+            threadwise_gemm.Run(a_thread_buf,
+                                make_tuple(I0, I1, I0),
+                                b_thread_buf,
+                                make_tuple(I0, I1, I0),
+                                c_thread_buf,
+                                make_tuple(I1, I0, I1, I0));
+
+            // read B_sub_1
+            b_thread_copy_.Run(b_k_n0_n1_block_desc_,
+                               make_tuple(k, I1, I0),
+                               b_block_buf,
+                               b_k_n0_n1_thread_desc_,
+                               make_tuple(I0, I1, I0),
+                               b_thread_buf);
+
+            // read A_sub_1
+            a_thread_copy_.Run(a_k_m0_m1_block_desc_,
+                               make_tuple(k, I1, I0),
+                               a_block_buf,
+                               a_k_m0_m1_thread_desc_,
+                               make_tuple(I0, I1, I0),
+                               a_thread_buf);
+
+            // C_sub_00 += transpose(A_sub_0) * B_sub_0
+            threadwise_gemm.Run(a_thread_buf,
+                                make_tuple(I0, I0, I0),
+                                b_thread_buf,
+                                make_tuple(I0, I0, I0),
+                                c_thread_buf,
+                                make_tuple(I0, I0, I0, I0));
+
+            // C_sub_01 += transpose(A_sub_0) * B_sub_1
+            threadwise_gemm.Run(a_thread_buf,
+                                make_tuple(I0, I0, I0),
+                                b_thread_buf,
+                                make_tuple(I0, I1, I0),
+                                c_thread_buf,
+                                make_tuple(I0, I0, I1, I0));
+        });
+
+        // C_sub_10 += transpose(A_sub_1) * B_sub_0
+        threadwise_gemm.Run(a_thread_buf,
+                            make_tuple(I0, I1, I0),
+                            b_thread_buf,
+                            make_tuple(I0, I0, I0),
+                            c_thread_buf,
+                            make_tuple(I1, I0, I0, I0));
+
+        // C_sub_11 += transpose(A_sub_1) * B_sub_1
+        threadwise_gemm.Run(a_thread_buf,
+                            make_tuple(I0, I1, I0),
+                            b_thread_buf,
+                            make_tuple(I0, I1, I0),
+                            c_thread_buf,
+                            make_tuple(I1, I0, I1, I0));
+    }
+
+    private:
+    // A[K, M0, M1]
+    static constexpr auto a_k_m0_m1_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<KPerThread>{}, Number<M0>{}, Number<M1PerThreadM11>{}));
+
+    // B[K, N0, N1]
+    static constexpr auto b_k_n0_n1_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<KPerThread>{}, Number<N0>{}, Number<N1PerThreadN11>{}));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatA,
+                                                         FloatA,
+                                                         decltype(a_k_m0_m1_block_desc_),
+                                                         decltype(a_k_m0_m1_thread_desc_),
+                                                         Sequence<KPerThread, 1, M1PerThreadM11>,
+                                                         Sequence<0, 1, 2>,
+                                                         2,
+                                                         AThreadCopyScalarPerVector_M11,
+                                                         1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatB,
+                                                         FloatB,
+                                                         decltype(b_k_n0_n1_block_desc_),
+                                                         decltype(b_k_n0_n1_thread_desc_),
+                                                         Sequence<KPerThread, 1, N1PerThreadN11>,
+                                                         Sequence<0, 1, 2>,
+                                                         2,
+                                                         BThreadCopyScalarPerVector_N11,
+                                                         1>;
+
+    CIndex c_thread_origin_data_idx_;
+
+    AThreadCopy a_thread_copy_;
+    BThreadCopy b_thread_copy_;
+};
+
+} // namespace ck
+#endif
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp
new file mode 100644
index 0000000000..0d092da516
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp
@@ -0,0 +1,178 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_BLOCKWISE_GEMM_DLOPS_V3_HPP
+#define CK_BLOCKWISE_GEMM_DLOPS_V3_HPP
+
+#include "common_header.hpp"
+#include "threadwise_gemm_dlops_v3.hpp"
+
+namespace ck {
+
+template <index_t BlockSize,
+          typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename ABlockDesc_E1_K1_E2,
+          typename BBlockDesc_E1_N_Ho_Wo_E2,
+          typename CThreadDesc_K_N_Ho_Wo,
+          index_t EPerThreadLoop,
+          index_t KPerThreadLoop>
+struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v3
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+
+    using AIndex = MultiIndex<3>;
+    using BIndex = MultiIndex<3>;
+    using CIndex = MultiIndex<4>;
+
+    static constexpr auto E1        = ABlockDesc_E1_K1_E2{}.GetLength(I0);
+    static constexpr auto KPerBlock = ABlockDesc_E1_K1_E2{}.GetLength(I1);
+    static constexpr auto E2        = ABlockDesc_E1_K1_E2{}.GetLength(I2);
+
+    static constexpr auto HoPerBlock = BBlockDesc_E1_N_Ho_Wo_E2{}.GetLength(I2);
+    static constexpr auto WoPerBlock = BBlockDesc_E1_N_Ho_Wo_E2{}.GetLength(I3);
+
+    static constexpr auto KPerThread  = CThreadDesc_K_N_Ho_Wo{}.GetLength(I0);
+    static constexpr auto HoPerThread = CThreadDesc_K_N_Ho_Wo{}.GetLength(I2);
+    static constexpr auto WoPerThread = CThreadDesc_K_N_Ho_Wo{}.GetLength(I3);
+
+    static constexpr auto a_thread_mtx_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<EPerThreadLoop>{}, Number<KPerThreadLoop>{}, Number<E2>{}));
+
+    static constexpr auto b_thread_mtx_ =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<EPerThreadLoop>{},
+                                                       Number<1>{},
+                                                       Number<HoPerThread>{},
+                                                       Number<WoPerThread>{},
+                                                       Number<E2>{}));
+
+    static constexpr auto c_thread_mtx_ = make_naive_tensor_descriptor_packed(make_tuple(
+        Number<KPerThreadLoop>{}, Number<1>{}, Number<HoPerThread>{}, Number<WoPerThread>{}));
+
+    __device__ BlockwiseGemmDlops_km_kn_m0m1n0n1_v3()
+        : c_thread_origin_data_idx_{GetBeginOfCThreadDesc_K_N_Ho_Wo(get_thread_local_1d_id())},
+          a_thread_copy_{make_tuple(0, c_thread_origin_data_idx_[I0] * KPerThread, 0)}
+    {
+        static_assert(ABlockDesc_E1_K1_E2::IsKnownAtCompileTime() &&
+                          BBlockDesc_E1_N_Ho_Wo_E2::IsKnownAtCompileTime() &&
+                          CThreadDesc_K_N_Ho_Wo::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(
+            ABlockDesc_E1_K1_E2{}.GetLength(I0) == BBlockDesc_E1_N_Ho_Wo_E2{}.GetLength(I0) &&
+                ABlockDesc_E1_K1_E2{}.GetLength(I2) == BBlockDesc_E1_N_Ho_Wo_E2{}.GetLength(I4),
+            "wrong! E dimension not consistent\n");
+
+        static_assert(E1 % EPerThreadLoop == 0, "");
+        static_assert(KPerThread % KPerThreadLoop == 0, "");
+
+        static_assert(KPerBlock % KPerThread == 0 && HoPerBlock % HoPerThread == 0 &&
+                          WoPerBlock % WoPerThread == 0,
+                      "wrong! Cannot evenly divide work among\n");
+
+        constexpr auto KThreadCluster = KPerBlock / KPerThread;
+        constexpr auto HThreadCluster = HoPerBlock / HoPerThread;
+        constexpr auto WThreadCluster = WoPerBlock / WoPerThread;
+
+        static_assert(BlockSize == KThreadCluster * HThreadCluster * WThreadCluster,
+                      "wrong! wrong blocksize\n");
+    }
+
+    __device__ static constexpr auto GetCThreadDesc_K_N_Ho_WoLengths()
+    {
+        return Sequence<KPerThread, I1, HoPerThread, WoPerThread>{};
+    }
+
+    __device__ static CIndex GetBeginOfCThreadDesc_K_N_Ho_Wo(index_t thread_id)
+    {
+        constexpr auto K0 = KPerBlock / KPerThread;
+        constexpr auto N0 = I1;
+        constexpr auto H0 = HoPerBlock / HoPerThread;
+        constexpr auto W0 = WoPerBlock / WoPerThread;
+
+        constexpr auto c_threadid_to_k_n_h_w_thread_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(K0, N0, H0, W0))),
+                make_tuple(Sequence<0, 1, 2, 3>{}),
+                make_tuple(Sequence<0>{}));
+
+        const auto c_k_n_h_w_thread_cluster_idx =
+            c_threadid_to_k_n_h_w_thread_cluster_adaptor.CalculateBottomIndex(
+                make_multi_index(thread_id));
+
+        return c_k_n_h_w_thread_cluster_idx;
+    }
+
+    template <typename ABlockBuffer, typename BThreadBuffer, typename CThreadBuffer>
+    __device__ void Run(const ABlockBuffer& a_block_buf,
+                        const BThreadBuffer& b_thread_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        static_assert(
+            is_same<remove_cvref_t<typename ABlockBuffer::type>, remove_cvref_t<FloatA>>::value &&
+            is_same<remove_cvref_t<typename BThreadBuffer::type>, remove_cvref_t<FloatB>>::value &&
+            is_same<remove_cvref_t<typename CThreadBuffer::type>, remove_cvref_t<FloatC>>::value &&
+            "wrong! inconsistent type");
+
+        constexpr auto a_block_mtx = ABlockDesc_E1_K1_E2{};
+
+        // thread A buffer for GEMM
+        StaticBuffer<AddressSpaceEnum::Vgpr, FloatA, a_thread_mtx_.GetElementSpaceSize(), true>
+            a_thread_buf;
+
+        constexpr auto threadwise_gemm = ThreadwiseGemmDlops_km_kn_mn_v3<FloatA,
+                                                                         FloatB,
+                                                                         FloatC,
+                                                                         decltype(a_thread_mtx_),
+                                                                         decltype(b_thread_mtx_),
+                                                                         decltype(c_thread_mtx_)>{};
+
+        static_for<0, E1, EPerThreadLoop>{}([&](auto e_begin) {
+            static_for<0, KPerThread, KPerThreadLoop>{}([&](auto k_begin) {
+                a_thread_copy_.Run(a_block_mtx,
+                                   make_tuple(e_begin, k_begin, I0),
+                                   a_block_buf,
+                                   a_thread_mtx_,
+                                   make_tuple(I0, I0, I0),
+                                   a_thread_buf);
+
+                threadwise_gemm.Run(a_thread_buf,
+                                    make_tuple(I0, I0, I0),
+                                    b_thread_buf,
+                                    make_tuple(e_begin, I0, I0, I0, I0),
+                                    c_thread_buf,
+                                    make_tuple(k_begin, I0, I0, I0));
+            });
+        });
+    }
+
+    template <typename ABlockSliceMoveStepIdx>
+    __device__ void MoveABlockSliceWindow(const ABlockSliceMoveStepIdx& a_block_slice_move_step_idx)
+    {
+        a_thread_copy_.MoveSrcSliceWindow(ABlockDesc_E1_K1_E2{}, a_block_slice_move_step_idx);
+    }
+
+    private:
+    using AThreadCopy =
+        ThreadwiseTensorSliceTransfer_v4<FloatA,
+                                         FloatA,
+                                         ABlockDesc_E1_K1_E2,
+                                         decltype(a_thread_mtx_),
+                                         Sequence<EPerThreadLoop, KPerThreadLoop, E2>,
+                                         Sequence<0, 1, 2>,
+                                         2,
+                                         E2,
+                                         E2>;
+
+    CIndex c_thread_origin_data_idx_;
+
+    AThreadCopy a_thread_copy_;
+};
+
+} // namespace ck
+#endif
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp
new file mode 100644
index 0000000000..3cdc19dcdc
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp
@@ -0,0 +1,801 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/warp/wmma_gemm.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#define CK_MNK_LOOP
+
+namespace ck {
+
+template <index_t BlockSize,
+          typename FloatA,
+          typename FloatB,
+          typename FloatAcc,
+          typename AK0MK1BlockDesc,
+          typename BK0NK1BlockDesc,
+          index_t MPerWMMA,
+          index_t NPerWMMA,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack>
+/* A: K0PerBlock x MPerBlock x K1
+ * B: K0PerBlock x NPerBlock x K1
+ * C: MRepeat x MWave x MSubGroup x NRepeat x NWave x NThreadPerSubGroup x MAccVgprs
+ * KPACK == WMMA_K = 16
+ */
+struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle
+{
+    static constexpr auto I0    = Number<0>{};
+    static constexpr auto I1    = Number<1>{};
+    static constexpr auto I2    = Number<2>{};
+    static constexpr auto I3    = Number<3>{};
+    static constexpr auto I4    = Number<4>{};
+    static constexpr auto WmmaK = Number<16>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    // Hardcode of WaveSize, since current HIP Runtime(5.4.0-10984) could not return correct one.
+    static constexpr index_t WaveSize = 32;
+
+    static constexpr index_t MPerBlock = AK0MK1BlockDesc{}.GetLength(I1);
+    static constexpr index_t NPerBlock = BK0NK1BlockDesc{}.GetLength(I1);
+    static constexpr index_t KPerBlock =
+        BK0NK1BlockDesc{}.GetLength(I0) * BK0NK1BlockDesc{}.GetLength(I2);
+
+    static constexpr index_t A_K0 = AK0MK1BlockDesc{}.GetLength(I0);
+    static constexpr index_t B_K0 = BK0NK1BlockDesc{}.GetLength(I0);
+    static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
+    static constexpr index_t B_K1 = BK0NK1BlockDesc{}.GetLength(I2);
+
+    static constexpr auto wmma_gemm =
+        WmmaGemm<FloatA, FloatB, FloatAcc, MPerWMMA, NPerWMMA, KPack>{};
+
+    static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerWMMA);
+    static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerWMMA);
+
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                              FloatAcc,
+                              MRepeat * NRepeat,
+                              wmma_gemm.GetRegSizePerWmma(),
+                              true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = ThisThreadBlock::GetThreadId();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+
+        const auto WMMA_a_idx = wmma_gemm.CalculateAThreadOriginDataIndex();
+        //  |KRepeat   |MRepeat|MWave      |MLane       |KPack
+        return make_multi_index(0, 0, waveId_m, WMMA_a_idx, 0);
+    }
+
+    __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_n = wave_idx[I1];
+
+        const auto WMMA_b_idx = wmma_gemm.CalculateBThreadOriginDataIndex();
+        //  |KRepeat   |NRepeat|Nwave      |NLane       |KPack
+        return make_multi_index(0, 0, waveId_n, WMMA_b_idx, 0);
+    }
+
+    template <index_t m0, index_t n0>
+    __device__ static auto CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = wmma_gemm.GetBeginOfThreadBlk();
+
+        constexpr auto mrepeat_mwave_mperWMMA_to_m_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerWMMA))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        constexpr auto nrepeat_nwave_nperWMMA_to_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerWMMA))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        const index_t c_thread_m = mrepeat_mwave_mperWMMA_to_m_adaptor.CalculateBottomIndex(
+            make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
+        const index_t c_thread_n = nrepeat_nwave_nperWMMA_to_n_adaptor.CalculateBottomIndex(
+            make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
+
+        return make_tuple(c_thread_m, c_thread_n);
+    }
+
+    __host__ __device__ BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle()
+    {
+        static_assert(AK0MK1BlockDesc::IsKnownAtCompileTime() &&
+                          BK0NK1BlockDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
+                      "ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
+
+        static_assert(MPerBlock % (MPerWMMA * MRepeat) == 0 &&
+                          NPerBlock % (NPerWMMA * NRepeat) == 0,
+                      "wrong!");
+    }
+
+    // Thread level, register decriptor. Vector-write
+    __host__ __device__ static constexpr auto
+    GetCThreadDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs()
+    {
+        constexpr auto c_msubgroup_nthreadpersubgroup_maccvgprs_tblk_lens =
+            wmma_gemm.GetCMSubGroupNThreadPerSubGroupMAccVgprsThreadBlkLengths();
+
+        constexpr auto MSubGroup          = c_msubgroup_nthreadpersubgroup_maccvgprs_tblk_lens[I0];
+        constexpr auto NThreadPerSubGroup = c_msubgroup_nthreadpersubgroup_maccvgprs_tblk_lens[I1];
+        constexpr auto MAccVgprs          = c_msubgroup_nthreadpersubgroup_maccvgprs_tblk_lens[I2];
+
+        return make_naive_tensor_descriptor_packed(
+            //        |MRepeat           |MWave |MSubGroup |NRepeat           |NWave
+            //        |NThreadPerSubGroup |MAccVgprs
+            make_tuple(Number<MRepeat>{},
+                       I1,
+                       MSubGroup,
+                       Number<NRepeat>{},
+                       I1,
+                       NThreadPerSubGroup,
+                       MAccVgprs));
+    }
+
+    // Provide dimension size
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs()
+    {
+        constexpr auto c_block_desc_mrepeat_mwave_mperwmma_nrepeat_nwave_nperwmma =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<MPerWMMA>{},
+                                                           Number<NRepeat>{},
+                                                           Number<NWaves>{},
+                                                           Number<NPerWMMA>{}));
+
+        return wmma_gemm
+            .MakeCDesc_MBlockxRepeat_MWave_MSubGroup_NBlockxRepeat_NWave_NThreadPerSubGroup_MAccVgprs(
+                c_block_desc_mrepeat_mwave_mperwmma_nrepeat_nwave_nperwmma);
+    }
+
+    __host__ __device__ static constexpr auto MakeABlockDescriptor_K0_M0_M1_M2_K1()
+    {
+        return transform_tensor_descriptor(
+            AK0MK1BlockDesc{},
+            make_tuple(make_pass_through_transform(Number<A_K0>{}),
+                       make_unmerge_transform(
+                           make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerWMMA>{})),
+                       make_pass_through_transform(Number<A_K1>{})),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}));
+    }
+
+    __host__ __device__ static constexpr auto MakeBBlockDescriptor_K0_N0_N1_N2_K1()
+    {
+        return transform_tensor_descriptor(
+            BK0NK1BlockDesc{},
+            make_tuple(make_pass_through_transform(Number<B_K0>{}),
+                       make_unmerge_transform(
+                           make_tuple(Number<NRepeat>{}, Number<NWaves>{}, Number<NPerWMMA>{})),
+                       make_pass_through_transform(Number<B_K1>{})),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}));
+    }
+
+    // M0_M1_M2 = MRepeat_MWave_MPerWmma, N0_N1_N2 = NRepeat_NWave_NPerWmma
+    static constexpr auto a_block_desc_k0_m0_m1_m2_k1 = MakeABlockDescriptor_K0_M0_M1_M2_K1();
+    static constexpr auto b_block_desc_k0_n0_n1_n2_k1 = MakeBBlockDescriptor_K0_N0_N1_N2_K1();
+
+    template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
+    __device__ void Run(const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_block_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        static_for<0, KPerBlock / WmmaK, 1>{}([&](auto k) { // k=0,1,2 instead of k=0,kpack*1, ...
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                // read A
+                a_thread_copy_.Run(a_block_desc_k0_m0_m1_m2_k1,
+                                   make_tuple(Number<k * WmmaK / A_K1>{}, m0, I0, I0, I0),
+                                   a_block_buf,
+                                   a_thread_desc_,
+                                   make_tuple(I0, m0, I0, I0, I0),
+                                   a_thread_buf);
+
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    // read B
+                    b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1,
+                                       make_tuple(Number<k * WmmaK / B_K1>{}, n0, I0, I0, I0),
+                                       b_block_buf,
+                                       b_thread_desc_,
+                                       make_tuple(I0, n0, I0, I0, I0),
+                                       b_thread_buf);
+                    vector_type<FloatA, WmmaK> a_thread_vec;
+                    vector_type<FloatB, WmmaK> b_thread_vec;
+
+                    static_for<0, WmmaK, 1>{}([&](auto i) {
+                        a_thread_vec.template AsType<FloatA>()(i) =
+                            a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                make_tuple(i / A_K1, m0, 0, 0, i % A_K1))>{}];
+                        b_thread_vec.template AsType<FloatB>()(i) =
+                            b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                make_tuple(i / B_K1, n0, 0, 0, i % B_K1))>{}];
+                    });
+
+                    using wmma_input_type_a = typename vector_type<FloatA, WmmaK>::type;
+                    using wmma_input_type_b = typename vector_type<FloatB, WmmaK>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    wmma_gemm.template Run(
+                        a_thread_vec.template AsType<wmma_input_type_a>()(Number<0>{}),
+                        b_thread_vec.template AsType<wmma_input_type_b>()(Number<0>{}),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+        });
+    }
+
+    protected:
+    // A[K0, M0, M1, M2, K1]
+    static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<WmmaK / A_K1>{}, Number<MRepeat>{}, I1, I1, Number<A_K1>{}));
+
+    // B[K0, N0, N1, N2, K1]
+    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<WmmaK / B_K1>{}, Number<NRepeat>{}, I1, I1, Number<B_K1>{}));
+
+    // C[M, N, NumRegWMMA]
+    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, wmma_gemm.GetRegSizePerWmma()));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatA,
+                                                         FloatA,
+                                                         decltype(a_block_desc_k0_m0_m1_m2_k1),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<WmmaK / A_K1, 1, 1, 1, A_K1>,
+                                                         Sequence<0, 1, 2, 3, 4>,
+                                                         4,
+                                                         A_K1,
+                                                         A_K1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatB,
+                                                         FloatB,
+                                                         decltype(b_block_desc_k0_n0_n1_n2_k1),
+                                                         decltype(b_thread_desc_),
+                                                         Sequence<WmmaK / B_K1, 1, 1, 1, B_K1>,
+                                                         Sequence<0, 1, 2, 3, 4>,
+                                                         4,
+                                                         B_K1,
+                                                         B_K1>;
+
+    AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
+    BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
+};
+
+// block wise level pipe designed for inline asm
+template <index_t BlockSize,
+          typename FloatA,
+          typename FloatB,
+          typename FloatAcc,
+          typename AK0MK1BlockDesc,
+          typename BK0NK1BlockDesc,
+          index_t MPerWMMA,
+          index_t NPerWMMA,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack>
+/* A: K0PerBlock x MPerBlock x K1
+ * B: K0PerBlock x NPerBlock x K1
+ * C: MRepeat x MWave x MSubGroup x NRepeat x NWave x NThreadPerSubGroup x MAccVgprs
+ * KPACK == WMMA_K = 16
+ */
+struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO
+{
+    static constexpr auto I0    = Number<0>{};
+    static constexpr auto I1    = Number<1>{};
+    static constexpr auto I2    = Number<2>{};
+    static constexpr auto I3    = Number<3>{};
+    static constexpr auto I4    = Number<4>{};
+    static constexpr auto WmmaK = Number<16>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    // Hardcode of WaveSize, since current HIP Runtime(5.4.0-10984) could not return correct one.
+    static constexpr index_t WaveSize = 32;
+
+    static constexpr index_t MPerBlock = AK0MK1BlockDesc{}.GetLength(I1);
+    static constexpr index_t NPerBlock = BK0NK1BlockDesc{}.GetLength(I1);
+    static constexpr index_t KPerBlock =
+        BK0NK1BlockDesc{}.GetLength(I0) * BK0NK1BlockDesc{}.GetLength(I2);
+
+    static constexpr index_t A_K0 = AK0MK1BlockDesc{}.GetLength(I0);
+    static constexpr index_t B_K0 = BK0NK1BlockDesc{}.GetLength(I0);
+    static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
+    static constexpr index_t B_K1 = BK0NK1BlockDesc{}.GetLength(I2);
+
+    static constexpr auto wmma_gemm =
+        WmmaGemm<FloatA, FloatB, FloatAcc, MPerWMMA, NPerWMMA, KPack>{};
+
+    static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerWMMA);
+    static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerWMMA);
+
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                              FloatAcc,
+                              MRepeat * NRepeat,
+                              wmma_gemm.GetRegSizePerWmma(),
+                              true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = ThisThreadBlock::GetThreadId();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+
+        const auto WMMA_a_idx = wmma_gemm.CalculateAThreadOriginDataIndex();
+        //  |KRepeat   |MRepeat|MWave      |MLane       |KPack
+        return make_tuple(0, 0, waveId_m, WMMA_a_idx, 0);
+    }
+
+    __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_n = wave_idx[I1];
+
+        const auto WMMA_b_idx = wmma_gemm.CalculateBThreadOriginDataIndex();
+        //  |KRepeat   |NRepeat|Nwave      |NLane       |KPack
+        return make_tuple(0, 0, waveId_n, WMMA_b_idx, 0);
+    }
+
+    template <index_t m0, index_t n0>
+    __device__ static auto CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = wmma_gemm.GetBeginOfThreadBlk();
+
+        constexpr auto mrepeat_mwave_mperWMMA_to_m_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerWMMA))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        constexpr auto nrepeat_nwave_nperWMMA_to_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerWMMA))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        const index_t c_thread_m = mrepeat_mwave_mperWMMA_to_m_adaptor.CalculateBottomIndex(
+            make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
+        const index_t c_thread_n = nrepeat_nwave_nperWMMA_to_n_adaptor.CalculateBottomIndex(
+            make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
+
+        return make_tuple(c_thread_m, c_thread_n);
+    }
+
+    __host__ __device__ BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO()
+    {
+        static_assert(AK0MK1BlockDesc::IsKnownAtCompileTime() &&
+                          BK0NK1BlockDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
+                      "ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
+
+        static_assert(MPerBlock % (MPerWMMA * MRepeat) == 0 &&
+                          NPerBlock % (NPerWMMA * NRepeat) == 0,
+                      "wrong!");
+    }
+    // Thread level, register decriptor. Vector-write
+    __host__ __device__ static constexpr auto
+    GetCThreadDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs()
+    {
+        constexpr auto c_msubgroup_nthreadpersubgroup_maccvgprs_tblk_lens =
+            wmma_gemm.GetCMSubGroupNThreadPerSubGroupMAccVgprsThreadBlkLengths();
+
+        constexpr auto MSubGroup          = c_msubgroup_nthreadpersubgroup_maccvgprs_tblk_lens[I0];
+        constexpr auto NThreadPerSubGroup = c_msubgroup_nthreadpersubgroup_maccvgprs_tblk_lens[I1];
+        constexpr auto MAccVgprs          = c_msubgroup_nthreadpersubgroup_maccvgprs_tblk_lens[I2];
+
+        return make_naive_tensor_descriptor_packed(
+            //        |MRepeat           |MWave |MSubGroup |NRepeat           |NWave
+            //        |NThreadPerSubGroup |MAccVgprs
+            make_tuple(Number<MRepeat>{},
+                       I1,
+                       MSubGroup,
+                       Number<NRepeat>{},
+                       I1,
+                       NThreadPerSubGroup,
+                       MAccVgprs));
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlockxRepeat_MWave_MSubGroup_NBlockxRepeat_NWave_NThreadPerSubGroup_MAccVgprs(
+        const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto c_grid_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma =
+            transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(
+                    make_unmerge_transform(make_tuple(M / (MWaves * MPerWMMA), MWaves, MPerWMMA)),
+                    make_unmerge_transform(make_tuple(N / (NWaves * NPerWMMA), NWaves, NPerWMMA))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
+
+        return wmma_gemm
+            .MakeCDesc_MBlockxRepeat_MWave_MSubGroup_NBlockxRepeat_NWave_NThreadPerSubGroup_MAccVgprs(
+                c_grid_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma);
+    }
+
+    // Provide dimension size
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs()
+    {
+        constexpr auto c_block_desc_mrepeat_mwave_mperwmma_nrepeat_nwave_nperwmma =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<MPerWMMA>{},
+                                                           Number<NRepeat>{},
+                                                           Number<NWaves>{},
+                                                           Number<NPerWMMA>{}));
+
+        return wmma_gemm
+            .MakeCDesc_MBlockxRepeat_MWave_MSubGroup_NBlockxRepeat_NWave_NThreadPerSubGroup_MAccVgprs(
+                c_block_desc_mrepeat_mwave_mperwmma_nrepeat_nwave_nperwmma);
+    }
+
+    __host__ __device__ static constexpr auto MakeABlockDescriptor_K0_M0_M1_M2_K1()
+    {
+        return transform_tensor_descriptor(
+            AK0MK1BlockDesc{},
+            make_tuple(make_pass_through_transform(Number<A_K0>{}),
+                       make_unmerge_transform(
+                           make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerWMMA>{})),
+                       make_pass_through_transform(Number<A_K1>{})),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}));
+    }
+
+    __host__ __device__ static constexpr auto MakeBBlockDescriptor_K0_N0_N1_N2_K1()
+    {
+        return transform_tensor_descriptor(
+            BK0NK1BlockDesc{},
+            make_tuple(make_pass_through_transform(Number<B_K0>{}),
+                       make_unmerge_transform(
+                           make_tuple(Number<NRepeat>{}, Number<NWaves>{}, Number<NPerWMMA>{})),
+                       make_pass_through_transform(Number<B_K1>{})),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2, 3>{}, Sequence<4>{}));
+    }
+
+    // M0_M1_M2 = MRepeat_MWave_MPerWmma, N0_N1_N2 = NRepeat_NWave_NPerWmma
+    static constexpr auto a_block_desc_k0_m0_m1_m2_k1 = MakeABlockDescriptor_K0_M0_M1_M2_K1();
+    static constexpr auto b_block_desc_k0_n0_n1_n2_k1 = MakeBBlockDescriptor_K0_N0_N1_N2_K1();
+
+    template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
+    __device__ void Run(const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_block_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatB>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        constexpr auto RepeatDiff = MRepeat - NRepeat;
+        // Read all Mrepeat, Nrepeat
+        static_for<0, NRepeat, 1>{}([&](auto iN) {
+            b_thread_copy_.Run(b_block_desc_k0_n0_n1_n2_k1,
+                               make_tuple(I0, Number<iN>{}, I0, I0, I0),
+                               b_block_buf,
+                               b_thread_desc_,
+                               make_tuple(I0, Number<iN>{}, I0, I0, I0),
+                               b_thread_buf);
+        });
+
+        static_for<0, MRepeat, 1>{}([&](auto iM) {
+            a_thread_copy_.Run(a_block_desc_k0_m0_m1_m2_k1,
+                               make_tuple(I0, Number<iM>{}, I0, I0, I0),
+                               a_block_buf,
+                               a_thread_desc_,
+                               make_tuple(I0, Number<iM>{}, I0, I0, I0),
+                               a_thread_buf);
+        });
+
+        // Stage 1: Cut to Repeat Retangle to Square, assume MRepeat > NRepeat
+        static_for<0, RepeatDiff, 1>{}([&](auto iCut) {
+            static_for<0, NRepeat, 1>{}([&](auto iN) {
+                vector_type<FloatA, WmmaK> a_thread_vec;
+                vector_type<FloatB, WmmaK> b_thread_vec;
+
+                static_for<0, WmmaK, 1>{}([&](auto iK) {
+                    a_thread_vec.template AsType<FloatA>()(iK) =
+                        a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                            make_tuple(iK / A_K1, iCut, 0, 0, iK % A_K1))>{}];
+                    b_thread_vec.template AsType<FloatB>()(iK) =
+                        b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                            make_tuple(iK / B_K1, iN, 0, 0, iK % B_K1))>{}];
+                });
+                using wmma_input_type_a = typename vector_type<FloatA, WmmaK>::type;
+                using wmma_input_type_b = typename vector_type<FloatB, WmmaK>::type;
+
+                constexpr index_t c_offset =
+                    c_thread_desc_.CalculateOffset(make_tuple(iCut, iN, 0));
+                // s_nop();
+                wmma_gemm.template Run(
+                    a_thread_vec.template AsType<wmma_input_type_a>()(Number<0>{}),
+                    b_thread_vec.template AsType<wmma_input_type_b>()(Number<0>{}),
+                    c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                // s_nop();
+            });
+            if constexpr(KPerBlock > WmmaK)
+            {
+                // Read Consumed Next inner loop A
+                a_thread_copy_.Run(a_block_desc_k0_m0_m1_m2_k1,
+                                   make_tuple(Number<WmmaK / A_K1>{}, Number<iCut>{}, I0, I0, I0),
+                                   a_block_buf,
+                                   a_thread_desc_,
+                                   make_tuple(I0, Number<iCut>{}, I0, I0, I0),
+                                   a_thread_buf);
+            }
+        });
+
+        static_for<WmmaK, KPerBlock, WmmaK>{}([&](auto iWmmaK) {
+            // Stage 2: Run FIFO fashion loopover in Square
+            static_for<0, NRepeat, 1>{}([&](auto WmmaInnerloop) {
+                // Row Repeatation
+                static_for<WmmaInnerloop, NRepeat, 1>{}([&](auto iN) {
+                    vector_type<FloatA, WmmaK> a_thread_vec;
+                    vector_type<FloatB, WmmaK> b_thread_vec;
+
+                    static_for<0, WmmaK, 1>{}([&](auto iK) {
+                        a_thread_vec.template AsType<FloatA>()(iK) =
+                            a_thread_buf[Number<a_thread_desc_.CalculateOffset(make_tuple(
+                                iK / A_K1, WmmaInnerloop + RepeatDiff, 0, 0, iK % A_K1))>{}];
+                        b_thread_vec.template AsType<FloatB>()(iK) =
+                            b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                make_tuple(iK / B_K1, iN, 0, 0, iK % B_K1))>{}];
+                    });
+                    using wmma_input_type_a = typename vector_type<FloatA, WmmaK>::type;
+                    using wmma_input_type_b = typename vector_type<FloatB, WmmaK>::type;
+
+                    constexpr index_t c_offset = c_thread_desc_.CalculateOffset(
+                        make_tuple(WmmaInnerloop + RepeatDiff, iN, 0));
+                    // s_nop();
+                    wmma_gemm.template Run(
+                        a_thread_vec.template AsType<wmma_input_type_a>()(Number<0>{}),
+                        b_thread_vec.template AsType<wmma_input_type_b>()(Number<0>{}),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    // s_nop();
+                });
+
+                // Read Consumed Next inner loop A
+                a_thread_copy_.Run(
+                    a_block_desc_k0_m0_m1_m2_k1,
+                    make_tuple(
+                        Number<iWmmaK / A_K1>{}, Number<WmmaInnerloop + RepeatDiff>{}, I0, I0, I0),
+                    a_block_buf,
+                    a_thread_desc_,
+                    make_tuple(I0, Number<WmmaInnerloop + RepeatDiff>{}, I0, I0, I0),
+                    a_thread_buf);
+
+                // Col Repeatation
+                static_for<WmmaInnerloop + 1 + RepeatDiff, MRepeat, 1>{}([&](auto iM) {
+                    vector_type<FloatA, WmmaK> a_thread_vec;
+                    vector_type<FloatB, WmmaK> b_thread_vec;
+
+                    static_for<0, WmmaK, 1>{}([&](auto iK) {
+                        a_thread_vec.template AsType<FloatA>()(iK) =
+                            a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                make_tuple(iK / A_K1, iM, 0, 0, iK % A_K1))>{}];
+                        b_thread_vec.template AsType<FloatB>()(iK) =
+                            b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                make_tuple(iK / B_K1, WmmaInnerloop, 0, 0, iK % B_K1))>{}];
+                    });
+                    using wmma_input_type_a = typename vector_type<FloatA, WmmaK>::type;
+                    using wmma_input_type_b = typename vector_type<FloatB, WmmaK>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(iM, WmmaInnerloop, 0));
+                    // s_nop();
+                    wmma_gemm.template Run(
+                        a_thread_vec.template AsType<wmma_input_type_a>()(Number<0>{}),
+                        b_thread_vec.template AsType<wmma_input_type_b>()(Number<0>{}),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    // s_nop();
+                });
+                // Read Consumed Next inner loop B
+                b_thread_copy_.Run(
+                    b_block_desc_k0_n0_n1_n2_k1,
+                    make_tuple(Number<iWmmaK / B_K1>{}, Number<WmmaInnerloop>{}, I0, I0, I0),
+                    b_block_buf,
+                    b_thread_desc_,
+                    make_tuple(I0, Number<WmmaInnerloop>{}, I0, I0, I0),
+                    b_thread_buf);
+            });
+
+            // Stage 1: Cut to Repeat Retangle to Square, assume MRepeat > NRepeat
+            static_for<0, RepeatDiff, 1>{}([&](auto iCut) {
+                static_for<0, NRepeat, 1>{}([&](auto iN) {
+                    vector_type<FloatA, WmmaK> a_thread_vec;
+                    vector_type<FloatB, WmmaK> b_thread_vec;
+
+                    static_for<0, WmmaK, 1>{}([&](auto iK) {
+                        a_thread_vec.template AsType<FloatA>()(iK) =
+                            a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                make_tuple(iK / A_K1, iCut, 0, 0, iK % A_K1))>{}];
+                        b_thread_vec.template AsType<FloatB>()(iK) =
+                            b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                make_tuple(iK / B_K1, iN, 0, 0, iK % B_K1))>{}];
+                    });
+                    using wmma_input_type_a = typename vector_type<FloatA, WmmaK>::type;
+                    using wmma_input_type_b = typename vector_type<FloatB, WmmaK>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(iCut, iN, 0));
+                    // s_nop();
+                    wmma_gemm.template Run(
+                        a_thread_vec.template AsType<wmma_input_type_a>()(Number<0>{}),
+                        b_thread_vec.template AsType<wmma_input_type_b>()(Number<0>{}),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    // s_nop();
+                });
+                if constexpr(KPerBlock > WmmaK)
+                {
+                    a_thread_copy_.Run(
+                        a_block_desc_k0_m0_m1_m2_k1,
+                        make_tuple(Number<(iWmmaK + WmmaK) / A_K1>{}, Number<iCut>{}, I0, I0, I0),
+                        a_block_buf,
+                        a_thread_desc_,
+                        make_tuple(I0, Number<iCut>{}, I0, I0, I0),
+                        a_thread_buf);
+                }
+            });
+        });
+
+        // Stage 2: Run FIFO fashion loopover in Square
+        static_for<0, NRepeat, 1>{}([&](auto WmmaInnerloop) {
+            // Row Repeatation
+            static_for<WmmaInnerloop, NRepeat, 1>{}([&](auto iN) {
+                vector_type<FloatA, WmmaK> a_thread_vec;
+                vector_type<FloatB, WmmaK> b_thread_vec;
+
+                static_for<0, WmmaK, 1>{}([&](auto iK) {
+                    a_thread_vec.template AsType<FloatA>()(iK) =
+                        a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                            make_tuple(iK / A_K1, WmmaInnerloop + RepeatDiff, 0, 0, iK % A_K1))>{}];
+                    b_thread_vec.template AsType<FloatB>()(iK) =
+                        b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                            make_tuple(iK / B_K1, iN, 0, 0, iK % B_K1))>{}];
+                });
+                using wmma_input_type_a = typename vector_type<FloatA, WmmaK>::type;
+                using wmma_input_type_b = typename vector_type<FloatB, WmmaK>::type;
+
+                constexpr index_t c_offset =
+                    c_thread_desc_.CalculateOffset(make_tuple(WmmaInnerloop + RepeatDiff, iN, 0));
+                // s_nop();
+                wmma_gemm.template Run(
+                    a_thread_vec.template AsType<wmma_input_type_a>()(Number<0>{}),
+                    b_thread_vec.template AsType<wmma_input_type_b>()(Number<0>{}),
+                    c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                // s_nop();
+            });
+
+            // Col Repeatation
+            static_for<WmmaInnerloop + 1 + RepeatDiff, MRepeat, 1>{}([&](auto iM) {
+                vector_type<FloatA, WmmaK> a_thread_vec;
+                vector_type<FloatB, WmmaK> b_thread_vec;
+
+                static_for<0, WmmaK, 1>{}([&](auto iK) {
+                    a_thread_vec.template AsType<FloatA>()(iK) =
+                        a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                            make_tuple(iK / A_K1, iM, 0, 0, iK % A_K1))>{}];
+                    b_thread_vec.template AsType<FloatB>()(iK) =
+                        b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                            make_tuple(iK / B_K1, WmmaInnerloop, 0, 0, iK % B_K1))>{}];
+                });
+                using wmma_input_type_a = typename vector_type<FloatA, WmmaK>::type;
+                using wmma_input_type_b = typename vector_type<FloatB, WmmaK>::type;
+
+                constexpr index_t c_offset =
+                    c_thread_desc_.CalculateOffset(make_tuple(iM, WmmaInnerloop, 0));
+                // s_nop();
+                wmma_gemm.template Run(
+                    a_thread_vec.template AsType<wmma_input_type_a>()(Number<0>{}),
+                    b_thread_vec.template AsType<wmma_input_type_b>()(Number<0>{}),
+                    c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                // s_nop();
+            });
+        });
+    }
+
+    protected:
+    // A[M0, M1, M2, K0 = WmmaK]
+    static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<WmmaK / A_K1>{}, Number<MRepeat>{}, I1, I1, Number<A_K1>{}));
+
+    // B[N0, N1, N2, K0 = WmmaK]
+    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<WmmaK / B_K1>{}, Number<NRepeat>{}, I1, I1, Number<B_K1>{}));
+
+    // C[M, N, NumRegWMMA]
+    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, wmma_gemm.GetRegSizePerWmma()));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatA,
+                                                         FloatA,
+                                                         decltype(a_block_desc_k0_m0_m1_m2_k1),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<WmmaK / A_K1, 1, 1, 1, A_K1>,
+                                                         Sequence<0, 1, 2, 3, 4>,
+                                                         4,
+                                                         A_K1,
+                                                         A_K1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatB,
+                                                         FloatB,
+                                                         decltype(b_block_desc_k0_n0_n1_n2_k1),
+                                                         decltype(b_thread_desc_),
+                                                         Sequence<WmmaK / B_K1, 1, 1, 1, B_K1>,
+                                                         Sequence<0, 1, 2, 3, 4>,
+                                                         4,
+                                                         B_K1,
+                                                         B_K1>;
+
+    AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
+    BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
new file mode 100644
index 0000000000..5403604a55
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
@@ -0,0 +1,1006 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/warp/xdlops_gemm.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+namespace ck {
+
+enum struct LoopScheduler
+{
+    Default,
+    Interwave,
+};
+
+constexpr LoopScheduler make_default_loop_scheduler()
+{
+#if CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING
+    return LoopScheduler::Interwave;
+#else
+    return LoopScheduler::Default;
+#endif // if CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING
+}
+
+template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
+__host__ __device__ static constexpr auto
+MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K(const TileDesc_K0_MN_K1&)
+{
+    constexpr index_t K0 = TileDesc_K0_MN_K1{}.GetLength(Number<0>{});
+    constexpr index_t K1 = TileDesc_K0_MN_K1{}.GetLength(Number<2>{});
+
+    return transform_tensor_descriptor(
+        TileDesc_K0_MN_K1{},
+        make_tuple(make_merge_transform_v3_division_mod(make_tuple(Number<K0>{}, Number<K1>{})),
+                   make_unmerge_transform(
+                       make_tuple(Number<MNXdlPerWave>{}, Number<MNWaves>{}, Number<MNPerXdl>{}))),
+        make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+        make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename AK0MK1BlockDesc,
+          typename BK0NK1BlockDesc,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack>
+struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    static constexpr index_t WaveSize = get_warp_size();
+
+    static constexpr index_t MPerBlock = AK0MK1BlockDesc{}.GetLength(I1);
+    static constexpr index_t NPerBlock = BK0NK1BlockDesc{}.GetLength(I1);
+    static constexpr index_t KPerBlock =
+        BK0NK1BlockDesc{}.GetLength(I0) * BK0NK1BlockDesc{}.GetLength(I2);
+
+    static constexpr index_t A_K0 = AK0MK1BlockDesc{}.GetLength(I0);
+    static constexpr index_t B_K0 = BK0NK1BlockDesc{}.GetLength(I0);
+    static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
+    static constexpr index_t B_K1 = BK0NK1BlockDesc{}.GetLength(I2);
+
+    static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack>{};
+
+    static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
+
+    static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
+    static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
+
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                              FloatAcc,
+                              MRepeat * NRepeat,
+                              xdlops_gemm.GetRegSizePerXdlops(),
+                              true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = ThisThreadBlock::GetThreadId();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+
+        const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
+
+        return make_multi_index(0, waveId_m, xdlops_a_idx[I1], KPerThread * xdlops_a_idx[I0]);
+    }
+
+    __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_n = wave_idx[I1];
+
+        const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
+
+        return make_multi_index(0, waveId_n, xdlops_b_idx[I1], KPerThread * xdlops_b_idx[I0]);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
+
+        constexpr auto mrepeat_mwave_mperxdl_to_m_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        constexpr auto nrepeat_nwave_nperxdl_to_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        const index_t c_thread_m = mrepeat_mwave_mperxdl_to_m_adaptor.CalculateBottomIndex(
+            make_multi_index(m0, waveId_m, blk_idx[I0]))[I0];
+
+        const index_t c_thread_n = nrepeat_nwave_nperxdl_to_n_adaptor.CalculateBottomIndex(
+            make_multi_index(n0, waveId_n, blk_idx[I1]))[I0];
+
+        return make_multi_index(c_thread_m, c_thread_n);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
+
+        return make_tuple(Number<m0>{},
+                          Number<n0>{},
+                          waveId_m,
+                          waveId_n,
+                          blk_idx[I0],
+                          blk_idx[I1],
+                          blk_idx[I2],
+                          blk_idx[I3]);
+    }
+
+    __host__ __device__ BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1()
+    {
+        static_assert(AK0MK1BlockDesc::IsKnownAtCompileTime() &&
+                          BK0NK1BlockDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
+                      "ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
+
+        static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
+                      "wrong!");
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_block_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_G_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+    {
+        const auto G = c_grid_desc_g_m_n.GetLength(I0);
+        const auto M = c_grid_desc_g_m_n.GetLength(I1);
+        const auto N = c_grid_desc_g_m_n.GetLength(I2);
+
+        const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_g_m_n,
+            make_tuple(make_pass_through_transform(G),
+                       make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_grid_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto MakeABlockDescriptor_M0_M1_M2_K()
+    {
+        return transform_tensor_descriptor(
+            AK0MK1BlockDesc{},
+            make_tuple(
+                make_merge_transform_v3_division_mod(make_tuple(Number<A_K0>{}, Number<A_K1>{})),
+                make_unmerge_transform(
+                    make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerXDL>{}))),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+    }
+
+    __host__ __device__ static constexpr auto MakeBBlockDescriptor_N0_N1_N2_K()
+    {
+        return transform_tensor_descriptor(
+            BK0NK1BlockDesc{},
+            make_tuple(
+                make_merge_transform_v3_division_mod(make_tuple(Number<B_K0>{}, Number<B_K1>{})),
+                make_unmerge_transform(
+                    make_tuple(Number<NRepeat>{}, Number<NWaves>{}, Number<NPerXDL>{}))),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+    }
+
+    static constexpr auto a_block_desc_m0_m1_m2_k = MakeABlockDescriptor_M0_M1_M2_K();
+    static constexpr auto b_block_desc_n0_n1_n2_k = MakeBBlockDescriptor_N0_N1_N2_K();
+
+    template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
+    __device__ void Run(const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_block_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
+            // read A
+            a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                               make_tuple(m0, I0, I0, I0),
+                               a_block_buf,
+                               a_thread_desc_,
+                               make_tuple(I0, I0, I0, I0),
+                               a_thread_buf);
+
+            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                // read B
+                b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                   make_tuple(n0, I0, I0, I0),
+                                   b_block_buf,
+                                   b_thread_desc_,
+                                   make_tuple(I0, I0, I0, I0),
+                                   b_thread_buf);
+
+                static_for<0, KPerThread, KPack>{}([&](auto k) {
+                    vector_type<FloatAB, KPack> a_thread_vec;
+                    vector_type<FloatAB, KPack> b_thread_vec;
+
+                    static_for<0, KPack, 1>{}([&](auto i) {
+                        a_thread_vec.template AsType<FloatAB>()(i) =
+                            a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                make_multi_index(0, 0, 0, k + i))>{}];
+                        b_thread_vec.template AsType<FloatAB>()(i) =
+                            b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                make_multi_index(0, 0, 0, k + i))>{}];
+                    });
+
+                    using mfma_input_type =
+                        typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_multi_index(m0, n0, 0));
+
+                    xdlops_gemm.template Run(
+                        a_thread_vec.template AsType<mfma_input_type>(),
+                        b_thread_vec.template AsType<mfma_input_type>(),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+        });
+    }
+
+    protected:
+    // A[M0, M1, M2, KPerThread]
+    static constexpr auto a_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
+
+    // B[N0, N1, N2, KPerThread]
+    static constexpr auto b_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
+
+    // C[M, N, NumRegXdlops]
+    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(a_block_desc_m0_m1_m2_k),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, KPerThread>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         A_K1,
+                                                         A_K1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(b_block_desc_n0_n1_n2_k),
+                                                         decltype(b_thread_desc_),
+                                                         Sequence<1, 1, 1, KPerThread>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         B_K1,
+                                                         B_K1>;
+
+    AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
+    BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
+};
+
+// Note: To facilitate the inter-wave loop scheduler, we need to explicitly set the macro
+// CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING=1 as a few intrinsics are not yet available in
+// the latest ROCm release. For unsupported compilers, inter-wave loop scheduler falls back to the
+// default loop scheduler which is given by the macro CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING=0
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename AK0MK1BlockDesc,
+          typename BK0NK1BlockDesc,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack,
+          index_t NumMacClusters = CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS>
+struct BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
+    : public BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                 FloatAB,
+                                                                 FloatAcc,
+                                                                 AK0MK1BlockDesc,
+                                                                 BK0NK1BlockDesc,
+                                                                 MPerXDL,
+                                                                 NPerXDL,
+                                                                 MRepeat,
+                                                                 NRepeat,
+                                                                 KPack>
+{
+    using Base = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                     FloatAB,
+                                                                     FloatAcc,
+                                                                     AK0MK1BlockDesc,
+                                                                     BK0NK1BlockDesc,
+                                                                     MPerXDL,
+                                                                     NPerXDL,
+                                                                     MRepeat,
+                                                                     NRepeat,
+                                                                     KPack>;
+
+#if CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING
+    using Base::a_block_desc_m0_m1_m2_k;
+    using Base::A_K1;
+    using Base::b_block_desc_n0_n1_n2_k;
+    using Base::B_K1;
+    using Base::c_thread_buf_;
+    using Base::c_thread_desc_;
+    using Base::CalculateAThreadOriginDataIndex;
+    using Base::CalculateBThreadOriginDataIndex;
+    using Base::I0;
+    using Base::I1;
+    using Base::KPerThread;
+    using Base::xdlops_gemm;
+
+    static constexpr index_t KPerInnerLoop = math::max(KPerThread / NumMacClusters, KPack);
+
+    // 2-wave optimized blockwise gemm
+    template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
+    __device__ void Run(const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_block_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        static_for<0, KPerThread, KPerInnerLoop>{}([&](auto k) {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                // read A
+                a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                   make_tuple(m0, I0, I0, k),
+                                   a_block_buf,
+                                   a_thread_desc_,
+                                   make_tuple(m0, I0, I0, I0),
+                                   a_thread_buf);
+            });
+            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                // read B
+                b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                   make_tuple(n0, I0, I0, k),
+                                   b_block_buf,
+                                   b_thread_desc_,
+                                   make_tuple(n0, I0, I0, I0),
+                                   b_thread_buf);
+            });
+            __builtin_amdgcn_sched_barrier(0);
+            // NOTE: Synchronize threads in a workgroup at the start of each MAC cluster, but except
+            // the first, as we can shorten non-MAC cluster a bit and there's no observable negative
+            // impact. The desired effect is waves in a workgroup executing MAC in sync. This avoids
+            // some out-of-sync waves hijacking MAC resource from other workgroups and reducing the
+            // chance of latency hiding by waiting for the rest of the workgroup at the eventual
+            // sync point.
+            if constexpr(k.value != 0 || KPerInnerLoop == KPerThread)
+            {
+                asm volatile("s_barrier" ::);
+                __builtin_amdgcn_sched_barrier(0);
+            }
+            static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<FloatAB, KPack> a_thread_vec;
+                        vector_type<FloatAB, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto i) {
+                            a_thread_vec.template AsType<FloatAB>()(i) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, 0, 0, k_ + i))>{}];
+                            b_thread_vec.template AsType<FloatAB>()(i) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, 0, 0, k_ + i))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        // The block_sync_lds() here performs double duty:
+                        // A) safeguard against data hazard because barrier from blockwise_gemm is
+                        // moved here B) reduce VMEM FIFO congestion by applying small delays to
+                        // different wavefronts It is performed near the end of MAC cluster to
+                        // minimize lgkmcnt penalty
+                        if constexpr(k.value == KPerThread - KPerInnerLoop &&
+                                     k_.value == KPerInnerLoop - KPack && m0.value == MRepeat - 1 &&
+                                     n0.value == NRepeat - 1)
+                        {
+                            __builtin_amdgcn_sched_barrier(0);
+                            block_sync_lds();
+                            __builtin_amdgcn_sched_barrier(0);
+                        }
+
+                        // TODO: insert setprio in more precise manner since we
+                        // could have more than >1 MFMA instructions in single call
+                        xdlops_gemm.template Run(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0)
+                        {
+                            __builtin_amdgcn_sched_barrier(0);
+                            __builtin_amdgcn_s_setprio(1);
+                            __builtin_amdgcn_sched_barrier(0);
+                        }
+                    });
+                });
+            });
+            __builtin_amdgcn_sched_barrier(0);
+            __builtin_amdgcn_s_setprio(0);
+            __builtin_amdgcn_sched_barrier(0);
+        });
+    }
+
+    protected:
+    // A[M0, M1, M2, KPerInnerLoop]
+    static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, I1, I1, Number<KPerInnerLoop>{}));
+
+    // B[N0, N1, N2, KPerInnerLoop]
+    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<NRepeat>{}, I1, I1, Number<KPerInnerLoop>{}));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(a_block_desc_m0_m1_m2_k),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, KPerInnerLoop>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         A_K1,
+                                                         A_K1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(b_block_desc_n0_n1_n2_k),
+                                                         decltype(b_thread_desc_),
+                                                         Sequence<1, 1, 1, KPerInnerLoop>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         B_K1,
+                                                         B_K1>;
+
+    AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
+    BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
+
+#endif // #if CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING
+};
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename AK0MK1BlockDesc,
+          typename BK0NK1BlockDesc,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack,
+          LoopScheduler LoopSched>
+constexpr auto BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector()
+{
+    if constexpr(LoopSched == LoopScheduler::Default)
+    {
+        return BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                   FloatAB,
+                                                                   FloatAcc,
+                                                                   AK0MK1BlockDesc,
+                                                                   BK0NK1BlockDesc,
+                                                                   MPerXDL,
+                                                                   NPerXDL,
+                                                                   MRepeat,
+                                                                   NRepeat,
+                                                                   KPack>{};
+    }
+    else if constexpr(LoopSched == LoopScheduler::Interwave)
+    {
+        return BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                            FloatAB,
+                                                                            FloatAcc,
+                                                                            AK0MK1BlockDesc,
+                                                                            BK0NK1BlockDesc,
+                                                                            MPerXDL,
+                                                                            NPerXDL,
+                                                                            MRepeat,
+                                                                            NRepeat,
+                                                                            KPack>{};
+    }
+};
+
+/**
+ * @brief Blockwise gemm
+ *
+ * Supports
+ * 1. regular XDL output M2_M3_M4_M2 and transposed XDL output M2_N2_N3_N4
+ * 2. decoupled input tile descriptor and mma tile descriptor in order to support both vgpr and LDS
+ * source buffer
+ * 3. configurable k index starting position and step size after each FMA/XDL instruction
+ */
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack,
+          bool TransposeC = false,
+          index_t AMmaKStride =
+              KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, TransposeC>{}.K0PerXdlops,
+          index_t BMmaKStride =
+              KPack* XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, TransposeC>{}.K0PerXdlops>
+struct BlockwiseGemmXdlops_v2
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    static constexpr index_t WaveSize = get_warp_size();
+
+    static constexpr index_t A_K0 = ATileDesc{}.GetLength(I0);
+    static constexpr index_t B_K0 = BTileDesc{}.GetLength(I0);
+    static constexpr index_t A_K1 = ATileDesc{}.GetLength(I2);
+    static constexpr index_t B_K1 = BTileDesc{}.GetLength(I2);
+
+    static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack, TransposeC>{};
+
+    static constexpr index_t KPerThread = KPerBlock / xdlops_gemm.K0PerXdlops;
+
+    static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
+    static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
+
+    static_assert(KPerThread % KPack == 0,
+                  "Wrong KPack setting; try increasing KPerThread or decreasing KPack");
+
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                              FloatAcc,
+                              MRepeat * NRepeat,
+                              xdlops_gemm.GetRegSizePerXdlops(),
+                              true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = ThisThreadBlock::GetThreadId();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+
+        const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
+
+        return make_multi_index(0, waveId_m, xdlops_a_idx[I1], KPack * xdlops_a_idx[I0]);
+    }
+
+    __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_n = wave_idx[I1];
+
+        const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
+
+        return make_multi_index(0, waveId_n, xdlops_b_idx[I1], KPack * xdlops_b_idx[I0]);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
+
+        constexpr auto mrepeat_mwave_mperxdl_to_m_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        constexpr auto nrepeat_nwave_nperxdl_to_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        const index_t c_thread_m = mrepeat_mwave_mperxdl_to_m_adaptor.CalculateBottomIndex(
+            make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
+        const index_t c_thread_n = nrepeat_nwave_nperxdl_to_n_adaptor.CalculateBottomIndex(
+            make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
+
+        return make_tuple(c_thread_m, c_thread_n);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex8D(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk4D(xdlops_i, blk_i);
+
+        return make_tuple(
+            m0, n0, waveId_m, waveId_n, blk_idx[I0], blk_idx[I1], blk_idx[I2], blk_idx[I3]);
+    }
+
+    using Array4 = decltype(CalculateAThreadOriginDataIndex());
+
+    __host__ __device__ BlockwiseGemmXdlops_v2(Array4 a_origin = CalculateAThreadOriginDataIndex(),
+                                               Array4 b_origin = CalculateBThreadOriginDataIndex())
+        : a_thread_copy_(a_origin), b_thread_copy_(b_origin)
+    {
+        static_assert(AMmaTileDesc::IsKnownAtCompileTime() && BMmaTileDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(ThisThreadBlock::GetNumOfThread() == MWaves * NWaves * WaveSize,
+                      "ThisThreadBlock::GetNumOfThread() != MWaves * NWaves * WaveSize\n");
+
+        static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
+                      "wrong!");
+    }
+
+    __host__ __device__ BlockwiseGemmXdlops_v2(const BlockwiseGemmXdlops_v2& other)
+        : a_thread_copy_(other.a_origin), b_thread_copy_(other.b_origin)
+    {
+    }
+
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, N, M0, M1, M2));
+    }
+
+    // XDL output supporting C_xdl = A_xdl * B_xdl
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    // transposed XDL output supporting C_xdl' = B_xdl' * A_xdl'
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    // XDL output supporting C_xdl = A_xdl * B_xdl
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_block_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_G_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+    {
+        const auto G = c_grid_desc_g_m_n.GetLength(I0);
+        const auto M = c_grid_desc_g_m_n.GetLength(I1);
+        const auto N = c_grid_desc_g_m_n.GetLength(I2);
+
+        const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_g_m_n,
+            make_tuple(make_pass_through_transform(G),
+                       make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_grid_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    static constexpr AMmaTileDesc a_block_desc_m0_m1_m2_k;
+    static constexpr BMmaTileDesc b_block_desc_n0_n1_n2_k;
+
+    template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
+    __device__ void Run(const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_block_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        static_for<0, KPerThread / KPack, 1>{}([&](auto k) { // k=0,1,2 instead of k=0,kpack*1, ...
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                // read A
+                a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                   make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                   a_block_buf,
+                                   a_thread_desc_,
+                                   make_tuple(I0, I0, I0, I0),
+                                   a_thread_buf);
+
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    // read B
+                    b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                       make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                       b_block_buf,
+                                       b_thread_desc_,
+                                       make_tuple(I0, I0, I0, I0),
+                                       b_thread_buf);
+                    vector_type<FloatAB, KPack> a_thread_vec;
+                    vector_type<FloatAB, KPack> b_thread_vec;
+
+                    static_for<0, KPack, 1>{}([&](auto i) {
+                        a_thread_vec.template AsType<FloatAB>()(i) = a_thread_buf
+                            [Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, i))>{}];
+                        b_thread_vec.template AsType<FloatAB>()(i) = b_thread_buf
+                            [Number<b_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, i))>{}];
+                    });
+
+                    using mfma_input_type =
+                        typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    xdlops_gemm.template Run(
+                        a_thread_vec.template AsType<mfma_input_type>(),
+                        b_thread_vec.template AsType<mfma_input_type>(),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+        });
+    }
+
+    protected:
+    // A[M0, M1, M2, KPack]
+    static constexpr auto a_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPack>{}));
+
+    // B[N0, N1, N2, KPack]
+    static constexpr auto b_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPack>{}));
+
+    // C[M, N, NumRegXdlops]
+    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(a_block_desc_m0_m1_m2_k),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         A_K1,
+                                                         A_K1>;
+
+    using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(b_block_desc_n0_n1_n2_k),
+                                                         decltype(b_thread_desc_),
+                                                         Sequence<1, 1, 1, KPack>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         B_K1,
+                                                         B_K1>;
+
+    AThreadCopy a_thread_copy_;
+    BThreadCopy b_thread_copy_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp
new file mode 100644
index 0000000000..8ae1ba3f34
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp
@@ -0,0 +1,321 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/warp/xdlops_gemm.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+namespace ck {
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename AK0MK1BlockDesc,
+          typename BK0K0BN0N1N2N3K1BlockDesc,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack>
+struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1r1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr index_t WaveSize = 64;
+
+    static constexpr index_t KPerBlock = K0PerBlock * KPack;
+
+    static constexpr index_t A_K0 = AK0MK1BlockDesc{}.GetLength(I0);
+    static constexpr index_t A_K1 = AK0MK1BlockDesc{}.GetLength(I2);
+
+    static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, KPack>{};
+
+    static constexpr index_t KPerThread  = KPerBlock / xdlops_gemm.K0PerXdlops;
+    static constexpr index_t K0PerThread = K0PerBlock / xdlops_gemm.K0PerXdlops;
+
+    static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
+    static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
+
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                              FloatAcc,
+                              MRepeat * NRepeat,
+                              xdlops_gemm.GetRegSizePerXdlops(),
+                              true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = get_thread_local_1d_id();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+
+        const auto xdlops_a_idx = xdlops_gemm.CalculateAThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_m, xdlops_a_idx[I1], KPerThread * xdlops_a_idx[I0]);
+    }
+
+    __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_n = wave_idx[I1];
+
+        const auto xdlops_b_idx = xdlops_gemm.CalculateBThreadOriginDataIndex();
+
+        return make_tuple(0, waveId_n, xdlops_b_idx[I1], KPerThread * xdlops_b_idx[I0]);
+    }
+
+    template <index_t m0, index_t n0, index_t xdlops_i, index_t blk_i>
+    __device__ static auto
+        CalculateCThreadOriginDataIndex(Number<m0>, Number<n0>, Number<xdlops_i>, Number<blk_i>)
+    {
+        const auto wave_idx = GetWaveIdx();
+
+        const auto waveId_m = wave_idx[I0];
+        const auto waveId_n = wave_idx[I1];
+
+        const auto blk_idx = xdlops_gemm.GetBeginOfThreadBlk(xdlops_i, blk_i);
+
+        constexpr auto mrepeat_mwave_mperxdl_to_m_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(MRepeat, MWaves, MPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        constexpr auto nrepeat_nwave_nperxdl_to_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(NRepeat, NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 2>{}));
+
+        const index_t c_thread_m = mrepeat_mwave_mperxdl_to_m_adaptor.CalculateBottomIndex(
+            make_tuple(m0, waveId_m, blk_idx[I0]))[I0];
+        const index_t c_thread_n = nrepeat_nwave_nperxdl_to_n_adaptor.CalculateBottomIndex(
+            make_tuple(n0, waveId_n, blk_idx[I1]))[I0];
+
+        return make_tuple(c_thread_m, c_thread_n);
+    }
+
+    __host__ __device__ BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1r1()
+    {
+        static_assert(AK0MK1BlockDesc::IsKnownAtCompileTime() &&
+                          BK0K0BN0N1N2N3K1BlockDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(BlockSize == MWaves * NWaves * WaveSize,
+                      "BlockSize != MWaves * NWaves * WaveSize\n");
+
+        static_assert(MPerBlock % (MPerXDL * MRepeat) == 0 && NPerBlock % (NPerXDL * NRepeat) == 0,
+                      "wrong!");
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_m0_m1_m2_n_tblk_lens = xdlops_gemm.GetCM0M1M2NThreadBlkLengths();
+
+        constexpr auto M0 = c_m0_m1_m2_n_tblk_lens[I0];
+        constexpr auto M1 = c_m0_m1_m2_n_tblk_lens[I1];
+        constexpr auto M2 = c_m0_m1_m2_n_tblk_lens[I2];
+        constexpr auto N  = c_m0_m1_m2_n_tblk_lens[I3];
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1, Number<MRepeat>{}, Number<NRepeat>{}, I1, I1, M0, M1, M2, N));
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_block_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2()
+    {
+        constexpr auto c_block_desc_g_m0_n0_m1_n1_m2_n2 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           Number<MRepeat>{},
+                                                           Number<NRepeat>{},
+                                                           Number<MWaves>{},
+                                                           Number<NWaves>{},
+                                                           Number<MPerXDL>{},
+                                                           Number<NPerXDL>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_block_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto c_grid_desc_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}));
+
+        return xdlops_gemm.MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m0_n0_m1_n1_m2_n2);
+    }
+
+    template <typename CGridDesc_G_M_N>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+    {
+        const auto G = c_grid_desc_g_m_n.GetLength(I0);
+        const auto M = c_grid_desc_g_m_n.GetLength(I1);
+        const auto N = c_grid_desc_g_m_n.GetLength(I2);
+
+        const auto c_grid_desc_g_m0_n0_m1_n1_m2_n2 = transform_tensor_descriptor(
+            c_grid_desc_g_m_n,
+            make_tuple(make_pass_through_transform(G),
+                       make_unmerge_transform(make_tuple(M / (MWaves * MPerXDL), MWaves, MPerXDL)),
+                       make_unmerge_transform(make_tuple(N / (NWaves * NPerXDL), NWaves, NPerXDL))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 3, 5>{}, Sequence<2, 4, 6>{}));
+
+        return xdlops_gemm.MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+            c_grid_desc_g_m0_n0_m1_n1_m2_n2);
+    }
+
+    __host__ __device__ static constexpr auto MakeABlockDescriptor_M0_M1_M2_K()
+    {
+        return transform_tensor_descriptor(
+            AK0MK1BlockDesc{},
+            make_tuple(
+                make_merge_transform_v3_division_mod(make_tuple(Number<A_K0>{}, Number<A_K1>{})),
+                make_unmerge_transform(
+                    make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerXDL>{}))),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
+    }
+
+    __device__ void MoveABlockSliceWindow()
+    {
+        a_thread_copy_.MoveSrcSliceWindow(a_block_desc_m0_m1_m2_k,
+                                          make_multi_index(0, 0, 0, K0PerBlock * KPack));
+    }
+    __device__ void ResetABlockStartWindow()
+    {
+        a_thread_copy_.SetSrcCoord(CalculateAThreadOriginDataIndex());
+    }
+
+    static constexpr auto a_block_desc_m0_m1_m2_k = MakeABlockDescriptor_M0_M1_M2_K();
+
+    template <typename ABlockBuffer, typename BBlockBuffer, typename CThreadBuffer>
+    __device__ void Run(const ABlockBuffer& a_block_buf,
+                        const BBlockBuffer& b_thread_buf,
+                        CThreadBuffer& c_thread_buf) const
+    {
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a_thread_desc_.GetElementSpaceSize());
+
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
+            // read A
+            a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                               make_tuple(m0, I0, I0, I0),
+                               a_block_buf,
+                               a_thread_desc_,
+                               make_tuple(I0, I0, I0, I0),
+                               a_thread_buf);
+
+            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                // read B
+                static_for<0, KPerThread, KPack>{}([&](auto k) {
+                    vector_type<FloatAB, KPack> a_thread_vec;
+                    vector_type<FloatAB, KPack> b_thread_vec;
+                    constexpr index_t k0 = k / KPack;
+                    static_for<0, KPack, 1>{}([&](auto i) {
+                        a_thread_vec.template AsType<FloatAB>()(i) = a_thread_buf
+                            [Number<a_thread_desc_.CalculateOffset(make_tuple(0, 0, 0, k + i))>{}];
+                        b_thread_vec.template AsType<FloatAB>()(i) = b_thread_buf
+                            [Number<b_thread_desc_.CalculateOffset(make_tuple(k0, n0, i))>{}];
+                    });
+
+                    using mfma_input_type =
+                        typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset =
+                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                    xdlops_gemm.template Run(
+                        a_thread_vec.template AsType<mfma_input_type>(),
+                        b_thread_vec.template AsType<mfma_input_type>(),
+                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                });
+            });
+        });
+    }
+
+    private:
+    // A[M0, M1, M2, KPerThread]
+    static constexpr auto a_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I1, Number<KPerThread>{}));
+
+    // B[N0, N1, N2, KPerThread]
+    static constexpr auto b_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<K0PerThread>{}, // KPerThread
+                                                       Number<NRepeat>{},     // repeat
+                                                       Number<KPack>{}));
+
+    // C[M, N, NumRegXdlops]
+    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, xdlops_gemm.GetRegSizePerXdlops()));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
+                                                         FloatAB,
+                                                         decltype(a_block_desc_m0_m1_m2_k),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, KPerThread>,
+                                                         Sequence<0, 1, 2, 3>,
+                                                         3,
+                                                         A_K1,
+                                                         A_K1>;
+
+    AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp b/include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp
new file mode 100644
index 0000000000..3d37560988
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp
@@ -0,0 +1,125 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+
+namespace ck {
+
+/**
+ * @brief Blockwise softmax
+ *
+ * @tparam BlockSize Block size
+ * @tparam AccDataType Accumulator data type
+ * @tparam ThreadMap_M_K Thread id to m_k
+ * @tparam ThreadClusterDesc_M_K Threadwise cluster descriptor
+ * @tparam ThreadSliceDesc_M_K Threadwise slices descriptor
+ * @tparam IgnoreNaN Flag to ignore NaN, false by default
+ */
+template <index_t BlockSize,
+          typename AccDataType,
+          typename ThreadMap_M_K, // thread_id to m_k
+          typename ThreadClusterDesc_M_K,
+          typename ThreadSliceDesc_M_K,
+          bool IgnoreNaN = false>
+struct BlockwiseSoftmax
+{
+    static constexpr auto I0         = Number<0>{};
+    static constexpr auto I1         = Number<1>{};
+    static constexpr index_t MRepeat = ThreadSliceDesc_M_K{}.GetLength(I0);
+    static constexpr index_t KRepeat = ThreadSliceDesc_M_K{}.GetLength(I1);
+
+    using ThreadSliceDesc_M = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(ThreadSliceDesc_M_K{}.GetLength(I0))));
+
+    using ThreadwiseMaxReduce = typename conditional<
+        IgnoreNaN,
+        ThreadwiseReduction<AccDataType,
+                            ThreadSliceDesc_M_K,
+                            ThreadSliceDesc_M,
+                            reduce::Max,
+                            false,
+                            detail::AccumulateWithNanIgnore<reduce::Max, AccDataType>>,
+        ThreadwiseReduction<AccDataType,
+                            ThreadSliceDesc_M_K,
+                            ThreadSliceDesc_M,
+                            reduce::Max,
+                            false>>::type;
+
+    using ThreadwiseSumReduce = typename conditional<
+        IgnoreNaN,
+        ThreadwiseReduction<AccDataType,
+                            ThreadSliceDesc_M_K,
+                            ThreadSliceDesc_M,
+                            reduce::Add,
+                            false,
+                            detail::AccumulateWithNanIgnore<reduce::Add, AccDataType>>,
+        ThreadwiseReduction<AccDataType,
+                            ThreadSliceDesc_M_K,
+                            ThreadSliceDesc_M,
+                            reduce::Add,
+                            false>>::type;
+
+    using ThreadClusterLengths_M_K = decltype(to_sequence(ThreadClusterDesc_M_K{}.GetLengths()));
+
+    using BlockwiseMaxReduce = PartitionedBlockwiseReduction_v2<AccDataType,
+                                                                BlockSize,
+                                                                ThreadClusterLengths_M_K,
+                                                                ThreadMap_M_K,
+                                                                reduce::Max,
+                                                                false>;
+
+    using BlockwiseSumReduce = PartitionedBlockwiseReduction_v2<AccDataType,
+                                                                BlockSize,
+                                                                ThreadClusterLengths_M_K,
+                                                                ThreadMap_M_K,
+                                                                reduce::Add,
+                                                                false>;
+
+    using BufferType = StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MRepeat, true>;
+
+    template <typename CThreadBuffer, typename WorkspaceBuffer>
+    __host__ __device__ void Run(CThreadBuffer& in_thread_buf, WorkspaceBuffer& reduce_work_buf)
+    {
+        // find max value
+        static_for<0, MRepeat, 1>{}([&](auto I) {
+            max_value_buf(I) = reduce::Max::template GetIdentityValue<AccDataType>();
+        });
+        ThreadwiseMaxReduce::Reduce(in_thread_buf, max_value_buf);
+        static_for<0, MRepeat, 1>{}([&](auto I) {
+            BlockwiseMaxReduce::Reduce(reduce_work_buf, max_value_buf(I));
+            block_sync_lds();
+        });
+
+        // calculate exp for elements, P=exp(s-max)
+        static_for<0, MRepeat, 1>{}([&](auto iM) {
+            static_for<0, KRepeat, 1>{}([&](auto iK) {
+                auto offset = Number<ThreadSliceDesc_M_K{}.CalculateOffset(make_tuple(iM, iK))>{};
+                in_thread_buf(offset) = IgnoreNaN && ck::math::isnan(in_thread_buf[offset])
+                                            ? 0
+                                            : math::exp(in_thread_buf[offset] - max_value_buf(iM));
+            });
+        });
+
+        // sum data
+        static_for<0, MRepeat, 1>{}([&](auto I) {
+            sum_value_buf(I) = reduce::Add::template GetIdentityValue<AccDataType>();
+        });
+        ThreadwiseSumReduce::Reduce(in_thread_buf, sum_value_buf);
+        static_for<0, MRepeat, 1>{}([&](auto I) {
+            BlockwiseSumReduce::Reduce(reduce_work_buf, sum_value_buf(I));
+            block_sync_lds();
+        });
+    }
+
+    BufferType max_value_buf;
+    BufferType sum_value_buf;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp b/include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp
new file mode 100644
index 0000000000..d8da134a34
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp
@@ -0,0 +1,156 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp"
+
+namespace ck {
+
+// this version does following things to avoid scratch memory issue
+// 1. Use StaticallyIndexedArray instead of C array for thread buffer
+// 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
+// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
+template <index_t BlockSize,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename BlockSliceLengths,
+          typename ThreadSliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          typename SrcVectorTensorLengths,
+          typename DstVectorTensorLengths,
+          typename SrcVectorTensorContiguousDimOrder,
+          typename DstVectorTensorContiguousDimOrder,
+          bool ThreadTransferSrcResetCoordinateAfterRun,
+          bool ThreadTransferDstResetCoordinateAfterRun>
+struct BlockwiseTensorSliceTransfer_v5r1
+{
+    static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
+
+    using Index = MultiIndex<nDim>;
+
+    __device__ constexpr BlockwiseTensorSliceTransfer_v5r1(const SrcDesc& src_desc,
+                                                           const Index& src_block_slice_origin,
+                                                           const DstDesc& dst_desc,
+                                                           const Index& dst_block_slice_origin)
+        : threadwise_transfer_(
+              src_desc, make_zero_multi_index<nDim>(), dst_desc, make_zero_multi_index<nDim>())
+
+    {
+        static_assert(nDim == remove_cvref_t<SrcDesc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
+                          nDim == BlockSliceLengths::Size() && nDim == ThreadSliceLengths::Size() &&
+                          nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == SrcDimAccessOrder::Size() && nDim == DstDimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<BlockSliceLengths, decltype(ThreadSliceLengths{} * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(BlockSize >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! BlockSize too small");
+
+        if(BlockSize == thread_cluster_desc_.GetElementSize() or
+           get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(get_thread_local_1d_id()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * ThreadSliceLengths{};
+
+            threadwise_transfer_.SetSrcSliceOrigin(src_desc,
+                                                   src_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetDstSliceOrigin(dst_desc,
+                                                   dst_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    template <typename SrcBuffer>
+    __device__ void RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf)
+    {
+        if(BlockSize == thread_cluster_desc_.GetElementSize() or
+           get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.RunRead(src_desc, src_buf);
+        }
+    }
+
+    template <typename DstBuffer>
+    __device__ void RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf)
+    {
+        if(BlockSize == thread_cluster_desc_.GetElementSize() or
+           get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.RunWrite(dst_desc, dst_buf);
+        }
+    }
+
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc, const Index& step)
+    {
+        if(BlockSize == thread_cluster_desc_.GetElementSize() or
+           get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrcSliceWindow(src_desc, step);
+        }
+    }
+
+    // SrcMoveSliceWindowStepHack to control index calculation move slice window
+    template <typename SrcMoveSliceWindowStepHack>
+    __device__ void
+    MoveSrcSliceWindow(const SrcDesc& src_desc,
+                       const Index& step,
+                       const SrcMoveSliceWindowStepHack& src_move_slice_window_step_hack)
+    {
+        if(BlockSize == thread_cluster_desc_.GetElementSize() or
+           get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrcSliceWindow(
+                src_desc, step, src_move_slice_window_step_hack);
+        }
+    }
+
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
+    {
+        if(BlockSize == thread_cluster_desc_.GetElementSize() or
+           get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
+        }
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v5r1<ThreadSliceLengths,
+                                           DstInMemOp,
+                                           SrcData,
+                                           DstData,
+                                           SrcDesc,
+                                           DstDesc,
+                                           SrcDimAccessOrder,
+                                           DstDimAccessOrder,
+                                           SrcVectorTensorLengths,
+                                           DstVectorTensorLengths,
+                                           SrcVectorTensorContiguousDimOrder,
+                                           DstVectorTensorContiguousDimOrder,
+                                           ThreadTransferSrcResetCoordinateAfterRun,
+                                           ThreadTransferDstResetCoordinateAfterRun>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_welford.hpp b/include/ck/tensor_operation/gpu/block/blockwise_welford.hpp
new file mode 100644
index 0000000000..820a08fc48
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/blockwise_welford.hpp
@@ -0,0 +1,110 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/utility/get_shift.hpp"
+
+namespace ck {
+
+// clang-format off
+// Assume:
+//  1) work_buffer is buffer (typically LDS) allocated outside as workspace
+//  2) work_buffer has T elements, and space size is no less than 3*BlockSize
+//  3) mean_value, var_value and count is the input data in vgpr from each thread
+//  4) mean_value, var_value and count is the over-written reduced output in vgpr for each thread
+//  5) Merge mean and M from ThreadwiseWelford
+// clang-format on
+template <typename T,
+          index_t BlockSize,
+          typename ThreadClusterLengths_M_K,
+          typename ThreadClusterArrangeOrder,
+          bool GetActualVariance = true>
+struct BlockwiseWelford
+{
+    static_assert(BlockSize == ThreadClusterLengths_M_K::At(0) * ThreadClusterLengths_M_K::At(1),
+                  "The product of cluster lengths should be same as BlockSize!");
+
+    static constexpr auto BufferLength_M = ThreadClusterLengths_M_K::At(0);
+    static constexpr auto BufferLength_K = ThreadClusterLengths_M_K::At(1);
+
+    static constexpr auto block_buf_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<BufferLength_M>{}, Number<BufferLength_K>{}));
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    template <typename CountDataType>
+    __device__ static inline void
+    Merge(T& mean_a, T& var_a, CountDataType& count_a, T mean_b, T var_b, CountDataType count_b)
+    {
+        CountDataType count  = count_a + count_b;
+        T count_b_over_count = count == 0 ? type_convert<T>(0) : type_convert<T>(count_b) / count;
+        T delta              = mean_b - mean_a;
+        mean_a += delta * count_b_over_count;
+        var_a += var_b + delta * delta * count_a * count_b_over_count;
+        count_a = count;
+    }
+
+    template <typename CountDataType>
+    __device__ static void Run(T& mean_value, T& var_value, CountDataType& count)
+    {
+        __shared__ T mean_block_buf[BlockSize];
+        __shared__ T var_block_buf[BlockSize];
+        __shared__ CountDataType count_block_buf[BlockSize];
+
+        constexpr auto cluster_len_shift = get_shift<BufferLength_K>();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[Number<0>{}];
+        const auto thread_k_cluster_id = thread_cluster_idx[Number<1>{}];
+
+        index_t offset1 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx);
+
+        mean_block_buf[offset1]  = mean_value;
+        var_block_buf[offset1]   = var_value;
+        count_block_buf[offset1] = count;
+
+        block_sync_lds();
+
+        static_for<0, cluster_len_shift, 1>{}([&](auto I) {
+            constexpr index_t indOffset = 1 << (cluster_len_shift - 1 - I());
+
+            if(thread_k_cluster_id < indOffset)
+            {
+                index_t offset2 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx +
+                                                                     make_tuple(0, indOffset));
+
+                T mean1              = mean_block_buf[offset1];
+                T var1               = var_block_buf[offset1];
+                CountDataType count1 = count_block_buf[offset1];
+
+                T mean2              = mean_block_buf[offset2];
+                T var2               = var_block_buf[offset2];
+                CountDataType count2 = count_block_buf[offset2];
+
+                Merge(mean1, var1, count1, mean2, var2, count2);
+
+                mean_block_buf[offset1]  = mean1;
+                var_block_buf[offset1]   = var1;
+                count_block_buf[offset1] = count1;
+            }
+
+            block_sync_lds();
+        });
+
+        index_t offset = block_buf_desc_m_k.CalculateOffset(make_tuple(thread_m_cluster_id, 0));
+
+        count      = count_block_buf[offset];
+        mean_value = mean_block_buf[offset];
+
+        if constexpr(GetActualVariance)
+            var_value = var_block_buf[offset] / count;
+        else
+            var_value = var_block_buf[offset];
+    };
+};
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp b/include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp
new file mode 100644
index 0000000000..82667e2352
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp
@@ -0,0 +1,244 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/utility/get_shift.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+
+namespace ck {
+
+// clang-format off
+// Assume:
+//  1) work_buffer is buffer (typically LDS) allocated outside as workspace, does not include any in/out data
+//  2) work_buffer has AccDataType elements, and space size is no less than BlockSize
+//  3) in_out_value is the input data in vgpr from each thread
+//  4) in_out_value is the over-written reduced output in vgpr for each thread
+// clang-format on
+template <typename AccDataType,
+          index_t BlockSize,
+          typename ThreadClusterLengths_M_K,
+          typename ThreadClusterArrangeOrder,
+          typename OpReduce,
+          bool PropagateNan,
+          typename Accumulation =
+              detail::AccumulateWithNanCheck<PropagateNan, OpReduce, AccDataType>>
+struct PartitionedBlockwiseReduction
+{
+    static_assert(BlockSize == ThreadClusterLengths_M_K::At(0) * ThreadClusterLengths_M_K::At(1),
+                  "The product of cluster lengths should be same as BlockSize!");
+
+    static constexpr auto BufferLength_M = ThreadClusterLengths_M_K::At(0);
+    static constexpr auto BufferLength_K = ThreadClusterLengths_M_K::At(1);
+
+    static_assert(BufferLength_K > 1, "Parallel reduction need work on at least two elements");
+
+    static constexpr auto block_buf_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<BufferLength_M>{}, Number<BufferLength_K>{}));
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    template <typename BufferType>
+    __device__ static void Reduce(BufferType& work_buffer, AccDataType& in_out_value)
+    {
+        static_assert(is_same<typename BufferType::type, AccDataType>{},
+                      "Buffer data type should be consistent as AccDataType!");
+
+        constexpr auto cluster_len_shift = get_shift<BufferLength_K>();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[Number<0>{}];
+        const auto thread_k_cluster_id = thread_cluster_idx[Number<1>{}];
+
+        work_buffer(block_buf_desc_m_k.CalculateOffset(thread_cluster_idx)) = in_out_value;
+
+        __syncthreads();
+
+        static_for<0, cluster_len_shift, 1>{}([&](auto I) {
+            constexpr index_t indOffset = 1 << (cluster_len_shift - 1 - I());
+
+            if(thread_k_cluster_id < indOffset)
+            {
+                index_t offset1 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx);
+                index_t offset2 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx +
+                                                                     make_tuple(0, indOffset));
+
+                AccDataType opData1 = work_buffer[offset1];
+                AccDataType opData2 = work_buffer[offset2];
+                Accumulation::Calculate(opData1, opData2);
+                work_buffer(offset1) = opData1;
+            }
+
+            __syncthreads();
+        });
+
+        index_t offset = block_buf_desc_m_k.CalculateOffset(make_tuple(thread_m_cluster_id, 0));
+
+        in_out_value = work_buffer[offset];
+    };
+};
+
+// clang-format off
+// Assume:
+//  1) work_buffer is buffer (typically LDS) allocated outside as workspace, does not include any in/out data
+//  2) work_buffer has AccDataType elements, and space size is no less than BlockSize
+//  3) in_out_value is the input data in vgpr from each thread
+//  4) in_out_value is the over-written reduced output in vgpr for each thread
+// clang-format on
+template <typename AccDataType,
+          index_t BlockSize,
+          typename ThreadClusterLengths_M_K,
+          typename ThreadClusterDesc,
+          typename OpReduce,
+          bool PropagateNan,
+          typename Accumulation =
+              detail::AccumulateWithNanCheck<PropagateNan, OpReduce, AccDataType>>
+struct PartitionedBlockwiseReduction_v2
+{
+    static_assert(BlockSize == ThreadClusterLengths_M_K::At(0) * ThreadClusterLengths_M_K::At(1),
+                  "The product of cluster lengths should be same as BlockSize!");
+
+    static constexpr auto BufferLength_M = ThreadClusterLengths_M_K::At(0);
+    static constexpr auto BufferLength_K = ThreadClusterLengths_M_K::At(1);
+
+    static_assert(BufferLength_K > 1, "Parallel reduction need work on at least two elements");
+
+    static constexpr auto block_buf_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<BufferLength_M>{}, Number<BufferLength_K>{}));
+
+    static constexpr auto thread_cluster_desc = ThreadClusterDesc{};
+
+    template <typename BufferType>
+    __device__ static void Reduce(BufferType& work_buffer, AccDataType& in_out_value)
+    {
+        static_assert(is_same<typename BufferType::type, AccDataType>{},
+                      "Buffer data type should be consistent as AccDataType!");
+
+        constexpr auto cluster_len_shift = get_shift<BufferLength_K>();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[Number<0>{}];
+        const auto thread_k_cluster_id = thread_cluster_idx[Number<1>{}];
+
+        work_buffer(block_buf_desc_m_k.CalculateOffset(thread_cluster_idx)) = in_out_value;
+
+        __syncthreads();
+
+        static_for<0, cluster_len_shift, 1>{}([&](auto I) {
+            constexpr index_t indOffset = 1 << (cluster_len_shift - 1 - I());
+
+            if(thread_k_cluster_id < indOffset)
+            {
+                index_t offset1 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx);
+                index_t offset2 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx +
+                                                                     make_tuple(0, indOffset));
+
+                AccDataType opData1 = work_buffer[offset1];
+                AccDataType opData2 = work_buffer[offset2];
+                Accumulation::Calculate(opData1, opData2);
+                work_buffer(offset1) = opData1;
+            }
+
+            __syncthreads();
+        });
+
+        index_t offset = block_buf_desc_m_k.CalculateOffset(make_tuple(thread_m_cluster_id, 0));
+
+        in_out_value = work_buffer[offset];
+    };
+};
+
+// clang-format off
+// Assume:
+//  1) work_val_buffer/work_idx_buffer is buffer (typically LDS) allocated outside as workspace, does not include any in/out data
+//  2) work_val_buffer/work_idx_buffer has AccDataType/IndexDataType elements, and space size is no less than BlockSize
+//  3) in_out_value/in_out_index is the input data in vgpr from each thread
+//  4) in_out_value/in_out_index is the over-written reduced output in vgpr for each thread
+// clang-format on
+template <
+    typename AccDataType,
+    typename IndexDataType,
+    index_t BlockSize,
+    typename ThreadClusterLengths_M_K,
+    typename ThreadClusterArrangeOrder,
+    typename OpReduce,
+    bool PropagateNan,
+    typename Accumulation =
+        detail::AccumulateWithIndexAndNanCheck<PropagateNan, OpReduce, AccDataType, IndexDataType>>
+struct PartitionedBlockwiseReductionWithIndex
+{
+    static_assert(BlockSize == ThreadClusterLengths_M_K::At(0) * ThreadClusterLengths_M_K::At(1),
+                  "The product of cluster lengths should be same as BlockSize!");
+
+    static constexpr auto BufferLength_M = ThreadClusterLengths_M_K::At(0);
+    static constexpr auto BufferLength_K = ThreadClusterLengths_M_K::At(1);
+
+    static_assert(BufferLength_K > 1, "Parallel reduction need work on at least two elements");
+
+    static constexpr auto block_buf_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<BufferLength_M>{}, Number<BufferLength_K>{}));
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    // This interface accumulates on both data values and indices
+    template <typename BufferType, typename IdxBufferType>
+    __device__ static void Reduce(BufferType& work_val_buffer,
+                                  IdxBufferType& work_idx_buffer,
+                                  AccDataType& in_out_value,
+                                  IndexDataType& in_out_index)
+    {
+        static_assert(is_same<typename BufferType::type, AccDataType>{},
+                      "Buffer data type should be consistent as AccDataType!");
+        static_assert(is_same<typename IdxBufferType::type, IndexDataType>{},
+                      "Buffer data type should be consistent as IndexDataType!");
+
+        constexpr auto cluster_len_shift = get_shift<BufferLength_K>();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[Number<0>{}];
+        const auto thread_k_cluster_id = thread_cluster_idx[Number<1>{}];
+
+        work_val_buffer(block_buf_desc_m_k.CalculateOffset(thread_cluster_idx)) = in_out_value;
+        work_idx_buffer(block_buf_desc_m_k.CalculateOffset(thread_cluster_idx)) = in_out_index;
+
+        __syncthreads();
+
+        static_for<0, cluster_len_shift, 1>{}([&](auto I) {
+            constexpr index_t indOffset = 1 << I();
+
+            if(thread_k_cluster_id % (indOffset * 2) == 0)
+            {
+                index_t offset1 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx);
+                index_t offset2 = block_buf_desc_m_k.CalculateOffset(thread_cluster_idx +
+                                                                     make_tuple(0, indOffset));
+
+                AccDataType opData1      = work_val_buffer[offset1];
+                AccDataType opData2      = work_val_buffer[offset2];
+                IndexDataType currIndex1 = work_idx_buffer[offset1];
+                IndexDataType currIndex2 = work_idx_buffer[offset2];
+
+                Accumulation::Calculate(opData1, opData2, currIndex1, currIndex2);
+                work_val_buffer(offset1) = opData1;
+                work_idx_buffer(offset1) = currIndex1;
+            }
+
+            __syncthreads();
+        });
+
+        index_t offset = block_buf_desc_m_k.CalculateOffset(make_tuple(thread_m_cluster_id, 0));
+
+        in_out_value = work_val_buffer[offset];
+        in_out_index = work_idx_buffer[offset];
+    };
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp
new file mode 100644
index 0000000000..b58aef90b8
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp
@@ -0,0 +1,197 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp"
+
+namespace ck {
+
+/**
+ * @brief Blockwise data transfer
+ *
+ * This version does following things to avoid scratch memory issue
+ * 1. Use StaticallyIndexedArray instead of C array for thread buffer
+ * 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
+ * 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
+ *
+ */
+template <typename ThreadGroup,
+          typename SrcElementwiseOperation,
+          typename DstElementwiseOperation,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename BlockSliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector,
+          index_t SrcScalarStrideInVector,
+          index_t DstScalarStrideInVector,
+          bool ThreadTransferSrcResetCoordinateAfterRun,
+          bool ThreadTransferDstResetCoordinateAfterRun,
+          index_t NumThreadScratch = 1>
+struct ThreadGroupTensorSliceTransfer_v4r1
+{
+    static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
+
+    static constexpr auto thread_slice_lengths = BlockSliceLengths{} / ThreadClusterLengths{};
+
+    using Index = MultiIndex<nDim>;
+
+#if 1 // debug
+    __host__ __device__ constexpr ThreadGroupTensorSliceTransfer_v4r1() : threadwise_transfer_{} {}
+#endif
+
+    __device__ constexpr ThreadGroupTensorSliceTransfer_v4r1(
+        const SrcDesc& src_desc,
+        const Index& src_block_slice_origin,
+        const SrcElementwiseOperation& src_element_op,
+        const DstDesc& dst_desc,
+        const Index& dst_block_slice_origin,
+        const DstElementwiseOperation& dst_element_op)
+        : threadwise_transfer_(src_desc,
+                               make_zero_multi_index<nDim>(),
+                               src_element_op,
+                               dst_desc,
+                               make_zero_multi_index<nDim>(),
+                               dst_element_op)
+
+    {
+        static_assert(nDim == remove_cvref_t<SrcDesc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
+                          nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == SrcDimAccessOrder::Size() && nDim == DstDimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<BlockSliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! ThreadGroup::GetNumOfThread() too small");
+
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(ThreadGroup::GetThreadId()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            threadwise_transfer_.SetSrcSliceOrigin(src_desc,
+                                                   src_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetDstSliceOrigin(dst_desc,
+                                                   dst_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_block_slice_origin)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(ThreadGroup::GetThreadId()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            threadwise_transfer_.SetSrcSliceOrigin(src_desc,
+                                                   src_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    template <typename SrcBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunRead(const SrcDesc& src_desc,
+                            const SrcBuffer& src_buf,
+                            Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.RunRead(src_desc, src_buf, thread_scratch_id);
+        }
+    }
+
+    template <typename DstBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunWrite(const DstDesc& dst_desc,
+                             DstBuffer& dst_buf,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.RunWrite(dst_desc, dst_buf, thread_scratch_id);
+        }
+    }
+
+    template <typename SrcBuffer, typename DstBuffer, index_t ThreadScratchId>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf,
+                        Number<ThreadScratchId> thread_scratch_id)
+    {
+        RunRead(src_desc, src_buf, thread_scratch_id);
+        RunWrite(dst_desc, dst_buf, thread_scratch_id);
+    }
+
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrcSliceWindow(src_desc, step);
+        }
+    }
+
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
+        }
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v3r1<decltype(thread_slice_lengths),
+                                           SrcElementwiseOperation,
+                                           DstElementwiseOperation,
+                                           DstInMemOp,
+                                           SrcData,
+                                           DstData,
+                                           SrcDesc,
+                                           DstDesc,
+                                           SrcDimAccessOrder,
+                                           DstDimAccessOrder,
+                                           SrcVectorDim,
+                                           DstVectorDim,
+                                           SrcScalarPerVector,
+                                           DstScalarPerVector,
+                                           SrcScalarStrideInVector,
+                                           DstScalarStrideInVector,
+                                           ThreadTransferSrcResetCoordinateAfterRun,
+                                           ThreadTransferDstResetCoordinateAfterRun,
+                                           NumThreadScratch>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp
new file mode 100644
index 0000000000..905a59f56e
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp
@@ -0,0 +1,134 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp"
+
+namespace ck {
+
+// this version does following things to avoid scratch memory issue
+// 1. Use StaticallyIndexedArray instead of C array for thread buffer
+// 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
+// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
+template <typename ThreadGroup,
+          typename ElementwiseOperation,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          bool ThreadTransferSrcResetCoordinateAfterRun,
+          bool ThreadTransferDstResetCoordinateAfterRun>
+struct ThreadGroupTensorSliceTransfer_v6r1
+{
+    static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
+
+    static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
+
+    using Index = MultiIndex<nDim>;
+
+    __device__ constexpr ThreadGroupTensorSliceTransfer_v6r1(const SrcDesc& src_desc,
+                                                             const Index& src_block_slice_origin,
+                                                             const DstDesc& dst_desc,
+                                                             const Index& dst_block_slice_origin,
+                                                             const ElementwiseOperation& element_op)
+        : threadwise_transfer_(src_desc,
+                               make_zero_multi_index<nDim>(),
+                               dst_desc,
+                               make_zero_multi_index<nDim>(),
+                               element_op)
+
+    {
+        static_assert(nDim == remove_cvref_t<SrcDesc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
+                          nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == DimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! ThreadGroup::GetNumOfThread() too small");
+
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(ThreadGroup::GetThreadId()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            threadwise_transfer_.SetSrcSliceOrigin(src_desc,
+                                                   src_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetDstSliceOrigin(dst_desc,
+                                                   dst_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    template <typename SrcBuffer, typename DstBuffer>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.Run(src_desc, src_buf, dst_desc, dst_buf);
+        }
+    }
+
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrcSliceWindow(src_desc, step);
+        }
+    }
+
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
+        }
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v6r1<SrcData,
+                                           DstData,
+                                           SrcDesc,
+                                           DstDesc,
+                                           ElementwiseOperation,
+                                           decltype(thread_slice_lengths),
+                                           DimAccessOrder,
+                                           VectorDim,
+                                           ScalarPerVector,
+                                           DstInMemOp,
+                                           ThreadTransferSrcResetCoordinateAfterRun,
+                                           ThreadTransferDstResetCoordinateAfterRun>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1r2.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1r2.hpp
new file mode 100644
index 0000000000..83cb9fb5de
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1r2.hpp
@@ -0,0 +1,164 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1r2.hpp"
+
+namespace ck {
+
+// this version does following things to avoid scratch memory issue
+// 1. Use StaticallyIndexedArray instead of C array for thread buffer
+// 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
+// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
+template <typename ThreadGroup,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          bool ThreadTransferSrcResetCoordinateAfterRun,
+          bool ThreadTransferDstResetCoordinateAfterRun>
+struct ThreadGroupTensorSliceTransfer_v6r1r2
+{
+    static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
+
+    static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
+
+    using Index = MultiIndex<nDim>;
+
+    __device__ constexpr ThreadGroupTensorSliceTransfer_v6r1r2(
+        const SrcDesc& src_desc,
+        const Index& src_block_slice_origin,
+        const DstDesc& dst_desc,
+        const Index& dst_block_slice_origin,
+        const ElementwiseOperation& element_op)
+        : threadwise_transfer_(src_desc,
+                               make_zero_multi_index<nDim>(),
+                               dst_desc,
+                               make_zero_multi_index<nDim>(),
+                               element_op)
+
+    {
+        static_assert(nDim == remove_cvref_t<SrcDesc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
+                          nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == DimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! ThreadGroup::GetNumOfThread() too small");
+
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(ThreadGroup::GetThreadId()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            threadwise_transfer_.SetSrcSliceOrigin(src_desc,
+                                                   src_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetDstSliceOrigin(dst_desc,
+                                                   dst_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    template <typename SrcBuffer, typename DstBuffer, InMemoryDataOperationEnum DstInMemOp>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.template Run<SrcBuffer, DstBuffer, DstInMemOp>(
+                src_desc, src_buf, dst_desc, dst_buf);
+        }
+    }
+
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrcSliceWindow(src_desc, step);
+        }
+    }
+
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
+        }
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_block_slice_origin)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(ThreadGroup::GetThreadId()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            threadwise_transfer_.SetSrcSliceOrigin(src_desc,
+                                                   src_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_block_slice_origin)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(ThreadGroup::GetThreadId()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            threadwise_transfer_.SetDstSliceOrigin(dst_desc,
+                                                   dst_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v6r1r2<SrcData,
+                                             DstData,
+                                             SrcDesc,
+                                             DstDesc,
+                                             ElementwiseOperation,
+                                             decltype(thread_slice_lengths),
+                                             DimAccessOrder,
+                                             VectorDim,
+                                             ScalarPerVector,
+                                             ThreadTransferSrcResetCoordinateAfterRun,
+                                             ThreadTransferDstResetCoordinateAfterRun>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp
new file mode 100644
index 0000000000..17110c8358
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp
@@ -0,0 +1,158 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp"
+
+namespace ck {
+
+// this version does following things to avoid scratch memory issue
+// 1. Use StaticallyIndexedArray instead of C array for thread buffer
+// 2. It does not keep reference to tensor descriptor
+// 3. Run() does not construct new tensor coordinate
+template <typename ThreadGroup,
+          typename ElementwiseOperation,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename Src0Data,
+          typename Src1Data,
+          typename DstData,
+          typename Src0Desc,
+          typename Src1Desc,
+          typename DstDesc,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          bool ThreadTransferSrc0ResetCoordinateAfterRun,
+          bool ThreadTransferSrc1ResetCoordinateAfterRun,
+          bool ThreadTransferDstResetCoordinateAfterRun>
+struct ThreadGroupTensorSliceTransfer_v6r2
+{
+    static constexpr index_t nDim = remove_reference_t<Src0Desc>::GetNumOfDimension();
+
+    static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
+
+    using Index = MultiIndex<nDim>;
+
+    __device__ constexpr ThreadGroupTensorSliceTransfer_v6r2(const Src0Desc& src0_desc,
+                                                             const Index& src0_block_slice_origin,
+                                                             const Src1Desc& src1_desc,
+                                                             const Index& src1_block_slice_origin,
+                                                             const DstDesc& dst_desc,
+                                                             const Index& dst_block_slice_origin,
+                                                             const ElementwiseOperation& element_op)
+        : threadwise_transfer_(src0_desc,
+                               make_zero_multi_index<nDim>(),
+                               src1_desc,
+                               make_zero_multi_index<nDim>(),
+                               dst_desc,
+                               make_zero_multi_index<nDim>(),
+                               element_op)
+
+    {
+        static_assert(nDim == remove_cvref_t<Src0Desc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<Src1Desc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
+                          nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == DimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! ThreadGroup::GetNumOfThread() too small");
+
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(ThreadGroup::GetThreadId()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            threadwise_transfer_.SetSrc0SliceOrigin(
+                src0_desc, src0_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetSrc1SliceOrigin(
+                src1_desc, src1_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetDstSliceOrigin(dst_desc,
+                                                   dst_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    template <typename Src0Buffer, typename Src1Buffer, typename DstBuffer>
+    __device__ void Run(const Src0Desc& src0_desc,
+                        const Src0Buffer& src0_buf,
+                        const Src1Desc& src1_desc,
+                        const Src1Buffer& src1_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.Run(src0_desc, src0_buf, src1_desc, src1_buf, dst_desc, dst_buf);
+        }
+    }
+
+    __device__ void MoveSrc0SliceWindow(const Src0Desc& src0_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrc0SliceWindow(src0_desc, step);
+        }
+    }
+
+    __device__ void MoveSrc1SliceWindow(const Src1Desc& src1_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrc1SliceWindow(src1_desc, step);
+        }
+    }
+
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
+        }
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v6r2<Src0Data,
+                                           Src1Data,
+                                           DstData,
+                                           Src0Desc,
+                                           Src1Desc,
+                                           DstDesc,
+                                           ElementwiseOperation,
+                                           decltype(thread_slice_lengths),
+                                           DimAccessOrder,
+                                           VectorDim,
+                                           ScalarPerVector,
+                                           DstInMemOp,
+                                           ThreadTransferSrc0ResetCoordinateAfterRun,
+                                           ThreadTransferSrc1ResetCoordinateAfterRun,
+                                           ThreadTransferDstResetCoordinateAfterRun>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp
new file mode 100644
index 0000000000..9a5317dd12
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp
@@ -0,0 +1,183 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp"
+
+namespace ck {
+
+// this version does following things to avoid scratch memory issue
+// 1. Use StaticallyIndexedArray instead of C array for thread buffer
+// 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
+// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
+template <typename ThreadGroup,
+          typename ElementwiseOperation,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename Src0Data,
+          typename Src1Data,
+          typename Src2Data,
+          typename DstData,
+          typename Src0Desc,
+          typename Src1Desc,
+          typename Src2Desc,
+          typename DstDesc,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          bool ThreadTransferSrc0ResetCoordinateAfterRun,
+          bool ThreadTransferSrc1ResetCoordinateAfterRun,
+          bool ThreadTransferSrc2ResetCoordinateAfterRun,
+          bool ThreadTransferDstResetCoordinateAfterRun>
+struct ThreadGroupTensorSliceTransfer_v6r3
+{
+    static constexpr index_t nDim = remove_reference_t<Src0Desc>::GetNumOfDimension();
+
+    static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
+
+    using Index = MultiIndex<nDim>;
+
+    __device__ constexpr ThreadGroupTensorSliceTransfer_v6r3(const Src0Desc& src0_desc,
+                                                             const Index& src0_block_slice_origin,
+                                                             const Src1Desc& src1_desc,
+                                                             const Index& src1_block_slice_origin,
+                                                             const Src2Desc& src2_desc,
+                                                             const Index& src2_block_slice_origin,
+                                                             const DstDesc& dst_desc,
+                                                             const Index& dst_block_slice_origin,
+                                                             const ElementwiseOperation& element_op)
+        : threadwise_transfer_(src0_desc,
+                               make_zero_multi_index<nDim>(),
+                               src1_desc,
+                               make_zero_multi_index<nDim>(),
+                               src2_desc,
+                               make_zero_multi_index<nDim>(),
+                               dst_desc,
+                               make_zero_multi_index<nDim>(),
+                               element_op)
+
+    {
+        static_assert(nDim == remove_cvref_t<Src0Desc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<Src1Desc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<Src2Desc>::GetNumOfDimension() &&
+                          nDim == remove_cvref_t<DstDesc>::GetNumOfDimension() &&
+                          nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == DimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! ThreadGroup::GetNumOfThread() too small");
+
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(get_thread_local_1d_id()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            threadwise_transfer_.SetSrc0SliceOrigin(
+                src0_desc, src0_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetSrc1SliceOrigin(
+                src1_desc, src1_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetSrc2SliceOrigin(
+                src2_desc, src2_block_slice_origin + thread_data_idx_begin);
+            threadwise_transfer_.SetDstSliceOrigin(dst_desc,
+                                                   dst_block_slice_origin + thread_data_idx_begin);
+        }
+    }
+
+    template <typename Src0Buffer, typename Src1Buffer, typename Src2Buffer, typename DstBuffer>
+    __device__ void Run(const Src0Desc& src0_desc,
+                        const Src0Buffer& src0_buf,
+                        const Src1Desc& src1_desc,
+                        const Src1Buffer& src1_buf,
+                        const Src2Desc& src2_desc,
+                        const Src2Buffer& src2_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.Run(
+                src0_desc, src0_buf, src1_desc, src1_buf, src2_desc, src2_buf, dst_desc, dst_buf);
+        }
+    }
+
+    __device__ void MoveSrc0SliceWindow(const Src0Desc& src0_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrc0SliceWindow(src0_desc, step);
+        }
+    }
+
+    __device__ void MoveSrc1SliceWindow(const Src1Desc& src1_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrc1SliceWindow(src1_desc, step);
+        }
+    }
+
+    __device__ void MoveSrc2SliceWindow(const Src2Desc& src2_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrc2SliceWindow(src2_desc, step);
+        }
+    }
+
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_desc, step);
+        }
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v6r3<Src0Data,
+                                           Src1Data,
+                                           Src2Data,
+                                           DstData,
+                                           Src0Desc,
+                                           Src1Desc,
+                                           Src2Desc,
+                                           DstDesc,
+                                           ElementwiseOperation,
+                                           decltype(thread_slice_lengths),
+                                           DimAccessOrder,
+                                           VectorDim,
+                                           ScalarPerVector,
+                                           DstInMemOp,
+                                           ThreadTransferSrc0ResetCoordinateAfterRun,
+                                           ThreadTransferSrc1ResetCoordinateAfterRun,
+                                           ThreadTransferSrc2ResetCoordinateAfterRun,
+                                           ThreadTransferDstResetCoordinateAfterRun>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp
new file mode 100644
index 0000000000..993d90e356
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp
@@ -0,0 +1,172 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp"
+
+namespace ck {
+
+// Thread-group level multi-source, multi-destination tensor slice data movement
+// Assume:
+//   1. All sources and destinations are DynamicBuffer
+//   2. Same VectorDim and ScalerPerVector for all sources and destinations
+//   3. DstInMemOps are per destination tensor
+//   4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
+//   5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
+//
+// Does following things to avoid scratch memory issue
+//   1. Pass tensor descritpors by reference (or tuple of references)
+//   2. Does not keep reference to tensor descriptor
+//   3. Does not construct new tensor coordinate when call Run()
+template <typename ThreadGroup,
+          typename SrcDatas,
+          typename DstDatas,
+          typename SrcDescs,
+          typename DstDescs,
+          typename ElementwiseOperation,
+          typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
+          typename SliceLengths,
+          typename ThreadClusterLengths,
+          typename ThreadClusterArrangeOrder,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          typename ThreadTransferSrcResetCoordinateAfterRunFlags,
+          typename ThreadTransferDstResetCoordinateAfterRunFlags>
+struct ThreadGroupTensorSliceTransfer_v7
+{
+    static constexpr index_t nDim =
+        remove_cvref_t<tuple_element_t<0, SrcDescs>>::GetNumOfDimension();
+
+    static constexpr index_t nSrc = remove_cvref_t<SrcDescs>::Size();
+    static constexpr index_t nDst = remove_cvref_t<DstDescs>::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{};
+
+    __device__ constexpr ThreadGroupTensorSliceTransfer_v7(
+        const SrcDescs& src_descs,
+        const StaticallyIndexedArray<Index, nSrc>& src_block_slice_origins,
+        const DstDescs& dst_descs,
+        const StaticallyIndexedArray<Index, nDst>& dst_block_slice_origins,
+        const ElementwiseOperation& element_op)
+        : threadwise_transfer_(src_descs,
+                               StaticallyIndexedArray<Index, nSrc>{},
+                               dst_descs,
+                               StaticallyIndexedArray<Index, nDst>{},
+                               element_op)
+    {
+        static_assert(nSrc == SrcDatas::Size() && nSrc == SrcDescs::Size() &&
+                          nSrc == ThreadTransferSrcResetCoordinateAfterRunFlags::Size() &&
+                          nDst == DstDatas::Size() && nDst == DstDescs::Size() &&
+                          nDst == ThreadTransferDstResetCoordinateAfterRunFlags::Size(),
+                      "wrong!");
+
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            static_assert(
+                nDim == remove_cvref_t<tuple_element_t<i.value, SrcDescs>>::GetNumOfDimension(),
+                "wrong!");
+        });
+
+        static_for<0, nDst, 1>{}([&](auto i) {
+            static_assert(
+                nDim == remove_cvref_t<tuple_element_t<i.value, DstDescs>>::GetNumOfDimension(),
+                "wrong!");
+        });
+
+        static_assert(nDim == ThreadClusterLengths::Size() &&
+                          nDim == ThreadClusterArrangeOrder::Size() &&
+                          nDim == DimAccessOrder::Size(),
+                      "wrong! nDim not consistent");
+
+        static_assert(
+            is_same<SliceLengths, decltype(thread_slice_lengths * ThreadClusterLengths{})>{},
+            "wrong! threads should be mapped to cover entire slicing window");
+
+        static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(),
+                      "wrong! ThreadGroup::GetNumOfThread() too small");
+
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex(
+                make_multi_index(get_thread_local_1d_id()));
+
+            const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths;
+
+            const auto src_thread_slice_origins = generate_tuple(
+                [&](auto i) { return src_block_slice_origins[i] + thread_data_idx_begin; },
+                Number<nSrc>{});
+
+            const auto dst_thread_slice_origins = generate_tuple(
+                [&](auto i) { return dst_block_slice_origins[i] + thread_data_idx_begin; },
+                Number<nDst>{});
+
+            threadwise_transfer_.SetSrcSliceOrigins(src_descs, src_thread_slice_origins);
+            threadwise_transfer_.SetDstSliceOrigins(dst_descs, dst_thread_slice_origins);
+        }
+    }
+
+    template <typename SrcBuffers, typename DstBuffers>
+    __device__ void Run(const SrcDescs& src_descs,
+                        const SrcBuffers& src_bufs,
+                        const DstDescs& dst_descs,
+                        DstBuffers dst_bufs)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.Run(src_descs, src_bufs, dst_descs, dst_bufs);
+        }
+    }
+
+    template <index_t ISrc>
+    __device__ void
+    MoveSrcSliceWindow(const SrcDescs& src_descs, Number<ISrc> iSrc, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveSrcSliceWindow(src_descs, iSrc, step);
+        }
+    }
+
+    template <index_t IDst>
+    __device__ void
+    MoveDstSliceWindow(const DstDescs& dst_descs, Number<IDst> iDst, const Index& step)
+    {
+        if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or
+           ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize())
+        {
+            threadwise_transfer_.MoveDstSliceWindow(dst_descs, iDst, step);
+        }
+    }
+
+    private:
+    static constexpr auto thread_cluster_desc_ =
+        make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
+
+    using ThreadwiseTransfer =
+        ThreadwiseTensorSliceTransfer_v7<SrcDatas,
+                                         DstDatas,
+                                         SrcDescs,
+                                         DstDescs,
+                                         ElementwiseOperation,
+                                         DstInMemOps,
+                                         decltype(thread_slice_lengths),
+                                         DimAccessOrder,
+                                         VectorDim,
+                                         ScalarPerVector,
+                                         ThreadTransferSrcResetCoordinateAfterRunFlags,
+                                         ThreadTransferDstResetCoordinateAfterRunFlags>;
+
+    ThreadwiseTransfer threadwise_transfer_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp b/include/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp
new file mode 100644
index 0000000000..cab5e21365
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp
@@ -0,0 +1,29 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum struct ConvolutionBackwardDataSpecialization
+{
+    Default,
+    Filter1x1Stride1Pad0,
+};
+
+inline std::string
+getConvBackwardDataSpecializationString(const ConvolutionBackwardDataSpecialization& s)
+{
+    switch(s)
+    {
+    case ConvolutionBackwardDataSpecialization::Default: return "Default";
+    case ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0: return "Filter1x1Stride1Pad0";
+    default: return "Unrecognized specialization!";
+    }
+}
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp b/include/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp
new file mode 100644
index 0000000000..01bb806789
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp
@@ -0,0 +1,33 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum struct ConvolutionBackwardWeightSpecialization
+{
+    Default,
+    Filter1x1Stride1Pad0,
+    Filter1x1Pad0,
+    OddC,
+};
+
+inline std::string
+getConvBackwardWeightSpecializationString(const ConvolutionBackwardWeightSpecialization& s)
+{
+    switch(s)
+    {
+    case ConvolutionBackwardWeightSpecialization::Default: return "Default";
+    case ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0:
+        return "Filter1x1Stride1Pad0";
+    case ConvolutionBackwardWeightSpecialization::Filter1x1Pad0: return "Filter1x1Pad0";
+    case ConvolutionBackwardWeightSpecialization::OddC: return "OddC";
+    default: return "Unrecognized specialization!";
+    }
+}
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp b/include/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp
new file mode 100644
index 0000000000..adfa1689c6
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <string>
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum struct ConvolutionForwardSpecialization
+{
+    Default,
+    Filter1x1Pad0,
+    Filter1x1Stride1Pad0,
+    OddC,
+};
+
+inline std::string getConvForwardSpecializationString(const ConvolutionForwardSpecialization& s)
+{
+    switch(s)
+    {
+    case ConvolutionForwardSpecialization::Default: return "Default";
+    case ConvolutionForwardSpecialization::Filter1x1Pad0: return "Filter1x1Pad0";
+    case ConvolutionForwardSpecialization::Filter1x1Stride1Pad0: return "Filter1x1Stride1Pad0";
+    case ConvolutionForwardSpecialization::OddC: return "OddC";
+    default: return "Unrecognized specialization!";
+    }
+}
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp b/include/ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp
new file mode 100644
index 0000000000..5a2b082c31
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp
@@ -0,0 +1,39 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NDimSpatial,
+          typename DOutDataType,
+          typename DInDataType,
+          typename DOutLayout,
+          typename DInLayout>
+struct DeviceAvgPoolBwd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_dout,
+                        void* p_din,
+                        std::vector<ck::index_t> dout_n_k_wos_lengths,
+                        std::vector<ck::index_t> dout_n_k_wos_strides,
+                        std::vector<ck::index_t> din_n_k_wos_length,
+                        std::vector<ck::index_t> din_n_k_wos_strides,
+                        std::vector<ck::index_t> window_k_c_xs_lengths,
+                        std::vector<ck::index_t> window_strides,
+                        std::vector<ck::index_t> window_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_base.hpp b/include/ck/tensor_operation/gpu/device/device_base.hpp
new file mode 100644
index 0000000000..1981690111
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_base.hpp
@@ -0,0 +1,73 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <string>
+#include <sstream>
+
+#include "ck/stream_config.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+struct BaseArgument
+{
+    BaseArgument()                    = default;
+    BaseArgument(const BaseArgument&) = default;
+    BaseArgument& operator=(const BaseArgument&) = default;
+
+    virtual ~BaseArgument() {}
+
+    void* p_workspace_ = nullptr;
+};
+
+struct BaseInvoker
+{
+    BaseInvoker()                   = default;
+    BaseInvoker(const BaseInvoker&) = default;
+    BaseInvoker& operator=(const BaseInvoker&) = default;
+
+    virtual float Run(const BaseArgument*, const StreamConfig& = StreamConfig{})
+    {
+        return float{0};
+    }
+
+    virtual ~BaseInvoker() {}
+};
+
+struct BaseOperator
+{
+    BaseOperator()                    = default;
+    BaseOperator(const BaseOperator&) = default;
+    BaseOperator& operator=(const BaseOperator&) = default;
+
+    virtual bool IsSupportedArgument(const BaseArgument*) { return false; }
+    virtual std::string GetTypeString() const { return ""; }
+
+    virtual std::string GetTypeIdName() const { return typeid(*this).name(); }
+
+    virtual std::string GetTypeIdHashCode() const
+    {
+        std::ostringstream oss;
+
+        oss << std::hex << typeid(*this).hash_code();
+
+        return oss.str();
+    };
+
+    virtual size_t GetWorkSpaceSize(const BaseArgument*) const { return 0; }
+
+    virtual void SetWorkSpacePointer(BaseArgument* p_arg, void* p_workspace) const
+    {
+        assert(p_arg);
+        p_arg->p_workspace_ = p_workspace;
+    }
+
+    virtual ~BaseOperator() {}
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp
new file mode 100644
index 0000000000..ee7af0117d
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp
@@ -0,0 +1,64 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceBatchedContractionMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::vector<index_t>& a_gs_ms_ns_lengths,
+                        const std::vector<index_t>& a_gs_ms_ks_strides,
+                        const std::vector<index_t>& b_gs_ns_ks_lengths,
+                        const std::vector<index_t>& b_gs_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                        const std::vector<index_t>& e_gs_ms_ns_lengths,
+                        const std::vector<index_t>& e_gs_ms_ns_strides,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm.hpp
new file mode 100644
index 0000000000..6cc2c7bb2f
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_batched_gemm.hpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceBatchedGemm : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        ck::index_t BatchStrideA,
+                        ck::index_t BatchStrideB,
+                        ck::index_t BatchStrideC,
+                        ck::index_t Batch,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+using DeviceBatchedGemmPtr = std::unique_ptr<DeviceBatchedGemm<ALayout,
+                                                               BLayout,
+                                                               CLayout,
+                                                               ADataType,
+                                                               BDataType,
+                                                               CDataType,
+                                                               AElementwiseOperation,
+                                                               BElementwiseOperation,
+                                                               CElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp
new file mode 100644
index 0000000000..acd779b2db
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp
@@ -0,0 +1,50 @@
+#pragma once
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+struct BatchedGemmEPermuteDesc
+{
+    ck::index_t G0_, G1_, M_, N_;
+    ck::index_t stride_G0_, stride_G1_, stride_M_, stride_N_;
+};
+
+template <typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceBatchedGemmEPermute : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_e,
+                        index_t M,
+                        index_t N,
+                        index_t K,
+                        index_t stride_A,
+                        index_t stride_B,
+                        index_t batch_stride_A,
+                        index_t batch_stride_B,
+                        BatchedGemmEPermuteDesc batched_gemm_e_permute_desc,
+                        index_t BatchCount,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp
new file mode 100644
index 0000000000..91b4b6b91b
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename B0Layout,
+          typename B1Layout,
+          typename CLayout,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename Acc0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceBatchedGemmGemm : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b0,
+                        const void* p_b1,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t O,
+                        ck::index_t Batch,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB0,
+                        ck::index_t StrideB1,
+                        ck::index_t StrideC,
+                        ck::index_t BatchStrideA,
+                        ck::index_t BatchStrideB0,
+                        ck::index_t BatchStrideB1,
+                        ck::index_t BatchStrideC,
+                        AElementwiseOperation a_element_op,
+                        B0ElementwiseOperation b0_element_op,
+                        Acc0ElementwiseOperation acc0_element_op,
+                        B1ElementwiseOperation b1_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp
new file mode 100644
index 0000000000..f18dc32906
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceBatchedGemmMultiD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static_assert(DsLayout::Size() == DsDataType::Size(), "wrong! inconsisiten NumDTensor");
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const std::array<const void*, NumDTensor>& p_ds,
+                        void* p_e,
+                        index_t M,
+                        index_t N,
+                        index_t K,
+                        index_t Batch,
+                        index_t StrideA,
+                        index_t StrideB,
+                        const std::array<ck::index_t, NumDTensor>& StrideDs,
+                        index_t StrideE,
+                        index_t BatchStrideA,
+                        index_t BatchStrideB,
+                        const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
+                        index_t BatchStrideE,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp
new file mode 100644
index 0000000000..8234e29486
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp
@@ -0,0 +1,72 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename A0Layout,
+          typename B0Layout,
+          typename D0sLayout,
+          typename B1Layout,
+          typename D1sLayout,
+          typename E1Layout,
+          typename A0DataType,
+          typename B0DataType,
+          typename D0sDataType,
+          typename B1DataType,
+          typename D1sDataType,
+          typename E1DataType,
+          typename A0ElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename CDE0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CDE1ElementwiseOperation>
+struct DeviceBatchedGemmMultipleDGemmMultipleD : public BaseOperator
+{
+    static constexpr index_t NumD0Tensor = D0sDataType::Size();
+    static constexpr index_t NumD1Tensor = D1sDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a0,
+                        const void* p_b0,
+                        std::array<const void*, NumD0Tensor> p_d0s,
+                        const void* p_b1,
+                        std::array<const void*, NumD1Tensor> p_d1s,
+                        void* p_e1,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t O,
+                        ck::index_t Batch,
+                        ck::index_t StrideA0,
+                        ck::index_t StrideB0,
+                        std::array<ck::index_t, NumD0Tensor> StrideD0s,
+                        ck::index_t StrideB1,
+                        std::array<ck::index_t, NumD1Tensor> StrideD1s,
+                        ck::index_t StrideE1,
+                        ck::index_t BatchStrideA0,
+                        ck::index_t BatchStrideB0,
+                        std::array<ck::index_t, NumD0Tensor> BatchStrideD0s,
+                        ck::index_t BatchStrideB1,
+                        std::array<ck::index_t, NumD1Tensor> BatchStrideD1s,
+                        ck::index_t BatchStrideE1,
+                        A0ElementwiseOperation a0_element_op,
+                        B0ElementwiseOperation b0_element_op,
+                        CDE0ElementwiseOperation cde0_element_op,
+                        B1ElementwiseOperation b1_element_op,
+                        CDE1ElementwiseOperation cde1_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp
new file mode 100644
index 0000000000..09259224e7
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename B0Layout,
+          typename B1Layout,
+          typename CLayout,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename Acc0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          bool MaskOutUpperTriangle> // TODO: enum for mask type
+struct DeviceBatchedGemmSoftmaxGemm : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b0,
+                        const void* p_b1,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t O,
+                        ck::index_t Batch,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB0,
+                        ck::index_t StrideB1,
+                        ck::index_t StrideC,
+                        ck::index_t BatchStrideA,
+                        ck::index_t BatchStrideB0,
+                        ck::index_t BatchStrideB1,
+                        ck::index_t BatchStrideC,
+                        AElementwiseOperation a_element_op,
+                        B0ElementwiseOperation b0_element_op,
+                        Acc0ElementwiseOperation acc0_element_op,
+                        B1ElementwiseOperation b1_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp
new file mode 100644
index 0000000000..be8105c967
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp
@@ -0,0 +1,70 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasDataType,
+          typename Acc1BiasDataType,
+          typename AElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename C0DEElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename C1DEElementwiseOperation,
+          MaskingSpecialization MaskingSpec>
+struct DeviceBatchedGemmSoftmaxGemmPermute : public BaseOperator
+{
+    static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
+    static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b0,
+        const void* p_b1,
+        void* p_c,
+        const std::array<void*, NumAcc0Bias> p_acc0_biases,
+        const std::array<void*, NumAcc1Bias> p_acc1_biases,
+        const std::vector<index_t>& a_gs_ms_ks_lengths,
+        const std::vector<index_t>& a_gs_ms_ks_strides,
+        const std::vector<index_t>& b_gs_ns_ks_lengths,
+        const std::vector<index_t>& b_gs_ns_ks_strides,
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+        const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+        const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+        const std::array<std::vector<index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides,
+        const std::array<std::vector<index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+        const std::array<std::vector<index_t>, NumAcc1Bias>
+            acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+        AElementwiseOperation a_element_op,
+        B0ElementwiseOperation b0_element_op,
+        C0DEElementwiseOperation c0de_element_op,
+        B1ElementwiseOperation b1_element_op,
+        C1DEElementwiseOperation c1de_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp b/include/ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp
new file mode 100644
index 0000000000..2c0da69257
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp
@@ -0,0 +1,77 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename XDataType,
+          typename DxDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+struct DeviceBatchNormBwd : public BaseOperator
+{
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> xyLengths,
+                        const std::array<index_t, Rank> xStrides,
+                        const std::array<index_t, Rank> dyStrides,
+                        const std::array<index_t, Rank> dxStrides,
+                        const std::array<int, NumBatchNormReduceDim> reduceDims,
+                        const std::array<ck::index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                        const std::array<ck::index_t, NumInvariantDim> bnScaleStrides,
+                        const std::array<ck::index_t, NumInvariantDim> bnDscaleDbiasStrides,
+                        const std::array<ck::index_t, NumInvariantDim> bnMeanVarStrides,
+                        const void* p_x,
+                        const void* p_dy,
+                        const void* p_scale,
+                        const void* p_savedMean,
+                        const void* p_savedInvVar,
+                        double epsilon,
+                        const DyElementwiseOp dy_elementwise_op,
+                        void* p_dx,
+                        void* p_dscale,
+                        void* p_dbias) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename XDataType,
+          typename DxDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+using DeviceBatchNormBwdPtr = std::unique_ptr<DeviceBatchNormBwd<XDataType,
+                                                                 DxDataType,
+                                                                 DyDataType,
+                                                                 AccDataType,
+                                                                 ScaleDataType,
+                                                                 DscaleDbiasDataType,
+                                                                 MeanVarDataType,
+                                                                 DyElementwiseOp,
+                                                                 Rank,
+                                                                 NumBatchNormReduceDim>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp b/include/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp
new file mode 100644
index 0000000000..e3962e177e
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp
@@ -0,0 +1,72 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+struct DeviceBatchNormFwd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, Rank> xyLengths,
+        const std::array<index_t, Rank> xStrides,
+        const std::array<index_t, Rank> yStrides,
+        const std::array<int, NumBatchNormReduceDim> reduceDims,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
+        const void* p_x,
+        const void* bnScale,
+        const void* bnBias,
+        double epsilon,
+        const YElementwiseOp y_elementwise_op,
+        void* p_y,
+        void* resultSaveMean,
+        void* resultSaveInvVariance,
+        double exponentialAverageFactor,
+        void* resultRunningMean,
+        void* resultRunningVariance) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+using DeviceBatchNormFwdPtr = std::unique_ptr<DeviceBatchNormFwd<XDataType,
+                                                                 YDataType,
+                                                                 AccDataType,
+                                                                 ScaleDataType,
+                                                                 BiasDataType,
+                                                                 MeanVarDataType,
+                                                                 YElementwiseOp,
+                                                                 Rank,
+                                                                 NumBatchNormReduceDim>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp b/include/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp
new file mode 100644
index 0000000000..69103b6f44
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp
@@ -0,0 +1,69 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+struct DeviceBatchNormInfer : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, Rank> xyLengths,
+        const std::array<index_t, Rank> xStrides,
+        const std::array<index_t, Rank> yStrides,
+        const std::array<int, NumBatchNormReduceDim> reduceDims,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
+        const void* p_x,
+        const void* bnScale,
+        const void* bnBias,
+        double epsilon,
+        const YElementwiseOp y_elementwise_op,
+        const void* estimatedMean,
+        const void* estimatedInvVariance,
+        void* p_y) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+using DeviceBatchNormInferPtr = std::unique_ptr<DeviceBatchNormInfer<XDataType,
+                                                                     YDataType,
+                                                                     AccDataType,
+                                                                     ScaleDataType,
+                                                                     BiasDataType,
+                                                                     MeanVarDataType,
+                                                                     YElementwiseOp,
+                                                                     Rank,
+                                                                     NumBatchNormReduceDim>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_cgemm.hpp b/include/ck/tensor_operation/gpu/device/device_cgemm.hpp
new file mode 100644
index 0000000000..8484212118
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_cgemm.hpp
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceCGemm : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a_real,
+                                                              const void* p_a_imag,
+                                                              const void* p_b_real,
+                                                              const void* p_b_imag,
+                                                              void* p_c_real,
+                                                              void* p_c_imag,
+                                                              void* p_workspace,
+                                                              ck::index_t M,
+                                                              ck::index_t N,
+                                                              ck::index_t K,
+                                                              ck::index_t StrideA,
+                                                              ck::index_t StrideB,
+                                                              ck::index_t StrideC,
+                                                              AElementwiseOperation a_element_op,
+                                                              BElementwiseOperation b_element_op,
+                                                              CElementwiseOperation c_element_op,
+                                                              ck::index_t KBatch = 1) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+    virtual std::size_t GetWorkspaceSize(index_t MRaw,
+                                         index_t NRaw,
+                                         index_t KRaw,
+                                         index_t StrideA,
+                                         index_t StrideB,
+                                         index_t StrideC)     = 0;
+};
+
+template <typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+using DeviceCGemmPtr = std::unique_ptr<
+    DeviceCGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp
new file mode 100644
index 0000000000..118ade8978
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp
@@ -0,0 +1,63 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceContractionMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::vector<index_t>& a_ms_ns_lengths,
+                        const std::vector<index_t>& a_ms_ks_strides,
+                        const std::vector<index_t>& b_ns_ks_lengths,
+                        const std::vector<index_t>& b_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_strides,
+                        const std::vector<index_t>& e_ms_ns_lengths,
+                        const std::vector<index_t>& e_ms_ns_strides,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp b/include/ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp
new file mode 100644
index 0000000000..eb1b85ec82
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceConvBwdData : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(void* p_in,
+                        const void* p_wei,
+                        const void* p_out,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_conv_fwd.hpp b/include/ck/tensor_operation/gpu/device/device_conv_fwd.hpp
new file mode 100644
index 0000000000..4dc11dbefd
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_conv_fwd.hpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceConvFwd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        const void* p_wei,
+                        void* p_out,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp b/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp
new file mode 100644
index 0000000000..7d3845666c
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <iostream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceConvFwdBiasActivation : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        const void* p_wei,
+                        void* p_out,
+                        const void* p_bias,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+using DeviceConvFwdBiasActivationPtr =
+    std::unique_ptr<DeviceConvFwdBiasActivation<InElementwiseOperation,
+                                                WeiElementwiseOperation,
+                                                OutElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp b/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp
new file mode 100644
index 0000000000..3a49ac632e
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <iostream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceConvFwdBiasActivationAdd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        const void* p_wei,
+                        void* p_out,
+                        const void* p_bias,
+                        const void* p_resi,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+using DeviceConvFwdBiasActivationAddPtr =
+    std::unique_ptr<DeviceConvFwdBiasActivationAdd<InElementwiseOperation,
+                                                   WeiElementwiseOperation,
+                                                   OutElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_elementwise.hpp b/include/ck/tensor_operation/gpu/device/device_elementwise.hpp
new file mode 100644
index 0000000000..db0e4bd83f
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_elementwise.hpp
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+#include <array>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataTypeTuple,
+          typename OutDataTypeTuple,
+          typename ElementwiseOperation,
+          index_t NumDim>
+struct DeviceElementwise : public BaseOperator
+{
+    static constexpr int NumInput  = InDataTypeTuple::Size();
+    static constexpr int NumOutput = OutDataTypeTuple::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, NumDim> lengths,
+                        const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
+                        const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
+                        const std::array<const void*, NumInput> in_dev_buffers,
+                        const std::array<void*, NumOutput> out_dev_buffers,
+                        ElementwiseOperation elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+}; // namespace device
+
+template <typename InDataTypeTuple,
+          typename OutDataTypeTuple,
+          typename ElementwiseOperation,
+          index_t NumDim>
+using DeviceElementwisePtr = std::unique_ptr<
+    DeviceElementwise<InDataTypeTuple, OutDataTypeTuple, ElementwiseOperation, NumDim>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp b/include/ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp
new file mode 100644
index 0000000000..c56a947ec9
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataTypeTuple,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType,
+          typename XElementwiseOperation,
+          typename YElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceElementwiseNormalization : public BaseOperator
+{
+    static constexpr int NumInput = InDataTypeTuple::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> lengths,
+                        const std::array<std::vector<index_t>, NumInput> inStridesArray,
+                        const std::vector<index_t> gammaStrides,
+                        const std::vector<index_t> betaStrides,
+                        const std::vector<index_t> yStrides,
+                        const std::vector<index_t> reduceDims,
+                        double epsilon,
+                        const std::array<const void*, NumInput> in_dev_buffers,
+                        const void* p_gamma,
+                        const void* p_beta,
+                        void* p_y,
+                        XElementwiseOperation x_elementwise_op,
+                        YElementwiseOperation y_elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename InDataTypeTuple,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType,
+          typename XElementwiseOperation,
+          typename YElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim>
+using DeviceElementwiseNormalizationPtr =
+    std::unique_ptr<DeviceElementwiseNormalization<InDataTypeTuple,
+                                                   GammaDataType,
+                                                   BetaDataType,
+                                                   AccDataType,
+                                                   YDataType,
+                                                   XElementwiseOperation,
+                                                   YElementwiseOperation,
+                                                   Rank,
+                                                   NumReduceDim>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_gemm.hpp
new file mode 100644
index 0000000000..adf909821d
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_gemm.hpp
@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGemm : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp
new file mode 100644
index 0000000000..a7f42c3b35
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+struct DEGridDesc_M0_M1_M2_N0_N1
+{
+    ck::index_t M0_, M1_, M2_, N0_, N1_;
+    ck::index_t stride_M0_, stride_M1_, stride_M2_, stride_N0_, stride_N1_;
+};
+
+// input : A[M, K], B[K, N],
+// input : D[M, N], ...
+// output : E[M, N]
+// C = a_op(A) * b_op(B)
+// E = cde_op(C, D)
+template <typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGemmBiasCPermute : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const void* p_d,
+                        void* p_e,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        DEGridDesc_M0_M1_M2_N0_N1 d_gride_desc,
+                        DEGridDesc_M0_M1_M2_N0_N1 e_gride_desc,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
new file mode 100644
index 0000000000..a44356dc24
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// GEMM:
+//   input : A[M, K], B[K, N],
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGemmMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        std::array<ck::index_t, NumDTensor> StrideDs,
+                        ck::index_t StrideE,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp
new file mode 100644
index 0000000000..0258858fe5
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   output : H[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+//   H = layernorm(E)
+// Assume:
+//   D0, D1, ... and E have the same layout
+//   Calculate mean & variance along N dimension in layernorm(E)
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename HLayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename HDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename HElementwiseOperation>
+struct DeviceGemmMultipleDLayernorm : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        const void* p_gamma,
+                        const void* p_beta,
+                        void* p_h,
+                        index_t MRaw,
+                        index_t NRaw,
+                        index_t KRaw,
+                        index_t StrideA,
+                        index_t StrideB,
+                        std::array<index_t, NumDTensor> StrideDs,
+                        index_t StrideH,
+                        double epsilon,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op,
+                        HElementwiseOperation h_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+}; // namespace device
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp
new file mode 100644
index 0000000000..539e83f7cb
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp
@@ -0,0 +1,97 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// FIXME: DeviceGemmReduce type need to well define the problem
+// GEMM:
+//   input : A[AK0, M, AK1]
+//   input : B[AK0, N, AK1]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   output : R0[M], R1[M], ...
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+//   Q0 = reduce0(q_op0(E)), Q1 = reduce1(q_op0(E)), ...
+//   R0 = r_op0(Q0), R1 = r_op1(Q1), ...
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename RsDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation>
+struct DeviceGemmMultipleDMultipleR : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+    static constexpr index_t NumRTensor = RsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        std::array<void*, NumRTensor> p_rs,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        std::array<ck::index_t, NumDTensor> StrideDs,
+                        ck::index_t StrideE,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op,
+                        QsElementwiseOperation qs_element_op,
+                        RsElementwiseOperation rs_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename RsDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation>
+using DeviceGemmMultipleDMultipleRPtr =
+    std::unique_ptr<DeviceGemmMultipleDMultipleR<ALayout,
+                                                 BLayout,
+                                                 DELayout,
+                                                 ADataType,
+                                                 BDataType,
+                                                 DsDataType,
+                                                 EDataType,
+                                                 RsDataType,
+                                                 AElementwiseOperation,
+                                                 BElementwiseOperation,
+                                                 CDEElementwiseOperation,
+                                                 QsElementwiseOperation,
+                                                 RsElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp
new file mode 100644
index 0000000000..eaa7671c64
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp
@@ -0,0 +1,46 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// FIXME: DeviceGemmReduce type need to well define the problem
+template <ck::index_t NumDTensor, ck::index_t NumReduce>
+struct DeviceGemmReduce : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const void* p_bias,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_c,
+                        std::array<void*, NumReduce> p_reduces,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        std::array<ck::index_t, NumDTensor> StrideDs,
+                        std::array<void*, 3> gemm_element_ops,
+                        std::array<void*, NumDTensor> d_element_ops,
+                        std::array<void*, NumReduce> reduce_in_element_ops,
+                        std::array<void*, NumReduce> reduce_out_element_ops,
+                        ck::index_t BatchCount = 1) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <ck::index_t NumDTensor, ck::index_t NumReduce>
+using DeviceGemmReducePtr = std::unique_ptr<DeviceGemmReduce<NumDTensor, NumReduce>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp
new file mode 100644
index 0000000000..6407aa7e09
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp
@@ -0,0 +1,64 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGemmSplitK : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                              const void* p_b,
+                                                              void* p_c,
+                                                              ck::index_t M,
+                                                              ck::index_t N,
+                                                              ck::index_t K,
+                                                              ck::index_t StrideA,
+                                                              ck::index_t StrideB,
+                                                              ck::index_t StrideC,
+                                                              AElementwiseOperation a_element_op,
+                                                              BElementwiseOperation b_element_op,
+                                                              CElementwiseOperation c_element_op,
+                                                              ck::index_t KBatch) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+using DeviceGemmSplitKPtr = std::unique_ptr<DeviceGemmSplitK<ALayout,
+                                                             BLayout,
+                                                             CLayout,
+                                                             ADataType,
+                                                             BDataType,
+                                                             CDataType,
+                                                             AElementwiseOperation,
+                                                             BElementwiseOperation,
+                                                             CElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_streamk.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_streamk.hpp
new file mode 100644
index 0000000000..ed081ad7fc
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_streamk.hpp
@@ -0,0 +1,64 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGemmStreamK : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                              const void* p_b,
+                                                              void* p_c,
+                                                              ck::index_t M,
+                                                              ck::index_t N,
+                                                              ck::index_t K,
+                                                              ck::index_t StrideA,
+                                                              ck::index_t StrideB,
+                                                              ck::index_t StrideC,
+                                                              AElementwiseOperation a_element_op,
+                                                              BElementwiseOperation b_element_op,
+                                                              CElementwiseOperation c_element_op,
+                                                              ck::index_t NumSKBlocks = 0) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+using DeviceGemmStreamKPtr = std::unique_ptr<DeviceGemmStreamK<ALayout,
+                                                               BLayout,
+                                                               CLayout,
+                                                               ADataType,
+                                                               BDataType,
+                                                               CDataType,
+                                                               AElementwiseOperation,
+                                                               BElementwiseOperation,
+                                                               CElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp
new file mode 100644
index 0000000000..ba81948440
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp
@@ -0,0 +1,72 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumDTensor>
+struct ContractionDesc
+{
+    std::vector<index_t> a_ms_ks_lengths;
+    std::vector<index_t> a_ms_ks_strides;
+
+    std::vector<index_t> b_ns_ks_lengths;
+    std::vector<index_t> b_ns_ks_strides;
+
+    std::array<std::vector<index_t>, NumDTensor> ds_ms_ns_lengths;
+    std::array<std::vector<index_t>, NumDTensor> ds_ms_ns_strides;
+
+    std::vector<index_t> e_ms_ns_lengths;
+    std::vector<index_t> e_ms_ns_strides;
+};
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGroupedContractionMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*> p_a_vec,
+                        std::vector<const void*> p_b_vec,
+                        std::vector<std::array<const void*, NumDTensor>> p_ds_vec,
+                        std::vector<void*> p_e_vec,
+                        std::vector<ContractionDesc<NumDTensor>> contraction_descs,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp
new file mode 100644
index 0000000000..7e4bca2bd6
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Conv backward data multiple D:
+//   input : output image A[G, N, K, Ho, Wo]
+//   input : weight B[G, K, C, Y, X],
+//   input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
+//   output : input image E[G, N, C, Hi, Wi],
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGroupedConvBwdDataMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static_assert(NumDTensor == DsLayout::Size(), "wrong! Inconsistent NumDTensor");
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,                                                 // output image
+        const void* p_b,                                                 // weight
+        const std::array<const void*, NumDTensor>& p_ds,                 // bias
+        void* p_e,                                                       // input image
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_lengths, // output image
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_strides, // output image
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,  // weight
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,  // weight
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+            ds_g_n_k_wos_lengths, // bias
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+            ds_g_n_k_wos_strides,                                        // bias
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_lengths, // input image
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_strides, // input image
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp
new file mode 100644
index 0000000000..ab9e6adb41
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp
@@ -0,0 +1,50 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceGroupedConvBwdWeight : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        void* p_wei,
+                        const void* p_out,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+                        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+                        const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op,
+                        ck::index_t split_k) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp
new file mode 100644
index 0000000000..025c43e75c
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp
@@ -0,0 +1,55 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Convolution Forward:
+//   input : input image A[G, N, C, Hi, Wi],
+//   input : weight B[G, K, C, Y, X],
+//   output : output image E[G, N, K, Ho, Wo]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+template <index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceGroupedConvFwd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,  // input image
+                        const void* p_wei, // weight
+                        void* p_out,       // output image
+                        const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& wei_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& wei_g_k_c_xs_strides,
+                        const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads,
+                        const InElementwiseOperation& in_element_op,
+                        const WeiElementwiseOperation& wei_element_op,
+                        const OutElementwiseOperation& out_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp
new file mode 100644
index 0000000000..1cc30fd9e6
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Convolution Forward:
+//   input : input image A[G, N, C, Hi, Wi],
+//   input : weight B[G, K, C, Y, X],
+//   input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
+//   output : output image E[G, N, K, Ho, Wo]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGroupedConvFwdMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static_assert(NumDTensor == DsLayout::Size(), "wrong! Inconsistent NumDTensor");
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a, // input image
+        const void* p_b, // weight
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e, // output image
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp
new file mode 100644
index 0000000000..1e03405536
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp
@@ -0,0 +1,55 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+struct GemmDesc
+{
+    ck::index_t M_, N_, K_;
+    ck::index_t stride_A_, stride_B_, stride_C_;
+
+    std::vector<ck::index_t> stride_Ds_;
+};
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGroupedGemm : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static_assert(DsLayout::Size() == DsDataType::Size(), "wrong! inconsistent NumDTensor");
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*>& p_a,
+                        std::vector<const void*>& p_b,
+                        std::vector<std::array<const void*, NumDTensor>>& p_ds,
+                        std::vector<void*>& p_e,
+                        std::vector<GemmDesc>& gemm_desc,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp
new file mode 100644
index 0000000000..fcb2ba6a4d
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp
@@ -0,0 +1,63 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <array>
+
+#include "device_grouped_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumDTensor = 0>
+struct GroupedGemmKernelArgument
+{
+    const void* p_a_grid;
+    const void* p_b_grid;
+    std::array<const void*, NumDTensor> p_ds_grid;
+    void* p_e_grid;
+
+    index_t M;
+    index_t N;
+    index_t K;
+
+    index_t StrideA;
+    index_t StrideB;
+    std::array<index_t, NumDTensor> StrideDs;
+    index_t StrideE;
+};
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGroupedGemmFixedNK : DeviceGroupedGemm<ALayout,
+                                                    BLayout,
+                                                    DsLayout,
+                                                    ELayout,
+                                                    ADataType,
+                                                    BDataType,
+                                                    DsDataType,
+                                                    EDataType,
+                                                    AElementwiseOperation,
+                                                    BElementwiseOperation,
+                                                    CElementwiseOperation>
+{
+    virtual void SetDeviceKernelArgs(BaseArgument* p_arg, const void* kernel_args) const = 0;
+    virtual size_t GetDeviceKernelArgSize(const BaseArgument* p_arg) const               = 0;
+    virtual void SetKBatch(BaseArgument* p_arg, index_t k_batch) const                   = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp
new file mode 100644
index 0000000000..fae6509740
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp
@@ -0,0 +1,75 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasDataType,
+          typename Acc1BiasDataType,
+          typename AElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename Acc0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          MaskingSpecialization MaskingSpec>
+struct DeviceGroupedGemmSoftmaxGemmPermute : public BaseOperator
+{
+    struct ProblemDesc
+    {
+        std::vector<index_t> a_gs_ms_ks_lengths;
+        std::vector<index_t> a_gs_ms_ks_strides;
+
+        std::vector<index_t> b0_gs_ns_ks_lengths;
+        std::vector<index_t> b0_gs_ns_ks_strides;
+
+        std::vector<index_t> b1_gs_os_ns_lengths;
+        std::vector<index_t> b1_gs_os_ns_strides;
+
+        std::vector<index_t> c_gs_ms_os_lengths;
+        std::vector<index_t> c_gs_ms_os_strides;
+
+        std::vector<std::vector<index_t>> acc0_biases_gs_ms_ns_lengths;
+        std::vector<std::vector<index_t>> acc0_biases_gs_ms_ns_strides;
+
+        std::vector<std::vector<index_t>> acc1_biases_gs_ms_os_lengths;
+        std::vector<std::vector<index_t>> acc1_biases_gs_ms_os_strides;
+    };
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*> p_a_vec,
+                        std::vector<const void*> p_b0_vec,
+                        std::vector<const void*> p_b1_vec,
+                        std::vector<void*> p_c_vec,
+                        std::vector<std::vector<const void*>> p_acc0_biases_vec,
+                        std::vector<std::vector<const void*>> p_acc1_biases_vec,
+                        std::vector<ProblemDesc> problem_desc_vec,
+                        AElementwiseOperation a_element_op,
+                        B0ElementwiseOperation b0_element_op,
+                        Acc0ElementwiseOperation acc0_element_op,
+                        B1ElementwiseOperation b1_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp
new file mode 100644
index 0000000000..06d180d30f
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp
@@ -0,0 +1,39 @@
+#pragma once
+#include <iostream>
+#include <vector>
+
+#include "device_grouped_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGroupedGemmSplitK : public DeviceGroupedGemm<ALayout,
+                                                          BLayout,
+                                                          DsLayout,
+                                                          ELayout,
+                                                          ADataType,
+                                                          BDataType,
+                                                          DsDataType,
+                                                          EDataType,
+                                                          AElementwiseOperation,
+                                                          BElementwiseOperation,
+                                                          CElementwiseOperation>
+{
+    virtual void SetKBatchSize(BaseArgument* p_arg, index_t kbatch) const = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_image_to_column.hpp b/include/ck/tensor_operation/gpu/device/device_image_to_column.hpp
new file mode 100644
index 0000000000..631d5189dd
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_image_to_column.hpp
@@ -0,0 +1,70 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+/**
+ * \brief Image to column.
+ *
+ * This Device operator converts image ([G, N, Di, Hi, Wi, C]) to the gemm
+ * problem([N * Do * Ho * Wo, Z *  Y * X * C]). G must be equal to 1.
+ *
+ * \tparam NDimSpatial Number of spatial dimensions.
+ * \tparam InputLayout Input Layout.
+ * \tparam InputDataType Input Data Type.
+ * \tparam OutputDataType Output Data Type.
+ */
+template <index_t NDimSpatial,
+          typename InputLayout,
+          typename InputDataType,
+          typename OutputDataType>
+struct DeviceImageToColumn : public BaseOperator
+{
+
+    /**
+     * \brief Make argument pointer for image to column.
+     *
+     * \param p_in A pointer to the device memory of the input image.
+     * \param p_out A pointer to the device memory of the output.
+     * \param N Convolution batch size.
+     * \param C Convolution number of channels.
+     * \param input_spatial_lengths Input spatial lengths.
+     * \param filter_spatial_lengths Filter spatial lengths.
+     * \param output_spatial_lengths Output spatial lengths.
+     * \param input_g_n_c_wis_strides Input strides in order [G, N, C, D, H, W].
+     * \param output_m_k_strides Output strides.
+     * \param conv_filter_strides Convolution filter strides.
+     * \param conv_filter_dilations Convolution filter dilations.
+     * \param input_left_pads Convolution left pads.
+     * \param input_right_pads Convolution right pads.
+     * \return Pointer to the argument.
+     */
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        void* p_out,
+                        const ck::index_t N,
+                        const ck::index_t C,
+                        const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+                        const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                        const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                        const std::array<index_t, 2>& output_m_k_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp b/include/ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp
new file mode 100644
index 0000000000..5a4a9cac1e
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp
@@ -0,0 +1,33 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// For pooling which used indexable operation, such as MaxPool, MinPool...etc
+template <typename DOutDataType, typename IndexDataType, typename DInDataType>
+struct DeviceMaxPoolBwd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_dout,
+                        const void* p_indices,
+                        void* p_din,
+                        index_t dout_length,
+                        index_t din_length,
+                        std::vector<ck::index_t> window_lengths,
+                        std::vector<ck::index_t> window_strides,
+                        std::vector<ck::index_t> window_dilations) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp b/include/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp
new file mode 100644
index 0000000000..f68022ca04
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include <array>
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t Rank,
+          index_t NumReduceDim,
+          index_t NumReduction,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple>
+struct DeviceMultipleReduce : public BaseOperator
+{
+    static constexpr index_t NumInputDim  = Rank;
+    static constexpr index_t NumOutputDim = (Rank - NumReduceDim > 1) ? Rank - NumReduceDim : 1;
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, NumInputDim> inLengths,
+        const std::array<index_t, NumInputDim> inStrides,
+        const std::array<index_t, NumOutputDim> outLengths,
+        const std::array<std::array<index_t, NumOutputDim>, NumReduction> outStrides,
+        const std::array<int, NumReduceDim> reduceDims,
+        const std::array<double, NumReduction> alphas,
+        const std::array<double, NumReduction> betas,
+        const void* in_dev,
+        const std::array<void*, NumReduction> out_dev_buffers,
+        const InElementwiseOperationTuple in_elementwise_op_tuple,
+        const AccElementwiseOperationTuple acc_elementwise_op_tuple) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <index_t Rank,
+          index_t NumReduceDim,
+          index_t NumReduction,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple>
+using DeviceMultipleReducePtr = std::unique_ptr<DeviceMultipleReduce<Rank,
+                                                                     NumReduceDim,
+                                                                     NumReduction,
+                                                                     InElementwiseOperationTuple,
+                                                                     AccElementwiseOperationTuple>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_normalization.hpp b/include/ck/tensor_operation/gpu/device/device_normalization.hpp
new file mode 100644
index 0000000000..1f178f9fcb
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_normalization.hpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename ComputeDataType,
+          typename YDataType,
+          typename YElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceNormalization : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> lengths,
+                        const std::vector<index_t> xStrides,
+                        const std::vector<index_t> gammaStrides,
+                        const std::vector<index_t> betaStrides,
+                        const std::vector<index_t> yStrides,
+                        const std::vector<index_t> reduceDims,
+                        double epsilon,
+                        const void* p_x,
+                        const void* p_gamma,
+                        const void* p_beta,
+                        void* p_y,
+                        void* p_savedMean,
+                        void* p_savedInvVar,
+                        YElementwiseOperation y_elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename ComputeDataType,
+          typename YDataType,
+          typename YElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim>
+using DeviceNormalizationPtr = std::unique_ptr<DeviceNormalization<XDataType,
+                                                                   GammaDataType,
+                                                                   BetaDataType,
+                                                                   ComputeDataType,
+                                                                   YDataType,
+                                                                   YElementwiseOperation,
+                                                                   Rank,
+                                                                   NumReduceDim>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_permute.hpp b/include/ck/tensor_operation/gpu/device/device_permute.hpp
new file mode 100644
index 0000000000..c994cf02c6
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_permute.hpp
@@ -0,0 +1,36 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include <memory>
+#include <type_traits>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumDim, typename InDataType, typename OutDataType, typename ElementwiseOperation>
+struct DevicePermute : BaseOperator
+{
+    using Lengths = std::array<index_t, NumDim>;
+    using Strides = Lengths;
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const Lengths& in_lengths,
+                        const Strides& in_strides,
+                        const Lengths& out_lengths,
+                        const Strides& out_strides,
+                        const void* in_dev_buffer,
+                        void* out_dev_buffer,
+                        ElementwiseOperation elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_pool_fwd.hpp b/include/ck/tensor_operation/gpu/device/device_pool_fwd.hpp
new file mode 100644
index 0000000000..62071c43a1
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_pool_fwd.hpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t InOutRank,
+          index_t WindowRank,
+          typename InDataType,
+          typename OutDataType,
+          typename IndexDataType,
+          typename InLayout,
+          typename OutLayout,
+          ReduceTensorOp ReduceOpId,
+          bool OutputIndex>
+struct DevicePoolFwd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_dev,
+                        void* p_out_dev,
+                        void* p_out_indices_dev,
+                        std::vector<ck::index_t> input_n_c_wis_lengths,
+                        std::vector<ck::index_t> window_xs_lengths,
+                        std::vector<ck::index_t> output_n_c_wos_lengths,
+                        std::vector<ck::index_t> input_n_c_wis_stride,
+                        std::vector<ck::index_t> output_n_c_wis_stride,
+                        std::vector<ck::index_t> indices_n_c_wis_stride,
+                        std::vector<ck::index_t> window_xs_strides,
+                        std::vector<ck::index_t> window_xs_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        std::vector<ck::index_t> pooling_dims) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_put_element.hpp b/include/ck/tensor_operation/gpu/device/device_put_element.hpp
new file mode 100644
index 0000000000..17df2de37b
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_put_element.hpp
@@ -0,0 +1,36 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// output[indices] = input
+template <typename InDataType,
+          typename IndexDataType,
+          typename OutDataType,
+          typename ElementwiseOperation,
+          InMemoryDataOperationEnum Op>
+struct DevicePutElement : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_input,
+                        const void* p_indices,
+                        void* p_output,
+                        index_t input_length,
+                        index_t output_length,
+                        ElementwiseOperation elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_reduce.hpp b/include/ck/tensor_operation/gpu/device/device_reduce.hpp
new file mode 100644
index 0000000000..c2721b1845
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_reduce.hpp
@@ -0,0 +1,71 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation,
+          bool PropagateNan,
+          bool OutputIndex>
+struct DeviceReduce : public BaseOperator
+{
+    static constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> inLengths,
+                        const std::array<index_t, Rank> inStrides,
+                        const std::array<index_t, NumOutDim> outLengths,
+                        const std::array<index_t, NumOutDim> outStrides,
+                        const std::array<int, NumReduceDim> reduceDims,
+                        double alpha,
+                        double beta,
+                        const void* in_dev,
+                        const void* in_index_dev,
+                        void* out_dev,
+                        void* out_index_dev,
+                        const InElementwiseOperation in_elementwise_op,
+                        const AccElementwiseOperation acc_elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation,
+          bool PropagateNan,
+          bool OutputIndex>
+using DeviceReducePtr = std::unique_ptr<DeviceReduce<InDataType,
+                                                     AccDataType,
+                                                     OutDataType,
+                                                     Rank,
+                                                     NumReduceDim,
+                                                     ReduceOperation,
+                                                     InElementwiseOperation,
+                                                     AccElementwiseOperation,
+                                                     PropagateNan,
+                                                     OutputIndex>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_softmax.hpp b/include/ck/tensor_operation/gpu/device/device_softmax.hpp
new file mode 100644
index 0000000000..1902fd09ee
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_softmax.hpp
@@ -0,0 +1,72 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          typename InElementwiseOp,
+          typename AccElementwiseOp,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceSoftmax : public BaseOperator
+{
+    //
+    // @brief      Makes a pointer to Argument class.
+    //
+    // @param[in]  inLengths           Input tensor extent(s) from high to low dimension
+    // @param[in]  inStrides           Input tensor stride(s) from high to low dimension
+    // @param[in]  reduceDims          The dimension(s) the normalization operation is applied
+    // @param[in]  alpha               double type value
+    // @param[in]  beta                double type value
+    // @param[in]  in_dev              Typeless const pointer in device memory storing the input
+    //                                 tensor
+    // @param      out_dev             Typeless pointer in device memory storing the output tensor
+    // @param[in]  in_elementwise_op   The input elementwise operation.
+    // @param[in]  acc_elementwise_op  The accumulation elementwise operation.
+    //
+    // @return     Unique pointer to the Argument class.
+    //
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> inLengths,
+                        const std::vector<index_t> inStrides,
+                        const std::vector<int> reduceDims,
+                        double alpha,
+                        double beta,
+                        const void* in_dev,
+                        void* out_dev,
+                        InElementwiseOp in_elementwise_op,
+                        AccElementwiseOp acc_elementwise_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          typename InElementwiseOp,
+          typename AccElementwiseOp,
+          index_t Rank,
+          index_t NumReduceDim>
+using DeviceSoftmaxPtr = std::unique_ptr<DeviceSoftmax<InDataType,
+                                                       AccDataType,
+                                                       OutDataType,
+                                                       InElementwiseOp,
+                                                       AccElementwiseOp,
+                                                       Rank,
+                                                       NumReduceDim>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp
new file mode 100644
index 0000000000..eeccd977cc
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceSplitKContractionMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::vector<index_t>& a_gs_ms_ns_lengths,
+                        const std::vector<index_t>& a_gs_ms_ks_strides,
+                        const std::vector<index_t>& b_gs_ns_ks_lengths,
+                        const std::vector<index_t>& b_gs_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                        const std::vector<index_t>& e_gs_ms_ns_lengths,
+                        const std::vector<index_t>& e_gs_ms_ns_strides,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op,
+                        index_t split_k) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/gemm_dl_algorithm.hpp b/include/ck/tensor_operation/gpu/device/gemm_dl_algorithm.hpp
new file mode 100644
index 0000000000..16ca582b89
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/gemm_dl_algorithm.hpp
@@ -0,0 +1,18 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum struct GemmDlAlgorithm
+{
+    Default, // Uses DOT vector instructions
+    Dpp8,    // Uses DOT vector instructions with DPP8 SEL modifier to reduce data loads from LDS
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/gemm_specialization.hpp b/include/ck/tensor_operation/gpu/device/gemm_specialization.hpp
new file mode 100644
index 0000000000..0bb45b18c3
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/gemm_specialization.hpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum struct GemmSpecialization
+{
+    // Gemm
+    Default,
+    MPadding,
+    NPadding,
+    KPadding,
+    MNPadding,
+    MKPadding,
+    NKPadding,
+    MNKPadding,
+    // Gemm + Gemm
+    OPadding,
+    MOPadding,
+    NOPadding,
+    KOPadding,
+    MNOPadding,
+    MKOPadding,
+    NKOPadding,
+    MNKOPadding,
+};
+
+inline std::string getGemmSpecializationString(const GemmSpecialization& s)
+{
+    switch(s)
+    {
+    case GemmSpecialization::Default: return "Default";
+    case GemmSpecialization::MPadding: return "MPadding";
+    case GemmSpecialization::NPadding: return "NPadding";
+    case GemmSpecialization::KPadding: return "KPadding";
+    case GemmSpecialization::MNPadding: return "MNPadding";
+    case GemmSpecialization::MKPadding: return "MKPadding";
+    case GemmSpecialization::NKPadding: return "NKPadding";
+    case GemmSpecialization::MNKPadding: return "MNKPadding";
+    case GemmSpecialization::OPadding: return "OPadding";
+    case GemmSpecialization::MOPadding: return "MOPadding";
+    case GemmSpecialization::NOPadding: return "NOPadding";
+    case GemmSpecialization::KOPadding: return "KOPadding";
+    case GemmSpecialization::MNOPadding: return "MNOPadding";
+    case GemmSpecialization::MKOPadding: return "MKOPadding";
+    case GemmSpecialization::NKOPadding: return "NKOPadding";
+    case GemmSpecialization::MNKOPadding: return "MNKOPadding";
+    default: return "Unrecognized specialization!";
+    }
+}
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp b/include/ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp
new file mode 100644
index 0000000000..3a2280a75c
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp
@@ -0,0 +1,575 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// In and Din = [N, C, Di, Hi, Wi]
+// Out and Dout = [N, C, Do, Ho, Wo]
+// Out = AvgPoolFwd(In)
+// Din = AvgPoolBwd(Dout)
+// Pooling dimension = D, H, W
+template <typename DOutDataType,
+          typename DInDataType,
+          typename ComputeDataType,
+          ck::index_t BlockSize,
+          ck::index_t MThreadClusterSize,
+          ck::index_t KThreadClusterSize,
+          ck::index_t MThreadSliceSize,
+          ck::index_t KThreadSliceSize,
+          ck::index_t InSrcOutDstVectorSize>
+struct DeviceAvgPool3dBwd_NDHWC_NDHWC : public DeviceAvgPoolBwd<3,
+                                                                DOutDataType,
+                                                                DInDataType,
+                                                                tensor_layout::convolution::NDHWC,
+                                                                tensor_layout::convolution::NDHWC>
+{
+    static constexpr ck::index_t NDimSpatial = 3;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr ck::index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr ck::index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto
+    Make3DGridDescriptor_Out_M_K_In_M(const std::vector<ck::index_t>& dout_n_c_wos_lengths,
+                                      const std::vector<ck::index_t>& din_n_c_wos_length,
+                                      const std::vector<ck::index_t>& dout_n_c_wos_strides,
+                                      const std::vector<ck::index_t>& din_n_c_wos_strides,
+                                      const std::vector<ck::index_t>& window_lengths,
+                                      const std::vector<ck::index_t>& window_strides,
+                                      const std::vector<ck::index_t>& window_dilations,
+                                      const std::vector<ck::index_t>& input_left_pads,
+                                      const std::vector<ck::index_t>& input_right_pads,
+                                      const std::vector<ck::index_t>& tildes)
+    {
+        index_t i_ztilde = tildes[0];
+        index_t i_ytilde = tildes[1];
+        index_t i_xtilde = tildes[2];
+
+        const index_t N = dout_n_c_wos_lengths[0];
+        const index_t C = dout_n_c_wos_lengths[1];
+
+        const index_t Di = din_n_c_wos_length[2];
+        const index_t Hi = din_n_c_wos_length[3];
+        const index_t Wi = din_n_c_wos_length[4];
+
+        const index_t Do = dout_n_c_wos_lengths[2];
+        const index_t Ho = dout_n_c_wos_lengths[3];
+        const index_t Wo = dout_n_c_wos_lengths[4];
+
+        const index_t Z = window_lengths[0];
+        const index_t Y = window_lengths[1];
+        const index_t X = window_lengths[2];
+
+        const index_t InLeftPadD = input_left_pads[0];
+        const index_t InLeftPadH = input_left_pads[1];
+        const index_t InLeftPadW = input_left_pads[2];
+
+        const index_t InRightPadD = input_right_pads[0];
+        const index_t InRightPadH = input_right_pads[1];
+        const index_t InRightPadW = input_right_pads[2];
+
+        const index_t ConvStrideD = window_strides[0];
+        const index_t ConvStrideH = window_strides[1];
+        const index_t ConvStrideW = window_strides[2];
+
+        const index_t ConvDilationD = window_dilations[0];
+        const index_t ConvDilationH = window_dilations[1];
+        const index_t ConvDilationW = window_dilations[2];
+
+        const auto out_n_do_ho_wo_c_grid_desc =
+            make_naive_tensor_descriptor(make_tuple(N, Do, Ho, Wo, C),
+                                         make_tuple(dout_n_c_wos_strides[0],
+                                                    dout_n_c_wos_strides[2],
+                                                    dout_n_c_wos_strides[3],
+                                                    dout_n_c_wos_strides[4],
+                                                    dout_n_c_wos_strides[1]));
+
+        const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
+        const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+        const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+        const auto ZTilde = ConvStrideD / GcdStrideDilationD;
+        const auto YTilde = ConvStrideH / GcdStrideDilationH;
+        const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+        const auto ZDot = math::integer_divide_ceil(Z, ZTilde);
+        const auto YDot = math::integer_divide_ceil(Y, YTilde);
+        const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+        const auto DTilde = Do + math::integer_divide_ceil(ConvDilationD * (Z - I1), ConvStrideD);
+        const auto HTilde = Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+        const auto WTilde = Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+        // only work on Tildes that contribute to non-padding area of input tensor
+        const auto IDTildeSliceBegin = math::integer_divide_floor(
+            math::max(I0, InLeftPadD - ConvDilationD * (ZTilde - I1)), ConvStrideD);
+        const auto IHTildeSliceBegin = math::integer_divide_floor(
+            math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+        const auto IWTildeSliceBegin = math::integer_divide_floor(
+            math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+        const auto IDTildeSliceEnd =
+            math::min(DTilde, math::integer_divide_ceil(InLeftPadD + Di - I1, ConvStrideD) + I1);
+        const auto IHTildeSliceEnd =
+            math::min(HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+        const auto IWTildeSliceEnd =
+            math::min(WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+        const auto DTildeSlice = IDTildeSliceEnd - IDTildeSliceBegin;
+        const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+        const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+        // ReduceK is different for each Reduce
+        const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
+        const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+        const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+        // Problem size of reduction kernel
+        const index_t MRaw = N * DTildeSlice * HTildeSlice * WTildeSlice * C;
+        const index_t MPad = math::integer_least_multiple(MRaw, M_BlockTileSize) - MRaw;
+
+        const index_t KRaw = ZDotSlice * YDotSlice * XDotSlice;
+        const index_t KPad = math::integer_least_multiple(KRaw, K_BlockTileSize) - KRaw;
+
+        // Out[ReduceM, ReduceK]
+        const auto out_n_dop_hop_wop_c_grid_desc = transform_tensor_descriptor(
+            out_n_do_ho_wo_c_grid_desc,
+            make_tuple(make_pass_through_transform(N),
+                       make_pad_transform(Do, I0, I0),
+                       make_pad_transform(Ho, I0, I0),
+                       make_pad_transform(Wo, I0, I0),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+        const auto out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_c_grid_desc =
+            transform_tensor_descriptor(
+                out_n_dop_hop_wop_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(ZDot, DTilde),
+                                         make_tuple(-ConvDilationD / GcdStrideDilationD, I1)),
+                    make_embed_transform(make_tuple(YDot, HTilde),
+                                         make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                    make_embed_transform(make_tuple(XDot, WTilde),
+                                         make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                    make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+        const auto
+            out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_c_grid_desc =
+                transform_tensor_descriptor(
+                    out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_c_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_slice_transform(ZDot, I0, ZDotSlice),
+                               make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                               make_slice_transform(YDot, I0, YDotSlice),
+                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                               make_slice_transform(XDot, I0, XDotSlice),
+                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{},
+                               Sequence<6>{},
+                               Sequence<7>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{},
+                               Sequence<6>{},
+                               Sequence<7>{}));
+
+        const auto out_grid_desc_reducemraw_reducekraw = transform_tensor_descriptor(
+            out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_c_grid_desc,
+            make_tuple(
+                make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice, C)),
+                make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice))),
+            make_tuple(Sequence<0, 2, 4, 6, 7>{}, Sequence<1, 3, 5>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto out_grid_desc_reducem_reducek = transform_tensor_descriptor(
+            out_grid_desc_reducemraw_reducekraw,
+            make_tuple(make_right_pad_transform(MRaw, MPad), make_right_pad_transform(KRaw, KPad)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        // In[ReduceM]
+        const auto in_n_di_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor(make_tuple(N, Di, Hi, Wi, C),
+                                         make_tuple(din_n_c_wos_strides[0],
+                                                    din_n_c_wos_strides[2],
+                                                    din_n_c_wos_strides[3],
+                                                    din_n_c_wos_strides[4],
+                                                    din_n_c_wos_strides[1]));
+
+        const auto in_n_dip_hip_wip_c_grid_desc = transform_tensor_descriptor(
+            in_n_di_hi_wi_c_grid_desc,
+            make_tuple(make_pass_through_transform(N),
+                       make_pad_transform(Di, InLeftPadD, InRightPadD),
+                       make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                       make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+        const auto in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc =
+            transform_tensor_descriptor(
+                in_n_dip_hip_wip_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(XTilde, DTilde),
+                                                make_tuple(ConvDilationD, ConvStrideD)),
+                           make_embed_transform(make_tuple(YTilde, HTilde),
+                                                make_tuple(ConvDilationH, ConvStrideH)),
+                           make_embed_transform(make_tuple(XTilde, WTilde),
+                                                make_tuple(ConvDilationW, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+        const auto in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc =
+            transform_tensor_descriptor(
+                in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_freeze_transform(i_ztilde),
+                           make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                           make_freeze_transform(i_ytilde),
+                           make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                           make_freeze_transform(i_xtilde),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{},
+                           Sequence<6>{},
+                           Sequence<7>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<>{},
+                           Sequence<1>{},
+                           Sequence<>{},
+                           Sequence<2>{},
+                           Sequence<>{},
+                           Sequence<3>{},
+                           Sequence<4>{}));
+
+        const auto in_grid_desc_reducemraw = transform_tensor_descriptor(
+            in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc,
+            make_tuple(
+                make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice, C))),
+            make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto in_grid_desc_reducem =
+            transform_tensor_descriptor(in_grid_desc_reducemraw,
+                                        make_tuple(make_right_pad_transform(MRaw, MPad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+
+        return make_tuple(out_grid_desc_reducem_reducek, in_grid_desc_reducem);
+    }
+
+    using DoutDinGridDesc = decltype(Make3DGridDescriptor_Out_M_K_In_M({0, 0, 0, 0, 0},
+                                                                       {0, 0, 0, 0, 0},
+                                                                       {0, 0, 0, 0, 0},
+                                                                       {0, 0, 0, 0, 0},
+                                                                       {0, 0, 0},
+                                                                       {0, 0, 0},
+                                                                       {0, 0, 0},
+                                                                       {0, 0, 0},
+                                                                       {0, 0, 0},
+                                                                       {0, 0, 0}));
+
+    using DoutGridDesc_M_K = remove_cvref_t<tuple_element_t<0, DoutDinGridDesc>>;
+    using DinGridDesc_M    = remove_cvref_t<tuple_element_t<1, DoutDinGridDesc>>;
+
+    // FIXME
+    // for NDHWC, the dim C is the fastest dimension, and is not reduced.
+    // Hence, it is in M dimension for reduction kernel.
+    static constexpr index_t OutSrcInDstVectorDim = 0; // 0: M, 1: K
+
+    using PassThrough = tensor_operation::element_wise::PassThrough;
+    using Div         = tensor_operation::element_wise::UnaryDivide;
+
+    using gridwise_reduce = GridwiseReduction_mk_to_m_threadwise<DOutDataType,
+                                                                 DInDataType,
+                                                                 ComputeDataType,
+                                                                 int,
+                                                                 DoutGridDesc_M_K,
+                                                                 DinGridDesc_M,
+                                                                 reduce::Add,
+                                                                 PassThrough,
+                                                                 Div,
+                                                                 InMemoryDataOperationEnum::Set,
+                                                                 false, // propagate_nan
+                                                                 BlockSize,
+                                                                 MThreadSliceSize,
+                                                                 KThreadSliceSize,
+                                                                 OutSrcInDstVectorDim,
+                                                                 InSrcOutDstVectorSize,
+                                                                 InSrcOutDstVectorSize>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const DOutDataType* p_dout,
+                 DInDataType* p_din,
+                 std::vector<ck::index_t> dout_n_c_wos_lengths,
+                 std::vector<ck::index_t> din_n_c_wos_length,
+                 std::vector<ck::index_t> dout_n_c_wos_strides,
+                 std::vector<ck::index_t> din_n_c_wos_strides,
+                 std::vector<ck::index_t> window_lengths,
+                 std::vector<ck::index_t> window_strides,
+                 std::vector<ck::index_t> window_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads)
+            : p_dout_grid_{p_dout},
+              p_din_grid_{p_din},
+              dout_n_c_wos_lengths_{dout_n_c_wos_lengths},
+              din_n_c_wos_length_{din_n_c_wos_length},
+              dout_n_c_wos_strides_{dout_n_c_wos_strides},
+              din_n_c_wos_strides_{din_n_c_wos_strides},
+              num_reduce_{1},
+              div_element_op_{window_lengths[0] * window_lengths[1] * window_lengths[2]}
+        {
+            std::vector<ck::index_t> Tildes(NDimSpatial);
+            for(int i = 0; i < NDimSpatial; ++i)
+            {
+                int GcdStrideDilation = math::gcd(window_strides[i], window_dilations[i]);
+                Tildes[i]             = window_strides[i] / GcdStrideDilation;
+                num_reduce_ *= Tildes[i];
+            }
+
+            for(index_t i_ztilde = 0; i_ztilde < Tildes[0]; ++i_ztilde)
+            {
+                for(index_t i_ytilde = 0; i_ytilde < Tildes[1]; ++i_ytilde)
+                {
+                    for(index_t i_xtilde = 0; i_xtilde < Tildes[2]; ++i_xtilde)
+                    {
+                        // check slice is valid
+                        const auto ZDotSlice =
+                            math::integer_divide_ceil(window_lengths[0] - i_ztilde, Tildes[0]);
+                        const auto YDotSlice =
+                            math::integer_divide_ceil(window_lengths[1] - i_ytilde, Tildes[1]);
+                        const auto XDotSlice =
+                            math::integer_divide_ceil(window_lengths[2] - i_xtilde, Tildes[2]);
+
+                        if(ZDotSlice * YDotSlice * XDotSlice <= 0)
+                        {
+                            continue;
+                        }
+
+                        const auto dout_din_grid_desc =
+                            Make3DGridDescriptor_Out_M_K_In_M(dout_n_c_wos_lengths,
+                                                              din_n_c_wos_length,
+                                                              dout_n_c_wos_strides,
+                                                              din_n_c_wos_strides,
+                                                              window_lengths,
+                                                              window_strides,
+                                                              window_dilations,
+                                                              input_left_pads,
+                                                              input_right_pads,
+                                                              {i_ztilde, i_ytilde, i_xtilde});
+
+                        dout_grid_desc_m_k_container_.push_back(dout_din_grid_desc[I0]);
+                        din_grid_desc_m_container_.push_back(dout_din_grid_desc[I1]);
+                    }
+                }
+            }
+        }
+
+        const DOutDataType* p_dout_grid_;
+        DInDataType* p_din_grid_;
+        std::vector<ck::index_t> dout_n_c_wos_lengths_;
+        std::vector<ck::index_t> din_n_c_wos_length_;
+        std::vector<ck::index_t> dout_n_c_wos_strides_;
+        std::vector<ck::index_t> din_n_c_wos_strides_;
+
+        int num_reduce_;
+        std::vector<DoutGridDesc_M_K> dout_grid_desc_m_k_container_;
+        std::vector<DinGridDesc_M> din_grid_desc_m_container_;
+
+        Div div_element_op_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            float ave_time = 0;
+
+            for(index_t i = 0; i < arg.num_reduce_; i++)
+            {
+                const auto kernel = kernel_reduce_threadwise<gridwise_reduce,
+                                                             false,
+                                                             false,
+                                                             false, // don't have index input
+                                                             DOutDataType,
+                                                             DInDataType,
+                                                             ComputeDataType,
+                                                             int,
+                                                             DoutGridDesc_M_K,
+                                                             DinGridDesc_M,
+                                                             PassThrough,
+                                                             Div>;
+
+                ck::index_t M           = arg.dout_grid_desc_m_k_container_[i].GetLength(I0);
+                const index_t grid_size = (M / M_BlockTileSize);
+
+                ave_time += launch_and_time_kernel(stream_config,
+                                                   kernel,
+                                                   dim3(grid_size),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.dout_grid_desc_m_k_container_[i],
+                                                   arg.din_grid_desc_m_container_[i],
+                                                   PassThrough{},
+                                                   arg.div_element_op_,
+                                                   float(1),
+                                                   arg.p_dout_grid_,
+                                                   nullptr,
+                                                   float(0),
+                                                   arg.p_din_grid_,
+                                                   nullptr);
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        constexpr index_t Rank = NDimSpatial + 2;
+        int doutFastestDim     = -1;
+        int dinFastestDim      = -1;
+
+        for(int i = 0; i < Rank; ++i)
+        {
+            if(arg.dout_n_c_wos_strides_[i] == 1)
+                doutFastestDim = i;
+            if(arg.din_n_c_wos_strides_[i] == 1)
+                dinFastestDim = i;
+        }
+
+        if(doutFastestDim == -1 || dinFastestDim == -1)
+        {
+            if constexpr(InSrcOutDstVectorSize != 1)
+                return false;
+        }
+        else
+        {
+            if(arg.dout_n_c_wos_lengths_[doutFastestDim] % InSrcOutDstVectorSize != 0)
+                return false;
+            if(arg.din_n_c_wos_length_[dinFastestDim] % InSrcOutDstVectorSize != 0)
+                return false;
+        }
+
+        return true;
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_dout,
+                        void* p_din,
+                        std::vector<ck::index_t> dout_n_c_wos_lengths,
+                        std::vector<ck::index_t> din_n_c_wos_length,
+                        std::vector<ck::index_t> dout_n_c_wos_strides,
+                        std::vector<ck::index_t> din_n_c_wos_strides,
+                        std::vector<ck::index_t> window_lengths,
+                        std::vector<ck::index_t> window_strides,
+                        std::vector<ck::index_t> window_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads) override
+    {
+        constexpr index_t Rank = NDimSpatial + 2;
+
+        if(dout_n_c_wos_strides.size() != Rank || din_n_c_wos_strides.size() != Rank ||
+           dout_n_c_wos_lengths.size() != Rank || din_n_c_wos_length.size() != Rank)
+            throw std::runtime_error("dimension is incorrect");
+
+        if(window_lengths.size() != NDimSpatial || window_strides.size() != NDimSpatial ||
+           window_dilations.size() != NDimSpatial || input_left_pads.size() != NDimSpatial ||
+           input_right_pads.size() != NDimSpatial)
+            throw std::runtime_error("dimension is incorrect");
+
+        return std::make_unique<Argument>(static_cast<const DOutDataType*>(p_dout),
+                                          static_cast<DInDataType*>(p_din),
+                                          dout_n_c_wos_lengths,
+                                          din_n_c_wos_length,
+                                          dout_n_c_wos_strides,
+                                          din_n_c_wos_strides,
+                                          window_lengths,
+                                          window_strides,
+                                          window_dilations,
+                                          input_left_pads,
+                                          input_right_pads);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceAvgPool3dBwd<" << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str <<"InSrcOutDstVectorSize_" << InSrcOutDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp
new file mode 100644
index 0000000000..4d599e8017
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp
@@ -0,0 +1,992 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+
+// NOTE: TensorSpecialization::Packed specialized tensor is "packed" in a sense that each inner
+// dimension in a dimension group (eg [G0, G1] in Gs, [M0, M1, M2] in Ms, etc.) are contiguous and
+// ordered. Not in a sense that the tensor [G0, G1, ..., M0, M1, ..., N0, N1...] can be permuted
+// while still being a contiguous, unpadded tensor. In other words, it merely degenerates into
+// TensorSpecialization::Default with NumDimG/M/N/K = 1
+//
+// Detail- Packed tensor satisfies
+//   stride_0 = 1
+//   stride_i = stride_{i - 1} * extent_{i - 1}
+// So tensor
+//   [G0, G1, G2, M, N]
+// transposed into tensor
+//   [G0, G2, G1, M, N]
+// with strides
+//   [G2 * G1 * M * N, G1 * M * N, M * N, N, 1]
+// is again a packed tensor. MakeGridDescriptor() currently just merges dimensions and ignores some
+// strides from input tensor extents so finer dimension information is lost. Merging dimensions is
+// essentially a degenerated case of TensorSpecialization::Default with NumDimG/M/N/K = 1.
+//
+// Might need to expose dimension order to the interface to fully support
+// TensorSpecialization::Packed in a traditional sense of "packed" tensor
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization DESpec,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerWMMA,
+          ck::index_t NPerWMMA,
+          ck::index_t MRepeat,
+          ck::index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          typename CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          ck::index_t NumPrefetch         = 1,
+          ck::LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>
+struct DeviceBatchedContractionMultipleD_Wmma_CShuffle
+    : public DeviceBatchedContractionMultipleD<NumDimG,
+                                               NumDimM,
+                                               NumDimN,
+                                               NumDimK,
+                                               ADataType,
+                                               BDataType,
+                                               DsDataType,
+                                               EDataType,
+                                               AElementwiseOperation,
+                                               BElementwiseOperation,
+                                               CDEElementwiseOperation>
+{
+    using DeviceOp                      = DeviceBatchedContractionMultipleD_Wmma_CShuffle;
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    // K1 = Max Vector Access Pixels
+    static constexpr auto K1Number = Number<K1>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock* K1};
+
+    // Assume: A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+    static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        assert(a_gs_ms_ks_lengths_vec.size() == NumDimG + NumDimM + NumDimK &&
+               a_gs_ms_ks_strides_vec.size() == NumDimG + NumDimM + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto a_ms_ks_lengths = to_tuple(
+            a_gs_ms_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimK>{});
+        const auto a_ms_ks_strides = to_tuple(
+            a_gs_ms_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimK>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimK, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(a_ms_ks_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(a_ms_ks_lengths, kDimIds);
+
+        if constexpr(ASpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto a_grid_desc_mraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(M, K),
+                make_tuple(a_ms_ks_strides[Number<NumDimM - 1>{}],
+                           a_ms_ks_strides[Number<NumDimM + NumDimK - 1>{}]));
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+        else
+        {
+            // naive tensor A[M0, M1, M2, ..., K0, K1, K2...]
+            const auto a_grid_desc_ms_ks =
+                make_naive_tensor_descriptor(a_ms_ks_lengths, a_ms_ks_strides);
+
+            // transformed tensor A[MRaw = M0 * M1 * M2 * ... , KRaw = K0 * K1 * K2 * ...]
+            const auto a_grid_desc_mraw_kraw = transform_tensor_descriptor(
+                a_grid_desc_ms_ks,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(kLengths)),
+                make_tuple(mDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+    }
+
+    // Assume: B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+    static auto MakeBGridDescriptor_N_K(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
+                                        const std::vector<index_t>& b_gs_ns_ks_strides_vec)
+    {
+        assert(b_gs_ns_ks_lengths_vec.size() == NumDimG + NumDimN + NumDimK &&
+               b_gs_ns_ks_strides_vec.size() == NumDimG + NumDimN + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto b_ns_ks_lengths = to_tuple(
+            b_gs_ns_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimN + NumDimK>{});
+        const auto b_ns_ks_strides = to_tuple(
+            b_gs_ns_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimN + NumDimK>{});
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<0, NumDimN, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimN, NumDimN + NumDimK, 1>::type{};
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(b_ns_ks_lengths, kDimIds);
+
+        // lengths for N0, N1, ...
+        const auto nLengths = get_container_subset(b_ns_ks_lengths, nDimIds);
+
+        if constexpr(BSpec == TensorSpecialization::Packed)
+        {
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto b_grid_desc_nraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(N, K),
+                make_tuple(b_ns_ks_strides[Number<NumDimN - 1>{}],
+                           b_ns_ks_strides[Number<NumDimN + NumDimK - 1>{}]));
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+        else
+        {
+            // naive tensor B[N0, N1, N2, ..., K0, K1, K2, ...]
+            const auto b_grid_desc_ns_ks =
+                make_naive_tensor_descriptor(b_ns_ks_lengths, b_ns_ks_strides);
+
+            // transformed tensor B[NRaw = N0 * N1 * N2 * ..., KRaw = K0 * K1 * K2 * ...]
+            const auto b_grid_desc_nraw_kraw = transform_tensor_descriptor(
+                b_grid_desc_ns_ks,
+                make_tuple(make_merge_transform(nLengths), make_merge_transform(kLengths)),
+                make_tuple(nDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+    }
+
+    // assume E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
+                                        const std::vector<index_t>& e_gs_ms_ns_strides_vec)
+    {
+        assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               e_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto e_ms_ns_lengths = to_tuple(
+            e_gs_ms_ns_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto e_ms_ns_strides = to_tuple(
+            e_gs_ms_ns_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(M, N),
+                make_tuple(e_ms_ns_strides[Number<NumDimM - 1>{}],
+                           e_ms_ns_strides[Number<NumDimM + NumDimN - 1>{}]));
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+        else
+        {
+            // naive tensor E[M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_ms_ns =
+                make_naive_tensor_descriptor(e_ms_ns_lengths, e_ms_ns_strides);
+
+            // transformed tensor E[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
+            const auto e_grid_desc_mraw_nraw = transform_tensor_descriptor(
+                e_grid_desc_ms_ns,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+                make_tuple(mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+    }
+
+    // assume E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_G_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
+                                          const std::vector<index_t>& e_gs_ms_ns_strides_vec)
+    {
+        assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               e_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto e_gs_ms_ns_lengths =
+            to_tuple(e_gs_ms_ns_lengths_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto e_gs_ms_ns_strides =
+            to_tuple(e_gs_ms_ns_strides_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // dimension Ids for G0, G1, ...
+        constexpr auto gDimIds = typename arithmetic_sequence_gen<0, NumDimG, 1>::type{};
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds =
+            typename arithmetic_sequence_gen<NumDimG, NumDimG + NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<NumDimG + NumDimM,
+                                                                  NumDimG + NumDimM + NumDimN,
+                                                                  1>::type{};
+
+        // lengths for G0, G1, ...
+        const auto gLengths = get_container_subset(e_gs_ms_ns_lengths, gDimIds);
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_gs_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_gs_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto G = container_reduce(gLengths, math::multiplies{}, Number<1>{});
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_g_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(G, M, N),
+                make_tuple(e_gs_ms_ns_strides[Number<NumDimG - 1>{}],
+                           e_gs_ms_ns_strides[Number<NumDimG + NumDimM - 1>{}],
+                           e_gs_ms_ns_strides[Number<NumDimG + NumDimM + NumDimN - 1>{}]));
+            // return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
+            return e_grid_desc_g_mraw_nraw;
+        }
+        else
+        {
+            // naive tensor E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_gs_ms_ns =
+                make_naive_tensor_descriptor(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+            // transformed tensor E[G = G0 * G1 * ..., MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
+            // N2 * ...]
+            const auto e_grid_desc_g_mraw_nraw = transform_tensor_descriptor(
+                e_grid_desc_gs_ms_ns,
+                make_tuple(make_merge_transform(gLengths),
+                           make_merge_transform(mLengths),
+                           make_merge_transform(nLengths)),
+                make_tuple(gDimIds, mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
+            return e_grid_desc_g_mraw_nraw;
+        }
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_M_N(ds_gs_ms_ns_lengths_vec[i],
+                                                         ds_gs_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    static auto MakeDsGridDescriptor_G_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_G_M_N(ds_gs_ms_ns_lengths_vec[i],
+                                                           ds_gs_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // Gridwise descriptor, mapping to whole given provblem.
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K({}, {}));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K({}, {}));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N({}, {}));
+
+    using DsGridDesc_G_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_G_M_N({}, {}))>;
+    using EGridDesc_G_M_N  = decltype(MakeEGridDescriptor_G_M_N({}, {}));
+
+    struct ComputePtrOffsetOfStridedBatch
+    {
+        ComputePtrOffsetOfStridedBatch(index_t batch_stride_A,
+                                       index_t batch_stride_B,
+                                       DsGridDesc_G_M_N ds_grid_desc_g_m_n,
+                                       EGridDesc_G_M_N e_grid_desc_g_m_n)
+            : batch_stride_A_(batch_stride_A),
+              batch_stride_B_(batch_stride_B),
+              ds_grid_desc_g_m_n_(ds_grid_desc_g_m_n),
+              e_grid_desc_g_m_n_(e_grid_desc_g_m_n)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) * batch_stride_A_;
+        }
+
+        __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) * batch_stride_B_;
+        }
+
+        __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+        {
+            std::array<long_index_t, NumDTensor> ds_offset;
+
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                ds_offset[i] = static_cast<long_index_t>(g_idx) *
+                               ds_grid_desc_g_m_n_[i].CalculateOffset(make_multi_index(1, 0, 0));
+            });
+
+            return ds_offset;
+        }
+
+        __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) *
+                   e_grid_desc_g_m_n_.CalculateOffset(make_multi_index(1, 0, 0));
+        }
+
+        private:
+        index_t batch_stride_A_;
+        index_t batch_stride_B_;
+        DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
+        EGridDesc_G_M_N e_grid_desc_g_m_n_;
+    };
+
+    // A desc for source in blockwise copy
+    template <typename AGridDesc_M_K>
+    __host__ __device__ static constexpr auto
+    MakeAGridDescriptor_K0_M_K1(const AGridDesc_M_K& a_grid_desc_m_k)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_unmerge_transform(make_tuple(AK0, K1)), make_pass_through_transform(M)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B desc for source in blockwise copy
+    template <typename BGridDesc_N_K>
+    __host__ __device__ static constexpr auto
+    MakeBGridDescriptor_K0_N_K1(const BGridDesc_N_K& b_grid_desc_n_k)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            b_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(BK0, K1)), make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    using AGridDesc_K0_M_K1 = decltype(DeviceOp::MakeAGridDescriptor_K0_M_K1(AGridDesc_M_K{}));
+    using BGridDesc_K0_N_K1 = decltype(DeviceOp::MakeBGridDescriptor_K0_N_K1(BGridDesc_N_K{}));
+
+    // GridwiseOp
+    using GridwiseOp = GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle<
+        // DataType Family
+        ADataType,
+        BDataType,
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        // InMemory Data Descriptor
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        DsGridDesc_M_N,
+        EGridDesc_M_N,
+        // ElementwiseOp Family
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        // Tiling Family
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerWMMA,
+        NPerWMMA,
+        K1,
+        MRepeat,
+        NRepeat,
+        // ThreadCluster Family
+        BlockSize,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMRepeatPerShuffle,
+        CShuffleNRepeatPerShuffle,
+        CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+        NumPrefetch,
+        LoopSched,
+        PipelineVer>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 const std::vector<index_t>& a_gs_ms_ks_lengths,
+                 const std::vector<index_t>& b_gs_ns_ks_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                 const std::vector<index_t>& e_gs_ms_ns_lengths,
+                 const std::vector<index_t>& a_gs_ms_ks_strides,
+                 const std::vector<index_t>& b_gs_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                 const std::vector<index_t>& e_gs_ms_ns_strides,
+                 index_t M01,
+                 index_t N01,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_m_k_{},
+              b_grid_desc_n_k_{},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{},
+              ds_grid_desc_g_m_n_{
+                  DeviceOp::MakeDsGridDescriptor_G_M_N(ds_gs_ms_ns_lengths, ds_gs_ms_ns_strides)},
+              e_grid_desc_g_m_n_{
+                  DeviceOp::MakeEGridDescriptor_G_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides)},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_mz_stride_{},
+              a_kz_stride_{},
+              b_nz_stride_{},
+              b_kz_stride_{},
+              ds_nz_stride_{},
+              e_nz_stride_{},
+              a_batch_stride_{a_gs_ms_ks_strides[NumDimG - 1]},
+              b_batch_stride_{b_gs_ns_ks_strides[NumDimG - 1]},
+              compute_ptr_offset_of_batch_{
+                  a_batch_stride_, b_batch_stride_, ds_grid_desc_g_m_n_, e_grid_desc_g_m_n_}
+        {
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+            });
+
+            a_grid_desc_m_k_ =
+                DeviceOp::MakeAGridDescriptor_M_K(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+            b_grid_desc_n_k_ =
+                DeviceOp::MakeBGridDescriptor_N_K(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
+
+            ds_grid_desc_m_n_ =
+                DeviceOp::MakeDsGridDescriptor_M_N(ds_gs_ms_ns_lengths, ds_gs_ms_ns_strides);
+
+            e_grid_desc_m_n_ =
+                DeviceOp::MakeEGridDescriptor_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+            a_grid_desc_k0_m_k1_ = DeviceOp::MakeAGridDescriptor_K0_M_K1(a_grid_desc_m_k_);
+            b_grid_desc_k0_n_k1_ = DeviceOp::MakeBGridDescriptor_K0_N_K1(b_grid_desc_n_k_);
+
+            block_2_ctile_map_ = GridwiseOp::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_, M01, N01);
+
+            ds_grid_desc_mblock_mperblock_nblock_nperblock =
+                GridwiseOp::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                    ds_grid_desc_m_n_);
+
+            e_grid_desc_mblock_mperblock_nblock_nperblock =
+                GridwiseOp::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n_);
+
+            // for sanity check of vector memory access
+            a_mz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM - 1];
+            a_kz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1];
+            b_nz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN - 1];
+            b_kz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1];
+
+            for(index_t i = 0; i < NumDTensor; ++i)
+            {
+                ds_nz_stride_[i] = ds_gs_ms_ns_strides[i][NumDimG + NumDimM + NumDimN - 1];
+            }
+
+            e_nz_stride_ = e_gs_ms_ns_strides[NumDimG + NumDimM + NumDimN - 1];
+        }
+
+        // Pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseOp::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // Tensor Descriptors
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+        DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
+        EGridDesc_G_M_N e_grid_desc_g_m_n_;
+
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+
+        typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock;
+        typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock;
+
+        // Block to Tile mapping
+        typename GridwiseOp::DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // Idle
+        index_t M01_;
+        index_t N01_;
+
+        // ElementwiseOp
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // Strides for the last M/N/K dimensions of A/B/Ds/E
+        //   for sanity check of vector load/store
+        index_t a_mz_stride_;
+        index_t a_kz_stride_;
+        index_t b_nz_stride_;
+        index_t b_kz_stride_;
+        std::array<index_t, NumDTensor> ds_nz_stride_;
+        index_t e_mz_stride_;
+        index_t e_nz_stride_;
+
+        index_t a_batch_stride_;
+        index_t b_batch_stride_;
+
+        // Batch Offset
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const index_t G = arg.e_grid_desc_g_m_n_.GetLength(I0);
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * G;
+
+            const auto K =
+                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_contraction_multiple_d_wmma_cshuffle<
+                    GridwiseOp,
+                    ADataType,
+                    BDataType,
+                    typename GridwiseOp::DsGridPointer,
+                    EDataType,
+                    DeviceOp::AGridDesc_K0_M_K1,
+                    DeviceOp::BGridDesc_K0_N_K1,
+                    typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    ComputePtrOffsetOfStridedBatch,
+                    typename GridwiseOp::DefaultBlock2CTileMap,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              G,
+                                              arg.a_grid_desc_k0_m_k1_,
+                                              arg.b_grid_desc_k0_n_k1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.compute_ptr_offset_of_batch_,
+                                              arg.block_2_ctile_map_);
+            };
+
+            if(GridwiseOp::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
+           ck::get_device_name() == "gfx1102")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        if(!GridwiseOp::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                      arg.b_grid_desc_k0_n_k1_,
+                                      arg.ds_grid_desc_m_n_,
+                                      arg.e_grid_desc_m_n_,
+                                      arg.block_2_ctile_map_))
+        {
+            return false;
+        }
+
+        // check vector access
+        static_assert((ABlockTransferSrcVectorDim == 1 || ABlockTransferSrcVectorDim == 2) &&
+                          (BBlockTransferSrcVectorDim == 1 || BBlockTransferSrcVectorDim == 2),
+                      "wrong!");
+
+        // vector memory access of A: could be on M or AK1 dimension
+        if constexpr(ABlockTransferSrcVectorDim == 1)
+        {
+            if(!(arg.a_mz_stride_ == 1 &&
+                 arg.a_grid_desc_k0_m_k1_.GetLength(I1) % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.a_kz_stride_ == 1 &&
+                 arg.a_grid_desc_k0_m_k1_.GetLength(I2) % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of B: could be on N or BK1 dimension
+        if constexpr(BBlockTransferSrcVectorDim == 1)
+        {
+            if(!(arg.b_nz_stride_ == 1 &&
+                 arg.b_grid_desc_k0_n_k1_.GetLength(I1) % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.b_kz_stride_ == 1 &&
+                 arg.b_grid_desc_k0_n_k1_.GetLength(I2) % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of Ds: always on NPerBlock dimension
+        bool valid_d_access = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            if(!(arg.ds_nz_stride_[i] == 1 &&
+                 arg.ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetLength(I3) %
+                         CDEShuffleBlockTransferScalarPerVector_NPerBlock ==
+                     0))
+            {
+                valid_d_access = false;
+            }
+        });
+
+        if(valid_d_access == false)
+        {
+            return false;
+        }
+
+        // vector memory access of E: always on NPerBlock dimension
+        if(!((arg.e_nz_stride_ == 1 &&
+              arg.e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I3) %
+                      CDEShuffleBlockTransferScalarPerVector_NPerBlock ==
+                  0) ||
+             CDEShuffleBlockTransferScalarPerVector_NPerBlock == 1))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto
+    MakeArgument(const void* p_a,
+                 const void* p_b,
+                 std::array<const void*, NumDTensor> p_ds,
+                 void* p_e,
+                 const std::vector<index_t>& a_gs_ms_ks_lengths,
+                 const std::vector<index_t>& a_gs_ms_ks_strides,
+                 const std::vector<index_t>& b_gs_ns_ks_lengths,
+                 const std::vector<index_t>& b_gs_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                 const std::vector<index_t>& e_gs_ms_ns_lengths,
+                 const std::vector<index_t>& e_gs_ms_ns_strides,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_gs_ms_ks_lengths,
+                        b_gs_ns_ks_lengths,
+                        ds_gs_ms_ns_lengths,
+                        e_gs_ms_ns_lengths,
+                        a_gs_ms_ks_strides,
+                        b_gs_ns_ks_strides,
+                        ds_gs_ms_ns_strides,
+                        e_gs_ms_ns_strides,
+                        1,
+                        1,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::vector<index_t>& a_gs_ms_ks_lengths,
+                        const std::vector<index_t>& a_gs_ms_ks_strides,
+                        const std::vector<index_t>& b_gs_ns_ks_lengths,
+                        const std::vector<index_t>& b_gs_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                        const std::vector<index_t>& e_gs_ms_ns_lengths,
+                        const std::vector<index_t>& e_gs_ms_ns_strides,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_gs_ms_ks_lengths,
+                                          b_gs_ns_ks_lengths,
+                                          ds_gs_ms_ns_lengths,
+                                          e_gs_ms_ns_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b_gs_ns_ks_strides,
+                                          ds_gs_ms_ns_strides,
+                                          e_gs_ms_ns_strides,
+                                          1,
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
+        // clang-format off
+        str << "DeviceBatchedContractionMultipleD_Wmma_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerWMMA << ", "
+            << NPerWMMA << ", "
+            << MRepeat << ", "
+            << NRepeat
+            << ">"
+            << " NumPrefetch: "
+            << NumPrefetch << ", "
+            << "LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..32c45bc57e
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,1049 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatDsPointer,
+          typename FloatE,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename ComputePtrOffsetOfBatch,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_contraction_multiple_d_xdl_cshuffle(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatDsPointer p_ds_grid,
+            FloatE* __restrict__ p_e_grid,
+            const index_t batch_count,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
+            const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    FloatDsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_ds_grid_grp,
+                                                  p_e_grid + e_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_etile_map;
+    ignore = compute_ptr_offset_of_batch;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+
+// NOTE: TensorSpecialization::Packed specialized tensor is "packed" in a sense that each inner
+// dimension in a dimension group (eg [G0, G1] in Gs, [M0, M1, M2] in Ms, etc.) are contiguous and
+// ordered. Not in a sense that the tensor [G0, G1, ..., M0, M1, ..., N0, N1...] can be permuted
+// while still being a contiguous, unpadded tensor. In other words, it merely degenerates into
+// TensorSpecialization::Default with NumDimG/M/N/K = 1
+//
+// Detail- Packed tensor satisfies
+//   stride_0 = 1
+//   stride_i = stride_{i - 1} * extent_{i - 1}
+// So tensor
+//   [G0, G1, G2, M, N]
+// transposed into tensor
+//   [G0, G2, G1, M, N]
+// with strides
+//   [G2 * G1 * M * N, G1 * M * N, M * N, N, 1]
+// is again a packed tensor. MakeGridDescriptor() currently just merges dimensions and ignores some
+// strides from input tensor extents so finer dimension information is lost. Merging dimensions is
+// essentially a degenerated case of TensorSpecialization::Default with NumDimG/M/N/K = 1.
+//
+// Might need to expose dimension order to the interface to fully support
+// TensorSpecialization::Packed in a traditional sense of "packed" tensor
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization DESpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceBatchedContractionMultipleD_Xdl_CShuffle
+    : public DeviceBatchedContractionMultipleD<NumDimG,
+                                               NumDimM,
+                                               NumDimN,
+                                               NumDimK,
+                                               ADataType,
+                                               BDataType,
+                                               DsDataType,
+                                               EDataType,
+                                               AElementwiseOperation,
+                                               BElementwiseOperation,
+                                               CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceBatchedContractionMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    // Assume: A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+    static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        assert(a_gs_ms_ks_lengths_vec.size() == NumDimG + NumDimM + NumDimK &&
+               a_gs_ms_ks_strides_vec.size() == NumDimG + NumDimM + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto a_ms_ks_lengths = to_tuple(
+            a_gs_ms_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimK>{});
+        const auto a_ms_ks_strides = to_tuple(
+            a_gs_ms_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimK>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimK, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(a_ms_ks_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(a_ms_ks_lengths, kDimIds);
+
+        if constexpr(ASpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto a_grid_desc_mraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(M, K),
+                make_tuple(a_ms_ks_strides[Number<NumDimM - 1>{}],
+                           a_ms_ks_strides[Number<NumDimM + NumDimK - 1>{}]));
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+        else
+        {
+            // naive tensor A[M0, M1, M2, ..., K0, K1, K2...]
+            const auto a_grid_desc_ms_ks =
+                make_naive_tensor_descriptor(a_ms_ks_lengths, a_ms_ks_strides);
+
+            // transformed tensor A[MRaw = M0 * M1 * M2 * ... , KRaw = K0 * K1 * K2 * ...]
+            const auto a_grid_desc_mraw_kraw = transform_tensor_descriptor(
+                a_grid_desc_ms_ks,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(kLengths)),
+                make_tuple(mDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+    }
+
+    // Assume: B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+    static auto MakeBGridDescriptor_N_K(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
+                                        const std::vector<index_t>& b_gs_ns_ks_strides_vec)
+    {
+        assert(b_gs_ns_ks_lengths_vec.size() == NumDimG + NumDimN + NumDimK &&
+               b_gs_ns_ks_strides_vec.size() == NumDimG + NumDimN + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto b_ns_ks_lengths = to_tuple(
+            b_gs_ns_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimN + NumDimK>{});
+        const auto b_ns_ks_strides = to_tuple(
+            b_gs_ns_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimN + NumDimK>{});
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<0, NumDimN, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimN, NumDimN + NumDimK, 1>::type{};
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(b_ns_ks_lengths, kDimIds);
+
+        // lengths for N0, N1, ...
+        const auto nLengths = get_container_subset(b_ns_ks_lengths, nDimIds);
+
+        if constexpr(BSpec == TensorSpecialization::Packed)
+        {
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto b_grid_desc_nraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(N, K),
+                make_tuple(b_ns_ks_strides[Number<NumDimN - 1>{}],
+                           b_ns_ks_strides[Number<NumDimN + NumDimK - 1>{}]));
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+        else
+        {
+            // naive tensor B[N0, N1, N2, ..., K0, K1, K2, ...]
+            const auto b_grid_desc_ns_ks =
+                make_naive_tensor_descriptor(b_ns_ks_lengths, b_ns_ks_strides);
+
+            // transformed tensor B[NRaw = N0 * N1 * N2 * ..., KRaw = K0 * K1 * K2 * ...]
+            const auto b_grid_desc_nraw_kraw = transform_tensor_descriptor(
+                b_grid_desc_ns_ks,
+                make_tuple(make_merge_transform(nLengths), make_merge_transform(kLengths)),
+                make_tuple(nDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+    }
+
+    // assume E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
+                                        const std::vector<index_t>& e_gs_ms_ns_strides_vec)
+    {
+        assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               e_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto e_ms_ns_lengths = to_tuple(
+            e_gs_ms_ns_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto e_ms_ns_strides = to_tuple(
+            e_gs_ms_ns_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(M, N),
+                make_tuple(e_ms_ns_strides[Number<NumDimM - 1>{}],
+                           e_ms_ns_strides[Number<NumDimM + NumDimN - 1>{}]));
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+        else
+        {
+            // naive tensor E[M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_ms_ns =
+                make_naive_tensor_descriptor(e_ms_ns_lengths, e_ms_ns_strides);
+
+            // transformed tensor E[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
+            const auto e_grid_desc_mraw_nraw = transform_tensor_descriptor(
+                e_grid_desc_ms_ns,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+                make_tuple(mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+    }
+
+    // assume E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_G_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
+                                          const std::vector<index_t>& e_gs_ms_ns_strides_vec)
+    {
+        assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               e_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto e_gs_ms_ns_lengths =
+            to_tuple(e_gs_ms_ns_lengths_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto e_gs_ms_ns_strides =
+            to_tuple(e_gs_ms_ns_strides_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // dimension Ids for G0, G1, ...
+        constexpr auto gDimIds = typename arithmetic_sequence_gen<0, NumDimG, 1>::type{};
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds =
+            typename arithmetic_sequence_gen<NumDimG, NumDimG + NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<NumDimG + NumDimM,
+                                                                  NumDimG + NumDimM + NumDimN,
+                                                                  1>::type{};
+
+        // lengths for G0, G1, ...
+        const auto gLengths = get_container_subset(e_gs_ms_ns_lengths, gDimIds);
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_gs_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_gs_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto G = container_reduce(gLengths, math::multiplies{}, Number<1>{});
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_g_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(G, M, N),
+                make_tuple(e_gs_ms_ns_strides[Number<NumDimG - 1>{}],
+                           e_gs_ms_ns_strides[Number<NumDimG + NumDimM - 1>{}],
+                           e_gs_ms_ns_strides[Number<NumDimG + NumDimM + NumDimN - 1>{}]));
+            // return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
+            return e_grid_desc_g_mraw_nraw;
+        }
+        else
+        {
+            // naive tensor E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_gs_ms_ns =
+                make_naive_tensor_descriptor(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+            // transformed tensor E[G = G0 * G1 * ..., MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
+            // N2 * ...]
+            const auto e_grid_desc_g_mraw_nraw = transform_tensor_descriptor(
+                e_grid_desc_gs_ms_ns,
+                make_tuple(make_merge_transform(gLengths),
+                           make_merge_transform(mLengths),
+                           make_merge_transform(nLengths)),
+                make_tuple(gDimIds, mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
+            return e_grid_desc_g_mraw_nraw;
+        }
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_M_N(ds_gs_ms_ns_lengths_vec[i],
+                                                         ds_gs_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    static auto MakeDsGridDescriptor_G_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_G_M_N(ds_gs_ms_ns_lengths_vec[i],
+                                                           ds_gs_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K({}, {}));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K({}, {}));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({{}}, {{}}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N({}, {}));
+
+    using DsGridDesc_G_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_G_M_N({}, {}))>;
+    using EGridDesc_G_M_N  = decltype(MakeEGridDescriptor_G_M_N({}, {}));
+
+    struct ComputePtrOffsetOfStridedBatch
+    {
+        ComputePtrOffsetOfStridedBatch(index_t batch_stride_A,
+                                       index_t batch_stride_B,
+                                       DsGridDesc_G_M_N ds_grid_desc_g_m_n,
+                                       EGridDesc_G_M_N e_grid_desc_g_m_n)
+            : batch_stride_A_(batch_stride_A),
+              batch_stride_B_(batch_stride_B),
+              ds_grid_desc_g_m_n_(ds_grid_desc_g_m_n),
+              e_grid_desc_g_m_n_(e_grid_desc_g_m_n)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) * batch_stride_A_;
+        }
+
+        __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) * batch_stride_B_;
+        }
+
+        __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+        {
+            std::array<long_index_t, NumDTensor> ds_offset;
+
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                ds_offset[i] = static_cast<long_index_t>(g_idx) *
+                               ds_grid_desc_g_m_n_[i].CalculateOffset(make_multi_index(1, 0, 0));
+            });
+
+            return ds_offset;
+        }
+
+        __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) *
+                   e_grid_desc_g_m_n_.CalculateOffset(make_multi_index(1, 0, 0));
+        }
+
+        private:
+        index_t batch_stride_A_;
+        index_t batch_stride_B_;
+        DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
+        EGridDesc_G_M_N e_grid_desc_g_m_n_;
+    };
+
+    using ComputeDataType = ADataType;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType,
+        BDataType,
+        ComputeDataType,
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // desc for blockwise copy
+    using AGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
+            AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
+            BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
+        decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+
+    // block-to-e-tile map
+    using Block2ETileMap =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 const std::vector<index_t>& a_gs_ms_ns_lengths,
+                 const std::vector<index_t>& a_gs_ms_ks_strides,
+                 const std::vector<index_t>& b_gs_ns_ks_lengths,
+                 const std::vector<index_t>& b_gs_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                 const std::vector<index_t>& e_gs_ms_ns_lengths,
+                 const std::vector<index_t>& e_gs_ms_ns_strides,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_m_k_{
+                  DeviceOp::MakeAGridDescriptor_M_K(a_gs_ms_ns_lengths, a_gs_ms_ks_strides)},
+              b_grid_desc_n_k_{
+                  DeviceOp::MakeBGridDescriptor_N_K(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{
+                  DeviceOp::MakeEGridDescriptor_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides)},
+              ds_grid_desc_g_m_n_{
+                  DeviceOp::MakeDsGridDescriptor_G_M_N(ds_gs_ms_ns_lengths, ds_gs_ms_ns_strides)},
+              e_grid_desc_g_m_n_{
+                  DeviceOp::MakeEGridDescriptor_G_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_mz_stride_{},
+              a_kz_stride_{},
+              b_nz_stride_{},
+              b_kz_stride_{},
+              ds_nz_stride_{},
+              e_nz_stride_{},
+              a_batch_stride_{a_gs_ms_ks_strides[NumDimG - 1]},
+              b_batch_stride_{b_gs_ns_ks_strides[NumDimG - 1]},
+              compute_ptr_offset_of_batch_{
+                  a_batch_stride_, b_batch_stride_, ds_grid_desc_g_m_n_, e_grid_desc_g_m_n_}
+        {
+            static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0, "");
+
+            // populate pointer, batch stride, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N(ds_gs_ms_ns_lengths[i],
+                                                                         ds_gs_ms_ns_strides[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           ds_grid_desc_m_n_,
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+            }
+
+            // for sanity check of vector memory access
+            a_mz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM - 1];
+            a_kz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1];
+            b_nz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN - 1];
+            b_kz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1];
+
+            for(index_t i = 0; i < NumDTensor; ++i)
+            {
+                ds_nz_stride_[i] = ds_gs_ms_ns_strides[i][NumDimG + NumDimM + NumDimN - 1];
+            }
+
+            e_nz_stride_ = e_gs_ms_ns_strides[NumDimG + NumDimM + NumDimN - 1];
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
+        EGridDesc_G_M_N e_grid_desc_g_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // Strides for the last M/N/K dimensions of A/B/Ds/E
+        //   for sanity check of vector load/store
+        index_t a_mz_stride_;
+        index_t a_kz_stride_;
+        index_t b_nz_stride_;
+        index_t b_kz_stride_;
+        std::array<index_t, NumDTensor> ds_nz_stride_;
+        index_t e_mz_stride_;
+        index_t e_nz_stride_;
+
+        index_t a_batch_stride_;
+        index_t b_batch_stride_;
+
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.ds_grid_desc_m_n_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemmMultipleD_xdl_cshuffle has invalid setting");
+            }
+
+            const index_t G = arg.e_grid_desc_g_m_n_.GetLength(I0);
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * G;
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_contraction_multiple_d_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    ComputePtrOffsetOfStridedBatch,
+                    DeviceOp::Block2ETileMap,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              G,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.compute_ptr_offset_of_batch_,
+                                              arg.block_2_etile_map_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                        arg.b_grid_desc_n_k_,
+                                        arg.ds_grid_desc_m_n_,
+                                        arg.e_grid_desc_m_n_,
+                                        arg.block_2_etile_map_))
+        {
+            return false;
+        }
+
+        // check vector access
+        static_assert((ABlockTransferSrcVectorDim == 1 || ABlockTransferSrcVectorDim == 2) &&
+                          (BBlockTransferSrcVectorDim == 1 || BBlockTransferSrcVectorDim == 2),
+                      "wrong!");
+
+        // vector memory access of A: could be on M or AK1 dimension
+        if constexpr(ABlockTransferSrcVectorDim == 1)
+        {
+            if(!(arg.a_mz_stride_ == 1 &&
+                 arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.a_kz_stride_ == 1 &&
+                 arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of B: could be on N or BK1 dimension
+        if constexpr(BBlockTransferSrcVectorDim == 1)
+        {
+            if(!(arg.b_nz_stride_ == 1 &&
+                 arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.b_kz_stride_ == 1 &&
+                 arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of Ds: always on NPerBlock dimension
+        bool valid_d_access = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            if(!(arg.ds_nz_stride_[i] == 1 &&
+                 arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_[i].GetLength(I3) %
+                         CDEBlockTransferScalarPerVector_NPerBlock ==
+                     0))
+            {
+                valid_d_access = false;
+            }
+        });
+
+        if(valid_d_access == false)
+        {
+            return false;
+        }
+
+        // vector memory access of E: always on NPerBlock dimension
+        if(!((arg.e_nz_stride_ == 1 &&
+              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_.GetLength(I3) %
+                      CDEBlockTransferScalarPerVector_NPerBlock ==
+                  0) ||
+             CDEBlockTransferScalarPerVector_NPerBlock == 1))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto
+    MakeArgument(const void* p_a,
+                 const void* p_b,
+                 std::array<const void*, NumDTensor> p_ds,
+                 void* p_e,
+                 const std::vector<index_t>& a_gs_ms_ns_lengths,
+                 const std::vector<index_t>& a_gs_ms_ks_strides,
+                 const std::vector<index_t>& b_gs_ns_ks_lengths,
+                 const std::vector<index_t>& b_gs_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                 const std::vector<index_t>& e_gs_ms_ns_lengths,
+                 const std::vector<index_t>& e_gs_ms_ns_strides,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_gs_ms_ns_lengths,
+                        a_gs_ms_ks_strides,
+                        b_gs_ns_ks_lengths,
+                        b_gs_ns_ks_strides,
+                        ds_gs_ms_ns_lengths,
+                        ds_gs_ms_ns_strides,
+                        e_gs_ms_ns_lengths,
+                        e_gs_ms_ns_strides,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::vector<index_t>& a_gs_ms_ns_lengths,
+                        const std::vector<index_t>& a_gs_ms_ks_strides,
+                        const std::vector<index_t>& b_gs_ns_ks_lengths,
+                        const std::vector<index_t>& b_gs_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                        const std::vector<index_t>& e_gs_ms_ns_lengths,
+                        const std::vector<index_t>& e_gs_ms_ns_strides,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_gs_ms_ns_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b_gs_ns_ks_lengths,
+                                          b_gs_ns_ks_strides,
+                                          ds_gs_ms_ns_lengths,
+                                          ds_gs_ms_ns_strides,
+                                          e_gs_ms_ns_lengths,
+                                          e_gs_ms_ns_strides,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedContractionMultipleD_Xdl_CShuffle"
+            << "<"
+            << NumDimG << ", "
+            << NumDimM << ", "
+            << NumDimN << ", "
+            << NumDimK << ", "
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << ABlockTransferSrcVectorDim << ", "
+            << BBlockTransferSrcVectorDim
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp
new file mode 100644
index 0000000000..ba22cf0bf8
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp
@@ -0,0 +1,697 @@
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for
+\link
+ * DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
+ * pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemmCPermute and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename EDataType,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename ComputePtrOffsetOfBatch,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_e_permute_xdl(const ABDataType* __restrict__ p_a_grid,
+                                          const ABDataType* __restrict__ p_b_grid,
+                                          EDataType* __restrict__ p_e_grid,
+                                          const index_t batch_count,
+                                          const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+                                          const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+                                          const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                                              e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                          const AElementwiseOperation a_element_op,
+                                          const BElementwiseOperation b_element_op,
+                                          const CDEElementwiseOperation cde_element_op,
+                                          const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
+                                          const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  ck::Tuple<>{},
+                                                  p_e_grid + e_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  ck::Tuple<>{},
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_etile_map;
+#endif
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumPrefetch,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceBatchedGemmEPermuteXdl : public DeviceBatchedGemmEPermute<ALayout,
+                                                                       BLayout,
+                                                                       ELayout,
+                                                                       ADataType,
+                                                                       BDataType,
+                                                                       EDataType,
+                                                                       AElementwiseOperation,
+                                                                       BElementwiseOperation,
+                                                                       CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceBatchedGemmEPermuteXdl;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    static auto
+    MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t stride_M, index_t stride_N)
+    {
+        const auto e_grid_desc_mraw_nraw =
+            make_naive_tensor_descriptor(make_tuple(MRaw, NRaw), make_tuple(stride_M, stride_N));
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    static auto MakeEGridDescriptor_G0_G1_M_N(index_t G0,
+                                              index_t G1,
+                                              index_t MRaw,
+                                              index_t NRaw,
+                                              index_t stride_G0,
+                                              index_t stride_G1,
+                                              index_t stride_M,
+                                              index_t stride_N)
+    {
+        const auto e_grid_desc_g0_g1_mraw_nraw = [&]() {
+            return make_naive_tensor_descriptor(
+                make_tuple(G0, G1, MRaw, NRaw),
+                make_tuple(stride_G0, stride_G1, stride_M, stride_N));
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto NPad = N - NRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(
+                e_grid_desc_g0_g1_mraw_nraw,
+                make_tuple(make_pass_through_transform(G0),
+                           make_pass_through_transform(G1),
+                           make_right_pad_transform(MRaw, MPad),
+                           make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                e_grid_desc_g0_g1_mraw_nraw,
+                make_tuple(make_pass_through_transform(G0),
+                           make_pass_through_transform(G1),
+                           make_right_pad_transform(MRaw, MPad),
+                           make_pass_through_transform(NRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                e_grid_desc_g0_g1_mraw_nraw,
+                make_tuple(make_pass_through_transform(G0),
+                           make_pass_through_transform(G1),
+                           make_pass_through_transform(MRaw),
+                           make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return e_grid_desc_g0_g1_mraw_nraw;
+        }
+    }
+
+    using AGridDesc_M_K       = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K       = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using EGridDesc_M_N       = decltype(MakeEGridDescriptor_M_N(1, 1, 1, 1));
+    using EGridDesc_G0_G1_M_N = decltype(MakeEGridDescriptor_G0_G1_M_N(1, 1, 1, 1, 1, 1, 1, 1));
+
+    struct ComputePtrOffsetOfStridedBatch
+    {
+        ComputePtrOffsetOfStridedBatch(index_t Batchstride_A,
+                                       index_t Batchstride_B,
+                                       EGridDesc_G0_G1_M_N e_grid_desc_g0_g1_m_n)
+            : Batchstride_A_(Batchstride_A),
+              Batchstride_B_(Batchstride_B),
+              e_grid_desc_g0_g1_m_n_(e_grid_desc_g0_g1_m_n)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(Batchstride_A_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(Batchstride_B_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
+        {
+            const index_t G1 = e_grid_desc_g0_g1_m_n_.GetLength(I1);
+            index_t b0       = g_idx / G1;
+            index_t b1       = g_idx - b0 * G1; // g_idx % G1
+            return e_grid_desc_g0_g1_m_n_.CalculateOffset(make_multi_index(b0, b1, 0, 0));
+        }
+
+        private:
+        index_t Batchstride_A_;
+        index_t Batchstride_B_;
+        EGridDesc_G0_G1_M_N e_grid_desc_g0_g1_m_n_;
+    };
+
+    using ComputeDataType = ADataType;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType,
+        BDataType,
+        ComputeDataType,
+        AccDataType,
+        CShuffleDataType,
+        ck::Tuple<>, // DsDataType,
+        EDataType,   // EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_M_K,
+        BGridDesc_N_K,
+        Tuple<>,
+        EGridDesc_M_N,
+        NumPrefetch,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    using AGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
+            AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
+            BGridDesc_N_K{}))>;
+
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}));
+    using Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 EDataType* p_e_grid,
+                 index_t M,
+                 index_t N,
+                 index_t K,
+                 index_t stride_A,
+                 index_t stride_B,
+                 index_t batch_stride_A,
+                 index_t batch_stride_B,
+                 BatchedGemmEPermuteDesc batched_gemm_e_permute_desc,
+                 index_t BatchCount,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_e_grid_{p_e_grid},
+              BatchCount_(BatchCount),
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(M, K, stride_A)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(K, N, stride_B)},
+              e_grid_desc_m_n_{
+                  DeviceOp::MakeEGridDescriptor_M_N(batched_gemm_e_permute_desc.M_,
+                                                    batched_gemm_e_permute_desc.N_,
+                                                    batched_gemm_e_permute_desc.stride_M_,
+                                                    batched_gemm_e_permute_desc.stride_N_)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              e_grid_desc_mblock_mperblock_nblock_nperblock{},
+              e_grid_desc_g0_g1_m_n_{
+                  DeviceOp::MakeEGridDescriptor_G0_G1_M_N(batched_gemm_e_permute_desc.G0_,
+                                                          batched_gemm_e_permute_desc.G1_,
+                                                          batched_gemm_e_permute_desc.M_,
+                                                          batched_gemm_e_permute_desc.N_,
+                                                          batched_gemm_e_permute_desc.stride_G0_,
+                                                          batched_gemm_e_permute_desc.stride_G1_,
+                                                          batched_gemm_e_permute_desc.stride_M_,
+                                                          batched_gemm_e_permute_desc.stride_N_)},
+              compute_ptr_offset_of_batch_{batch_stride_A, batch_stride_B, e_grid_desc_g0_g1_m_n_},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           ck::Tuple<>{},
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            std::cout << "C[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        EDataType* p_e_grid_;
+
+        // batch count
+        index_t BatchCount_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock;
+        EGridDesc_G0_G1_M_N e_grid_desc_g0_g1_m_n_;
+
+        // for calculating Batch offset
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            ck::Tuple<>{},
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseBatchedGemmCPermute_km_kn_m0m1n0n1_xdlops_v2r3 has invalid "
+                    "setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * arg.BatchCount_;
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel = kernel_batched_gemm_e_permute_xdl<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    EDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_AK0_M_AK1>,
+                    remove_reference_t<DeviceOp::BGridDesc_BK0_N_BK1>,
+                    typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    ComputePtrOffsetOfStridedBatch,
+                    remove_reference_t<Block2ETileMap>,
+                    has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_e_grid_,
+                                              arg.BatchCount_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.compute_ptr_offset_of_batch_,
+                                              arg.block_2_etile_map_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           ck::Tuple<>{},
+                                           arg.e_grid_desc_m_n_,
+                                           arg.block_2_etile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             EDataType* p_e,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t stride_A,
+                             index_t stride_B,
+                             index_t batch_stride_A,
+                             index_t batch_stride_B,
+                             BatchedGemmEPermuteDesc batched_gemm_e_permute_desc,
+                             index_t BatchCount,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_e,
+                        M,
+                        N,
+                        K,
+                        stride_A,
+                        stride_B,
+                        batch_stride_A,
+                        batch_stride_B,
+                        batched_gemm_e_permute_desc,
+                        BatchCount,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_e,
+                        index_t M,
+                        index_t N,
+                        index_t K,
+                        index_t stride_A,
+                        index_t stride_B,
+                        index_t batch_stride_A,
+                        index_t batch_stride_B,
+                        BatchedGemmEPermuteDesc batched_gemm_e_permute_desc,
+                        index_t BatchCount,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<EDataType*>(p_e),
+                                          M,
+                                          N,
+                                          K,
+                                          stride_A,
+                                          stride_B,
+                                          batch_stride_A,
+                                          batch_stride_B,
+                                          batched_gemm_e_permute_desc,
+                                          BatchCount,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedGemmEPermuteXdl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..3dbe8c6722
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp
@@ -0,0 +1,748 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename B1GridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2CTileMap,
+          typename ComputeBasePtrOfStridedBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_gemm_xdl_cshuffle_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            const FloatAB* __restrict__ p_b1_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const AccElementwiseOperation acc_element_op,
+            const B1ElementwiseOperation b1_element_op,
+            const CElementwiseOperation c_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const Block2CTileMap block_2_ctile_map,
+            const index_t batch_count,
+            const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetABasePtr(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetBBasePtr(g_idx)));
+    const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetB1BasePtr(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetCBasePtr(g_idx)));
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_b1_grid + b1_batch_offset,
+                                                  p_c_grid + c_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  acc_element_op,
+                                                  b1_element_op,
+                                                  c_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  b1_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_b1_grid;
+    ignore = p_c_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = acc_element_op;
+    ignore = b1_element_op;
+    ignore = c_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = b1_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_ctile_map;
+    ignore = batch_count;
+    ignore = compute_base_ptr_of_batch;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// Computes C = A * B0 * B1
+//              ^^^^^^ (Acc0)
+//              ^^^^^^^^^^^ (Acc1)
+template <typename ALayout,
+          typename BLayout, // B0Layout
+          typename B1Layout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename B1DataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock, // Gemm0NPerBlock
+          index_t KPerBlock, // Gemm0KPerBlock
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t B1K1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = LoopScheduler::Default>
+struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout,
+                                                                         BLayout,
+                                                                         B1Layout,
+                                                                         CLayout,
+                                                                         ADataType,
+                                                                         BDataType,
+                                                                         B1DataType,
+                                                                         CDataType,
+                                                                         AElementwiseOperation,
+                                                                         BElementwiseOperation,
+                                                                         AccElementwiseOperation,
+                                                                         B1ElementwiseOperation,
+                                                                         CElementwiseOperation>
+{
+    using DeviceOp = DeviceBatchedGemmGemm_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto matrix_padder =
+        GemmGemmPadder<GemmSpec, index_t, index_t, index_t, index_t>{
+            MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock};
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto a_grid_desc_m_k = matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto b_grid_desc_n_k = matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // Args: Gemm1KRaw, Gemm1NRaw, StrideB1
+    static auto MakeB1GridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b1_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, B1Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, B1Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto b1_grid_desc_n_k = matrix_padder.PadB1Descriptor_N_K(b1_grid_desc_nraw_kraw);
+
+        const auto N = b1_grid_desc_n_k.GetLength(I0);
+        const auto K = b1_grid_desc_n_k.GetLength(I1);
+
+        const auto B1K0 = K / B1K1;
+
+        return transform_tensor_descriptor(
+            b1_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
+                       make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideC));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(c_grid_desc_mraw_nraw);
+    }
+
+    struct ComputeBasePtrOfStridedBatch
+    {
+        ComputeBasePtrOfStridedBatch(index_t BatchStrideA,
+                                     index_t BatchStrideB,
+                                     index_t BatchStrideB1,
+                                     index_t BatchStrideC)
+            : BatchStrideA_(BatchStrideA),
+              BatchStrideB_(BatchStrideB),
+              BatchStrideB1_(BatchStrideB1),
+              BatchStrideC_(BatchStrideC)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideA_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB1_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideC_);
+        }
+
+        private:
+        index_t BatchStrideA_;
+        index_t BatchStrideB_;
+        index_t BatchStrideB1_;
+        index_t BatchStrideC_;
+    };
+
+    using AGridDesc_AK0_M_AK1  = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
+    using BGridDesc_BK0_N_BK1  = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
+    using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1(1, 1, 1));
+    using CGridDesc_M_N        = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseBatchedGemmGemm_Xdl_CShuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        AccElementwiseOperation,
+        B1ElementwiseOperation,
+        CElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        B1GridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        Gemm1NPerBlock,
+        Gemm1KPerBlock,
+        AK1,
+        BK1,
+        B1K1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        Gemm1NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        true,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        true,
+        BBlockLdsExtraN,
+        B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+        B1BlockTransferThreadClusterArrangeOrder,
+        B1BlockTransferSrcAccessOrder,
+        B1BlockTransferSrcVectorDim,
+        B1BlockTransferSrcScalarPerVector,
+        B1BlockTransferDstScalarPerVector_BK1,
+        false,
+        B1BlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 const B1DataType* p_b1_grid,
+                 CDataType* p_c_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t Gemm1NRaw, // = ORaw
+                 index_t Batch,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideB1,
+                 index_t StrideC,
+                 index_t BatchStrideA,
+                 index_t BatchStrideB,
+                 index_t BatchStrideB1,
+                 index_t BatchStrideC,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 AccElementwiseOperation acc_element_op,
+                 B1ElementwiseOperation b1_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_b1_grid_{p_b1_grid},
+              p_c_grid_{p_c_grid},
+              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
+              b1_grid_desc_bk0_n_bk1_{
+                  DeviceOp::MakeB1GridDescriptor_BK0_N_BK1(NRaw, Gemm1NRaw, StrideB1)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, Gemm1NRaw, StrideC)},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              acc_element_op_{acc_element_op},
+              b1_element_op_{b1_element_op},
+              c_element_op_{c_element_op},
+              batch_count_(Batch),
+              compute_base_ptr_of_batch_{BatchStrideA, BatchStrideB, BatchStrideB1, BatchStrideC},
+              raw_lengths_m_n_k_o_{MRaw, NRaw, KRaw, Gemm1NRaw}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           b1_grid_desc_bk0_n_bk1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[AK0, M, AK1]: " << a_grid_desc_ak0_m_ak1_ << std::endl;
+            std::cout << "B0[BK0, N, BK1]: " << b_grid_desc_bk0_n_bk1_ << std::endl;
+            std::cout << "B1[BK0, N, BK1]: " << b1_grid_desc_bk0_n_bk1_ << std::endl;
+            std::cout << "C[M, N]: " << c_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        const B1DataType* p_b1_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        AccElementwiseOperation acc_element_op_;
+        B1ElementwiseOperation b1_element_op_;
+        CElementwiseOperation c_element_op_;
+        index_t batch_count_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
+
+        // For robust IsSupportedArgument() check
+        std::vector<index_t> raw_lengths_m_n_k_o_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!DeviceOp::IsSupportedArgument(arg))
+            {
+                throw std::runtime_error("wrong! unsupported argument");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.batch_count_;
+
+            // Gemm0_K
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel = kernel_gemm_gemm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    AccElementwiseOperation,
+                    B1ElementwiseOperation,
+                    CElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    DeviceOp::B1GridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    ComputeBasePtrOfStridedBatch,
+                    has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_b1_grid_,
+                                              arg.p_c_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.acc_element_op_,
+                                              arg.b1_element_op_,
+                                              arg.c_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.b1_grid_desc_bk0_n_bk1_,
+                                              arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_ctile_map_,
+                                              arg.batch_count_,
+                                              arg.compute_base_ptr_of_batch_);
+            };
+
+            // Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
+            // to concern Gemm0's loop
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        // Note: we need raw lengths since threadwise copy can not handle vector load when part of
+        // vector is out of bounds
+        const auto MRaw      = arg.raw_lengths_m_n_k_o_[0];
+        const auto NRaw      = arg.raw_lengths_m_n_k_o_[1];
+        const auto KRaw      = arg.raw_lengths_m_n_k_o_[2];
+        const auto Gemm1NRaw = arg.raw_lengths_m_n_k_o_[3];
+
+        // Check scalar per vector requirement
+        const auto a_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, ALayout> ? KRaw : MRaw;
+        const auto b_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, BLayout> ? NRaw : KRaw;
+        const auto b1_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, B1Layout> ? Gemm1NRaw : NRaw;
+        const auto c_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, CLayout> ? Gemm1NRaw : MRaw;
+
+        if(!(a_extent_lowest % ABlockTransferSrcScalarPerVector == 0 &&
+             b_extent_lowest % BBlockTransferSrcScalarPerVector == 0 &&
+             b1_extent_lowest % B1BlockTransferSrcScalarPerVector == 0 &&
+             c_extent_lowest % CShuffleBlockTransferScalarPerVector_NPerBlock == 0))
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.b1_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             const B1DataType* p_b1,
+                             CDataType* p_c,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t Gemm1NRaw,
+                             index_t Batch,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideB1,
+                             index_t StrideC,
+                             index_t BatchStrideA,
+                             index_t BatchStrideB,
+                             index_t BatchStrideB1,
+                             index_t BatchStrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             AccElementwiseOperation acc_element_op,
+                             B1ElementwiseOperation b1_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,           p_b,          p_b1,         p_c,          MRaw,
+                        NRaw,          KRaw,         Gemm1NRaw,    Batch,        StrideA,
+                        StrideB,       StrideB1,     StrideC,      BatchStrideA, BatchStrideB,
+                        BatchStrideB1, BatchStrideC, a_element_op, b_element_op, acc_element_op,
+                        b1_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      const void* p_b1,
+                                                      void* p_c,
+                                                      index_t MRaw,
+                                                      index_t NRaw,
+                                                      index_t KRaw,
+                                                      index_t Gemm1NRaw,
+                                                      index_t Batch,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideB1,
+                                                      index_t StrideC,
+                                                      index_t BatchStrideA,
+                                                      index_t BatchStrideB,
+                                                      index_t BatchStrideB1,
+                                                      index_t BatchStrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      AccElementwiseOperation acc_element_op,
+                                                      B1ElementwiseOperation b1_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<const B1DataType*>(p_b1),
+                                          static_cast<CDataType*>(p_c),
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          Gemm1NRaw,
+                                          Batch,
+                                          StrideA,
+                                          StrideB,
+                                          StrideB1,
+                                          StrideC,
+                                          BatchStrideA,
+                                          BatchStrideB,
+                                          BatchStrideB1,
+                                          BatchStrideC,
+                                          a_element_op,
+                                          b_element_op,
+                                          acc_element_op,
+                                          b1_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedGemmGemm_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerBlock << ", "
+            << Gemm1NPerBlock << ", "
+            << Gemm1KPerBlock << ", "
+            << B1K1 << ", "
+            << getGemmSpecializationString(GemmSpec) << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp
new file mode 100644
index 0000000000..545d7e576f
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp
@@ -0,0 +1,725 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for
+ * \link DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the
+ * computing of pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename ComputePtrOffsetOfBatch,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_xdl(const ABDataType* __restrict__ p_a_grid,
+                                const ABDataType* __restrict__ p_b_grid,
+                                DsPointer p_ds_grid,
+                                EDataType* __restrict__ p_e_grid,
+                                const index_t batch_count,
+                                const AElementwiseOperation a_element_op,
+                                const BElementwiseOperation b_element_op,
+                                const CDEElementwiseOperation cde_element_op,
+                                const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
+                                const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
+                                const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                                    ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                                    e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
+                                const Block2ETileMap block_2_etile_map)
+{
+
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_ds_grid_grp,
+                                                  p_e_grid + e_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_etile_map;
+#endif
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceBatchedGemmMultiD_Xdl : public DeviceBatchedGemmMultiD<ALayout,
+                                                                    BLayout,
+                                                                    DsLayout,
+                                                                    ELayout,
+                                                                    ADataType,
+                                                                    BDataType,
+                                                                    DsDataType,
+                                                                    EDataType,
+                                                                    AElementwiseOperation,
+                                                                    BElementwiseOperation,
+                                                                    CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceBatchedGemmMultiD_Xdl;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    template <typename ELay>
+    static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                                         const std::array<index_t, NumDTensor>& NRaws,
+                                         const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
+
+    struct ComputePtrOffsetOfStridedBatch
+    {
+        ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                       index_t BatchStrideB,
+                                       std::array<ck::index_t, NumDTensor> BatchStrideDs,
+                                       index_t BatchStrideE)
+            : BatchStrideA_(BatchStrideA),
+              BatchStrideB_(BatchStrideB),
+              BatchStrideDs_(BatchStrideDs),
+              BatchStrideE_(BatchStrideE)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideA_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB_);
+        }
+
+        __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+        {
+            std::array<long_index_t, NumDTensor> ds_offset;
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                ds_offset[i] = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]);
+            });
+            return ds_offset;
+        }
+
+        __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideE_);
+        }
+
+        private:
+        index_t BatchStrideA_;
+        index_t BatchStrideB_;
+        std::array<ck::index_t, NumDTensor> BatchStrideDs_;
+        index_t BatchStrideE_;
+    };
+
+    using ComputeDataType = ADataType;
+
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeDataType,
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // desc for blockwise copy
+    using AGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
+            AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
+            BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
+        decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+
+    // block-to-e-tile map
+    using Block2ETileMap =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t Batch,
+                 index_t StrideA,
+                 index_t StrideB,
+                 const std::array<ck::index_t, NumDTensor>& StrideDs,
+                 index_t StrideE,
+                 index_t BatchStrideA,
+                 index_t BatchStrideB,
+                 const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
+                 index_t BatchStrideE,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              Batch_(Batch),
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(MRaw, KRaw, StrideA)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(KRaw, NRaw, StrideB)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(MRaw, NRaw, StrideE)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              compute_ptr_offset_of_batch_{BatchStrideA, BatchStrideB, BatchStrideDs, BatchStrideE},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+            // populate pointer, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) =
+                    DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaw, NRaw, StrideDs[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           ds_grid_desc_m_n_,
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // Batch
+        index_t Batch_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // for calculating batch offset
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceBatchedGemmMultiD_Xdl::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.ds_grid_desc_m_n_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * arg.Batch_;
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel =
+                    kernel_batched_gemm_xdl<GridwiseGemm,
+                                            ADataType, // TODO: distiguish A/B datatype
+                                            typename GridwiseGemm::DsGridPointer,
+                                            EDataType,
+                                            AElementwiseOperation,
+                                            BElementwiseOperation,
+                                            CDEElementwiseOperation,
+                                            DeviceOp::AGridDesc_AK0_M_AK1,
+                                            DeviceOp::BGridDesc_BK0_N_BK1,
+                                            DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                                            DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                                            ComputePtrOffsetOfStridedBatch,
+                                            Block2ETileMap,
+                                            has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.Batch_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.compute_ptr_offset_of_batch_,
+                                              arg.block_2_etile_map_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           arg.ds_grid_desc_m_n_,
+                                           arg.e_grid_desc_m_n_,
+                                           arg.block_2_etile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             const std::array<const void*, NumDTensor>& p_ds,
+                             void* p_e,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t Batch,
+                             index_t StrideA,
+                             index_t StrideB,
+                             const std::array<index_t, NumDTensor>& StrideDs,
+                             index_t StrideE,
+                             index_t BatchStrideA,
+                             index_t BatchStrideB,
+                             const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
+                             index_t BatchStrideE,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        M,
+                        N,
+                        K,
+                        Batch,
+                        StrideA,
+                        StrideB,
+                        StrideDs,
+                        StrideE,
+                        BatchStrideA,
+                        BatchStrideB,
+                        BatchStrideDs,
+                        BatchStrideE,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const std::array<const void*, NumDTensor>& p_ds,
+                        void* p_e,
+                        index_t M,
+                        index_t N,
+                        index_t K,
+                        index_t Batch,
+                        index_t StrideA,
+                        index_t StrideB,
+                        const std::array<ck::index_t, NumDTensor>& StrideDs,
+                        index_t StrideE,
+                        index_t BatchStrideA,
+                        index_t BatchStrideB,
+                        const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
+                        index_t BatchStrideE,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          M,
+                                          N,
+                                          K,
+                                          Batch,
+                                          StrideA,
+                                          StrideB,
+                                          StrideDs,
+                                          StrideE,
+                                          BatchStrideA,
+                                          BatchStrideB,
+                                          BatchStrideDs,
+                                          BatchStrideE,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedGemmMultiD_Xdl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << getGemmSpecializationString(GemmSpec)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp
new file mode 100644
index 0000000000..b51c600476
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp
@@ -0,0 +1,795 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+/*
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for
+ * \link DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the
+ * computing of pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ */
+
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_K0_M0_M1_K1,
+          typename BGridDesc_K0_N0_N1_K1,
+          typename DsGridDesc_M0_M10_M11_N0_N10_N11,
+          typename CGridDesc_M0_M10_M11_N0_N10_N11,
+          typename ComputePtrOffsetOfBatch,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop,
+          bool HasDoubleTailKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_dl_multiple_d(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const index_t batch_count,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1,
+            const BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1,
+            const DsGridDesc_M0_M10_M11_N0_N10_N11 ds_grid_desc_m0_m10_m11_n0_n10_n11,
+            const CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||             \
+    defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx1030__) || defined(__gfx1100__) || \
+    defined(__gfx1101__) || defined(__gfx1102__))
+
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor = DsGridDesc_M0_M10_M11_N0_N10_N11::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseGemm::Run(p_a_grid + a_batch_offset,
+                      p_b_grid + b_batch_offset,
+                      p_ds_grid_grp,
+                      p_e_grid + e_batch_offset,
+                      p_shared,
+                      a_element_op,
+                      b_element_op,
+                      cde_element_op,
+                      a_grid_desc_k0_m0_m1_k1,
+                      b_grid_desc_k0_n0_n1_k1,
+                      ds_grid_desc_m0_m10_m11_n0_n10_n11,
+                      e_grid_desc_m0_m10_m11_n0_n10_n11,
+                      block_2_ctile_map,
+                      integral_constant<bool, HasMainKBlockLoop>{},
+                      integral_constant<bool, HasDoubleTailKBlockLoop>{});
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = a_grid_desc_k0_m0_m1_k1;
+    ignore = b_grid_desc_k0_n0_n1_k1;
+    ignore = ds_grid_desc_m0_m10_m11_n0_n10_n11;
+    ignore = e_grid_desc_m0_m10_m11_n0_n10_n11;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+
+#endif
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t K1,
+          index_t M1PerThread,
+          index_t N1PerThread,
+          index_t KPerThread,
+          typename M1N1ThreadClusterM1Xs,
+          typename M1N1ThreadClusterN1Xs,
+          typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+          typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+          typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+          typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+          typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          enable_if_t<
+              is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+                  is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+              bool> = false>
+struct DeviceBatchedGemmMultipleD_Dl : public DeviceBatchedGemmMultiD<ALayout,
+                                                                      BLayout,
+                                                                      DsLayout,
+                                                                      ELayout,
+                                                                      ADataType,
+                                                                      BDataType,
+                                                                      DsDataType,
+                                                                      EDataType,
+                                                                      AElementwiseOperation,
+                                                                      BElementwiseOperation,
+                                                                      CDEElementwiseOperation>
+
+{
+    using DeviceOp                      = DeviceBatchedGemmMultipleD_Dl;
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static constexpr auto K1Number = Number<K1>{};
+
+    static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
+    {
+        const index_t K0 = K / K1;
+
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(M, PadM)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
+    {
+        const index_t K0 = K / K1;
+
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    template <typename ELay>
+    static auto MakeEGridDescriptor_M_N(index_t M, index_t N, index_t StrideE)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideE));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                                         const std::array<index_t, NumDTensor>& NRaws,
+                                         const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
+    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
+    using DsGridDesc_M_N    = decltype(MakeDsGridDescriptor_M_N({}, {}, {}));
+    using EGridDesc_M_N     = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
+
+    struct ComputePtrOffsetOfStridedBatch
+    {
+        ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                       index_t BatchStrideB,
+                                       std::array<ck::index_t, NumDTensor> BatchStrideDs,
+                                       index_t BatchStrideE)
+            : BatchStrideA_(BatchStrideA),
+              BatchStrideB_(BatchStrideB),
+              BatchStrideDs_(BatchStrideDs),
+              BatchStrideE_(BatchStrideE)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideA_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB_);
+        }
+
+        __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+        {
+            std::array<long_index_t, NumDTensor> ds_offset;
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                ds_offset[i] = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]);
+            });
+            return ds_offset;
+        }
+
+        __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideE_);
+        }
+
+        private:
+        index_t BatchStrideA_;
+        index_t BatchStrideB_;
+        std::array<ck::index_t, NumDTensor> BatchStrideDs_;
+        index_t BatchStrideE_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm =
+        GridwiseGemmDlMultipleD_km_kn_mn<BlockSize,
+                                         ADataType,
+                                         AccDataType,
+                                         DsDataType,
+                                         EDataType,
+                                         AElementwiseOperation,
+                                         BElementwiseOperation,
+                                         CDEElementwiseOperation,
+                                         InMemoryDataOperationEnum::Set,
+                                         AGridDesc_K0_M_K1,
+                                         BGridDesc_K0_N_K1,
+                                         EGridDesc_M_N,
+                                         MPerBlock,
+                                         NPerBlock,
+                                         K0PerBlock,
+                                         K1,
+                                         M1PerThread,
+                                         N1PerThread,
+                                         KPerThread,
+                                         M1N1ThreadClusterM1Xs,
+                                         M1N1ThreadClusterN1Xs,
+                                         ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+                                         ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+                                         ABlockTransferThreadClusterArrangeOrder,
+                                         ABlockTransferSrcAccessOrder,
+                                         ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+                                         ABlockTransferSrcVectorTensorContiguousDimOrder,
+                                         ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+                                         BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+                                         BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+                                         BBlockTransferThreadClusterArrangeOrder,
+                                         BBlockTransferSrcAccessOrder,
+                                         BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+                                         BBlockTransferSrcVectorTensorContiguousDimOrder,
+                                         BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+                                         CThreadTransferSrcDstAccessOrder,
+                                         CThreadTransferSrcDstVectorDim,
+                                         CThreadTransferDstScalarPerVector>;
+
+    using AGridDesc_K0_M0_M1_K1 =
+        decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_K0_M_K1{}));
+    using BGridDesc_K0_N0_N1_K1 =
+        decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{}));
+    using DsGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(DsGridDesc_M_N{}));
+    using EGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(EGridDesc_M_N{}));
+    using DefaultBlock2CTileMap =
+        decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(EGridDesc_M_N{}));
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 index_t M,
+                 index_t N,
+                 index_t K,
+                 index_t Batch,
+                 index_t StrideA,
+                 index_t StrideB,
+                 std::array<index_t, NumDTensor> StrideDs,
+                 index_t StrideE,
+                 index_t BatchStrideA,
+                 index_t BatchStrideB,
+                 const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
+                 index_t BatchStrideE,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              K_(K),
+              Batch_(Batch),
+              a_grid_desc_k0_m0_m1_k1_{},
+              b_grid_desc_k0_n0_n1_k1_{},
+              e_grid_desc_m0_m10_m11_n0_n10_n11_{},
+              compute_ptr_offset_of_batch_{BatchStrideA, BatchStrideB, BatchStrideDs, BatchStrideE},
+              block_2_ctile_map_{},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+            a_grid_desc_k0_m_k1_ =
+                DeviceBatchedGemmMultipleD_Dl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
+            b_grid_desc_k0_n_k1_ =
+                DeviceBatchedGemmMultipleD_Dl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) =
+                    DeviceOp::MakeEGridDescriptor_M_N<DLayout>(M, N, StrideDs[i]);
+            });
+            e_grid_desc_m_n_ =
+                DeviceBatchedGemmMultipleD_Dl::MakeEGridDescriptor_M_N<ELayout>(M, N, StrideE);
+
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, e_grid_desc_m_n_))
+            {
+                a_grid_desc_k0_m0_m1_k1_ =
+                    GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(a_grid_desc_k0_m_k1_);
+                b_grid_desc_k0_n0_n1_k1_ =
+                    GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(b_grid_desc_k0_n_k1_);
+
+                ds_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(ds_grid_desc_m_n_);
+
+                e_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(e_grid_desc_m_n_);
+
+                block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        index_t K_;
+
+        // Batch
+        index_t Batch_;
+
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1_;
+        BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1_;
+        DsGridDesc_M0_M10_M11_N0_N10_N11 ds_grid_desc_m0_m10_m11_n0_n10_n11_;
+        EGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11_;
+
+        // for calculating batch offset
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+
+        DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // TODO: unused since gridwise_gemm_dl_v1r3 does NOT support prologue for the time being.
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceBatchedGemmMultipleD_Dl::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            {
+                std::cout << "arg.a_grid_desc_k0_m0_m1_k1_{"
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I0) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n0_n1_k1_{"
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I0) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.e_grid_desc_m_n_{ " << arg.e_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.e_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+
+            if(!GridwiseGemm::CheckValidity(
+                   arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.e_grid_desc_m_n_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemmDlMultipleD_km_kn_mn has invalid setting");
+            }
+
+            const index_t grid_size =
+                GridwiseGemm::CalculateGridSize(arg.e_grid_desc_m_n_.GetLength(I0),
+                                                arg.e_grid_desc_m_n_.GetLength(I1)) *
+                arg.Batch_;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop,
+                                     auto has_double_tail_k_block_loop) {
+                constexpr bool has_main_loop   = has_main_k_block_loop.value;
+                constexpr bool has_double_loop = has_double_tail_k_block_loop.value;
+
+                const auto kernel =
+                    kernel_gemm_dl_multiple_d<GridwiseGemm,
+                                              ADataType,
+                                              typename GridwiseGemm::DsGridPointer,
+                                              EDataType,
+                                              AElementwiseOperation,
+                                              BElementwiseOperation,
+                                              CDEElementwiseOperation,
+                                              DeviceOp::AGridDesc_K0_M0_M1_K1,
+                                              DeviceOp::BGridDesc_K0_N0_N1_K1,
+                                              DeviceOp::DsGridDesc_M0_M10_M11_N0_N10_N11,
+                                              DeviceOp::EGridDesc_M0_M10_M11_N0_N10_N11,
+                                              ComputePtrOffsetOfStridedBatch,
+                                              DefaultBlock2CTileMap,
+                                              has_main_loop,
+                                              has_double_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.Batch_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_k0_m0_m1_k1_,
+                                              arg.b_grid_desc_k0_n0_n1_k1_,
+                                              arg.ds_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                              arg.e_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                              arg.compute_ptr_offset_of_batch_,
+                                              arg.block_2_ctile_map_);
+            };
+
+            const auto K0                    = arg.a_grid_desc_k0_m0_m1_k1_.GetLength(I0);
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
+            const bool has_double_tail_k_block_loop =
+                GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
+
+            if(has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, true>{});
+            }
+            else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, false>{});
+            }
+            else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        // TODO: Enable for gfx90a after complier fix
+        if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx90a" ||
+           ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx1030" ||
+           ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx1100" ||
+           ck::get_device_name() == "gfx1101" || ck::get_device_name() == "gfx1102")
+        {
+            bool pass = true;
+            pass      = pass && arg.K_ % K1 == 0;
+
+            pass = pass && GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                                       arg.b_grid_desc_k0_n_k1_,
+                                                       arg.e_grid_desc_m_n_);
+
+            return pass;
+        }
+        else
+        {
+            return false;
+        }
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             std::array<const void*, NumDTensor> p_ds,
+                             void* p_e,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t Batch,
+                             index_t StrideA,
+                             index_t StrideB,
+                             std::array<ck::index_t, NumDTensor> StrideDs,
+                             index_t StrideE,
+                             index_t BatchStrideA,
+                             index_t BatchStrideB,
+                             const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
+                             index_t BatchStrideE,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        M,
+                        N,
+                        K,
+                        Batch,
+                        StrideA,
+                        StrideB,
+                        StrideDs,
+                        StrideE,
+                        BatchStrideA,
+                        BatchStrideB,
+                        BatchStrideDs,
+                        BatchStrideE,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const std::array<const void*, NumDTensor>& p_ds,
+                        void* p_e,
+                        index_t M,
+                        index_t N,
+                        index_t K,
+                        index_t Batch,
+                        index_t StrideA,
+                        index_t StrideB,
+                        const std::array<ck::index_t, NumDTensor>& StrideDs,
+                        index_t StrideE,
+                        index_t BatchStrideA,
+                        index_t BatchStrideB,
+                        const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
+                        index_t BatchStrideE,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          M,
+                                          N,
+                                          K,
+                                          Batch,
+                                          StrideA,
+                                          StrideB,
+                                          StrideDs,
+                                          StrideE,
+                                          BatchStrideA,
+                                          BatchStrideB,
+                                          BatchStrideDs,
+                                          BatchStrideE,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedGemmMultipleD_Dl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << M1PerThread << ", "
+            << N1PerThread << ", "
+            << KPerThread
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..9455ec48b3
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,995 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename A0B0B1DataType,
+          typename D0sPointer,
+          typename D1sPointer,
+          typename E1DataType,
+          typename A0ElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename CDE0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CDE1ElementwiseOperation,
+          typename A0GridDesc_AK0_M_AK1,
+          typename B0GridDesc_BK0_N_BK1,
+          typename D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5,
+          typename B1GridDesc_BK0_N_BK1,
+          typename D1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename E1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2E1TileMap,
+          typename ComputeBasePtrOfStridedBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_gemm_xdl_cshuffle_v1(
+            const A0B0B1DataType* __restrict__ p_a0_grid,
+            const A0B0B1DataType* __restrict__ p_b0_grid,
+            D0sPointer p_d0s_grid,
+            const A0B0B1DataType* __restrict__ p_b1_grid,
+            D1sPointer p_d1s_grid,
+            E1DataType* __restrict__ p_e1_grid,
+            const A0ElementwiseOperation a0_element_op,
+            const B0ElementwiseOperation b0_element_op,
+            const CDE0ElementwiseOperation cde0_element_op,
+            const B1ElementwiseOperation b1_element_op,
+            const CDE1ElementwiseOperation cde1_element_op,
+            const A0GridDesc_AK0_M_AK1 a0_grid_desc_ak0_m_ak1,
+            const B0GridDesc_BK0_N_BK1 b0_grid_desc_bk0_n_bk1,
+            const D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
+                d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+            const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
+            const D1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                d1s_grid_desc_mblock_mperblock_nblock_nperblock,
+            const E1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e1_grid_desc_mblock_mperblock_nblock_nperblock,
+            const Block2E1TileMap block_2_e1tile_map,
+            const index_t batch_count,
+            const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetABasePtr(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetBBasePtr(g_idx)));
+    const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetB1BasePtr(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetCBasePtr(g_idx)));
+
+    static_for<0, p_d0s_grid.Size(), 1>{}([&](auto In) {
+        const long_index_t d0_batch_offset = __builtin_amdgcn_readfirstlane(
+            static_cast<long_index_t>(compute_base_ptr_of_batch.GetD0BasePtr(g_idx, In)));
+        p_d0s_grid(In) = p_d0s_grid(In) + d0_batch_offset;
+    });
+
+    static_for<0, p_d1s_grid.Size(), 1>{}([&](auto In) {
+        const long_index_t d1_batch_offset = __builtin_amdgcn_readfirstlane(
+            static_cast<long_index_t>(compute_base_ptr_of_batch.GetD1BasePtr(g_idx, In)));
+        p_d1s_grid(In) = p_d1s_grid(In) + d1_batch_offset;
+    });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a0_grid + a_batch_offset,
+                                                  p_b0_grid + b_batch_offset,
+                                                  p_d0s_grid,
+                                                  p_b1_grid + b1_batch_offset,
+                                                  p_d1s_grid,
+                                                  p_e1_grid + c_batch_offset,
+                                                  p_shared,
+                                                  a0_element_op,
+                                                  b0_element_op,
+                                                  cde0_element_op,
+                                                  b1_element_op,
+                                                  cde1_element_op,
+                                                  a0_grid_desc_ak0_m_ak1,
+                                                  b0_grid_desc_bk0_n_bk1,
+                                                  d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                                                  b1_grid_desc_bk0_n_bk1,
+                                                  d1s_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e1_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_e1tile_map);
+#else
+    ignore = p_a0_grid;
+    ignore = p_b0_grid;
+    ignore = p_d0s_grid;
+    ignore = p_b1_grid;
+    ignore = p_d1s_grid;
+    ignore = p_e1_grid;
+    ignore = a0_element_op;
+    ignore = b0_element_op;
+    ignore = cde0_element_op;
+    ignore = b1_element_op;
+    ignore = cde1_element_op;
+    ignore = a0_grid_desc_ak0_m_ak1;
+    ignore = b0_grid_desc_bk0_n_bk1;
+    ignore = d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5;
+    ignore = b1_grid_desc_bk0_n_bk1;
+    ignore = d1s_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e1_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_e1tile_map;
+    ignore = batch_count;
+    ignore = compute_base_ptr_of_batch;
+#endif
+}
+
+// Computes C = A * B0 * B1
+//              ^^^^^^ (Acc0)
+//              ^^^^^^^^^^^ (Acc1)
+template <typename A0Layout,
+          typename B0Layout, // B0Layout
+          typename D0sLayout,
+          typename B1Layout,
+          typename D1sLayout,
+          typename E1Layout,
+          typename A0DataType,
+          typename B0DataType,
+          typename Acc0DataType,
+          typename D0sDataType,
+          typename B1DataType,
+          typename Acc1DataType,
+          typename C1ShuffleDataType,
+          typename D1sDataType,
+          typename E1DataType,
+          typename A0ElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename CDE0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CDE1ElementwiseOperation,
+          bool PadGemm0M,
+          bool PadGemm0N,
+          bool PadGemm0K,
+          bool PadGemm1N,
+          bool PadGemm1K,
+          index_t NumGemm0KPrefetchStage,
+          index_t BlockSize,
+          index_t Gemm0MPerBlock,
+          index_t Gemm0NPerBlock,
+          index_t Gemm0KPerBlock,
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t A0K1,
+          index_t B0K1,
+          index_t B1K1,
+          index_t Gemm0MPerXdl,
+          index_t Gemm0NPerXdl,
+          index_t Gemm0MXdlPerWave,
+          index_t Gemm0NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename A0BlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename A0BlockTransferThreadClusterArrangeOrder,
+          typename A0BlockTransferSrcAccessOrder,
+          index_t A0BlockTransferSrcVectorDim,
+          index_t A0BlockTransferSrcScalarPerVector,
+          index_t A0BlockTransferDstScalarPerVector_AK1,
+          bool A0BlockLdsExtraM,
+          typename B0BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B0BlockTransferThreadClusterArrangeOrder,
+          typename B0BlockTransferSrcAccessOrder,
+          index_t B0BlockTransferSrcVectorDim,
+          index_t B0BlockTransferSrcScalarPerVector,
+          index_t B0BlockTransferDstScalarPerVector_BK1,
+          bool B0BlockLdsExtraN,
+          index_t CDE0BlockTransferSrcVectorDim,
+          index_t CDE0BlockTransferSrcScalaerPerVector,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1BlockLdsExtraN,
+          index_t C1ShuffleMXdlPerWavePerShuffle,
+          index_t C1ShuffleGemm0NXdlPerWavePerShuffle,
+          typename CDE1ShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDE1ShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = LoopScheduler::Default>
+struct DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
+    : public DeviceBatchedGemmMultipleDGemmMultipleD<A0Layout,
+                                                     B0Layout,
+                                                     D0sLayout,
+                                                     B1Layout,
+                                                     D1sLayout,
+                                                     E1Layout,
+                                                     A0DataType,
+                                                     B0DataType,
+                                                     D0sDataType,
+                                                     B1DataType,
+                                                     D1sDataType,
+                                                     E1DataType,
+                                                     A0ElementwiseOperation,
+                                                     B0ElementwiseOperation,
+                                                     CDE0ElementwiseOperation,
+                                                     B1ElementwiseOperation,
+                                                     CDE1ElementwiseOperation>
+{
+    using DeviceOp = DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumD0Tensor = D0sDataType::Size();
+    static constexpr index_t NumD1Tensor = D1sDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+    static constexpr auto I8 = Number<8>{};
+    static constexpr auto I9 = Number<9>{};
+
+    static constexpr auto gemm0_padder =
+        GemmPadder_v2<PadGemm0M, PadGemm0N, PadGemm0K, index_t, index_t, index_t>{
+            Gemm0MPerBlock, Gemm0NPerBlock, Gemm0KPerBlock};
+
+    static constexpr auto gemm1_padder =
+        GemmPadder_v2<PadGemm0M, PadGemm1N, PadGemm1K, index_t, index_t, index_t>{
+            Gemm0MPerBlock, Gemm1NPerBlock, Gemm1KPerBlock};
+
+    // for Gemm0
+    static auto MakeA0GridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA0)
+    {
+        const auto a0_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, A0Layout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA0, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, A0Layout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA0));
+            }
+        }();
+
+        return gemm0_padder.PadADescriptor_M_K(a0_grid_desc_mraw_kraw);
+    }
+
+    // for Gemm0
+    static auto MakeB0GridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b0_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, B0Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, B0Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return gemm0_padder.PadBDescriptor_N_K(b0_grid_desc_nraw_kraw);
+    }
+
+    // for Gemm0
+    template <typename DLay>
+    static auto MakeD0GridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideD0)
+    {
+        const auto d0_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, DLay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideD0, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, DLay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideD0));
+            }
+        }();
+
+        return gemm0_padder.PadCDescriptor_M_N(d0_grid_desc_mraw_nraw);
+    }
+
+    // for Gemm1
+    static auto MakeB1GridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b1_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, B1Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, B1Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return gemm1_padder.PadBDescriptor_N_K(b1_grid_desc_nraw_kraw);
+    }
+
+    // for Gemm1
+    template <typename ELay>
+    static auto MakeE1GridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE1)
+    {
+        const auto e1_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE1, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE1));
+            }
+        }();
+
+        return gemm1_padder.PadCDescriptor_M_N(e1_grid_desc_mraw_nraw);
+    }
+
+    static auto MakeD0sGridDescriptor_M_N(const std::array<index_t, NumD1Tensor>& MRaws,
+                                          const std::array<index_t, NumD1Tensor>& NRaws,
+                                          const std::array<index_t, NumD1Tensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, D0sLayout>>;
+
+                return DeviceOp::MakeD0GridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumD0Tensor>{});
+    }
+
+    static auto MakeD1sGridDescriptor_M_N(const std::array<index_t, NumD1Tensor>& MRaws,
+                                          const std::array<index_t, NumD1Tensor>& NRaws,
+                                          const std::array<index_t, NumD1Tensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, D1sLayout>>;
+
+                return DeviceOp::MakeE1GridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumD1Tensor>{});
+    }
+
+    struct ComputeBasePtrOfStridedBatch
+    {
+        ComputeBasePtrOfStridedBatch(index_t BatchStrideA0,
+                                     index_t BatchStrideB0,
+                                     std::array<index_t, NumD0Tensor> BatchStrideD0s,
+                                     index_t BatchStrideB1,
+                                     std::array<index_t, NumD1Tensor> BatchStrideD1s,
+                                     index_t BatchStrideE1)
+            : BatchStrideA0_(BatchStrideA0),
+              BatchStrideB0_(BatchStrideB0),
+              BatchStrideD0s_(BatchStrideD0s),
+              BatchStrideB1_(BatchStrideB1),
+              BatchStrideD1s_(BatchStrideD1s),
+              BatchStrideE1_(BatchStrideE1)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideA0_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB0_);
+        }
+
+        template <index_t I>
+        __host__ __device__ constexpr long_index_t GetD0BasePtr(index_t g_idx,
+                                                                Number<I> d1_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideD0s_[d1_idx]);
+        }
+
+        __host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB1_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideE1_);
+        }
+
+        template <index_t I>
+        __host__ __device__ constexpr auto GetD1BasePtr(index_t g_idx, Number<I> d1_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideD1s_[d1_idx]);
+        }
+
+        private:
+        index_t BatchStrideA0_;
+        index_t BatchStrideB0_;
+        std::array<index_t, NumD0Tensor> BatchStrideD0s_;
+        index_t BatchStrideB1_;
+        std::array<index_t, NumD1Tensor> BatchStrideD1s_;
+        index_t BatchStrideE1_;
+    };
+
+    using A0GridDesc_M_K  = decltype(MakeA0GridDescriptor_M_K(1, 1, 1));
+    using B0GridDesc_N_K  = decltype(MakeB0GridDescriptor_N_K(1, 1, 1));
+    using D0sGridDesc_M_N = remove_cvref_t<decltype(MakeD0sGridDescriptor_M_N({}, {}, {}))>;
+    using B1GridDesc_N_K  = decltype(MakeB1GridDescriptor_N_K(1, 1, 1));
+    using D1sGridDesc_M_N = remove_cvref_t<decltype(MakeD1sGridDescriptor_M_N({}, {}, {}))>;
+    using E1GridDesc_M_N  = decltype(MakeE1GridDescriptor_M_N<E1Layout>(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<
+        A0DataType, // TODO: distinguish A/B datatype
+        Acc0DataType,
+        D0sDataType,
+        Acc1DataType,
+        C1ShuffleDataType,
+        D1sDataType,
+        E1DataType,
+        A0ElementwiseOperation,
+        B0ElementwiseOperation,
+        CDE0ElementwiseOperation,
+        B1ElementwiseOperation,
+        CDE1ElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        A0GridDesc_M_K,
+        B0GridDesc_N_K,
+        D0sGridDesc_M_N,
+        B1GridDesc_N_K,
+        D1sGridDesc_M_N,
+        E1GridDesc_M_N,
+        NumGemm0KPrefetchStage,
+        BlockSize,
+        Gemm0MPerBlock,
+        Gemm0NPerBlock,
+        Gemm0KPerBlock,
+        Gemm1NPerBlock,
+        Gemm1KPerBlock,
+        A0K1,
+        B0K1,
+        B1K1,
+        Gemm0MPerXdl,
+        Gemm0NPerXdl,
+        Gemm0MXdlPerWave,
+        Gemm0NXdlPerWave,
+        Gemm1NXdlPerWave,
+        A0BlockTransferThreadClusterLengths_AK0_M_AK1,
+        A0BlockTransferThreadClusterArrangeOrder,
+        A0BlockTransferSrcAccessOrder,
+        A0BlockTransferSrcVectorDim,
+        A0BlockTransferSrcScalarPerVector,
+        A0BlockTransferDstScalarPerVector_AK1,
+        true,
+        A0BlockLdsExtraM,
+        B0BlockTransferThreadClusterLengths_BK0_N_BK1,
+        B0BlockTransferThreadClusterArrangeOrder,
+        B0BlockTransferSrcAccessOrder,
+        B0BlockTransferSrcVectorDim,
+        B0BlockTransferSrcScalarPerVector,
+        B0BlockTransferDstScalarPerVector_BK1,
+        true,
+        B0BlockLdsExtraN,
+        CDE0BlockTransferSrcVectorDim,
+        CDE0BlockTransferSrcScalaerPerVector,
+        B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+        B1BlockTransferThreadClusterArrangeOrder,
+        B1BlockTransferSrcAccessOrder,
+        B1BlockTransferSrcVectorDim,
+        B1BlockTransferSrcScalarPerVector,
+        B1BlockTransferDstScalarPerVector_BK1,
+        false,
+        B1BlockLdsExtraN,
+        C1ShuffleMXdlPerWavePerShuffle,
+        C1ShuffleGemm0NXdlPerWavePerShuffle,
+        CDE1ShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDE1ShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    using A0GridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultA0GridDescriptor_AK0_M_AK1(
+            A0GridDesc_M_K{}))>;
+    using B0GridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultB0GridDescriptor_BK0_N_BK1(
+            B0GridDesc_N_K{}))>;
+    using B1GridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultB1GridDescriptor_BK0_N_BK1(
+            B1GridDesc_N_K{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const A0DataType* p_a0_grid,
+                 const B0DataType* p_b0_grid,
+                 std::array<const void*, NumD0Tensor> p_d0s_grid,
+                 const B1DataType* p_b1_grid,
+                 std::array<const void*, NumD1Tensor> p_d1s_grid,
+                 E1DataType* p_e1_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t Gemm1NRaw, // = ORaw
+                 index_t Batch,
+                 index_t StrideA0,
+                 index_t StrideB0,
+                 std::array<index_t, NumD0Tensor> StrideD0s,
+                 index_t StrideB1,
+                 std::array<index_t, NumD1Tensor> StrideD1s,
+                 index_t StrideE1,
+                 index_t BatchStrideA0,
+                 index_t BatchStrideB0,
+                 std::array<index_t, NumD0Tensor> BatchStrideD0s,
+                 index_t BatchStrideB1,
+                 std::array<index_t, NumD1Tensor> BatchStrideD1s,
+                 index_t BatchStrideE1,
+                 A0ElementwiseOperation a0_element_op,
+                 B0ElementwiseOperation b0_element_op,
+                 CDE0ElementwiseOperation cde0_element_op,
+                 B1ElementwiseOperation b1_element_op,
+                 CDE1ElementwiseOperation cde1_element_op)
+            : p_a0_grid_{p_a0_grid},
+              p_b0_grid_{p_b0_grid},
+              p_d0s_grid_{},
+              p_b1_grid_{p_b1_grid},
+              p_d1s_grid_{},
+              p_e1_grid_{p_e1_grid},
+              a0_grid_desc_m_k_{DeviceOp::MakeA0GridDescriptor_M_K(MRaw, KRaw, StrideA0)},
+              b0_grid_desc_n_k_{DeviceOp::MakeB0GridDescriptor_N_K(KRaw, NRaw, StrideB0)},
+              d0s_grid_desc_m_n_{},
+              b1_grid_desc_n_k_{DeviceOp::MakeB1GridDescriptor_N_K(NRaw, Gemm1NRaw, StrideB1)},
+              d1s_grid_desc_m_n_{},
+              e1_grid_desc_m_n_{
+                  DeviceOp::MakeE1GridDescriptor_M_N<E1Layout>(MRaw, Gemm1NRaw, StrideE1)},
+              a0_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultA0GridDescriptor_AK0_M_AK1(a0_grid_desc_m_k_)},
+              b0_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultB0GridDescriptor_BK0_N_BK1(b0_grid_desc_n_k_)},
+              d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_{},
+              b1_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultB1GridDescriptor_BK0_N_BK1(b1_grid_desc_n_k_)},
+              d1s_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e1_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_e1tile_map_{GridwiseGemm::MakeDefaultBlock2E1TileMap(e1_grid_desc_m_n_)},
+              a0_element_op_{a0_element_op},
+              b0_element_op_{b0_element_op},
+              cde0_element_op_{cde0_element_op},
+              b1_element_op_{b1_element_op},
+              cde1_element_op_{cde1_element_op},
+              batch_count_(Batch),
+              compute_base_ptr_of_batch_{BatchStrideA0,
+                                         BatchStrideB0,
+                                         BatchStrideD0s,
+                                         BatchStrideB1,
+                                         BatchStrideD1s,
+                                         BatchStrideE1}
+        {
+#if DEBUG_LOG
+            std::cout << "a0_grid_desc_m_k_{" << a0_grid_desc_m_k_.GetLength(I0) << ", "
+                      << a0_grid_desc_m_k_.GetLength(I1) << "}" << std::endl;
+            std::cout << "b0_grid_desc_n_k_{" << b0_grid_desc_n_k_.GetLength(I0) << ", "
+                      << b0_grid_desc_n_k_.GetLength(I1) << "}" << std::endl;
+            std::cout << "d0s_grid_desc_m_n_[I0]{" << d0s_grid_desc_m_n_[I0].GetLength(I0) << ", "
+                      << d0s_grid_desc_m_n_[I0].GetLength(I1) << "}" << std::endl;
+            std::cout << "b1_grid_desc_n_k_{" << b1_grid_desc_n_k_.GetLength(I0) << ", "
+                      << b1_grid_desc_n_k_.GetLength(I1) << "}" << std::endl;
+            std::cout << "d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_{"
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I0) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I1) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I2) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I3) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I4) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I5) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I6) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I7) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I8) << ", "
+                      << d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_[I0].GetLength(I9) << "}"
+                      << std::endl;
+            std::cout << "e1_grid_desc_m_n_{" << e1_grid_desc_m_n_.GetLength(I0) << ", "
+                      << e1_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+#endif
+
+            static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                using D0Layout   = remove_cvref_t<tuple_element_t<i.value, D0sLayout>>;
+                using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
+
+                // D0 pointer
+                p_d0s_grid_(i) = static_cast<const D0DataType*>(p_d0s_grid[i]);
+
+                // D0 desc
+                d0s_grid_desc_m_n_(i) =
+                    DeviceOp::MakeD0GridDescriptor_M_N<D0Layout>(MRaw, NRaw, StrideD0s[i]);
+            });
+
+            static_for<0, NumD1Tensor, 1>{}([&](auto i) {
+                using D1Layout   = remove_cvref_t<tuple_element_t<i.value, D1sLayout>>;
+                using D1DataType = remove_cvref_t<tuple_element_t<i.value, D1sDataType>>;
+
+                // D1 pointer
+                p_d1s_grid_(i) = static_cast<const D1DataType*>(p_d1s_grid[i]);
+
+                // D1 desc
+                d1s_grid_desc_m_n_(i) =
+                    DeviceOp::MakeE1GridDescriptor_M_N<D1Layout>(MRaw, Gemm1NRaw, StrideD1s[i]);
+            });
+
+            if(GridwiseGemm::CheckValidity(a0_grid_desc_m_k_,
+                                           b0_grid_desc_n_k_,
+                                           b1_grid_desc_n_k_,
+                                           e1_grid_desc_m_n_,
+                                           block_2_e1tile_map_))
+            {
+                e1_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeE1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e1_grid_desc_m_n_);
+
+                d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_ =
+                    GridwiseGemm::MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(
+                        d0s_grid_desc_m_n_);
+
+                d1s_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeD1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        d1s_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        // pointers
+        const A0DataType* p_a0_grid_;
+        const B0DataType* p_b0_grid_;
+        typename GridwiseGemm::D0sGridPointer p_d0s_grid_;
+        const B1DataType* p_b1_grid_;
+        typename GridwiseGemm::D1sGridPointer p_d1s_grid_;
+        E1DataType* p_e1_grid_;
+
+        // tensor descriptors for problem definiton
+        A0GridDesc_M_K a0_grid_desc_m_k_;
+        B0GridDesc_N_K b0_grid_desc_n_k_;
+        D0sGridDesc_M_N d0s_grid_desc_m_n_;
+        B1GridDesc_N_K b1_grid_desc_n_k_;
+        D1sGridDesc_M_N d1s_grid_desc_m_n_;
+        E1GridDesc_M_N e1_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        A0GridDesc_AK0_M_AK1 a0_grid_desc_ak0_m_ak1_;
+        B0GridDesc_BK0_N_BK1 b0_grid_desc_bk0_n_bk1_;
+        typename GridwiseGemm::D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
+            d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_;
+        B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
+        typename GridwiseGemm::D1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            d1s_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::E1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e1_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e1-tile map
+        typename GridwiseGemm::DefaultBlock2E1TileMap block_2_e1tile_map_;
+
+        // element-wise op
+        A0ElementwiseOperation a0_element_op_;
+        B0ElementwiseOperation b0_element_op_;
+        CDE0ElementwiseOperation cde0_element_op_;
+        B1ElementwiseOperation b1_element_op_;
+        CDE1ElementwiseOperation cde1_element_op_;
+
+        // batch
+        index_t batch_count_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a0_grid_desc_m_k_,
+                                            arg.b0_grid_desc_n_k_,
+                                            arg.b1_grid_desc_n_k_,
+                                            arg.e1_grid_desc_m_n_,
+                                            arg.block_2_e1tile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_e1tile_map_.CalculateGridSize(arg.e1_grid_desc_m_n_) * arg.batch_count_;
+
+            // Gemm0_K
+            const auto K = arg.a0_grid_desc_m_k_.GetLength(I1);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel = kernel_batched_gemm_gemm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    A0DataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::D0sGridPointer,
+                    typename GridwiseGemm::D1sGridPointer,
+                    E1DataType,
+                    A0ElementwiseOperation,
+                    B0ElementwiseOperation,
+                    CDE0ElementwiseOperation,
+                    B1ElementwiseOperation,
+                    CDE1ElementwiseOperation,
+                    DeviceOp::A0GridDesc_AK0_M_AK1,
+                    DeviceOp::B0GridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5,
+                    DeviceOp::B1GridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::D1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::E1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::DefaultBlock2E1TileMap,
+                    ComputeBasePtrOfStridedBatch,
+                    has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a0_grid_,
+                                              arg.p_b0_grid_,
+                                              arg.p_d0s_grid_,
+                                              arg.p_b1_grid_,
+                                              arg.p_d1s_grid_,
+                                              arg.p_e1_grid_,
+                                              arg.a0_element_op_,
+                                              arg.b0_element_op_,
+                                              arg.cde0_element_op_,
+                                              arg.b1_element_op_,
+                                              arg.cde1_element_op_,
+                                              arg.a0_grid_desc_ak0_m_ak1_,
+                                              arg.b0_grid_desc_bk0_n_bk1_,
+                                              arg.d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_,
+                                              arg.b1_grid_desc_bk0_n_bk1_,
+                                              arg.d1s_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e1_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_e1tile_map_,
+                                              arg.batch_count_,
+                                              arg.compute_base_ptr_of_batch_);
+            };
+
+            // Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
+            // to concern Gemm0's loop
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    // check if DsLayout is supported
+    template <typename RefLayout, typename DsLayout, const index_t NumDTensor>
+    static bool CheckDLayout()
+    {
+        static bool valid = true;
+        // iterate over DLayout tuple
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+            // if RefLayout and DLayout are same, keep valid true, otherwise false
+            valid = valid && is_same_v<RefLayout, DLayout>;
+        });
+        return valid;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        // Check supported layouts
+        // A0 - Row
+        // B0 - Col
+        // D0s - Rows
+        // B1 - Row or Col
+        // D1s - Rows
+        // E1 - Row
+        if(!(is_same_v<tensor_layout::gemm::RowMajor, A0Layout> &&
+             is_same_v<tensor_layout::gemm::ColumnMajor, B0Layout> &&
+             CheckDLayout<tensor_layout::gemm::RowMajor, D0sLayout, NumD0Tensor>() &&
+             (is_same_v<tensor_layout::gemm::RowMajor, B1Layout> ||
+              is_same_v<tensor_layout::gemm::ColumnMajor,
+                        B1Layout>)&&CheckDLayout<tensor_layout::gemm::RowMajor,
+                                                 D1sLayout,
+                                                 NumD1Tensor>() &&
+             is_same_v<tensor_layout::gemm::RowMajor, E1Layout>))
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a0_grid_desc_m_k_,
+                                           arg.b0_grid_desc_n_k_,
+                                           arg.b1_grid_desc_n_k_,
+                                           arg.e1_grid_desc_m_n_,
+                                           arg.block_2_e1tile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const A0DataType* p_a0,
+                             const B0DataType* p_b0,
+                             std::array<const void*, NumD0Tensor> p_d0s,
+                             const B1DataType* p_b1,
+                             std::array<const void*, NumD1Tensor> p_d1s,
+                             E1DataType* p_e1,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t Gemm1NRaw,
+                             index_t Batch,
+                             index_t StrideA0,
+                             index_t StrideB0,
+                             std::array<index_t, NumD0Tensor> StrideD0s,
+                             index_t StrideB1,
+                             std::array<index_t, NumD1Tensor> StrideD1s,
+                             index_t StrideE1,
+                             index_t BatchStrideA0,
+                             index_t BatchStrideB0,
+                             std::array<index_t, NumD0Tensor> BatchStrideD0s,
+                             index_t BatchStrideB1,
+                             std::array<index_t, NumD1Tensor> BatchStrideD1s,
+                             index_t BatchStrideE1,
+                             A0ElementwiseOperation a0_element_op,
+                             B0ElementwiseOperation b0_element_op,
+                             CDE0ElementwiseOperation cde0_element_op,
+                             B1ElementwiseOperation b1_element_op,
+                             CDE1ElementwiseOperation cde1_element_op)
+    {
+        return Argument{p_a0,          p_b0,
+                        p_d0s,         p_b1,
+                        p_d1s,         p_e1,
+                        MRaw,          NRaw,
+                        KRaw,          Gemm1NRaw,
+                        Batch,         StrideA0,
+                        StrideB0,      StrideD0s,
+                        StrideB1,      StrideD1s,
+                        StrideE1,      BatchStrideA0,
+                        BatchStrideB0, BatchStrideD0s,
+                        BatchStrideB1, BatchStrideD1s,
+                        BatchStrideE1, a0_element_op,
+                        b0_element_op, cde0_element_op,
+                        b1_element_op, cde1_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a0,
+                        const void* p_b0,
+                        std::array<const void*, NumD0Tensor> p_d0s,
+                        const void* p_b1,
+                        std::array<const void*, NumD1Tensor> p_d1s,
+                        void* p_e1,
+                        index_t MRaw,
+                        index_t NRaw,
+                        index_t KRaw,
+                        index_t Gemm1NRaw,
+                        index_t Batch,
+                        index_t StrideA0,
+                        index_t StrideB0,
+                        std::array<ck::index_t, NumD0Tensor> StrideD0s,
+                        index_t StrideB1,
+                        std::array<ck::index_t, NumD1Tensor> StrideD1s,
+                        index_t StrideE1,
+                        index_t BatchStrideA0,
+                        index_t BatchStrideB0,
+                        std::array<ck::index_t, NumD0Tensor> BatchStrideD0s,
+                        index_t BatchStrideB1,
+                        std::array<ck::index_t, NumD1Tensor> BatchStrideD1s,
+                        index_t BatchStrideE1,
+                        A0ElementwiseOperation a0_element_op,
+                        B0ElementwiseOperation b0_element_op,
+                        CDE0ElementwiseOperation cde0_element_op,
+                        B1ElementwiseOperation b1_element_op,
+                        CDE1ElementwiseOperation cde1_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const A0DataType*>(p_a0),
+                                          static_cast<const B0DataType*>(p_b0),
+                                          p_d0s,
+                                          static_cast<const B1DataType*>(p_b1),
+                                          p_d1s,
+                                          static_cast<E1DataType*>(p_e1),
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          Gemm1NRaw,
+                                          Batch,
+                                          StrideA0,
+                                          StrideB0,
+                                          StrideD0s,
+                                          StrideB1,
+                                          StrideD1s,
+                                          StrideE1,
+                                          BatchStrideA0,
+                                          BatchStrideB0,
+                                          BatchStrideD0s,
+                                          BatchStrideB1,
+                                          BatchStrideD1s,
+                                          BatchStrideE1,
+                                          a0_element_op,
+                                          b0_element_op,
+                                          cde0_element_op,
+                                          b1_element_op,
+                                          cde1_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << Gemm0MPerBlock << ", "
+            << Gemm0NPerBlock << ", "
+            << Gemm0KPerBlock << ", "
+            << A0K1 << ", "
+            << B0K1 << ", "
+            << B1K1 << ", "
+            << Gemm0MPerXdl << ", "
+            << Gemm0NPerXdl << ", "
+            << Gemm0MXdlPerWave << ", "
+            << Gemm0NXdlPerWave << ", "
+            << Gemm1NXdlPerWave << "> ";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..c4567108c9
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp
@@ -0,0 +1,1007 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename ReduceGridDescriptor_MBlock_MPerBlock,
+          typename ComputeBasePrtOfBatch,
+          typename Block2CTileMap,
+          bool HasMainK0BlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_reduce_xdl_cshuffle_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            ReducePtrsGlobal p_reduces_grid,
+            const index_t batch_count,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const ReduceInElementwiseOperations reduce_in_element_ops,
+            const ReduceAccElementwiseOperations reduce_out_element_ops,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const ReduceGridDescriptor_MBlock_MPerBlock reduce_grid_desc_mblock_mperblock,
+            const ComputeBasePrtOfBatch compute_base_ptr_of_batch_,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch_.GetABasePtr(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch_.GetBBasePtr(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch_.GetCBasePtr(g_idx)));
+
+    static_for<0, p_reduces_grid.Size(), 1>{}([&](auto In) {
+        const long_index_t d_batch_offset = __builtin_amdgcn_readfirstlane(
+            static_cast<long_index_t>(compute_base_ptr_of_batch_.GetDBasePtr(g_idx, In)));
+        p_reduces_grid(In) = p_reduces_grid(In) + d_batch_offset;
+    });
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainK0BlockLoop>(p_a_grid + a_batch_offset,
+                                                   p_b_grid + b_batch_offset,
+                                                   p_c_grid + c_batch_offset,
+                                                   p_reduces_grid,
+                                                   p_shared,
+                                                   a_element_op,
+                                                   b_element_op,
+                                                   c_element_op,
+                                                   reduce_in_element_ops,
+                                                   reduce_out_element_ops,
+                                                   a_grid_desc_ak0_m_ak1,
+                                                   b_grid_desc_bk0_n_bk1,
+                                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                   reduce_grid_desc_mblock_mperblock,
+                                                   block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_reduces_grid;
+    ignore = batch_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = reduce_in_element_ops;
+    ignore = reduce_out_element_ops;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = reduce_grid_desc_mblock_mperblock;
+    ignore = compute_base_ptr_of_batch_;
+    ignore = block_2_ctile_map;
+#endif
+}
+
+// Note: inter-wave loop scheduler is rolled out to c-shuffle version first. Becuase non c-shuffle
+// version currently has compiler issues with register spill which further causes validation
+// failures.
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename ReduceAccDataType,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename ReduceOperations,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          typename ReduceGlobalMemoryDataOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+          index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+          index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<0, ReduceOperations::Size()>
+{
+    using DeviceOp = DeviceBatchedGemmReduce_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(M)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_right_pad_transform(MRaw, MPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto NPad = N - NRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(NRaw, NPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(N)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_right_pad_transform(NRaw, NPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideC));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto NPad = N - NRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(NRaw, NPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_grid_desc_mraw_nraw;
+        }
+    }
+
+    // assume D is packed tensor
+    static auto MakeReduceGridDescriptor_M(index_t MRaw)
+    {
+        const auto d_grid_desc_mraw = make_naive_tensor_descriptor_packed(make_tuple(MRaw));
+
+        const auto M    = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto MPad = M - MRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                     GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M
+            return transform_tensor_descriptor(d_grid_desc_mraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad)),
+                                               make_tuple(Sequence<0>{}),
+                                               make_tuple(Sequence<0>{}));
+        }
+        else
+        {
+            // not pad M
+            return d_grid_desc_mraw;
+        }
+    }
+
+    using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
+    using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
+    using CGridDesc_M_N       = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+    using ReduceGridDesc_M    = decltype(MakeReduceGridDescriptor_M(1));
+
+    struct ComputeBasePtrOfStridedBatch
+    {
+        ComputeBasePtrOfStridedBatch(index_t BatchStrideA,
+                                     index_t BatchStrideB,
+                                     index_t BatchStrideC,
+                                     index_t BatchStrideD)
+            : BatchStrideA_(BatchStrideA),
+              BatchStrideB_(BatchStrideB),
+              BatchStrideC_(BatchStrideC),
+              BatchStrideD_(BatchStrideD)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideA_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideC_);
+        }
+
+        template <index_t I>
+        __host__ __device__ constexpr long_index_t GetDBasePtr(index_t g_idx,
+                                                               Number<I> reduction_idx) const
+        {
+            // TODO - Support sequence of StrideD in MakeArgument()
+            (void)reduction_idx;
+            return g_idx * static_cast<long_index_t>(BatchStrideD_);
+        }
+
+        private:
+        index_t BatchStrideA_;
+        index_t BatchStrideB_;
+        index_t BatchStrideC_;
+        index_t BatchStrideD_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        ReduceAccDataType,
+        ReducePtrsGlobal,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        ReduceOperations,
+        ReduceInElementwiseOperations,
+        ReduceAccElementwiseOperations,
+        InMemoryDataOperationEnum::Set,
+        ReduceGlobalMemoryDataOperation,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        ReduceGridDesc_M,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+        CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+        CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+        LoopSched>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 ReducePtrsGlobal p_reduces_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op,
+                 ReduceInElementwiseOperations reduce_in_element_ops,
+                 ReduceAccElementwiseOperations reduce_out_element_ops,
+                 index_t Batch)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              p_reduces_grid_{p_reduces_grid},
+              Batch_(Batch),
+              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, NRaw, StrideC)},
+              reduce_grid_desc_m_{DeviceOp::MakeReduceGridDescriptor_M(MRaw)},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              reduce_grid_desc_mblock_mperblock_{},
+              compute_base_ptr_of_batch_{
+                  type_convert<index_t>(a_grid_desc_ak0_m_ak1_.GetElementSpaceSize()),
+                  type_convert<index_t>(b_grid_desc_bk0_n_bk1_.GetElementSpaceSize()),
+                  type_convert<index_t>(c_grid_desc_m_n_.GetElementSpaceSize()),
+                  type_convert<index_t>(reduce_grid_desc_m_.GetElementSpaceSize())},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op},
+              reduce_in_element_ops_{reduce_in_element_ops},
+              reduce_out_element_ops_{reduce_out_element_ops}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+
+                reduce_grid_desc_mblock_mperblock_ =
+                    GridwiseGemm::MakeReduceGridDescriptor_MBlock_MPerBlock(reduce_grid_desc_m_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        ReducePtrsGlobal p_reduces_grid_;
+        index_t Batch_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        ReduceGridDesc_M reduce_grid_desc_m_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock
+            reduce_grid_desc_mblock_mperblock_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+        ReduceInElementwiseOperations reduce_in_element_ops_;
+        ReduceAccElementwiseOperations reduce_out_element_ops_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if DEBUG_LOG
+            {
+                std::cout << "arg.Batch_ = " << arg.Batch_ << std::endl;
+
+                std::cout << "arg.a_grid_desc_ak0_m_ak1_{"
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_bk0_n_bk1_{"
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I0) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+
+                std::cout << "arg.reduce_grid_desc_m_{ " << arg.reduce_grid_desc_m_.GetLength(I0)
+                          << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                            arg.b_grid_desc_bk0_n_bk1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.Batch_;
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float elapsed_time = 0.0f;
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_batched_gemm_reduce_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    ReducePtrsGlobal,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    ReduceInElementwiseOperations,
+                    ReduceAccElementwiseOperations,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock,
+                    ComputeBasePtrOfStridedBatch,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    true>;
+
+                elapsed_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_reduces_grid_,
+                                           arg.Batch_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.reduce_in_element_ops_,
+                                           arg.reduce_out_element_ops_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.reduce_grid_desc_mblock_mperblock_,
+                                           arg.compute_base_ptr_of_batch_,
+                                           arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_batched_gemm_reduce_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    ReducePtrsGlobal,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    ReduceInElementwiseOperations,
+                    ReduceAccElementwiseOperations,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock,
+                    ComputeBasePtrOfStridedBatch,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    false>;
+
+                elapsed_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_reduces_grid_,
+                                           arg.Batch_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.reduce_in_element_ops_,
+                                           arg.reduce_out_element_ops_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.reduce_grid_desc_mblock_mperblock_,
+                                           arg.compute_base_ptr_of_batch_,
+                                           arg.block_2_ctile_map_);
+            }
+
+            return elapsed_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        auto casted_p_arg = dynamic_cast<const Argument*>(p_arg);
+        if(casted_p_arg == nullptr)
+        {
+            return false;
+        }
+        else
+        {
+            return IsSupportedArgument(*casted_p_arg);
+        }
+    }
+
+    static constexpr int NumReduce = ReduceOperations::Size();
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             const void* p_bias,
+                             std::array<const void*, 0> p_ds,
+                             void* p_c,
+                             std::array<void*, NumReduce> p_reduces,
+                             ck::index_t M,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t StrideA,
+                             ck::index_t StrideB,
+                             ck::index_t StrideC,
+                             std::array<ck::index_t, 0> StrideDs,
+                             std::array<void*, 3> gemm_element_ops,
+                             std::array<void*, 0> d_element_ops,
+                             std::array<void*, NumReduce> reduce_in_element_op,
+                             std::array<void*, NumReduce> reduce_out_element_op,
+                             index_t Batch)
+    {
+        (void)p_bias;
+        (void)p_ds;
+        (void)StrideDs;
+        (void)d_element_ops;
+
+        ReducePtrsGlobal reduce_tuple = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReducePtrsGlobal{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return static_cast<T*>(p_reduces[I]);
+            },
+            Number<NumReduce>{});
+
+        ReduceInElementwiseOperations reduce_in_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceInElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_in_element_op[I]));
+            },
+            Number<NumReduce>{});
+        ReduceAccElementwiseOperations reduce_out_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceAccElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_out_element_op[I]));
+            },
+            Number<NumReduce>{});
+
+        AElementwiseOperation a_element_op =
+            *(static_cast<AElementwiseOperation*>(gemm_element_ops[0]));
+        BElementwiseOperation b_element_op =
+            *(static_cast<BElementwiseOperation*>(gemm_element_ops[1]));
+        CElementwiseOperation c_element_op =
+            *(static_cast<CElementwiseOperation*>(gemm_element_ops[2]));
+
+        return Argument{static_cast<const ADataType*>(p_a),
+                        static_cast<const BDataType*>(p_b),
+                        static_cast<CDataType*>(p_c),
+                        reduce_tuple,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op,
+                        reduce_in_element_ops,
+                        reduce_out_element_ops,
+                        Batch};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const void* p_bias,
+                        std::array<const void*, 0> p_ds,
+                        void* p_c,
+                        std::array<void*, NumReduce> p_reduces,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        std::array<ck::index_t, 0> StrideDs,
+                        std::array<void*, 3> gemm_element_ops,
+                        std::array<void*, 0> d_element_ops,
+                        std::array<void*, NumReduce> reduce_in_element_op,
+                        std::array<void*, NumReduce> reduce_out_element_op,
+                        index_t Batch = 1) override
+    {
+        (void)p_bias;
+        (void)p_ds;
+        (void)StrideDs;
+        (void)d_element_ops;
+
+        ReducePtrsGlobal reduce_tuple = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReducePtrsGlobal{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return static_cast<T*>(p_reduces[I]);
+            },
+            Number<NumReduce>{});
+
+        ReduceInElementwiseOperations reduce_in_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceInElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_in_element_op[I]));
+            },
+            Number<NumReduce>{});
+        ReduceAccElementwiseOperations reduce_out_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceAccElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_out_element_op[I]));
+            },
+            Number<NumReduce>{});
+
+        AElementwiseOperation a_element_op =
+            *(static_cast<AElementwiseOperation*>(gemm_element_ops[0]));
+        BElementwiseOperation b_element_op =
+            *(static_cast<BElementwiseOperation*>(gemm_element_ops[1]));
+        CElementwiseOperation c_element_op =
+            *(static_cast<CElementwiseOperation*>(gemm_element_ops[2]));
+
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          reduce_tuple,
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op,
+                                          reduce_in_element_ops,
+                                          reduce_out_element_ops,
+                                          Batch);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedGemmReduce_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..c38b7e1c5b
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
@@ -0,0 +1,954 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename D0sPointer,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename C0DEElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename C1DEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename B1GridDesc_BK0_N_BK1,
+          typename C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5,
+          typename Block2CTileMap,
+          typename ComputeBasePtrOfStridedBatch,
+          typename C0MatrixMask,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_softmax_gemm_xdl_cshuffle_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            const FloatAB* __restrict__ p_b1_grid,
+            FloatC* __restrict__ p_c_grid,
+            D0sPointer p_d0s_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const C0DEElementwiseOperation c0de_element_op,
+            const B1ElementwiseOperation b1_element_op,
+            const C1DEElementwiseOperation c1de_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
+            const C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c1_grid_desc_mblock_mperblock_nblock_nperblock,
+            const D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
+                d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+            const Block2CTileMap block_2_ctile_map,
+            const index_t batch_count,
+            const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
+            const C0MatrixMask c0_matrix_mask)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetABasePtr(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetBBasePtr(g_idx)));
+    const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetB1BasePtr(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetCBasePtr(g_idx)));
+
+    static_for<0, p_d0s_grid.Size(), 1>{}([&](auto In) {
+        const long_index_t d0_batch_offset = __builtin_amdgcn_readfirstlane(
+            static_cast<long_index_t>(compute_base_ptr_of_batch.GetD0BasePtr(g_idx, In)));
+        p_d0s_grid(In) = p_d0s_grid(In) + d0_batch_offset;
+    });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_b1_grid + b1_batch_offset,
+                                                  p_c_grid + c_batch_offset,
+                                                  p_d0s_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c0de_element_op,
+                                                  b1_element_op,
+                                                  c1de_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  b1_grid_desc_bk0_n_bk1,
+                                                  c1_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                                                  block_2_ctile_map,
+                                                  c0_matrix_mask);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_b1_grid;
+    ignore = p_c_grid;
+    ignore = p_d0s_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c0de_element_op;
+    ignore = b1_element_op;
+    ignore = c1de_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = b1_grid_desc_bk0_n_bk1;
+    ignore = c1_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5;
+    ignore = block_2_ctile_map;
+    ignore = batch_count;
+    ignore = compute_base_ptr_of_batch;
+    ignore = c0_matrix_mask;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// Computes C = A * B0 * B1
+//              ^^^^^^ (Acc0)
+//              ^^^^^^^^^^^ (Acc1)
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO, // NumDimGemm1N
+          typename ADataType,
+          typename BDataType,
+          typename B1DataType,
+          typename CDataType,
+          typename D0sDataType,
+          typename D1sDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename C0DEElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename C1DEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization B1Spec,
+          TensorSpecialization CSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock, // Gemm0NPerBlock
+          index_t KPerBlock, // Gemm0KPerBlock
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t B1K1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          MaskingSpecialization MaskingSpec,
+          int D0sTransferSrcScalarPerVector = 4,
+          LoopScheduler LoopSched           = LoopScheduler::Default>
+struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
+    : public DeviceBatchedGemmSoftmaxGemmPermute<NumDimG,
+                                                 NumDimM,
+                                                 NumDimN,
+                                                 NumDimK,
+                                                 NumDimO,
+                                                 ADataType,
+                                                 BDataType,
+                                                 B1DataType,
+                                                 CDataType,
+                                                 D0sDataType,
+                                                 D1sDataType,
+                                                 AElementwiseOperation,
+                                                 BElementwiseOperation,
+                                                 C0DEElementwiseOperation,
+                                                 B1ElementwiseOperation,
+                                                 C1DEElementwiseOperation,
+                                                 MaskingSpec>
+{
+    static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0 && NumDimO > 0,
+                  "Number of dimension must be greater than 0");
+
+    static constexpr index_t NumD0Tensor = D0sDataType::Size();
+    static constexpr index_t NumD1Tensor = D1sDataType::Size();
+
+    // TODO ANT: implement bias combination
+    static_assert(NumD1Tensor == 0, "Gemm1 Bias addition is unimplemented");
+
+#if 0
+    // TODO ANT: use alias
+    static constexpr index_t NumDimGemm0M = NumDimM;
+    static constexpr index_t NumDimGemm0N = NumDimN;
+    static constexpr index_t NumDimGemm0K = NumDimK;
+    static constexpr index_t NumDimGemm1M = NumDimM;
+    static constexpr index_t NumDimGemm1N = NumDimO;
+    static constexpr index_t NumDimGemm1K = NumDimN;
+#endif
+
+    using DeviceOp = DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    using Transform = TransformBatchedContractionContractionToBatchedGemmGemm<
+        Sequence<NumDimG, NumDimM, NumDimN, NumDimK, NumDimO>,
+        Sequence<MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock>,
+        GemmSpec,
+        ASpec,
+        BSpec,
+        B1Spec,
+        CSpec>;
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                              const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        return Transform::MakeAGridDescriptor_AK0_M_AK1(
+            Transform::MakeAGridDescriptor_M_K(a_gs_ms_ks_lengths_vec, a_gs_ms_ks_strides_vec),
+            Number<AK1>{});
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
+                                              const std::vector<index_t>& b_gs_ns_ks_strides_vec)
+    {
+        return Transform::MakeB0GridDescriptor_BK0_N_BK1(
+            Transform::MakeB0GridDescriptor_N_K(b_gs_ns_ks_lengths_vec, b_gs_ns_ks_strides_vec),
+            Number<BK1>{});
+    }
+
+    static auto
+    MakeB1GridDescriptor_BK0_N_BK1(const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths_vec,
+                                   const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides_vec)
+    {
+        return Transform::MakeB1GridDescriptor_BK0_N_BK1(
+            Transform::MakeB1GridDescriptor_N_K(b1_gs_gemm1ns_gemm1ks_lengths_vec,
+                                                b1_gs_gemm1ns_gemm1ks_strides_vec),
+            Number<B1K1>{});
+    }
+
+    static auto MakeD0sGridDescriptor_M_N(
+        const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_strides)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return Transform::MakeCGridDescriptor_M_N(acc0_biases_gs_ms_ns_lengths[i],
+                                                          acc0_biases_gs_ms_ns_strides[i]);
+            },
+            Number<NumD0Tensor>{});
+    }
+
+    static auto MakeD0sGridDescriptor_G_M_N(
+        const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_strides)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return Transform::MakeCGridDescriptor_G_M_N(acc0_biases_gs_ms_ns_lengths[i],
+                                                            acc0_biases_gs_ms_ns_strides[i]);
+            },
+            Number<NumD0Tensor>{});
+    }
+
+    using AGridDesc_AK0_M_AK1  = decltype(MakeAGridDescriptor_AK0_M_AK1({}, {}));
+    using BGridDesc_BK0_N_BK1  = decltype(MakeBGridDescriptor_BK0_N_BK1({}, {}));
+    using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1({}, {}));
+    using C1GridDesc_M_N       = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
+    using AGridDesc_G_M_K      = decltype(Transform::MakeAGridDescriptor_G_M_K({}, {}));
+    using BGridDesc_G_N_K      = decltype(Transform::MakeB0GridDescriptor_G_N_K({}, {}));
+    using B1GridDesc_G_N_K     = decltype(Transform::MakeB1GridDescriptor_G_N_K({}, {}));
+    using C1GridDesc_G_M_N     = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
+    using D0sGridDesc_M_N      = decltype(MakeD0sGridDescriptor_M_N({}, {}));
+    using D0sGridDesc_G_M_N    = decltype(MakeD0sGridDescriptor_G_M_N({}, {}));
+
+    constexpr static auto make_MaskOutPredicate()
+    {
+        if constexpr(MaskingSpec == MaskingSpecialization::MaskDisabled)
+        {
+            return MaskDisabledPredicate{};
+        }
+        else if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+        {
+            return MaskOutUpperTrianglePredicate{};
+        }
+    }
+    using C0MatrixMask = C0MatrixMask_impl<decltype(make_MaskOutPredicate())>;
+
+    struct ComputeBasePtrOfStridedBatch
+    {
+        ComputeBasePtrOfStridedBatch(const AGridDesc_G_M_K& a_grid_desc_g_m_k,
+                                     const BGridDesc_G_N_K& b_grid_desc_g_n_k,
+                                     const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
+                                     const C1GridDesc_G_M_N& c1_grid_desc_g_m_n,
+                                     const D0sGridDesc_G_M_N& d0s_grid_desc_g_m_n)
+            : a_grid_desc_g_m_k_(a_grid_desc_g_m_k),
+              b_grid_desc_g_n_k_(b_grid_desc_g_n_k),
+              b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
+              c1_grid_desc_g_m_n_(c1_grid_desc_g_m_n),
+              d0s_grid_desc_g_m_n_(d0s_grid_desc_g_m_n)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
+        {
+            return a_grid_desc_g_m_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
+        {
+            return b_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
+        {
+            return b1_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
+        {
+            return c1_grid_desc_g_m_n_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        template <index_t I>
+        __host__ __device__ constexpr long_index_t GetD0BasePtr(index_t g_idx,
+                                                                Number<I> d0_idx) const
+        {
+            return d0s_grid_desc_g_m_n_[d0_idx].CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        private:
+        AGridDesc_G_M_K a_grid_desc_g_m_k_;
+        BGridDesc_G_N_K b_grid_desc_g_n_k_;
+        B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
+        C1GridDesc_G_M_N c1_grid_desc_g_m_n_;
+        D0sGridDesc_G_M_N d0s_grid_desc_g_m_n_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        D0sDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        C0DEElementwiseOperation,
+        B1ElementwiseOperation,
+        C1DEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        B1GridDesc_BK0_N_BK1,
+        C1GridDesc_M_N,
+        D0sGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        Gemm1NPerBlock,
+        Gemm1KPerBlock,
+        AK1,
+        BK1,
+        B1K1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        Gemm1NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        true,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        true,
+        BBlockLdsExtraN,
+        B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+        B1BlockTransferThreadClusterArrangeOrder,
+        B1BlockTransferSrcAccessOrder,
+        B1BlockTransferSrcVectorDim,
+        B1BlockTransferSrcScalarPerVector,
+        B1BlockTransferDstScalarPerVector_BK1,
+        false,
+        B1BlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched,
+        Transform::matrix_padder.PadN,
+        MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle,
+        D0sTransferSrcScalarPerVector>;
+
+    // Argument
+    // FIXME: constness
+    struct Argument : public BaseArgument
+    {
+        Argument(
+            const ADataType* p_a_grid,
+            const BDataType* p_b_grid,
+            const B1DataType* p_b1_grid,
+            CDataType* p_c_grid,
+            const std::array<void*, NumD0Tensor> p_acc0_biases,
+            const std::array<void*, NumD1Tensor> p_acc1_biases,
+            const std::vector<index_t>& a_gs_ms_ks_lengths,
+            const std::vector<index_t>& a_gs_ms_ks_strides,
+            const std::vector<index_t>& b_gs_ns_ks_lengths,
+            const std::vector<index_t>& b_gs_ns_ks_strides,
+            const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+            const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+            const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+            const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+            const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_lengths,
+            const std::array<std::vector<ck::index_t>, NumD0Tensor>& acc0_biases_gs_ms_ns_strides,
+            const std::array<std::vector<ck::index_t>, NumD1Tensor>&
+                acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+            const std::array<std::vector<ck::index_t>, NumD1Tensor>&
+                acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+            AElementwiseOperation a_element_op,
+            BElementwiseOperation b_element_op,
+            C0DEElementwiseOperation c0de_element_op,
+            B1ElementwiseOperation b1_element_op,
+            C1DEElementwiseOperation c1de_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_b1_grid_{p_b1_grid},
+              p_c_grid_{p_c_grid},
+              p_d0s_grid_{},
+              a_grid_desc_ak0_m_ak1_{
+                  DeviceOp::MakeAGridDescriptor_AK0_M_AK1(a_gs_ms_ks_lengths, a_gs_ms_ks_strides)},
+              b_grid_desc_bk0_n_bk1_{
+                  DeviceOp::MakeBGridDescriptor_BK0_N_BK1(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
+              b1_grid_desc_bk0_n_bk1_{DeviceOp::MakeB1GridDescriptor_BK0_N_BK1(
+                  b1_gs_gemm1ns_gemm1ks_lengths, b1_gs_gemm1ns_gemm1ks_strides)},
+              c1_grid_desc_m_n_{Transform::MakeCGridDescriptor_M_N(c_gs_ms_gemm1ns_lengths,
+                                                                   c_gs_ms_gemm1ns_strides)},
+              a_grid_desc_g_m_k_{
+                  Transform::MakeAGridDescriptor_G_M_K(a_gs_ms_ks_lengths, a_gs_ms_ks_strides)},
+              b_grid_desc_g_n_k_{
+                  Transform::MakeB0GridDescriptor_G_N_K(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
+              b1_grid_desc_g_n_k_{Transform::MakeB1GridDescriptor_G_N_K(
+                  b1_gs_gemm1ns_gemm1ks_lengths, b1_gs_gemm1ns_gemm1ks_strides)},
+              c1_grid_desc_g_m_n_{Transform::MakeCGridDescriptor_G_M_N(c_gs_ms_gemm1ns_lengths,
+                                                                       c_gs_ms_gemm1ns_strides)},
+              d0s_grid_desc_g_m_n_{DeviceOp::MakeD0sGridDescriptor_G_M_N(
+                  acc0_biases_gs_ms_ns_lengths, acc0_biases_gs_ms_ns_strides)},
+              c1_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_{},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c1_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c0de_element_op_{c0de_element_op},
+              b1_element_op_{b1_element_op},
+              c1de_element_op_{c1de_element_op},
+              c0_matrix_mask_{b_grid_desc_g_n_k_.GetLength(I1)},
+              raw_lengths_mz_nz_kz_gemm1nz_{a_gs_ms_ks_lengths[NumDimG + NumDimM - 1],
+                                            b_gs_ns_ks_lengths[NumDimG + NumDimN - 1],
+                                            b_gs_ns_ks_lengths[NumDimG + NumDimN + NumDimK - 1],
+                                            b1_gs_gemm1ns_gemm1ks_lengths[NumDimG + NumDimO - 1]},
+              a_mz_kz_strides_{a_gs_ms_ks_strides[NumDimG + NumDimM - 1],
+                               a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1]},
+              b_nz_kz_strides_{b_gs_ns_ks_strides[NumDimG + NumDimN - 1],
+                               b_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1]},
+              b1_nz_kz_strides_{b1_gs_gemm1ns_gemm1ks_strides[NumDimG + NumDimO - 1],
+                                b1_gs_gemm1ns_gemm1ks_strides[NumDimG + NumDimO + NumDimN - 1]},
+              c_mz_gemm1nz_strides_{c_gs_ms_gemm1ns_strides[NumDimG + NumDimM - 1],
+                                    c_gs_ms_gemm1ns_strides[NumDimG + NumDimM + NumDimO - 1]},
+              batch_count_{c1_grid_desc_g_m_n_.GetLength(I0)},
+              compute_base_ptr_of_batch_{a_grid_desc_g_m_k_,
+                                         b_grid_desc_g_n_k_,
+                                         b1_grid_desc_g_n_k_,
+                                         c1_grid_desc_g_m_n_,
+                                         d0s_grid_desc_g_m_n_}
+        {
+            // TODO ANT: implement bias addition
+            ignore = p_acc1_biases;
+            ignore = acc1_biases_gs_ms_gemm1ns_lengths;
+            ignore = acc1_biases_gs_ms_gemm1ns_strides;
+
+            static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
+                // D0 pointer
+                p_d0s_grid_(i) = static_cast<const D0DataType*>(p_acc0_biases[i]);
+                // for  check
+                d0s_nl_ns_lengths_strides_[i].push_back(
+                    acc0_biases_gs_ms_ns_lengths[i][NumDimG + NumDimM]);
+                d0s_nl_ns_lengths_strides_[i].push_back(
+                    acc0_biases_gs_ms_ns_strides[i][NumDimG + NumDimM]);
+            });
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           b1_grid_desc_bk0_n_bk1_,
+                                           c1_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c1_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeC1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c1_grid_desc_m_n_);
+
+                D0sGridDesc_M_N d0s_grid_desc_m_n{DeviceOp::MakeD0sGridDescriptor_M_N(
+                    acc0_biases_gs_ms_ns_lengths, acc0_biases_gs_ms_ns_strides)};
+                d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_ =
+                    GridwiseGemm::MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(
+                        d0s_grid_desc_m_n);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "a_grid_desc_g_m_k_: " << a_grid_desc_g_m_k_.GetLength(I0) << ", "
+                      << a_grid_desc_g_m_k_.GetLength(I1) << ", "
+                      << a_grid_desc_g_m_k_.GetLength(I2) << '\n';
+            std::cout << "b_grid_desc_g_n_k_: " << b_grid_desc_g_n_k_.GetLength(I0) << ", "
+                      << b_grid_desc_g_n_k_.GetLength(I1) << ", "
+                      << b_grid_desc_g_n_k_.GetLength(I2) << '\n';
+            std::cout << "b1_grid_desc_g_n_k_: " << b1_grid_desc_g_n_k_.GetLength(I0) << ", "
+                      << b1_grid_desc_g_n_k_.GetLength(I1) << ", "
+                      << b1_grid_desc_g_n_k_.GetLength(I2) << '\n';
+            std::cout << "c1_grid_desc_g_m_n_: " << c1_grid_desc_g_m_n_.GetLength(I0) << ", "
+                      << c1_grid_desc_g_m_n_.GetLength(I1) << ", "
+                      << c1_grid_desc_g_m_n_.GetLength(I2) << '\n';
+        }
+
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        const B1DataType* p_b1_grid_;
+        CDataType* p_c_grid_;
+        typename GridwiseGemm::D0sGridPointer p_d0s_grid_;
+
+        // tensor descriptor
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
+        C1GridDesc_M_N c1_grid_desc_m_n_;
+        AGridDesc_G_M_K a_grid_desc_g_m_k_;
+        BGridDesc_G_N_K b_grid_desc_g_n_k_;
+        B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
+        C1GridDesc_G_M_N c1_grid_desc_g_m_n_;
+        D0sGridDesc_G_M_N d0s_grid_desc_g_m_n_;
+
+        typename GridwiseGemm::C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c1_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
+            d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_;
+
+        // block-to-c-tile map
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        C0DEElementwiseOperation c0de_element_op_;
+        B1ElementwiseOperation b1_element_op_;
+        C1DEElementwiseOperation c1de_element_op_;
+
+        // check C0 masking and padding
+        C0MatrixMask c0_matrix_mask_;
+
+        // For robust IsSupportedArgument() check
+        std::vector<index_t> raw_lengths_mz_nz_kz_gemm1nz_;
+        std::vector<index_t> a_mz_kz_strides_;
+        std::vector<index_t> b_nz_kz_strides_;
+        std::vector<index_t> b1_nz_kz_strides_;
+        std::vector<index_t> c_mz_gemm1nz_strides_;
+        std::array<std::vector<ck::index_t>, NumD0Tensor> d0s_nl_ns_lengths_strides_;
+
+        index_t batch_count_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!DeviceOp::IsSupportedArgument(arg))
+            {
+                throw std::runtime_error("wrong! unsupported argument");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c1_grid_desc_m_n_) * arg.batch_count_;
+
+            // Gemm0_K
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel = kernel_batched_gemm_softmax_gemm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    typename GridwiseGemm::D0sGridPointer,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    C0DEElementwiseOperation,
+                    B1ElementwiseOperation,
+                    C1DEElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    DeviceOp::B1GridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    ComputeBasePtrOfStridedBatch,
+                    C0MatrixMask,
+                    has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_b1_grid_,
+                                              arg.p_c_grid_,
+                                              arg.p_d0s_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.c0de_element_op_,
+                                              arg.b1_element_op_,
+                                              arg.c1de_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.b1_grid_desc_bk0_n_bk1_,
+                                              arg.c1_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_,
+                                              arg.block_2_ctile_map_,
+                                              arg.batch_count_,
+                                              arg.compute_base_ptr_of_batch_,
+                                              arg.c0_matrix_mask_);
+            };
+
+            // Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
+            // to concern Gemm0's loop
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+#if DEBUG_LOG
+        arg.Print();
+#endif
+
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        // TODO ANT: Check if tensor specialization & strides mismatch
+
+        // Check if C permute dimension matches GEMM + GEMM shape
+        const index_t c_g       = arg.c1_grid_desc_g_m_n_.GetLength(I0); // unpadded
+        const index_t c_m       = arg.c1_grid_desc_m_n_.GetLength(I0);
+        const index_t c_gemm1n  = arg.c1_grid_desc_m_n_.GetLength(I1);
+        const index_t a_m       = arg.a_grid_desc_ak0_m_ak1_.GetLength(I1);
+        const index_t b1_gemm1n = arg.b1_grid_desc_bk0_n_bk1_.GetLength(I1);
+
+        if(!(c_g == arg.batch_count_ && c_m == a_m && c_gemm1n == b1_gemm1n))
+        {
+            return false;
+        }
+
+        // Note: we need raw lengths since threadwise copy can not handle vector load when part of
+        // vector is out of bounds
+        // Note: need lowest dim in Ms/Ns/Ks/Os, not merged M/N/K/O
+        const auto MzRaw      = arg.raw_lengths_mz_nz_kz_gemm1nz_[0];
+        const auto NzRaw      = arg.raw_lengths_mz_nz_kz_gemm1nz_[1];
+        const auto KzRaw      = arg.raw_lengths_mz_nz_kz_gemm1nz_[2];
+        const auto Gemm1NzRaw = arg.raw_lengths_mz_nz_kz_gemm1nz_[3];
+
+        // Check scalar per vector requirement
+        const auto a_extent_lowest  = ABlockTransferSrcVectorDim == 2 ? KzRaw : MzRaw;
+        const auto b_extent_lowest  = BBlockTransferSrcVectorDim == 2 ? KzRaw : NzRaw;
+        const auto b1_extent_lowest = B1BlockTransferSrcVectorDim == 2 ? NzRaw : Gemm1NzRaw;
+        const auto c_extent_lowest  = Gemm1NzRaw;
+
+        if(!(a_extent_lowest % ABlockTransferSrcScalarPerVector == 0 &&
+             b_extent_lowest % BBlockTransferSrcScalarPerVector == 0 &&
+             b1_extent_lowest % B1BlockTransferSrcScalarPerVector == 0 &&
+             c_extent_lowest % CShuffleBlockTransferScalarPerVector_NPerBlock == 0))
+        {
+            return false;
+        }
+
+        // Check vector load/store requirement
+        const auto a_stride_lowest =
+            ABlockTransferSrcVectorDim == 2 ? arg.a_mz_kz_strides_[1] : arg.a_mz_kz_strides_[0];
+        const auto b_stride_lowest =
+            BBlockTransferSrcVectorDim == 2 ? arg.b_nz_kz_strides_[1] : arg.b_nz_kz_strides_[0];
+        const auto b1_stride_lowest =
+            B1BlockTransferSrcVectorDim == 2 ? arg.b1_nz_kz_strides_[1] : arg.b1_nz_kz_strides_[0];
+        const auto c_stride_lowest =
+            arg.c_mz_gemm1nz_strides_[1]; // cshuffle assumes lowest dim in Gemm1Ns to be contiguous
+
+        if(!(a_stride_lowest == 1 || b_stride_lowest == 1 || b1_stride_lowest == 1 ||
+             c_stride_lowest == 1))
+        {
+            return false;
+        }
+        for(int i = 0; i < NumD0Tensor; i++)
+        {
+            if(arg.d0s_nl_ns_lengths_strides_[i][1] == 1 &&
+               arg.d0s_nl_ns_lengths_strides_[i][0] % D0sTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+            if(arg.d0s_nl_ns_lengths_strides_[i][1] != 1 && D0sTransferSrcScalarPerVector != 1)
+            {
+                return false;
+            }
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.b1_grid_desc_bk0_n_bk1_,
+                                           arg.c1_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(
+        const ADataType* p_a,
+        const BDataType* p_b,
+        const B1DataType* p_b1,
+        CDataType* p_c,
+        const std::array<void*, NumD0Tensor> p_acc0_biases,
+        const std::array<void*, NumD1Tensor> p_acc1_biases,
+        const std::vector<index_t>& a_gs_ms_ks_lengths,
+        const std::vector<index_t>& a_gs_ms_ks_strides,
+        const std::vector<index_t>& b_gs_ns_ks_lengths,
+        const std::vector<index_t>& b_gs_ns_ks_strides,
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+        const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+        const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+        const std::array<std::vector<ck::index_t>, NumD0Tensor> acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<ck::index_t>, NumD0Tensor> acc0_biases_gs_ms_ns_strides,
+        const std::array<std::vector<ck::index_t>, NumD1Tensor>
+            acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+        const std::array<std::vector<ck::index_t>, NumD1Tensor>
+            acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+        AElementwiseOperation a_element_op,
+        BElementwiseOperation b_element_op,
+        C0DEElementwiseOperation c0de_element_op,
+        B1ElementwiseOperation b1_element_op,
+        C1DEElementwiseOperation c1de_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_b1,
+                        p_c,
+                        p_acc0_biases,
+                        p_acc1_biases,
+                        a_gs_ms_ks_lengths,
+                        a_gs_ms_ks_strides,
+                        b_gs_ns_ks_lengths,
+                        b_gs_ns_ks_strides,
+                        b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+                        b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+                        c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+                        c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+                        acc0_biases_gs_ms_ns_lengths,
+                        acc0_biases_gs_ms_ns_strides,
+                        acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+                        acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+                        a_element_op,
+                        b_element_op,
+                        c0de_element_op,
+                        b1_element_op,
+                        c1de_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    // FIXME: constness
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const void* p_b1,
+        void* p_c,
+        const std::array<void*, NumD0Tensor> p_acc0_biases,
+        const std::array<void*, NumD1Tensor> p_acc1_biases,
+        const std::vector<index_t>& a_gs_ms_ks_lengths,
+        const std::vector<index_t>& a_gs_ms_ks_strides,
+        const std::vector<index_t>& b_gs_ns_ks_lengths,
+        const std::vector<index_t>& b_gs_ns_ks_strides,
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+        const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+        const std::vector<index_t>& c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+        const std::vector<index_t>& c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+        const std::array<std::vector<ck::index_t>, NumD0Tensor> acc0_biases_gs_ms_ns_lengths,
+        const std::array<std::vector<ck::index_t>, NumD0Tensor> acc0_biases_gs_ms_ns_strides,
+        const std::array<std::vector<ck::index_t>, NumD1Tensor>
+            acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
+        const std::array<std::vector<ck::index_t>, NumD1Tensor>
+            acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
+        AElementwiseOperation a_element_op,
+        BElementwiseOperation b_element_op,
+        C0DEElementwiseOperation c0de_element_op,
+        B1ElementwiseOperation b1_element_op,
+        C1DEElementwiseOperation c1de_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<const B1DataType*>(p_b1),
+                                          static_cast<CDataType*>(p_c),
+                                          p_acc0_biases, // cast in struct Argument
+                                          p_acc1_biases, // cast in struct Argument
+                                          a_gs_ms_ks_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b_gs_ns_ks_lengths,
+                                          b_gs_ns_ks_strides,
+                                          b1_gs_gemm1ns_gemm1ks_lengths, // b1_gs_os_ns_lengths
+                                          b1_gs_gemm1ns_gemm1ks_strides, // b1_gs_os_ns_strides
+                                          c_gs_ms_gemm1ns_lengths,       // c_gs_ms_os_lengths
+                                          c_gs_ms_gemm1ns_strides,       // c_gs_ms_os_strides
+                                          acc0_biases_gs_ms_ns_lengths,
+                                          acc0_biases_gs_ms_ns_strides,
+                                          acc1_biases_gs_ms_gemm1ns_lengths,
+                                          acc1_biases_gs_ms_gemm1ns_strides,
+                                          a_element_op,
+                                          b_element_op,
+                                          c0de_element_op,
+                                          b1_element_op,
+                                          c1de_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerBlock << ", "
+            << Gemm1NPerBlock << ", "
+            << Gemm1KPerBlock << ", "
+            << B1K1 << ", "
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << "ASpec" << getTensorSpecializationString(ASpec) << ", "
+            << "B0Spec" << getTensorSpecializationString(BSpec) << ", "
+            << "B1Spec" << getTensorSpecializationString(B1Spec) << ", "
+            << "CSpec" << getTensorSpecializationString(CSpec) << ", "
+            << getMaskingSpecializationString(MaskingSpec) << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..f12e05fd34
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp
@@ -0,0 +1,772 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename B1GridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2CTileMap,
+          typename ComputeBasePtrOfStridedBatch,
+          typename C0MatrixMask,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_softmax_gemm_xdl_cshuffle_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            const FloatAB* __restrict__ p_b1_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const AccElementwiseOperation acc_element_op,
+            const B1ElementwiseOperation b1_element_op,
+            const CElementwiseOperation c_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const Block2CTileMap block_2_ctile_map,
+            const index_t batch_count,
+            const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch,
+            const C0MatrixMask c0_matrix_mask)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetABasePtr(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetBBasePtr(g_idx)));
+    const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetB1BasePtr(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_base_ptr_of_batch.GetCBasePtr(g_idx)));
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_b1_grid + b1_batch_offset,
+                                                  p_c_grid + c_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  acc_element_op,
+                                                  b1_element_op,
+                                                  c_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  b1_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_ctile_map,
+                                                  c0_matrix_mask);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_b1_grid;
+    ignore = p_c_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = acc_element_op;
+    ignore = b1_element_op;
+    ignore = c_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = b1_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_ctile_map;
+    ignore = batch_count;
+    ignore = compute_base_ptr_of_batch;
+    ignore = c0_matrix_mask;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// Computes C = A * B0 * B1
+//              ^^^^^^ (Acc0)
+//              ^^^^^^^^^^^ (Acc1)
+
+// When using NPadding as GemmSpecialization, AccElementwiseOperation should be set to
+// ScaleAndResetNaNToMinusInfinity.
+// if !isNan(AccElement)
+//     AccElement *= scale
+// else
+//     AccElement = -INFINITY
+// Otherwise, result may be wrong.
+
+template <typename ALayout,
+          typename BLayout, // B0Layout
+          typename B1Layout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename B1DataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock, // Gemm0NPerBlock
+          index_t KPerBlock, // Gemm0KPerBlock
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t B1K1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          bool MaskOutUpperTriangle,
+          LoopScheduler LoopSched = LoopScheduler::Default>
+struct DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle
+    : public DeviceBatchedGemmSoftmaxGemm<ALayout,
+                                          BLayout,
+                                          B1Layout,
+                                          CLayout,
+                                          ADataType,
+                                          BDataType,
+                                          B1DataType,
+                                          CDataType,
+                                          AElementwiseOperation,
+                                          BElementwiseOperation,
+                                          AccElementwiseOperation,
+                                          B1ElementwiseOperation,
+                                          CElementwiseOperation,
+                                          MaskOutUpperTriangle>
+{
+    using DeviceOp = DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto matrix_padder =
+        GemmGemmPadder<GemmSpec, index_t, index_t, index_t, index_t>{
+            MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock};
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto a_grid_desc_m_k = matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto b_grid_desc_n_k = matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // Args: Gemm1KRaw, Gemm1NRaw, StrideB1
+    static auto MakeB1GridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b1_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, B1Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, B1Layout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto b1_grid_desc_n_k = matrix_padder.PadB1Descriptor_N_K(b1_grid_desc_nraw_kraw);
+
+        const auto N = b1_grid_desc_n_k.GetLength(I0);
+        const auto K = b1_grid_desc_n_k.GetLength(I1);
+
+        const auto B1K0 = K / B1K1;
+
+        return transform_tensor_descriptor(
+            b1_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
+                       make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideC));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(c_grid_desc_mraw_nraw);
+    }
+
+    struct ComputeBasePtrOfStridedBatch
+    {
+        ComputeBasePtrOfStridedBatch(index_t BatchStrideA,
+                                     index_t BatchStrideB,
+                                     index_t BatchStrideB1,
+                                     index_t BatchStrideC)
+            : BatchStrideA_(BatchStrideA),
+              BatchStrideB_(BatchStrideB),
+              BatchStrideB1_(BatchStrideB1),
+              BatchStrideC_(BatchStrideC)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideA_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB1_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideC_);
+        }
+
+        private:
+        index_t BatchStrideA_;
+        index_t BatchStrideB_;
+        index_t BatchStrideB1_;
+        index_t BatchStrideC_;
+    };
+
+    using AGridDesc_AK0_M_AK1  = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
+    using BGridDesc_BK0_N_BK1  = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
+    using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1(1, 1, 1));
+    using CGridDesc_M_N        = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+
+    using C0MatrixMask = conditional_t<MaskOutUpperTriangle,
+                                       C0MatrixMask_impl<MaskOutUpperTrianglePredicate>,
+                                       C0MatrixMask_impl<MaskDisabledPredicate>>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        AccElementwiseOperation,
+        B1ElementwiseOperation,
+        CElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        B1GridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        Gemm1NPerBlock,
+        Gemm1KPerBlock,
+        AK1,
+        BK1,
+        B1K1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        Gemm1NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        true,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        true,
+        BBlockLdsExtraN,
+        B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+        B1BlockTransferThreadClusterArrangeOrder,
+        B1BlockTransferSrcAccessOrder,
+        B1BlockTransferSrcVectorDim,
+        B1BlockTransferSrcScalarPerVector,
+        B1BlockTransferDstScalarPerVector_BK1,
+        false,
+        B1BlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched,
+        matrix_padder.PadN,
+        MaskOutUpperTriangle>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 const B1DataType* p_b1_grid,
+                 CDataType* p_c_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t Gemm1NRaw, // = ORaw
+                 index_t Batch,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideB1,
+                 index_t StrideC,
+                 index_t BatchStrideA,
+                 index_t BatchStrideB,
+                 index_t BatchStrideB1,
+                 index_t BatchStrideC,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 AccElementwiseOperation acc_element_op,
+                 B1ElementwiseOperation b1_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_b1_grid_{p_b1_grid},
+              p_c_grid_{p_c_grid},
+              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
+              b1_grid_desc_bk0_n_bk1_{
+                  DeviceOp::MakeB1GridDescriptor_BK0_N_BK1(NRaw, Gemm1NRaw, StrideB1)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, Gemm1NRaw, StrideC)},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              acc_element_op_{acc_element_op},
+              b1_element_op_{b1_element_op},
+              c_element_op_{c_element_op},
+              batch_count_(Batch),
+              compute_base_ptr_of_batch_{BatchStrideA, BatchStrideB, BatchStrideB1, BatchStrideC},
+              c0_matrix_mask_{NRaw},
+              raw_lengths_m_n_k_o_{MRaw, NRaw, KRaw, Gemm1NRaw}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           b1_grid_desc_bk0_n_bk1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        const B1DataType* p_b1_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        AccElementwiseOperation acc_element_op_;
+        B1ElementwiseOperation b1_element_op_;
+        CElementwiseOperation c_element_op_;
+        index_t batch_count_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
+
+        // check C0 masking and padding
+        C0MatrixMask c0_matrix_mask_;
+
+        // For robust IsSupportedArgument() check
+        std::vector<index_t> raw_lengths_m_n_k_o_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                            arg.b_grid_desc_bk0_n_bk1_,
+                                            arg.b1_grid_desc_bk0_n_bk1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.batch_count_;
+
+            // Gemm0_K
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel = kernel_batched_gemm_softmax_gemm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    AccElementwiseOperation,
+                    B1ElementwiseOperation,
+                    CElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    DeviceOp::B1GridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    ComputeBasePtrOfStridedBatch,
+                    C0MatrixMask,
+                    has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_b1_grid_,
+                                              arg.p_c_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.acc_element_op_,
+                                              arg.b1_element_op_,
+                                              arg.c_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.b1_grid_desc_bk0_n_bk1_,
+                                              arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_ctile_map_,
+                                              arg.batch_count_,
+                                              arg.compute_base_ptr_of_batch_,
+                                              arg.c0_matrix_mask_);
+            };
+
+            // Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
+            // to concern Gemm0's loop
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        // Note: we need raw lengths since threadwise copy can not handle vector load when part of
+        // vector is out of bounds
+        const auto MRaw      = arg.raw_lengths_m_n_k_o_[0];
+        const auto NRaw      = arg.raw_lengths_m_n_k_o_[1];
+        const auto KRaw      = arg.raw_lengths_m_n_k_o_[2];
+        const auto Gemm1NRaw = arg.raw_lengths_m_n_k_o_[3];
+
+        // Check scalar per vector requirement
+        const auto a_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, ALayout> ? KRaw : MRaw;
+        const auto b_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, BLayout> ? NRaw : KRaw;
+        const auto b1_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, B1Layout> ? Gemm1NRaw : NRaw;
+        const auto c_extent_lowest =
+            is_same_v<tensor_layout::gemm::RowMajor, CLayout> ? Gemm1NRaw : MRaw;
+
+        if(!(a_extent_lowest % ABlockTransferSrcScalarPerVector == 0 &&
+             b_extent_lowest % BBlockTransferSrcScalarPerVector == 0 &&
+             b1_extent_lowest % B1BlockTransferSrcScalarPerVector == 0 &&
+             c_extent_lowest % CShuffleBlockTransferScalarPerVector_NPerBlock == 0))
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.b1_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             const B1DataType* p_b1,
+                             CDataType* p_c,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t Gemm1NRaw,
+                             index_t Batch,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideB1,
+                             index_t StrideC,
+                             index_t BatchStrideA,
+                             index_t BatchStrideB,
+                             index_t BatchStrideB1,
+                             index_t BatchStrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             AccElementwiseOperation acc_element_op,
+                             B1ElementwiseOperation b1_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,           p_b,          p_b1,         p_c,          MRaw,
+                        NRaw,          KRaw,         Gemm1NRaw,    Batch,        StrideA,
+                        StrideB,       StrideB1,     StrideC,      BatchStrideA, BatchStrideB,
+                        BatchStrideB1, BatchStrideC, a_element_op, b_element_op, acc_element_op,
+                        b1_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      const void* p_b1,
+                                                      void* p_c,
+                                                      index_t MRaw,
+                                                      index_t NRaw,
+                                                      index_t KRaw,
+                                                      index_t Gemm1NRaw,
+                                                      index_t Batch,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideB1,
+                                                      index_t StrideC,
+                                                      index_t BatchStrideA,
+                                                      index_t BatchStrideB,
+                                                      index_t BatchStrideB1,
+                                                      index_t BatchStrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      AccElementwiseOperation acc_element_op,
+                                                      B1ElementwiseOperation b1_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<const B1DataType*>(p_b1),
+                                          static_cast<CDataType*>(p_c),
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          Gemm1NRaw,
+                                          Batch,
+                                          StrideA,
+                                          StrideB,
+                                          StrideB1,
+                                          StrideC,
+                                          BatchStrideA,
+                                          BatchStrideB,
+                                          BatchStrideB1,
+                                          BatchStrideC,
+                                          a_element_op,
+                                          b_element_op,
+                                          acc_element_op,
+                                          b1_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerBlock << ", "
+            << Gemm1NPerBlock << ", "
+            << Gemm1KPerBlock << ", "
+            << B1K1 << ", "
+            << getGemmSpecializationString(GemmSpec) << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp
new file mode 100644
index 0000000000..408c56d490
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp
@@ -0,0 +1,435 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2CTileMap Block2CTileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
+ * DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
+ * pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2CTileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename DeviceOp, typename GridwiseGemm, bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_xdlops_v2r3(const typename DeviceOp::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / karg.Batch);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(karg.compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(karg.compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(karg.compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const auto a_grid_desc_k0_m_k1 =
+        amd_wave_read_first_lane(GridwiseGemm::MakeAGridDescriptor_K0_M_K1(
+            karg.M, karg.MPadded, karg.K, karg.K0, karg.StrideA));
+    const auto b_grid_desc_k0_n_k1 =
+        amd_wave_read_first_lane(GridwiseGemm::MakeBGridDescriptor_K0_N_K1(
+            karg.K, karg.N, karg.NPadded, karg.K0, karg.StrideB));
+    const auto c_grid_desc_m_n = amd_wave_read_first_lane(GridwiseGemm::MakeCGridDescriptor_M_N(
+        karg.M, karg.MPadded, karg.N, karg.NPadded, karg.StrideC));
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(karg.p_a_grid + a_batch_offset,
+                                                  karg.p_b_grid + b_batch_offset,
+                                                  karg.p_c_grid + c_batch_offset,
+                                                  p_shared,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  c_grid_desc_m_n);
+#else
+    ignore = karg;
+#endif
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector,
+          ck::index_t NumGemmKPrefetchStage = 1,
+          ck::LoopScheduler LoopSched       = make_default_loop_scheduler(),
+          ck::PipelineVersion PipelineVer   = ck::PipelineVersion::v1>
+struct DeviceBatchedGemmXdl : public DeviceBatchedGemm<ALayout,
+                                                       BLayout,
+                                                       CLayout,
+                                                       ADataType,
+                                                       BDataType,
+                                                       CDataType,
+                                                       AElementwiseOperation,
+                                                       BElementwiseOperation,
+                                                       CElementwiseOperation>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto K1Number = Number<K1>{};
+
+    struct ComputePtrOffsetOfStridedBatch
+    {
+        ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                       index_t BatchStrideB,
+                                       index_t BatchStrideC)
+            : BatchStrideA_(BatchStrideA), BatchStrideB_(BatchStrideB), BatchStrideC_(BatchStrideC)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideA_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideB_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(BatchStrideC_);
+        }
+
+        private:
+        index_t BatchStrideA_;
+        index_t BatchStrideB_;
+        index_t BatchStrideC_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        ALayout,
+        BLayout,
+        CLayout,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        GemmSpecialization::MNPadding,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        Sequence<2, 3, 0, 1, 7, 5, 4, 6>,
+        CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector,
+        NumGemmKPrefetchStage,
+        LoopSched,
+        PipelineVer>;
+
+    using Problem = typename GridwiseGemm::Problem;
+
+    // Argument
+    struct Argument : public Problem, public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid_,
+                 const BDataType* p_b_grid_,
+                 CDataType* p_c_grid_,
+                 index_t M_,
+                 index_t N_,
+                 index_t K_,
+                 index_t StrideA_,
+                 index_t StrideB_,
+                 index_t StrideC_,
+                 index_t BatchStrideA,
+                 index_t BatchStrideB,
+                 index_t BatchStrideC,
+                 index_t Batch_)
+            : Problem{M_, N_, K_, StrideA_, StrideB_, StrideC_},
+              p_a_grid{p_a_grid_},
+              p_b_grid{p_b_grid_},
+              p_c_grid{p_c_grid_},
+              Batch(Batch_),
+              compute_ptr_offset_of_batch{BatchStrideA, BatchStrideB, BatchStrideC}
+        {
+        }
+
+        const ADataType* p_a_grid;
+        const BDataType* p_b_grid;
+        CDataType* p_c_grid;
+        index_t Batch;
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceBatchedGemmXdl::Argument;
+
+        float Run(const Argument& karg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                karg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(karg))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext has invalid setting");
+            }
+
+            auto [gdx, gdy, gdz] = GridwiseGemm::CalculateGridSize(karg.M, karg.N);
+            gdx *= karg.Batch;
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(karg.K))
+            {
+                const auto kernel =
+                    kernel_batched_gemm_xdlops_v2r3<DeviceBatchedGemmXdl, GridwiseGemm, true>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg);
+            }
+            else
+            {
+                const auto kernel =
+                    kernel_batched_gemm_xdlops_v2r3<DeviceBatchedGemmXdl, GridwiseGemm, false>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Problem& problem)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if(problem.K % K1 != 0)
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(problem);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             index_t BatchStrideA,
+                             index_t BatchStrideB,
+                             index_t BatchStrideC,
+                             index_t Batch)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        BatchStrideA,
+                        BatchStrideB,
+                        BatchStrideC,
+                        Batch};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      index_t BatchStrideA,
+                                                      index_t BatchStrideB,
+                                                      index_t BatchStrideC,
+                                                      index_t Batch,
+                                                      AElementwiseOperation,
+                                                      BElementwiseOperation,
+                                                      CElementwiseOperation) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          BatchStrideA,
+                                          BatchStrideB,
+                                          BatchStrideC,
+                                          Batch);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
+        // clang-format off
+        str << "DeviceBatchedGemmXdl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock
+            << ">"
+            << " NumGemmKPrefetchStage: "
+            << NumGemmKPrefetchStage << ", "
+            << "LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp
new file mode 100644
index 0000000000..f46237e005
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp
@@ -0,0 +1,874 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp"
+#include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp"
+#include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/welford_helper.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename XDataType,
+          typename DxDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim,
+          bool UseMultiblockInK,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XDyDxVectorDim,
+          index_t XSrcVectorSize,
+          index_t DySrcVectorSize,
+          index_t DxDstVectorSize,
+          index_t ScaleSrcVectorSize,
+          index_t DscaleDbiasDstVectorSize,
+          index_t MeanVarSrcVectorSize>
+struct DeviceBatchNormBwdImpl : public DeviceBatchNormBwd<XDataType,
+                                                          DxDataType,
+                                                          DyDataType,
+                                                          AccDataType,
+                                                          ScaleDataType,
+                                                          DscaleDbiasDataType,
+                                                          MeanVarDataType,
+                                                          DyElementwiseOp,
+                                                          Rank,
+                                                          NumBatchNormReduceDim>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
+                  "Invalid thread cluster size assignments!");
+
+    static_assert((XDyDxVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0 &&
+                   MThreadSliceSize % DySrcVectorSize == 0 &&
+                   MThreadSliceSize % DxDstVectorSize == 0) ||
+                      (XDyDxVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0 &&
+                       KThreadSliceSize % DySrcVectorSize == 0 &&
+                       KThreadSliceSize % DxDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto MakeXY2dDescriptor(const std::array<index_t, Rank>& xyLengths,
+                                   const std::array<index_t, Rank>& xyStrides,
+                                   int blkGroupSize,
+                                   int numBlockTileIteration)
+    {
+        const auto tupleXYLengths =
+            generate_tuple([&](auto I) { return xyLengths[I]; }, Number<Rank>{});
+        const auto tupleXYStrides =
+            generate_tuple([&](auto I) { return xyStrides[I]; }, Number<Rank>{});
+
+        const auto raw_grid_desc = make_naive_tensor_descriptor(tupleXYLengths, tupleXYStrides);
+
+        const auto grid_desc_m_k = [&]() {
+            using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+            using ReduceDims    = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+            const auto reduceDimLengths =
+                generate_tuple([&](auto I) { return xyLengths[NumInvariantDim + I]; },
+                               Number<NumBatchNormReduceDim>{});
+            const auto invariantDimLengths =
+                generate_tuple([&](auto I) { return xyLengths[I]; }, Number<NumInvariantDim>{});
+
+            return transform_tensor_descriptor(raw_grid_desc,
+                                               make_tuple(make_merge_transform(invariantDimLengths),
+                                                          make_merge_transform(reduceDimLengths)),
+                                               make_tuple(InvariantDims{}, ReduceDims{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }();
+
+        const auto invariantLength = grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = grid_desc_m_k.GetLength(Number<1>{});
+
+        const int workSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto kPad = workSizePerBlock * blkGroupSize - reduceLength;
+
+        auto grid_desc_m_k_padded =
+            transform_tensor_descriptor(grid_desc_m_k,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_right_pad_transform(reduceLength, kPad)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_k_padded);
+    };
+
+    static auto MakeMultiblockFirstReduceOutputMG2dDescriptor(int invariantLength, int blkGroupSize)
+    {
+        const auto grid_desc_m_g =
+            make_naive_tensor_descriptor_packed(make_tuple(invariantLength, blkGroupSize));
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto grid_desc_m_g_padded =
+            transform_tensor_descriptor(grid_desc_m_g,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_pass_through_transform(blkGroupSize)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_g_padded);
+    };
+
+    static auto MakeMultiblockFinalReduceInputMK2dDescriptor(int invariantLength, int blkGroupSize)
+    {
+        const auto reduceLength = blkGroupSize;
+        const auto grid_desc_m_k =
+            make_naive_tensor_descriptor_packed(make_tuple(invariantLength, reduceLength));
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto kPad =
+            math::integer_least_multiple(reduceLength, KThreadClusterSize) - reduceLength;
+
+        auto grid_desc_m_k_padded =
+            transform_tensor_descriptor(grid_desc_m_k,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_right_pad_transform(reduceLength, kPad)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_k_padded);
+    };
+
+    static auto
+    MakeScaleBiasMeanVar1dDescriptor(const std::array<index_t, NumInvariantDim>& lengths,
+                                     const std::array<index_t, NumInvariantDim>& strides)
+    {
+        const auto tupleLengths =
+            generate_tuple([&](auto I) { return lengths[I]; }, Number<NumInvariantDim>{});
+        const auto tupleStrides =
+            generate_tuple([&](auto I) { return strides[I]; }, Number<NumInvariantDim>{});
+
+        auto raw_grid_desc = make_naive_tensor_descriptor(tupleLengths, tupleStrides);
+
+        auto grid_desc_m = transform_tensor_descriptor(
+            raw_grid_desc,
+            make_tuple(make_merge_transform(tupleLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = grid_desc_m.GetLength(Number<0>{});
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto grid_desc_m_padded =
+            transform_tensor_descriptor(grid_desc_m,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return (grid_desc_m_padded);
+    };
+
+    using XYGridDesc_M_K      = decltype(MakeXY2dDescriptor({1}, {1}, 1, 1));
+    using ScaleBiasGridDesc_M = decltype(MakeScaleBiasMeanVar1dDescriptor({1}, {1}));
+    using MeanVarGridDesc_M   = ScaleBiasGridDesc_M;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> xyLengths,
+                 const std::array<index_t, Rank> xStrides,
+                 const std::array<index_t, Rank> dyStrides,
+                 const std::array<index_t, Rank> dxStrides,
+                 const std::array<int, NumBatchNormReduceDim> reduceDims,
+                 const std::array<ck::index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                 const std::array<ck::index_t, NumInvariantDim> bnScaleStrides,
+                 const std::array<ck::index_t, NumInvariantDim> bnDscaleDbiasStrides,
+                 const std::array<ck::index_t, NumInvariantDim> bnMeanVarStrides,
+                 const XDataType* p_x,
+                 const DyDataType* p_dy,
+                 const ScaleDataType* p_scale,
+                 const MeanVarDataType* p_savedMean,
+                 const MeanVarDataType* p_savedInvVar,
+                 const DyElementwiseOp dy_elementwise_op,
+                 double epsilon,
+                 DxDataType* p_dx,
+                 DscaleDbiasDataType* p_dscale,
+                 DscaleDbiasDataType* p_dbias)
+            : bnScaleBiasMeanVarLengths_(bnScaleBiasMeanVarLengths),
+              bnScaleStrides_(bnScaleStrides),
+              bnDscaleDbiasStrides_(bnDscaleDbiasStrides),
+              bnMeanVarStrides_(bnMeanVarStrides),
+              p_x_(p_x),
+              p_dy_(p_dy),
+              p_scale_(p_scale),
+              p_savedMean_(p_savedMean),
+              p_savedInvVar_(p_savedInvVar),
+              dy_elementwise_op_(dy_elementwise_op),
+              p_dx_(p_dx),
+              p_dscale_(p_dscale),
+              p_dbias_(p_dbias)
+        {
+            xyLengths_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(xyLengths, reduceDims);
+            xStrides_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(xStrides, reduceDims);
+            dyStrides_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(dyStrides, reduceDims);
+            dxStrides_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(dxStrides, reduceDims);
+
+            std::tie(invariant_length, reduce_length) =
+                get_2d_lengths<Rank, NumBatchNormReduceDim>(xyLengths_);
+
+            epsilon_ = type_convert<AccDataType>(epsilon);
+
+            haveSavedMeanInvVar_ = (p_savedMean_ != nullptr && p_savedInvVar_ != nullptr);
+
+            if(UseMultiblockInK)
+            {
+                int iterations = 1;
+                while(true)
+                {
+                    int testBlkGroupSize = (reduce_length + (K_BlockTileSize * iterations) - 1) /
+                                           (K_BlockTileSize * iterations);
+
+                    // we want the blkGroupSize be not more than 128
+                    if(testBlkGroupSize <= 128)
+                        break;
+
+                    iterations++;
+                };
+
+                blkGroupSize = (reduce_length + (K_BlockTileSize * iterations) - 1) /
+                               (K_BlockTileSize * iterations);
+
+                numBlockTileIteration = iterations;
+            }
+            else
+            {
+                blkGroupSize          = 1;
+                numBlockTileIteration = (reduce_length + K_BlockTileSize - 1) / K_BlockTileSize;
+            };
+
+            gridSize = (invariant_length + M_BlockTileSize - 1) / M_BlockTileSize * blkGroupSize;
+
+            x_grid_desc_m_k =
+                MakeXY2dDescriptor(xyLengths_, xStrides_, blkGroupSize, numBlockTileIteration);
+            dy_grid_desc_m_k =
+                MakeXY2dDescriptor(xyLengths_, dyStrides_, blkGroupSize, numBlockTileIteration);
+            dx_grid_desc_m_k =
+                MakeXY2dDescriptor(xyLengths_, dxStrides_, blkGroupSize, numBlockTileIteration);
+            scale_grid_desc_m =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnScaleStrides);
+            dscale_dbias_grid_desc_m =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnDscaleDbiasStrides);
+            mean_var_grid_desc_m =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnMeanVarStrides);
+        }
+
+        AccDataType epsilon_;
+
+        bool haveSavedMeanInvVar_;
+
+        std::array<index_t, Rank> xyLengths_;
+        std::array<index_t, Rank> xStrides_;
+        std::array<index_t, Rank> dyStrides_;
+        std::array<index_t, Rank> dxStrides_;
+
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnDscaleDbiasStrides_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides_;
+
+        const XDataType* p_x_;
+        const DyDataType* p_dy_;
+        const ScaleDataType* p_scale_;
+        const MeanVarDataType* p_savedMean_;
+        const MeanVarDataType* p_savedInvVar_;
+        const DyElementwiseOp dy_elementwise_op_;
+        DxDataType* p_dx_;
+        DscaleDbiasDataType* p_dscale_;
+        DscaleDbiasDataType* p_dbias_;
+
+        long_index_t invariant_length;
+        long_index_t reduce_length;
+
+        int blkGroupSize;
+        int numBlockTileIteration;
+        size_t gridSize;
+
+        XYGridDesc_M_K x_grid_desc_m_k;
+        XYGridDesc_M_K dy_grid_desc_m_k;
+        XYGridDesc_M_K dx_grid_desc_m_k;
+        ScaleBiasGridDesc_M scale_grid_desc_m;
+        ScaleBiasGridDesc_M dscale_dbias_grid_desc_m;
+        MeanVarGridDesc_M mean_var_grid_desc_m;
+
+        void* workspace_mean;
+        void* workspace_variance;
+        void* workspace_count;
+
+        void* workspace_savedMean;
+        void* workspace_savedInvVar;
+
+        void* workspace_reduce_dscale;
+        void* workspace_reduce_dbias;
+    };
+
+    size_t GetWorkSpaceSize(const BaseArgument* pArg) const override
+    {
+        const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
+
+        size_t workspace_size = 0;
+
+        if(UseMultiblockInK && pArg_->blkGroupSize > 1)
+        {
+            // workspace for the partial reduced result for dscale
+            workspace_size +=
+                pArg_->invariant_length * pArg_->blkGroupSize * sizeof(DscaleDbiasDataType) + 64;
+
+            // workspace for the partial reduced result for dbias
+            workspace_size +=
+                pArg_->invariant_length * pArg_->blkGroupSize * sizeof(DscaleDbiasDataType) + 64;
+
+            if(!pArg_->haveSavedMeanInvVar_)
+            {
+                // workspace for welford intermediate mean
+                workspace_size +=
+                    pArg_->invariant_length * pArg_->blkGroupSize * sizeof(MeanVarDataType) + 64;
+
+                // workspace for welford intermediate variance
+                workspace_size +=
+                    pArg_->invariant_length * pArg_->blkGroupSize * sizeof(MeanVarDataType) + 64;
+
+                // workspace for welford intermediate count
+                workspace_size +=
+                    pArg_->invariant_length * pArg_->blkGroupSize * sizeof(int32_t) + 64;
+
+                // workspace for welford result mean
+                workspace_size += pArg_->invariant_length * sizeof(MeanVarDataType) + 64;
+
+                // workspace for welford result inv_variance
+                workspace_size += pArg_->invariant_length * sizeof(MeanVarDataType) + 64;
+            };
+        }
+
+        return (workspace_size);
+    };
+
+    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    {
+        Argument* pArg_ = dynamic_cast<Argument*>(pArg);
+
+        pArg_->p_workspace_ = p_workspace;
+
+        index_t space_sz;
+
+        // setup buffer for the partial reduced result for dscale
+        pArg_->workspace_reduce_dscale = pArg_->p_workspace_;
+
+        space_sz = pArg_->invariant_length * pArg_->blkGroupSize * sizeof(DscaleDbiasDataType);
+        space_sz = math::integer_least_multiple(space_sz, 64);
+
+        // setup buffer for the partial reduced result for dbias
+        pArg_->workspace_reduce_dbias =
+            reinterpret_cast<char*>(pArg_->workspace_reduce_dscale) + space_sz;
+
+        if(UseMultiblockInK && pArg_->blkGroupSize > 1)
+        {
+            space_sz = pArg_->invariant_length * pArg_->blkGroupSize * sizeof(DscaleDbiasDataType);
+            space_sz = math::integer_least_multiple(space_sz, 64);
+
+            // setup buffer for welford intermediate mean
+            pArg_->workspace_mean =
+                reinterpret_cast<char*>(pArg_->workspace_reduce_dbias) + space_sz;
+
+            space_sz = pArg_->invariant_length * pArg_->blkGroupSize * sizeof(MeanVarDataType);
+            space_sz = math::integer_least_multiple(space_sz, 64);
+
+            // setup buffer for welford intermediate varirance
+            pArg_->workspace_variance = reinterpret_cast<char*>(pArg_->workspace_mean) + space_sz;
+
+            space_sz = pArg_->invariant_length * pArg_->blkGroupSize * sizeof(MeanVarDataType);
+            space_sz = math::integer_least_multiple(space_sz, 64);
+
+            // setup buffer for welford intermediate count
+            pArg_->workspace_count = reinterpret_cast<char*>(pArg_->workspace_variance) + space_sz;
+
+            space_sz = pArg_->invariant_length * pArg_->blkGroupSize * sizeof(int32_t);
+            space_sz = math::integer_least_multiple(space_sz, 64);
+
+            // setup buffer for welford result mean
+            pArg_->workspace_savedMean = reinterpret_cast<char*>(pArg_->workspace_count) + space_sz;
+
+            space_sz = pArg_->invariant_length * sizeof(MeanVarDataType);
+            space_sz = math::integer_least_multiple(space_sz, 64);
+
+            // setup buffer for welford result inv_variance
+            pArg_->workspace_savedInvVar =
+                reinterpret_cast<char*>(pArg_->workspace_savedMean) + space_sz;
+        };
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            float avg_time = 0;
+
+            const auto mean_var_count_grid_desc_m_g =
+                DeviceBatchNormBwdImpl::MakeMultiblockFirstReduceOutputMG2dDescriptor(
+                    arg.invariant_length, arg.blkGroupSize);
+
+            const auto dscale_dbias_grid_desc_m_g =
+                DeviceBatchNormBwdImpl::MakeMultiblockFirstReduceOutputMG2dDescriptor(
+                    arg.invariant_length, arg.blkGroupSize);
+
+            const auto mean_var_count_grid_desc_m_k =
+                DeviceBatchNormBwdImpl::MakeMultiblockFinalReduceInputMK2dDescriptor(
+                    arg.invariant_length, arg.blkGroupSize);
+
+            const auto dscale_dbias_grid_desc_m_k =
+                DeviceBatchNormBwdImpl::MakeMultiblockFinalReduceInputMK2dDescriptor(
+                    arg.invariant_length, arg.blkGroupSize);
+
+            using MeanVarCountGridDesc_M_G = decltype(mean_var_count_grid_desc_m_g);
+            using MeanVarCountGridDesc_M_K = decltype(mean_var_count_grid_desc_m_k);
+            using DscaleDbiasGridDesc_M_G  = decltype(dscale_dbias_grid_desc_m_g);
+            using DscaleDbiasGridDesc_M_K  = decltype(dscale_dbias_grid_desc_m_k);
+
+            using GridwiseWelfordSecondHalfReduceFirstHalf_ =
+                GridwiseWelfordSecondHalfReduceFirstHalf<XDataType,
+                                                         DyDataType,
+                                                         AccDataType,
+                                                         ScaleDataType,
+                                                         DscaleDbiasDataType,
+                                                         MeanVarDataType,
+                                                         DyElementwiseOp,
+                                                         XYGridDesc_M_K,
+                                                         MeanVarGridDesc_M,
+                                                         MeanVarCountGridDesc_M_K,
+                                                         DscaleDbiasGridDesc_M_G,
+                                                         BlockSize,
+                                                         MThreadClusterSize,
+                                                         KThreadClusterSize,
+                                                         MThreadSliceSize,
+                                                         KThreadSliceSize,
+                                                         XDyDxVectorDim,
+                                                         XSrcVectorSize,
+                                                         DySrcVectorSize,
+                                                         MeanVarSrcVectorSize>;
+
+            using GridwiseReduceSecondHalfBatchNormBwdFinal_ =
+                GridwiseReduceSecondHalfBatchNormBackwardFinal<XDataType,
+                                                               DyDataType,
+                                                               DxDataType,
+                                                               AccDataType,
+                                                               ScaleDataType,
+                                                               DscaleDbiasDataType,
+                                                               MeanVarDataType,
+                                                               DyElementwiseOp,
+                                                               XYGridDesc_M_K,
+                                                               DscaleDbiasGridDesc_M_K,
+                                                               MeanVarGridDesc_M,
+                                                               ScaleBiasGridDesc_M,
+                                                               BlockSize,
+                                                               MThreadClusterSize,
+                                                               KThreadClusterSize,
+                                                               MThreadSliceSize,
+                                                               KThreadSliceSize,
+                                                               XDyDxVectorDim,
+                                                               XSrcVectorSize,
+                                                               DySrcVectorSize,
+                                                               DxDstVectorSize,
+                                                               ScaleSrcVectorSize,
+                                                               DscaleDbiasDstVectorSize,
+                                                               MeanVarSrcVectorSize>;
+
+            if(UseMultiblockInK && arg.blkGroupSize > 1)
+            {
+                using GetReduceCountPerThreadFunctor =
+                    GetReduceCountPerThreadForMultiblockWelford<K_BlockTileSize, KThreadSliceSize>;
+
+                GetReduceCountPerThreadFunctor get_reduce_count_per_thread(
+                    arg.blkGroupSize, arg.numBlockTileIteration, arg.reduce_length);
+
+                if(!arg.haveSavedMeanInvVar_)
+                {
+                    using GridwiseMultiblockWelfordFirstHalf_ =
+                        GridwiseMultiblockWelfordFirstHalf<XDataType,
+                                                           AccDataType,
+                                                           MeanVarDataType,
+                                                           XYGridDesc_M_K,
+                                                           MeanVarCountGridDesc_M_G,
+                                                           GetReduceCountPerThreadFunctor,
+                                                           BlockSize,
+                                                           MThreadClusterSize,
+                                                           KThreadClusterSize,
+                                                           MThreadSliceSize,
+                                                           KThreadSliceSize,
+                                                           XDyDxVectorDim,
+                                                           XSrcVectorSize>;
+
+                    const auto kern_multiblock_welford_first_half =
+                        kernel_multiblock_welford_first_half<GridwiseMultiblockWelfordFirstHalf_,
+                                                             XDataType,
+                                                             MeanVarDataType,
+                                                             XYGridDesc_M_K,
+                                                             MeanVarCountGridDesc_M_G,
+                                                             GetReduceCountPerThreadFunctor>;
+
+                    avg_time += launch_and_time_kernel(
+                        stream_config,
+                        kern_multiblock_welford_first_half,
+                        dim3(arg.gridSize),
+                        dim3(BlockSize),
+                        0,
+                        arg.x_grid_desc_m_k,
+                        mean_var_count_grid_desc_m_g,
+                        get_reduce_count_per_thread,
+                        arg.numBlockTileIteration,
+                        arg.p_x_,
+                        static_cast<MeanVarDataType*>(arg.workspace_mean),
+                        static_cast<MeanVarDataType*>(arg.workspace_variance),
+                        static_cast<int32_t*>(arg.workspace_count));
+                };
+
+                const auto kern_welford_second_half_reduce_first_half =
+                    kernel_welford_second_half_reduce_first_half<
+                        GridwiseWelfordSecondHalfReduceFirstHalf_,
+                        XDataType,
+                        DyDataType,
+                        AccDataType,
+                        ScaleDataType,
+                        DscaleDbiasDataType,
+                        MeanVarDataType,
+                        DyElementwiseOp,
+                        XYGridDesc_M_K,
+                        MeanVarGridDesc_M,
+                        MeanVarCountGridDesc_M_K,
+                        DscaleDbiasGridDesc_M_G>;
+
+                const auto kern_reduce_second_half_batchnorm_backward_final =
+                    kernel_reduce_second_half_batchnorm_backward_final<
+                        GridwiseReduceSecondHalfBatchNormBwdFinal_,
+                        XDataType,
+                        DyDataType,
+                        DxDataType,
+                        ScaleDataType,
+                        DscaleDbiasDataType,
+                        MeanVarDataType,
+                        DyElementwiseOp,
+                        XYGridDesc_M_K,
+                        DscaleDbiasGridDesc_M_K,
+                        MeanVarGridDesc_M,
+                        ScaleBiasGridDesc_M>;
+
+                index_t numDscaleDbiasBlockTileIteration =
+                    (arg.blkGroupSize + KThreadClusterSize - 1) / KThreadClusterSize;
+
+                avg_time += launch_and_time_kernel(
+                    stream_config,
+                    kern_welford_second_half_reduce_first_half,
+                    dim3(arg.gridSize),
+                    dim3(BlockSize),
+                    0,
+                    arg.x_grid_desc_m_k,
+                    arg.dy_grid_desc_m_k,
+                    arg.mean_var_grid_desc_m,
+                    mean_var_count_grid_desc_m_k,
+                    dscale_dbias_grid_desc_m_g,
+                    arg.blkGroupSize,
+                    arg.numBlockTileIteration,
+                    numDscaleDbiasBlockTileIteration,
+                    arg.epsilon_,
+                    arg.haveSavedMeanInvVar_,
+                    arg.haveSavedMeanInvVar_ ? arg.p_savedMean_ : nullptr,
+                    arg.haveSavedMeanInvVar_ ? arg.p_savedInvVar_ : nullptr,
+                    arg.haveSavedMeanInvVar_
+                        ? nullptr
+                        : static_cast<const MeanVarDataType*>(arg.workspace_mean),
+                    arg.haveSavedMeanInvVar_
+                        ? nullptr
+                        : static_cast<const MeanVarDataType*>(arg.workspace_variance),
+                    arg.haveSavedMeanInvVar_ ? nullptr
+                                             : static_cast<const int32_t*>(arg.workspace_count),
+                    arg.dy_elementwise_op_,
+                    arg.haveSavedMeanInvVar_
+                        ? nullptr
+                        : static_cast<MeanVarDataType*>(arg.workspace_savedMean),
+                    arg.haveSavedMeanInvVar_
+                        ? nullptr
+                        : static_cast<MeanVarDataType*>(arg.workspace_savedInvVar),
+                    arg.p_x_,
+                    arg.p_dy_,
+                    static_cast<DscaleDbiasDataType*>(arg.workspace_reduce_dscale),
+                    static_cast<DscaleDbiasDataType*>(arg.workspace_reduce_dbias));
+
+                avg_time += launch_and_time_kernel(
+                    stream_config,
+                    kern_reduce_second_half_batchnorm_backward_final,
+                    dim3(arg.gridSize),
+                    dim3(BlockSize),
+                    0,
+                    arg.x_grid_desc_m_k,
+                    arg.dy_grid_desc_m_k,
+                    arg.dx_grid_desc_m_k,
+                    dscale_dbias_grid_desc_m_k,
+                    arg.mean_var_grid_desc_m,
+                    arg.scale_grid_desc_m,
+                    arg.dscale_dbias_grid_desc_m,
+                    arg.blkGroupSize,
+                    arg.reduce_length,
+                    arg.numBlockTileIteration,
+                    numDscaleDbiasBlockTileIteration,
+                    static_cast<const DscaleDbiasDataType*>(arg.workspace_reduce_dscale),
+                    static_cast<const DscaleDbiasDataType*>(arg.workspace_reduce_dbias),
+                    arg.haveSavedMeanInvVar_
+                        ? arg.p_savedMean_
+                        : static_cast<const MeanVarDataType*>(arg.workspace_savedMean),
+                    arg.haveSavedMeanInvVar_
+                        ? arg.p_savedInvVar_
+                        : static_cast<const MeanVarDataType*>(arg.workspace_savedInvVar),
+                    arg.p_x_,
+                    arg.p_dy_,
+                    arg.p_scale_,
+                    arg.dy_elementwise_op_,
+                    arg.p_dx_,
+                    arg.p_dscale_,
+                    arg.p_dbias_);
+            }
+            else
+            {
+                using GetReduceCountPerThreadFunctor =
+                    GetReduceCountPerThreadForBlockwiseWelford<K_BlockTileSize, KThreadSliceSize>;
+
+                GetReduceCountPerThreadFunctor get_reduce_count_per_thread(
+                    arg.numBlockTileIteration, arg.reduce_length);
+
+                using GridwiseBatchNormBackwardWithBlockwiseWelford_ =
+                    GridwiseBatchNormBackwardWithBlockwiseWelford<XDataType,
+                                                                  DyDataType,
+                                                                  DxDataType,
+                                                                  AccDataType,
+                                                                  ScaleDataType,
+                                                                  DscaleDbiasDataType,
+                                                                  MeanVarDataType,
+                                                                  DyElementwiseOp,
+                                                                  XYGridDesc_M_K,
+                                                                  ScaleBiasGridDesc_M,
+                                                                  MeanVarGridDesc_M,
+                                                                  GetReduceCountPerThreadFunctor,
+                                                                  BlockSize,
+                                                                  MThreadClusterSize,
+                                                                  KThreadClusterSize,
+                                                                  MThreadSliceSize,
+                                                                  KThreadSliceSize,
+                                                                  XDyDxVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  DySrcVectorSize,
+                                                                  DxDstVectorSize,
+                                                                  ScaleSrcVectorSize,
+                                                                  DscaleDbiasDstVectorSize,
+                                                                  MeanVarSrcVectorSize>;
+
+                const auto kern_batchnorm_bwd = kernel_batchnorm_backward_with_blockwise_welford<
+                    GridwiseBatchNormBackwardWithBlockwiseWelford_,
+                    XDataType,
+                    DyDataType,
+                    DxDataType,
+                    AccDataType,
+                    ScaleDataType,
+                    DscaleDbiasDataType,
+                    MeanVarDataType,
+                    DyElementwiseOp,
+                    XYGridDesc_M_K,
+                    ScaleBiasGridDesc_M,
+                    MeanVarGridDesc_M,
+                    GetReduceCountPerThreadFunctor>;
+
+                avg_time += launch_and_time_kernel(stream_config,
+                                                   kern_batchnorm_bwd,
+                                                   dim3(arg.gridSize),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.x_grid_desc_m_k,
+                                                   arg.dy_grid_desc_m_k,
+                                                   arg.dx_grid_desc_m_k,
+                                                   arg.scale_grid_desc_m,
+                                                   arg.dscale_dbias_grid_desc_m,
+                                                   arg.mean_var_grid_desc_m,
+                                                   get_reduce_count_per_thread,
+                                                   arg.reduce_length,
+                                                   arg.numBlockTileIteration,
+                                                   arg.epsilon_,
+                                                   arg.p_x_,
+                                                   arg.p_dy_,
+                                                   arg.p_scale_,
+                                                   arg.haveSavedMeanInvVar_,
+                                                   arg.p_savedMean_,
+                                                   arg.p_savedInvVar_,
+                                                   arg.dy_elementwise_op_,
+                                                   arg.p_dx_,
+                                                   arg.p_dscale_,
+                                                   arg.p_dbias_);
+            };
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* pArg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(pArg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* pArg) override
+    {
+        const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
+
+        if constexpr(XDyDxVectorDim == 0)
+        {
+            if(pArg_->xStrides_[NumInvariantDim - 1] != 1 ||
+               pArg_->dyStrides_[NumInvariantDim - 1] != 1 ||
+               pArg_->dxStrides_[NumInvariantDim - 1] != 1)
+                return false;
+
+            if(pArg_->xyLengths_[NumInvariantDim - 1] % XSrcVectorSize != 0 ||
+               pArg_->xyLengths_[NumInvariantDim - 1] % DySrcVectorSize != 0 ||
+               pArg_->xyLengths_[NumInvariantDim - 1] % DxDstVectorSize != 0)
+                return false;
+        }
+        else
+        {
+            if(pArg_->xStrides_[Rank - 1] != 1 || pArg_->dyStrides_[Rank - 1] != 1 ||
+               pArg_->dxStrides_[Rank - 1] != 1)
+                return false;
+
+            if(pArg_->xyLengths_[Rank - 1] % XSrcVectorSize != 0 ||
+               pArg_->xyLengths_[Rank - 1] % DySrcVectorSize != 0 ||
+               pArg_->xyLengths_[Rank - 1] % DxDstVectorSize != 0)
+                return false;
+        };
+
+        if(pArg_->bnScaleStrides_[NumInvariantDim - 1] != 1 && ScaleSrcVectorSize != 1)
+            return false;
+
+        if(pArg_->bnDscaleDbiasStrides_[NumInvariantDim - 1] != 1 && DscaleDbiasDstVectorSize != 1)
+            return false;
+
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % ScaleSrcVectorSize != 0)
+            return false;
+
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % DscaleDbiasDstVectorSize != 0)
+            return false;
+
+        if(pArg_->haveSavedMeanInvVar_)
+        {
+            if(pArg_->bnMeanVarStrides_[NumInvariantDim - 1] != 1 && MeanVarSrcVectorSize != 1)
+                return false;
+
+            if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % MeanVarSrcVectorSize != 0)
+                return false;
+        };
+
+        bool is_valid = true;
+
+        static_for<0, NumInvariantDim, 1>{}([&](auto I) {
+            if(pArg_->xyLengths_[I] != pArg_->bnScaleBiasMeanVarLengths_[I])
+                is_valid = false;
+        });
+
+        if(!is_valid)
+            return false;
+
+        return true;
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> xyLengths,
+                        const std::array<index_t, Rank> xStrides,
+                        const std::array<index_t, Rank> dyStrides,
+                        const std::array<index_t, Rank> dxStrides,
+                        const std::array<int, NumBatchNormReduceDim> reduceDims,
+                        const std::array<ck::index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                        const std::array<ck::index_t, NumInvariantDim> bnScaleStrides,
+                        const std::array<ck::index_t, NumInvariantDim> bnDscaleDbiasStrides,
+                        const std::array<ck::index_t, NumInvariantDim> bnMeanVarStrides,
+                        const void* p_x,
+                        const void* p_dy,
+                        const void* p_scale,
+                        const void* p_savedMean,
+                        const void* p_savedInvVar,
+                        double epsilon,
+                        const DyElementwiseOp dy_elementwise_op,
+                        void* p_dx,
+                        void* p_dscale,
+                        void* p_dbias) override
+    {
+        return std::make_unique<Argument>(xyLengths,
+                                          xStrides,
+                                          dyStrides,
+                                          dxStrides,
+                                          reduceDims,
+                                          bnScaleBiasMeanVarLengths,
+                                          bnScaleStrides,
+                                          bnDscaleDbiasStrides,
+                                          bnMeanVarStrides,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const DyDataType*>(p_dy),
+                                          static_cast<const ScaleDataType*>(p_scale),
+                                          static_cast<const MeanVarDataType*>(p_savedMean),
+                                          static_cast<const MeanVarDataType*>(p_savedInvVar),
+                                          dy_elementwise_op,
+                                          epsilon,
+                                          static_cast<DxDataType*>(p_dx),
+                                          static_cast<DscaleDbiasDataType*>(p_dscale),
+                                          static_cast<DscaleDbiasDataType*>(p_dbias));
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchNormBwdImpl<" << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "XDyDxVectorDim_" << XDyDxVectorDim  << ",";
+        str << "VectorSize_X" << XSrcVectorSize << "_scale_" << ScaleSrcVectorSize << "_bias_" << DscaleDbiasDstVectorSize << "_mean_var_" << MeanVarSrcVectorSize << "_Dx_" << DxDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+}; // namespace device
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp
new file mode 100644
index 0000000000..ad8e795603
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp
@@ -0,0 +1,822 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/device/welford_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_batchnorm_forward.hpp"
+#include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp"
+#include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/hip_check_error.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim,
+          bool UseMultiblockInK,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcYDstVectorDim,
+          index_t XSrcVectorSize,
+          index_t YDstVectorSize,
+          index_t ScaleSrcVectorSize,
+          index_t BiasSrcVectorSize,
+          index_t MeanVarSrcDstVectorSize>
+struct DeviceBatchNormFwdImpl : public DeviceBatchNormFwd<XDataType,
+                                                          YDataType,
+                                                          AccDataType,
+                                                          ScaleDataType,
+                                                          BiasDataType,
+                                                          MeanVarDataType,
+                                                          YElementwiseOp,
+                                                          Rank,
+                                                          NumBatchNormReduceDim>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
+                  "Invalid thread cluster size assignments!");
+
+    static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcYDstVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto MakeXY2dDescriptor(const std::array<index_t, Rank>& xyLengths,
+                                   const std::array<index_t, Rank>& xyStrides,
+                                   int blkGroupSize,
+                                   int numBlockTileIteration)
+    {
+        const auto tupleXYLengths =
+            generate_tuple([&](auto I) { return xyLengths[I]; }, Number<Rank>{});
+        const auto tupleXYStrides =
+            generate_tuple([&](auto I) { return xyStrides[I]; }, Number<Rank>{});
+
+        const auto raw_grid_desc = make_naive_tensor_descriptor(tupleXYLengths, tupleXYStrides);
+
+        const auto grid_desc_m_k = [&]() {
+            using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+            using ReduceDims    = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+            const auto reduceDimLengths =
+                generate_tuple([&](auto I) { return xyLengths[NumInvariantDim + I]; },
+                               Number<NumBatchNormReduceDim>{});
+            const auto invariantDimLengths =
+                generate_tuple([&](auto I) { return xyLengths[I]; }, Number<NumInvariantDim>{});
+
+            return transform_tensor_descriptor(raw_grid_desc,
+                                               make_tuple(make_merge_transform(invariantDimLengths),
+                                                          make_merge_transform(reduceDimLengths)),
+                                               make_tuple(InvariantDims{}, ReduceDims{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }();
+
+        const auto invariantLength = grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = grid_desc_m_k.GetLength(Number<1>{});
+
+        const int workSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto kPad = workSizePerBlock * blkGroupSize - reduceLength;
+
+        auto grid_desc_m_k_padded =
+            transform_tensor_descriptor(grid_desc_m_k,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_right_pad_transform(reduceLength, kPad)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_k_padded);
+    };
+
+    static auto MakeMeanVarCountOutputMG2dDescriptor(int invariantLength, int blkGroupSize)
+    {
+        const auto grid_desc_m_g = make_naive_tensor_descriptor(
+            make_tuple(invariantLength, blkGroupSize), make_tuple(1, invariantLength));
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto grid_desc_m_g_padded =
+            transform_tensor_descriptor(grid_desc_m_g,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_pass_through_transform(blkGroupSize)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_g_padded);
+    };
+
+    static auto MakeMeanVarCountInputMK2dDescriptor(int invariantLength, int blkGroupSize)
+    {
+        const auto reduceLength  = blkGroupSize;
+        const auto grid_desc_m_k = make_naive_tensor_descriptor(
+            make_tuple(invariantLength, reduceLength), make_tuple(1, invariantLength));
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto kPad =
+            math::integer_least_multiple(reduceLength, KThreadClusterSize) - reduceLength;
+
+        auto grid_desc_m_k_padded =
+            transform_tensor_descriptor(grid_desc_m_k,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_right_pad_transform(reduceLength, kPad)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_k_padded);
+    };
+
+    static auto
+    MakeScaleBiasMeanVar1dDescriptor(const std::array<index_t, NumInvariantDim>& lengths,
+                                     const std::array<index_t, NumInvariantDim>& strides)
+    {
+        const auto tupleLengths =
+            generate_tuple([&](auto I) { return lengths[I]; }, Number<NumInvariantDim>{});
+        const auto tupleStrides =
+            generate_tuple([&](auto I) { return strides[I]; }, Number<NumInvariantDim>{});
+
+        auto raw_grid_desc = make_naive_tensor_descriptor(tupleLengths, tupleStrides);
+
+        auto grid_desc_m = transform_tensor_descriptor(
+            raw_grid_desc,
+            make_tuple(make_merge_transform(tupleLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = grid_desc_m.GetLength(Number<0>{});
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto grid_desc_m_padded =
+            transform_tensor_descriptor(grid_desc_m,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return (grid_desc_m_padded);
+    };
+
+    using XYGridDesc_M_K             = decltype(MakeXY2dDescriptor({1}, {1}, 1, 1));
+    using ScaleBiasMeanVarGridDesc_M = decltype(MakeScaleBiasMeanVar1dDescriptor({1}, {1}));
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> xyLengths,
+                 const std::array<index_t, Rank> xStrides,
+                 const std::array<index_t, Rank> yStrides,
+                 const std::array<int, NumBatchNormReduceDim> reduceDims,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
+                 const XDataType* p_x,
+                 const ScaleDataType* p_scale,
+                 const BiasDataType* p_bias,
+                 const YElementwiseOp y_elementwise_op,
+                 double epsilon,
+                 YDataType* p_y,
+                 MeanVarDataType* resultSaveMean,
+                 MeanVarDataType* resultSaveInvVariance,
+                 double averageFactor,
+                 MeanVarDataType* resultRunningMean,
+                 MeanVarDataType* resultRunningVariance)
+            : bnScaleBiasMeanVarLengths_(bnScaleBiasMeanVarLengths),
+              bnScaleStrides_(bnScaleStrides),
+              bnBiasStrides_(bnBiasStrides),
+              bnMeanVarStrides_(bnMeanVarStrides),
+              p_x_(p_x),
+              p_scale_(p_scale),
+              p_bias_(p_bias),
+              y_elementwise_op_(y_elementwise_op),
+              p_y_(p_y),
+              resultSaveMean_(resultSaveMean),
+              resultSaveInvVariance_(resultSaveInvVariance),
+              resultRunningMean_(resultRunningMean),
+              resultRunningVariance_(resultRunningVariance)
+        {
+            xyLengths_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(xyLengths, reduceDims);
+            xStrides_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(xStrides, reduceDims);
+            yStrides_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(yStrides, reduceDims);
+
+            std::tie(invariant_length_, reduce_length_) =
+                get_2d_lengths<Rank, NumBatchNormReduceDim>(xyLengths_);
+
+            epsilon_       = type_convert<AccDataType>(epsilon);
+            averageFactor_ = type_convert<AccDataType>(averageFactor);
+
+            updateMovingAverage_ =
+                (resultRunningMean != nullptr && resultRunningVariance != nullptr);
+            saveMeanInvVariance_ = (resultSaveMean != nullptr && resultSaveInvVariance_ != nullptr);
+
+            if(UseMultiblockInK)
+            {
+                int iterations = 1;
+                while(true)
+                {
+                    int testBlkGroupSize = (reduce_length_ + (K_BlockTileSize * iterations) - 1) /
+                                           (K_BlockTileSize * iterations);
+
+                    // we want the blkGroupSize be not more than 16
+                    if(testBlkGroupSize <= 16)
+                        break;
+
+                    iterations++;
+                };
+
+                blkGroupSize_ = (reduce_length_ + (K_BlockTileSize * iterations) - 1) /
+                                (K_BlockTileSize * iterations);
+
+                numBlockTileIteration_ = iterations;
+            }
+            else
+            {
+                blkGroupSize_          = 1;
+                numBlockTileIteration_ = (reduce_length_ + K_BlockTileSize - 1) / K_BlockTileSize;
+            };
+
+            gridSize_ = (invariant_length_ + M_BlockTileSize - 1) / M_BlockTileSize * blkGroupSize_;
+
+            x_grid_desc_m_k_ =
+                MakeXY2dDescriptor(xyLengths_, xStrides_, blkGroupSize_, numBlockTileIteration_);
+            y_grid_desc_m_k_ =
+                MakeXY2dDescriptor(xyLengths_, yStrides_, blkGroupSize_, numBlockTileIteration_);
+            scale_grid_desc_m_ =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnScaleStrides_);
+            bias_grid_desc_m_ =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnBiasStrides_);
+            mean_var_grid_desc_m_ =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnMeanVarStrides_);
+        }
+
+        AccDataType epsilon_;
+        AccDataType averageFactor_;
+
+        bool updateMovingAverage_;
+        bool saveMeanInvVariance_;
+
+        std::array<index_t, Rank> xyLengths_;
+        std::array<index_t, Rank> xStrides_;
+        std::array<index_t, Rank> yStrides_;
+
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides_;
+
+        const XDataType* p_x_;
+        const ScaleDataType* p_scale_;
+        const BiasDataType* p_bias_;
+        const YElementwiseOp y_elementwise_op_;
+        YDataType* p_y_;
+
+        MeanVarDataType* resultSaveMean_;
+        MeanVarDataType* resultSaveInvVariance_;
+
+        MeanVarDataType* resultRunningMean_;
+        MeanVarDataType* resultRunningVariance_;
+
+        long_index_t invariant_length_;
+        long_index_t reduce_length_;
+
+        int blkGroupSize_;
+        int numBlockTileIteration_;
+        size_t gridSize_;
+
+        XYGridDesc_M_K x_grid_desc_m_k_;
+        XYGridDesc_M_K y_grid_desc_m_k_;
+        ScaleBiasMeanVarGridDesc_M scale_grid_desc_m_;
+        ScaleBiasMeanVarGridDesc_M bias_grid_desc_m_;
+        ScaleBiasMeanVarGridDesc_M mean_var_grid_desc_m_;
+
+        void* workspace_mean_;
+        void* workspace_variance_;
+        void* workspace_count_;
+
+        void* control_;
+    };
+
+    size_t GetWorkSpaceSize(const BaseArgument* pArg) const override
+    {
+        const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
+
+        size_t workspace_size = 0;
+
+        if(UseMultiblockInK && pArg_->blkGroupSize_ > 1)
+        {
+            // workspace for welford intermediate mean
+            workspace_size +=
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType) + 64;
+
+            // workspace for welford intermediate variance
+            workspace_size +=
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType) + 64;
+
+            // workspace for welford intermediate count
+            workspace_size +=
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(int32_t) + 64;
+
+            // workspace for barrier objects, each barrier object consists of two integers
+            // TODO: allocate barrier object memory globally to reuse it by other operators
+            workspace_size += (pArg_->invariant_length_ + M_BlockTileSize - 1) / M_BlockTileSize *
+                              sizeof(int) * 2;
+        }
+
+        return (workspace_size);
+    };
+
+    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    {
+        Argument* pArg_ = dynamic_cast<Argument*>(pArg);
+
+        pArg_->p_workspace_ = p_workspace;
+
+        if(UseMultiblockInK && pArg_->blkGroupSize_ > 1)
+        {
+            // setup buffer used for intermediate welford mean
+            pArg_->workspace_mean_ = static_cast<char*>(pArg_->p_workspace_);
+
+            index_t mean_space_sz =
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType);
+
+            mean_space_sz = math::integer_least_multiple(mean_space_sz, 64);
+
+            // setup buffer used for intermediate welford varirance
+            pArg_->workspace_variance_ =
+                reinterpret_cast<char*>(pArg_->workspace_mean_) + mean_space_sz;
+
+            index_t variance_space_sz =
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType);
+
+            variance_space_sz = math::integer_least_multiple(variance_space_sz, 64);
+
+            // setup buffer used for intermediate welfor count
+            pArg_->workspace_count_ =
+                reinterpret_cast<char*>(pArg_->workspace_variance_) + variance_space_sz;
+
+            index_t count_space_sz =
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(int32_t);
+
+            count_space_sz = math::integer_least_multiple(count_space_sz, 64);
+
+            pArg_->control_ = reinterpret_cast<char*>(pArg_->workspace_count_) + count_space_sz;
+
+            index_t control_space_sz = (pArg_->invariant_length_ + M_BlockTileSize - 1) /
+                                       M_BlockTileSize * sizeof(int) * 2;
+
+            hip_check_error(hipMemset(pArg_->control_, 0, control_space_sz));
+        };
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            float avg_time = 0;
+
+            if(UseMultiblockInK && arg.blkGroupSize_ > 1)
+            {
+                using GetReduceCountPerThreadFunctor =
+                    GetReduceCountPerThreadForMultiblockWelford<K_BlockTileSize, KThreadSliceSize>;
+
+                GetReduceCountPerThreadFunctor get_reduce_count_per_thread(
+                    arg.blkGroupSize_, arg.numBlockTileIteration_, arg.reduce_length_);
+
+                const auto mean_var_count_grid_desc_m_g =
+                    DeviceBatchNormFwdImpl::MakeMeanVarCountOutputMG2dDescriptor(
+                        arg.invariant_length_, arg.blkGroupSize_);
+
+                const auto mean_var_count_grid_desc_m_k =
+                    DeviceBatchNormFwdImpl::MakeMeanVarCountInputMK2dDescriptor(
+                        arg.invariant_length_, arg.blkGroupSize_);
+
+                using MeanVarCountGridDesc_M_G = decltype(mean_var_count_grid_desc_m_g);
+                using MeanVarCountGridDesc_M_K = decltype(mean_var_count_grid_desc_m_k);
+
+                using GridwiseMultiblockBatchNormForward_ =
+                    GridwiseMultiblockBatchNormForward<XDataType,
+                                                       YDataType,
+                                                       AccDataType,
+                                                       ScaleDataType,
+                                                       BiasDataType,
+                                                       MeanVarDataType,
+                                                       YElementwiseOp,
+                                                       XYGridDesc_M_K,
+                                                       MeanVarCountGridDesc_M_G,
+                                                       MeanVarCountGridDesc_M_K,
+                                                       ScaleBiasMeanVarGridDesc_M,
+                                                       ScaleBiasMeanVarGridDesc_M,
+                                                       GetReduceCountPerThreadFunctor,
+                                                       BlockSize,
+                                                       MThreadClusterSize,
+                                                       KThreadClusterSize,
+                                                       MThreadSliceSize,
+                                                       KThreadSliceSize,
+                                                       XSrcYDstVectorDim,
+                                                       XSrcVectorSize,
+                                                       YDstVectorSize,
+                                                       ScaleSrcVectorSize,
+                                                       BiasSrcVectorSize,
+                                                       MeanVarSrcDstVectorSize>;
+
+                using GridwiseMultiblockWelfordFirstHalf_ =
+                    GridwiseMultiblockWelfordFirstHalf<XDataType,
+                                                       AccDataType,
+                                                       MeanVarDataType,
+                                                       XYGridDesc_M_K,
+                                                       MeanVarCountGridDesc_M_G,
+                                                       GetReduceCountPerThreadFunctor,
+                                                       BlockSize,
+                                                       MThreadClusterSize,
+                                                       KThreadClusterSize,
+                                                       MThreadSliceSize,
+                                                       KThreadSliceSize,
+                                                       XSrcYDstVectorDim,
+                                                       XSrcVectorSize>;
+
+                using GridwiseWelfordSecondHalfBatchNormForwardFinal_ =
+                    GridwiseWelfordSecondHalfBatchNormForwardFinal<XDataType,
+                                                                   YDataType,
+                                                                   AccDataType,
+                                                                   ScaleDataType,
+                                                                   BiasDataType,
+                                                                   MeanVarDataType,
+                                                                   YElementwiseOp,
+                                                                   XYGridDesc_M_K,
+                                                                   MeanVarCountGridDesc_M_K,
+                                                                   ScaleBiasMeanVarGridDesc_M,
+                                                                   ScaleBiasMeanVarGridDesc_M,
+                                                                   BlockSize,
+                                                                   MThreadClusterSize,
+                                                                   KThreadClusterSize,
+                                                                   MThreadSliceSize,
+                                                                   KThreadSliceSize,
+                                                                   XSrcYDstVectorDim,
+                                                                   XSrcVectorSize,
+                                                                   YDstVectorSize,
+                                                                   ScaleSrcVectorSize,
+                                                                   BiasSrcVectorSize,
+                                                                   MeanVarSrcDstVectorSize>;
+
+                // It is found that:
+                // 1) gfx1030 does not support the GLC enabled vector load/store, so using the
+                //    two-kernel method for gfx1030
+                // 2) Profiler on gfx908 could hang even though it works when running examples
+                // 3) Single-kernel method works on gfx1100, but the performance it not better
+                //    than two-kernel method (due to more warps participating the barrier)
+                if(ck::get_device_name() == "gfx90a")
+                {
+                    const auto kern_multiblock_batchnorm_fwd_ =
+                        kernel_multiblock_batchnorm_forward<GridwiseMultiblockBatchNormForward_,
+                                                            XDataType,
+                                                            YDataType,
+                                                            AccDataType,
+                                                            ScaleDataType,
+                                                            BiasDataType,
+                                                            MeanVarDataType,
+                                                            YElementwiseOp,
+                                                            XYGridDesc_M_K,
+                                                            MeanVarCountGridDesc_M_G,
+                                                            MeanVarCountGridDesc_M_K,
+                                                            ScaleBiasMeanVarGridDesc_M,
+                                                            ScaleBiasMeanVarGridDesc_M,
+                                                            GetReduceCountPerThreadFunctor>;
+
+                    avg_time += launch_and_time_kernel(
+                        stream_config,
+                        kern_multiblock_batchnorm_fwd_,
+                        dim3(arg.gridSize_),
+                        dim3(BlockSize),
+                        0,
+                        arg.x_grid_desc_m_k_,
+                        arg.y_grid_desc_m_k_,
+                        mean_var_count_grid_desc_m_g, // for writing to mean/variance/count
+                                                      // workspace by multiple workgroups
+                        mean_var_count_grid_desc_m_k, // for reading from mean/variance/count
+                                                      // workspace by each workgroup
+                        arg.scale_grid_desc_m_,
+                        arg.bias_grid_desc_m_,
+                        arg.mean_var_grid_desc_m_,
+                        get_reduce_count_per_thread,
+                        arg.numBlockTileIteration_,
+                        arg.epsilon_,
+                        arg.p_x_,
+                        static_cast<MeanVarDataType*>(arg.workspace_mean_),
+                        static_cast<MeanVarDataType*>(arg.workspace_variance_),
+                        static_cast<int32_t*>(arg.workspace_count_),
+                        static_cast<int*>(arg.control_),
+                        arg.p_scale_,
+                        arg.p_bias_,
+                        arg.y_elementwise_op_,
+                        arg.p_y_,
+                        arg.updateMovingAverage_, // true or false
+                        arg.averageFactor_,
+                        arg.resultRunningMean_,
+                        arg.resultRunningVariance_,
+                        arg.saveMeanInvVariance_, // true or false
+                        arg.resultSaveMean_,
+                        arg.resultSaveInvVariance_);
+                }
+                else
+                {
+                    const auto kern_multiblock_welford_first_half =
+                        kernel_multiblock_welford_first_half<GridwiseMultiblockWelfordFirstHalf_,
+                                                             XDataType,
+                                                             MeanVarDataType,
+                                                             XYGridDesc_M_K,
+                                                             MeanVarCountGridDesc_M_G,
+                                                             GetReduceCountPerThreadFunctor>;
+
+                    const auto kern_welford_second_half_batchnorm_forward_final =
+                        kernel_welford_second_half_batchnorm_forward_final<
+                            GridwiseWelfordSecondHalfBatchNormForwardFinal_,
+                            XDataType,
+                            YDataType,
+                            AccDataType,
+                            ScaleDataType,
+                            BiasDataType,
+                            MeanVarDataType,
+                            YElementwiseOp,
+                            XYGridDesc_M_K,
+                            MeanVarCountGridDesc_M_K,
+                            ScaleBiasMeanVarGridDesc_M,
+                            ScaleBiasMeanVarGridDesc_M>;
+
+                    avg_time += launch_and_time_kernel(
+                        stream_config,
+                        kern_multiblock_welford_first_half,
+                        dim3(arg.gridSize_),
+                        dim3(BlockSize),
+                        0,
+                        arg.x_grid_desc_m_k_,
+                        mean_var_count_grid_desc_m_g,
+                        get_reduce_count_per_thread,
+                        arg.numBlockTileIteration_,
+                        arg.p_x_,
+                        static_cast<MeanVarDataType*>(arg.workspace_mean_),
+                        static_cast<MeanVarDataType*>(arg.workspace_variance_),
+                        static_cast<int32_t*>(arg.workspace_count_));
+
+                    avg_time += launch_and_time_kernel(
+                        stream_config,
+                        kern_welford_second_half_batchnorm_forward_final,
+                        dim3(arg.gridSize_),
+                        dim3(BlockSize),
+                        0,
+                        arg.x_grid_desc_m_k_,
+                        arg.y_grid_desc_m_k_,
+                        mean_var_count_grid_desc_m_k,
+                        arg.scale_grid_desc_m_,
+                        arg.bias_grid_desc_m_,
+                        arg.mean_var_grid_desc_m_,
+                        arg.blkGroupSize_,
+                        arg.numBlockTileIteration_,
+                        arg.epsilon_,
+                        static_cast<MeanVarDataType*>(arg.workspace_mean_),
+                        static_cast<MeanVarDataType*>(arg.workspace_variance_),
+                        static_cast<int32_t*>(arg.workspace_count_),
+                        arg.p_x_,
+                        arg.p_scale_,
+                        arg.p_bias_,
+                        arg.y_elementwise_op_,
+                        arg.p_y_,
+                        arg.updateMovingAverage_,
+                        arg.averageFactor_,
+                        arg.resultRunningMean_,
+                        arg.resultRunningVariance_,
+                        arg.saveMeanInvVariance_,
+                        arg.resultSaveMean_,
+                        arg.resultSaveInvVariance_);
+                };
+            }
+            else
+            {
+                using GetReduceCountPerThreadFunctor =
+                    GetReduceCountPerThreadForBlockwiseWelford<K_BlockTileSize, KThreadSliceSize>;
+
+                GetReduceCountPerThreadFunctor get_reduce_count_per_thread(
+                    arg.numBlockTileIteration_, arg.reduce_length_);
+
+                using GridwiseBatchNormForwardWithBlockwiseWelford_ =
+                    GridwiseBatchNormForwardWithBlockwiseWelford<XDataType,
+                                                                 YDataType,
+                                                                 AccDataType,
+                                                                 ScaleDataType,
+                                                                 BiasDataType,
+                                                                 MeanVarDataType,
+                                                                 YElementwiseOp,
+                                                                 XYGridDesc_M_K,
+                                                                 ScaleBiasMeanVarGridDesc_M,
+                                                                 ScaleBiasMeanVarGridDesc_M,
+                                                                 GetReduceCountPerThreadFunctor,
+                                                                 BlockSize,
+                                                                 MThreadClusterSize,
+                                                                 KThreadClusterSize,
+                                                                 MThreadSliceSize,
+                                                                 KThreadSliceSize,
+                                                                 XSrcYDstVectorDim,
+                                                                 XSrcVectorSize,
+                                                                 YDstVectorSize,
+                                                                 ScaleSrcVectorSize,
+                                                                 BiasSrcVectorSize,
+                                                                 MeanVarSrcDstVectorSize>;
+
+                const auto kern_batchnorm_fwd = kernel_batchnorm_forward_with_blockwise_welford<
+                    GridwiseBatchNormForwardWithBlockwiseWelford_,
+                    XDataType,
+                    YDataType,
+                    AccDataType,
+                    ScaleDataType,
+                    BiasDataType,
+                    MeanVarDataType,
+                    YElementwiseOp,
+                    XYGridDesc_M_K,
+                    ScaleBiasMeanVarGridDesc_M,
+                    ScaleBiasMeanVarGridDesc_M,
+                    GetReduceCountPerThreadFunctor>;
+
+                avg_time += launch_and_time_kernel(stream_config,
+                                                   kern_batchnorm_fwd,
+                                                   dim3(arg.gridSize_),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.x_grid_desc_m_k_,
+                                                   arg.y_grid_desc_m_k_,
+                                                   arg.scale_grid_desc_m_,
+                                                   arg.bias_grid_desc_m_,
+                                                   arg.mean_var_grid_desc_m_,
+                                                   get_reduce_count_per_thread,
+                                                   arg.numBlockTileIteration_,
+                                                   arg.epsilon_,
+                                                   arg.p_x_,
+                                                   arg.p_scale_,
+                                                   arg.p_bias_,
+                                                   arg.y_elementwise_op_,
+                                                   arg.p_y_,
+                                                   arg.updateMovingAverage_, // true or false
+                                                   arg.averageFactor_,
+                                                   arg.resultRunningMean_,
+                                                   arg.resultRunningVariance_,
+                                                   arg.saveMeanInvVariance_, // true or false
+                                                   arg.resultSaveMean_,
+                                                   arg.resultSaveInvVariance_);
+            };
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* pArg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(pArg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* pArg) override
+    {
+        const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
+
+        if constexpr(XSrcYDstVectorDim == 0)
+        {
+            if(pArg_->xStrides_[NumInvariantDim - 1] != 1 ||
+               pArg_->yStrides_[NumInvariantDim - 1] != 1)
+                return false;
+
+            if(pArg_->xyLengths_[NumInvariantDim - 1] % XSrcVectorSize != 0 ||
+               pArg_->xyLengths_[NumInvariantDim - 1] % YDstVectorSize != 0)
+                return false;
+        }
+        else
+        {
+            if(pArg_->xStrides_[Rank - 1] != 1 || pArg_->yStrides_[Rank - 1] != 1)
+                return false;
+
+            if(pArg_->xyLengths_[Rank - 1] % XSrcVectorSize != 0 ||
+               pArg_->xyLengths_[Rank - 1] % YDstVectorSize != 0)
+                return false;
+        };
+
+        if(pArg_->bnScaleStrides_[NumInvariantDim - 1] != 1 && ScaleSrcVectorSize != 1)
+            return false;
+        if(pArg_->bnBiasStrides_[NumInvariantDim - 1] != 1 && BiasSrcVectorSize != 1)
+            return false;
+
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % ScaleSrcVectorSize != 0)
+            return false;
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % BiasSrcVectorSize != 0)
+            return false;
+
+        if(pArg_->bnMeanVarStrides_[NumInvariantDim - 1] != 1 && MeanVarSrcDstVectorSize != 1)
+            return false;
+
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % MeanVarSrcDstVectorSize != 0)
+            return false;
+
+        bool is_valid = true;
+
+        static_for<0, NumInvariantDim, 1>{}([&](auto I) {
+            if(pArg_->xyLengths_[I] != pArg_->bnScaleBiasMeanVarLengths_[I])
+                is_valid = false;
+        });
+
+        if(!is_valid)
+            return false;
+
+        return true;
+    };
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, Rank> xyLengths,
+        const std::array<index_t, Rank> xStrides,
+        const std::array<index_t, Rank> yStrides,
+        const std::array<int, NumBatchNormReduceDim> reduceDims,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
+        const void* p_x,
+        const void* p_scale,
+        const void* p_bias,
+        double epsilon,
+        const YElementwiseOp y_elementwise_op,
+        void* p_y,
+        void* resultSaveMean,
+        void* resultSaveInvVariance,
+        double averageFactor,
+        void* resultRunningMean,
+        void* resultRunningVariance) override
+    {
+        return std::make_unique<Argument>(xyLengths,
+                                          xStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          bnScaleBiasMeanVarLengths,
+                                          bnScaleStrides,
+                                          bnBiasStrides,
+                                          bnMeanVarStrides,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const ScaleDataType*>(p_scale),
+                                          static_cast<const BiasDataType*>(p_bias),
+                                          y_elementwise_op,
+                                          epsilon,
+                                          static_cast<YDataType*>(p_y),
+                                          static_cast<MeanVarDataType*>(resultSaveMean),
+                                          static_cast<MeanVarDataType*>(resultSaveInvVariance),
+                                          averageFactor,
+                                          static_cast<MeanVarDataType*>(resultRunningMean),
+                                          static_cast<MeanVarDataType*>(resultRunningVariance));
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchNormFwdImpl<" << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "XSrcYDstVectorDim_" << XSrcYDstVectorDim  << ",";
+        str << "VectorSize_X" << XSrcVectorSize << "_scale_" << ScaleSrcVectorSize << "_bias_" << BiasSrcVectorSize << "_mean_var_" << MeanVarSrcDstVectorSize << "_Y" << YDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl_obsolete.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl_obsolete.hpp
new file mode 100644
index 0000000000..b826793c27
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl_obsolete.hpp
@@ -0,0 +1,714 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/device/welford_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp"
+#include "ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim,
+          bool UseMultiblockInK,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcYDstVectorDim,
+          index_t XSrcVectorSize,
+          index_t YDstVectorSize,
+          index_t ScaleSrcVectorSize,
+          index_t BiasSrcVectorSize,
+          index_t MeanVarSrcDstVectorSize>
+struct DeviceBatchNormFwdImpl : public DeviceBatchNormFwd<XDataType,
+                                                          YDataType,
+                                                          AccDataType,
+                                                          ScaleDataType,
+                                                          BiasDataType,
+                                                          MeanVarDataType,
+                                                          YElementwiseOp,
+                                                          Rank,
+                                                          NumBatchNormReduceDim>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
+                  "Invalid thread cluster size assignments!");
+
+    static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcYDstVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto MakeXY2dDescriptor(const std::array<index_t, Rank>& xyLengths,
+                                   const std::array<index_t, Rank>& xyStrides,
+                                   int blkGroupSize,
+                                   int numBlockTileIteration)
+    {
+        const auto tupleXYLengths =
+            generate_tuple([&](auto I) { return xyLengths[I]; }, Number<Rank>{});
+        const auto tupleXYStrides =
+            generate_tuple([&](auto I) { return xyStrides[I]; }, Number<Rank>{});
+
+        const auto raw_grid_desc = make_naive_tensor_descriptor(tupleXYLengths, tupleXYStrides);
+
+        const auto grid_desc_m_k = [&]() {
+            using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+            using ReduceDims    = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+            const auto reduceDimLengths =
+                generate_tuple([&](auto I) { return xyLengths[NumInvariantDim + I]; },
+                               Number<NumBatchNormReduceDim>{});
+            const auto invariantDimLengths =
+                generate_tuple([&](auto I) { return xyLengths[I]; }, Number<NumInvariantDim>{});
+
+            return transform_tensor_descriptor(raw_grid_desc,
+                                               make_tuple(make_merge_transform(invariantDimLengths),
+                                                          make_merge_transform(reduceDimLengths)),
+                                               make_tuple(InvariantDims{}, ReduceDims{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }();
+
+        const auto invariantLength = grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = grid_desc_m_k.GetLength(Number<1>{});
+
+        const int workSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto kPad = workSizePerBlock * blkGroupSize - reduceLength;
+
+        auto grid_desc_m_k_padded =
+            transform_tensor_descriptor(grid_desc_m_k,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_right_pad_transform(reduceLength, kPad)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_k_padded);
+    };
+
+    static auto MakeMeanVarCountOutputMG2dDescriptor(int invariantLength, int blkGroupSize)
+    {
+        const auto grid_desc_m_g = make_naive_tensor_descriptor(
+            make_tuple(invariantLength, blkGroupSize), make_tuple(1, invariantLength));
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto grid_desc_m_g_padded =
+            transform_tensor_descriptor(grid_desc_m_g,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_pass_through_transform(blkGroupSize)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_g_padded);
+    };
+
+    static auto MakeMeanVarCountInputMK2dDescriptor(int invariantLength, int blkGroupSize)
+    {
+        const auto reduceLength  = blkGroupSize;
+        const auto grid_desc_m_k = make_naive_tensor_descriptor(
+            make_tuple(invariantLength, reduceLength), make_tuple(1, invariantLength));
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto kPad =
+            math::integer_least_multiple(reduceLength, KThreadClusterSize) - reduceLength;
+
+        auto grid_desc_m_k_padded =
+            transform_tensor_descriptor(grid_desc_m_k,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad),
+                                                   make_right_pad_transform(reduceLength, kPad)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (grid_desc_m_k_padded);
+    };
+
+    static auto
+    MakeScaleBiasMeanVar1dDescriptor(const std::array<index_t, NumInvariantDim>& lengths,
+                                     const std::array<index_t, NumInvariantDim>& strides)
+    {
+        const auto tupleLengths =
+            generate_tuple([&](auto I) { return lengths[I]; }, Number<NumInvariantDim>{});
+        const auto tupleStrides =
+            generate_tuple([&](auto I) { return strides[I]; }, Number<NumInvariantDim>{});
+
+        auto raw_grid_desc = make_naive_tensor_descriptor(tupleLengths, tupleStrides);
+
+        auto grid_desc_m = transform_tensor_descriptor(
+            raw_grid_desc,
+            make_tuple(make_merge_transform(tupleLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = grid_desc_m.GetLength(Number<0>{});
+
+        const auto mPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto grid_desc_m_padded =
+            transform_tensor_descriptor(grid_desc_m,
+                                        make_tuple(make_right_pad_transform(invariantLength, mPad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return (grid_desc_m_padded);
+    };
+
+    using XYGridDesc_M_K             = decltype(MakeXY2dDescriptor({1}, {1}, 1, 1));
+    using ScaleBiasMeanVarGridDesc_M = decltype(MakeScaleBiasMeanVar1dDescriptor({1}, {1}));
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> xyLengths,
+                 const std::array<index_t, Rank> xStrides,
+                 const std::array<index_t, Rank> yStrides,
+                 const std::array<int, NumBatchNormReduceDim> reduceDims,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+                 const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
+                 const XDataType* p_x,
+                 const ScaleDataType* p_scale,
+                 const BiasDataType* p_bias,
+                 const YElementwiseOp y_elementwise_op,
+                 double epsilon,
+                 YDataType* p_y,
+                 MeanVarDataType* resultSaveMean,
+                 MeanVarDataType* resultSaveInvVariance,
+                 double averageFactor,
+                 MeanVarDataType* resultRunningMean,
+                 MeanVarDataType* resultRunningVariance)
+            : bnScaleBiasMeanVarLengths_(bnScaleBiasMeanVarLengths),
+              bnScaleStrides_(bnScaleStrides),
+              bnBiasStrides_(bnBiasStrides),
+              bnMeanVarStrides_(bnMeanVarStrides),
+              p_x_(p_x),
+              p_scale_(p_scale),
+              p_bias_(p_bias),
+              y_elementwise_op_(y_elementwise_op),
+              p_y_(p_y),
+              resultSaveMean_(resultSaveMean),
+              resultSaveInvVariance_(resultSaveInvVariance),
+              resultRunningMean_(resultRunningMean),
+              resultRunningVariance_(resultRunningVariance)
+        {
+            xyLengths_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(xyLengths, reduceDims);
+            xStrides_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(xStrides, reduceDims);
+            yStrides_ =
+                shuffle_tensor_dimensions<Rank, NumBatchNormReduceDim>(yStrides, reduceDims);
+
+            std::tie(invariant_length_, reduce_length_) =
+                get_2d_lengths<Rank, NumBatchNormReduceDim>(xyLengths_);
+
+            epsilon_       = type_convert<AccDataType>(epsilon);
+            averageFactor_ = type_convert<AccDataType>(averageFactor);
+
+            updateMovingAverage_ =
+                (resultRunningMean != nullptr && resultRunningVariance != nullptr);
+            saveMeanInvVariance_ = (resultSaveMean != nullptr && resultSaveInvVariance_ != nullptr);
+
+            if(UseMultiblockInK)
+            {
+                int iterations = 1;
+                while(true)
+                {
+                    int testBlkGroupSize = (reduce_length_ + (K_BlockTileSize * iterations) - 1) /
+                                           (K_BlockTileSize * iterations);
+
+                    // we want the blkGroupSize be not more than 16
+                    if(testBlkGroupSize <= 16)
+                        break;
+
+                    iterations++;
+                };
+
+                blkGroupSize_ = (reduce_length_ + (K_BlockTileSize * iterations) - 1) /
+                                (K_BlockTileSize * iterations);
+
+                numBlockTileIteration_ = iterations;
+            }
+            else
+            {
+                blkGroupSize_          = 1;
+                numBlockTileIteration_ = (reduce_length_ + K_BlockTileSize - 1) / K_BlockTileSize;
+            };
+
+            gridSize_ = (invariant_length_ + M_BlockTileSize - 1) / M_BlockTileSize * blkGroupSize_;
+
+            x_grid_desc_m_k_ =
+                MakeXY2dDescriptor(xyLengths_, xStrides_, blkGroupSize_, numBlockTileIteration_);
+            y_grid_desc_m_k_ =
+                MakeXY2dDescriptor(xyLengths_, yStrides_, blkGroupSize_, numBlockTileIteration_);
+            scale_grid_desc_m_ =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnScaleStrides_);
+            bias_grid_desc_m_ =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnBiasStrides_);
+            mean_var_grid_desc_m_ =
+                MakeScaleBiasMeanVar1dDescriptor(bnScaleBiasMeanVarLengths, bnMeanVarStrides_);
+        }
+
+        AccDataType epsilon_;
+        AccDataType averageFactor_;
+
+        bool updateMovingAverage_;
+        bool saveMeanInvVariance_;
+
+        std::array<index_t, Rank> xyLengths_;
+        std::array<index_t, Rank> xStrides_;
+        std::array<index_t, Rank> yStrides_;
+
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides_;
+        std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides_;
+
+        const XDataType* p_x_;
+        const ScaleDataType* p_scale_;
+        const BiasDataType* p_bias_;
+        const YElementwiseOp y_elementwise_op_;
+        YDataType* p_y_;
+
+        MeanVarDataType* resultSaveMean_;
+        MeanVarDataType* resultSaveInvVariance_;
+
+        MeanVarDataType* resultRunningMean_;
+        MeanVarDataType* resultRunningVariance_;
+
+        long_index_t invariant_length_;
+        long_index_t reduce_length_;
+
+        int blkGroupSize_;
+        int numBlockTileIteration_;
+        size_t gridSize_;
+
+        XYGridDesc_M_K x_grid_desc_m_k_;
+        XYGridDesc_M_K y_grid_desc_m_k_;
+        ScaleBiasMeanVarGridDesc_M scale_grid_desc_m_;
+        ScaleBiasMeanVarGridDesc_M bias_grid_desc_m_;
+        ScaleBiasMeanVarGridDesc_M mean_var_grid_desc_m_;
+
+        void* workspace_mean_;
+        void* workspace_variance_;
+        void* workspace_count_;
+    };
+
+    size_t GetWorkSpaceSize(const BaseArgument* pArg) const override
+    {
+        const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
+
+        size_t workspace_size = 0;
+
+        if(UseMultiblockInK && pArg_->blkGroupSize_ > 1)
+        {
+            // workspace for welford intermediate mean
+            workspace_size +=
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType) + 64;
+
+            // workspace for welford intermediate variance
+            workspace_size +=
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType) + 64;
+
+            // workspace for welford intermediate count
+            workspace_size +=
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(int32_t) + 64;
+        }
+
+        return (workspace_size);
+    };
+
+    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    {
+        Argument* pArg_ = dynamic_cast<Argument*>(pArg);
+
+        pArg_->p_workspace_ = p_workspace;
+
+        if(UseMultiblockInK && pArg_->blkGroupSize_ > 1)
+        {
+
+            // setup buffer used for intermediate welford mean
+            pArg_->workspace_mean_ = static_cast<char*>(pArg_->p_workspace_);
+
+            index_t mean_space_sz =
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType);
+
+            mean_space_sz = math::integer_least_multiple(mean_space_sz, 64);
+
+            // setup buffer used for intermediate welford varirance
+            pArg_->workspace_variance_ =
+                reinterpret_cast<char*>(pArg_->workspace_mean_) + mean_space_sz;
+
+            index_t variance_space_sz =
+                pArg_->invariant_length_ * pArg_->blkGroupSize_ * sizeof(MeanVarDataType);
+
+            variance_space_sz = math::integer_least_multiple(variance_space_sz, 64);
+
+            // setup buffer used for intermediate welfor count
+            pArg_->workspace_count_ =
+                reinterpret_cast<char*>(pArg_->workspace_variance_) + variance_space_sz;
+        };
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            float avg_time = 0;
+
+            if(UseMultiblockInK && arg.blkGroupSize_ > 1)
+            {
+                using GetReduceCountPerThreadFunctor =
+                    GetReduceCountPerThreadForMultiblockWelford<K_BlockTileSize, KThreadSliceSize>;
+
+                GetReduceCountPerThreadFunctor get_reduce_count_per_thread(
+                    arg.blkGroupSize_, arg.numBlockTileIteration_, arg.reduce_length_);
+
+                const auto mean_var_count_grid_desc_m_g =
+                    DeviceBatchNormFwdImpl::MakeMeanVarCountOutputMG2dDescriptor(
+                        arg.invariant_length_, arg.blkGroupSize_);
+
+                const auto mean_var_count_grid_desc_m_k =
+                    DeviceBatchNormFwdImpl::MakeMeanVarCountInputMK2dDescriptor(
+                        arg.invariant_length_, arg.blkGroupSize_);
+
+                using MeanVarCountGridDesc_M_G = decltype(mean_var_count_grid_desc_m_g);
+                using MeanVarCountGridDesc_M_K = decltype(mean_var_count_grid_desc_m_k);
+
+                using GridwiseMultiblockWelfordFirstHalf_ =
+                    GridwiseMultiblockWelfordFirstHalf<XDataType,
+                                                       AccDataType,
+                                                       MeanVarDataType,
+                                                       XYGridDesc_M_K,
+                                                       MeanVarCountGridDesc_M_G,
+                                                       GetReduceCountPerThreadFunctor,
+                                                       BlockSize,
+                                                       MThreadClusterSize,
+                                                       KThreadClusterSize,
+                                                       MThreadSliceSize,
+                                                       KThreadSliceSize,
+                                                       XSrcYDstVectorDim,
+                                                       XSrcVectorSize>;
+
+                using GridwiseWelfordSecondHalfBatchNormForwardFinal_ =
+                    GridwiseWelfordSecondHalfBatchNormForwardFinal<XDataType,
+                                                                   YDataType,
+                                                                   AccDataType,
+                                                                   ScaleDataType,
+                                                                   BiasDataType,
+                                                                   MeanVarDataType,
+                                                                   YElementwiseOp,
+                                                                   XYGridDesc_M_K,
+                                                                   MeanVarCountGridDesc_M_K,
+                                                                   ScaleBiasMeanVarGridDesc_M,
+                                                                   ScaleBiasMeanVarGridDesc_M,
+                                                                   BlockSize,
+                                                                   MThreadClusterSize,
+                                                                   KThreadClusterSize,
+                                                                   MThreadSliceSize,
+                                                                   KThreadSliceSize,
+                                                                   XSrcYDstVectorDim,
+                                                                   XSrcVectorSize,
+                                                                   YDstVectorSize,
+                                                                   ScaleSrcVectorSize,
+                                                                   BiasSrcVectorSize,
+                                                                   MeanVarSrcDstVectorSize>;
+
+                const auto kern_multiblock_welford_first_half =
+                    kernel_multiblock_welford_first_half<GridwiseMultiblockWelfordFirstHalf_,
+                                                         XDataType,
+                                                         MeanVarDataType,
+                                                         XYGridDesc_M_K,
+                                                         MeanVarCountGridDesc_M_G,
+                                                         GetReduceCountPerThreadFunctor>;
+
+                const auto kern_welford_second_half_batchnorm_forward_final =
+                    kernel_welford_second_half_batchnorm_forward_final<
+                        GridwiseWelfordSecondHalfBatchNormForwardFinal_,
+                        XDataType,
+                        YDataType,
+                        AccDataType,
+                        ScaleDataType,
+                        BiasDataType,
+                        MeanVarDataType,
+                        YElementwiseOp,
+                        XYGridDesc_M_K,
+                        MeanVarCountGridDesc_M_K,
+                        ScaleBiasMeanVarGridDesc_M,
+                        ScaleBiasMeanVarGridDesc_M>;
+
+                avg_time +=
+                    launch_and_time_kernel(stream_config,
+                                           kern_multiblock_welford_first_half,
+                                           dim3(arg.gridSize_),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.x_grid_desc_m_k_,
+                                           mean_var_count_grid_desc_m_g,
+                                           get_reduce_count_per_thread,
+                                           arg.numBlockTileIteration_,
+                                           arg.p_x_,
+                                           static_cast<MeanVarDataType*>(arg.workspace_mean_),
+                                           static_cast<MeanVarDataType*>(arg.workspace_variance_),
+                                           static_cast<int32_t*>(arg.workspace_count_));
+
+                avg_time +=
+                    launch_and_time_kernel(stream_config,
+                                           kern_welford_second_half_batchnorm_forward_final,
+                                           dim3(arg.gridSize_),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.x_grid_desc_m_k_,
+                                           arg.y_grid_desc_m_k_,
+                                           mean_var_count_grid_desc_m_k,
+                                           arg.scale_grid_desc_m_,
+                                           arg.bias_grid_desc_m_,
+                                           arg.mean_var_grid_desc_m_,
+                                           arg.blkGroupSize_,
+                                           arg.numBlockTileIteration_,
+                                           arg.epsilon_,
+                                           static_cast<MeanVarDataType*>(arg.workspace_mean_),
+                                           static_cast<MeanVarDataType*>(arg.workspace_variance_),
+                                           static_cast<int32_t*>(arg.workspace_count_),
+                                           arg.p_x_,
+                                           arg.p_scale_,
+                                           arg.p_bias_,
+                                           arg.y_elementwise_op_,
+                                           arg.p_y_,
+                                           arg.updateMovingAverage_,
+                                           arg.averageFactor_,
+                                           arg.resultRunningMean_,
+                                           arg.resultRunningVariance_,
+                                           arg.saveMeanInvVariance_,
+                                           arg.resultSaveMean_,
+                                           arg.resultSaveInvVariance_);
+            }
+            else
+            {
+                using GetReduceCountPerThreadFunctor =
+                    GetReduceCountPerThreadForBlockwiseWelford<K_BlockTileSize, KThreadSliceSize>;
+
+                GetReduceCountPerThreadFunctor get_reduce_count_per_thread(
+                    arg.numBlockTileIteration_, arg.reduce_length_);
+
+                using GridwiseBatchNormForwardWithBlockwiseWelford_ =
+                    GridwiseBatchNormForwardWithBlockwiseWelford<XDataType,
+                                                                 YDataType,
+                                                                 AccDataType,
+                                                                 ScaleDataType,
+                                                                 BiasDataType,
+                                                                 MeanVarDataType,
+                                                                 YElementwiseOp,
+                                                                 XYGridDesc_M_K,
+                                                                 ScaleBiasMeanVarGridDesc_M,
+                                                                 ScaleBiasMeanVarGridDesc_M,
+                                                                 GetReduceCountPerThreadFunctor,
+                                                                 BlockSize,
+                                                                 MThreadClusterSize,
+                                                                 KThreadClusterSize,
+                                                                 MThreadSliceSize,
+                                                                 KThreadSliceSize,
+                                                                 XSrcYDstVectorDim,
+                                                                 XSrcVectorSize,
+                                                                 YDstVectorSize,
+                                                                 ScaleSrcVectorSize,
+                                                                 BiasSrcVectorSize,
+                                                                 MeanVarSrcDstVectorSize>;
+
+                const auto kern_batchnorm_fwd = kernel_batchnorm_forward_with_blockwise_welford<
+                    GridwiseBatchNormForwardWithBlockwiseWelford_,
+                    XDataType,
+                    YDataType,
+                    AccDataType,
+                    ScaleDataType,
+                    BiasDataType,
+                    MeanVarDataType,
+                    YElementwiseOp,
+                    XYGridDesc_M_K,
+                    ScaleBiasMeanVarGridDesc_M,
+                    ScaleBiasMeanVarGridDesc_M,
+                    GetReduceCountPerThreadFunctor>;
+
+                avg_time += launch_and_time_kernel(stream_config,
+                                                   kern_batchnorm_fwd,
+                                                   dim3(arg.gridSize_),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.x_grid_desc_m_k_,
+                                                   arg.y_grid_desc_m_k_,
+                                                   arg.scale_grid_desc_m_,
+                                                   arg.bias_grid_desc_m_,
+                                                   arg.mean_var_grid_desc_m_,
+                                                   get_reduce_count_per_thread,
+                                                   arg.numBlockTileIteration_,
+                                                   arg.epsilon_,
+                                                   arg.p_x_,
+                                                   arg.p_scale_,
+                                                   arg.p_bias_,
+                                                   arg.y_elementwise_op_,
+                                                   arg.p_y_,
+                                                   arg.updateMovingAverage_, // true or false
+                                                   arg.averageFactor_,
+                                                   arg.resultRunningMean_,
+                                                   arg.resultRunningVariance_,
+                                                   arg.saveMeanInvVariance_, // true or false
+                                                   arg.resultSaveMean_,
+                                                   arg.resultSaveInvVariance_);
+            };
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* pArg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(pArg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* pArg) override
+    {
+        const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
+
+        if constexpr(XSrcYDstVectorDim == 0)
+        {
+            if(pArg_->xStrides_[NumInvariantDim - 1] != 1 ||
+               pArg_->yStrides_[NumInvariantDim - 1] != 1)
+                return false;
+
+            if(pArg_->xyLengths_[NumInvariantDim - 1] % XSrcVectorSize != 0 ||
+               pArg_->xyLengths_[NumInvariantDim - 1] % YDstVectorSize != 0)
+                return false;
+        }
+        else
+        {
+            if(pArg_->xStrides_[Rank - 1] != 1 || pArg_->yStrides_[Rank - 1] != 1)
+                return false;
+
+            if(pArg_->xyLengths_[Rank - 1] % XSrcVectorSize != 0 ||
+               pArg_->xyLengths_[Rank - 1] % YDstVectorSize != 0)
+                return false;
+        };
+
+        if(pArg_->bnScaleStrides_[NumInvariantDim - 1] != 1 && ScaleSrcVectorSize != 1)
+            return false;
+        if(pArg_->bnBiasStrides_[NumInvariantDim - 1] != 1 && BiasSrcVectorSize != 1)
+            return false;
+
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % ScaleSrcVectorSize != 0)
+            return false;
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % BiasSrcVectorSize != 0)
+            return false;
+
+        if(pArg_->bnMeanVarStrides_[NumInvariantDim - 1] != 1 && MeanVarSrcDstVectorSize != 1)
+            return false;
+
+        if(pArg_->bnScaleBiasMeanVarLengths_[NumInvariantDim - 1] % MeanVarSrcDstVectorSize != 0)
+            return false;
+
+        bool is_valid = true;
+
+        static_for<0, NumInvariantDim, 1>{}([&](auto I) {
+            if(pArg_->xyLengths_[I] != pArg_->bnScaleBiasMeanVarLengths_[I])
+                is_valid = false;
+        });
+
+        if(!is_valid)
+            return false;
+
+        return true;
+    };
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, Rank> xyLengths,
+        const std::array<index_t, Rank> xStrides,
+        const std::array<index_t, Rank> yStrides,
+        const std::array<int, NumBatchNormReduceDim> reduceDims,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleBiasMeanVarLengths,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnScaleStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnBiasStrides,
+        const std::array<index_t, Rank - NumBatchNormReduceDim> bnMeanVarStrides,
+        const void* p_x,
+        const void* p_scale,
+        const void* p_bias,
+        double epsilon,
+        const YElementwiseOp y_elementwise_op,
+        void* p_y,
+        void* resultSaveMean,
+        void* resultSaveInvVariance,
+        double averageFactor,
+        void* resultRunningMean,
+        void* resultRunningVariance) override
+    {
+        return std::make_unique<Argument>(xyLengths,
+                                          xStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          bnScaleBiasMeanVarLengths,
+                                          bnScaleStrides,
+                                          bnBiasStrides,
+                                          bnMeanVarStrides,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const ScaleDataType*>(p_scale),
+                                          static_cast<const BiasDataType*>(p_bias),
+                                          y_elementwise_op,
+                                          epsilon,
+                                          static_cast<YDataType*>(p_y),
+                                          static_cast<MeanVarDataType*>(resultSaveMean),
+                                          static_cast<MeanVarDataType*>(resultSaveInvVariance),
+                                          averageFactor,
+                                          static_cast<MeanVarDataType*>(resultRunningMean),
+                                          static_cast<MeanVarDataType*>(resultRunningVariance));
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceBatchNormFwdImpl<" << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "XSrcYDstVectorDim_" << XSrcYDstVectorDim  << ",";
+        str << "VectorSize_X" << XSrcVectorSize << "_scale_" << ScaleSrcVectorSize << "_bias_" << BiasSrcVectorSize << "_mean_var_" << MeanVarSrcDstVectorSize << "_Y" << YDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..a095521161
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp
@@ -0,0 +1,587 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_cgemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <
+    typename ALayout,
+    typename BLayout,
+    typename CLayout,
+    typename ADataType,
+    typename BDataType,
+    typename CDataType,
+    typename GemmAccDataType,
+    typename CShuffleDataType,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    GemmSpecialization GemmSpec,
+    index_t NumGemmKPrefetchStage,
+    index_t BlockSize,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t KPerBlock,
+    index_t AK1,
+    index_t BK1,
+    index_t MPerXDL,
+    index_t NPerXDL,
+    index_t MXdlPerWave,
+    index_t NXdlPerWave,
+    typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    index_t ABlockTransferSrcVectorDim,
+    index_t ABlockTransferSrcScalarPerVector,
+    index_t ABlockTransferDstScalarPerVector_AK1,
+    bool ABlockLdsExtraM,
+    typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    index_t BBlockTransferSrcVectorDim,
+    index_t BBlockTransferSrcScalarPerVector,
+    index_t BBlockTransferDstScalarPerVector_BK1,
+    bool BBlockLdsExtraN,
+    index_t CShuffleMXdlPerWavePerShuffle,
+    index_t CShuffleNXdlPerWavePerShuffle,
+    typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+    index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+    LoopScheduler LoopSched = make_default_loop_scheduler(),
+    enable_if_t<
+        is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<CElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+        bool> = false>
+struct DeviceCGemm_4Gemm_Xdl_CShuffle
+    : public DeviceCGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>
+{
+    using DeviceOp = DeviceCGemm_4Gemm_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto MPerThread       = Number<4>{};
+    static constexpr auto AScalarPerVector = Number<4>{};
+    static constexpr auto BScalarPerVector = Number<4>{};
+    static constexpr auto CScalarPerVector = Number<4>{};
+
+    template <typename Desc_M>
+    static auto PadDescriptor_M_1d(Desc_M desc_m, index_t gridSize, index_t blockSize)
+    {
+        const auto M            = desc_m.GetLength(I0);
+        const index_t loop_step = gridSize * blockSize * MPerThread;
+        const auto pad          = math::integer_least_multiple(M, loop_step) - M;
+        const auto desc_m_pad =
+            transform_tensor_descriptor(desc_m,
+                                        make_tuple(make_right_pad_transform(M, pad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return desc_m_pad;
+    }
+
+    static auto MakeDescriptor_M(const std::vector<index_t>& lengths,
+                                 const std::vector<index_t>& strides,
+                                 index_t gridSize,
+                                 index_t blockSize)
+    {
+        auto tupleOfShape  = generate_tuple([&](auto I) { return lengths[I]; }, Number<2>{});
+        auto tupleOfStride = generate_tuple([&](auto I) { return strides[I]; }, Number<2>{});
+
+        // nd desc - [s0, s1, s2, ...]
+        const auto desc   = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
+        const auto desc_m = transform_tensor_descriptor(
+            desc,
+            make_tuple(make_merge_transform(tupleOfShape)),
+            make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<2>{})),
+            make_tuple(Sequence<0>{}));
+
+        return PadDescriptor_M_1d(desc_m, gridSize, blockSize);
+    }
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ALayout,
+        BLayout,
+        CLayout,
+        ADataType,
+        BDataType,
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        GemmSpec,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    using CGridDesc_M = decltype(MakeDescriptor_M({1, 1}, {1, 1}, 1, 1));
+
+    // Argument
+    struct Argument : public tensor_operation::device::BaseArgument, public GridwiseGemm::Problem
+    {
+        using Problem = typename GridwiseGemm::Problem;
+
+        Argument(const ADataType* p_a_grid_real_,
+                 const ADataType* p_a_grid_imag_,
+                 const BDataType* p_b_grid_real_,
+                 const BDataType* p_b_grid_imag_,
+                 CDataType* p_c_grid_real_,
+                 CDataType* p_c_grid_imag_,
+                 CDataType* p_workspace,
+                 index_t M_,
+                 index_t N_,
+                 index_t K_,
+                 index_t StrideA_,
+                 index_t StrideB_,
+                 index_t StrideC_)
+            : Problem{M_, N_, K_, StrideA_, StrideB_, StrideC_},
+              p_a_grid_real{p_a_grid_real_},
+              p_a_grid_imag{p_a_grid_imag_},
+              p_b_grid_real{p_b_grid_real_},
+              p_b_grid_imag{p_b_grid_imag_},
+              p_c_grid_real{p_c_grid_real_},
+              p_c_grid_imag{p_c_grid_imag_},
+              p_aux_grid{p_workspace}
+        {
+            const index_t grid_size = std::get<1>(GridwiseGemm::CalculateGridSize(M_, N_));
+
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                c_grid_desc_m =
+                    DeviceOp::MakeDescriptor_M({M_, N_}, {StrideC_, I1}, grid_size, BlockSize);
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                c_grid_desc_m =
+                    DeviceOp::MakeDescriptor_M({M_, N_}, {I1, StrideC_}, grid_size, BlockSize);
+            }
+
+            p_aux_2_grid = p_workspace + GetCElementSpaceSize(M_, N_, StrideC_);
+        }
+
+        //  private:
+        const ADataType* p_a_grid_real;
+        const ADataType* p_a_grid_imag;
+        const BDataType* p_b_grid_real;
+        const BDataType* p_b_grid_imag;
+        CDataType* p_c_grid_real;
+        CDataType* p_c_grid_imag;
+        CDataType* p_aux_grid;
+        CDataType* p_aux_2_grid;
+        CGridDesc_M c_grid_desc_m;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            index_t gdx, gdy, gdz;
+            std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N);
+
+            const auto K = GridwiseGemm::CalculateAK0(arg.K) * AK1;
+
+            float ave_time = 0;
+
+            using Add      = ck::tensor_operation::element_wise::Add;
+            using Subtract = ck::tensor_operation::element_wise::Subtract;
+
+            using GridwiseBinAdd =
+                GridwiseElementwise_1D<Tuple<CGridDesc_M, CGridDesc_M>,
+                                       Tuple<CGridDesc_M>,
+                                       Tuple<const CDataType*, const CDataType*>,
+                                       Tuple<CDataType*>,
+                                       Add,
+                                       MPerThread,
+                                       Sequence<AScalarPerVector, BScalarPerVector>,
+                                       Sequence<CScalarPerVector>>;
+
+            using GridwiseBinSubtract =
+                GridwiseElementwise_1D<Tuple<CGridDesc_M, CGridDesc_M>,
+                                       Tuple<CGridDesc_M>,
+                                       Tuple<const CDataType*, const CDataType*>,
+                                       Tuple<CDataType*>,
+                                       Subtract,
+                                       MPerThread,
+                                       Sequence<AScalarPerVector, BScalarPerVector>,
+                                       Sequence<CScalarPerVector>>;
+
+            const auto add_kernel = kernel_elementwise_1d<GridwiseBinAdd,
+                                                          Tuple<CGridDesc_M, CGridDesc_M>,
+                                                          Tuple<CGridDesc_M>,
+                                                          Tuple<const CDataType*, const CDataType*>,
+                                                          Tuple<CDataType*>,
+                                                          Add>;
+
+            const auto subtract_kernel =
+                kernel_elementwise_1d<GridwiseBinSubtract,
+                                      Tuple<CGridDesc_M, CGridDesc_M>,
+                                      Tuple<CGridDesc_M>,
+                                      Tuple<const CDataType*, const CDataType*>,
+                                      Tuple<CDataType*>,
+                                      Subtract>;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_xdl_cshuffle_v1<GridwiseGemm,
+                                                                ADataType,
+                                                                BDataType,
+                                                                CDataType,
+                                                                true>;
+
+                ave_time += launch_and_time_kernel(stream_config,
+                                                   kernel,
+                                                   dim3(gdx, gdy, gdz),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.p_a_grid_real,
+                                                   arg.p_b_grid_real,
+                                                   arg.p_aux_grid,
+                                                   arg);
+
+                ave_time += launch_and_time_kernel(stream_config,
+                                                   kernel,
+                                                   dim3(gdx, gdy, gdz),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.p_a_grid_imag,
+                                                   arg.p_b_grid_imag,
+                                                   arg.p_aux_2_grid,
+                                                   arg);
+
+                // c_real = aux - aux_2
+                ave_time += launch_and_time_kernel(
+                    stream_config,
+                    subtract_kernel,
+                    dim3(gdx, gdy, gdz),
+                    dim3(BlockSize),
+                    0,
+                    make_tuple(arg.c_grid_desc_m, arg.c_grid_desc_m),
+                    make_tuple(arg.c_grid_desc_m),
+                    make_tuple(const_cast<const CDataType*>(arg.p_aux_grid),
+                               const_cast<const CDataType*>(arg.p_aux_2_grid)),
+                    make_tuple(arg.p_c_grid_real),
+                    Subtract{});
+
+                ave_time += launch_and_time_kernel(stream_config,
+                                                   kernel,
+                                                   dim3(gdx, gdy, gdz),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.p_a_grid_real,
+                                                   arg.p_b_grid_imag,
+                                                   arg.p_aux_grid,
+                                                   arg);
+
+                ave_time += launch_and_time_kernel(stream_config,
+                                                   kernel,
+                                                   dim3(gdx, gdy, gdz),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.p_a_grid_imag,
+                                                   arg.p_b_grid_real,
+                                                   arg.p_aux_2_grid,
+                                                   arg);
+
+                // c_imag = aux + aux_2
+                ave_time += launch_and_time_kernel(
+                    stream_config,
+                    add_kernel,
+                    dim3(gdx, gdy, gdz),
+                    dim3(BlockSize),
+                    0,
+                    make_tuple(arg.c_grid_desc_m, arg.c_grid_desc_m),
+                    make_tuple(arg.c_grid_desc_m),
+                    make_tuple(const_cast<const CDataType*>(arg.p_aux_grid),
+                               const_cast<const CDataType*>(arg.p_aux_2_grid)),
+                    make_tuple(arg.p_c_grid_imag),
+                    Add{});
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdl_cshuffle_v1<GridwiseGemm,
+                                                                ADataType,
+                                                                BDataType,
+                                                                CDataType,
+                                                                false>;
+
+                ave_time += launch_and_time_kernel(stream_config,
+                                                   kernel,
+                                                   dim3(gdx, gdy, gdz),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.p_a_grid_real,
+                                                   arg.p_b_grid_real,
+                                                   arg.p_aux_grid,
+                                                   arg);
+
+                ave_time += launch_and_time_kernel(stream_config,
+                                                   kernel,
+                                                   dim3(gdx, gdy, gdz),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.p_a_grid_imag,
+                                                   arg.p_b_grid_imag,
+                                                   arg.p_aux_2_grid,
+                                                   arg);
+
+                // c_real = aux - aux_2
+                ave_time += launch_and_time_kernel(
+                    stream_config,
+                    subtract_kernel,
+                    dim3(gdx, gdy, gdz),
+                    dim3(BlockSize),
+                    0,
+                    make_tuple(arg.c_grid_desc_m, arg.c_grid_desc_m),
+                    make_tuple(arg.c_grid_desc_m),
+                    make_tuple(const_cast<const CDataType*>(arg.p_aux_grid),
+                               const_cast<const CDataType*>(arg.p_aux_2_grid)),
+                    make_tuple(arg.p_c_grid_real),
+                    Subtract{});
+
+                ave_time += launch_and_time_kernel(stream_config,
+                                                   kernel,
+                                                   dim3(gdx, gdy, gdz),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.p_a_grid_real,
+                                                   arg.p_b_grid_imag,
+                                                   arg.p_aux_grid,
+                                                   arg);
+
+                ave_time += launch_and_time_kernel(stream_config,
+                                                   kernel,
+                                                   dim3(gdx, gdy, gdz),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.p_a_grid_imag,
+                                                   arg.p_b_grid_real,
+                                                   arg.p_aux_2_grid,
+                                                   arg);
+
+                // c_imag = aux + aux_2
+                ave_time += launch_and_time_kernel(
+                    stream_config,
+                    add_kernel,
+                    dim3(gdx, gdy, gdz),
+                    dim3(BlockSize),
+                    0,
+                    make_tuple(arg.c_grid_desc_m, arg.c_grid_desc_m),
+                    make_tuple(arg.c_grid_desc_m),
+                    make_tuple(const_cast<const CDataType*>(arg.p_aux_grid),
+                               const_cast<const CDataType*>(arg.p_aux_2_grid)),
+                    make_tuple(arg.p_c_grid_imag),
+                    Add{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a_real,
+                             const ADataType* p_a_imag,
+                             const BDataType* p_b_real,
+                             const BDataType* p_b_imag,
+                             CDataType* p_c_real,
+                             CDataType* p_c_imag,
+                             CDataType* p_workspace,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation,
+                             BElementwiseOperation,
+                             CElementwiseOperation)
+    {
+        return Argument{p_a_real,
+                        p_a_imag,
+                        p_b_real,
+                        p_b_imag,
+                        p_c_real,
+                        p_c_imag,
+                        p_workspace,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a_real,
+                                                      const void* p_a_imag,
+                                                      const void* p_b_real,
+                                                      const void* p_b_imag,
+                                                      void* p_c_real,
+                                                      void* p_c_imag,
+                                                      void* p_workspace,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation,
+                                                      BElementwiseOperation,
+                                                      CElementwiseOperation,
+                                                      index_t /* KBatch */ = 1) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a_real),
+                                          static_cast<const ADataType*>(p_a_imag),
+                                          static_cast<const BDataType*>(p_b_real),
+                                          static_cast<const BDataType*>(p_b_imag),
+                                          static_cast<CDataType*>(p_c_real),
+                                          static_cast<CDataType*>(p_c_imag),
+                                          static_cast<CDataType*>(p_workspace),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceCGemm_4Gemm_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    static std::size_t GetCElementSpaceSize(index_t M, index_t N, index_t StrideC)
+    {
+        const auto c_grid_desc_m_n = GridwiseGemm::MakeCGridDescriptor_M_N(
+            M, GridwiseGemm::CalculateMPadded(M), N, GridwiseGemm::CalculateNPadded(N), StrideC);
+
+        return c_grid_desc_m_n.GetElementSpaceSize();
+    }
+
+    std::size_t GetWorkspaceSize(index_t M,
+                                 index_t N,
+                                 [[maybe_unused]] index_t K,
+                                 [[maybe_unused]] index_t StrideA,
+                                 [[maybe_unused]] index_t StrideB,
+                                 index_t StrideC) override
+    {
+        return 2 * sizeof(CDataType) * GetCElementSpaceSize(M, N, StrideC);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..5ae06836fa
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,793 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatDsPointer,
+          typename FloatE,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_contraction_multiple_d_xdl_cshuffle(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatDsPointer p_ds_grid,
+            FloatE* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_ds_grid,
+                                                  p_e_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_etile_map;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceContractionMultipleD_Xdl_CShuffle
+    : public DeviceContractionMultipleD<NumDimM,
+                                        NumDimN,
+                                        NumDimK,
+                                        ADataType,
+                                        BDataType,
+                                        DsDataType,
+                                        EDataType,
+                                        AElementwiseOperation,
+                                        BElementwiseOperation,
+                                        CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceContractionMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    // Assume: A[M0, M1, M2, ..., K0, K1, K2, ...]
+    static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_ms_ks_strides_vec)
+    {
+        assert(a_ms_ks_lengths_vec.size() == NumDimM + NumDimK &&
+               a_ms_ks_strides_vec.size() == NumDimM + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto a_ms_ns_lengths = to_tuple(a_ms_ks_lengths_vec, Number<NumDimM + NumDimK>{});
+        const auto a_ms_ks_strides = to_tuple(a_ms_ks_strides_vec, Number<NumDimM + NumDimK>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimK, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(a_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(a_ms_ns_lengths, kDimIds);
+
+        // naive tensor A[M0, M1, M2, ..., K0, K1, K2...]
+        const auto a_grid_desc_ms_ks =
+            make_naive_tensor_descriptor(a_ms_ns_lengths, a_ms_ks_strides);
+
+        // transformed tensor A[MRaw = M0 * M1 * M2 * ... , KRaw = K0 * K1 * K2 * ...]
+        const auto a_grid_desc_mraw_kraw = transform_tensor_descriptor(
+            a_grid_desc_ms_ks,
+            make_tuple(make_merge_transform(mLengths), make_merge_transform(kLengths)),
+            make_tuple(mDimIds, kDimIds),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    // Assume: B[N0, N1, N2, ..., K0, K1, K2, ...]
+    static auto MakeBGridDescriptor_N_K(const std::vector<index_t>& b_ns_ks_lengths_vec,
+                                        const std::vector<index_t>& b_ns_ks_strides_vec)
+    {
+        assert(b_ns_ks_lengths_vec.size() == NumDimN + NumDimK &&
+               b_ns_ks_strides_vec.size() == NumDimN + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto b_ns_ks_lengths = to_tuple(b_ns_ks_lengths_vec, Number<NumDimN + NumDimK>{});
+        const auto b_ns_ks_strides = to_tuple(b_ns_ks_strides_vec, Number<NumDimN + NumDimK>{});
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<0, NumDimN, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimN, NumDimN + NumDimK, 1>::type{};
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(b_ns_ks_lengths, kDimIds);
+
+        // lengths for N0, N1, ...
+        const auto nLengths = get_container_subset(b_ns_ks_lengths, nDimIds);
+
+        // naive tensor B[N0, N1, N2, ..., K0, K1, K2, ...]
+        const auto b_grid_desc_ns_ks =
+            make_naive_tensor_descriptor(b_ns_ks_lengths, b_ns_ks_strides);
+
+        // transformed tensor B[NRaw = N0 * N1 * N2 * ..., KRaw = K0 * K1 * K2 * ...]
+        const auto b_grid_desc_nraw_kraw = transform_tensor_descriptor(
+            b_grid_desc_ns_ks,
+            make_tuple(make_merge_transform(nLengths), make_merge_transform(kLengths)),
+            make_tuple(nDimIds, kDimIds),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    // assume E[M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_M_N(const std::vector<index_t>& e_ms_ns_lengths_vec,
+                                        const std::vector<index_t>& e_ms_ns_strides_vec)
+    {
+        assert(e_ms_ns_lengths_vec.size() == NumDimM + NumDimN &&
+               e_ms_ns_strides_vec.size() == NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto e_ms_ns_lengths = to_tuple(e_ms_ns_lengths_vec, Number<NumDimM + NumDimN>{});
+        const auto e_ms_ns_strides = to_tuple(e_ms_ns_strides_vec, Number<NumDimM + NumDimN>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_ms_ns_lengths, nDimIds);
+
+        // naive tensor E[M0, M1, M2, ..., N0, N1, N2...]
+        const auto e_grid_desc_ms_ns =
+            make_naive_tensor_descriptor(e_ms_ns_lengths, e_ms_ns_strides);
+
+        // transformed tensor E[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
+        const auto e_grid_desc_mraw_nraw = transform_tensor_descriptor(
+            e_grid_desc_ms_ns,
+            make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+            make_tuple(mDimIds, nDimIds),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_M_N(ds_ms_ns_lengths_vec[i],
+                                                         ds_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K({}, {}));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K({}, {}));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({{}}, {{}}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N({}, {}));
+
+    using ComputeDataType = ADataType;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeDataType,
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // desc for blockwise copy
+    using AGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
+            AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
+            BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
+        decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+
+    // block-to-e-tile map
+    using Block2ETileMap =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 const std::vector<index_t>& a_ms_ns_lengths,
+                 const std::vector<index_t>& a_ms_ks_strides,
+                 const std::vector<index_t>& b_ns_ks_lengths,
+                 const std::vector<index_t>& b_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_strides,
+                 const std::vector<index_t>& e_ms_ns_lengths,
+                 const std::vector<index_t>& e_ms_ns_strides,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(a_ms_ns_lengths, a_ms_ks_strides)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(b_ns_ks_lengths, b_ns_ks_strides)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N(e_ms_ns_lengths, e_ms_ns_strides)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_mz_stride_{},
+              a_kz_stride_{},
+              b_nz_stride_{},
+              b_kz_stride_{},
+              ds_nz_stride_{},
+              e_nz_stride_{}
+        {
+            // populate pointer, batch stride, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) =
+                    DeviceOp::MakeEGridDescriptor_M_N(ds_ms_ns_lengths[i], ds_ms_ns_strides[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           ds_grid_desc_m_n_,
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+            }
+
+            // for sanity check of vector memory access
+            a_mz_stride_ = a_ms_ks_strides[NumDimM - 1];
+            a_kz_stride_ = a_ms_ks_strides[NumDimM + NumDimK - 1];
+
+            b_nz_stride_ = b_ns_ks_strides[NumDimN - 1];
+            b_kz_stride_ = b_ns_ks_strides[NumDimN + NumDimK - 1];
+
+            for(index_t i = 0; i < NumDTensor; ++i)
+            {
+                ds_nz_stride_[i] = ds_ms_ns_strides[i][NumDimM + NumDimN - 1];
+            }
+
+            e_nz_stride_ = e_ms_ns_strides[NumDimM + NumDimN - 1];
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // Strides for the last M/N/K dimensions of A/B/Ds/E
+        //   for sanity check of vector load/store
+        index_t a_mz_stride_;
+        index_t a_kz_stride_;
+        index_t b_nz_stride_;
+        index_t b_kz_stride_;
+        std::array<index_t, NumDTensor> ds_nz_stride_;
+        index_t e_mz_stride_;
+        index_t e_nz_stride_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.ds_grid_desc_m_n_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemmMultipleD_xdl_cshuffle has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_contraction_multiple_d_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    DeviceOp::Block2ETileMap,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_etile_map_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if(ck::get_device_name() != "gfx90a" && std::is_same<ADataType, double>::value)
+        {
+            return false;
+        }
+
+        if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                        arg.b_grid_desc_n_k_,
+                                        arg.ds_grid_desc_m_n_,
+                                        arg.e_grid_desc_m_n_,
+                                        arg.block_2_etile_map_))
+        {
+            return false;
+        }
+
+        // check vector access
+        static_assert((ABlockTransferSrcVectorDim == 1 || ABlockTransferSrcVectorDim == 2) &&
+                          (BBlockTransferSrcVectorDim == 1 || BBlockTransferSrcVectorDim == 2),
+                      "wrong!");
+
+        // vector memory access of A: could be on M or AK1 dimension
+        if constexpr(ABlockTransferSrcVectorDim == 1)
+        {
+            if(!(arg.a_mz_stride_ == 1 &&
+                 arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.a_kz_stride_ == 1 &&
+                 arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of B: could be on N or BK1 dimension
+        if constexpr(BBlockTransferSrcVectorDim == 1)
+        {
+            if(!(arg.b_nz_stride_ == 1 &&
+                 arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.b_kz_stride_ == 1 &&
+                 arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of Ds: always on NPerBlock dimension
+        bool valid_d_access = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            if(!(arg.ds_nz_stride_[i] == 1 &&
+                 arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_[i].GetLength(I3) %
+                         CDEBlockTransferScalarPerVector_NPerBlock ==
+                     0))
+            {
+                valid_d_access = false;
+            }
+        });
+
+        if(valid_d_access == false)
+        {
+            return false;
+        }
+
+        // vector memory access of E: always on NPerBlock dimension
+        if(!(arg.e_nz_stride_ == 1 &&
+             arg.e_grid_desc_mblock_mperblock_nblock_nperblock_.GetLength(I3) %
+                     CDEBlockTransferScalarPerVector_NPerBlock ==
+                 0))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             std::array<const void*, NumDTensor> p_ds,
+                             void* p_e,
+                             const std::vector<index_t>& a_ms_ns_lengths,
+                             const std::vector<index_t>& a_ms_ks_strides,
+                             const std::vector<index_t>& b_ns_ks_lengths,
+                             const std::vector<index_t>& b_ns_ks_strides,
+                             const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_lengths,
+                             const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_strides,
+                             const std::vector<index_t>& e_ms_ns_lengths,
+                             const std::vector<index_t>& e_ms_ns_strides,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_ms_ns_lengths,
+                        a_ms_ks_strides,
+                        b_ns_ks_lengths,
+                        b_ns_ks_strides,
+                        ds_ms_ns_lengths,
+                        ds_ms_ns_strides,
+                        e_ms_ns_lengths,
+                        e_ms_ns_strides,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::vector<index_t>& a_ms_ns_lengths,
+                        const std::vector<index_t>& a_ms_ks_strides,
+                        const std::vector<index_t>& b_ns_ks_lengths,
+                        const std::vector<index_t>& b_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_strides,
+                        const std::vector<index_t>& e_ms_ns_lengths,
+                        const std::vector<index_t>& e_ms_ns_strides,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_ms_ns_lengths,
+                                          a_ms_ks_strides,
+                                          b_ns_ks_lengths,
+                                          b_ns_ks_strides,
+                                          ds_ms_ns_lengths,
+                                          ds_ms_ns_strides,
+                                          e_ms_ns_lengths,
+                                          e_ms_ns_strides,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceContractionMultipleD_Xdl_CShuffle"
+            << "<"
+            << NumDimM << ", "
+            << NumDimN << ", "
+            << NumDimK << ", "
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << ABlockTransferSrcVectorDim << ", "
+            << BBlockTransferSrcVectorDim
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000000..23440e24f6
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,808 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_weight.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+template <typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXdl,
+          ck::index_t NPerXdl,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXdl>
+struct DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
+    : public DeviceConvBwdWeight<2,
+                                 ck::tensor_layout::convolution::NHWC,
+                                 ck::tensor_layout::convolution::KYXC,
+                                 ck::tensor_layout::convolution::NHWK,
+                                 InDataType,
+                                 WeiDataType,
+                                 OutDataType,
+                                 InElementwiseOperation,
+                                 WeiElementwiseOperation,
+                                 OutElementwiseOperation>
+{
+    static constexpr ck::index_t NDimSpatial = 2;
+
+    using DeviceOp =
+        DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
+
+    using ADataType = OutDataType;
+    using BDataType = InDataType;
+    using CDataType = WeiDataType;
+
+    using AElementwiseOperation = OutElementwiseOperation;
+    using BElementwiseOperation = InElementwiseOperation;
+    using CElementwiseOperation = WeiElementwiseOperation;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    static constexpr auto N1Number = K1Number;
+
+    // Bytes per 32 lds bank: 32 * 4 bytes
+    static constexpr auto BankLength = 128;
+    static constexpr auto ElePerBank = BankLength / sizeof(ADataType);
+
+    // M1 & M0
+    static constexpr auto ABlockLdsM1PerBlock = ElePerBank / K1;
+    static constexpr auto ABlockLdsM0PerBlock = MPerBlock / ABlockLdsM1PerBlock;
+    static constexpr auto ABlockLdsM1Padding  = 4;
+
+    // N1 & N0
+    static constexpr auto BBlockLdsN1PerBlock = ElePerBank / K1;
+    static constexpr auto BBlockLdsN0PerBlock = NPerBlock / BBlockLdsN1PerBlock;
+    static constexpr auto BBlockLdsN1Padding  = 4;
+
+    static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        ck::index_t N,
+        ck::index_t K,
+        ck::index_t C,
+        std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
+        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
+        std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
+        std::array<ck::index_t, NDimSpatial> conv_filter_strides,
+        std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
+        std::array<ck::index_t, NDimSpatial> input_left_pads,
+        std::array<ck::index_t, NDimSpatial> input_right_pads,
+        ck::index_t batch_k)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t GemmKTotal = N * Ho * Wo;
+        const index_t GemmM      = K;
+        const index_t GemmN      = C * X * Y;
+
+        const index_t GemmKBatch = batch_k;
+        const index_t GemmK0 =
+            math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
+            K0PerBlock;
+
+        const auto in_n_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+        // A: output tensor
+        const index_t N0          = N / N1Number;
+        const index_t GemmK0Total = N0 * Ho * Wo;
+
+        const index_t GemmK0S =
+            math::integer_divide_ceil(GemmK0Total, K0PerBlock * GemmKBatch) * K0PerBlock;
+        const index_t GemmK0Pad = GemmKBatch * GemmK0S;
+        const auto out_n_ho_wo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Ho * Wo, K));
+
+        const auto out_n0_ho_wo_k_n1_grid_desc =
+            transform_tensor_descriptor(out_n_ho_wo_k_grid_desc,
+                                        make_tuple(make_unmerge_transform(make_tuple(N0, N1Number)),
+                                                   make_pass_through_transform(Ho * Wo),
+                                                   make_pass_through_transform(K)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                                        make_tuple(Sequence<0, 3>{}, Sequence<1>{}, Sequence<2>{}));
+
+        const auto out_gemmk0total_gemmm_gemmk1_grid_desc =
+            transform_tensor_descriptor(out_n0_ho_wo_k_n1_grid_desc,
+                                        make_tuple(make_merge_transform(make_tuple(N0, Ho * Wo)),
+                                                   make_pass_through_transform(K),
+                                                   make_pass_through_transform(N1Number)),
+                                        make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<3>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        const auto out_gemmk0pad_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+            out_gemmk0total_gemmm_gemmk1_grid_desc,
+            make_tuple(make_right_pad_transform(GemmK0Total, GemmK0Pad - GemmK0Total),
+                       make_pass_through_transform(GemmM),
+                       make_pass_through_transform(N1Number)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+            out_gemmk0pad_gemmm_gemmk1_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0)),
+                       make_pass_through_transform(GemmM),
+                       make_pass_through_transform(N1Number)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<3>{}));
+
+        // B: input tensor
+        const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+            in_n_hi_wi_c_grid_desc,
+            make_tuple(make_pass_through_transform(N),
+                       make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                       make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+        const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+            in_n_hip_wip_c_grid_desc,
+            make_tuple(
+                make_pass_through_transform(N),
+                make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+        const auto in_n0_y_ho_x_wo_c_n1_grid_desc =
+            transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                        make_tuple(make_unmerge_transform(make_tuple(N0, N1Number)),
+                                                   make_pass_through_transform(Y),
+                                                   make_pass_through_transform(Ho),
+                                                   make_pass_through_transform(X),
+                                                   make_pass_through_transform(Wo),
+                                                   make_pass_through_transform(C)),
+                                        make_tuple(Sequence<0>{},
+                                                   Sequence<1>{},
+                                                   Sequence<2>{},
+                                                   Sequence<3>{},
+                                                   Sequence<4>{},
+                                                   Sequence<5>{}),
+                                        make_tuple(Sequence<0, 6>{},
+                                                   Sequence<1>{},
+                                                   Sequence<2>{},
+                                                   Sequence<3>{},
+                                                   Sequence<4>{},
+                                                   Sequence<5>{}));
+
+        const auto in_gemmk0total_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+            in_n0_y_ho_x_wo_c_n1_grid_desc,
+            make_tuple(make_merge_transform(make_tuple(N0, Ho, Wo)),
+                       make_merge_transform(make_tuple(Y, X, C)),
+                       make_pass_through_transform(N1Number)),
+            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}, Sequence<6>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        const auto in_gemmk0pad_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+            in_gemmk0total_gemmn_gemmk1_grid_desc,
+            make_tuple(make_right_pad_transform(GemmK0Total, GemmK0Pad - GemmK0Total),
+                       make_pass_through_transform(GemmN),
+                       make_pass_through_transform(N1Number)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+            in_gemmk0pad_gemmn_gemmk1_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0)),
+                       make_pass_through_transform(GemmN),
+                       make_pass_through_transform(N1Number)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<3>{}));
+
+        // C: weight tensor
+        const auto wei_gemmm_gemmn_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+        return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                          in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                          wei_gemmm_gemmn_grid_desc);
+    }
+
+    using ABCGridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, 1));
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXdl,
+        NPerXdl,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        ABlockLdsM1PerBlock,
+        ABlockLdsM0PerBlock,
+        ABlockLdsM1Padding,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        BBlockLdsN1PerBlock,
+        BBlockLdsN0PerBlock,
+        BBlockLdsN1Padding,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CBlockTransferScalarPerVector_NWaveNPerXdl,
+        CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        true,
+        true>;
+
+    using GridwiseGemmAtomicAdd = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::AtomicAdd,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXdl,
+        NPerXdl,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        ABlockLdsM1PerBlock,
+        ABlockLdsM0PerBlock,
+        ABlockLdsM1Padding,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        BBlockLdsN1PerBlock,
+        BBlockLdsN0PerBlock,
+        BBlockLdsN1Padding,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CBlockTransferScalarPerVector_NWaveNPerXdl,
+        CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        true,
+        true>;
+    // Argument
+    using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        decltype(GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}));
+
+    using Block2CTileMap =
+        decltype(GridwiseGemm::MakeCBlockClusterAdaptor(CGridDesc_M_N{}, 1, 1, 1));
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 WeiDataType* p_wei_grid,
+                 const OutDataType* p_out_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
+                 std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
+                 std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
+                 std::array<ck::index_t, NDimSpatial> conv_filter_strides,
+                 std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
+                 std::array<ck::index_t, NDimSpatial> input_left_pads,
+                 std::array<ck::index_t, NDimSpatial> input_right_pads,
+                 ck::index_t M01,
+                 ck::index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op,
+                 ck::index_t split_k)
+            : p_a_grid_{p_out_grid},
+              p_b_grid_{p_in_grid},
+              p_c_grid_{p_wei_grid},
+              a_grid_desc_kbatch_k0_m_k1_{},
+              b_grid_desc_kbatch_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{out_element_op},
+              b_element_op_{in_element_op},
+              c_element_op_{wei_element_op},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              output_spatial_lengths_{output_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads},
+              k_batch_{split_k}
+        {
+            const auto descs =
+                DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(N,
+                                                                          K,
+                                                                          C,
+                                                                          input_spatial_lengths,
+                                                                          filter_spatial_lengths,
+                                                                          output_spatial_lengths,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads,
+                                                                          k_batch_);
+
+            a_grid_desc_kbatch_k0_m_k1_ = descs[I0];
+            b_grid_desc_kbatch_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_            = descs[I2];
+
+            block_2_ctile_map_ =
+                GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_);
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_kbatch_k0_m_k1_,
+                                           b_grid_desc_kbatch_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n_);
+            }
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_kbatch_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_kbatch_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        Block2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        InElementwiseOperation a_element_op_;
+        OutElementwiseOperation b_element_op_;
+        WeiElementwiseOperation c_element_op_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+        std::array<index_t, NDimSpatial> output_spatial_lengths_;
+        std::array<index_t, NDimSpatial> filter_spatial_lengths_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+        index_t k_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        void Print(const Argument& arg)
+        {
+            std::cout << "arg.a_grid_desc_kbatch_k0_m_k1_{"
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I2) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I3) << "}" << std::endl;
+
+            std::cout << "arg.b_grid_desc_kbatch_k0_n_k1_{"
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I0) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I1) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I2) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I3) << "}" << std::endl;
+
+            std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                      << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+        }
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                Print(arg);
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
+                                            arg.b_grid_desc_kbatch_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight has invalid setting");
+            }
+            const auto kbatch = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0);
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K0 = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1);
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            float ave_time = 0;
+
+            const auto Run = [&](const auto& kernel) {
+                hipGetErrorString(hipMemsetAsync(
+                    arg.p_c_grid_,
+                    0,
+                    arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
+                        sizeof(CDataType),
+                    stream_config.stream_id_));
+
+                ave_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.a_grid_desc_kbatch_k0_m_k1_,
+                                           arg.b_grid_desc_kbatch_k0_n_k1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.block_2_ctile_map_);
+            };
+
+            if(has_main_k0_block_loop)
+            {
+                if(kbatch == 1)
+                {
+                    const auto kernel = kernel_gemm_xdlops_bwd_weight<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        OutElementwiseOperation,
+                        InElementwiseOperation,
+                        WeiElementwiseOperation,
+                        remove_reference_t<DeviceOp::Block2CTileMap>,
+                        true>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel = kernel_gemm_xdlops_bwd_weight<
+                        GridwiseGemmAtomicAdd,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        OutElementwiseOperation,
+                        InElementwiseOperation,
+                        WeiElementwiseOperation,
+                        remove_reference_t<DeviceOp::Block2CTileMap>,
+                        true>;
+
+                    Run(kernel);
+                }
+            }
+            else
+            {
+                if(kbatch == 1)
+                {
+                    const auto kernel = kernel_gemm_xdlops_bwd_weight<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        OutElementwiseOperation,
+                        InElementwiseOperation,
+                        WeiElementwiseOperation,
+                        remove_reference_t<DeviceOp::Block2CTileMap>,
+                        false>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel = kernel_gemm_xdlops_bwd_weight<
+                        GridwiseGemmAtomicAdd,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        OutElementwiseOperation,
+                        InElementwiseOperation,
+                        WeiElementwiseOperation,
+                        remove_reference_t<DeviceOp::Block2CTileMap>,
+                        false>;
+
+                    Run(kernel);
+                }
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 2 &&
+             arg.Conv_K_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // unmerge N to N0 and N1, where N1 equals to K1
+        if(!(arg.Conv_N_ % K1 == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_C_ % CBlockTransferScalarPerVector_NWaveNPerXdl == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
+                                           arg.b_grid_desc_kbatch_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in_grid,
+                             WeiDataType* p_wei_grid,
+                             const OutDataType* p_out_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
+                             std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
+                             std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
+                             std::array<ck::index_t, NDimSpatial> conv_filter_strides,
+                             std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
+                             std::array<ck::index_t, NDimSpatial> input_left_pads,
+                             std::array<ck::index_t, NDimSpatial> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op,
+                             ck::index_t split_k)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op,
+                        split_k};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        void* p_wei_grid,
+                        const void* p_out_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::array<ck::index_t, NDimSpatial> input_spatial_lengths,
+                        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths,
+                        std::array<ck::index_t, NDimSpatial> output_spatial_lengths,
+                        std::array<ck::index_t, NDimSpatial> conv_filter_strides,
+                        std::array<ck::index_t, NDimSpatial> conv_filter_dilations,
+                        std::array<ck::index_t, NDimSpatial> input_left_pads,
+                        std::array<ck::index_t, NDimSpatial> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op,
+                        ck::index_t split_k) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<WeiDataType*>(p_wei_grid),
+                                          static_cast<const OutDataType*>(p_out_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op,
+                                          split_k);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv2dBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << ABlockTransferDstScalarPerVector_K1 << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferDstScalarPerVector_K1 << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle << ", "
+            << CBlockTransferScalarPerVector_NWaveNPerXdl
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000000..e22c5a2aa5
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,765 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+template <typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ConvolutionBackwardDataSpecialization ConvBackwardDataSpecialization,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXdl,
+          ck::index_t NPerXdl,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector>
+struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
+    : public DeviceConvBwdData<2,
+                               ck::tensor_layout::convolution::NHWC,
+                               ck::tensor_layout::convolution::KYXC,
+                               ck::tensor_layout::convolution::NHWK,
+                               InDataType,
+                               WeiDataType,
+                               OutDataType,
+                               InElementwiseOperation,
+                               WeiElementwiseOperation,
+                               OutElementwiseOperation>
+{
+    using DeviceOp = DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
+
+    using ADataType = OutDataType;
+    using BDataType = WeiDataType;
+    using CDataType = InDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr index_t NDimSpatial = 2;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static_assert((K1 % ABlockTransferThreadClusterLengths_K0_M_K1{}[I2]) %
+                      ABlockTransferSrcScalarPerVector ==
+                  0);
+    static_assert((NPerBlock / BBlockTransferThreadClusterLengths_K0_N_K1{}[I1]) %
+                      BBlockTransferSrcScalarPerVector ==
+                  0);
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads,
+                                                    index_t i_ytilde,
+                                                    index_t i_xtilde)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const auto K0 = K / K1;
+
+        const auto out_n_ho_wo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Ho, Wo, K));
+        const auto wei_k_y_x_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Y, X, C));
+        const auto in_n_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+                make_tuple(make_pass_through_transform(N * Ho * Wo),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: input tensor
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_freeze_transform(I0),
+                           make_freeze_transform(I0),
+                           make_merge_transform(make_tuple(N, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<1>{}, Sequence<3>{}, Sequence<0, 2, 4>{}, Sequence<5>{}),
+                make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto YDot = math::integer_divide_ceil(Y, YTilde);
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // A: output tensor
+            const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                out_n_ho_wo_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Ho, I0, I0),
+                           make_pad_transform(Wo, I0, I0),
+                           make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto out_n_ydot_htilde_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
+                out_n_hop_wop_k_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(YDot, HTilde),
+                                         make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                    make_embed_transform(make_tuple(XDot, WTilde),
+                                         make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                    make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc =
+                transform_tensor_descriptor(
+                    out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_slice_transform(YDot, I0, YDotSlice),
+                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                               make_slice_transform(XDot, I0, XDotSlice),
+                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                               make_unmerge_transform(make_tuple(K0, K1))),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5, 6>{}));
+
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                           make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B weight tensor
+            const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
+                wei_k_y_x_c_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_embed_transform(make_tuple(YDot, YTilde),
+                                                make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                           make_embed_transform(make_tuple(XDot, XTilde),
+                                                make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto wei_k0_k1_ydotslice_xdotslice_c_grid_desc =
+                transform_tensor_descriptor(wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_slice_transform(YDot, I0, YDotSlice),
+                                                       make_slice_transform(XDot, I0, XDotSlice),
+                                                       make_freeze_transform(i_ytilde),
+                                                       make_freeze_transform(i_xtilde),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{},
+                                                       Sequence<1>{},
+                                                       Sequence<3>{},
+                                                       Sequence<2>{},
+                                                       Sequence<4>{},
+                                                       Sequence<5>{}),
+                                            make_tuple(Sequence<0, 1>{},
+                                                       Sequence<2>{},
+                                                       Sequence<3>{},
+                                                       Sequence<>{},
+                                                       Sequence<>{},
+                                                       Sequence<4>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_k0_k1_ydotslice_xdotslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                           make_pass_through_transform(C),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<2, 3, 0>{}, Sequence<4>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // C: input tensor
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(YTilde, HTilde),
+                                                make_tuple(ConvDilationH, ConvStrideH)),
+                           make_embed_transform(make_tuple(XTilde, WTilde),
+                                                make_tuple(ConvDilationW, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
+                in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_freeze_transform(i_ytilde),
+                           make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                           make_freeze_transform(i_xtilde),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<>{},
+                           Sequence<1>{},
+                           Sequence<>{},
+                           Sequence<2>{},
+                           Sequence<3>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_htildeslice_wtildeslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+
+    using ABCGridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, 0, 0));
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXdl,
+        NPerXdl,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        Sequence<2, 3, 0, 1, 7, 5, 4, 6>, // CThreadTransferSrcDstAccessOrder,
+        7,                                // CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 const OutDataType* p_out_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads)
+            : p_a_grid_{p_out_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_in_grid},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              input_spatial_lengths_{input_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              output_spatial_lengths_{output_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            const index_t ConvStrideH = conv_filter_strides[0];
+            const index_t ConvStrideW = conv_filter_strides[1];
+
+            const index_t ConvDilationH = conv_filter_dilations[0];
+            const index_t ConvDilationW = conv_filter_dilations[1];
+
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
+            {
+                for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+                {
+                    // check slice is valid
+                    const index_t Y      = filter_spatial_lengths_[0];
+                    const index_t X      = filter_spatial_lengths_[1];
+                    const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+                    const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+                    if(YDotSlice * XDotSlice <= 0)
+                    {
+                        continue;
+                    }
+
+                    const auto descs = DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        i_ytilde,
+                        i_xtilde);
+                    a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
+                    b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
+                    c_grid_desc_m_n_container_.push_back(descs[I2]);
+                }
+            }
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        std::vector<AGridDesc_K0_M_K1> a_grid_desc_k0_m_k1_container_;
+        std::vector<BGridDesc_K0_N_K1> b_grid_desc_k0_n_k1_container_;
+        std::vector<CGridDesc_M_N> c_grid_desc_m_n_container_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+
+        std::vector<ck::index_t> input_spatial_lengths_;
+        std::vector<ck::index_t> filter_spatial_lengths_;
+        std::vector<ck::index_t> output_spatial_lengths_;
+        std::vector<ck::index_t> conv_filter_strides_;
+        std::vector<ck::index_t> conv_filter_dilations_;
+        std::vector<ck::index_t> input_left_pads_;
+        std::vector<ck::index_t> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            float ave_time = 0;
+            for(size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
+            {
+#if DEBUG_LOG
+                {
+                    std::cout << "arg.a_grid_desc_k0_m_k1_container_{"
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) << ", "
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I1) << ", "
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.b_grid_desc_k0_n_k1_container_{"
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I0) << ", "
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I1) << ", "
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I2) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.c_grid_desc_m_n_container_{ "
+                              << arg.c_grid_desc_m_n_container_[i].GetLength(I0) << ", "
+                              << arg.c_grid_desc_m_n_container_[i].GetLength(I1) << "}"
+                              << std::endl;
+                }
+#endif
+
+                if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
+                                                arg.b_grid_desc_k0_n_k1_container_[i],
+                                                arg.c_grid_desc_m_n_container_[i]))
+                {
+                    throw std::runtime_error(
+                        "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
+                }
+
+                const auto [gdx, gdy, gdz] =
+                    GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_container_[i]);
+
+                const auto K = arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) *
+                               arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2);
+
+                if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+                {
+                    const auto kernel =
+                        kernel_gemm_xdlops_v2r3<GridwiseGemm,
+                                                ADataType, // TODO: distiguish A/B datatype
+                                                CDataType,
+                                                DeviceOp::AGridDesc_K0_M_K1,
+                                                DeviceOp::BGridDesc_K0_N_K1,
+                                                DeviceOp::CGridDesc_M_N,
+                                                true>;
+
+                    ave_time += launch_and_time_kernel(stream_config,
+                                                       kernel,
+                                                       dim3(gdx, gdy, gdz),
+                                                       dim3(BlockSize),
+                                                       0,
+                                                       arg.p_a_grid_,
+                                                       arg.p_b_grid_,
+                                                       arg.p_c_grid_,
+                                                       arg.a_grid_desc_k0_m_k1_container_[i],
+                                                       arg.b_grid_desc_k0_n_k1_container_[i],
+                                                       arg.c_grid_desc_m_n_container_[i]);
+                }
+                else
+                {
+                    const auto kernel =
+                        kernel_gemm_xdlops_v2r3<GridwiseGemm,
+                                                ADataType, // TODO: distiguish A/B datatype
+                                                CDataType,
+                                                DeviceOp::AGridDesc_K0_M_K1,
+                                                DeviceOp::BGridDesc_K0_N_K1,
+                                                DeviceOp::CGridDesc_M_N,
+                                                false>;
+
+                    ave_time += launch_and_time_kernel(stream_config,
+                                                       kernel,
+                                                       dim3(gdx, gdy, gdz),
+                                                       dim3(BlockSize),
+                                                       0,
+                                                       arg.p_a_grid_,
+                                                       arg.p_b_grid_,
+                                                       arg.p_c_grid_,
+                                                       arg.a_grid_desc_k0_m_k1_container_[i],
+                                                       arg.b_grid_desc_k0_n_k1_container_[i],
+                                                       arg.c_grid_desc_m_n_container_[i]);
+                }
+            }
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 pad = 0 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.conv_filter_strides_[0] == 1 && arg.conv_filter_strides_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 1 &&
+             arg.Conv_K_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_C_ % CThreadTransferDstScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        for(std::size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
+                                            arg.b_grid_desc_k0_n_k1_container_[i],
+                                            arg.c_grid_desc_m_n_container_[i]))
+            {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             const OutDataType* p_out_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(void* p_in_grid,
+                        const void* p_wei_grid,
+                        const void* p_out_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation,
+                        WeiElementwiseOperation,
+                        OutElementwiseOperation) override
+    {
+        return std::make_unique<Argument>(static_cast<InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<const OutDataType*>(p_out_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << ABlockTransferDstScalarPerVector_K1 << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferDstScalarPerVector_K1
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000000..c9e8940edc
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,985 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] =
+//     activate(in[N, Hi, Wi, C] * wei[K, Y, X, C] + bias[K]) + residual[N, Ho, Wo, K]
+template <
+    typename InDataType,
+    typename WeiDataType,
+    typename OutDataType,
+    typename AccDataType,
+    typename InElementwiseOperation,
+    typename WeiElementwiseOperation,
+    typename OutElementwiseOperation,
+    ConvolutionForwardSpecialization ConvForwardSpecialization,
+    ck::index_t BlockSize,
+    ck::index_t MPerBlock,
+    ck::index_t NPerBlock,
+    ck::index_t K0PerBlock,
+    ck::index_t K1,
+    ck::index_t MPerXDL,
+    ck::index_t NPerXDL,
+    ck::index_t MXdlPerWave,
+    ck::index_t NXdlPerWave,
+    typename ABlockTransferThreadClusterLengths_K0_M_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    ck::index_t ABlockTransferSrcVectorDim,
+    ck::index_t ABlockTransferSrcScalarPerVector,
+    ck::index_t ABlockTransferDstScalarPerVector_K1,
+    bool ABlockLdsAddExtraM,
+    typename BBlockTransferThreadClusterLengths_K0_N_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    ck::index_t BBlockTransferSrcVectorDim,
+    ck::index_t BBlockTransferSrcScalarPerVector,
+    ck::index_t BBlockTransferDstScalarPerVector_K1,
+    bool BBlockLdsAddExtraN,
+    index_t CShuffleMXdlPerWavePerShuffle,
+    index_t CShuffleNXdlPerWavePerShuffle,
+    typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+    index_t CBlockTransferScalarPerVector_NWaveNPerXdl>
+struct
+    DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
+    : public DeviceConvFwdBiasActivationAdd<InElementwiseOperation,
+                                            WeiElementwiseOperation,
+                                            OutElementwiseOperation>
+{
+    using DeviceOp =
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
+
+    using ADataType = InDataType;
+    using BDataType = WeiDataType;
+    using CDataType = OutDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    // TODO make it support any # of spatial dimensions
+    static constexpr index_t NDimSpatial = 2;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t GemmMRaw = N * Ho * Wo;
+        const index_t GemmN    = K;
+
+        const auto GemmM    = math::integer_least_multiple(GemmMRaw, MPerBlock);
+        const auto GemmMPad = GemmM - GemmMRaw;
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        { // 1x1, stride=1, pad=0
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_gemmmraw_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, C));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmmraw_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_right_pad_transform(GemmMRaw, GemmMPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            // C1: residual tensor: assume same layout as output tensor
+            const auto resi_grid_desc_gemmm_gemmn = out_gemmm_gemmn_grid_desc;
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn,
+                              resi_grid_desc_gemmm_gemmn);
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        { // 1x1, pad=0
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_ho_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_n_ho_wo_c_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_merge_transform(make_tuple(N, Ho, Wo))),
+                make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            // C1: residual tensor: assume same layout as output tensor
+            const auto resi_grid_desc_gemmm_gemmn = out_gemmm_gemmn_grid_desc;
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn,
+                              resi_grid_desc_gemmm_gemmn);
+        }
+        else if constexpr(ConvForwardSpecialization == ConvolutionForwardSpecialization::OddC)
+        { // C = odd value
+            const index_t GemmKRaw = Y * X * C;
+            const index_t GemmK = math::integer_least_multiple(GemmKRaw, K0PerBlock * GemmK1Number);
+            const index_t GemmKPad = GemmK - GemmKRaw;
+            const index_t GemmK0   = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmkraw_gemmmraw_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk_gemmm_grid_desc = transform_tensor_descriptor(
+                in_gemmkraw_gemmmraw_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKRaw, GemmKPad),
+                           make_right_pad_transform(GemmMRaw, GemmMPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: weight tensor
+            const auto wei_k_yxc_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+                wei_k_yxc_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_right_pad_transform(GemmKRaw, GemmKPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmk_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_nhowo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmmraw_gemmn_grid_desc =
+                transform_tensor_descriptor(out_nhowo_k_grid_desc,
+                                            make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                                       make_pass_through_transform(K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            // C1: residual tensor: assume same layout as output tensor
+            const auto resi_grid_desc_gemmm_gemmn = out_gemmm_gemmn_grid_desc;
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn,
+                              resi_grid_desc_gemmm_gemmn);
+        }
+        else
+        {
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmk_gemmmraw_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk_gemmmraw_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmMRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_k_yxc_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+                wei_k_yxc_grid_desc,
+                make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmk_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_nhowo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmmraw_gemmn_grid_desc =
+                transform_tensor_descriptor(out_nhowo_k_grid_desc,
+                                            make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                                       make_pass_through_transform(K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            // C1: residual tensor: assume same layout as output tensor
+            const auto resi_grid_desc_gemmm_gemmn = out_gemmm_gemmn_grid_desc;
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn,
+                              resi_grid_desc_gemmm_gemmn);
+        }
+    }
+
+    using GridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}));
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(GridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(GridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(GridDescs{}[I2])>;
+    using C0GridDesc_M_N    = remove_cvref_t<decltype(GridDescs{}[I3])>;
+    using C1GridDesc_M_N    = remove_cvref_t<decltype(GridDescs{}[I4])>;
+
+    using Block2CTileMap = BlockToCTileMap_M00_N0_M01<MPerBlock, NPerBlock, CGridDesc_M_N>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        C0GridDesc_M_N,
+        C1GridDesc_M_N,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
+        2,                 // ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        Sequence<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // BBlockTransferSrcAccessOrder,
+        2,                 // BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+        CBlockTransferScalarPerVector_NWaveNPerXdl>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 OutDataType* p_out_grid,
+                 const OutDataType* p_bias_grid,
+                 const OutDataType* p_resi_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_in_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_out_grid},
+              p_c0_grid_{p_bias_grid},
+              p_c1_grid_{p_resi_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c0_grid_desc_m_n_{},
+              c1_grid_desc_m_n_{},
+              c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
+              c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
+              c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
+              block_2_ctile_map_{},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              input_spatial_lengths_{input_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              output_spatial_lengths_{output_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            const auto descs =
+                DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(N,
+                                                                          K,
+                                                                          C,
+                                                                          input_spatial_lengths,
+                                                                          filter_spatial_lengths,
+                                                                          output_spatial_lengths,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads);
+
+            a_grid_desc_k0_m_k1_ = descs[I0];
+            b_grid_desc_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_     = descs[I2];
+            c0_grid_desc_m_n_    = descs[I3];
+            c1_grid_desc_m_n_    = descs[I4];
+
+            block_2_ctile_map_ = Block2CTileMap{c_grid_desc_m_n_};
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
+                                           b_grid_desc_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
+                    GridwiseGemm::
+                        MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                            c_grid_desc_m_n_);
+
+                c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
+                    GridwiseGemm::
+                        MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                            c0_grid_desc_m_n_);
+
+                c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
+                    GridwiseGemm::
+                        MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                            c1_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        const CDataType* p_c0_grid_;
+        const CDataType* p_c1_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        C0GridDesc_M_N c0_grid_desc_m_n_;
+        C1GridDesc_M_N c1_grid_desc_m_n_;
+        typename GridwiseGemm::
+            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
+        typename GridwiseGemm::
+            C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
+        typename GridwiseGemm::
+            C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
+        Block2CTileMap block_2_ctile_map_;
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+        std::vector<index_t> input_spatial_lengths_;
+        std::vector<index_t> filter_spatial_lengths_;
+        std::vector<index_t> output_spatial_lengths_;
+        std::vector<index_t> conv_filter_strides_;
+        std::vector<index_t> conv_filter_dilations_;
+        std::vector<index_t> input_left_pads_;
+        std::vector<index_t> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if DEBUG_LOG
+            {
+                std::cout << DeviceOp{}.GetTypeString() << std::endl;
+                std::cout << "N " << arg.Conv_N_ << ", "
+                          << "K " << arg.Conv_K_ << ", "
+                          << "C " << arg.Conv_C_ << ", " << std::endl;
+                std::cout << "Y X " << arg.filter_spatial_lengths_[0] << ", "
+                          << arg.filter_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Hi Wi " << arg.input_spatial_lengths_[0] << ", "
+                          << arg.input_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Ho Wo " << arg.output_spatial_lengths_[0] << ", "
+                          << arg.output_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Strides " << arg.conv_filter_strides_[0] << ", "
+                          << arg.conv_filter_strides_[1] << ", " << std::endl;
+                std::cout << "Dilations " << arg.conv_filter_dilations_[0] << ", "
+                          << arg.conv_filter_dilations_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_left_pads_[0] << ", "
+                          << arg.input_left_pads_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_right_pads_[0] << ", "
+                          << arg.input_right_pads_[1] << ", " << std::endl;
+            }
+
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+
+                std::cout << "arg.c0_grid_desc_m_n_{ " << arg.c0_grid_desc_m_n_.GetLength(I0)
+                          << ", " << arg.c0_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+
+                std::cout << "arg.c1_grid_desc_m_n_{ " << arg.c1_grid_desc_m_n_.GetLength(I0)
+                          << ", " << arg.c1_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r3 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_xdlops_v3r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    Block2CTileMap,
+                    true>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    arg.p_a_grid_,
+                    arg.p_b_grid_,
+                    arg.p_c_grid_,
+                    arg.p_c0_grid_,
+                    arg.p_c1_grid_,
+                    arg.a_grid_desc_k0_m_k1_,
+                    arg.b_grid_desc_k0_n_k1_,
+                    arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.in_element_op_,
+                    arg.wei_element_op_,
+                    arg.out_element_op_,
+                    arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_v3r3<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    Block2CTileMap,
+                    false>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    arg.p_a_grid_,
+                    arg.p_b_grid_,
+                    arg.p_c_grid_,
+                    arg.p_c0_grid_,
+                    arg.p_c1_grid_,
+                    arg.a_grid_desc_k0_m_k1_,
+                    arg.b_grid_desc_k0_n_k1_,
+                    arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.in_element_op_,
+                    arg.wei_element_op_,
+                    arg.out_element_op_,
+                    arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.conv_filter_strides_[0] == 1 && arg.conv_filter_strides_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 2 &&
+             arg.Conv_C_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_K_ % CBlockTransferScalarPerVector_NWaveNPerXdl == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             OutDataType* p_out_grid,
+                             const OutDataType* p_bias_grid,
+                             const OutDataType* p_resi_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        p_bias_grid,
+                        p_resi_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        const void* p_wei_grid,
+                        void* p_out_grid,
+                        const void* p_bias_grid,
+                        const void* p_resi_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<OutDataType*>(p_out_grid),
+                                          static_cast<const OutDataType*>(p_bias_grid),
+                                          static_cast<const OutDataType*>(p_resi_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << ABlockTransferDstScalarPerVector_K1 << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferDstScalarPerVector_K1 << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle << ", "
+            << CBlockTransferScalarPerVector_NWaveNPerXdl
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000000..28fceb428e
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,942 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <vector>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] =
+//     activate(in[N, Hi, Wi, C] * wei[K, Y, X, C] + bias[K])
+template <
+    typename InDataType,
+    typename WeiDataType,
+    typename OutDataType,
+    typename AccDataType,
+    typename InElementwiseOperation,
+    typename WeiElementwiseOperation,
+    typename OutElementwiseOperation,
+    InMemoryDataOperationEnum OutGlobalMemoryDataOperation,
+    ConvolutionForwardSpecialization ConvForwardSpecialization,
+    ck::index_t BlockSize,
+    ck::index_t MPerBlock,
+    ck::index_t NPerBlock,
+    ck::index_t K0PerBlock,
+    ck::index_t K1,
+    ck::index_t MPerXDL,
+    ck::index_t NPerXDL,
+    ck::index_t MXdlPerWave,
+    ck::index_t NXdlPerWave,
+    typename ABlockTransferThreadClusterLengths_K0_M_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    ck::index_t ABlockTransferSrcVectorDim,
+    ck::index_t ABlockTransferSrcScalarPerVector,
+    ck::index_t ABlockTransferDstScalarPerVector_K1,
+    bool ABlockLdsAddExtraM,
+    typename BBlockTransferThreadClusterLengths_K0_N_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    ck::index_t BBlockTransferSrcVectorDim,
+    ck::index_t BBlockTransferSrcScalarPerVector,
+    ck::index_t BBlockTransferDstScalarPerVector_K1,
+    bool BBlockLdsAddExtraN,
+    index_t CShuffleMXdlPerWavePerShuffle,
+    index_t CShuffleNXdlPerWavePerShuffle,
+    typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+    index_t CBlockTransferScalarPerVector_NWaveNPerXdl>
+struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
+    : public DeviceConvFwdBiasActivation<InElementwiseOperation,
+                                         WeiElementwiseOperation,
+                                         OutElementwiseOperation>
+{
+    using DeviceOp =
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
+
+    using ADataType = InDataType;
+    using BDataType = WeiDataType;
+    using CDataType = OutDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    // TODO make it support any # of spatial dimensions
+    static constexpr index_t NDimSpatial = 2;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t GemmMRaw = N * Ho * Wo;
+        const index_t GemmN    = K;
+
+        const auto GemmM    = math::integer_least_multiple(GemmMRaw, MPerBlock);
+        const auto GemmMPad = GemmM - GemmMRaw;
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        { // 1x1, stride=1, pad=0
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_gemmmraw_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, C));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmmraw_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_right_pad_transform(GemmMRaw, GemmMPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn);
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        { // 1x1, pad=0
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_ho_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_n_ho_wo_c_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_merge_transform(make_tuple(N, Ho, Wo))),
+                make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn);
+        }
+        else if constexpr(ConvForwardSpecialization == ConvolutionForwardSpecialization::OddC)
+        { // C = odd value
+            const index_t GemmKRaw = Y * X * C;
+            const index_t GemmK = math::integer_least_multiple(GemmKRaw, K0PerBlock * GemmK1Number);
+            const index_t GemmKPad = GemmK - GemmKRaw;
+            const index_t GemmK0   = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmkraw_gemmmraw_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk_gemmm_grid_desc = transform_tensor_descriptor(
+                in_gemmkraw_gemmmraw_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKRaw, GemmKPad),
+                           make_right_pad_transform(GemmMRaw, GemmMPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: weight tensor
+            const auto wei_k_yxc_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+                wei_k_yxc_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_right_pad_transform(GemmKRaw, GemmKPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmk_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_nhowo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmmraw_gemmn_grid_desc =
+                transform_tensor_descriptor(out_nhowo_k_grid_desc,
+                                            make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                                       make_pass_through_transform(K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn);
+        }
+        else
+        {
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmk_gemmmraw_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk_gemmmraw_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmMRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_k_yxc_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+                wei_k_yxc_grid_desc,
+                make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmk_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_nhowo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmmraw_gemmn_grid_desc =
+                transform_tensor_descriptor(out_nhowo_k_grid_desc,
+                                            make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                                       make_pass_through_transform(K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            // C0: bias tensor: assume a contiguous vector
+            const auto bias_grid_desc_gemmm_gemmn =
+                make_naive_tensor_descriptor(make_tuple(GemmM, GemmN), make_tuple(I0, I1));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc,
+                              bias_grid_desc_gemmm_gemmn);
+        }
+    }
+
+    using ABCGridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}));
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+    using C0GridDesc_M_N    = remove_cvref_t<decltype(ABCGridDescs{}[I3])>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        OutGlobalMemoryDataOperation,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        C0GridDesc_M_N,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
+        2,                 // ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        Sequence<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // BBlockTransferSrcAccessOrder,
+        2,                 // BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+        CBlockTransferScalarPerVector_NWaveNPerXdl>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 OutDataType* p_out_grid,
+                 const OutDataType* p_bias_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 ck::index_t M01,
+                 ck::index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_in_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_out_grid},
+              p_c0_grid_{p_bias_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c0_grid_desc_m_n_{},
+              c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
+              c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              input_spatial_lengths_{input_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              output_spatial_lengths_{output_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            const auto descs =
+                DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(N,
+                                                                          K,
+                                                                          C,
+                                                                          input_spatial_lengths,
+                                                                          filter_spatial_lengths,
+                                                                          output_spatial_lengths,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads);
+
+            a_grid_desc_k0_m_k1_ = descs[I0];
+            b_grid_desc_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_     = descs[I2];
+            c0_grid_desc_m_n_    = descs[I3];
+            block_2_ctile_map_ =
+                GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
+                                           b_grid_desc_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
+                    GridwiseGemm::
+                        MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                            c_grid_desc_m_n_);
+
+                c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
+                    GridwiseGemm::
+                        MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                            c0_grid_desc_m_n_);
+            }
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        const CDataType* p_c0_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        C0GridDesc_M_N c0_grid_desc_m_n_;
+        typename GridwiseGemm::
+            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
+        typename GridwiseGemm::
+            C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+        std::vector<index_t> input_spatial_lengths_;
+        std::vector<index_t> filter_spatial_lengths_;
+        std::vector<index_t> output_spatial_lengths_;
+        std::vector<index_t> conv_filter_strides_;
+        std::vector<index_t> conv_filter_dilations_;
+        std::vector<index_t> input_left_pads_;
+        std::vector<index_t> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if DEBUG_LOG
+            {
+                std::cout << DeviceOp{}.GetTypeString() << std::endl;
+                std::cout << "N " << arg.Conv_N_ << ", "
+                          << "K " << arg.Conv_K_ << ", "
+                          << "C " << arg.Conv_C_ << ", " << std::endl;
+                std::cout << "Y X " << arg.filter_spatial_lengths_[0] << ", "
+                          << arg.filter_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Hi Wi " << arg.input_spatial_lengths_[0] << ", "
+                          << arg.input_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Ho Wo " << arg.output_spatial_lengths_[0] << ", "
+                          << arg.output_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Strides " << arg.conv_filter_strides_[0] << ", "
+                          << arg.conv_filter_strides_[1] << ", " << std::endl;
+                std::cout << "Dilations " << arg.conv_filter_dilations_[0] << ", "
+                          << arg.conv_filter_dilations_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_left_pads_[0] << ", "
+                          << arg.input_left_pads_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_right_pads_[0] << ", "
+                          << arg.input_right_pads_[1] << ", " << std::endl;
+            }
+
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+
+                std::cout << "arg.c0_grid_desc_m_n_{ " << arg.c0_grid_desc_m_n_.GetLength(I0)
+                          << ", " << arg.c0_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r2 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_xdlops_v3r2<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    true>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    arg.p_a_grid_,
+                    arg.p_b_grid_,
+                    arg.p_c_grid_,
+                    arg.p_c0_grid_,
+                    arg.a_grid_desc_k0_m_k1_,
+                    arg.b_grid_desc_k0_n_k1_,
+                    arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.in_element_op_,
+                    arg.wei_element_op_,
+                    arg.out_element_op_,
+                    arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_v3r2<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    false>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    arg.p_a_grid_,
+                    arg.p_b_grid_,
+                    arg.p_c_grid_,
+                    arg.p_c0_grid_,
+                    arg.a_grid_desc_k0_m_k1_,
+                    arg.b_grid_desc_k0_n_k1_,
+                    arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.in_element_op_,
+                    arg.wei_element_op_,
+                    arg.out_element_op_,
+                    arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.conv_filter_strides_[0] == 1 && arg.conv_filter_strides_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 2 &&
+             arg.Conv_C_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_K_ % CBlockTransferScalarPerVector_NWaveNPerXdl == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             OutDataType* p_out_grid,
+                             const OutDataType* p_bias_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        p_bias_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        const void* p_wei_grid,
+                        void* p_out_grid,
+                        const void* p_bias_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<OutDataType*>(p_out_grid),
+                                          static_cast<const OutDataType*>(p_bias_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << ABlockTransferDstScalarPerVector_K1 << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferDstScalarPerVector_K1 << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle << ", "
+            << CBlockTransferScalarPerVector_NWaveNPerXdl
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000000..ca291d3b11
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,908 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+template <
+    typename InDataType,
+    typename WeiDataType,
+    typename OutDataType,
+    typename AccDataType,
+    typename InElementwiseOperation,
+    typename WeiElementwiseOperation,
+    typename OutElementwiseOperation,
+    ConvolutionForwardSpecialization ConvForwardSpecialization,
+    ck::index_t BlockSize,
+    ck::index_t MPerBlock,
+    ck::index_t NPerBlock,
+    ck::index_t K0PerBlock,
+    ck::index_t K1,
+    ck::index_t MPerXdl,
+    ck::index_t NPerXdl,
+    ck::index_t MXdlPerWave,
+    ck::index_t NXdlPerWave,
+    typename ABlockTransferThreadClusterLengths_K0_M_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    ck::index_t ABlockTransferSrcVectorDim,
+    ck::index_t ABlockTransferSrcScalarPerVector,
+    ck::index_t ABlockTransferDstScalarPerVector_K1,
+    bool ABlockLdsAddExtraM,
+    typename BBlockTransferThreadClusterLengths_K0_N_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    ck::index_t BBlockTransferSrcVectorDim,
+    ck::index_t BBlockTransferSrcScalarPerVector,
+    ck::index_t BBlockTransferDstScalarPerVector_K1,
+    bool BBlockLdsAddExtraN,
+    index_t CShuffleMXdlPerWavePerShuffle,
+    index_t CShuffleNXdlPerWavePerShuffle,
+    typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+    index_t CBlockTransferScalarPerVector_NWaveNPerXdl>
+struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
+    : public DeviceConvFwd<2,
+                           ck::tensor_layout::convolution::NHWC,
+                           ck::tensor_layout::convolution::KYXC,
+                           ck::tensor_layout::convolution::NHWK,
+                           InDataType,
+                           WeiDataType,
+                           OutDataType,
+                           InElementwiseOperation,
+                           WeiElementwiseOperation,
+                           OutElementwiseOperation>
+{
+    using DeviceOp = DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
+
+    using ADataType = InDataType;
+    using BDataType = WeiDataType;
+    using CDataType = OutDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr index_t NDimSpatial = 2;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t GemmMRaw = N * Ho * Wo;
+        const index_t GemmN    = K;
+
+        const auto GemmM    = math::integer_least_multiple(GemmMRaw, MPerBlock);
+        const auto GemmMPad = GemmM - GemmMRaw;
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        { // 1x1, stride=1, pad=0
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_gemmmraw_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, C));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmmraw_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_right_pad_transform(GemmMRaw, GemmMPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc);
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        { // 1x1, pad=0
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_ho_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_n_ho_wo_c_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_merge_transform(make_tuple(N, Ho, Wo))),
+                make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc);
+        }
+        else if constexpr(ConvForwardSpecialization == ConvolutionForwardSpecialization::OddC)
+        { // C = odd value
+            const index_t GemmKRaw = Y * X * C;
+            const index_t GemmK = math::integer_least_multiple(GemmKRaw, K0PerBlock * GemmK1Number);
+            const index_t GemmKPad = GemmK - GemmKRaw;
+            const index_t GemmK0   = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmkraw_gemmmraw_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk_gemmm_grid_desc = transform_tensor_descriptor(
+                in_gemmkraw_gemmmraw_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKRaw, GemmKPad),
+                           make_right_pad_transform(GemmMRaw, GemmMPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: weight tensor
+            const auto wei_k_yxc_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+                wei_k_yxc_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_right_pad_transform(GemmKRaw, GemmKPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmk_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_nhowo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmmraw_gemmn_grid_desc =
+                transform_tensor_descriptor(out_nhowo_k_grid_desc,
+                                            make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                                       make_pass_through_transform(K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const index_t GemmK = Y * X * C;
+            assert(GemmK % GemmK1Number == 0);
+
+            const index_t GemmK0 = GemmK / GemmK1Number;
+
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmk_gemmmraw_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk_gemmmraw_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmMRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_k_yxc_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+                wei_k_yxc_grid_desc,
+                make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmk_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_nhowo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmmraw_gemmn_grid_desc =
+                transform_tensor_descriptor(out_nhowo_k_grid_desc,
+                                            make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                                       make_pass_through_transform(K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc);
+        }
+    }
+
+    using ABCGridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}));
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    using Block2CTileMap = BlockToCTileMap_M00_N0_M01<MPerBlock, NPerBlock, CGridDesc_M_N>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType, // TODO: Add ShuffleType for DeviceConv2d
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock * K1,
+        K1, // AK1
+        K1, // BK1
+        MPerXdl,
+        NPerXdl,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
+        2,                 // ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        Sequence<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // BBlockTransferSrcAccessOrder,
+        2,                 // BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+        CBlockTransferScalarPerVector_NWaveNPerXdl>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 OutDataType* p_out_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_in_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_out_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
+              block_2_ctile_map_{},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              input_spatial_lengths_{input_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              output_spatial_lengths_{output_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            const auto descs =
+                DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(N,
+                                                                          K,
+                                                                          C,
+                                                                          input_spatial_lengths,
+                                                                          filter_spatial_lengths,
+                                                                          output_spatial_lengths,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads);
+
+            a_grid_desc_k0_m_k1_ = descs[I0];
+            b_grid_desc_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_     = descs[I2];
+
+            block_2_ctile_map_ = Block2CTileMap{c_grid_desc_m_n_};
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
+                                           b_grid_desc_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
+                    GridwiseGemm::
+                        MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                            c_grid_desc_m_n_);
+            }
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        typename GridwiseGemm::
+            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
+        Block2CTileMap block_2_ctile_map_;
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+        std::vector<index_t> input_spatial_lengths_;
+        std::vector<index_t> filter_spatial_lengths_;
+        std::vector<index_t> output_spatial_lengths_;
+        std::vector<index_t> conv_filter_strides_;
+        std::vector<index_t> conv_filter_dilations_;
+        std::vector<index_t> input_left_pads_;
+        std::vector<index_t> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if DEBUG_LOG
+            {
+                std::cout << DeviceOp{}.GetTypeString() << std::endl;
+                std::cout << "N " << arg.Conv_N_ << ", "
+                          << "K " << arg.Conv_K_ << ", "
+                          << "C " << arg.Conv_C_ << ", " << std::endl;
+                std::cout << "Y X " << arg.filter_spatial_lengths_[0] << ", "
+                          << arg.filter_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Hi Wi " << arg.input_spatial_lengths_[0] << ", "
+                          << arg.input_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Ho Wo " << arg.output_spatial_lengths_[0] << ", "
+                          << arg.output_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Strides " << arg.conv_filter_strides_[0] << ", "
+                          << arg.conv_filter_strides_[1] << ", " << std::endl;
+                std::cout << "Dilations " << arg.conv_filter_dilations_[0] << ", "
+                          << arg.conv_filter_dilations_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_left_pads_[0] << ", "
+                          << arg.input_left_pads_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_right_pads_[0] << ", "
+                          << arg.input_right_pads_[1] << ", " << std::endl;
+            }
+
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+
+                std::cout
+                    << "arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_"
+                       "nwavenperxdl_{ "
+                    << arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_
+                           .GetLength(I0)
+                    << ", "
+                    << arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_
+                           .GetLength(I1)
+                    << ", "
+                    << arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_
+                           .GetLength(I2)
+                    << ", "
+                    << arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_
+                           .GetLength(I3)
+                    << ", "
+                    << arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_
+                           .GetLength(I4)
+                    << ", "
+                    << arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_
+                           .GetLength(I5)
+                    << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_xdlops_v3r1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    Block2CTileMap,
+                    true>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    arg.p_a_grid_,
+                    arg.p_b_grid_,
+                    arg.p_c_grid_,
+                    arg.a_grid_desc_k0_m_k1_,
+                    arg.b_grid_desc_k0_n_k1_,
+                    arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.in_element_op_,
+                    arg.wei_element_op_,
+                    arg.out_element_op_,
+                    arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_v3r1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseGemm::
+                            CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    Block2CTileMap,
+                    false>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    arg.p_a_grid_,
+                    arg.p_b_grid_,
+                    arg.p_c_grid_,
+                    arg.a_grid_desc_k0_m_k1_,
+                    arg.b_grid_desc_k0_n_k1_,
+                    arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
+                    arg.in_element_op_,
+                    arg.wei_element_op_,
+                    arg.out_element_op_,
+                    arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.conv_filter_strides_[0] == 1 && arg.conv_filter_strides_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 2 &&
+             arg.Conv_C_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_K_ % CBlockTransferScalarPerVector_NWaveNPerXdl == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             OutDataType* p_out_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        const void* p_wei_grid,
+                        void* p_out_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<OutDataType*>(p_out_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization) << ", "
+            << K1 << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << ABlockTransferDstScalarPerVector_K1 << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferDstScalarPerVector_K1 << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle << ", "
+            << CBlockTransferScalarPerVector_NWaveNPerXdl
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000000..ef94120f4e
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,679 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+template <typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ConvolutionForwardSpecialization ConvForwardSpecialization,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector>
+struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
+    : public DeviceConvFwd<2,
+                           ck::tensor_layout::convolution::NHWC,
+                           ck::tensor_layout::convolution::KYXC,
+                           ck::tensor_layout::convolution::NHWK,
+                           InDataType,
+                           WeiDataType,
+                           OutDataType,
+                           InElementwiseOperation,
+                           WeiElementwiseOperation,
+                           OutElementwiseOperation>
+{
+    using DeviceOp = DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
+
+    using ADataType = InDataType;
+    using BDataType = WeiDataType;
+    using CDataType = OutDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr index_t NDimSpatial = 2;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t GemmMRaw = N * Ho * Wo;
+        const index_t GemmN    = K;
+        const index_t GemmK    = Y * X * C;
+
+        const auto GemmMPad = math::integer_least_multiple(GemmMRaw, MPerBlock) - GemmMRaw;
+
+        assert(GemmK % GemmK1Number == 0);
+
+        const index_t GemmK0 = GemmK / GemmK1Number;
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: input tensor
+            const auto in_gemmmraw_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, C));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmmraw_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_right_pad_transform(GemmMRaw, GemmMPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc);
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_ho_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_n_ho_wo_c_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_merge_transform(make_tuple(N, Ho, Wo))),
+                make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmn_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, C));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmn_gemmk_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_gemmmraw_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            // A: input tensor
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmk_gemmmraw_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmmraw_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk_gemmmraw_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmMRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            const auto in_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmk0_gemmmraw_gemmk1_grid_desc,
+                make_tuple(make_pass_through_transform(GemmK0),
+                           make_right_pad_transform(GemmMRaw, GemmMPad),
+                           make_pass_through_transform(GemmK1Number)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B: weight tensor
+            const auto wei_k_yxc_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            const auto wei_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+                wei_k_yxc_grid_desc,
+                make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_gemmk_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: output tensor
+            const auto out_nhowo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmmraw_gemmn_grid_desc =
+                transform_tensor_descriptor(out_nhowo_k_grid_desc,
+                                            make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                                       make_pass_through_transform(K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmm_gemmn_grid_desc =
+                transform_tensor_descriptor(out_gemmmraw_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmMRaw, GemmMPad),
+                                                       make_pass_through_transform(GemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              out_gemmm_gemmn_grid_desc);
+        }
+    }
+
+    using ABCGridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}));
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
+        2,                 // ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        Sequence<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // BBlockTransferSrcAccessOrder,
+        2,                 // BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        Sequence<2, 3, 0, 1, 7, 5, 4, 6>, // CThreadTransferSrcDstAccessOrder,
+        7,                                // CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 OutDataType* p_out_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads)
+            : p_a_grid_{p_in_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_out_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            const auto descs =
+                DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(N,
+                                                                          K,
+                                                                          C,
+                                                                          input_spatial_lengths,
+                                                                          filter_spatial_lengths,
+                                                                          output_spatial_lengths,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads);
+
+            a_grid_desc_k0_m_k1_ = descs[I0];
+            b_grid_desc_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_     = descs[I2];
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+        std::vector<index_t> filter_spatial_lengths_;
+        std::vector<index_t> conv_filter_strides_;
+        std::vector<index_t> input_left_pads_;
+        std::vector<index_t> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if DEBUG_LOG
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+#endif
+            if(!GridwiseGemm::CheckValidity(
+                   arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.c_grid_desc_m_n_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
+            }
+
+            const auto [gdx, gdy, gdz] = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel =
+                    kernel_gemm_xdlops_v2r3<GridwiseGemm,
+                                            ADataType, // TODO: distiguish A/B datatype
+                                            CDataType,
+                                            DeviceOp::AGridDesc_K0_M_K1,
+                                            DeviceOp::BGridDesc_K0_N_K1,
+                                            DeviceOp::CGridDesc_M_N,
+                                            true>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(gdx, gdy, gdz),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m_n_);
+            }
+            else
+            {
+                const auto kernel =
+                    kernel_gemm_xdlops_v2r3<GridwiseGemm,
+                                            ADataType, // TODO: distiguish A/B datatype
+                                            CDataType,
+                                            DeviceOp::AGridDesc_K0_M_K1,
+                                            DeviceOp::BGridDesc_K0_N_K1,
+                                            DeviceOp::CGridDesc_M_N,
+                                            false>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(gdx, gdy, gdz),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m_n_);
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.conv_filter_strides_[0] == 1 && arg.conv_filter_strides_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
+                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
+                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
+            {
+                return false;
+            }
+        }
+
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 2 &&
+             arg.Conv_C_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_K_ % CThreadTransferDstScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        return GridwiseGemm::CheckValidity(
+            arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.c_grid_desc_m_n_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             OutDataType* p_out_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        const void* p_wei_grid,
+                        void* p_out_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation,
+                        WeiElementwiseOperation,
+                        OutElementwiseOperation) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<OutDataType*>(p_out_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization) << ", "
+            << K1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << ABlockTransferDstScalarPerVector_K1 << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferDstScalarPerVector_K1
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp
new file mode 100644
index 0000000000..cd89f3232c
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp
@@ -0,0 +1,268 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef DEVICE_CONV3D_FWD_NAIVE_HPP
+#define DEVICE_CONV3D_FWD_NAIVE_HPP
+
+#include <iostream>
+#include <memory>
+#include <sstream>
+#include "conv_util.hpp"
+#include "device.hpp"
+#include "device_conv_fwd.hpp"
+#include "common_header.hpp"
+#include "naive_conv_fwd.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// specialization for #D conv: in[n, di, hi, wi, c] * wei[k, z, y, x, c] = out[n, do, ho, wo, k]
+template <typename InDataType,
+          typename WeiDataType, // WeiDataType must be the same as InDataType
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceConv3dFwdNaive_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K
+    : public DeviceConvFwd<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>
+
+{
+    using DeviceOp = DeviceConv3dFwdNaive_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K;
+
+    using ADataType = InDataType;
+    using BDataType = WeiDataType;
+    using CDataType = OutDataType;
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in,
+                 const WeiDataType* p_wei,
+                 OutDataType* p_out,
+                 const index_t N,
+                 const index_t K,
+                 const index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : params_{3,
+                      N,
+                      K,
+                      C,
+                      filter_spatial_lengths,
+                      input_spatial_lengths,
+                      conv_filter_strides,
+                      conv_filter_dilations,
+                      input_left_pads,
+                      input_right_pads},
+              out_spatial_lengths_{output_spatial_lengths},
+              p_in_{p_in},
+              p_wei_{p_wei},
+              p_out_{p_out},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+
+        {
+        }
+
+        //  private:
+        utils::conv::ConvParams params_;
+        std::vector<index_t> out_spatial_lengths_;
+
+        const InDataType* p_in_;
+        const WeiDataType* p_wei_;
+        OutDataType* p_out_;
+
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto naive_conv3d_fwd =
+                ref::naive_conv_fwd_ndhwc_kzyxc_ndhwk<InDataType,
+                                                      WeiDataType,
+                                                      OutDataType,
+                                                      AccDataType,
+                                                      InElementwiseOperation,
+                                                      WeiElementwiseOperation,
+                                                      OutElementwiseOperation>;
+
+            float ave_time = launch_and_time_kernel(stream_config,
+                                                    naive_conv3d_fwd,
+                                                    dim3(256),
+                                                    dim3(256),
+                                                    0,
+                                                    arg.p_in_,
+                                                    arg.p_wei_,
+                                                    arg.p_out_,
+                                                    arg.N_,
+                                                    arg.K_,
+                                                    arg.C_,
+                                                    arg.in_spatial_lengths_[0],
+                                                    arg.in_spatial_lengths_[1],
+                                                    arg.in_spatial_lengths_[2],
+                                                    arg.filter_spatial_lengths_[0],
+                                                    arg.filter_spatial_lengths_[1],
+                                                    arg.filter_spatial_lengths_[2],
+                                                    arg.out_spatial_lengths_[0],
+                                                    arg.out_spatial_lengths_[1],
+                                                    arg.out_spatial_lengths_[2],
+                                                    arg.conv_filter_strides_[0],
+                                                    arg.conv_filter_strides_[1],
+                                                    arg.conv_filter_strides_[2],
+                                                    arg.conv_filter_dilations_[0],
+                                                    arg.conv_filter_dilations_[1],
+                                                    arg.conv_filter_dilations_[2],
+                                                    arg.in_left_pads_[0],
+                                                    arg.in_left_pads_[1],
+                                                    arg.in_left_pads_[2]);
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        std::vector<index_t> out_spatial_lengths = arg.params_.GetOutputSpatialLengths();
+
+        bool out_lengths_are_consistent = out_spatial_lengths[0] == arg.out_spatial_lengths_[0] &&
+                                          out_spatial_lengths[1] == arg.out_spatial_lengths_[1] &&
+                                          out_spatial_lengths[2] == arg.out_spatial_lengths_[2];
+        return out_lengths_are_consistent;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in,
+                             const WeiDataType* p_wei,
+                             OutDataType* p_out,
+                             const index_t N,
+                             const index_t K,
+                             const index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in,
+                        p_wei,
+                        p_out,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        const void* p_wei,
+                        void* p_out,
+                        const index_t N,
+                        const index_t K,
+                        const index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in),
+                                          static_cast<const WeiDataType*>(p_wei),
+                                          static_cast<OutDataType*>(p_out),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv3dFwdNaive_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K<>";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp
new file mode 100644
index 0000000000..a8e586b20c
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp
@@ -0,0 +1,662 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef DEVICE_CONV3D_FWD_XDL_HPP
+#define DEVICE_CONV3D_FWD_XDL_HPP
+
+#include <iostream>
+#include <memory>
+#include <sstream>
+#include "device.hpp"
+#include "device_conv_fwd.hpp"
+#include "common_header.hpp"
+#include "tensor_layout.hpp"
+#include "convolution_forward_specialization.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp"
+#include "gridwise_gemm_xdlops_v2r3.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+/*
+ * \see \link impl/device_batched_gemm_xdl.hpp kernel_batched_gemm_xdlops_v2r3() \endlink.
+ */
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v2r3_for_conv3d(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const index_t num_batches,
+            const index_t a_batch_stride,
+            const index_t b_batch_stride,
+            const index_t c_batch_stride,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+            const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / num_batches);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset =
+        __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(a_batch_stride) * g_idx);
+    const long_index_t b_batch_offset =
+        __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(b_batch_stride) * g_idx);
+    const long_index_t c_batch_offset =
+        __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(c_batch_stride) * g_idx);
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_c_grid + c_batch_offset,
+                                                  p_shared,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  block_2_ctile_map);
+
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = num_batches;
+    ignore = a_batch_stride;
+    ignore = b_batch_stride;
+    ignore = c_batch_stride;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// specialization for #D conv: in[n, di, hi, wi, c] * wei[k, z, y, x, c] = out[n, do, ho, wo, k]
+template <typename InDataType,
+          typename WeiDataType, // WeiDataType must be the same as InDataType
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ConvolutionForwardSpecialization ConvForwardSpecialization,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector>
+struct DeviceConv3dFwdXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K
+    : public DeviceConvFwd<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>
+
+{
+    using DeviceOp = DeviceConv3dFwdXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K;
+
+    using ADataType = InDataType;
+    using BDataType = WeiDataType;
+    using CDataType = OutDataType;
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    /*
+     * \brief Split the number of batches, \p N, into N = B * N1, such that the memory
+     * space of input and output tensors stays with the value range of index_t, and each subbatch
+     * can be dealed with GridwiseGemm.
+     */
+    static index_t GetMaxAllowableSubBatchSize(const index_t N,
+                                               const index_t K,
+                                               const index_t C,
+                                               std::vector<ck::index_t> input_spatial_lengths,
+                                               std::vector<ck::index_t> output_spatial_lengths)
+    {
+        const index_t Di = input_spatial_lengths[0];
+        const index_t Hi = input_spatial_lengths[1];
+        const index_t Wi = input_spatial_lengths[2];
+
+        const index_t Do = output_spatial_lengths[0];
+        const index_t Ho = output_spatial_lengths[1];
+        const index_t Wo = output_spatial_lengths[2];
+
+        // N1 should satisfy that
+        //   1) N % N1 = 0;
+        //   2) N1 * (Do * Ho * Wo * K) < (2^31 - 1)
+        //   3) N1 * (Di * Hi * Wi * C) < (2^31 - 1)
+        //
+        // Do NOT confuse (B, N1) in this function with (B, N1) in gridewise GEMM.
+        auto N1 = N + 1;
+
+        const auto stride =
+            math::max(long_index_t(Do) * Ho * Wo * K, long_index_t(Di) * Hi * Wi * C);
+        const index_t max_stride = NumericLimits<index_t>::Max();
+
+        for(index_t n0 = 1; n0 <= N; ++n0)
+        {
+            index_t n1 = N / n0;
+            if(n0 * n1 == N && long_index_t(n1) * long_index_t(stride) < max_stride)
+            {
+                N1 = n1;
+                break;
+            }
+        }
+
+        const auto B = N / N1;
+        if(B * N1 != N)
+        {
+            throw std::runtime_error(__func__ +
+                                     std::string(": failed to find num_subbatches for conv3d.\n"));
+        }
+
+        return N1;
+    }
+
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(const index_t N,
+                                                    const index_t K,
+                                                    const index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads)
+    {
+        assert(input_spatial_lengths.size() > 2);
+        assert(filter_spatial_lengths.size() > 2);
+        assert(conv_filter_strides.size() > 2);
+        assert(conv_filter_dilations.size() > 2);
+        assert(input_left_pads.size() > 2);
+        assert(input_right_pads.size() > 2);
+
+        const index_t Di = input_spatial_lengths[0];
+        const index_t Hi = input_spatial_lengths[1];
+        const index_t Wi = input_spatial_lengths[2];
+        const index_t Z  = filter_spatial_lengths[0];
+        const index_t Y  = filter_spatial_lengths[1];
+        const index_t X  = filter_spatial_lengths[2];
+
+        const index_t Do = output_spatial_lengths[0];
+        const index_t Ho = output_spatial_lengths[1];
+        const index_t Wo = output_spatial_lengths[2];
+
+        static_assert(ConvForwardSpecialization == ConvolutionForwardSpecialization::Default,
+                      "Wrong! This specialization not implemented!");
+
+        const auto in_desc_n_di_hi_wi_c =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
+        const auto wei_desc_k_z_y_x_c =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Z, Y, X, C));
+        const auto out_desc_n_do_ho_wo_k =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Do, Ho, Wo, K));
+
+        const auto descs = transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk_pad(
+            in_desc_n_di_hi_wi_c,
+            wei_desc_k_z_y_x_c,
+            out_desc_n_do_ho_wo_k,
+            make_tuple(conv_filter_strides[0], conv_filter_strides[1], conv_filter_strides[2]),
+            make_tuple(
+                conv_filter_dilations[0], conv_filter_dilations[1], conv_filter_dilations[2]),
+            make_tuple(input_left_pads[0], input_left_pads[1], input_left_pads[2]),
+            make_tuple(input_right_pads[0], input_right_pads[1], input_right_pads[2]),
+            Number<K1>{});
+
+        return descs;
+    }
+
+    using ABCGridDescs = remove_cvref_t<decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        1, 1, 1, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}))>;
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
+        BlockSize,
+        InDataType,
+        AccDataType,
+        OutDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
+        2,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        Sequence<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder,
+        Sequence<1, 0, 2>, // ABlockTransferSrcAccessOrder,
+        2,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        Sequence<2, 3, 0, 1, 7, 5, 4, 6>,
+        7,
+        CThreadTransferDstScalarPerVector>;
+
+    using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
+    using Block2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in,
+                 const WeiDataType* p_wei,
+                 OutDataType* p_out,
+                 const index_t N,
+                 const index_t K,
+                 const index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 index_t M01,
+                 index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_in},
+              p_b_grid_{p_wei},
+              p_c_grid_{p_out},
+              M01_{M01},
+              N01_{N01},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+        {
+            const index_t subbatch_size =
+                GetMaxAllowableSubBatchSize(N, K, C, input_spatial_lengths, output_spatial_lengths);
+            num_subbatches_ = N / subbatch_size;
+
+            const auto descs =
+                MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(subbatch_size,
+                                                                K,
+                                                                C,
+                                                                input_spatial_lengths,
+                                                                filter_spatial_lengths,
+                                                                output_spatial_lengths,
+                                                                conv_filter_strides,
+                                                                conv_filter_dilations,
+                                                                input_left_pads,
+                                                                input_right_pads);
+
+            a_grid_desc_k0_m_k1_ = descs[I0];
+            b_grid_desc_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_     = descs[I2];
+
+            block_2_ctile_map_ =
+                GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
+
+            a_batch_stride_ = a_grid_desc_k0_m_k1_.GetElementSpaceSize();
+            b_batch_stride_ = 0;
+            c_batch_stride_ = c_grid_desc_m_n_.GetElementSpaceSize();
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
+                                           b_grid_desc_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const InDataType* p_a_grid_;
+        const WeiDataType* p_b_grid_;
+        OutDataType* p_c_grid_;
+        index_t num_subbatches_;
+        index_t a_batch_stride_;
+        index_t b_batch_stride_;
+        index_t c_batch_stride_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
+        Block2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if DEBUG_LOG
+            {
+                std::cout << "num_batches_of_GEMM = " << arg.num_subbatches_ << std::endl;
+                std::cout << "a_grid_desc_k0_m_k1{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "b_grid_desc_k0_n_k1{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "c_grid_desc_m_n{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) *
+                arg.num_subbatches_;
+
+            const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            float ave_time = 0;
+            if(has_main_k0_block_loop)
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3_for_conv3d<
+                    GridwiseGemm,
+                    InDataType,
+                    OutDataType,
+                    remove_reference_t<AGridDesc_K0_M_K1>,
+                    remove_reference_t<BGridDesc_K0_N_K1>,
+                    remove_reference_t<CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<Block2CTileMap>,
+                    true>;
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.num_subbatches_,
+                                                  arg.a_batch_stride_,
+                                                  arg.b_batch_stride_,
+                                                  arg.c_batch_stride_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.in_element_op_,
+                                                  arg.wei_element_op_,
+                                                  arg.out_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3_for_conv3d<
+                    GridwiseGemm,
+                    InDataType,
+                    OutDataType,
+                    remove_reference_t<AGridDesc_K0_M_K1>,
+                    remove_reference_t<BGridDesc_K0_N_K1>,
+                    remove_reference_t<CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    OutElementwiseOperation,
+                    remove_reference_t<Block2CTileMap>,
+                    false>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.num_subbatches_,
+                                                  arg.a_batch_stride_,
+                                                  arg.b_batch_stride_,
+                                                  arg.c_batch_stride_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.in_element_op_,
+                                                  arg.wei_element_op_,
+                                                  arg.out_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const InDataType* p_in,
+                             const WeiDataType* p_wei,
+                             OutDataType* p_out,
+                             const index_t N,
+                             const index_t K,
+                             const index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in,
+                        p_wei,
+                        p_out,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        const void* p_wei,
+                        void* p_out,
+                        const index_t N,
+                        const index_t K,
+                        const index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in),
+                                          static_cast<const WeiDataType*>(p_wei),
+                                          static_cast<OutDataType*>(p_out),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConv3dFwdXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << ABlockTransferDstScalarPerVector_K1 << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferDstScalarPerVector_K1 << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle << ", "
+            << CBlockTransferScalarPerVector_NWaveNPerXdl
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp
new file mode 100644
index 0000000000..3178f73f4b
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp
@@ -0,0 +1,1588 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ConvolutionBackwardDataSpecialization ConvBackwardDataSpecialization,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          index_t M1PerThread,
+          index_t N1PerThread,
+          index_t KPerThread,
+          typename M1N1ThreadClusterM1Xs,
+          typename M1N1ThreadClusterN1Xs,
+          typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+          typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+          typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+          typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+          typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector>
+struct DeviceConvNdBwdDataNwcKxcNwk_Dl
+    : public DeviceConvBwdData<
+          NDimSpatial,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::NWC,
+                                        ck::tensor_layout::convolution::NHWC,
+                                        ck::tensor_layout::convolution::NDHWC>>,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::KXC,
+                                        ck::tensor_layout::convolution::KYXC,
+                                        ck::tensor_layout::convolution::KZYXC>>,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::NWK,
+                                        ck::tensor_layout::convolution::NHWK,
+                                        ck::tensor_layout::convolution::NDHWK>>,
+          InDataType,
+          WeiDataType,
+          OutDataType,
+          InElementwiseOperation,
+          WeiElementwiseOperation,
+          OutElementwiseOperation>
+{
+    using DeviceOp = DeviceConvNdBwdDataNwcKxcNwk_Dl;
+
+    using ADataType = OutDataType;
+    using BDataType = WeiDataType;
+    using CDataType = InDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads,
+                                                    std::vector<ck::index_t> tildes)
+    {
+        using namespace ck;
+
+        index_t i_xtilde = tildes[0];
+
+        const index_t Wi            = input_spatial_lengths[0];
+        const index_t Wo            = output_spatial_lengths[0];
+        const index_t X             = filter_spatial_lengths[0];
+        const index_t InLeftPadW    = input_left_pads[0];
+        const index_t InRightPadW   = input_right_pads[0];
+        const index_t ConvStrideW   = conv_filter_strides[0];
+        const index_t ConvDilationW = conv_filter_dilations[0];
+
+        const auto K0 = K / K1;
+
+        const auto in_n_wi_c_grid_desc = make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                make_naive_tensor_descriptor_packed(make_tuple(N * Wo, K)),
+                make_tuple(make_pass_through_transform(N * Wo),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: input tensor
+            const auto in_n_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_x_wo_c_grid_desc,
+                make_tuple(make_freeze_transform(I0),
+                           make_merge_transform(make_tuple(N, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto out_n_wo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Wo, K));
+            const auto wei_k_x_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, X, C));
+
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // A: output tensor
+            const auto out_n_wop_k_grid_desc = transform_tensor_descriptor(
+                out_n_wo_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wo, I0, I0),
+                           make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto out_n_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
+                out_n_wop_k_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(XDot, WTilde),
+                                         make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                    make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto out_n_xdotslice_wtildeslice_k0_k1_grid_desc = transform_tensor_descriptor(
+                out_n_xdot_wtilde_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_slice_transform(XDot, I0, XDotSlice),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4>{}));
+
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_n_xdotslice_wtildeslice_k0_k1_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(XDotSlice, K0)),
+                           make_merge_transform(make_tuple(N, WTildeSlice)),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<1, 3>{}, Sequence<0, 2>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B weight tensor
+            const auto wei_k_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
+                wei_k_x_c_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_embed_transform(make_tuple(XDot, XTilde),
+                                                make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto wei_k0_k1_xdotslice_c_grid_desc = transform_tensor_descriptor(
+                wei_k_xdot_xtilde_c_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                           make_slice_transform(XDot, I0, XDotSlice),
+                           make_freeze_transform(i_xtilde),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<>{}, Sequence<3>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_k0_k1_xdotslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(XDotSlice, K0)),
+                           make_pass_through_transform(C),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<2, 0>{}, Sequence<3>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // C: input tensor
+            const auto in_n_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_n_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
+                in_n_wip_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(XTilde, WTilde),
+                                                make_tuple(ConvDilationW, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_n_wtildeslice_c_grid_desc = transform_tensor_descriptor(
+                in_n_xtilde_wtilde_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_freeze_transform(i_xtilde),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_wtildeslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, WTildeSlice)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads,
+                                                    std::vector<ck::index_t> tildes)
+    {
+        using namespace ck;
+
+        index_t i_ytilde = tildes[0];
+        index_t i_xtilde = tildes[1];
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const auto K0 = K / K1;
+
+        const auto out_n_ho_wo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Ho, Wo, K));
+        const auto wei_k_y_x_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Y, X, C));
+        const auto in_n_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+                make_tuple(make_pass_through_transform(N * Ho * Wo),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: input tensor
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_freeze_transform(I0),
+                           make_freeze_transform(I0),
+                           make_merge_transform(make_tuple(N, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<1>{}, Sequence<3>{}, Sequence<0, 2, 4>{}, Sequence<5>{}),
+                make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto YDot = math::integer_divide_ceil(Y, YTilde);
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // A: output tensor
+            const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                out_n_ho_wo_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Ho, I0, I0),
+                           make_pad_transform(Wo, I0, I0),
+                           make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto out_n_ydot_htilde_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
+                out_n_hop_wop_k_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(YDot, HTilde),
+                                         make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                    make_embed_transform(make_tuple(XDot, WTilde),
+                                         make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                    make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc =
+                transform_tensor_descriptor(
+                    out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_slice_transform(YDot, I0, YDotSlice),
+                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                               make_slice_transform(XDot, I0, XDotSlice),
+                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                               make_unmerge_transform(make_tuple(K0, K1))),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5, 6>{}));
+
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                           make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B weight tensor
+            const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
+                wei_k_y_x_c_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_embed_transform(make_tuple(YDot, YTilde),
+                                                make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                           make_embed_transform(make_tuple(XDot, XTilde),
+                                                make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto wei_k0_k1_ydotslice_xdotslice_c_grid_desc =
+                transform_tensor_descriptor(wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_slice_transform(YDot, I0, YDotSlice),
+                                                       make_slice_transform(XDot, I0, XDotSlice),
+                                                       make_freeze_transform(i_ytilde),
+                                                       make_freeze_transform(i_xtilde),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{},
+                                                       Sequence<1>{},
+                                                       Sequence<3>{},
+                                                       Sequence<2>{},
+                                                       Sequence<4>{},
+                                                       Sequence<5>{}),
+                                            make_tuple(Sequence<0, 1>{},
+                                                       Sequence<2>{},
+                                                       Sequence<3>{},
+                                                       Sequence<>{},
+                                                       Sequence<>{},
+                                                       Sequence<4>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_k0_k1_ydotslice_xdotslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                           make_pass_through_transform(C),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<2, 3, 0>{}, Sequence<4>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // C: input tensor
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(YTilde, HTilde),
+                                                make_tuple(ConvDilationH, ConvStrideH)),
+                           make_embed_transform(make_tuple(XTilde, WTilde),
+                                                make_tuple(ConvDilationW, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
+                in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_freeze_transform(i_ytilde),
+                           make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                           make_freeze_transform(i_xtilde),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<>{},
+                           Sequence<1>{},
+                           Sequence<>{},
+                           Sequence<2>{},
+                           Sequence<3>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_htildeslice_wtildeslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads,
+                                                    std::vector<ck::index_t> tildes)
+    {
+        using namespace ck;
+
+        const index_t i_ztilde = tildes[0];
+        const index_t i_ytilde = tildes[1];
+        const index_t i_xtilde = tildes[2];
+
+        const index_t Di = input_spatial_lengths[0];
+        const index_t Hi = input_spatial_lengths[1];
+        const index_t Wi = input_spatial_lengths[2];
+
+        const index_t Do = output_spatial_lengths[0];
+        const index_t Ho = output_spatial_lengths[1];
+        const index_t Wo = output_spatial_lengths[2];
+
+        const index_t Z = filter_spatial_lengths[0];
+        const index_t Y = filter_spatial_lengths[1];
+        const index_t X = filter_spatial_lengths[2];
+
+        const index_t InLeftPadD = input_left_pads[0];
+        const index_t InLeftPadH = input_left_pads[1];
+        const index_t InLeftPadW = input_left_pads[2];
+
+        const index_t InRightPadD = input_right_pads[0];
+        const index_t InRightPadH = input_right_pads[1];
+        const index_t InRightPadW = input_right_pads[2];
+
+        const index_t ConvStrideD = conv_filter_strides[0];
+        const index_t ConvStrideH = conv_filter_strides[1];
+        const index_t ConvStrideW = conv_filter_strides[2];
+
+        const index_t ConvDilationD = conv_filter_dilations[0];
+        const index_t ConvDilationH = conv_filter_dilations[1];
+        const index_t ConvDilationW = conv_filter_dilations[2];
+
+        const auto K0 = K / K1;
+
+        const auto out_n_do_ho_wo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Do, Ho, Wo, K));
+        const auto wei_k_z_y_x_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Z, Y, X, C));
+        const auto in_n_di_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K)),
+                make_tuple(make_pass_through_transform(N * Do * Ho * Wo),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: input tensor
+            const auto in_n_z_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(I1, Do), make_tuple(I1, ConvStrideD)),
+                           make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_z_do_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_freeze_transform(I0),
+                           make_freeze_transform(I0),
+                           make_freeze_transform(I0),
+                           make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<1>{},
+                           Sequence<3>{},
+                           Sequence<5>{},
+                           Sequence<0, 2, 4, 6>{},
+                           Sequence<7>{}),
+                make_tuple(Sequence<>{}, Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto ZDot = math::integer_divide_ceil(Z, ZTilde);
+            const auto YDot = math::integer_divide_ceil(Y, YTilde);
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto DTilde =
+                Do + math::integer_divide_ceil(ConvDilationD * (Z - I1), ConvStrideD);
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IDTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadD - ConvDilationD * (ZTilde - I1)), ConvStrideD);
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IDTildeSliceEnd = math::min(
+                DTilde, math::integer_divide_ceil(InLeftPadD + Di - I1, ConvStrideD) + I1);
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto DTildeSlice = IDTildeSliceEnd - IDTildeSliceBegin;
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
+            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // A: output tensor
+            const auto out_n_dop_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                out_n_do_ho_wo_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Do, I0, I0),
+                           make_pad_transform(Ho, I0, I0),
+                           make_pad_transform(Wo, I0, I0),
+                           make_pass_through_transform(K)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_k_grid_desc =
+                transform_tensor_descriptor(
+                    out_n_dop_hop_wop_k_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(N),
+                        make_embed_transform(make_tuple(ZDot, DTilde),
+                                             make_tuple(-ConvDilationD / GcdStrideDilationD, I1)),
+                        make_embed_transform(make_tuple(YDot, HTilde),
+                                             make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                        make_embed_transform(make_tuple(XDot, WTilde),
+                                             make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                        make_pass_through_transform(K)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1, 2>{},
+                               Sequence<3, 4>{},
+                               Sequence<5, 6>{},
+                               Sequence<7>{}));
+
+            const auto
+                out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc =
+                    transform_tensor_descriptor(
+                        out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_k_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_slice_transform(ZDot, I0, ZDotSlice),
+                                   make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                                   make_slice_transform(YDot, I0, YDotSlice),
+                                   make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                                   make_slice_transform(XDot, I0, XDotSlice),
+                                   make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                                   make_unmerge_transform(make_tuple(K0, K1))),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<5>{},
+                                   Sequence<6>{},
+                                   Sequence<7>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<5>{},
+                                   Sequence<6>{},
+                                   Sequence<7, 8>{}));
+
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
+                make_tuple(
+                    make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, K0)),
+                    make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice)),
+                    make_pass_through_transform(K1)),
+                make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}, Sequence<8>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B weight tensor
+            const auto wei_k_zdot_ztilde_ydot_ytilde_xdot_xtilde_c_grid_desc =
+                transform_tensor_descriptor(
+                    wei_k_z_y_x_c_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(K),
+                        make_embed_transform(make_tuple(ZDot, ZTilde),
+                                             make_tuple(ConvStrideD / GcdStrideDilationD, I1)),
+                        make_embed_transform(make_tuple(YDot, YTilde),
+                                             make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                        make_embed_transform(make_tuple(XDot, XTilde),
+                                             make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                        make_pass_through_transform(C)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1, 2>{},
+                               Sequence<3, 4>{},
+                               Sequence<5, 6>{},
+                               Sequence<7>{}));
+
+            const auto wei_k0_k1_zdotslice_ydotslice_xdotslice_c_grid_desc =
+                transform_tensor_descriptor(wei_k_zdot_ztilde_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_slice_transform(ZDot, I0, ZDotSlice),
+                                                       make_slice_transform(YDot, I0, YDotSlice),
+                                                       make_slice_transform(XDot, I0, XDotSlice),
+                                                       make_freeze_transform(i_ztilde),
+                                                       make_freeze_transform(i_ytilde),
+                                                       make_freeze_transform(i_xtilde),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{},
+                                                       Sequence<1>{},
+                                                       Sequence<3>{},
+                                                       Sequence<5>{},
+                                                       Sequence<2>{},
+                                                       Sequence<4>{},
+                                                       Sequence<6>{},
+                                                       Sequence<7>{}),
+                                            make_tuple(Sequence<0, 1>{},
+                                                       Sequence<2>{},
+                                                       Sequence<3>{},
+                                                       Sequence<4>{},
+                                                       Sequence<>{},
+                                                       Sequence<>{},
+                                                       Sequence<>{},
+                                                       Sequence<5>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_k0_k1_zdotslice_ydotslice_xdotslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, K0)),
+                           make_pass_through_transform(C),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<2, 3, 4, 0>{}, Sequence<5>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // C: input tensor
+            const auto in_n_dip_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Di, InLeftPadD, InRightPadD),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc =
+                transform_tensor_descriptor(
+                    in_n_dip_hip_wip_c_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_embed_transform(make_tuple(ZTilde, DTilde),
+                                                    make_tuple(ConvDilationD, ConvStrideD)),
+                               make_embed_transform(make_tuple(YTilde, HTilde),
+                                                    make_tuple(ConvDilationH, ConvStrideH)),
+                               make_embed_transform(make_tuple(XTilde, WTilde),
+                                                    make_tuple(ConvDilationW, ConvStrideW)),
+                               make_pass_through_transform(C)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1, 2>{},
+                               Sequence<3, 4>{},
+                               Sequence<5, 6>{},
+                               Sequence<7>{}));
+
+            const auto in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc =
+                transform_tensor_descriptor(
+                    in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_freeze_transform(i_ztilde),
+                               make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                               make_freeze_transform(i_ytilde),
+                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                               make_freeze_transform(i_xtilde),
+                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{},
+                               Sequence<6>{},
+                               Sequence<7>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<>{},
+                               Sequence<1>{},
+                               Sequence<>{},
+                               Sequence<2>{},
+                               Sequence<>{},
+                               Sequence<3>{},
+                               Sequence<4>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc,
+                make_tuple(
+                    make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<1>(
+            1, 1, 1, {1}, {1}, {1}, {1}, {1}, {1}, {1}, {0});
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<2>(
+            1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {0, 0});
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<3>(1,
+                                                                  1,
+                                                                  1,
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {0, 0, 0});
+    }
+
+    using ABCGridDescs = decltype(GetABCGridDesc<NDimSpatial>());
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    // GridwiseGemm
+    using GridwiseGemm =
+        GridwiseGemmDl_km_kn_mn_v1r3<BlockSize,
+                                     ADataType,
+                                     AccDataType,
+                                     CDataType,
+                                     InMemoryDataOperationEnum::Set,
+                                     AGridDesc_K0_M_K1,
+                                     BGridDesc_K0_N_K1,
+                                     CGridDesc_M_N,
+                                     MPerBlock,
+                                     NPerBlock,
+                                     K0PerBlock,
+                                     K1,
+                                     M1PerThread,
+                                     N1PerThread,
+                                     KPerThread,
+                                     M1N1ThreadClusterM1Xs,
+                                     M1N1ThreadClusterN1Xs,
+                                     ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+                                     ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+                                     ABlockTransferThreadClusterArrangeOrder,
+                                     ABlockTransferSrcAccessOrder,
+                                     ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+                                     ABlockTransferSrcVectorTensorContiguousDimOrder,
+                                     ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+                                     BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+                                     BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+                                     BBlockTransferThreadClusterArrangeOrder,
+                                     BBlockTransferSrcAccessOrder,
+                                     BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+                                     BBlockTransferSrcVectorTensorContiguousDimOrder,
+                                     BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+                                     CThreadTransferSrcDstAccessOrder,
+                                     CThreadTransferSrcDstVectorDim,
+                                     CThreadTransferDstScalarPerVector>;
+
+    using AGridDesc_K0_M0_M1_K1 =
+        decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_K0_M_K1{}));
+    using BGridDesc_K0_N0_N1_K1 =
+        decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{}));
+    using CGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
+    using DefaultBlock2CTileMap =
+        decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(CGridDesc_M_N{}));
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 const OutDataType* p_out_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_out_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_in_grid},
+              a_element_op_{out_element_op},
+              b_element_op_{wei_element_op},
+              c_element_op_{in_element_op},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              input_spatial_lengths_{input_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              output_spatial_lengths_{output_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            CreateABCDesc<NDimSpatial>();
+        }
+
+        template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+        void CreateABCDesc()
+        {
+            const index_t ConvStrideW     = conv_filter_strides_[0];
+            const index_t ConvDilationW   = conv_filter_dilations_[0];
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+            const auto XTilde             = ConvStrideW / GcdStrideDilationW;
+
+            const index_t X = filter_spatial_lengths_[0];
+
+            for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+            {
+                // check slice is valid
+                const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+                if(XDotSlice <= 0)
+                {
+                    continue;
+                }
+
+                const auto descs =
+                    DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
+                        Conv_N_,
+                        Conv_K_,
+                        Conv_C_,
+                        input_spatial_lengths_,
+                        filter_spatial_lengths_,
+                        output_spatial_lengths_,
+                        conv_filter_strides_,
+                        conv_filter_dilations_,
+                        input_left_pads_,
+                        input_right_pads_,
+                        {i_xtilde});
+                a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
+                b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
+                c_grid_desc_m_n_container_.push_back(descs[I2]);
+
+                if(GridwiseGemm::CheckValidity(descs[I0], descs[I1], descs[I2]))
+                {
+                    a_grid_desc_k0_m0_m1_k1_container_.push_back(
+                        GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(descs[I0]));
+                    b_grid_desc_k0_n0_n1_k1_container_.push_back(
+                        GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(descs[I1]));
+                    c_grid_desc_m0_m10_m11_n0_n10_n11_container_.push_back(
+                        GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(descs[I2]));
+
+                    block_2_ctile_map_container_.push_back(
+                        GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2]));
+                }
+            }
+        }
+        template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+        void CreateABCDesc()
+        {
+            const index_t ConvStrideH = conv_filter_strides_[0];
+            const index_t ConvStrideW = conv_filter_strides_[1];
+
+            const index_t ConvDilationH = conv_filter_dilations_[0];
+            const index_t ConvDilationW = conv_filter_dilations_[1];
+
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const index_t Y = filter_spatial_lengths_[0];
+            const index_t X = filter_spatial_lengths_[1];
+            for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
+            {
+                for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+                {
+                    // check slice is valid
+                    const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+                    const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+                    if(YDotSlice * XDotSlice <= 0)
+                    {
+                        continue;
+                    }
+
+                    const auto descs =
+                        DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
+                            Conv_N_,
+                            Conv_K_,
+                            Conv_C_,
+                            input_spatial_lengths_,
+                            filter_spatial_lengths_,
+                            output_spatial_lengths_,
+                            conv_filter_strides_,
+                            conv_filter_dilations_,
+                            input_left_pads_,
+                            input_right_pads_,
+                            {i_ytilde, i_xtilde});
+                    a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
+                    b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
+                    c_grid_desc_m_n_container_.push_back(descs[I2]);
+
+                    if(GridwiseGemm::CheckValidity(descs[I0], descs[I1], descs[I2]))
+                    {
+                        a_grid_desc_k0_m0_m1_k1_container_.push_back(
+                            GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(descs[I0]));
+                        b_grid_desc_k0_n0_n1_k1_container_.push_back(
+                            GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(descs[I1]));
+                        c_grid_desc_m0_m10_m11_n0_n10_n11_container_.push_back(
+                            GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(descs[I2]));
+
+                        block_2_ctile_map_container_.push_back(
+                            GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2]));
+                    }
+                }
+            }
+        }
+        template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+        void CreateABCDesc()
+        {
+            const index_t ConvStrideD = conv_filter_strides_[0];
+            const index_t ConvStrideH = conv_filter_strides_[1];
+            const index_t ConvStrideW = conv_filter_strides_[2];
+
+            const index_t ConvDilationD = conv_filter_dilations_[0];
+            const index_t ConvDilationH = conv_filter_dilations_[1];
+            const index_t ConvDilationW = conv_filter_dilations_[2];
+
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const index_t Z = filter_spatial_lengths_[0];
+            const index_t Y = filter_spatial_lengths_[1];
+            const index_t X = filter_spatial_lengths_[2];
+            for(index_t i_ztilde = 0; i_ztilde < ZTilde; ++i_ztilde)
+            {
+                for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
+                {
+                    for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+                    {
+                        // check slice is valid
+                        const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
+                        const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+                        const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+                        if(ZDotSlice * YDotSlice * XDotSlice <= 0)
+                        {
+                            continue;
+                        }
+
+                        const auto descs =
+                            DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
+                                Conv_N_,
+                                Conv_K_,
+                                Conv_C_,
+                                input_spatial_lengths_,
+                                filter_spatial_lengths_,
+                                output_spatial_lengths_,
+                                conv_filter_strides_,
+                                conv_filter_dilations_,
+                                input_left_pads_,
+                                input_right_pads_,
+                                {i_ztilde, i_ytilde, i_xtilde});
+                        a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
+                        b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
+                        c_grid_desc_m_n_container_.push_back(descs[I2]);
+
+                        if(GridwiseGemm::CheckValidity(descs[I0], descs[I1], descs[I2]))
+                        {
+                            a_grid_desc_k0_m0_m1_k1_container_.push_back(
+                                GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(descs[I0]));
+                            b_grid_desc_k0_n0_n1_k1_container_.push_back(
+                                GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(descs[I1]));
+                            c_grid_desc_m0_m10_m11_n0_n10_n11_container_.push_back(
+                                GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(descs[I2]));
+
+                            block_2_ctile_map_container_.push_back(
+                                GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2]));
+                        }
+                    }
+                }
+            }
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        std::vector<AGridDesc_K0_M_K1> a_grid_desc_k0_m_k1_container_;
+        std::vector<BGridDesc_K0_N_K1> b_grid_desc_k0_n_k1_container_;
+        std::vector<CGridDesc_M_N> c_grid_desc_m_n_container_;
+
+        std::vector<AGridDesc_K0_M0_M1_K1> a_grid_desc_k0_m0_m1_k1_container_;
+        std::vector<BGridDesc_K0_N0_N1_K1> b_grid_desc_k0_n0_n1_k1_container_;
+        std::vector<CGridDesc_M0_M10_M11_N0_N10_N11> c_grid_desc_m0_m10_m11_n0_n10_n11_container_;
+
+        std::vector<DefaultBlock2CTileMap> block_2_ctile_map_container_;
+
+        // element-wise op
+        OutElementwiseOperation a_element_op_;
+        WeiElementwiseOperation b_element_op_;
+        InElementwiseOperation c_element_op_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+
+        std::vector<ck::index_t> input_spatial_lengths_;
+        std::vector<ck::index_t> filter_spatial_lengths_;
+        std::vector<ck::index_t> output_spatial_lengths_;
+        std::vector<ck::index_t> conv_filter_strides_;
+        std::vector<ck::index_t> conv_filter_dilations_;
+        std::vector<ck::index_t> input_left_pads_;
+        std::vector<ck::index_t> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            float ave_time = 0;
+            for(size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
+            {
+#if DEBUG_LOG
+                {
+                    std::cout << "arg.a_grid_desc_k0_m_k1_container_{"
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) << ", "
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I1) << ", "
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.b_grid_desc_k0_n_k1_container_{"
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I0) << ", "
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I1) << ", "
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I2) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.c_grid_desc_m_n_container_{ "
+                              << arg.c_grid_desc_m_n_container_[i].GetLength(I0) << ", "
+                              << arg.c_grid_desc_m_n_container_[i].GetLength(I1) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_( "
+                              << arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_[i].GetLength(I0)
+                              << ", "
+                              << arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_[i].GetLength(I1)
+                              << ", "
+                              << arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_[i].GetLength(I2)
+                              << ", "
+                              << arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_[i].GetLength(I3)
+                              << ", "
+                              << arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_[i].GetLength(I4)
+                              << ", "
+                              << arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_[i].GetLength(I5)
+                              << " ) " << std::endl;
+                }
+#endif
+
+                if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
+                                                arg.b_grid_desc_k0_n_k1_container_[i],
+                                                arg.c_grid_desc_m_n_container_[i]))
+                {
+                    throw std::runtime_error(
+                        "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
+                }
+
+                const index_t grid_size = arg.block_2_ctile_map_container_[i].CalculateGridSize(
+                    arg.c_grid_desc_m_n_container_[i]);
+
+                auto launch_kernel = [&](auto has_main_k_block_loop,
+                                         auto has_double_tail_k_block_loop) {
+                    constexpr bool has_main_loop   = has_main_k_block_loop.value;
+                    constexpr bool has_double_loop = has_double_tail_k_block_loop;
+
+                    const auto kernel = kernel_gemm_dl_v1r3<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M0_M1_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N0_N1_K1>,
+                        remove_reference_t<DeviceOp::CGridDesc_M0_M10_M11_N0_N10_N11>,
+                        remove_reference_t<DeviceOp::DefaultBlock2CTileMap>,
+                        has_main_loop,
+                        has_double_loop>;
+
+                    ave_time +=
+                        launch_and_time_kernel(stream_config,
+                                               kernel,
+                                               dim3(grid_size),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.p_a_grid_,
+                                               arg.p_b_grid_,
+                                               arg.p_c_grid_,
+                                               arg.a_grid_desc_k0_m0_m1_k1_container_[i],
+                                               arg.b_grid_desc_k0_n0_n1_k1_container_[i],
+                                               arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_[i],
+                                               arg.block_2_ctile_map_container_[i]);
+                };
+
+                const auto K0 = arg.a_grid_desc_k0_m0_m1_k1_container_[i].GetLength(I0);
+                const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
+                const bool has_double_tail_k_block_loop =
+                    GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
+
+                if(has_main_k_block_loop && has_double_tail_k_block_loop)
+                {
+                    launch_kernel(integral_constant<bool, true>{}, integral_constant<bool, true>{});
+                }
+                else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
+                {
+                    launch_kernel(integral_constant<bool, true>{},
+                                  integral_constant<bool, false>{});
+                }
+                else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
+                {
+                    launch_kernel(integral_constant<bool, false>{},
+                                  integral_constant<bool, true>{});
+                }
+                else
+                {
+                    launch_kernel(integral_constant<bool, false>{},
+                                  integral_constant<bool, false>{});
+                }
+            }
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        // check device
+        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
+             ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
+             ck::get_device_name() == "gfx1102"))
+        {
+            return false;
+        }
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 pad = 0 conv
+            for(int i = 0; i < NDimSpatial; i++)
+            {
+                if(!(arg.filter_spatial_lengths_[i] == 1 && arg.conv_filter_strides_[i] == 1 &&
+                     arg.input_left_pads_[i] == 0 && arg.input_right_pads_[i] == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // matrix A
+        {
+            auto srcVectorLengths = ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1{};
+            if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1)
+            {
+                return false;
+            }
+            if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0)
+            {
+                return false;
+            }
+
+            const index_t K = arg.Conv_K_;
+
+            if(K % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0)
+            {
+                return false;
+            }
+        }
+
+        // matrix B
+        {
+            auto srcLoadLenghts   = BBlockTransferThreadSliceLengths_K0_N0_N1_K1{};
+            auto srcVectorLengths = BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1{};
+            if(srcVectorLengths[I0] != 1 || srcVectorLengths[I3] != 1)
+            {
+                return false;
+            }
+            if(srcLoadLenghts[I1] % srcVectorLengths[I1] != 0 ||
+               srcLoadLenghts[I2] % srcVectorLengths[I2] != 0)
+            {
+                return false;
+            }
+
+            const index_t C = arg.Conv_K_;
+
+            if(C % (srcVectorLengths[I1] * srcVectorLengths[I2]) != 0)
+            {
+                return false;
+            }
+        }
+        // vector store C matrix into global memory
+        if(!(arg.Conv_C_ % CThreadTransferDstScalarPerVector == 0))
+        {
+            std::cout << "Not surpport,because: arg.Conv_C_ % CThreadTransferDstScalarPerVector = "
+                      << arg.Conv_C_ % CThreadTransferDstScalarPerVector << std::endl;
+            return false;
+        }
+
+        // Gridwise GEMM size
+        for(std::size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
+                                            arg.b_grid_desc_k0_n_k1_container_[i],
+                                            arg.c_grid_desc_m_n_container_[i]))
+            {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             const OutDataType* p_out_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(void* p_in_grid,
+                        const void* p_wei_grid,
+                        const void* p_out_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<const OutDataType*>(p_out_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConvNdBwdDataNwcKxcNwk_Dl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1
+            << ">";
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0){
+
+            str<< " Filter1x1Stride1Pad0";
+        }
+
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp
new file mode 100644
index 0000000000..ee3f0cea1b
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp
@@ -0,0 +1,1475 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ConvolutionBackwardDataSpecialization ConvBackwardDataSpecialization,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXdl,
+          ck::index_t NPerXdl,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector>
+struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
+    : public DeviceConvBwdData<
+          NDimSpatial,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::NWC,
+                                        ck::tensor_layout::convolution::NHWC,
+                                        ck::tensor_layout::convolution::NDHWC>>,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::KXC,
+                                        ck::tensor_layout::convolution::KYXC,
+                                        ck::tensor_layout::convolution::KZYXC>>,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::NWK,
+                                        ck::tensor_layout::convolution::NHWK,
+                                        ck::tensor_layout::convolution::NDHWK>>,
+          InDataType,
+          WeiDataType,
+          OutDataType,
+          InElementwiseOperation,
+          WeiElementwiseOperation,
+          OutElementwiseOperation>
+{
+    using DeviceOp = DeviceConvNdBwdDataNwcKxcNwk_Xdl;
+
+    using ADataType = OutDataType;
+    using BDataType = WeiDataType;
+    using CDataType = InDataType;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    static_assert((K1 % ABlockTransferThreadClusterLengths_K0_M_K1{}[I2]) %
+                      ABlockTransferSrcScalarPerVector ==
+                  0);
+    static_assert((NPerBlock / BBlockTransferThreadClusterLengths_K0_N_K1{}[I1]) %
+                      BBlockTransferSrcScalarPerVector ==
+                  0);
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads,
+                                                    std::vector<ck::index_t> tildes)
+    {
+        using namespace ck;
+
+        index_t i_xtilde = tildes[0];
+
+        const index_t Wi            = input_spatial_lengths[0];
+        const index_t Wo            = output_spatial_lengths[0];
+        const index_t X             = filter_spatial_lengths[0];
+        const index_t InLeftPadW    = input_left_pads[0];
+        const index_t InRightPadW   = input_right_pads[0];
+        const index_t ConvStrideW   = conv_filter_strides[0];
+        const index_t ConvDilationW = conv_filter_dilations[0];
+
+        const auto K0 = K / K1;
+
+        const auto in_n_wi_c_grid_desc = make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                make_naive_tensor_descriptor_packed(make_tuple(N * Wo, K)),
+                make_tuple(make_pass_through_transform(N * Wo),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: input tensor
+            const auto in_n_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_x_wo_c_grid_desc,
+                make_tuple(make_freeze_transform(I0),
+                           make_merge_transform(make_tuple(N, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto out_n_wo_k_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Wo, K));
+            const auto wei_k_x_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, X, C));
+
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // A: output tensor
+            const auto out_n_wop_k_grid_desc = transform_tensor_descriptor(
+                out_n_wo_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wo, I0, I0),
+                           make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto out_n_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
+                out_n_wop_k_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(XDot, WTilde),
+                                         make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                    make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto out_n_xdotslice_wtildeslice_k0_k1_grid_desc = transform_tensor_descriptor(
+                out_n_xdot_wtilde_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_slice_transform(XDot, I0, XDotSlice),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4>{}));
+
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_n_xdotslice_wtildeslice_k0_k1_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(XDotSlice, K0)),
+                           make_merge_transform(make_tuple(N, WTildeSlice)),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<1, 3>{}, Sequence<0, 2>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B weight tensor
+            const auto wei_k_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
+                wei_k_x_c_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_embed_transform(make_tuple(XDot, XTilde),
+                                                make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto wei_k0_k1_xdotslice_c_grid_desc = transform_tensor_descriptor(
+                wei_k_xdot_xtilde_c_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                           make_slice_transform(XDot, I0, XDotSlice),
+                           make_freeze_transform(i_xtilde),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0, 1>{}, Sequence<2>{}, Sequence<>{}, Sequence<3>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_k0_k1_xdotslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(XDotSlice, K0)),
+                           make_pass_through_transform(C),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<2, 0>{}, Sequence<3>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // C: input tensor
+            const auto in_n_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_n_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
+                in_n_wip_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(XTilde, WTilde),
+                                                make_tuple(ConvDilationW, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_n_wtildeslice_c_grid_desc = transform_tensor_descriptor(
+                in_n_xtilde_wtilde_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_freeze_transform(i_xtilde),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_wtildeslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, WTildeSlice)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads,
+                                                    std::vector<ck::index_t> tildes)
+    {
+        using namespace ck;
+
+        index_t i_ytilde = tildes[0];
+        index_t i_xtilde = tildes[1];
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const auto K0 = K / K1;
+
+        const auto out_n_ho_wo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Ho, Wo, K));
+        const auto wei_k_y_x_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Y, X, C));
+        const auto in_n_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+                make_tuple(make_pass_through_transform(N * Ho * Wo),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: input tensor
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_freeze_transform(I0),
+                           make_freeze_transform(I0),
+                           make_merge_transform(make_tuple(N, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<1>{}, Sequence<3>{}, Sequence<0, 2, 4>{}, Sequence<5>{}),
+                make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto YDot = math::integer_divide_ceil(Y, YTilde);
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // A: output tensor
+            const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                out_n_ho_wo_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Ho, I0, I0),
+                           make_pad_transform(Wo, I0, I0),
+                           make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto out_n_ydot_htilde_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
+                out_n_hop_wop_k_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(YDot, HTilde),
+                                         make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                    make_embed_transform(make_tuple(XDot, WTilde),
+                                         make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                    make_pass_through_transform(K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc =
+                transform_tensor_descriptor(
+                    out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_slice_transform(YDot, I0, YDotSlice),
+                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                               make_slice_transform(XDot, I0, XDotSlice),
+                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                               make_unmerge_transform(make_tuple(K0, K1))),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5, 6>{}));
+
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                           make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B weight tensor
+            const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
+                wei_k_y_x_c_grid_desc,
+                make_tuple(make_pass_through_transform(K),
+                           make_embed_transform(make_tuple(YDot, YTilde),
+                                                make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                           make_embed_transform(make_tuple(XDot, XTilde),
+                                                make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto wei_k0_k1_ydotslice_xdotslice_c_grid_desc =
+                transform_tensor_descriptor(wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_slice_transform(YDot, I0, YDotSlice),
+                                                       make_slice_transform(XDot, I0, XDotSlice),
+                                                       make_freeze_transform(i_ytilde),
+                                                       make_freeze_transform(i_xtilde),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{},
+                                                       Sequence<1>{},
+                                                       Sequence<3>{},
+                                                       Sequence<2>{},
+                                                       Sequence<4>{},
+                                                       Sequence<5>{}),
+                                            make_tuple(Sequence<0, 1>{},
+                                                       Sequence<2>{},
+                                                       Sequence<3>{},
+                                                       Sequence<>{},
+                                                       Sequence<>{},
+                                                       Sequence<4>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_k0_k1_ydotslice_xdotslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                           make_pass_through_transform(C),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<2, 3, 0>{}, Sequence<4>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // C: input tensor
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(YTilde, HTilde),
+                                                make_tuple(ConvDilationH, ConvStrideH)),
+                           make_embed_transform(make_tuple(XTilde, WTilde),
+                                                make_tuple(ConvDilationW, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
+                in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_freeze_transform(i_ytilde),
+                           make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                           make_freeze_transform(i_xtilde),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<>{},
+                           Sequence<1>{},
+                           Sequence<>{},
+                           Sequence<2>{},
+                           Sequence<3>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_htildeslice_wtildeslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    static auto
+    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
+                                                    ck::index_t K,
+                                                    ck::index_t C,
+                                                    std::vector<ck::index_t> input_spatial_lengths,
+                                                    std::vector<ck::index_t> filter_spatial_lengths,
+                                                    std::vector<ck::index_t> output_spatial_lengths,
+                                                    std::vector<ck::index_t> conv_filter_strides,
+                                                    std::vector<ck::index_t> conv_filter_dilations,
+                                                    std::vector<ck::index_t> input_left_pads,
+                                                    std::vector<ck::index_t> input_right_pads,
+                                                    std::vector<ck::index_t> tildes)
+    {
+        using namespace ck;
+
+        const index_t i_ztilde = tildes[0];
+        const index_t i_ytilde = tildes[1];
+        const index_t i_xtilde = tildes[2];
+
+        const index_t Di = input_spatial_lengths[0];
+        const index_t Hi = input_spatial_lengths[1];
+        const index_t Wi = input_spatial_lengths[2];
+
+        const index_t Do = output_spatial_lengths[0];
+        const index_t Ho = output_spatial_lengths[1];
+        const index_t Wo = output_spatial_lengths[2];
+
+        const index_t Z = filter_spatial_lengths[0];
+        const index_t Y = filter_spatial_lengths[1];
+        const index_t X = filter_spatial_lengths[2];
+
+        const index_t InLeftPadD = input_left_pads[0];
+        const index_t InLeftPadH = input_left_pads[1];
+        const index_t InLeftPadW = input_left_pads[2];
+
+        const index_t InRightPadD = input_right_pads[0];
+        const index_t InRightPadH = input_right_pads[1];
+        const index_t InRightPadW = input_right_pads[2];
+
+        const index_t ConvStrideD = conv_filter_strides[0];
+        const index_t ConvStrideH = conv_filter_strides[1];
+        const index_t ConvStrideW = conv_filter_strides[2];
+
+        const index_t ConvDilationD = conv_filter_dilations[0];
+        const index_t ConvDilationH = conv_filter_dilations[1];
+        const index_t ConvDilationW = conv_filter_dilations[2];
+
+        const auto K0 = K / K1;
+
+        const auto out_n_do_ho_wo_k_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Do, Ho, Wo, K));
+        const auto wei_k_z_y_x_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, Z, Y, X, C));
+        const auto in_n_di_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K)),
+                make_tuple(make_pass_through_transform(N * Do * Ho * Wo),
+                           make_unmerge_transform(make_tuple(K0, K1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            // B: weight tensor
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // C: input tensor
+            const auto in_n_z_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(I1, Do), make_tuple(I1, ConvStrideD)),
+                           make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_z_do_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_freeze_transform(I0),
+                           make_freeze_transform(I0),
+                           make_freeze_transform(I0),
+                           make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<1>{},
+                           Sequence<3>{},
+                           Sequence<5>{},
+                           Sequence<0, 2, 4, 6>{},
+                           Sequence<7>{}),
+                make_tuple(Sequence<>{}, Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto ZDot = math::integer_divide_ceil(Z, ZTilde);
+            const auto YDot = math::integer_divide_ceil(Y, YTilde);
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto DTilde =
+                Do + math::integer_divide_ceil(ConvDilationD * (Z - I1), ConvStrideD);
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IDTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadD - ConvDilationD * (ZTilde - I1)), ConvStrideD);
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IDTildeSliceEnd = math::min(
+                DTilde, math::integer_divide_ceil(InLeftPadD + Di - I1, ConvStrideD) + I1);
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto DTildeSlice = IDTildeSliceEnd - IDTildeSliceBegin;
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
+            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // A: output tensor
+            const auto out_n_dop_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                out_n_do_ho_wo_k_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Do, I0, I0),
+                           make_pad_transform(Ho, I0, I0),
+                           make_pad_transform(Wo, I0, I0),
+                           make_pass_through_transform(K)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_k_grid_desc =
+                transform_tensor_descriptor(
+                    out_n_dop_hop_wop_k_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(N),
+                        make_embed_transform(make_tuple(ZDot, DTilde),
+                                             make_tuple(-ConvDilationD / GcdStrideDilationD, I1)),
+                        make_embed_transform(make_tuple(YDot, HTilde),
+                                             make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                        make_embed_transform(make_tuple(XDot, WTilde),
+                                             make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                        make_pass_through_transform(K)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1, 2>{},
+                               Sequence<3, 4>{},
+                               Sequence<5, 6>{},
+                               Sequence<7>{}));
+
+            const auto
+                out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc =
+                    transform_tensor_descriptor(
+                        out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_k_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_slice_transform(ZDot, I0, ZDotSlice),
+                                   make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                                   make_slice_transform(YDot, I0, YDotSlice),
+                                   make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                                   make_slice_transform(XDot, I0, XDotSlice),
+                                   make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                                   make_unmerge_transform(make_tuple(K0, K1))),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<5>{},
+                                   Sequence<6>{},
+                                   Sequence<7>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<5>{},
+                                   Sequence<6>{},
+                                   Sequence<7, 8>{}));
+
+            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
+                make_tuple(
+                    make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, K0)),
+                    make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice)),
+                    make_pass_through_transform(K1)),
+                make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}, Sequence<8>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // B weight tensor
+            const auto wei_k_zdot_ztilde_ydot_ytilde_xdot_xtilde_c_grid_desc =
+                transform_tensor_descriptor(
+                    wei_k_z_y_x_c_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(K),
+                        make_embed_transform(make_tuple(ZDot, ZTilde),
+                                             make_tuple(ConvStrideD / GcdStrideDilationD, I1)),
+                        make_embed_transform(make_tuple(YDot, YTilde),
+                                             make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                        make_embed_transform(make_tuple(XDot, XTilde),
+                                             make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                        make_pass_through_transform(C)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1, 2>{},
+                               Sequence<3, 4>{},
+                               Sequence<5, 6>{},
+                               Sequence<7>{}));
+
+            const auto wei_k0_k1_zdotslice_ydotslice_xdotslice_c_grid_desc =
+                transform_tensor_descriptor(wei_k_zdot_ztilde_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                       make_slice_transform(ZDot, I0, ZDotSlice),
+                                                       make_slice_transform(YDot, I0, YDotSlice),
+                                                       make_slice_transform(XDot, I0, XDotSlice),
+                                                       make_freeze_transform(i_ztilde),
+                                                       make_freeze_transform(i_ytilde),
+                                                       make_freeze_transform(i_xtilde),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{},
+                                                       Sequence<1>{},
+                                                       Sequence<3>{},
+                                                       Sequence<5>{},
+                                                       Sequence<2>{},
+                                                       Sequence<4>{},
+                                                       Sequence<6>{},
+                                                       Sequence<7>{}),
+                                            make_tuple(Sequence<0, 1>{},
+                                                       Sequence<2>{},
+                                                       Sequence<3>{},
+                                                       Sequence<4>{},
+                                                       Sequence<>{},
+                                                       Sequence<>{},
+                                                       Sequence<>{},
+                                                       Sequence<5>{}));
+
+            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                wei_k0_k1_zdotslice_ydotslice_xdotslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, K0)),
+                           make_pass_through_transform(C),
+                           make_pass_through_transform(K1)),
+                make_tuple(Sequence<2, 3, 4, 0>{}, Sequence<5>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // C: input tensor
+            const auto in_n_dip_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Di, InLeftPadD, InRightPadD),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc =
+                transform_tensor_descriptor(
+                    in_n_dip_hip_wip_c_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_embed_transform(make_tuple(ZTilde, DTilde),
+                                                    make_tuple(ConvDilationD, ConvStrideD)),
+                               make_embed_transform(make_tuple(YTilde, HTilde),
+                                                    make_tuple(ConvDilationH, ConvStrideH)),
+                               make_embed_transform(make_tuple(XTilde, WTilde),
+                                                    make_tuple(ConvDilationW, ConvStrideW)),
+                               make_pass_through_transform(C)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1, 2>{},
+                               Sequence<3, 4>{},
+                               Sequence<5, 6>{},
+                               Sequence<7>{}));
+
+            const auto in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc =
+                transform_tensor_descriptor(
+                    in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_freeze_transform(i_ztilde),
+                               make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                               make_freeze_transform(i_ytilde),
+                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                               make_freeze_transform(i_xtilde),
+                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{},
+                               Sequence<6>{},
+                               Sequence<7>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<>{},
+                               Sequence<1>{},
+                               Sequence<>{},
+                               Sequence<2>{},
+                               Sequence<>{},
+                               Sequence<3>{},
+                               Sequence<4>{}));
+
+            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc,
+                make_tuple(
+                    make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                              wei_gemmk0_gemmn_gemmk1_grid_desc,
+                              in_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<1>(
+            1, 1, 1, {1}, {1}, {1}, {1}, {1}, {1}, {1}, {0});
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<2>(
+            1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {0, 0});
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<3>(1,
+                                                                  1,
+                                                                  1,
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {0, 0, 0});
+    }
+
+    using ABCGridDescs = decltype(GetABCGridDesc<NDimSpatial>());
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        InElementwiseOperation,
+        WeiElementwiseOperation,
+        OutElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXdl,
+        NPerXdl,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        Sequence<2, 3, 0, 1, 7, 5, 4, 6>, // CThreadTransferSrcDstAccessOrder,
+        7,                                // CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(InDataType* p_in_grid,
+                 const WeiDataType* p_wei_grid,
+                 const OutDataType* p_out_grid,
+                 ck::index_t N,
+                 ck::index_t K,
+                 ck::index_t C,
+                 std::vector<ck::index_t> input_spatial_lengths,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> output_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads)
+            : p_a_grid_{p_out_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_in_grid},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              input_spatial_lengths_{input_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              output_spatial_lengths_{output_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            CreateABCDesc<NDimSpatial>();
+        }
+
+        template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+        void CreateABCDesc()
+        {
+            const index_t ConvStrideW     = conv_filter_strides_[0];
+            const index_t ConvDilationW   = conv_filter_dilations_[0];
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+            const auto XTilde             = ConvStrideW / GcdStrideDilationW;
+
+            const index_t X = filter_spatial_lengths_[0];
+
+            for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+            {
+                // check slice is valid
+                const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+                if(XDotSlice <= 0)
+                {
+                    continue;
+                }
+
+                const auto descs =
+                    DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
+                        Conv_N_,
+                        Conv_K_,
+                        Conv_C_,
+                        input_spatial_lengths_,
+                        filter_spatial_lengths_,
+                        output_spatial_lengths_,
+                        conv_filter_strides_,
+                        conv_filter_dilations_,
+                        input_left_pads_,
+                        input_right_pads_,
+                        {i_xtilde});
+                a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
+                b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
+                c_grid_desc_m_n_container_.push_back(descs[I2]);
+            }
+        }
+        template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+        void CreateABCDesc()
+        {
+            const index_t ConvStrideH = conv_filter_strides_[0];
+            const index_t ConvStrideW = conv_filter_strides_[1];
+
+            const index_t ConvDilationH = conv_filter_dilations_[0];
+            const index_t ConvDilationW = conv_filter_dilations_[1];
+
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const index_t Y = filter_spatial_lengths_[0];
+            const index_t X = filter_spatial_lengths_[1];
+            for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
+            {
+                for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+                {
+                    // check slice is valid
+                    const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+                    const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+                    if(YDotSlice * XDotSlice <= 0)
+                    {
+                        continue;
+                    }
+
+                    const auto descs =
+                        DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
+                            Conv_N_,
+                            Conv_K_,
+                            Conv_C_,
+                            input_spatial_lengths_,
+                            filter_spatial_lengths_,
+                            output_spatial_lengths_,
+                            conv_filter_strides_,
+                            conv_filter_dilations_,
+                            input_left_pads_,
+                            input_right_pads_,
+                            {i_ytilde, i_xtilde});
+                    a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
+                    b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
+                    c_grid_desc_m_n_container_.push_back(descs[I2]);
+                }
+            }
+        }
+        template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+        void CreateABCDesc()
+        {
+            const index_t ConvStrideD = conv_filter_strides_[0];
+            const index_t ConvStrideH = conv_filter_strides_[1];
+            const index_t ConvStrideW = conv_filter_strides_[2];
+
+            const index_t ConvDilationD = conv_filter_dilations_[0];
+            const index_t ConvDilationH = conv_filter_dilations_[1];
+            const index_t ConvDilationW = conv_filter_dilations_[2];
+
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const index_t Z = filter_spatial_lengths_[0];
+            const index_t Y = filter_spatial_lengths_[1];
+            const index_t X = filter_spatial_lengths_[2];
+            for(index_t i_ztilde = 0; i_ztilde < ZTilde; ++i_ztilde)
+            {
+                for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
+                {
+                    for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+                    {
+                        // check slice is valid
+                        const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
+                        const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+                        const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+                        if(ZDotSlice * YDotSlice * XDotSlice <= 0)
+                        {
+                            continue;
+                        }
+
+                        const auto descs =
+                            DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
+                                Conv_N_,
+                                Conv_K_,
+                                Conv_C_,
+                                input_spatial_lengths_,
+                                filter_spatial_lengths_,
+                                output_spatial_lengths_,
+                                conv_filter_strides_,
+                                conv_filter_dilations_,
+                                input_left_pads_,
+                                input_right_pads_,
+                                {i_ztilde, i_ytilde, i_xtilde});
+                        a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
+                        b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
+                        c_grid_desc_m_n_container_.push_back(descs[I2]);
+                    }
+                }
+            }
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        std::vector<AGridDesc_K0_M_K1> a_grid_desc_k0_m_k1_container_;
+        std::vector<BGridDesc_K0_N_K1> b_grid_desc_k0_n_k1_container_;
+        std::vector<CGridDesc_M_N> c_grid_desc_m_n_container_;
+        OutElementwiseOperation a_element_op_;
+        WeiElementwiseOperation b_element_op_;
+        InElementwiseOperation c_element_op_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+
+        std::vector<ck::index_t> input_spatial_lengths_;
+        std::vector<ck::index_t> filter_spatial_lengths_;
+        std::vector<ck::index_t> output_spatial_lengths_;
+        std::vector<ck::index_t> conv_filter_strides_;
+        std::vector<ck::index_t> conv_filter_dilations_;
+        std::vector<ck::index_t> input_left_pads_;
+        std::vector<ck::index_t> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            float ave_time = 0;
+            for(size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
+            {
+#if DEBUG_LOG
+                {
+                    std::cout << "arg.a_grid_desc_k0_m_k1{"
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) << ", "
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I1) << ", "
+                              << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.b_grid_desc_k0_n_k1{"
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I0) << ", "
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I1) << ", "
+                              << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I2) << "}"
+                              << std::endl;
+
+                    std::cout << "arg.c_grid_desc_m_n{"
+                              << arg.c_grid_desc_m_n_container_[i].GetLength(I0) << ", "
+                              << arg.c_grid_desc_m_n_container_[i].GetLength(I1) << "}"
+                              << std::endl;
+                }
+#endif
+
+                if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
+                                                arg.b_grid_desc_k0_n_k1_container_[i],
+                                                arg.c_grid_desc_m_n_container_[i]))
+                {
+                    throw std::runtime_error(
+                        "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
+                }
+
+                const auto [gdx, gdy, gdz] =
+                    GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_container_[i]);
+
+                const auto K = arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) *
+                               arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2);
+
+                if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+                {
+                    const auto kernel =
+                        kernel_gemm_xdlops_v2r3<GridwiseGemm,
+                                                ADataType, // TODO: distiguish A/B datatype
+                                                CDataType,
+                                                DeviceOp::AGridDesc_K0_M_K1,
+                                                DeviceOp::BGridDesc_K0_N_K1,
+                                                DeviceOp::CGridDesc_M_N,
+                                                true>;
+
+                    ave_time += launch_and_time_kernel(stream_config,
+                                                       kernel,
+                                                       dim3(gdx, gdy, gdz),
+                                                       dim3(BlockSize),
+                                                       0,
+                                                       arg.p_a_grid_,
+                                                       arg.p_b_grid_,
+                                                       arg.p_c_grid_,
+                                                       arg.a_grid_desc_k0_m_k1_container_[i],
+                                                       arg.b_grid_desc_k0_n_k1_container_[i],
+                                                       arg.c_grid_desc_m_n_container_[i]);
+                }
+                else
+                {
+                    const auto kernel =
+                        kernel_gemm_xdlops_v2r3<GridwiseGemm,
+                                                ADataType, // TODO: distiguish A/B datatype
+                                                CDataType,
+                                                DeviceOp::AGridDesc_K0_M_K1,
+                                                DeviceOp::BGridDesc_K0_N_K1,
+                                                DeviceOp::CGridDesc_M_N,
+                                                false>;
+
+                    ave_time += launch_and_time_kernel(stream_config,
+                                                       kernel,
+                                                       dim3(gdx, gdy, gdz),
+                                                       dim3(BlockSize),
+                                                       0,
+                                                       arg.p_a_grid_,
+                                                       arg.p_b_grid_,
+                                                       arg.p_c_grid_,
+                                                       arg.a_grid_desc_k0_m_k1_container_[i],
+                                                       arg.b_grid_desc_k0_n_k1_container_[i],
+                                                       arg.c_grid_desc_m_n_container_[i]);
+                }
+            }
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 pad = 0 conv
+            for(int i = 0; i < NDimSpatial; i++)
+            {
+                if(!(arg.filter_spatial_lengths_[i] == 1 && arg.conv_filter_strides_[i] == 1 &&
+                     arg.input_left_pads_[i] == 0 && arg.input_right_pads_[i] == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 1 &&
+             arg.Conv_K_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_C_ % CThreadTransferDstScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        for(std::size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
+                                            arg.b_grid_desc_k0_n_k1_container_[i],
+                                            arg.c_grid_desc_m_n_container_[i]))
+            {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(InDataType* p_in_grid,
+                             const WeiDataType* p_wei_grid,
+                             const OutDataType* p_out_grid,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t C,
+                             std::vector<ck::index_t> input_spatial_lengths,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> output_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(void* p_in_grid,
+                        const void* p_wei_grid,
+                        const void* p_out_grid,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t C,
+                        std::vector<ck::index_t> input_spatial_lengths,
+                        std::vector<ck::index_t> filter_spatial_lengths,
+                        std::vector<ck::index_t> output_spatial_lengths,
+                        std::vector<ck::index_t> conv_filter_strides,
+                        std::vector<ck::index_t> conv_filter_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        InElementwiseOperation,
+                        WeiElementwiseOperation,
+                        OutElementwiseOperation) override
+    {
+        return std::make_unique<Argument>(static_cast<InDataType*>(p_in_grid),
+                                          static_cast<const WeiDataType*>(p_wei_grid),
+                                          static_cast<const OutDataType*>(p_out_grid),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceConvNdBwdDataNwcKxcNwk_Xdl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << ABlockTransferDstScalarPerVector_K1 << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferDstScalarPerVector_K1
+            << ">";
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0){
+
+            str<< " Filter1x1Stride1Pad0";
+        }
+
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp
new file mode 100644
index 0000000000..02ef29e32d
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp
@@ -0,0 +1,338 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/math.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/stream_utility.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+template <typename InDataTypeTuple,
+          typename OutDataTypeTuple,
+          typename ElementwiseOperation,
+          index_t NumDim_m,
+          index_t NumDim_n,
+          index_t MPerThread,
+          index_t NPerThread,
+          typename InScalarPerVectorSeq,
+          typename OutScalarPerVectorSeq>
+struct DeviceElementwise2dImpl : public DeviceElementwise<InDataTypeTuple,
+                                                          OutDataTypeTuple,
+                                                          ElementwiseOperation,
+                                                          NumDim_m + NumDim_n>
+{
+    static constexpr index_t NumDim = NumDim_m + NumDim_n;
+
+    static constexpr int NumInput  = InDataTypeTuple::Size();
+    static constexpr int NumOutput = OutDataTypeTuple::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static_assert(NumInput == InScalarPerVectorSeq::Size() &&
+                      NumOutput == OutScalarPerVectorSeq::Size(),
+                  "Tuple size is inconsistent with the number of in/out!");
+
+    static auto GenerateInDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+
+                return static_cast<const DataType*>(nullptr);
+            },
+            Number<NumInput>{});
+    };
+
+    static auto GenerateOutDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+
+                return static_cast<DataType*>(nullptr);
+            },
+            Number<NumOutput>{});
+    };
+
+    using InDataTypePointerTuple  = decltype(GenerateInDataTypePointerTuple());
+    using OutDataTypePointerTuple = decltype(GenerateOutDataTypePointerTuple());
+
+    template <typename Desc_MN>
+    static auto PadDescriptor_MN_2d(Desc_MN desc_mn,
+                                    index_t gridSize,
+                                    index_t blockSize,
+                                    index_t num_threads_m,
+                                    index_t num_threads_n)
+    {
+        std::ignore               = blockSize;
+        std::ignore               = gridSize;
+        const auto m              = desc_mn.GetLength(I0);
+        const auto n              = desc_mn.GetLength(I1);
+        const index_t loop_step_m = num_threads_m * MPerThread;
+        const index_t loop_step_n = num_threads_n * NPerThread;
+        const auto pad_m          = math::integer_least_multiple(m, loop_step_m) - m;
+        const auto pad_n          = math::integer_least_multiple(n, loop_step_n) - n;
+
+        const auto desc_mn_pad = transform_tensor_descriptor(
+            desc_mn,
+            make_tuple(make_right_pad_transform(m, pad_m), make_right_pad_transform(n, pad_n)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+        return desc_mn_pad;
+    }
+
+    static auto MakeDescriptor_MN(const std::array<index_t, NumDim>& lengths,
+                                  const std::array<index_t, NumDim>& stride,
+                                  index_t gridSize,
+                                  index_t blockSize,
+                                  index_t num_threads_m,
+                                  index_t num_threads_n)
+    {
+        auto tupleOfShape  = generate_tuple([&](auto I) { return lengths[I]; }, Number<NumDim>{});
+        auto tupleOfStride = generate_tuple([&](auto I) { return stride[I]; }, Number<NumDim>{});
+
+        // nd desc - [s0, s1, s2, ...]
+        const auto desc = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
+
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDim_m, 1>::type();
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDim_m, NumDim_m + NumDim_n, 1>::type();
+
+        const auto mLengths = get_container_subset(tupleOfShape, mDimIds);
+        const auto nLengths = get_container_subset(tupleOfShape, nDimIds);
+
+        // merge nd to 2d desc - [s0 * s1 * ...]
+
+        if constexpr(NumDim > 2)
+        {
+            const auto desc_mn = transform_tensor_descriptor(
+                desc,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+                make_tuple(mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return PadDescriptor_MN_2d(desc_mn, gridSize, blockSize, num_threads_m, num_threads_n);
+        }
+        else
+            return PadDescriptor_MN_2d(desc, gridSize, blockSize, num_threads_m, num_threads_n);
+    }
+
+    template <index_t TupleSize>
+    static auto GenerateInOutGrid2dDescTuple(Number<TupleSize>)
+    {
+        return generate_tuple(
+            [&](auto) {
+                if constexpr(NumDim > 2)
+                {
+                    return MakeDescriptor_MN({1, 1}, {1, 1}, 1, 1, 1, 1);
+                }
+                else
+                {
+                    return MakeDescriptor_MN({1}, {1}, 1, 1, 1, 1);
+                };
+            },
+            Number<TupleSize>{});
+    };
+
+    using OutGrid2dDescTuple = decltype(GenerateInOutGrid2dDescTuple(Number<NumOutput>{}));
+    using InGrid2dDescTuple  = decltype(GenerateInOutGrid2dDescTuple(Number<NumInput>{}));
+
+    using GridwiseElementwise = GridwiseElementwise_2D<InGrid2dDescTuple,
+                                                       OutGrid2dDescTuple,
+                                                       InDataTypePointerTuple,
+                                                       OutDataTypePointerTuple,
+                                                       ElementwiseOperation,
+                                                       MPerThread,
+                                                       NPerThread,
+                                                       InScalarPerVectorSeq,
+                                                       OutScalarPerVectorSeq>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, NumDim> lengths,
+                 const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
+                 const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
+                 const std::array<const void*, NumInput> in_dev_buffers,
+                 const std::array<void*, NumOutput> out_dev_buffers,
+                 ElementwiseOperation elementwise_op)
+
+            : lengths_(lengths),
+              inStridesArray_(inStridesArray),
+              outStridesArray_(outStridesArray),
+              elementwise_op_(elementwise_op),
+              blockSize_(256)
+        {
+            static_assert(NumDim_m > 0, "");
+            static_assert(NumDim_n > 0, "");
+
+            in_dev_buffers_ = generate_tuple(
+                [&](auto I) {
+                    using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+                    return static_cast<const DataType*>(in_dev_buffers[I.value]);
+                },
+                Number<NumInput>{});
+
+            out_dev_buffers_ = generate_tuple(
+                [&](auto I) {
+                    using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+                    return static_cast<DataType*>(out_dev_buffers[I.value]);
+                },
+                Number<NumOutput>{});
+        }
+
+        InDataTypePointerTuple in_dev_buffers_;
+        OutDataTypePointerTuple out_dev_buffers_;
+
+        std::array<index_t, NumDim> lengths_;
+        std::array<std::array<index_t, NumDim>, NumInput> inStridesArray_;
+        std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray_;
+
+        ElementwiseOperation elementwise_op_;
+        index_t blockSize_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            index_t gridSize      = getAvailableComputeUnitCount(stream_config);
+            index_t num_threads_m = (gridSize * arg.blockSize_) / 16;
+            index_t num_threads_n = 16;
+
+            auto in_grid_2d_desc_tuple = generate_tuple(
+                [&](auto I) {
+                    return MakeDescriptor_MN(arg.lengths_,
+                                             arg.inStridesArray_[I.value],
+                                             gridSize,
+                                             arg.blockSize_,
+                                             num_threads_m,
+                                             num_threads_n);
+                },
+                Number<NumInput>{});
+
+            auto out_grid_2d_desc_tuple = generate_tuple(
+                [&](auto I) {
+                    return MakeDescriptor_MN(arg.lengths_,
+                                             arg.outStridesArray_[I.value],
+                                             gridSize,
+                                             arg.blockSize_,
+                                             num_threads_m,
+                                             num_threads_n);
+                },
+                Number<NumOutput>{});
+
+            const auto kernel = kernel_elementwise_2d<GridwiseElementwise,
+                                                      InGrid2dDescTuple,
+                                                      OutGrid2dDescTuple,
+                                                      InDataTypePointerTuple,
+                                                      OutDataTypePointerTuple,
+                                                      ElementwiseOperation>;
+
+            float elapsed_time = launch_and_time_kernel(stream_config,
+                                                        kernel,
+                                                        dim3(gridSize),
+                                                        dim3(arg.blockSize_),
+                                                        0,
+                                                        in_grid_2d_desc_tuple,
+                                                        out_grid_2d_desc_tuple,
+                                                        arg.in_dev_buffers_,
+                                                        arg.out_dev_buffers_,
+                                                        arg.elementwise_op_,
+                                                        num_threads_m,
+                                                        num_threads_n);
+            return elapsed_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        if(pArg == nullptr)
+            return false;
+
+        if(pArg->lengths_.back() % MPerThread != 0)
+            return false;
+
+        auto IsScalarPerVectorValid = [&](const std::array<index_t, NumDim>& lengths,
+                                          const std::array<index_t, NumDim>& strides,
+                                          index_t scalarPerVector,
+                                          index_t vectorDim) {
+            if(strides[vectorDim] == 1 &&
+               (lengths[vectorDim] % scalarPerVector == 0 ||
+                lengths[vectorDim] % scalarPerVector == lengths[vectorDim]))
+            {
+                return true;
+            }
+            if(strides[vectorDim] != 1 && scalarPerVector == strides[vectorDim])
+            {
+                return true;
+            }
+            return false;
+        };
+
+        bool valid = true;
+        static_for<0, NumInput, 1>{}([&](auto I) {
+            if(!IsScalarPerVectorValid(pArg->lengths_,
+                                       pArg->inStridesArray_[I.value],
+                                       InScalarPerVectorSeq::At(I),
+                                       NumDim_m - 1))
+                valid = false;
+        });
+
+        static_for<0, NumOutput, 1>{}([&](auto I) {
+            if(!IsScalarPerVectorValid(pArg->lengths_,
+                                       pArg->outStridesArray_[I.value],
+                                       OutScalarPerVectorSeq::At(I),
+                                       NumDim - 1))
+                valid = false;
+        });
+
+        return valid;
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, NumDim> lengths,
+                        const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
+                        const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
+                        const std::array<const void*, NumInput> in_dev_buffers,
+                        const std::array<void*, NumOutput> out_dev_buffers,
+                        ElementwiseOperation elementwise_op) override
+    {
+        return std::make_unique<Argument>(lengths,
+                                          inStridesArray,
+                                          outStridesArray,
+                                          in_dev_buffers,
+                                          out_dev_buffers,
+                                          elementwise_op);
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+}; // namespace device
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp
new file mode 100644
index 0000000000..e203468c61
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp
@@ -0,0 +1,303 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/math.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/stream_utility.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataTypeTuple,
+          typename OutDataTypeTuple,
+          typename ElementwiseOperation,
+          index_t NumDim,
+          index_t MPerThread,
+          typename InScalarPerVectorSeq,
+          typename OutScalarPerVectorSeq>
+struct DeviceElementwiseImpl
+    : public DeviceElementwise<InDataTypeTuple, OutDataTypeTuple, ElementwiseOperation, NumDim>
+{
+    static constexpr int NumInput  = InDataTypeTuple::Size();
+    static constexpr int NumOutput = OutDataTypeTuple::Size();
+
+    static_assert(NumInput == InScalarPerVectorSeq::Size() &&
+                      NumOutput == OutScalarPerVectorSeq::Size(),
+                  "Tuple size is inconsistent with the number of in/out!");
+
+    static auto GenerateInDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+
+                return static_cast<const DataType*>(nullptr);
+            },
+            Number<NumInput>{});
+    };
+
+    static auto GenerateOutDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+
+                return static_cast<DataType*>(nullptr);
+            },
+            Number<NumOutput>{});
+    };
+
+    using InDataTypePointerTuple  = decltype(GenerateInDataTypePointerTuple());
+    using OutDataTypePointerTuple = decltype(GenerateOutDataTypePointerTuple());
+
+    template <typename Desc_M>
+    static auto PadDescriptor_M_1d(Desc_M desc_m, index_t gridSize, index_t blockSize)
+    {
+        constexpr auto I0 = Number<0>{};
+
+        const auto m            = desc_m.GetLength(I0);
+        const index_t loop_step = gridSize * blockSize * MPerThread;
+        const auto pad          = math::integer_least_multiple(m, loop_step) - m;
+        const auto desc_m_pad =
+            transform_tensor_descriptor(desc_m,
+                                        make_tuple(make_right_pad_transform(m, pad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return desc_m_pad;
+    }
+
+    static auto MakeDescriptor_M(const std::array<index_t, NumDim>& lengths,
+                                 const std::array<index_t, NumDim>& stride,
+                                 index_t gridSize,
+                                 index_t blockSize)
+    {
+        auto tupleOfShape  = generate_tuple([&](auto I) { return lengths[I]; }, Number<NumDim>{});
+        auto tupleOfStride = generate_tuple([&](auto I) { return stride[I]; }, Number<NumDim>{});
+
+        // nd desc - [s0, s1, s2, ...]
+        const auto desc = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
+
+        // merge nd to 1d desc - [s0 * s1 * ...]
+        if constexpr(NumDim > 1)
+        {
+            const auto desc_m = transform_tensor_descriptor(
+                desc,
+                make_tuple(make_merge_transform(tupleOfShape)),
+                make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<NumDim>{})),
+                make_tuple(Sequence<0>{}));
+
+            return PadDescriptor_M_1d(desc_m, gridSize, blockSize);
+        }
+        else
+            return PadDescriptor_M_1d(desc, gridSize, blockSize);
+    }
+
+    template <index_t TupleSize>
+    static auto GenerateInOutGrid1dDescTuple(Number<TupleSize>)
+    {
+        return generate_tuple(
+            [&](auto) {
+                if constexpr(NumDim > 1)
+                {
+                    return MakeDescriptor_M({1, 1}, {1, 1}, 1, 1);
+                }
+                else
+                {
+                    return MakeDescriptor_M({1}, {1}, 1, 1);
+                };
+            },
+            Number<TupleSize>{});
+    };
+
+    using InGrid1dDescTuple  = decltype(GenerateInOutGrid1dDescTuple(Number<NumInput>{}));
+    using OutGrid1dDescTuple = decltype(GenerateInOutGrid1dDescTuple(Number<NumOutput>{}));
+
+    using GridwiseElementwise = GridwiseElementwise_1D<InGrid1dDescTuple,
+                                                       OutGrid1dDescTuple,
+                                                       InDataTypePointerTuple,
+                                                       OutDataTypePointerTuple,
+                                                       ElementwiseOperation,
+                                                       MPerThread,
+                                                       InScalarPerVectorSeq,
+                                                       OutScalarPerVectorSeq>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, NumDim> lengths,
+                 const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
+                 const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
+                 const std::array<const void*, NumInput> in_dev_buffers,
+                 const std::array<void*, NumOutput> out_dev_buffers,
+                 ElementwiseOperation elementwise_op)
+
+            : lengths_(lengths),
+              inStridesArray_(inStridesArray),
+              outStridesArray_(outStridesArray),
+              elementwise_op_(elementwise_op),
+              blockSize_(256)
+        {
+            in_dev_buffers_ = generate_tuple(
+                [&](auto I) {
+                    using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+                    return static_cast<const DataType*>(in_dev_buffers[I.value]);
+                },
+                Number<NumInput>{});
+
+            out_dev_buffers_ = generate_tuple(
+                [&](auto I) {
+                    using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+                    return static_cast<DataType*>(out_dev_buffers[I.value]);
+                },
+                Number<NumOutput>{});
+        }
+
+        InDataTypePointerTuple in_dev_buffers_;
+        OutDataTypePointerTuple out_dev_buffers_;
+
+        std::array<index_t, NumDim> lengths_;
+        std::array<std::array<index_t, NumDim>, NumInput> inStridesArray_;
+        std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray_;
+
+        ElementwiseOperation elementwise_op_;
+        index_t blockSize_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            index_t gridSize = getAvailableComputeUnitCount(stream_config);
+
+            auto in_grid_1d_desc_tuple = generate_tuple(
+                [&](auto I) {
+                    return MakeDescriptor_M(
+                        arg.lengths_, arg.inStridesArray_[I.value], gridSize, arg.blockSize_);
+                },
+                Number<NumInput>{});
+
+            auto out_grid_1d_desc_tuple = generate_tuple(
+                [&](auto I) {
+                    return MakeDescriptor_M(
+                        arg.lengths_, arg.outStridesArray_[I.value], gridSize, arg.blockSize_);
+                },
+                Number<NumOutput>{});
+
+            const auto kernel = kernel_elementwise_1d<GridwiseElementwise,
+                                                      InGrid1dDescTuple,
+                                                      OutGrid1dDescTuple,
+                                                      InDataTypePointerTuple,
+                                                      OutDataTypePointerTuple,
+                                                      ElementwiseOperation>;
+
+            float elapsed_time = launch_and_time_kernel(stream_config,
+                                                        kernel,
+                                                        dim3(gridSize),
+                                                        dim3(arg.blockSize_),
+                                                        0,
+                                                        in_grid_1d_desc_tuple,
+                                                        out_grid_1d_desc_tuple,
+                                                        arg.in_dev_buffers_,
+                                                        arg.out_dev_buffers_,
+                                                        arg.elementwise_op_);
+            return elapsed_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(arg.lengths_.back() % MPerThread != 0)
+            return false;
+
+        auto IsScalarPerVectorValid = [&](const std::array<index_t, NumDim>& lengths,
+                                          const std::array<index_t, NumDim>& strides,
+                                          index_t scalarPerVector) {
+            if(strides.back() == 1 && lengths.back() % scalarPerVector == 0)
+                return true;
+
+            if(strides.back() != 1 && scalarPerVector == 1)
+                return true;
+
+            return false;
+        };
+
+        bool valid = true;
+        static_for<0, NumInput, 1>{}([&](auto I) {
+            if(!IsScalarPerVectorValid(
+                   arg.lengths_, arg.inStridesArray_[I.value], InScalarPerVectorSeq::At(I)))
+                valid = false;
+        });
+
+        static_for<0, NumOutput, 1>{}([&](auto I) {
+            if(!IsScalarPerVectorValid(
+                   arg.lengths_, arg.outStridesArray_[I.value], OutScalarPerVectorSeq::At(I)))
+                valid = false;
+        });
+
+        return valid;
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto
+    MakeArgument(const std::array<index_t, NumDim> lengths,
+                 const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
+                 const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
+                 const std::array<const void*, NumInput> in_dev_buffers,
+                 const std::array<void*, NumOutput> out_dev_buffers,
+                 ElementwiseOperation elementwise_op)
+    {
+        return Argument{lengths,
+                        inStridesArray,
+                        outStridesArray,
+                        in_dev_buffers,
+                        out_dev_buffers,
+                        elementwise_op};
+    }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, NumDim> lengths,
+                        const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
+                        const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
+                        const std::array<const void*, NumInput> in_dev_buffers,
+                        const std::array<void*, NumOutput> out_dev_buffers,
+                        ElementwiseOperation elementwise_op) override
+    {
+        return std::make_unique<Argument>(lengths,
+                                          inStridesArray,
+                                          outStridesArray,
+                                          in_dev_buffers,
+                                          out_dev_buffers,
+                                          elementwise_op);
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+}; // namespace device
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp
new file mode 100644
index 0000000000..c3416758d1
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp
@@ -0,0 +1,597 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/math.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/reduction_operator.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+// X = Elementwise(input1, input2, input3, ...)
+// Y = Normalization(X, beta, gamma)
+namespace ck {
+template <typename GridwiseElementwiseReduction,
+          typename InDataTypePointerTuple, // Datatype tuple of inputs
+          typename XDataType,              // Datatype of X
+          typename GammaDataType,          // Datatype of Gamma
+          typename BetaDataType,           // Datatype of Beta
+          typename YDataType,              // Datatype of Y
+          typename AccDataType,            // AccDatatype
+          typename XElementwiseOperation,  // Operation of input
+          typename YElementwiseOperation,  // Operation of output of normalization
+          typename InGrid2dDescTuple,      // Descriptor tuple of inputs
+          typename GridDesc_M_K>           // Descriptor of inputs, Gamma, Beta
+__global__ void kernel_elementwise_layernorm(
+    const InGrid2dDescTuple in_grid_2d_desc_tuple,          // Descriptor tuple of inputs
+    const GridDesc_M_K x_grid_desc_m_k,                     // Descriptor of X
+    const GridDesc_M_K gamma_grid_desc_m_k,                 // Descriptor of gamma
+    const GridDesc_M_K beta_grid_desc_m_k,                  // Descriptor of beta
+    const GridDesc_M_K y_grid_desc_m_k,                     // Descriptor of Y
+    index_t num_k_block_tile_iteration,                     //
+    AccDataType epsilon,                                    // Datatype of epsilon
+    const InDataTypePointerTuple p_in_global_tuple,         // Ptr tuple of input matrixs
+    const GammaDataType* const __restrict__ p_gamma_global, // Ptr of gamma
+    const BetaDataType* const __restrict__ p_beta_global,   // Ptr of beta
+    YDataType* const __restrict__ p_y_global,               // Ptr of y
+    const XElementwiseOperation x_elementwise_op,           // Operation of input
+    const YElementwiseOperation y_elementwise_op)           // Operation of output of normalization
+{
+    extern __shared__ XDataType p_x_lds[];
+    GridwiseElementwiseReduction::Run(in_grid_2d_desc_tuple,      // Descriptor tuple of inputs
+                                      x_grid_desc_m_k,            // Descriptor of X
+                                      gamma_grid_desc_m_k,        // Descriptor of Gamma
+                                      beta_grid_desc_m_k,         // Descriptor of Beta
+                                      y_grid_desc_m_k,            // Descriptor of Y
+                                      num_k_block_tile_iteration, //
+                                      epsilon,                    // epsilon
+                                      p_in_global_tuple,          // Ptr tuple of inputs
+                                      p_x_lds,                    // Ptr of X
+                                      p_gamma_global,             // Ptr of gamma
+                                      p_beta_global,              // Ptr of beta
+                                      p_y_global,                 // Ptr of Y
+                                      x_elementwise_op,           // Operation of input
+                                      y_elementwise_op); // Operation of output of normalization
+};
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Y = LayerNorm(A + B, Beta, Gamma)
+template <typename InDataTypeTuple,       // Datatype of inputs
+          typename GammaDataType,         // Datatype of gamma
+          typename BetaDataType,          // Datatype of beta
+          typename AccDataType,           //
+          typename YDataType,             //
+          typename XElementwiseOperation, //
+          typename YElementwiseOperation, //
+          index_t Rank,                   //
+          index_t NumReduceDim,           //
+          index_t BlockSize,              //
+          index_t MThreadClusterSize,     // Num of threads in a block on M direction
+          index_t KThreadClusterSize,     // Num of threads in a block on N direction
+          index_t MThreadSliceSize,       // Each thread calculate rows
+          index_t KThreadSliceSize,       // Each thread calculate columns
+          index_t XYSrcVectorDim,         // Dimension to do reduce
+          index_t XSrcVectorSize,         // Size to fetch source x
+          index_t GammaSrcVectorDim,      // Dimension for gamma to do reduce
+          index_t GammaSrcVectorSize,     // Size to fetch source gamma
+          index_t BetaSrcVectorDim,       // Dimension for beta to do reduce
+          index_t BetaSrcVectorSize,      // Size to fetch source beta
+          index_t YDstVectorSize>         // Size to write destination Y
+struct DeviceElementwiseNormalizationImpl
+    : public DeviceElementwiseNormalization<InDataTypeTuple,
+                                            GammaDataType,
+                                            BetaDataType,
+                                            AccDataType,
+                                            YDataType,
+                                            XElementwiseOperation,
+                                            YElementwiseOperation,
+                                            Rank,
+                                            NumReduceDim>
+{
+    static constexpr int NumInput = InDataTypeTuple::Size();
+
+    using XDataType = YDataType;
+
+    static_assert(
+        (KThreadSliceSize % GammaSrcVectorSize == 0),
+        "Invalid thread slice sizes and/or gamma vector sizes configuration, please check!");
+
+    static_assert(
+        (KThreadSliceSize % BetaSrcVectorSize == 0),
+        "Invalid thread slice sizes and/or beta vector sizes configuration, please check!");
+
+    static constexpr index_t M_BlockTileSize =
+        MThreadClusterSize * MThreadSliceSize; // num of rows calculated in a block
+    static constexpr index_t K_BlockTileSize =
+        KThreadClusterSize * KThreadSliceSize; // num of columns calculated in a block
+
+    static auto GenerateInDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+                return static_cast<const DataType*>(nullptr);
+            },
+            Number<NumInput>{});
+    };
+
+    using InDataTypePointerTuple = decltype(GenerateInDataTypePointerTuple());
+
+    static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
+                                    const std::vector<index_t>& inStrides,
+                                    int blkGroupSize,
+                                    int numBlockTileIteration)
+    {
+        constexpr index_t NumInvariantDim  = Rank - NumReduceDim;
+        static constexpr index_t numSrcDim = Rank;
+        static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+        const auto tupleSrcLengths = make_tuple_from_array(inLengths, Number<numSrcDim>{});
+        const auto tupleSrcStrides = make_tuple_from_array(inStrides, Number<numSrcDim>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, numSrcDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths =
+                    make_tuple_from_array_and_index_seq(inLengths, ReduceDims{});
+                const auto invariantDimLengths =
+                    make_tuple_from_array_and_index_seq(inLengths, InvariantDims{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    template <index_t TupleSize>
+    static auto GenerateSrcGrid2dDescTuple(Number<TupleSize>)
+    {
+        return generate_tuple([&](auto) { return MakeSrc2dDescriptor({1}, {1}, 1, 1); },
+                              Number<TupleSize>{});
+    };
+
+    using InGrid2dDescTuple = decltype(GenerateSrcGrid2dDescTuple(Number<NumInput>{}));
+
+    using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1));
+
+    using GridwiseReduceLayernormGeneric =
+        GridwiseElementwiseLayernormWelfordVariance_mk_to_mk<InDataTypePointerTuple,
+                                                             XDataType,
+                                                             GammaDataType,
+                                                             BetaDataType,
+                                                             YDataType,
+                                                             AccDataType,
+                                                             XElementwiseOperation,
+                                                             YElementwiseOperation,
+                                                             InGrid2dDescTuple,
+                                                             GridDesc_M_K,
+                                                             BlockSize,
+                                                             MThreadClusterSize,
+                                                             KThreadClusterSize,
+                                                             MThreadSliceSize,
+                                                             KThreadSliceSize,
+                                                             XYSrcVectorDim,
+                                                             XSrcVectorSize,
+                                                             GammaSrcVectorDim,
+                                                             GammaSrcVectorSize,
+                                                             BetaSrcVectorDim,
+                                                             BetaSrcVectorSize,
+                                                             XYSrcVectorDim,
+                                                             YDstVectorSize,
+                                                             false>;
+
+    using GridwiseReduceLayernormSweepOnce =
+        GridwiseElementwiseLayernormWelfordVariance_mk_to_mk<InDataTypePointerTuple,
+                                                             XDataType,
+                                                             GammaDataType,
+                                                             BetaDataType,
+                                                             YDataType,
+                                                             AccDataType,
+                                                             XElementwiseOperation,
+                                                             YElementwiseOperation,
+                                                             InGrid2dDescTuple,
+                                                             GridDesc_M_K,
+                                                             BlockSize,
+                                                             MThreadClusterSize,
+                                                             KThreadClusterSize,
+                                                             MThreadSliceSize,
+                                                             KThreadSliceSize,
+                                                             XYSrcVectorDim,
+                                                             XSrcVectorSize,
+                                                             GammaSrcVectorDim,
+                                                             GammaSrcVectorSize,
+                                                             BetaSrcVectorDim,
+                                                             BetaSrcVectorSize,
+                                                             XYSrcVectorDim,
+                                                             YDstVectorSize,
+                                                             true>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::vector<index_t> lengths,
+                 const std::array<std::vector<index_t>, NumInput> inStridesArray,
+                 const std::vector<index_t> gammaStrides,
+                 const std::vector<index_t> betaStrides,
+                 const std::vector<index_t> yStrides,
+                 const std::vector<index_t> reduceDims,
+                 XElementwiseOperation x_elementwise_op,
+                 YElementwiseOperation y_elementwise_op,
+                 double epsilon,
+                 const std::array<const void*, NumInput> in_dev_buffers,
+                 const GammaDataType* p_gamma,
+                 const BetaDataType* p_beta,
+                 YDataType* p_y)
+            : p_gamma_(p_gamma),
+              p_beta_(p_beta),
+              p_y_(p_y),
+              x_elementwise_op_(x_elementwise_op),
+              y_elementwise_op_(y_elementwise_op)
+        {
+            epsilon_ = static_cast<AccDataType>(epsilon);
+
+            Lengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(lengths, reduceDims);
+            for(int i = 0; i < NumInput; i++)
+            {
+                inStridesArray_[i] =
+                    shuffle_tensor_dimensions<Rank, NumReduceDim>(inStridesArray[i], reduceDims);
+            }
+
+            yStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(yStrides, reduceDims);
+            xStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(yStrides, reduceDims);
+
+            gammaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(gammaStrides, reduceDims);
+            betaStrides_  = shuffle_tensor_dimensions<Rank, NumReduceDim>(betaStrides, reduceDims);
+
+            in_dev_buffers_ = generate_tuple(
+                [&](auto I) {
+                    using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
+                    return static_cast<const DataType*>(in_dev_buffers[I.value]);
+                },
+                Number<NumInput>{});
+
+            long_index_t invariant_total_length;
+            long_index_t reduce_total_length;
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(Lengths_);
+
+            blkGroupSize_          = 1;
+            numBlockTileIteration_ = (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
+
+            gridSize_ = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                        M_BlockTileSize * blkGroupSize_;
+
+            in_grid_2d_desc_tuple_ = generate_tuple(
+                [&](auto I) {
+                    return MakeSrc2dDescriptor(
+                        Lengths_, inStridesArray_[I.value], blkGroupSize_, numBlockTileIteration_);
+                },
+                Number<NumInput>{});
+
+            x_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, xStrides_, blkGroupSize_, numBlockTileIteration_);
+
+            gamma_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, gammaStrides_, blkGroupSize_, numBlockTileIteration_);
+
+            beta_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, betaStrides_, blkGroupSize_, numBlockTileIteration_);
+
+            y_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, yStrides_, blkGroupSize_, numBlockTileIteration_);
+
+            sweep_once_ =
+                x_grid_desc_m_k_.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize;
+
+            if(!sweep_once_) // if not sweep once, compute memory size for matrix X in lds for
+                             // store Intermediate results
+            {
+                int block_TileSize = M_BlockTileSize * reduce_total_length;
+                x_lds_size_        = block_TileSize * sizeof(XDataType);
+            }
+            else
+                x_lds_size_ = 0;
+        }
+
+        AccDataType epsilon_;
+
+        InDataTypePointerTuple in_dev_buffers_;
+        const GammaDataType* p_gamma_;
+        const BetaDataType* p_beta_;
+        YDataType* p_y_;
+
+        std::vector<index_t> Lengths_;
+        std::array<std::vector<index_t>, NumInput> inStridesArray_;
+        std::vector<index_t> xStrides_;
+        std::vector<index_t> gammaStrides_;
+        std::vector<index_t> betaStrides_;
+        std::vector<index_t> yStrides_;
+
+        XElementwiseOperation x_elementwise_op_;
+        YElementwiseOperation y_elementwise_op_;
+
+        int blkGroupSize_;
+        int numBlockTileIteration_;
+        size_t gridSize_;
+
+        InGrid2dDescTuple in_grid_2d_desc_tuple_;
+        GridDesc_M_K x_grid_desc_m_k_;
+        GridDesc_M_K gamma_grid_desc_m_k_;
+        GridDesc_M_K beta_grid_desc_m_k_;
+        GridDesc_M_K y_grid_desc_m_k_;
+        bool sweep_once_;
+        int x_lds_size_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto kernel_main =
+                arg.sweep_once_ ? kernel_elementwise_layernorm<GridwiseReduceLayernormSweepOnce,
+                                                               InDataTypePointerTuple,
+                                                               XDataType,
+                                                               GammaDataType,
+                                                               BetaDataType,
+                                                               YDataType,
+                                                               AccDataType,
+                                                               XElementwiseOperation,
+                                                               YElementwiseOperation,
+                                                               InGrid2dDescTuple,
+                                                               GridDesc_M_K>
+                                : kernel_elementwise_layernorm<GridwiseReduceLayernormGeneric,
+                                                               InDataTypePointerTuple,
+                                                               XDataType,
+                                                               GammaDataType,
+                                                               BetaDataType,
+                                                               YDataType,
+                                                               AccDataType,
+                                                               XElementwiseOperation,
+                                                               YElementwiseOperation,
+                                                               InGrid2dDescTuple,
+                                                               GridDesc_M_K>;
+
+            float avg_time = 0;
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize_),
+                                               dim3(BlockSize),
+                                               arg.x_lds_size_,
+                                               arg.in_grid_2d_desc_tuple_,
+                                               arg.x_grid_desc_m_k_,
+                                               arg.gamma_grid_desc_m_k_,
+                                               arg.beta_grid_desc_m_k_,
+                                               arg.y_grid_desc_m_k_,
+                                               arg.numBlockTileIteration_,
+                                               arg.epsilon_,
+                                               arg.in_dev_buffers_,
+                                               arg.p_gamma_,
+                                               arg.p_beta_,
+                                               arg.p_y_,
+                                               arg.x_elementwise_op_,
+                                               arg.y_elementwise_op_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
+
+        constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+        if constexpr(XYSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return false;
+            }
+            else
+            {
+                for(int i = 0; i < NumInput; i++)
+                {
+                    if(p_arg_->inStridesArray_[i][NumInvariantDim - 1] != 1)
+                        return false;
+                }
+
+                if(p_arg_->inStridesArray_[0][NumInvariantDim - 1] != 1 &&
+                   p_arg_->inStridesArray_[1][NumInvariantDim - 1] != 1)
+                    return false;
+
+                if(p_arg_->invariant_lowest_length % XSrcVectorSize != 0)
+                    return false;
+            };
+        }
+        else
+        {
+            for(int i = 0; i < NumInput; i++)
+            {
+                if(p_arg_->inStridesArray_[i][Rank - 1] != 1)
+                    return false;
+            }
+
+            if(p_arg_->Lengths_[Rank - 1] % XSrcVectorSize != 0)
+                return false;
+        };
+
+        if(p_arg_->Lengths_[Rank - 1] % YDstVectorSize != 0)
+        {
+            return false;
+        }
+
+        auto IsScalarPerVectorValid = [](bool isLastDimensionCoalesced, int scalarPerVector) {
+            bool ret = true;
+
+            if(!isLastDimensionCoalesced)
+                ret = scalarPerVector == 1;
+            else
+                ret = KThreadSliceSize % scalarPerVector == 0;
+
+            return ret;
+        };
+
+        if(!IsScalarPerVectorValid(p_arg_->gammaStrides_.back() == 1, GammaSrcVectorSize))
+            return false;
+
+        if(!IsScalarPerVectorValid(p_arg_->betaStrides_.back() == 1, BetaSrcVectorSize))
+            return false;
+
+        // if fastest dim is not reduced
+        if constexpr(XYSrcVectorDim == 0) //
+        {
+            if(p_arg_->gammaStrides_[NumInvariantDim - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % GammaSrcVectorSize != 0)
+                return (false);
+        }
+        else // if fastest dim is reduced
+        {
+            if(p_arg_->gammaStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % GammaSrcVectorSize != 0)
+                return (false);
+        }
+
+        // if fastest dim is not reduced
+        if constexpr(XYSrcVectorDim == 0)
+        {
+            if(p_arg_->betaStrides_[NumInvariantDim - 1] != 1)
+                return (false);
+
+            if(p_arg_->invariant_lowest_length % BetaSrcVectorSize != 0)
+                return (false);
+        }
+        else // if fastest dim is reduced
+        {
+            if(p_arg_->betaStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % BetaSrcVectorSize != 0)
+                return (false);
+        }
+
+        if(p_arg_->x_lds_size_ >= 65536)
+        {
+            return (false);
+        }
+
+        return true;
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> lengths,
+                        const std::array<std::vector<index_t>, NumInput> inStridesArray,
+                        const std::vector<index_t> gammaStrides,
+                        const std::vector<index_t> betaStrides,
+                        const std::vector<index_t> yStrides,
+                        const std::vector<index_t> reduceDims,
+                        double epsilon,
+                        const std::array<const void*, NumInput> in_dev_buffers,
+                        const void* p_gamma,
+                        const void* p_beta,
+                        void* p_y,
+                        XElementwiseOperation x_elementwise_op,
+                        YElementwiseOperation y_elementwise_op) override
+    {
+        return std::make_unique<Argument>(lengths,
+                                          inStridesArray,
+                                          gammaStrides,
+                                          betaStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          x_elementwise_op,
+                                          y_elementwise_op,
+                                          epsilon,
+                                          in_dev_buffers,
+                                          static_cast<const GammaDataType*>(p_gamma),
+                                          static_cast<const BetaDataType*>(p_beta),
+                                          static_cast<YDataType*>(p_y));
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceElementwiseNormalizationImpl<" << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "XYSrcVectorDim_" << XYSrcVectorDim  << ",";
+        str << "VectorSize_X" << XSrcVectorSize << "_Gamma" << GammaSrcVectorSize << "_Beta" << BetaSrcVectorSize << "_Y" << YDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..5bfef60af2
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp
@@ -0,0 +1,888 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Note: inter-wave loop scheduler is rolled out to c-shuffle version first. Becuase non c-shuffle
+// version currently has compiler issues with register spill which further causes validation
+// failures.
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename BiasDataType,
+          typename D0DataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename ReduceAccDataType,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename D0ElementwiseOperation,
+          typename ReduceOperations,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          typename ReduceGlobalMemoryDataOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+          index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+          index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGemmBiasAddReduce_Xdl_CShuffle : public DeviceGemmReduce<1, ReduceOperations::Size()>
+{
+    using DeviceOp = DeviceGemmBiasAddReduce_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(M)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_right_pad_transform(MRaw, MPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto NPad = N - NRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(NRaw, NPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(N)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_right_pad_transform(NRaw, NPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideC));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto NPad = N - NRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(NRaw, NPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_grid_desc_mraw_nraw;
+        }
+    }
+
+    // assume D is packed tensor
+    static auto MakeReduceGridDescriptor_M(index_t MRaw)
+    {
+        const auto d_grid_desc_mraw = make_naive_tensor_descriptor_packed(make_tuple(MRaw));
+
+        const auto M    = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto MPad = M - MRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                     GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M
+            return transform_tensor_descriptor(d_grid_desc_mraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad)),
+                                               make_tuple(Sequence<0>{}),
+                                               make_tuple(Sequence<0>{}));
+        }
+        else
+        {
+            // not pad M
+            return d_grid_desc_mraw;
+        }
+    }
+
+    using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
+    using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
+    using CGridDesc_M_N       = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+    using C0GridDesc_M_N      = decltype(MakeCGridDescriptor_M_N(1, 1, 0));
+    using C1GridDesc_M_N      = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+    using ReduceGridDesc_M    = decltype(MakeReduceGridDescriptor_M(1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmBiasAddReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        BiasDataType,
+        D0DataType,
+        ReduceAccDataType,
+        ReducePtrsGlobal,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        D0ElementwiseOperation,
+        ReduceOperations,
+        ReduceInElementwiseOperations,
+        ReduceAccElementwiseOperations,
+        InMemoryDataOperationEnum::Set,
+        ReduceGlobalMemoryDataOperation,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        C0GridDesc_M_N,
+        C1GridDesc_M_N,
+        ReduceGridDesc_M,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+        CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+        CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+        LoopSched>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 const BiasDataType* p_bias_grid,
+                 const D0DataType* p_d0_grid,
+                 ReducePtrsGlobal p_reduces_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 index_t StrideC1,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op,
+                 D0ElementwiseOperation d0_element_op,
+                 ReduceInElementwiseOperations reduce_in_element_ops,
+                 ReduceAccElementwiseOperations reduce_out_element_ops)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              p_bias_grid_{p_bias_grid},
+              p_d0_grid_{p_d0_grid},
+              p_reduces_grid_{p_reduces_grid},
+              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, NRaw, StrideC)},
+              c0_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, NRaw, 0)},
+              c1_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, NRaw, StrideC1)},
+              reduce_grid_desc_m_{DeviceOp::MakeReduceGridDescriptor_M(MRaw)},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              c0_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              c1_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              reduce_grid_desc_mblock_mperblock_{},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op},
+              d0_element_op_{d0_element_op},
+              reduce_in_element_ops_{reduce_in_element_ops},
+              reduce_out_element_ops_{reduce_out_element_ops}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+
+                c0_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c0_grid_desc_m_n_);
+
+                c1_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c1_grid_desc_m_n_);
+
+                reduce_grid_desc_mblock_mperblock_ =
+                    GridwiseGemm::MakeReduceGridDescriptor_MBlock_MPerBlock(reduce_grid_desc_m_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        const BiasDataType* p_bias_grid_;
+        const D0DataType* p_d0_grid_;
+        ReducePtrsGlobal p_reduces_grid_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        C0GridDesc_M_N c0_grid_desc_m_n_;
+        C1GridDesc_M_N c1_grid_desc_m_n_;
+        ReduceGridDesc_M reduce_grid_desc_m_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c0_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c1_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock
+            reduce_grid_desc_mblock_mperblock_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+        D0ElementwiseOperation d0_element_op_;
+        ReduceInElementwiseOperations reduce_in_element_ops_;
+        ReduceAccElementwiseOperations reduce_out_element_ops_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                            arg.b_grid_desc_bk0_n_bk1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float elapsed_time = 0.0f;
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_bias_add_reduce_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    BiasDataType,
+                    D0DataType,
+                    ReducePtrsGlobal,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    D0ElementwiseOperation,
+                    ReduceInElementwiseOperations,
+                    ReduceAccElementwiseOperations,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    true>;
+
+                elapsed_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_bias_grid_,
+                                           arg.p_d0_grid_,
+                                           arg.p_reduces_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.d0_element_op_,
+                                           arg.reduce_in_element_ops_,
+                                           arg.reduce_out_element_ops_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.c0_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.c1_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.reduce_grid_desc_mblock_mperblock_,
+                                           arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_bias_add_reduce_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    BiasDataType,
+                    D0DataType,
+                    ReducePtrsGlobal,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    D0ElementwiseOperation,
+                    ReduceInElementwiseOperations,
+                    ReduceAccElementwiseOperations,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    false>;
+
+                elapsed_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_bias_grid_,
+                                           arg.p_d0_grid_,
+                                           arg.p_reduces_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.d0_element_op_,
+                                           arg.reduce_in_element_ops_,
+                                           arg.reduce_out_element_ops_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.c0_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.c1_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.reduce_grid_desc_mblock_mperblock_,
+                                           arg.block_2_ctile_map_);
+            }
+
+            return elapsed_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static constexpr int NumReduce = ReduceOperations::Size();
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             const void* p_bias,
+                             std::array<const void*, 1> p_ds,
+                             void* p_c,
+                             std::array<void*, NumReduce> p_reduces,
+                             ck::index_t M,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t StrideA,
+                             ck::index_t StrideB,
+                             ck::index_t StrideC,
+                             std::array<ck::index_t, 1> StrideDs,
+                             std::array<void*, 3> gemm_element_ops,
+                             std::array<void*, 1> d_element_ops,
+                             std::array<void*, NumReduce> reduce_in_element_op,
+                             std::array<void*, NumReduce> reduce_out_element_op)
+    {
+        ReducePtrsGlobal reduce_tuple = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReducePtrsGlobal{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return static_cast<T*>(p_reduces[I]);
+            },
+            Number<NumReduce>{});
+
+        ReduceInElementwiseOperations reduce_in_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceInElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_in_element_op[I]));
+            },
+            Number<NumReduce>{});
+        ReduceAccElementwiseOperations reduce_out_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceAccElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_out_element_op[I]));
+            },
+            Number<NumReduce>{});
+
+        AElementwiseOperation a_element_op =
+            *(static_cast<AElementwiseOperation*>(gemm_element_ops[0]));
+        BElementwiseOperation b_element_op =
+            *(static_cast<BElementwiseOperation*>(gemm_element_ops[1]));
+        CElementwiseOperation c_element_op =
+            *(static_cast<CElementwiseOperation*>(gemm_element_ops[2]));
+        D0ElementwiseOperation d_element_op =
+            *(static_cast<D0ElementwiseOperation*>(d_element_ops[0]));
+
+        return Argument{static_cast<const ADataType*>(p_a),
+                        static_cast<const BDataType*>(p_b),
+                        static_cast<CDataType*>(p_c),
+                        static_cast<const BiasDataType*>(p_bias),
+                        static_cast<const D0DataType*>(p_ds[0]),
+                        reduce_tuple,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        StrideDs[0],
+                        a_element_op,
+                        b_element_op,
+                        c_element_op,
+                        d_element_op,
+                        reduce_in_element_ops,
+                        reduce_out_element_ops};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const void* p_bias,
+                        std::array<const void*, 1> p_ds,
+                        void* p_c,
+                        std::array<void*, NumReduce> p_reduces,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        std::array<ck::index_t, 1> StrideDs,
+                        std::array<void*, 3> gemm_element_ops,
+                        std::array<void*, 1> d_element_ops,
+                        std::array<void*, NumReduce> reduce_in_element_op,
+                        std::array<void*, NumReduce> reduce_out_element_op,
+                        index_t /* KBatch */ = 1) override
+    {
+        ReducePtrsGlobal reduce_tuple = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReducePtrsGlobal{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return static_cast<T*>(p_reduces[I]);
+            },
+            Number<NumReduce>{});
+
+        ReduceInElementwiseOperations reduce_in_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceInElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_in_element_op[I]));
+            },
+            Number<NumReduce>{});
+        ReduceAccElementwiseOperations reduce_out_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceAccElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_out_element_op[I]));
+            },
+            Number<NumReduce>{});
+
+        AElementwiseOperation a_element_op =
+            *(static_cast<AElementwiseOperation*>(gemm_element_ops[0]));
+        BElementwiseOperation b_element_op =
+            *(static_cast<BElementwiseOperation*>(gemm_element_ops[1]));
+        CElementwiseOperation c_element_op =
+            *(static_cast<CElementwiseOperation*>(gemm_element_ops[2]));
+        D0ElementwiseOperation d_element_op =
+            *(static_cast<D0ElementwiseOperation*>(d_element_ops[0]));
+
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          static_cast<const BiasDataType*>(p_bias),
+                                          static_cast<const D0DataType*>(p_ds[0]),
+                                          reduce_tuple,
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          StrideDs[0],
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op,
+                                          d_element_op,
+                                          reduce_in_element_ops,
+                                          reduce_out_element_ops);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmBiasAddReduce_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp
new file mode 100644
index 0000000000..eedf384cd9
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp
@@ -0,0 +1,613 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_dl_algorithm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <
+    typename ADataType,
+    typename BDataType,
+    typename CDataType,
+    typename AccDataType,
+    typename ALayout,
+    typename BLayout,
+    typename CLayout,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    GemmSpecialization GemmSpec,
+    index_t BlockSize,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t K0PerBlock,
+    index_t K1,
+    index_t M1PerThread,
+    index_t N1PerThread,
+    index_t KPerThread,
+    typename M1N1ThreadClusterM1Xs,
+    typename M1N1ThreadClusterN1Xs,
+    typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+    typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+    typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+    typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+    typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+    typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+    typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+    typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+    typename CThreadTransferSrcDstAccessOrder,
+    index_t CThreadTransferSrcDstVectorDim,
+    index_t CThreadTransferDstScalarPerVector,
+    GemmDlAlgorithm GemmDlAlg = GemmDlAlgorithm::Default,
+    enable_if_t<
+        is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<CElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+        bool> = false>
+struct DeviceGemmDl : public DeviceGemm<ALayout,
+                                        BLayout,
+                                        CLayout,
+                                        ADataType,
+                                        BDataType,
+                                        CDataType,
+                                        AElementwiseOperation,
+                                        BElementwiseOperation,
+                                        CElementwiseOperation>
+
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static constexpr auto K1Number = Number<K1>{};
+
+    static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(M, PadM)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+    }
+
+    using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
+    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
+    using CGridDesc_M_N     = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm =
+        GridwiseGemmDl_km_kn_mn_v1r3<BlockSize,
+                                     ADataType,
+                                     AccDataType,
+                                     CDataType,
+                                     InMemoryDataOperationEnum::Set,
+                                     AGridDesc_K0_M_K1,
+                                     BGridDesc_K0_N_K1,
+                                     CGridDesc_M_N,
+                                     MPerBlock,
+                                     NPerBlock,
+                                     K0PerBlock,
+                                     K1,
+                                     M1PerThread,
+                                     N1PerThread,
+                                     KPerThread,
+                                     M1N1ThreadClusterM1Xs,
+                                     M1N1ThreadClusterN1Xs,
+                                     ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+                                     ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+                                     ABlockTransferThreadClusterArrangeOrder,
+                                     ABlockTransferSrcAccessOrder,
+                                     ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+                                     ABlockTransferSrcVectorTensorContiguousDimOrder,
+                                     ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+                                     BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+                                     BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+                                     BBlockTransferThreadClusterArrangeOrder,
+                                     BBlockTransferSrcAccessOrder,
+                                     BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+                                     BBlockTransferSrcVectorTensorContiguousDimOrder,
+                                     BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+                                     CThreadTransferSrcDstAccessOrder,
+                                     CThreadTransferSrcDstVectorDim,
+                                     CThreadTransferDstScalarPerVector,
+                                     GemmDlAlg>;
+
+    using AGridDesc_K0_M0_M1_K1 =
+        decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_K0_M_K1{}));
+    using BGridDesc_K0_N0_N1_K1 =
+        decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{}));
+    using CGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
+    using DefaultBlock2CTileMap =
+        decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(CGridDesc_M_N{}));
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 index_t M,
+                 index_t N,
+                 index_t K,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 index_t M01,
+                 index_t N01,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              a_grid_desc_k0_m0_m1_k1_{},
+              b_grid_desc_k0_n0_n1_k1_{},
+              c_grid_desc_m0_m10_m11_n0_n10_n11_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+            a_grid_desc_k0_m_k1_ = DeviceGemmDl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
+            b_grid_desc_k0_n_k1_ = DeviceGemmDl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
+            c_grid_desc_m_n_     = DeviceGemmDl::MakeCGridDescriptor_M_N(M, N, StrideC);
+
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_))
+            {
+                a_grid_desc_k0_m0_m1_k1_ =
+                    GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(a_grid_desc_k0_m_k1_);
+                b_grid_desc_k0_n0_n1_k1_ =
+                    GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(b_grid_desc_k0_n_k1_);
+                c_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(c_grid_desc_m_n_);
+
+                block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+
+        AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1_;
+        BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1_;
+        CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11_;
+
+        DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // TODO: unused, but may be useful in future.
+        index_t M01_;
+        index_t N01_;
+
+        // TODO: unused since gridwise_gemm_dl_v1r3 does NOT support prologue for the time being.
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceGemmDl::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if DEBUG_LOG
+            {
+                std::cout << "arg.a_grid_desc_k0_m0_m1_k1_{"
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I0) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n0_n1_k1_{"
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I0) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(
+                   arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.c_grid_desc_m_n_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdl_v2r3 has invalid setting");
+            }
+
+            const index_t grid_size = GridwiseGemm::CalculateGridSize(
+                arg.c_grid_desc_m_n_.GetLength(I0), arg.c_grid_desc_m_n_.GetLength(I1));
+
+            const auto K0                    = arg.a_grid_desc_k0_m0_m1_k1_.GetLength(I0);
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
+            const bool has_double_tail_k_block_loop =
+                GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
+
+            float ave_time = 0;
+
+            if(has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                const auto kernel =
+                    kernel_gemm_dl_v1r3<GridwiseGemm,
+                                        ADataType,
+                                        CDataType,
+                                        remove_reference_t<AGridDesc_K0_M0_M1_K1>,
+                                        remove_reference_t<BGridDesc_K0_N0_N1_K1>,
+                                        remove_reference_t<CGridDesc_M0_M10_M11_N0_N10_N11>,
+                                        remove_reference_t<DefaultBlock2CTileMap>,
+                                        true,
+                                        true,
+                                        GemmDlAlg>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m0_m1_k1_,
+                                                  arg.b_grid_desc_k0_n0_n1_k1_,
+                                                  arg.c_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
+            {
+                const auto kernel =
+                    kernel_gemm_dl_v1r3<GridwiseGemm,
+                                        ADataType,
+                                        CDataType,
+                                        remove_reference_t<AGridDesc_K0_M0_M1_K1>,
+                                        remove_reference_t<BGridDesc_K0_N0_N1_K1>,
+                                        remove_reference_t<CGridDesc_M0_M10_M11_N0_N10_N11>,
+                                        remove_reference_t<DefaultBlock2CTileMap>,
+                                        true,
+                                        false,
+                                        GemmDlAlg>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m0_m1_k1_,
+                                                  arg.b_grid_desc_k0_n0_n1_k1_,
+                                                  arg.c_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                const auto kernel =
+                    kernel_gemm_dl_v1r3<GridwiseGemm,
+                                        ADataType,
+                                        CDataType,
+                                        remove_reference_t<AGridDesc_K0_M0_M1_K1>,
+                                        remove_reference_t<BGridDesc_K0_N0_N1_K1>,
+                                        remove_reference_t<CGridDesc_M0_M10_M11_N0_N10_N11>,
+                                        remove_reference_t<DefaultBlock2CTileMap>,
+                                        false,
+                                        true,
+                                        GemmDlAlg>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m0_m1_k1_,
+                                                  arg.b_grid_desc_k0_n0_n1_k1_,
+                                                  arg.c_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel =
+                    kernel_gemm_dl_v1r3<GridwiseGemm,
+                                        ADataType,
+                                        CDataType,
+                                        remove_reference_t<AGridDesc_K0_M0_M1_K1>,
+                                        remove_reference_t<BGridDesc_K0_N0_N1_K1>,
+                                        remove_reference_t<CGridDesc_M0_M10_M11_N0_N10_N11>,
+                                        remove_reference_t<DefaultBlock2CTileMap>,
+                                        false,
+                                        false,
+                                        GemmDlAlg>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m0_m1_k1_,
+                                                  arg.b_grid_desc_k0_n0_n1_k1_,
+                                                  arg.c_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                                  arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if constexpr(GemmDlAlg == GemmDlAlgorithm::Dpp8)
+        {
+            if(ck::get_device_name() == "gfx1030")
+            {
+                return GridwiseGemm::CheckValidity(
+                    arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.c_grid_desc_m_n_);
+            }
+            return false;
+        }
+
+        if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
+           ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
+           ck::get_device_name() == "gfx1102")
+        {
+            return GridwiseGemm::CheckValidity(
+                arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.c_grid_desc_m_n_);
+        }
+        return false;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        1,
+                        1,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          1,
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    virtual std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmDl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << M1PerThread << ", "
+            << N1PerThread << ", "
+            << KPerThread
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl_dpp8.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl_dpp8.hpp
new file mode 100644
index 0000000000..336ce31e82
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl_dpp8.hpp
@@ -0,0 +1,133 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_dl_algorithm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <
+    typename ADataType,
+    typename BDataType,
+    typename CDataType,
+    typename AccDataType,
+    typename ALayout,
+    typename BLayout,
+    typename CLayout,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    GemmSpecialization GemmSpec,
+    index_t BlockSize,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t K0PerBlock,
+    index_t K1,
+    index_t M1PerThread,
+    index_t N1PerThread,
+    index_t KPerThread,
+    typename M1N1ThreadClusterM1Xs,
+    typename M1N1ThreadClusterN1Xs,
+    typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+    typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+    typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+    typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+    typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+    typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+    typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+    typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+    typename CThreadTransferSrcDstAccessOrder,
+    index_t CThreadTransferSrcDstVectorDim,
+    index_t CThreadTransferDstScalarPerVector,
+    enable_if_t<
+        is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<CElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+        bool> = false>
+struct DeviceGemmDlDpp8 : public DeviceGemmDl<ADataType,
+                                              BDataType,
+                                              CDataType,
+                                              AccDataType,
+                                              ALayout,
+                                              BLayout,
+                                              CLayout,
+                                              AElementwiseOperation,
+                                              BElementwiseOperation,
+                                              CElementwiseOperation,
+                                              GemmSpec,
+                                              BlockSize,
+                                              MPerBlock,
+                                              NPerBlock,
+                                              K0PerBlock,
+                                              K1,
+                                              M1PerThread,
+                                              N1PerThread,
+                                              KPerThread,
+                                              M1N1ThreadClusterM1Xs,
+                                              M1N1ThreadClusterN1Xs,
+                                              ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+                                              ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+                                              ABlockTransferThreadClusterArrangeOrder,
+                                              ABlockTransferSrcAccessOrder,
+                                              ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+                                              ABlockTransferSrcVectorTensorContiguousDimOrder,
+                                              ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+                                              BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+                                              BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+                                              BBlockTransferThreadClusterArrangeOrder,
+                                              BBlockTransferSrcAccessOrder,
+                                              BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+                                              BBlockTransferSrcVectorTensorContiguousDimOrder,
+                                              BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+                                              CThreadTransferSrcDstAccessOrder,
+                                              CThreadTransferSrcDstVectorDim,
+                                              CThreadTransferDstScalarPerVector,
+                                              GemmDlAlgorithm::Dpp8>
+
+{
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmDlDpp8"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << M1PerThread << ", "
+            << N1PerThread << ", "
+            << KPerThread
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp
new file mode 100644
index 0000000000..ad51096db7
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp
@@ -0,0 +1,673 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_K0_M0_M1_K1,
+          typename BGridDesc_K0_N0_N1_K1,
+          typename DsGridDesc_M0_M10_M11_N0_N10_N11,
+          typename CGridDesc_M0_M10_M11_N0_N10_N11,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop,
+          bool HasDoubleTailKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_dl_multiple_d(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1,
+            const BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1,
+            const DsGridDesc_M0_M10_M11_N0_N10_N11 ds_grid_desc_m0_m10_m11_n0_n10_n11,
+            const CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||             \
+    defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx1030__) || defined(__gfx1100__) || \
+    defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx941__) || defined(__gfx942__))
+
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);
+
+    __shared__ ABDataType p_shared[shared_block_size];
+
+    GridwiseGemm::Run(p_a_grid,
+                      p_b_grid,
+                      p_ds_grid,
+                      p_e_grid,
+                      p_shared,
+                      a_element_op,
+                      b_element_op,
+                      cde_element_op,
+                      a_grid_desc_k0_m0_m1_k1,
+                      b_grid_desc_k0_n0_n1_k1,
+                      ds_grid_desc_m0_m10_m11_n0_n10_n11,
+                      e_grid_desc_m0_m10_m11_n0_n10_n11,
+                      block_2_ctile_map,
+                      integral_constant<bool, HasMainKBlockLoop>{},
+                      integral_constant<bool, HasDoubleTailKBlockLoop>{});
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = a_grid_desc_k0_m0_m1_k1;
+    ignore = b_grid_desc_k0_n0_n1_k1;
+    ignore = ds_grid_desc_m0_m10_m11_n0_n10_n11;
+    ignore = e_grid_desc_m0_m10_m11_n0_n10_n11;
+    ignore = block_2_ctile_map;
+#endif
+}
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t K1,
+          index_t M1PerThread,
+          index_t N1PerThread,
+          index_t KPerThread,
+          typename M1N1ThreadClusterM1Xs,
+          typename M1N1ThreadClusterN1Xs,
+          typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+          typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+          typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+          typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+          typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          enable_if_t<
+              is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+                  is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+              bool> = false>
+struct DeviceGemmMultipleD_Dl : public DeviceGemmMultipleD<ALayout,
+                                                           BLayout,
+                                                           DsLayout,
+                                                           ELayout,
+                                                           ADataType,
+                                                           BDataType,
+                                                           DsDataType,
+                                                           EDataType,
+                                                           AElementwiseOperation,
+                                                           BElementwiseOperation,
+                                                           CDEElementwiseOperation>
+
+{
+    using DeviceOp                      = DeviceGemmMultipleD_Dl;
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static constexpr auto K1Number = Number<K1>{};
+
+    static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(M, PadM)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    template <typename ELay>
+    static auto MakeEGridDescriptor_M_N(index_t M, index_t N, index_t StrideE)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideE));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                                         const std::array<index_t, NumDTensor>& NRaws,
+                                         const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
+    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
+    using DsGridDesc_M_N    = decltype(MakeDsGridDescriptor_M_N({}, {}, {}));
+    using EGridDesc_M_N     = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm =
+        GridwiseGemmDlMultipleD_km_kn_mn<BlockSize,
+                                         ADataType,
+                                         AccDataType,
+                                         DsDataType,
+                                         EDataType,
+                                         AElementwiseOperation,
+                                         BElementwiseOperation,
+                                         CDEElementwiseOperation,
+                                         InMemoryDataOperationEnum::Set,
+                                         AGridDesc_K0_M_K1,
+                                         BGridDesc_K0_N_K1,
+                                         EGridDesc_M_N,
+                                         MPerBlock,
+                                         NPerBlock,
+                                         K0PerBlock,
+                                         K1,
+                                         M1PerThread,
+                                         N1PerThread,
+                                         KPerThread,
+                                         M1N1ThreadClusterM1Xs,
+                                         M1N1ThreadClusterN1Xs,
+                                         ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+                                         ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+                                         ABlockTransferThreadClusterArrangeOrder,
+                                         ABlockTransferSrcAccessOrder,
+                                         ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+                                         ABlockTransferSrcVectorTensorContiguousDimOrder,
+                                         ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+                                         BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+                                         BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+                                         BBlockTransferThreadClusterArrangeOrder,
+                                         BBlockTransferSrcAccessOrder,
+                                         BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+                                         BBlockTransferSrcVectorTensorContiguousDimOrder,
+                                         BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+                                         CThreadTransferSrcDstAccessOrder,
+                                         CThreadTransferSrcDstVectorDim,
+                                         CThreadTransferDstScalarPerVector>;
+
+    using AGridDesc_K0_M0_M1_K1 =
+        decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_K0_M_K1{}));
+    using BGridDesc_K0_N0_N1_K1 =
+        decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{}));
+    using DsGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(DsGridDesc_M_N{}));
+    using EGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(EGridDesc_M_N{}));
+    using DefaultBlock2CTileMap =
+        decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(EGridDesc_M_N{}));
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 index_t M,
+                 index_t N,
+                 index_t K,
+                 index_t StrideA,
+                 index_t StrideB,
+                 std::array<index_t, NumDTensor> StrideDs,
+                 index_t StrideE,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_k0_m0_m1_k1_{},
+              b_grid_desc_k0_n0_n1_k1_{},
+              e_grid_desc_m0_m10_m11_n0_n10_n11_{},
+              block_2_ctile_map_{},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+            a_grid_desc_k0_m_k1_ =
+                DeviceGemmMultipleD_Dl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
+            b_grid_desc_k0_n_k1_ =
+                DeviceGemmMultipleD_Dl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) =
+                    DeviceOp::MakeEGridDescriptor_M_N<DLayout>(M, N, StrideDs[i]);
+            });
+            e_grid_desc_m_n_ =
+                DeviceGemmMultipleD_Dl::MakeEGridDescriptor_M_N<ELayout>(M, N, StrideE);
+
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, e_grid_desc_m_n_))
+            {
+                a_grid_desc_k0_m0_m1_k1_ =
+                    GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(a_grid_desc_k0_m_k1_);
+                b_grid_desc_k0_n0_n1_k1_ =
+                    GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(b_grid_desc_k0_n_k1_);
+
+                ds_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(ds_grid_desc_m_n_);
+
+                e_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(e_grid_desc_m_n_);
+
+                block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1_;
+        BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1_;
+        DsGridDesc_M0_M10_M11_N0_N10_N11 ds_grid_desc_m0_m10_m11_n0_n10_n11_;
+        EGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11_;
+
+        DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // TODO: unused since gridwise_gemm_dl_v1r3 does NOT support prologue for the time being.
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceGemmMultipleD_Dl::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            {
+                std::cout << "arg.a_grid_desc_k0_m0_m1_k1_{"
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I0) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n0_n1_k1_{"
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I0) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.e_grid_desc_m_n_{ " << arg.e_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.e_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+
+            if(!GridwiseGemm::CheckValidity(
+                   arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.e_grid_desc_m_n_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemmDlMultipleD_km_kn_mn has invalid setting");
+            }
+
+            const index_t grid_size = GridwiseGemm::CalculateGridSize(
+                arg.e_grid_desc_m_n_.GetLength(I0), arg.e_grid_desc_m_n_.GetLength(I1));
+
+            auto launch_kernel = [&](auto has_main_k_block_loop,
+                                     auto has_double_tail_k_block_loop) {
+                constexpr bool has_main_loop   = has_main_k_block_loop.value;
+                constexpr bool has_double_loop = has_double_tail_k_block_loop.value;
+
+                const auto kernel =
+                    kernel_gemm_dl_multiple_d<GridwiseGemm,
+                                              ADataType,
+                                              typename GridwiseGemm::DsGridPointer,
+                                              EDataType,
+                                              AElementwiseOperation,
+                                              BElementwiseOperation,
+                                              CDEElementwiseOperation,
+                                              DeviceOp::AGridDesc_K0_M0_M1_K1,
+                                              DeviceOp::BGridDesc_K0_N0_N1_K1,
+                                              DeviceOp::DsGridDesc_M0_M10_M11_N0_N10_N11,
+                                              DeviceOp::EGridDesc_M0_M10_M11_N0_N10_N11,
+                                              DefaultBlock2CTileMap,
+                                              has_main_loop,
+                                              has_double_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_k0_m0_m1_k1_,
+                                              arg.b_grid_desc_k0_n0_n1_k1_,
+                                              arg.ds_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                              arg.e_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                              arg.block_2_ctile_map_);
+            };
+
+            const auto K0                    = arg.a_grid_desc_k0_m0_m1_k1_.GetLength(I0);
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
+            const bool has_double_tail_k_block_loop =
+                GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
+
+            if(has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, true>{});
+            }
+            else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, false>{});
+            }
+            else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx908" ||
+           ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx1030" ||
+           ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx1100" ||
+           ck::get_device_name() == "gfx1101" || ck::get_device_name() == "gfx1102" ||
+           ck::get_device_name() == "gfx941" || ck::get_device_name() == "gfx942")
+        {
+            return GridwiseGemm::CheckValidity(
+                arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.e_grid_desc_m_n_);
+        }
+        else
+        {
+            return false;
+        }
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             std::array<const void*, NumDTensor> p_ds,
+                             void* p_e,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             std::array<ck::index_t, NumDTensor> StrideDs,
+                             index_t StrideE,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideDs,
+                        StrideE,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        index_t M,
+                        index_t N,
+                        index_t K,
+                        index_t StrideA,
+                        index_t StrideB,
+                        std::array<ck::index_t, NumDTensor> StrideDs,
+                        index_t StrideE,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideDs,
+                                          StrideE,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmMultipleD_Dl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << M1PerThread << ", "
+            << N1PerThread << ", "
+            << KPerThread
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..b0efa9d4e4
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp
@@ -0,0 +1,1087 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_welford_second_half_layernorm2d.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemmWelford,
+          typename ABDataType,
+          typename DsPointer,
+          typename EMeanVarDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename MeanVarGridDescriptor_MBlock_MPerBlock_NBlock,
+          typename CountGridDescriptor_MBlock_MPerBlock_NBlock,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_multiple_d_welford_first_half_xdl_cshuffle(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EMeanVarDataType* __restrict__ p_e_grid,
+            EMeanVarDataType* __restrict__ p_welford_mean_grid,
+            EMeanVarDataType* __restrict__ p_welford_var_grid,
+            int32_t* __restrict__ p_welford_count_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const MeanVarGridDescriptor_MBlock_MPerBlock_NBlock
+                mean_var_grid_desc_mblock_mperblock_nblock,
+            const CountGridDescriptor_MBlock_MPerBlock_NBlock
+                count_grid_desc_mblock_mperblock_nblock,
+            const Block2ETileMap block_2_etile_map,
+            index_t NRaw)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemmWelford::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemmWelford::template Run<HasMainKBlockLoop>(
+        p_a_grid,
+        p_b_grid,
+        p_ds_grid,
+        p_e_grid,
+        p_welford_mean_grid,
+        p_welford_var_grid,
+        p_welford_count_grid,
+        p_shared,
+        a_element_op,
+        b_element_op,
+        cde_element_op,
+        a_grid_desc_ak0_m_ak1,
+        b_grid_desc_bk0_n_bk1,
+        ds_grid_desc_mblock_mperblock_nblock_nperblock,
+        e_grid_desc_mblock_mperblock_nblock_nperblock,
+        mean_var_grid_desc_mblock_mperblock_nblock,
+        count_grid_desc_mblock_mperblock_nblock,
+        block_2_etile_map,
+        NRaw);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = p_welford_mean_grid;
+    ignore = p_welford_var_grid;
+    ignore = p_welford_count_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = mean_var_grid_desc_mblock_mperblock_nblock;
+    ignore = count_grid_desc_mblock_mperblock_nblock;
+    ignore = block_2_etile_map;
+    ignore = NRaw;
+#endif
+}
+
+template <typename GridwiseWelfordLayernorm,
+          typename EMeanVarDataType,
+          typename HDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename ComputeDataType,
+          typename EHGridDesc_M_N,
+          typename LayernormMeanVarGridDesc_M_NBlock,
+          typename LayernormCountGridDesc_M_NBlock,
+          typename GammaBetaGridDesc_N,
+          typename HElementwiseOperation>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_welford_layernorm2d_second_half(
+            const EMeanVarDataType* __restrict__ p_e_grid,
+            const EMeanVarDataType* __restrict__ p_in_welford_mean_grid,
+            const EMeanVarDataType* __restrict__ p_in_welford_var_grid,
+            const int32_t* __restrict__ p_in_welford_count_grid,
+            const GammaDataType* __restrict__ p_gamma_grid,
+            const BetaDataType* __restrict__ p_beta_grid,
+            HDataType* __restrict__ p_h_grid,
+            const EHGridDesc_M_N e_grid_desc_m_n,
+            const EHGridDesc_M_N h_grid_desc_m_n,
+            const LayernormMeanVarGridDesc_M_NBlock mean_var_grid_desc_m_nblock,
+            const LayernormCountGridDesc_M_NBlock count_grid_desc_m_nblock,
+            const GammaBetaGridDesc_N gamma_grid_desc_n,
+            const GammaBetaGridDesc_N beta_grid_desc_n,
+            index_t numMeanVarCountBlockTileIteration_N,
+            index_t NBlockClusterLength,
+            ComputeDataType epsilon,
+            HElementwiseOperation h_element_op)
+{
+    GridwiseWelfordLayernorm::Run(p_e_grid,
+                                  p_in_welford_mean_grid,
+                                  p_in_welford_var_grid,
+                                  p_in_welford_count_grid,
+                                  p_gamma_grid,
+                                  p_beta_grid,
+                                  p_h_grid,
+                                  e_grid_desc_m_n,
+                                  h_grid_desc_m_n,
+                                  mean_var_grid_desc_m_nblock,
+                                  count_grid_desc_m_nblock,
+                                  gamma_grid_desc_n,
+                                  beta_grid_desc_n,
+                                  numMeanVarCountBlockTileIteration_N,
+                                  NBlockClusterLength,
+                                  epsilon,
+                                  h_element_op);
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   output : H[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+//   H = layernorm(E)
+// Assume:
+//   D0, D1, ... and E have the same layout
+//   Calculate mean & variance along N dimension in layernorm(E)
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename HLayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EMeanVarDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename HDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename HElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t GemmMPerBlock,
+          index_t GemmNPerBlock,
+          index_t GemmKPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename PostShuffleThreadClusterSize_M_N,
+          index_t PostShuffleScalarPerVector,
+          typename LayernormThreadClusterSize_M_N,
+          index_t LayernormThreadSliceSize_M,
+          LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
+    : public DeviceGemmMultipleDLayernorm<ALayout,
+                                          BLayout,
+                                          DsLayout,
+                                          HLayout,
+                                          ADataType,
+                                          BDataType,
+                                          DsDataType,
+                                          GammaDataType,
+                                          BetaDataType,
+                                          HDataType,
+                                          AElementwiseOperation,
+                                          BElementwiseOperation,
+                                          CDEElementwiseOperation,
+                                          HElementwiseOperation>
+{
+    // EDataType, MeanDataType and VarDataType must be the same.
+    // eg. M, N, K = [1, 1, 1],
+    // in case of layernorm, divisor = 1 / sqrt(var + 1e-5) = 316.227783
+    // if (x - mean) != 0, (x - mean) * divisor * gamma might be too large
+    // However, (x - mean) * divisor * gamma should be 0 in this case
+
+    using DeviceOp = DeviceGemmMultipleDLayernorm_Xdl_CShuffle;
+    using ELayout  = HLayout;
+
+    static constexpr index_t NumDTensor                  = DsDataType::Size();
+    static constexpr index_t LayernormHDstVectorSize     = PostShuffleScalarPerVector;
+    static constexpr index_t LayernormGammaSrcVectorSize = PostShuffleScalarPerVector;
+    static constexpr index_t LayernormBetaSrcVectorSize  = PostShuffleScalarPerVector;
+    static constexpr index_t LayernormESrcVectorSize     = PostShuffleScalarPerVector;
+    static constexpr index_t LayernormThreadSliceSize_N  = PostShuffleScalarPerVector;
+    using LayernormBlockTileSize_M_N =
+        Sequence<LayernormThreadClusterSize_M_N::At(0) * LayernormThreadSliceSize_M,
+                 LayernormThreadClusterSize_M_N::At(1) * LayernormThreadSliceSize_N>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto matrix_padder = MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+        GemmMPerBlock, GemmNPerBlock, GemmKPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    template <typename DoPads, index_t MPerTile, index_t NPerTile>
+    static auto MakeEHGridDescriptor_M_N(index_t M, index_t N, index_t Stride)
+    {
+        // Only support row major for E and H
+        const auto grid_desc_m_n =
+            make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(Stride, I1));
+        return PadTensorDescriptor(grid_desc_m_n, make_tuple(MPerTile, NPerTile), DoPads{});
+    }
+
+    static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                                         const std::array<index_t, NumDTensor>& NRaws,
+                                         const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                static_assert(is_same<tensor_layout::gemm::RowMajor, DLayout>::value);
+
+                return DeviceOp::
+                    MakeEHGridDescriptor_M_N<Sequence<true, true>, GemmMPerBlock, GemmNPerBlock>(
+                        MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    template <typename DoPads, index_t MPerTile, index_t NPerTile>
+    static auto MakeMeanVarDescriptor_M_N(index_t M, index_t N)
+    {
+        const auto grid_desc_m_n =
+            make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(N, I1));
+        return PadTensorDescriptor(grid_desc_m_n, make_tuple(MPerTile, NPerTile), DoPads{});
+    }
+
+    template <typename DoPads, index_t MPerTile, index_t NPerTile>
+    static auto MakeCountDescriptor_M_N(index_t M, index_t N)
+    {
+        // We will broadcast [N] to [M, N] in this descriptor
+        // Hence, 1st stride is 0
+        const auto grid_desc_m_n =
+            make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I0, I1));
+        return PadTensorDescriptor(grid_desc_m_n, make_tuple(MPerTile, NPerTile), DoPads{});
+    }
+
+    template <index_t XPerTile>
+    static auto MakeDescriptor_X(index_t X)
+    {
+        const auto grid_desc_x = make_naive_tensor_descriptor_packed(make_tuple(X));
+        return PadTensorDescriptor(grid_desc_x, make_tuple(XPerTile), Sequence<true>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
+    // We have to separate mean var descriptor for gemm and layernorm bacause of different grid
+    // layout(different padding)
+    using GemmMeanVarGridDesc_M_NBlock =
+        decltype(MakeMeanVarDescriptor_M_N<Sequence<true, false>, GemmMPerBlock, GemmNPerBlock>(1,
+                                                                                                1));
+
+    using GemmCountGridDesc_M_NBlock =
+        decltype(MakeCountDescriptor_M_N<Sequence<true, false>, GemmMPerBlock, GemmNPerBlock>(1,
+                                                                                              1));
+
+    using LayernormMeanVarGridDesc_M_NBlock =
+        decltype(MakeMeanVarDescriptor_M_N<Sequence<true, true>,
+                                           LayernormBlockTileSize_M_N::At(0),
+                                           LayernormBlockTileSize_M_N::At(1)>(1, 1));
+
+    using LayernormCountGridDesc_M_NBlock =
+        decltype(MakeCountDescriptor_M_N<Sequence<true, true>,
+                                         LayernormBlockTileSize_M_N::At(0),
+                                         LayernormBlockTileSize_M_N::At(1)>(1, 1));
+
+    using GammaBetaGridDesc_N = decltype(MakeDescriptor_X<LayernormBlockTileSize_M_N::At(1)>(1));
+    using EHGridDesc_M_N = decltype(MakeEHGridDescriptor_M_N<Sequence<true, true>, 1, 1>(1, 1, 1));
+
+    using GridwiseGemmWelford = GridwiseGemmMultipleDWelfordFirstHalf_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EMeanVarDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_M_K,
+        BGridDesc_N_K,
+        DsGridDesc_M_N,
+        EHGridDesc_M_N,
+        GemmMeanVarGridDesc_M_NBlock,
+        GemmCountGridDesc_M_NBlock,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        GemmMPerBlock,
+        GemmNPerBlock,
+        GemmKPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        PostShuffleThreadClusterSize_M_N,
+        PostShuffleScalarPerVector,
+        LoopSched,
+        PipelineVer>;
+
+    using Block2ETileMap = typename GridwiseGemmWelford::DefaultBlock2ETileMap;
+
+    using GridwiseWelfordLayernorm =
+        GridwiseWelfordSecondHalfLayernorm2d<EMeanVarDataType,
+                                             HDataType,
+                                             GammaDataType,
+                                             BetaDataType,
+                                             AccDataType,
+                                             EHGridDesc_M_N,
+                                             LayernormMeanVarGridDesc_M_NBlock,
+                                             LayernormCountGridDesc_M_NBlock,
+                                             GammaBetaGridDesc_N,
+                                             HElementwiseOperation,
+                                             BlockSize,
+                                             LayernormThreadClusterSize_M_N::At(I0),
+                                             LayernormThreadClusterSize_M_N::At(I1),
+                                             LayernormThreadSliceSize_M,
+                                             LayernormThreadSliceSize_N,
+                                             LayernormESrcVectorSize,
+                                             LayernormHDstVectorSize,
+                                             LayernormGammaSrcVectorSize,
+                                             LayernormBetaSrcVectorSize>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 const void* p_gamma_grid,
+                 const void* p_beta_grid,
+                 void* p_h_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 std::array<index_t, NumDTensor> StrideDs,
+                 index_t StrideH,
+                 double epsilon,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op,
+                 HElementwiseOperation h_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_workspace_e_grid_{nullptr},
+              p_workspace_mean_{nullptr},
+              p_workspace_var_{nullptr},
+              p_workspace_count_{nullptr},
+              p_gamma_grid_{static_cast<const GammaDataType*>(p_gamma_grid)},
+              p_beta_grid_{static_cast<const BetaDataType*>(p_beta_grid)},
+              p_h_grid_{static_cast<HDataType*>(p_h_grid)},
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(MRaw, KRaw, StrideA)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(KRaw, NRaw, StrideB)},
+              ds_grid_desc_m_n_{},
+              gemm_e_grid_desc_m_n_{
+                  DeviceOp::MakeEHGridDescriptor_M_N<Sequence<true, true>,
+                                                     GemmMPerBlock,
+                                                     GemmNPerBlock>(MRaw, NRaw, StrideH)},
+              layernorm_e_grid_desc_m_n_{
+                  DeviceOp::MakeEHGridDescriptor_M_N<Sequence<true, true>,
+                                                     LayernormBlockTileSize_M_N::At(0),
+                                                     LayernormBlockTileSize_M_N::At(1)>(
+                      MRaw, NRaw, StrideH)},
+              gemm_mean_var_grid_desc_m_nblock_{},
+              gemm_count_grid_desc_m_nblock_{},
+              layernorm_mean_var_grid_desc_m_nblock_{},
+              layernorm_count_grid_desc_m_nblock_{},
+              gamma_grid_desc_n_{
+                  DeviceOp::MakeDescriptor_X<LayernormBlockTileSize_M_N::At(1)>(NRaw)},
+              beta_grid_desc_n_{
+                  DeviceOp::MakeDescriptor_X<LayernormBlockTileSize_M_N::At(1)>(NRaw)},
+              h_grid_desc_m_n_{
+                  DeviceOp::MakeEHGridDescriptor_M_N<Sequence<true, true>,
+                                                     LayernormBlockTileSize_M_N::At(0),
+                                                     LayernormBlockTileSize_M_N::At(1)>(
+                      MRaw, NRaw, StrideH)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemmWelford::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemmWelford::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              block_2_etile_map_{
+                  GridwiseGemmWelford::MakeDefaultBlock2ETileMap(gemm_e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              h_element_op_{h_element_op},
+              MRaw_{MRaw},
+              NRaw_{NRaw},
+              KRaw_{KRaw},
+              gemm_nblock_{math::integer_divide_ceil(NRaw, GemmNPerBlock)},
+              epsilon_{static_cast<AccDataType>(epsilon)}
+        {
+            // We don't need to pad in N dimension in gemm for mean/var/count. Set NPerTile 1.
+            gemm_mean_var_grid_desc_m_nblock_ =
+                DeviceOp::MakeMeanVarDescriptor_M_N<Sequence<true, false>, GemmMPerBlock, 1>(
+                    MRaw, gemm_nblock_);
+
+            gemm_count_grid_desc_m_nblock_ =
+                DeviceOp::MakeCountDescriptor_M_N<Sequence<true, false>, GemmMPerBlock, 1>(
+                    MRaw, gemm_nblock_);
+
+            layernorm_mean_var_grid_desc_m_nblock_ =
+                DeviceOp::MakeMeanVarDescriptor_M_N<Sequence<true, true>,
+                                                    LayernormBlockTileSize_M_N::At(0),
+                                                    LayernormBlockTileSize_M_N::At(1)>(
+                    MRaw, gemm_nblock_);
+
+            layernorm_count_grid_desc_m_nblock_ =
+                DeviceOp::MakeCountDescriptor_M_N<Sequence<true, true>,
+                                                  LayernormBlockTileSize_M_N::At(0),
+                                                  LayernormBlockTileSize_M_N::At(1)>(MRaw,
+                                                                                     gemm_nblock_);
+
+            // populate pointer, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) =
+                    DeviceOp::MakeEHGridDescriptor_M_N<Sequence<true, true>,
+                                                       GemmMPerBlock,
+                                                       GemmNPerBlock>(MRaw, NRaw, StrideDs[i]);
+            });
+
+            // populate desc for Ds/E/mean/var/count
+            if(GridwiseGemmWelford::CheckValidity(a_grid_desc_m_k_,
+                                                  b_grid_desc_n_k_,
+                                                  ds_grid_desc_m_n_,
+                                                  gemm_e_grid_desc_m_n_,
+                                                  block_2_etile_map_))
+            {
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemmWelford::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemmWelford::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        gemm_e_grid_desc_m_n_);
+
+                gemm_mean_var_grid_desc_mblock_mperblock_nblock_ =
+                    GridwiseGemmWelford::MakeMeanVarCountGridDescriptor_MBlock_MPerBlock_NBlock(
+                        gemm_mean_var_grid_desc_m_nblock_);
+
+                gemm_count_grid_desc_mblock_mperblock_nblock_ =
+                    GridwiseGemmWelford::MakeMeanVarCountGridDescriptor_MBlock_MPerBlock_NBlock(
+                        gemm_count_grid_desc_m_nblock_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << gemm_e_grid_desc_m_n_ << std::endl;
+            std::cout << "H[M, N]: " << h_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemmWelford::DsGridPointer p_ds_grid_;
+        void* p_workspace_e_grid_;
+        void* p_workspace_mean_;
+        void* p_workspace_var_;
+        void* p_workspace_count_;
+        const GammaDataType* p_gamma_grid_;
+        const BetaDataType* p_beta_grid_;
+        HDataType* p_h_grid_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EHGridDesc_M_N gemm_e_grid_desc_m_n_;
+        EHGridDesc_M_N layernorm_e_grid_desc_m_n_;
+        GemmMeanVarGridDesc_M_NBlock gemm_mean_var_grid_desc_m_nblock_;
+        GemmCountGridDesc_M_NBlock gemm_count_grid_desc_m_nblock_;
+        LayernormMeanVarGridDesc_M_NBlock layernorm_mean_var_grid_desc_m_nblock_;
+        LayernormCountGridDesc_M_NBlock layernorm_count_grid_desc_m_nblock_;
+        GammaBetaGridDesc_N gamma_grid_desc_n_;
+        GammaBetaGridDesc_N beta_grid_desc_n_;
+        EHGridDesc_M_N h_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        typename GridwiseGemmWelford::DefaultAGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        typename GridwiseGemmWelford::DefaultBGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        typename GridwiseGemmWelford::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemmWelford::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemmWelford::MeanVarGridDescriptor_MBlock_MPerBlock_NBlock
+            gemm_mean_var_grid_desc_mblock_mperblock_nblock_;
+        typename GridwiseGemmWelford::CountGridDescriptor_MBlock_MPerBlock_NBlock
+            gemm_count_grid_desc_mblock_mperblock_nblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+        HElementwiseOperation h_element_op_;
+
+        index_t MRaw_;
+        index_t NRaw_;
+        index_t KRaw_;
+        index_t gemm_nblock_;
+        AccDataType epsilon_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            float avg_time = 0;
+
+            if(!GridwiseGemmWelford::CheckValidity(arg.a_grid_desc_m_k_,
+                                                   arg.b_grid_desc_n_k_,
+                                                   arg.ds_grid_desc_m_n_,
+                                                   arg.gemm_e_grid_desc_m_n_,
+                                                   arg.block_2_etile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemmWelford has invalid setting");
+            }
+            if(arg.p_workspace_e_grid_ == nullptr || arg.p_workspace_mean_ == nullptr ||
+               arg.p_workspace_var_ == nullptr || arg.p_workspace_count_ == nullptr)
+                throw std::runtime_error("wrong! WorkSpace pointer has not been set");
+
+            index_t grid_size = arg.block_2_etile_map_.CalculateGridSize(arg.gemm_e_grid_desc_m_n_);
+
+            const auto M = arg.h_grid_desc_m_n_.GetLength(I0);
+            const auto N = arg.h_grid_desc_m_n_.GetLength(I1);
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel_gemm_welford =
+                    kernel_gemm_multiple_d_welford_first_half_xdl_cshuffle<
+                        GridwiseGemmWelford,
+                        ADataType, // TODO: distiguish A/B datatype
+                        typename GridwiseGemmWelford::DsGridPointer,
+                        EMeanVarDataType,
+                        AElementwiseOperation,
+                        BElementwiseOperation,
+                        CDEElementwiseOperation,
+                        typename GridwiseGemmWelford::DefaultAGridDesc_AK0_M_AK1,
+                        typename GridwiseGemmWelford::DefaultBGridDesc_BK0_N_BK1,
+                        typename GridwiseGemmWelford::
+                            DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                        typename GridwiseGemmWelford::
+                            EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                        typename GridwiseGemmWelford::MeanVarGridDescriptor_MBlock_MPerBlock_NBlock,
+                        typename GridwiseGemmWelford::CountGridDescriptor_MBlock_MPerBlock_NBlock,
+                        typename GridwiseGemmWelford::DefaultBlock2ETileMap,
+                        has_main_loop>;
+
+                const auto kernel_welford_layernorm =
+                    kernel_welford_layernorm2d_second_half<GridwiseWelfordLayernorm,
+                                                           EMeanVarDataType,
+                                                           HDataType,
+                                                           GammaDataType,
+                                                           BetaDataType,
+                                                           AccDataType,
+                                                           EHGridDesc_M_N,
+                                                           LayernormMeanVarGridDesc_M_NBlock,
+                                                           LayernormCountGridDesc_M_NBlock,
+                                                           GammaBetaGridDesc_N,
+                                                           HElementwiseOperation>;
+
+                avg_time +=
+                    launch_and_time_kernel(stream_config,
+                                           kernel_gemm_welford,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_ds_grid_,
+                                           static_cast<EMeanVarDataType*>(arg.p_workspace_e_grid_),
+                                           static_cast<EMeanVarDataType*>(arg.p_workspace_mean_),
+                                           static_cast<EMeanVarDataType*>(arg.p_workspace_var_),
+                                           static_cast<int32_t*>(arg.p_workspace_count_),
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.cde_element_op_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.gemm_mean_var_grid_desc_mblock_mperblock_nblock_,
+                                           arg.gemm_count_grid_desc_mblock_mperblock_nblock_,
+                                           arg.block_2_etile_map_,
+                                           arg.NRaw_);
+
+                index_t MBlockClusterLength =
+                    math::integer_divide_ceil(M, LayernormBlockTileSize_M_N::At(0));
+                index_t NBlockClusterLength =
+                    math::integer_divide_ceil(N, LayernormBlockTileSize_M_N::At(1));
+                grid_size = MBlockClusterLength * NBlockClusterLength;
+
+                index_t numMeanVarCountBlockTileIteration_N = math::integer_divide_ceil(
+                    arg.gemm_nblock_, LayernormThreadClusterSize_M_N::At(I1));
+
+                avg_time += launch_and_time_kernel(
+                    stream_config,
+                    kernel_welford_layernorm,
+                    dim3(grid_size),
+                    dim3(BlockSize),
+                    0,
+                    static_cast<EMeanVarDataType*>(arg.p_workspace_e_grid_),
+                    static_cast<const EMeanVarDataType*>(arg.p_workspace_mean_),
+                    static_cast<const EMeanVarDataType*>(arg.p_workspace_var_),
+                    static_cast<const int32_t*>(arg.p_workspace_count_),
+                    arg.p_gamma_grid_,
+                    arg.p_beta_grid_,
+                    arg.p_h_grid_,
+                    arg.layernorm_e_grid_desc_m_n_,
+                    arg.h_grid_desc_m_n_,
+                    arg.layernorm_mean_var_grid_desc_m_nblock_,
+                    arg.layernorm_count_grid_desc_m_nblock_,
+                    arg.gamma_grid_desc_n_,
+                    arg.beta_grid_desc_n_,
+                    numMeanVarCountBlockTileIteration_N,
+                    NBlockClusterLength,
+                    arg.epsilon_,
+                    arg.h_element_op_);
+
+                return avg_time;
+            };
+
+            if(GridwiseGemmWelford::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    size_t GetWorkSpaceSize(const BaseArgument* pArg) const override
+    {
+        const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
+
+        size_t workspace_size = 0;
+
+        int gemm_welford_size = pArg_->MRaw_ * pArg_->gemm_nblock_;
+
+        // workspace for welford intermediate mean
+        workspace_size += gemm_welford_size * sizeof(EMeanVarDataType) + 64;
+
+        // workspace for welford intermediate variance
+        workspace_size += gemm_welford_size * sizeof(EMeanVarDataType) + 64;
+
+        // workspace for welford intermediate count
+        workspace_size += pArg_->gemm_nblock_ * sizeof(int32_t) + 64;
+
+        if constexpr(!is_same_v<EMeanVarDataType, HDataType>)
+            workspace_size += pArg_->MRaw_ * pArg_->NRaw_ * sizeof(EMeanVarDataType);
+
+        return (workspace_size);
+    };
+
+    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    {
+        Argument* pArg_ = dynamic_cast<Argument*>(pArg);
+
+        pArg_->p_workspace_ = p_workspace;
+
+        int gemm_welford_size = pArg_->MRaw_ * pArg_->gemm_nblock_;
+
+        // setup buffer used for intermediate welford mean
+        pArg_->p_workspace_mean_ = static_cast<char*>(pArg_->p_workspace_);
+
+        index_t mean_space_sz = gemm_welford_size * sizeof(EMeanVarDataType);
+        mean_space_sz         = math::integer_least_multiple(mean_space_sz, 64);
+
+        // setup buffer used for intermediate welford varirance
+        pArg_->p_workspace_var_ = reinterpret_cast<char*>(pArg_->p_workspace_mean_) + mean_space_sz;
+
+        index_t variance_space_sz = gemm_welford_size * sizeof(EMeanVarDataType);
+        variance_space_sz         = math::integer_least_multiple(variance_space_sz, 64);
+
+        // setup buffer used for intermediate welford count
+        pArg_->p_workspace_count_ =
+            reinterpret_cast<char*>(pArg_->p_workspace_var_) + variance_space_sz;
+
+        index_t count_space_sz = gemm_welford_size * sizeof(int32_t);
+        count_space_sz         = math::integer_least_multiple(count_space_sz, 64);
+
+        if constexpr(!is_same_v<EMeanVarDataType, HDataType>)
+            pArg_->p_workspace_e_grid_ =
+                reinterpret_cast<char*>(pArg_->p_workspace_count_) + count_space_sz;
+        else
+            pArg_->p_workspace_e_grid_ = static_cast<void*>(pArg_->p_h_grid_);
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        // check vector load/store
+        {
+            using Row = ck::tensor_layout::gemm::RowMajor;
+            using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+            // check vector load of A
+            if constexpr(is_same_v<ALayout, Row> && ABlockTransferSrcVectorDim == 2)
+            {
+                if(arg.KRaw_ % ABlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else if constexpr(is_same_v<ALayout, Col> && ABlockTransferSrcVectorDim == 1)
+            {
+                // FIXME: not rigorous
+                if(arg.MRaw_ % ABlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+
+            // check vector laod of B
+            if constexpr(is_same_v<BLayout, Col> && BBlockTransferSrcVectorDim == 2)
+            {
+                if(arg.KRaw_ % BBlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else if constexpr(is_same_v<BLayout, Row> && BBlockTransferSrcVectorDim == 1)
+            {
+                // FIXME: not rigorous
+                if(arg.NRaw_ % BBlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+
+            // check vector load of Ds
+            // only support RowMajor for now
+            bool all_valid = true;
+
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                if constexpr(!is_same_v<DLayout, Row>)
+                {
+                    all_valid = false;
+                }
+            });
+
+            if(!all_valid)
+            {
+                return false;
+            }
+
+            // check vector store of E
+            // E and H only support RowMajor for now
+            if constexpr(is_same_v<ELayout, Row> && is_same_v<HLayout, Row>)
+            {
+                if(arg.NRaw_ % PostShuffleScalarPerVector != 0 ||
+                   arg.NRaw_ % LayernormGammaSrcVectorSize != 0 ||
+                   arg.NRaw_ % LayernormBetaSrcVectorSize != 0 ||
+                   arg.NRaw_ % LayernormHDstVectorSize != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+        }
+
+        return GridwiseGemmWelford::CheckValidity(arg.a_grid_desc_m_k_,
+                                                  arg.b_grid_desc_n_k_,
+                                                  arg.ds_grid_desc_m_n_,
+                                                  arg.gemm_e_grid_desc_m_n_,
+                                                  arg.block_2_etile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             std::array<const void*, NumDTensor> p_ds,
+                             const void* p_gamma,
+                             const void* p_beta,
+                             void* p_h,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t StrideA,
+                             index_t StrideB,
+                             std::array<index_t, NumDTensor> StrideDs,
+                             index_t StrideH,
+                             double epsilon,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op,
+                             HElementwiseOperation h_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_gamma,
+                        p_beta,
+                        p_h,
+                        MRaw,
+                        NRaw,
+                        KRaw,
+                        StrideA,
+                        StrideB,
+                        StrideDs,
+                        StrideH,
+                        epsilon,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op,
+                        h_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      std::array<const void*, NumDTensor> p_ds,
+                                                      const void* p_gamma,
+                                                      const void* p_beta,
+                                                      void* p_h,
+                                                      index_t MRaw,
+                                                      index_t NRaw,
+                                                      index_t KRaw,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      std::array<index_t, NumDTensor> StrideDs,
+                                                      index_t StrideH,
+                                                      double epsilon,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CDEElementwiseOperation cde_element_op,
+                                                      HElementwiseOperation h_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_gamma,
+                                          p_beta,
+                                          p_h,
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          StrideA,
+                                          StrideB,
+                                          StrideDs,
+                                          StrideH,
+                                          epsilon,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op,
+                                          h_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
+        // clang-format off
+        str << "DeviceGemmMultipleDLayernorm_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << GemmMPerBlock << ", "
+            << GemmNPerBlock << ", "
+            << GemmKPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << PostShuffleThreadClusterSize_M_N::At(I0) << ", "
+            << PostShuffleThreadClusterSize_M_N::At(I1) << ", "
+            << LayernormThreadClusterSize_M_N::At(I0) << ", "
+            << LayernormThreadClusterSize_M_N::At(I1) << ", "
+            << LayernormThreadSliceSize_M
+            << ">"
+            << " LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
+        // clang-format on
+
+        return str.str();
+    }
+}; // namespace device
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..916f29a904
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
@@ -0,0 +1,693 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatDsPointer,
+          typename FloatE,
+          typename FloatRsPointer,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename RsGridDescriptor_MBlock_MPerBlock,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_multiple_d_multiple_r_xdl_cshuffle(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatDsPointer p_ds_grid,
+            FloatE* __restrict__ p_e_grid,
+            FloatRsPointer p_rs_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const QsElementwiseOperation qs_element_op,
+            const RsElementwiseOperation rs_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const RsGridDescriptor_MBlock_MPerBlock rs_grid_desc_mblock_mperblock,
+            const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_ds_grid,
+                                                  p_e_grid,
+                                                  p_rs_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  qs_element_op,
+                                                  rs_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  rs_grid_desc_mblock_mperblock,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = p_rs_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = qs_element_op;
+    ignore = rs_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = rs_grid_desc_mblock_mperblock;
+    ignore = block_2_etile_map;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// GEMM:
+//   input : A[AK0, M, AK1]
+//   input : B[AK0, N, AK1]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   output : R0[M], R1[M], ...
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+//   Q0 = reduce0(q_op0(E)), Q1 = reduce1(q_op0(E)), ...
+//   R0 = r_op0(Q0), R1 = r_op1(Q1), ...
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename ADataType,
+          typename BDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ReduceAccDataType,
+          typename RsDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation,
+          typename ThreadReduceOperations,
+          typename RsGlobalMemoryDataOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDRThreadTransferClusterLengths_MPerBlock_NPerBlock,
+          index_t CDEReduceThreadTransferScalarPerVector_NPerBlock,
+          index_t RThreadTransferDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGemmMultipleDMultipleR_Xdl_CShuffle
+    : public DeviceGemmMultipleDMultipleR<ALayout,
+                                          BLayout,
+                                          DELayout,
+                                          ADataType,
+                                          BDataType,
+                                          DsDataType,
+                                          EDataType,
+                                          RsDataType,
+                                          AElementwiseOperation,
+                                          BElementwiseOperation,
+                                          CDEElementwiseOperation,
+                                          QsElementwiseOperation,
+                                          RsElementwiseOperation>
+{
+    using DeviceOp = DeviceGemmMultipleDMultipleR_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+    static constexpr index_t NumRTensor = RsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, DELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, DELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    // assume D is packed tensor
+    static auto MakeRGridDescriptor_M(index_t MRaw)
+    {
+        const auto r_grid_desc_mraw = make_naive_tensor_descriptor_packed(make_tuple(MRaw));
+
+        const auto M    = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto MPad = M - MRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                     GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M
+            return transform_tensor_descriptor(r_grid_desc_mraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad)),
+                                               make_tuple(Sequence<0>{}),
+                                               make_tuple(Sequence<0>{}));
+        }
+        else
+        {
+            // not pad M
+            return r_grid_desc_mraw;
+        }
+    }
+
+    using AGridDesc_M_K = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N(1, 1, 1));
+    using RGridDesc_M   = decltype(MakeRGridDescriptor_M(1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleDMultipleR_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        ReduceAccDataType,
+        RsDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        QsElementwiseOperation,
+        RsElementwiseOperation,
+        ThreadReduceOperations,
+        InMemoryDataOperationEnum::Set,
+        RsGlobalMemoryDataOperation,
+        AGridDesc_M_K,
+        BGridDesc_N_K,
+        EGridDesc_M_N,
+        RGridDesc_M,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDRThreadTransferClusterLengths_MPerBlock_NPerBlock,
+        CDEReduceThreadTransferScalarPerVector_NPerBlock,
+        RThreadTransferDstScalarPerVector_MPerBlock,
+        LoopSched>;
+
+    using AGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
+            AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
+            BGridDesc_N_K{}))>;
+
+    using Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 std::array<void*, NumRTensor> p_rs_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 std::array<index_t, NumDTensor> StrideDs,
+                 index_t StrideE,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op,
+                 QsElementwiseOperation qs_element_op,
+                 RsElementwiseOperation rs_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{}, // FIXME
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              p_rs_grid_{}, // FIXME
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(MRaw, KRaw, StrideA)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(KRaw, NRaw, StrideB)},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N(MRaw, NRaw, StrideE)},
+              r_grid_desc_m_{DeviceOp::MakeRGridDescriptor_M(MRaw)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              rs_grid_desc_mblock_mperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              qs_element_op_{qs_element_op},
+              rs_element_op_{rs_element_op}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           e_grid_desc_m_n_,
+                                           r_grid_desc_m_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+
+                static_for<0, NumDTensor, 1>{}([&](auto i) {
+                    using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                    p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                    const auto d_grid_desc_m_n =
+                        DeviceOp::MakeEGridDescriptor_M_N(MRaw, NRaw, StrideDs[i]);
+
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock_(i) =
+                        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                            d_grid_desc_m_n);
+                });
+
+                static_for<0, NumRTensor, 1>{}([&](auto i) {
+                    using RDataType = remove_cvref_t<tuple_element_t<i.value, RsDataType>>;
+
+                    p_rs_grid_(i) = static_cast<RDataType*>(p_rs_grid[i]);
+
+                    rs_grid_desc_mblock_mperblock_(i) =
+                        GridwiseGemm::MakeRGridDescriptor_MBlock_MPerBlock(r_grid_desc_m_);
+                });
+            }
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+        typename GridwiseGemm::RsGridPointer p_rs_grid_;
+
+        // tensor descriptors
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+        RGridDesc_M r_grid_desc_m_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        StaticallyIndexedArray<
+            typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+            NumDTensor>
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_; // FIXME: Ds desc may be of different
+                                                             // type from E
+        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        StaticallyIndexedArray<typename GridwiseGemm::RGridDescriptor_MBlock_MPerBlock, NumRTensor>
+            rs_grid_desc_mblock_mperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+        QsElementwiseOperation qs_element_op_;
+        RsElementwiseOperation rs_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.r_grid_desc_m_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_gemm_multiple_d_multiple_r_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    typename GridwiseGemm::RsGridPointer,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    QsElementwiseOperation,
+                    RsElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    ck::StaticallyIndexedArray<
+                        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                        NumDTensor>,
+                    typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    ck::StaticallyIndexedArray<
+                        typename GridwiseGemm::RGridDescriptor_MBlock_MPerBlock,
+                        NumRTensor>,
+                    typename GridwiseGemm::DefaultBlock2ETileMap,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.p_rs_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.qs_element_op_,
+                                              arg.rs_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.rs_grid_desc_mblock_mperblock_,
+                                              arg.block_2_etile_map_);
+            };
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           arg.e_grid_desc_m_n_,
+                                           arg.r_grid_desc_m_,
+                                           arg.block_2_etile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             std::array<const void*, NumDTensor> p_ds,
+                             void* p_e,
+                             std::array<void*, NumRTensor> p_rs,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t StrideA,
+                             index_t StrideB,
+                             std::array<index_t, NumDTensor> StrideDs,
+                             index_t StrideE,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op,
+                             QsElementwiseOperation qs_element_op,
+                             RsElementwiseOperation rs_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        p_rs,
+                        MRaw,
+                        NRaw,
+                        KRaw,
+                        StrideA,
+                        StrideB,
+                        StrideDs,
+                        StrideE,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op,
+                        qs_element_op,
+                        rs_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      std::array<const void*, NumDTensor> p_ds,
+                                                      void* p_e,
+                                                      std::array<void*, NumRTensor> p_rs,
+                                                      index_t MRaw,
+                                                      index_t NRaw,
+                                                      index_t KRaw,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      std::array<index_t, NumDTensor> StrideDs,
+                                                      index_t StrideE,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CDEElementwiseOperation cde_element_op,
+                                                      QsElementwiseOperation qs_element_op,
+                                                      RsElementwiseOperation rs_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          p_rs,
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          StrideA,
+                                          StrideB,
+                                          StrideDs,
+                                          StrideE,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op,
+                                          qs_element_op,
+                                          rs_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmMultipleDMultipleR_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp
new file mode 100644
index 0000000000..44b3518e2c
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp
@@ -0,0 +1,706 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerWMMA,
+          ck::index_t NPerWMMA,
+          ck::index_t MRepeat,
+          ck::index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          typename CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          ck::index_t NumPrefetch         = 1,
+          ck::LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>
+struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
+                                                                      BLayout,
+                                                                      DsLayout,
+                                                                      ELayout,
+                                                                      ADataType,
+                                                                      BDataType,
+                                                                      DsDataType,
+                                                                      EDataType,
+                                                                      AElementwiseOperation,
+                                                                      BElementwiseOperation,
+                                                                      CDEElementwiseOperation>
+{
+    using DeviceOp                      = DeviceGemmMultipleD_Wmma_CShuffle;
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    // K1 = Max Vector Access Pixels
+    static constexpr auto K1Number = Number<K1>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock* K1};
+
+    static auto MakeAGridDescriptor_K0_M_K1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto a_grid_desc_m_k = matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        const auto M               = a_grid_desc_m_k.GetLength(I0);
+        const auto K               = a_grid_desc_m_k.GetLength(I1);
+        assert(K % K1 == 0);
+        const index_t K0 = K / K1;
+
+        return transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                       make_pass_through_transform(M)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    static auto MakeBGridDescriptor_K0_N_K1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto b_grid_desc_n_k = matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        const auto N               = b_grid_desc_n_k.GetLength(I0);
+        const auto K               = b_grid_desc_n_k.GetLength(I1);
+        assert(K % K1 == 0);
+        const index_t K0 = K / K1;
+
+        return transform_tensor_descriptor(
+            b_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                       make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    template <typename ELayout_>
+    static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout_>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout_>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& Ms,
+                                         const std::array<index_t, NumDTensor>& Ns,
+                                         const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(Ms[i], Ns[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // Gridwise descriptor, mapping to whole given provblem.
+    using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
+    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
+    using DsGridDesc_M_N    = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
+    using EGridDesc_M_N     = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
+
+    // GridwiseOp
+    using GridwiseOp = GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle<
+        // DataType Family
+        ADataType,
+        BDataType,
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        // InMemory Data Descriptor
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        DsGridDesc_M_N,
+        EGridDesc_M_N,
+        // ElementwiseOp Family
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        // Tiling Family
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerWMMA,
+        NPerWMMA,
+        K1,
+        MRepeat,
+        NRepeat,
+        // ThreadCluster Family
+        BlockSize,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMRepeatPerShuffle,
+        CShuffleNRepeatPerShuffle,
+        CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+        NumPrefetch,
+        LoopSched,
+        PipelineVer>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 index_t M,
+                 index_t N,
+                 index_t K,
+                 index_t StrideA,
+                 index_t StrideB,
+                 std::array<index_t, NumDTensor> StrideDs,
+                 index_t StrideE,
+                 index_t M01,
+                 index_t N01,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              MRaw_{M},
+              NRaw_{N},
+              KRaw_{K}
+        {
+            a_grid_desc_k0_m_k1_ = DeviceOp::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
+            b_grid_desc_k0_n_k1_ = DeviceOp::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) =
+                    DeviceOp::MakeEGridDescriptor_M_N<DLayout>(M, N, StrideDs[i]);
+            });
+            e_grid_desc_m_n_ = DeviceOp::MakeEGridDescriptor_M_N<ELayout>(M, N, StrideE);
+
+            block_2_ctile_map_ = GridwiseOp::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_, M01, N01);
+
+            if(GridwiseOp::CheckValidity(a_grid_desc_k0_m_k1_,
+                                         b_grid_desc_k0_n_k1_,
+                                         ds_grid_desc_m_n_,
+                                         e_grid_desc_m_n_,
+                                         block_2_ctile_map_))
+            {
+                ds_grid_desc_mblock_mperblock_nblock_nperblock =
+                    GridwiseOp::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+
+                e_grid_desc_mblock_mperblock_nblock_nperblock =
+                    GridwiseOp::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+            }
+        }
+
+        // Pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseOp::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // Tensor Descriptors
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+        typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock;
+        typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock;
+
+        // Block to Tile mapping
+        typename GridwiseOp::DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // Idle
+        index_t M01_;
+        index_t N01_;
+
+        // ElementwiseOp
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // for checking vector load/store
+        index_t MRaw_;
+        index_t NRaw_;
+        index_t KRaw_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if 0
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) 
+                          << ", " << arg.c_grid_desc_m_n_.GetLength(I1) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I2) << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseOp::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                          arg.b_grid_desc_k0_n_k1_,
+                                          arg.ds_grid_desc_m_n_,
+                                          arg.e_grid_desc_m_n_,
+                                          arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_m0nm1_wmma_v1r1 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            if(GridwiseOp::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_mupltipe_d_wmma_cshuffle<
+                    GridwiseOp,
+                    ADataType,
+                    BDataType,
+                    typename GridwiseOp::DsGridPointer,
+                    EDataType,
+                    remove_reference_t<typename DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<typename DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
+                    remove_reference_t<
+                        typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    remove_reference_t<typename GridwiseOp::DefaultBlock2CTileMap>,
+                    true>; // Last Option is W/O
+
+                ave_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_ds_grid_,
+                                           arg.p_e_grid_,
+                                           arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                           arg.e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.cde_element_op_,
+                                           arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_mupltipe_d_wmma_cshuffle<
+                    GridwiseOp,
+                    ADataType,
+                    BDataType,
+                    typename GridwiseOp::DsGridPointer,
+                    EDataType,
+                    remove_reference_t<typename DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<typename DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
+                    remove_reference_t<
+                        typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    remove_reference_t<typename GridwiseOp::DefaultBlock2CTileMap>,
+                    false>;
+
+                ave_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_ds_grid_,
+                                           arg.p_e_grid_,
+                                           arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                           arg.e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.cde_element_op_,
+                                           arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
+           ck::get_device_name() == "gfx1102")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+        // check vector load/store
+        {
+            using Row = ck::tensor_layout::gemm::RowMajor;
+            using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+            // check vector load of A
+            if constexpr(is_same_v<ALayout, Row> && ABlockTransferSrcVectorDim == 2)
+            {
+                if(arg.KRaw_ % ABlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else if constexpr(is_same_v<ALayout, Col> && ABlockTransferSrcVectorDim == 1)
+            {
+                // FIXME: not rigorous
+                if(arg.MRaw_ % ABlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+
+            // check vector laod of B
+            if constexpr(is_same_v<BLayout, Col> && BBlockTransferSrcVectorDim == 2)
+            {
+                if(arg.KRaw_ % BBlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else if constexpr(is_same_v<BLayout, Row> && BBlockTransferSrcVectorDim == 1)
+            {
+                // FIXME: not rigorous
+                if(arg.NRaw_ % BBlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+
+            // check vector load of Ds
+            // only support RowMajor for now
+            bool all_valid = true;
+
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                if constexpr(!is_same_v<DLayout, Row>)
+                {
+                    all_valid = false;
+                }
+            });
+
+            if(!all_valid)
+            {
+                return false;
+            }
+
+            // check vector store of E
+            // only support RowMajor for now
+            if constexpr(is_same_v<ELayout, Row>)
+            {
+                if(arg.NRaw_ % CDEShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+        }
+
+        return GridwiseOp::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                         arg.b_grid_desc_k0_n_k1_,
+                                         arg.ds_grid_desc_m_n_,
+                                         arg.e_grid_desc_m_n_,
+                                         arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             std::array<const void*, NumDTensor> p_ds,
+                             void* p_e,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             std::array<ck::index_t, NumDTensor> StrideDs,
+                             index_t StrideE,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideDs,
+                        StrideE,
+                        1,
+                        1,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        index_t M,
+                        index_t N,
+                        index_t K,
+                        index_t StrideA,
+                        index_t StrideB,
+                        std::array<ck::index_t, NumDTensor> StrideDs,
+                        index_t StrideE,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideDs,
+                                          StrideE,
+                                          1,
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
+        // clang-format off
+        str << "DeviceGemmMultipleD_Wmma_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerWMMA << ", "
+            << NPerWMMA << ", "
+            << MRepeat << ", "
+            << NRepeat
+            << ">"
+            << " NumPrefetch: "
+            << NumPrefetch << ", "
+            << "LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..c90c28f5a8
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,716 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename ADataType,
+          typename BDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_multiple_d_xdl_cshuffle(const ADataType* __restrict__ p_a_grid,
+                                            const BDataType* __restrict__ p_b_grid,
+                                            DsPointer p_ds_grid,
+                                            EDataType* __restrict__ p_e_grid,
+                                            const AElementwiseOperation a_element_op,
+                                            const BElementwiseOperation b_element_op,
+                                            const CDEElementwiseOperation cde_element_op,
+                                            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+                                            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+                                            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                                                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                                                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                            const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_ds_grid,
+                                                  p_e_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_etile_map;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
+                                                                     BLayout,
+                                                                     DsLayout,
+                                                                     ELayout,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     DsDataType,
+                                                                     EDataType,
+                                                                     AElementwiseOperation,
+                                                                     BElementwiseOperation,
+                                                                     CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGemmMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    template <typename ELay>
+    static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                                         const std::array<index_t, NumDTensor>& NRaws,
+                                         const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // desc for problem definition
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
+
+    using ComputeDataType = EDataType;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeDataType,
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched,
+        PipelineVer>;
+
+    // desc for blockwise copy
+    using AGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
+            AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
+            BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
+        decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+
+    // block-to-e-tile map
+    using Block2ETileMap =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 std::array<index_t, NumDTensor> StrideDs,
+                 index_t StrideE,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(MRaw, KRaw, StrideA)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(KRaw, NRaw, StrideB)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(MRaw, NRaw, StrideE)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              MRaw_{MRaw},
+              NRaw_{NRaw},
+              KRaw_{KRaw}
+        {
+            // populate pointer, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) =
+                    DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaw, NRaw, StrideDs[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           ds_grid_desc_m_n_,
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // for checking vector load/store
+        index_t MRaw_;
+        index_t NRaw_;
+        index_t KRaw_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.ds_grid_desc_m_n_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_gemm_multiple_d_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    BDataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    DeviceOp::Block2ETileMap,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_etile_map_);
+            };
+
+            const auto K = arg.a_grid_desc_m_k_.GetLength(I1);
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        // check vector load/store
+        {
+            using Row = ck::tensor_layout::gemm::RowMajor;
+            using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+            // check vector load of A
+            if constexpr(is_same_v<ALayout, Row> && ABlockTransferSrcVectorDim == 2)
+            {
+                if(arg.KRaw_ % ABlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else if constexpr(is_same_v<ALayout, Col> && ABlockTransferSrcVectorDim == 1)
+            {
+                // FIXME: not rigorous
+                if(arg.MRaw_ % ABlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+
+            // check vector laod of B
+            if constexpr(is_same_v<BLayout, Col> && BBlockTransferSrcVectorDim == 2)
+            {
+                if(arg.KRaw_ % BBlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else if constexpr(is_same_v<BLayout, Row> && BBlockTransferSrcVectorDim == 1)
+            {
+                // FIXME: not rigorous
+                if(arg.NRaw_ % BBlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+
+            // check vector load of Ds
+            // only support RowMajor for now
+            bool all_valid = true;
+
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                if constexpr(!is_same_v<DLayout, Row>)
+                {
+                    all_valid = false;
+                }
+            });
+
+            if(!all_valid)
+            {
+                return false;
+            }
+
+            // check vector store of E
+            // only support RowMajor for now
+            if constexpr(is_same_v<ELayout, Row>)
+            {
+                if(arg.NRaw_ % CDEBlockTransferScalarPerVector_NPerBlock != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           arg.ds_grid_desc_m_n_,
+                                           arg.e_grid_desc_m_n_,
+                                           arg.block_2_etile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             std::array<const void*, NumDTensor> p_ds,
+                             void* p_e,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t StrideA,
+                             index_t StrideB,
+                             std::array<index_t, NumDTensor> StrideDs,
+                             index_t StrideE,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        MRaw,
+                        NRaw,
+                        KRaw,
+                        StrideA,
+                        StrideB,
+                        StrideDs,
+                        StrideE,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        index_t MRaw,
+                        index_t NRaw,
+                        index_t KRaw,
+                        index_t StrideA,
+                        index_t StrideB,
+                        std::array<ck::index_t, NumDTensor> StrideDs,
+                        index_t StrideE,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          StrideA,
+                                          StrideB,
+                                          StrideDs,
+                                          StrideE,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
+        // clang-format off
+        str << "DeviceGemmMultipleD_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle << ", "
+            << getGemmSpecializationString(GemmSpec)
+            << ">"
+            << " LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..b590fba98f
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp
@@ -0,0 +1,848 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Note: inter-wave loop scheduler is rolled out to c-shuffle version first. Becuase non c-shuffle
+// version currently has compiler issues with register spill which further causes validation
+// failures.
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename ReduceAccDataType,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename ReduceOperations,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          typename ReduceGlobalMemoryDataOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+          index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+          index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<0, ReduceOperations::Size()>
+{
+    using DeviceOp = DeviceGemmReduce_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(M)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_right_pad_transform(MRaw, MPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto NPad = N - NRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(NRaw, NPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(N)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_right_pad_transform(NRaw, NPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideC));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto NPad = N - NRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(NRaw, NPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_grid_desc_mraw_nraw;
+        }
+    }
+
+    // assume Reduce is packed tensor
+    static auto MakeReduceGridDescriptor_M(index_t MRaw)
+    {
+        const auto d_grid_desc_mraw = make_naive_tensor_descriptor_packed(make_tuple(MRaw));
+
+        const auto M    = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto MPad = M - MRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                     GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M
+            return transform_tensor_descriptor(d_grid_desc_mraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad)),
+                                               make_tuple(Sequence<0>{}),
+                                               make_tuple(Sequence<0>{}));
+        }
+        else
+        {
+            // not pad M
+            return d_grid_desc_mraw;
+        }
+    }
+
+    using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
+    using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
+    using CGridDesc_M_N       = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+    using ReduceGridDesc_M    = decltype(MakeReduceGridDescriptor_M(1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        ReduceAccDataType,
+        ReducePtrsGlobal,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        ReduceOperations,
+        ReduceInElementwiseOperations,
+        ReduceAccElementwiseOperations,
+        InMemoryDataOperationEnum::Set,
+        ReduceGlobalMemoryDataOperation,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        ReduceGridDesc_M,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+        CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+        CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+        LoopSched>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 ReducePtrsGlobal p_reduces_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op,
+                 ReduceInElementwiseOperations reduce_in_element_ops,
+                 ReduceAccElementwiseOperations reduce_out_element_ops)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              p_reduces_grid_{p_reduces_grid},
+              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, NRaw, StrideC)},
+              reduce_grid_desc_m_{DeviceOp::MakeReduceGridDescriptor_M(MRaw)},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              reduce_grid_desc_mblock_mperblock_{},
+              block_2_ctile_map_{GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op},
+              reduce_in_element_ops_{reduce_in_element_ops},
+              reduce_out_element_ops_{reduce_out_element_ops}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+
+                reduce_grid_desc_mblock_mperblock_ =
+                    GridwiseGemm::MakeReduceGridDescriptor_MBlock_MPerBlock(reduce_grid_desc_m_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        ReducePtrsGlobal p_reduces_grid_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        ReduceGridDesc_M reduce_grid_desc_m_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock
+            reduce_grid_desc_mblock_mperblock_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+        ReduceInElementwiseOperations reduce_in_element_ops_;
+        ReduceAccElementwiseOperations reduce_out_element_ops_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if DEBUG_LOG
+            {
+                std::cout << "arg.a_grid_desc_ak0_m_ak1_{"
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_bk0_n_bk1_{"
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I0) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+
+                std::cout << "arg.reduce_grid_desc_m_{ " << arg.reduce_grid_desc_m_.GetLength(I0)
+                          << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                            arg.b_grid_desc_bk0_n_bk1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float elapsed_time = 0.0f;
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_reduce_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    ReducePtrsGlobal,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    ReduceInElementwiseOperations,
+                    ReduceAccElementwiseOperations,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    true>;
+
+                elapsed_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_reduces_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.reduce_in_element_ops_,
+                                           arg.reduce_out_element_ops_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.reduce_grid_desc_mblock_mperblock_,
+                                           arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_reduce_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    ReducePtrsGlobal,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    ReduceInElementwiseOperations,
+                    ReduceAccElementwiseOperations,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::ReduceGridDescriptor_MBlock_MPerBlock,
+                    typename GridwiseGemm::DefaultBlock2CTileMap,
+                    false>;
+
+                elapsed_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_reduces_grid_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.c_element_op_,
+                                           arg.reduce_in_element_ops_,
+                                           arg.reduce_out_element_ops_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.reduce_grid_desc_mblock_mperblock_,
+                                           arg.block_2_ctile_map_);
+            }
+
+            return elapsed_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static constexpr int NumReduce = ReduceOperations::Size();
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             const void* p_bias,
+                             std::array<const void*, 0> p_ds,
+                             void* p_c,
+                             std::array<void*, NumReduce> p_reduces,
+                             ck::index_t M,
+                             ck::index_t N,
+                             ck::index_t K,
+                             ck::index_t StrideA,
+                             ck::index_t StrideB,
+                             ck::index_t StrideC,
+                             std::array<ck::index_t, 0> StrideDs,
+                             std::array<void*, 3> gemm_element_ops,
+                             std::array<void*, 0> d_element_ops,
+                             std::array<void*, NumReduce> reduce_in_element_op,
+                             std::array<void*, NumReduce> reduce_out_element_op)
+    {
+        (void)p_bias;
+        (void)p_ds;
+        (void)StrideDs;
+        (void)d_element_ops;
+
+        ReducePtrsGlobal reduce_tuple = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReducePtrsGlobal{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return static_cast<T*>(p_reduces[I]);
+            },
+            Number<NumReduce>{});
+
+        ReduceInElementwiseOperations reduce_in_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceInElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_in_element_op[I]));
+            },
+            Number<NumReduce>{});
+        ReduceAccElementwiseOperations reduce_out_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceAccElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_out_element_op[I]));
+            },
+            Number<NumReduce>{});
+
+        AElementwiseOperation a_element_op =
+            *(static_cast<AElementwiseOperation*>(gemm_element_ops[0]));
+        BElementwiseOperation b_element_op =
+            *(static_cast<BElementwiseOperation*>(gemm_element_ops[1]));
+        CElementwiseOperation c_element_op =
+            *(static_cast<CElementwiseOperation*>(gemm_element_ops[2]));
+
+        return Argument{static_cast<const ADataType*>(p_a),
+                        static_cast<const BDataType*>(p_b),
+                        static_cast<CDataType*>(p_c),
+                        reduce_tuple,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op,
+                        reduce_in_element_ops,
+                        reduce_out_element_ops};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        const void* p_bias,
+                        std::array<const void*, 0> p_ds,
+                        void* p_c,
+                        std::array<void*, NumReduce> p_reduces,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        std::array<ck::index_t, 0> StrideDs,
+                        std::array<void*, 3> gemm_element_ops,
+                        std::array<void*, 0> d_element_ops,
+                        std::array<void*, NumReduce> reduce_in_element_op,
+                        std::array<void*, NumReduce> reduce_out_element_op,
+                        ck::index_t = 1) override
+    {
+        (void)p_bias;
+        (void)p_ds;
+        (void)StrideDs;
+        (void)d_element_ops;
+
+        ReducePtrsGlobal reduce_tuple = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReducePtrsGlobal{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return static_cast<T*>(p_reduces[I]);
+            },
+            Number<NumReduce>{});
+
+        ReduceInElementwiseOperations reduce_in_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceInElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_in_element_op[I]));
+            },
+            Number<NumReduce>{});
+        ReduceAccElementwiseOperations reduce_out_element_ops = generate_tuple(
+            [&](auto I) {
+                auto tmp = ReduceAccElementwiseOperations{}[I];
+                using T  = remove_pointer_t<decltype(tmp)>;
+                return *(static_cast<T*>(reduce_out_element_op[I]));
+            },
+            Number<NumReduce>{});
+
+        AElementwiseOperation a_element_op =
+            *(static_cast<AElementwiseOperation*>(gemm_element_ops[0]));
+        BElementwiseOperation b_element_op =
+            *(static_cast<BElementwiseOperation*>(gemm_element_ops[1]));
+        CElementwiseOperation c_element_op =
+            *(static_cast<CElementwiseOperation*>(gemm_element_ops[2]));
+
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          reduce_tuple,
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op,
+                                          reduce_in_element_ops,
+                                          reduce_out_element_ops);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmReduce_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle 
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp
new file mode 100644
index 0000000000..f64450b75f
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp
@@ -0,0 +1,606 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerWMMA,
+          ck::index_t NPerWMMA,
+          ck::index_t MRepeat,
+          ck::index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          ck::index_t NumPrefetch         = 1,
+          ck::LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>
+struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
+                                                   BLayout,
+                                                   CLayout,
+                                                   ADataType,
+                                                   BDataType,
+                                                   CDataType,
+                                                   AElementwiseOperation,
+                                                   BElementwiseOperation,
+                                                   CElementwiseOperation>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    // K1 = Max Vector Access Pixels
+    static constexpr auto K1Number = Number<K1>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock* K1};
+
+    static auto MakeAGridDescriptor_K0_M_K1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto a_grid_desc_m_k = matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        const auto M               = a_grid_desc_m_k.GetLength(I0);
+        const auto K               = a_grid_desc_m_k.GetLength(I1);
+        assert(K % K1 == 0);
+        const index_t K0 = K / K1;
+
+        return transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                       make_pass_through_transform(M)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    static auto MakeBGridDescriptor_K0_N_K1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto b_grid_desc_n_k = matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        const auto N               = b_grid_desc_n_k.GetLength(I0);
+        const auto K               = b_grid_desc_n_k.GetLength(I1);
+        assert(K % K1 == 0);
+        const index_t K0 = K / K1;
+
+        return transform_tensor_descriptor(
+            b_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                       make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideC));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(c_grid_desc_mraw_nraw);
+    }
+
+    // Gridwise descriptor, mapping to whole given provblem.
+    using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
+    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
+    using CGridDesc_M_N     = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_wmma<
+        BlockSize,
+        ADataType,
+        BDataType,
+        AccDataType,
+        CShuffleDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerWMMA,
+        NPerWMMA,
+        K1,
+        MRepeat,
+        NRepeat,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMRepeatPerShuffle,
+        CShuffleNRepeatPerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        NumPrefetch,
+        LoopSched,
+        PipelineVer>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 index_t M,
+                 index_t N,
+                 index_t K,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 index_t M01,
+                 index_t N01,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_mblock_mperblock_nblock_nperblock{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op},
+              MRaw_{M},
+              NRaw_{N},
+              KRaw_{K}
+        {
+            a_grid_desc_k0_m_k1_ =
+                DeviceGemmWmma_CShuffle::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
+            b_grid_desc_k0_n_k1_ =
+                DeviceGemmWmma_CShuffle::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
+            c_grid_desc_m_n_ = DeviceGemmWmma_CShuffle::MakeCGridDescriptor_M_N(M, N, StrideC);
+
+            block_2_ctile_map_ =
+                GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
+                                           b_grid_desc_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+        // for checking vector load/store
+        index_t MRaw_;
+        index_t NRaw_;
+        index_t KRaw_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceGemmWmma_CShuffle::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if 0
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) 
+                          << ", " << arg.c_grid_desc_m_n_.GetLength(I1) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I2) << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_m0nm1_wmma_v1r1 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_wmma<
+                    GridwiseGemm,
+                    ADataType,
+                    BDataType,
+                    CDataType,
+                    remove_reference_t<DeviceGemmWmma_CShuffle::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceGemmWmma_CShuffle::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    true>; // Last Option is W/O
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  arg.a_element_op_,
+                                                  arg.b_element_op_,
+                                                  arg.c_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_wmma<
+                    GridwiseGemm,
+                    ADataType,
+                    BDataType,
+                    CDataType,
+                    remove_reference_t<DeviceGemmWmma_CShuffle::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceGemmWmma_CShuffle::BGridDesc_K0_N_K1>,
+                    remove_reference_t<
+                        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    false>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_n_k1_,
+                                                  arg.c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  arg.a_element_op_,
+                                                  arg.b_element_op_,
+                                                  arg.c_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
+           ck::get_device_name() == "gfx1102")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector load/store
+        {
+            using Row = ck::tensor_layout::gemm::RowMajor;
+            using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+            // check vector load of A
+            if constexpr(is_same_v<ALayout, Row> && ABlockTransferSrcVectorDim == 2)
+            {
+                if(arg.KRaw_ % ABlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else if constexpr(is_same_v<ALayout, Col> && ABlockTransferSrcVectorDim == 1)
+            {
+                // FIXME: not rigorous
+                if(arg.MRaw_ % ABlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+
+            // check vector laod of B
+            if constexpr(is_same_v<BLayout, Col> && BBlockTransferSrcVectorDim == 2)
+            {
+                if(arg.KRaw_ % BBlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else if constexpr(is_same_v<BLayout, Row> && BBlockTransferSrcVectorDim == 1)
+            {
+                // FIXME: not rigorous
+                if(arg.NRaw_ % BBlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+
+            // check vector store of C
+            // only support RowMajor for now
+            if constexpr(is_same_v<CLayout, Row>)
+            {
+                if(arg.NRaw_ % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        1,
+                        1,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          1,
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
+        // clang-format off
+        str << "DeviceGemmWmma_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerWMMA << ", "
+            << NPerWMMA << ", "
+            << MRepeat << ", "
+            << NRepeat
+            << ">"
+            << " NumPrefetch: "
+            << NumPrefetch << ", "
+            << "LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp
new file mode 100644
index 0000000000..ebbb1d38c0
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp
@@ -0,0 +1,305 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector,
+          ck::index_t NumPrefetch         = 1,
+          ck::LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>
+struct DeviceGemmXdl : public DeviceGemm<ALayout,
+                                         BLayout,
+                                         CLayout,
+                                         ADataType,
+                                         BDataType,
+                                         CDataType,
+                                         AElementwiseOperation,
+                                         BElementwiseOperation,
+                                         CElementwiseOperation>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto K1Number = Number<K1>{};
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        ALayout,
+        BLayout,
+        CLayout,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        GemmSpec,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        Sequence<0, 2, 4, 5, 6, 1, 3, 7>, // CThreadTransferSrcDstAccessOrder,
+        CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector,
+        NumPrefetch,
+        LoopSched,
+        PipelineVer>;
+
+    using Argument = typename GridwiseGemm::Argument;
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& karg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                karg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(karg))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext has invalid setting");
+            }
+
+            const auto [gdx, gdy, gdz] = GridwiseGemm::CalculateGridSize(karg.M, karg.N);
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(karg.K))
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3<GridwiseGemm, true>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_v2r3<GridwiseGemm, false>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& karg)
+    {
+        if(ck::get_device_name() == "gfx908")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t>))
+            {
+                return false;
+            }
+        }
+        else if(ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        if(karg.K % K1 != 0)
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(karg);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation,
+                             BElementwiseOperation,
+                             CElementwiseOperation)
+    {
+        return Argument{p_a, p_b, p_c, M, N, K, StrideA, StrideB, StrideC};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation,
+                                                      BElementwiseOperation,
+                                                      CElementwiseOperation) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
+        // clang-format off
+        str << "DeviceGemmXdl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << ABlockTransferDstScalarPerVector_K1 << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferDstScalarPerVector_K1
+            << ">"
+            << " NumPrefetch: "
+            << NumPrefetch << ", "
+            << "LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..86bc7b5bbb
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp
@@ -0,0 +1,306 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Note: inter-wave loop scheduler is rolled out to c-shuffle version first. Becuase non c-shuffle
+// version currently has compiler issues with register spill which further causes validation
+// failures.
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          PipelineVersion PipelineVer = PipelineVersion::v1,
+          typename ComputeType        = CDataType>
+struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
+                                                   BLayout,
+                                                   CLayout,
+                                                   ADataType,
+                                                   BDataType,
+                                                   CDataType,
+                                                   AElementwiseOperation,
+                                                   BElementwiseOperation,
+                                                   CElementwiseOperation>
+{
+    using DeviceOp = DeviceGemm_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ALayout,
+        BLayout,
+        CLayout,
+        ADataType,
+        BDataType,
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        GemmSpec,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched,
+        PipelineVer,
+        ComputeType>;
+
+    using Argument = typename GridwiseGemm::Argument;
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            index_t gdx, gdy, gdz;
+            std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N);
+
+            const auto K = GridwiseGemm::CalculateAK0(arg.K) * AK1;
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_xdl_cshuffle_v1<GridwiseGemm, true>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdl_cshuffle_v1<GridwiseGemm, false>;
+
+                ave_time = launch_and_time_kernel(
+                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
+                                                       GemmSpec == GemmSpecialization::NKPadding ||
+                                                       GemmSpec == GemmSpecialization::MNKPadding ||
+                                                       GemmSpec == GemmSpecialization::KPadding))
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation,
+                             BElementwiseOperation,
+                             CElementwiseOperation)
+    {
+        return Argument{p_a, p_b, p_c, M, N, K, StrideA, StrideB, StrideC};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation,
+                                                      BElementwiseOperation,
+                                                      CElementwiseOperation) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<LoopScheduler, std::string> LoopSchedToString{
+            {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
+
+        std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
+                                                                       {PipelineVersion::v2, "v2"}};
+
+        // clang-format off
+        str << "DeviceGemm_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle
+            << ">"
+            << " LoopScheduler: "
+            << LoopSchedToString[LoopSched] << ", "
+            << "PipelineVersion: "
+            << PipelineVersionToString[PipelineVer];;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp
new file mode 100644
index 0000000000..a8cd39080c
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp
@@ -0,0 +1,781 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// The GEMM + Layernorm implementation is a specialized kernel which allows fusing both layers
+// together given the condition GEMM extents N of MNK is spanned by a single workgroup. For example,
+// a kernel configured with NPerBlock = 128 allows to operate on all GEMM sizes if N <= 128
+//
+// Note: inter-wave loop scheduler is rolled out to c-shuffle version first. Becuase non c-shuffle
+// version currently has compiler issues with register spill which further causes validation
+// failures.
+//
+// D = Layernorm(acc_element_op(A * B + broadcast(bias)) + add) * broadcast(gamma) + broadcast(beta)
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename C0DataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename ReduceAccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+          index_t CReduceThreadCopySrcDstScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGemmLayerNorm_Xdl_CShuffle : public BaseOperator
+{
+    using DeviceOp = DeviceGemmLayerNorm_Xdl_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(M)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_right_pad_transform(MRaw, MPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            assert(K % AK1 == 0);
+
+            const auto AK0 = K / AK1;
+
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_m_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            assert(KRaw % AK1 == 0);
+
+            const auto AK0 = KRaw / AK1;
+
+            const auto a_grid_desc_ak0_m_ak1 =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                       make_pass_through_transform(MRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
+
+        const auto NPad = N - NRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(NRaw, NPad),
+                                                       make_right_pad_transform(KRaw, KPad)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(N)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_right_pad_transform(NRaw, NPad)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            assert(K % BK1 == 0);
+
+            const auto BK0 = K / BK1;
+
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_n_k,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            assert(KRaw % BK1 == 0);
+
+            const auto BK0 = KRaw / BK1;
+
+            const auto b_grid_desc_bk0_n_bk1 =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(NRaw)),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideC));
+            }
+        }();
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+
+        const auto MPad = M - MRaw;
+        const auto NPad = N - NRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(NRaw, NPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_grid_desc_mraw_nraw;
+        }
+    }
+
+    static auto MakeGridDescriptor_N(index_t NRaw)
+    {
+        const auto grid_desc_nraw = make_naive_tensor_descriptor_packed(make_tuple(NRaw));
+
+        const auto N    = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+        const auto NPad = N - NRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                     GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad N
+            return transform_tensor_descriptor(grid_desc_nraw,
+                                               make_tuple(make_right_pad_transform(NRaw, NPad)),
+                                               make_tuple(Sequence<0>{}),
+                                               make_tuple(Sequence<0>{}));
+        }
+        else
+        {
+            // not pad N
+            return grid_desc_nraw;
+        }
+    }
+
+    using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
+    using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
+    using CGridDesc_M_N       = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+    using C0GridDesc_N        = decltype(MakeGridDescriptor_N(1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmLayernorm_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        C0DataType,
+        ReduceAccDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        AccElementwiseOperation,
+        CElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        C0GridDesc_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+        CReduceThreadCopySrcDstScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    using Block2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 const C0DataType* p_c0_grid_add,
+                 const C0DataType* p_c0_grid_bias,
+                 const C0DataType* p_c0_grid_gamma,
+                 const C0DataType* p_c0_grid_beta,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 AccElementwiseOperation acc_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              p_c0_grid_bias_{p_c0_grid_bias},
+              p_c0_grid_add_{p_c0_grid_add},
+              p_c0_grid_gamma_{p_c0_grid_gamma},
+              p_c0_grid_beta_{p_c0_grid_beta},
+              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(MRaw, KRaw, StrideA)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(KRaw, NRaw, StrideB)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(MRaw, NRaw, StrideC)},
+              c0_grid_desc_n_{MakeGridDescriptor_N(NRaw)},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              c0_grid_desc_nblock_nperblock_{},
+              block_2_ctile_map_{Block2CTileMap(c_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              acc_element_op_{acc_element_op},
+              c_element_op_{c_element_op}
+        {
+            if(GridwiseGemm::CheckValidity(a_grid_desc_ak0_m_ak1_,
+                                           b_grid_desc_bk0_n_bk1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n_);
+
+                c0_grid_desc_nblock_nperblock_ =
+                    GridwiseGemm::MakeC0GridDescriptor_NBlock_NPerBlock(c0_grid_desc_n_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        const C0DataType* p_c0_grid_bias_;
+        const C0DataType* p_c0_grid_add_;
+        const C0DataType* p_c0_grid_gamma_;
+        const C0DataType* p_c0_grid_beta_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        C0GridDesc_N c0_grid_desc_n_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::C0GridDescriptor_NBlock_NPerBlock c0_grid_desc_nblock_nperblock_;
+        Block2CTileMap block_2_ctile_map_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        AccElementwiseOperation acc_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if DEBUG_LOG
+            {
+                std::cout << "arg.a_grid_desc_ak0_m_ak1_{"
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_bk0_n_bk1_{"
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I0) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                            arg.b_grid_desc_bk0_n_bk1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            float ave_time = 0;
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                const auto kernel = kernel_gemm_layernorm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    C0DataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    AccElementwiseOperation,
+                    CElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::C0GridDescriptor_NBlock_NPerBlock,
+                    Block2CTileMap,
+                    true>;
+
+                ave_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_c0_grid_bias_,
+                                           arg.p_c0_grid_add_,
+                                           arg.p_c0_grid_gamma_,
+                                           arg.p_c0_grid_beta_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.acc_element_op_,
+                                           arg.c_element_op_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.c0_grid_desc_nblock_nperblock_,
+                                           arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_layernorm_xdl_cshuffle_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    C0DataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    AccElementwiseOperation,
+                    CElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::C0GridDescriptor_NBlock_NPerBlock,
+                    Block2CTileMap,
+                    false>;
+                ave_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(grid_size),
+                                           dim3(BlockSize),
+                                           0,
+                                           arg.p_a_grid_,
+                                           arg.p_b_grid_,
+                                           arg.p_c_grid_,
+                                           arg.p_c0_grid_bias_,
+                                           arg.p_c0_grid_add_,
+                                           arg.p_c0_grid_gamma_,
+                                           arg.p_c0_grid_beta_,
+                                           arg.a_element_op_,
+                                           arg.b_element_op_,
+                                           arg.acc_element_op_,
+                                           arg.c_element_op_,
+                                           arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                           arg.c0_grid_desc_nblock_nperblock_,
+                                           arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                           arg.b_grid_desc_bk0_n_bk1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             const C0DataType* p_c0_bias,
+                             const C0DataType* p_c0_add,
+                             const C0DataType* p_c0_gamma,
+                             const C0DataType* p_c0_beta,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             AccElementwiseOperation acc_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        p_c0_bias,
+                        p_c0_add,
+                        p_c0_gamma,
+                        p_c0_beta,
+                        MRaw,
+                        NRaw,
+                        KRaw,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        a_element_op,
+                        b_element_op,
+                        acc_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      const void* p_c0_bias,
+                                                      const void* p_c0_add,
+                                                      const void* p_c0_gamma,
+                                                      const void* p_c0_beta,
+                                                      index_t MRaw,
+                                                      index_t NRaw,
+                                                      index_t KRaw,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      AccElementwiseOperation acc_element_op,
+                                                      CElementwiseOperation c_element_op,
+                                                      index_t /* KBatch */ = 1)
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          static_cast<const C0DataType*>(p_c0_bias),
+                                          static_cast<const C0DataType*>(p_c0_add),
+                                          static_cast<const C0DataType*>(p_c0_gamma),
+                                          static_cast<const C0DataType*>(p_c0_beta),
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          a_element_op,
+                                          b_element_op,
+                                          acc_element_op,
+                                          c_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmLayerNorm_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp
new file mode 100644
index 0000000000..d54ddf433e
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp
@@ -0,0 +1,530 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp"
+
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockBufferSize,
+          ck::index_t CThreadTransferSrcDstVectorDim,
+          ck::index_t CThreadTransferDstScalarPerVector>
+struct DeviceGemmXdlSkipBLds : public DeviceGemm<ALayout,
+                                                 BLayout,
+                                                 CLayout,
+                                                 ADataType,
+                                                 BDataType,
+                                                 CDataType,
+                                                 AElementwiseOperation,
+                                                 BElementwiseOperation,
+                                                 CElementwiseOperation>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto K1Number = Number<K1>{};
+    static_assert(BBlockBufferSize >= 2);
+
+    static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(M, PadM)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+    }
+
+    using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
+    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
+    using CGridDesc_M_N     = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_skip_b_lds_v1<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferSrcScalarPerVector,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockBufferSize,
+        Sequence<0, 2, 4, 5, 6, 1, 3, 7>, // CThreadTransferSrcDstAccessOrder,
+        CThreadTransferSrcDstVectorDim,
+        CThreadTransferDstScalarPerVector>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 CDataType* p_c_grid,
+                 index_t M,
+                 index_t N,
+                 index_t K,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideC,
+                 index_t M01,
+                 index_t N01,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : p_a_grid_{p_a_grid},
+              p_b_grid_{p_b_grid},
+              p_c_grid_{p_c_grid},
+              a_grid_desc_k0_m_k1_{},
+              b_grid_desc_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+            a_grid_desc_k0_m_k1_ =
+                DeviceGemmXdlSkipBLds::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
+            b_grid_desc_k0_n_k1_ =
+                DeviceGemmXdlSkipBLds::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
+            c_grid_desc_m_n_ = DeviceGemmXdlSkipBLds::MakeCGridDescriptor_M_N(M, N, StrideC);
+
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
+            {
+                c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
+
+                block_2_ctile_map_ =
+                    GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
+
+                b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3_ =
+                    GridwiseGemm::MakeBGridDescriptor_K0_K1_K2_N0_N1_N2_N3_K3(b_grid_desc_k0_n_k1_);
+            }
+        }
+
+        //  private:
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        typename GridwiseGemm::BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3
+            b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3_;
+        typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2
+            c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
+        typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceGemmXdlSkipBLds::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if DEBUG_LOG
+            {
+                std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
+                          << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
+                          << ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                            arg.b_grid_desc_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.M01_,
+                                            arg.N01_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
+            }
+
+            const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
+
+            const auto K0 = arg.a_grid_desc_k0_m_k1_.GetLength(I0);
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            float ave_time = 0;
+
+            if(has_main_k0_block_loop)
+            {
+                const auto kernel = kernel_gemm_xdlops_skip_b_lds_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceGemmXdlSkipBLds::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceGemmXdlSkipBLds::BGridDesc_K0_N_K1>,
+                    remove_reference_t<typename GridwiseGemm::BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3>,
+                    remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    true>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.a_element_op_,
+                                                  arg.b_element_op_,
+                                                  arg.c_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+            else
+            {
+                const auto kernel = kernel_gemm_xdlops_skip_b_lds_v1<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceGemmXdlSkipBLds::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceGemmXdlSkipBLds::BGridDesc_K0_N_K1>,
+                    remove_reference_t<typename GridwiseGemm::BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3>,
+                    remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CElementwiseOperation,
+                    remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                    false>;
+
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  dim3(grid_size),
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  arg.p_a_grid_,
+                                                  arg.p_b_grid_,
+                                                  arg.p_c_grid_,
+                                                  arg.a_grid_desc_k0_m_k1_,
+                                                  arg.b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3_,
+                                                  arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
+                                                  arg.a_element_op_,
+                                                  arg.b_element_op_,
+                                                  arg.c_element_op_,
+                                                  arg.block_2_ctile_map_);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
+                                           arg.b_grid_desc_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.M01_,
+                                           arg.N01_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        1,
+                        1,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          1,
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemmXdlSkipBLds"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
new file mode 100644
index 0000000000..e7365daed2
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
@@ -0,0 +1,329 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXDL,
+          typename ComputeType        = CDataType,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+
+struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
+                                                             BLayout,
+                                                             CLayout,
+                                                             ADataType,
+                                                             BDataType,
+                                                             CDataType,
+                                                             AElementwiseOperation,
+                                                             BElementwiseOperation,
+                                                             CElementwiseOperation>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    // TODO: should be exposed as Tparams.
+    static constexpr index_t NumGemmKPrefetchStage = 1;
+    static constexpr LoopScheduler LoopSched       = make_default_loop_scheduler();
+
+    using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
+        BlockSize,
+        ADataType,
+        BDataType,
+        AccDataType,
+        CDataType,
+        ALayout,
+        BLayout,
+        CLayout,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        GemmSpec,
+        NumGemmKPrefetchStage,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMRepeatPerShuffle,
+        CShuffleNRepeatPerShuffle,
+        CBlockTransferScalarPerVector_NWaveNPerXDL,
+        CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        LoopSched,
+        PipelineVer,
+        ComputeType>;
+
+    using Argument              = typename GridwiseGemm::Argument;
+    using DefaultBlock2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+
+        void Print(const Argument& karg) { karg.Print(); }
+
+        float Run(const Argument& karg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                Print(karg);
+            }
+
+            const auto kbatch = karg.k_batch;
+
+            if(!GridwiseGemm::CheckValidity(karg))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 has invalid "
+                    "setting");
+            }
+
+            const auto b2c_map = DefaultBlock2CTileMap{};
+            index_t gdx, gdy, gdz;
+            std::tie(gdx, gdy, gdz) = b2c_map.CalculateGridSize(karg.M, karg.N, karg.k_batch);
+            const auto K0           = karg.K0;
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            float ave_time = 0;
+
+            const auto Run = [&](const auto& kernel) {
+                if(kbatch > 1)
+                    hipGetErrorString(hipMemsetAsync(karg.p_c_grid,
+                                                     0,
+                                                     karg.M * karg.N * sizeof(CDataType),
+                                                     stream_config.stream_id_));
+
+                ave_time = launch_and_time_kernel(
+                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg, b2c_map);
+            };
+
+            if(has_main_k0_block_loop)
+            {
+                if(kbatch == 1)
+                {
+                    const auto kernel =
+                        kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
+                                                             true,
+                                                             InMemoryDataOperationEnum::Set,
+                                                             DefaultBlock2CTileMap>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel =
+                        kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
+                                                             true,
+                                                             InMemoryDataOperationEnum::AtomicAdd,
+                                                             DefaultBlock2CTileMap>;
+
+                    Run(kernel);
+                }
+            }
+            else
+            {
+                if(kbatch == 1)
+                {
+                    const auto kernel =
+                        kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
+                                                             false,
+                                                             InMemoryDataOperationEnum::Set,
+                                                             DefaultBlock2CTileMap>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel =
+                        kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
+                                                             false,
+                                                             InMemoryDataOperationEnum::AtomicAdd,
+                                                             DefaultBlock2CTileMap>;
+
+                    Run(kernel);
+                }
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& karg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(karg);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation,
+                             BElementwiseOperation,
+                             CElementwiseOperation,
+                             index_t KBatch)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        GridwiseGemm::CalculateMPadded(M),
+                        GridwiseGemm::CalculateNPadded(N),
+                        GridwiseGemm::CalculateKPadded(K, KBatch),
+                        GridwiseGemm::CalculateK0(K, KBatch),
+                        KBatch};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation,
+                                                      BElementwiseOperation,
+                                                      CElementwiseOperation,
+                                                      ck::index_t KBatch = 1) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          GridwiseGemm::CalculateMPadded(M),
+                                          GridwiseGemm::CalculateNPadded(N),
+                                          GridwiseGemm::CalculateKPadded(K, KBatch),
+                                          GridwiseGemm::CalculateK0(K, KBatch),
+                                          KBatch);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override { return GridwiseGemm::GetTypeString(); }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_streamk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_streamk.hpp
new file mode 100644
index 0000000000..8de42ba9ef
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_streamk.hpp
@@ -0,0 +1,360 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_streamk.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/hip_check_error.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          ck::index_t ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          ck::index_t BBlockLdsAddExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXDL>
+struct DeviceGemmXdlStreamK : public DeviceGemmStreamK<ALayout,
+                                                       BLayout,
+                                                       CLayout,
+                                                       ADataType,
+                                                       BDataType,
+                                                       CDataType,
+                                                       AElementwiseOperation,
+                                                       BElementwiseOperation,
+                                                       CElementwiseOperation>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_streamk<
+        BlockSize,
+        BlockToCTileMap_GemmStreamK<MPerBlock,
+                                    NPerBlock,
+                                    K0PerBlock * K1,
+                                    StreamKReductionStrategy::Atomic>,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        ALayout,
+        BLayout,
+        CLayout,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        CShuffleMRepeatPerShuffle,
+        CShuffleNRepeatPerShuffle,
+        CBlockTransferScalarPerVector_NWaveNPerXDL,
+        CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
+
+    using Argument = typename GridwiseGemm::Argument;
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        void Print(const Argument& karg) { karg.Print(); }
+
+        float Run(const Argument& karg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                Print(karg);
+            }
+            if(!GridwiseGemm::CheckValidity(karg))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2 has invalid "
+                    "setting");
+            }
+
+            dim3 grid_dims = karg.block_mapping.get_grid_dims();
+
+            float ave_time = 0;
+
+            const auto kernel = kernel_gemm_xdlops_streamk<GridwiseGemm>;
+
+            // TODO: remove clear buffer for streamk kernels
+            if constexpr(GridwiseGemm::Block2CTileMap::ReductionStrategy ==
+                         StreamKReductionStrategy::Atomic)
+            {
+                hipGetErrorString(hipMemsetAsync(karg.p_c_grid,
+                                                 0,
+                                                 karg.M * karg.N * sizeof(CDataType),
+                                                 stream_config.stream_id_));
+                ave_time = launch_and_time_kernel(stream_config,
+                                                  kernel,
+                                                  grid_dims,
+                                                  dim3(BlockSize),
+                                                  0,
+                                                  karg.p_a_grid,
+                                                  karg.p_b_grid,
+                                                  karg.p_c_grid,
+                                                  karg.p_workspace_,
+                                                  karg.M,
+                                                  karg.N,
+                                                  karg.K,
+                                                  karg.StrideA,
+                                                  karg.StrideB,
+                                                  karg.StrideC,
+                                                  karg.block_mapping);
+            }
+            else if constexpr(GridwiseGemm::Block2CTileMap::ReductionStrategy ==
+                              StreamKReductionStrategy::Reduction)
+            {
+                char* workspace_semaphore = reinterpret_cast<char*>(karg.p_workspace_) +
+                                            karg.block_mapping.get_workspace_size_for_acc(
+                                                sizeof(typename GridwiseGemm::FloatAcc));
+                auto preprocess = [&]() {
+                    hipGetErrorString(
+                        hipMemsetAsync(workspace_semaphore,
+                                       0,
+                                       karg.block_mapping.get_workspace_size_for_semaphore(),
+                                       stream_config.stream_id_));
+                };
+
+                ave_time = launch_and_time_kernel_with_preprocess(stream_config,
+                                                                  preprocess,
+                                                                  kernel,
+                                                                  grid_dims,
+                                                                  dim3(BlockSize),
+                                                                  0,
+                                                                  karg.p_a_grid,
+                                                                  karg.p_b_grid,
+                                                                  karg.p_c_grid,
+                                                                  karg.p_workspace_,
+                                                                  karg.M,
+                                                                  karg.N,
+                                                                  karg.K,
+                                                                  karg.StrideA,
+                                                                  karg.StrideB,
+                                                                  karg.StrideC,
+                                                                  karg.block_mapping);
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    size_t GetWorkSpaceSize(const BaseArgument* pArg) const override
+    {
+        const Argument* p_arg = dynamic_cast<const Argument*>(pArg);
+        if constexpr(GridwiseGemm::Block2CTileMap::ReductionStrategy ==
+                     StreamKReductionStrategy::Reduction)
+        {
+            return p_arg->block_mapping.get_workspace_size(sizeof(typename GridwiseGemm::FloatAcc));
+        }
+        else
+        {
+            return 0;
+        }
+    }
+
+    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    {
+        Argument* pArg_ = dynamic_cast<Argument*>(pArg);
+
+        pArg_->p_workspace_ = p_workspace;
+    }
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& karg)
+    {
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
+             ck::get_device_name() == "gfx942"))
+        {
+            return false;
+        }
+        return GridwiseGemm::CheckValidity(karg);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideC,
+                             AElementwiseOperation,
+                             BElementwiseOperation,
+                             CElementwiseOperation,
+                             uint32_t NumSKBlocks = 0xffffffff)
+    {
+        const auto kernel = kernel_gemm_xdlops_streamk<GridwiseGemm>;
+        int occupancy, num_cu;
+        hipError_t rtn;
+        rtn = hipOccupancyMaxActiveBlocksPerMultiprocessor(
+            &occupancy, kernel, BlockSize, GridwiseGemm::GetSharedMemoryNumberOfByte());
+        hip_check_error(rtn);
+
+        hipDeviceProp_t dev_prop;
+        hipDevice_t dev;
+        rtn = hipGetDevice(&dev);
+        hip_check_error(rtn);
+        rtn = hipGetDeviceProperties(&dev_prop, dev);
+        hip_check_error(rtn);
+        num_cu = dev_prop.multiProcessorCount;
+
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideB,
+                        StrideC,
+                        static_cast<uint32_t>(num_cu),
+                        static_cast<uint32_t>(occupancy),
+                        NumSKBlocks};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_b,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideB,
+                                                      index_t StrideC,
+                                                      AElementwiseOperation,
+                                                      BElementwiseOperation,
+                                                      CElementwiseOperation,
+                                                      index_t NumSKBlocks = 0) override
+    {
+        const auto kernel = kernel_gemm_xdlops_streamk<GridwiseGemm>;
+        int occupancy, num_cu;
+        hipError_t rtn;
+        rtn = hipOccupancyMaxActiveBlocksPerMultiprocessor(
+            &occupancy, kernel, BlockSize, GridwiseGemm::GetSharedMemoryNumberOfByte());
+        hip_check_error(rtn);
+
+        hipDeviceProp_t dev_prop;
+        hipDevice_t dev;
+        rtn = hipGetDevice(&dev);
+        hip_check_error(rtn);
+        rtn = hipGetDeviceProperties(&dev_prop, dev);
+        hip_check_error(rtn);
+        num_cu = dev_prop.multiProcessorCount;
+
+        return std::make_unique<Argument>(reinterpret_cast<const ADataType*>(p_a),
+                                          reinterpret_cast<const BDataType*>(p_b),
+                                          reinterpret_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideC,
+                                          static_cast<uint32_t>(num_cu),
+                                          static_cast<uint32_t>(occupancy),
+                                          static_cast<uint32_t>(NumSKBlocks));
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override { return GridwiseGemm::GetTypeString(); }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp
new file mode 100644
index 0000000000..1b34e2dba2
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp
@@ -0,0 +1,527 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename EElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_WAVELET_MAX_THREAD_PER_BLOCK, CK_WAVELET_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdl_waveletmodel_cshuffle(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const EElementwiseOperation e_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_e_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  e_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_e_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = e_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_etile_map;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename GemmAcEDataType,
+          typename CShuffleDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t TileLoadThreadGroupSize,
+          index_t TileMathThreadGroupSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
+struct DeviceGemm_Xdl_WaveletModel_CShuffle : public DeviceGemm<ALayout,
+                                                                BLayout,
+                                                                ELayout,
+                                                                ADataType,
+                                                                BDataType,
+                                                                EDataType,
+                                                                AElementwiseOperation,
+                                                                BElementwiseOperation,
+                                                                CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGemm_Xdl_WaveletModel_CShuffle;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    template <typename ELay>
+    static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    using AGridDesc_M_K = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAcEDataType,
+        CShuffleDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_M_K,
+        BGridDesc_N_K,
+        EGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        TileLoadThreadGroupSize,
+        TileMathThreadGroupSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock>;
+
+    using AGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
+            AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
+            BGridDesc_N_K{}))>;
+
+    using Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const ADataType* p_a_grid,
+                 const BDataType* p_b_grid,
+                 EDataType* p_e_grid,
+                 index_t MRaw,
+                 index_t NRaw,
+                 index_t KRaw,
+                 index_t StrideA,
+                 index_t StrideB,
+                 index_t StrideE,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(MRaw, KRaw, StrideA)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(KRaw, NRaw, StrideB)},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(MRaw, NRaw, StrideE)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_m_k_, b_grid_desc_n_k_, e_grid_desc_m_n_, block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+#if 0
+            {
+                std::cout << "arg.a_grid_desc_ak0_m_ak1_{"
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) << ", "
+                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.b_grid_desc_bk0_n_bk1_{"
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I0) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) << ", "
+                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) << "}" << std::endl;
+
+                std::cout << "arg.e_grid_desc_m_n_{ " << arg.e_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.e_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+            }
+#endif
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.e_grid_desc_m_n_);
+            const auto K            = arg.a_grid_desc_m_k_.GetLength(I1);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_gemm_xdl_waveletmodel_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::DefaultBlock2ETileMap,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(grid_size),
+                    dim3(TileLoadThreadGroupSize + TileMathThreadGroupSize),
+                    0,
+                    arg.p_a_grid_,
+                    arg.p_b_grid_,
+                    arg.p_e_grid_,
+                    arg.a_element_op_,
+                    arg.b_element_op_,
+                    arg.cde_element_op_,
+                    arg.a_grid_desc_ak0_m_ak1_,
+                    arg.b_grid_desc_bk0_n_bk1_,
+                    arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                    arg.block_2_etile_map_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           arg.e_grid_desc_m_n_,
+                                           arg.block_2_etile_map_);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const BDataType* p_b,
+                             EDataType* p_e,
+                             index_t MRaw,
+                             index_t NRaw,
+                             index_t KRaw,
+                             index_t StrideA,
+                             index_t StrideB,
+                             index_t StrideE,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_e,
+                        MRaw,
+                        NRaw,
+                        KRaw,
+                        StrideA,
+                        StrideB,
+                        StrideE,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_e,
+                        index_t MRaw,
+                        index_t NRaw,
+                        index_t KRaw,
+                        index_t StrideA,
+                        index_t StrideB,
+                        index_t StrideE,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<EDataType*>(p_e),
+                                          MRaw,
+                                          NRaw,
+                                          KRaw,
+                                          StrideA,
+                                          StrideB,
+                                          StrideE,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGemm_Xdl_WaveletModel_CShuffle"
+            << "<"
+	    << TileLoadThreadGroupSize << ", "
+            << TileMathThreadGroupSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..4823f5d489
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,917 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename ContractionMultiDKernelArg,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_contraction_multiple_d_xdl_cshuffle(
+            const void CK_CONSTANT_ADDRESS_SPACE* contraction_args,
+            const index_t group_count,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t block_id = get_block_1d_id();
+
+    const auto contraction_arg_ptr = reinterpret_cast<const ContractionMultiDKernelArg*>(
+        cast_pointer_to_generic_address_space(contraction_args));
+
+    index_t left     = 0;
+    index_t right    = group_count;
+    index_t group_id = index_t((left + right) / 2);
+
+    while((!(block_id >= contraction_arg_ptr[group_id].block_start_ &&
+             block_id < contraction_arg_ptr[group_id].block_end_)) &&
+          left <= right)
+    {
+        if(block_id < contraction_arg_ptr[group_id].block_start_)
+        {
+            right = group_id;
+        }
+        else
+        {
+            left = group_id;
+        }
+        group_id = index_t((left + right) / 2);
+    }
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        contraction_arg_ptr[group_id].p_a_grid_,
+        contraction_arg_ptr[group_id].p_b_grid_,
+        contraction_arg_ptr[group_id].p_ds_grid_,
+        contraction_arg_ptr[group_id].p_e_grid_,
+        p_shared,
+        a_element_op,
+        b_element_op,
+        cde_element_op,
+        contraction_arg_ptr[group_id].a_grid_desc_ak0_m_ak1_,
+        contraction_arg_ptr[group_id].b_grid_desc_bk0_n_bk1_,
+        contraction_arg_ptr[group_id].ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+        contraction_arg_ptr[group_id].e_grid_desc_mblock_mperblock_nblock_nperblock_,
+        contraction_arg_ptr[group_id].block_2_etile_map_);
+#else
+    ignore = contraction_args;
+    ignore = group_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization DESpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedContractionMultipleD_Xdl_CShuffle
+    : public DeviceGroupedContractionMultipleD<NumDimM,
+                                               NumDimN,
+                                               NumDimK,
+                                               ADataType,
+                                               BDataType,
+                                               DsDataType,
+                                               EDataType,
+                                               AElementwiseOperation,
+                                               BElementwiseOperation,
+                                               CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedContractionMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    // Assume: A[M0, M1, M2, ..., K0, K1, K2, ...]
+    static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_ms_ks_strides_vec)
+    {
+        assert(a_ms_ks_lengths_vec.size() == NumDimM + NumDimK &&
+               a_ms_ks_strides_vec.size() == NumDimM + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto a_ms_ks_lengths = to_tuple(a_ms_ks_lengths_vec, Number<NumDimM + NumDimK>{});
+        const auto a_ms_ks_strides = to_tuple(a_ms_ks_strides_vec, Number<NumDimM + NumDimK>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimK, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(a_ms_ks_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(a_ms_ks_lengths, kDimIds);
+
+        if constexpr(ASpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto a_grid_desc_mraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(M, K),
+                make_tuple(a_ms_ks_strides[Number<NumDimM - 1>{}],
+                           a_ms_ks_strides[Number<NumDimM + NumDimK - 1>{}]));
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+        else
+        {
+            // naive tensor A[M0, M1, M2, ..., K0, K1, K2...]
+            const auto a_grid_desc_ms_ks =
+                make_naive_tensor_descriptor(a_ms_ks_lengths, a_ms_ks_strides);
+
+            // transformed tensor A[MRaw = M0 * M1 * M2 * ... , KRaw = K0 * K1 * K2 * ...]
+            const auto a_grid_desc_mraw_kraw = transform_tensor_descriptor(
+                a_grid_desc_ms_ks,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(kLengths)),
+                make_tuple(mDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+    }
+
+    // Assume: B[N0, N1, N2, ..., K0, K1, K2, ...]
+    static auto MakeBGridDescriptor_N_K(const std::vector<index_t>& b_ns_ks_lengths_vec,
+                                        const std::vector<index_t>& b_ns_ks_strides_vec)
+    {
+        assert(b_ns_ks_lengths_vec.size() == NumDimN + NumDimK &&
+               b_ns_ks_strides_vec.size() == NumDimN + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto b_ns_ks_lengths = to_tuple(b_ns_ks_lengths_vec, Number<NumDimN + NumDimK>{});
+        const auto b_ns_ks_strides = to_tuple(b_ns_ks_strides_vec, Number<NumDimN + NumDimK>{});
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<0, NumDimN, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimN, NumDimN + NumDimK, 1>::type{};
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(b_ns_ks_lengths, kDimIds);
+
+        // lengths for N0, N1, ...
+        const auto nLengths = get_container_subset(b_ns_ks_lengths, nDimIds);
+
+        if constexpr(BSpec == TensorSpecialization::Packed)
+        {
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto b_grid_desc_nraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(N, K),
+                make_tuple(b_ns_ks_strides[Number<NumDimN - 1>{}],
+                           b_ns_ks_strides[Number<NumDimN + NumDimK - 1>{}]));
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+        else
+        {
+            // naive tensor B[N0, N1, N2, ..., K0, K1, K2, ...]
+            const auto b_grid_desc_ns_ks =
+                make_naive_tensor_descriptor(b_ns_ks_lengths, b_ns_ks_strides);
+
+            // transformed tensor B[NRaw = N0 * N1 * N2 * ..., KRaw = K0 * K1 * K2 * ...]
+            const auto b_grid_desc_nraw_kraw = transform_tensor_descriptor(
+                b_grid_desc_ns_ks,
+                make_tuple(make_merge_transform(nLengths), make_merge_transform(kLengths)),
+                make_tuple(nDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+    }
+
+    // assume E[M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_M_N(const std::vector<index_t>& e_ms_ns_lengths_vec,
+                                        const std::vector<index_t>& e_ms_ns_strides_vec)
+    {
+        assert(e_ms_ns_lengths_vec.size() == NumDimM + NumDimN &&
+               e_ms_ns_strides_vec.size() == NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto num) {
+            return generate_tuple([&](auto i) { return vec[i]; }, num);
+        };
+
+        const auto e_ms_ns_lengths = to_tuple(e_ms_ns_lengths_vec, Number<NumDimM + NumDimN>{});
+        const auto e_ms_ns_strides = to_tuple(e_ms_ns_strides_vec, Number<NumDimM + NumDimN>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(M, N),
+                make_tuple(e_ms_ns_strides[Number<NumDimM - 1>{}],
+                           e_ms_ns_strides[Number<NumDimM + NumDimN - 1>{}]));
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+        else
+        {
+            // naive tensor E[M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_ms_ns =
+                make_naive_tensor_descriptor(e_ms_ns_lengths, e_ms_ns_strides);
+
+            // transformed tensor E[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
+            const auto e_grid_desc_mraw_nraw = transform_tensor_descriptor(
+                e_grid_desc_ms_ns,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+                make_tuple(mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_M_N(ds_ms_ns_lengths_vec[i],
+                                                         ds_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K({}, {}));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K({}, {}));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({{}}, {{}}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N({}, {}));
+
+    using ComputeDataType = ADataType;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeDataType,
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // desc for blockwise copy
+    using AGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
+            AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
+            BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
+        decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+
+    struct GroupedContractionBlock2ETileMap
+    {
+        // block-to-e-tile map
+        using Block2ETileMap =
+            remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+        GroupedContractionBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n,
+                                         ck::index_t BlockStart)
+        {
+            default_block_2_etile_map_ = GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n);
+            block_start_               = BlockStart;
+        }
+
+        template <typename TopIdx>
+        __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+        {
+            return default_block_2_etile_map_.CalculateBottomIndex(
+                make_multi_index(idx_top[I0] - block_start_));
+        }
+
+        // it's actually E-Tile
+        template <typename CTileIdx, typename CTileDim>
+        __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                                 const CTileDim& c_tile_dim) const
+        {
+            return default_block_2_etile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
+        }
+
+        __host__ bool CheckValidity(const EGridDesc_M_N& e_grid_desc_m_n) const
+        {
+            return default_block_2_etile_map_.CheckValidity(e_grid_desc_m_n);
+        }
+
+        Block2ETileMap default_block_2_etile_map_;
+        ck::index_t block_start_;
+    };
+
+    struct ContractionMultiDKernelArg
+    {
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // lock-to-e-tile map
+        GroupedContractionBlock2ETileMap block_2_etile_map_;
+
+        ck::index_t block_start_, block_end_;
+    };
+
+    struct ContractionMultiDDeviceArg
+    {
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // Strides for the last M/N/K dimensions of A/B/Ds/E
+        //   for sanity check of vector load/store
+        index_t a_mz_stride_;
+        index_t a_kz_stride_;
+        index_t b_nz_stride_;
+        index_t b_kz_stride_;
+        std::array<index_t, NumDTensor> ds_nz_stride_;
+        // index_t e_mz_stride_;
+        index_t e_nz_stride_;
+    };
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(std::vector<const void*> p_a_vec,
+                 std::vector<const void*> p_b_vec,
+                 std::vector<std::array<const void*, NumDTensor>> p_ds_vec,
+                 std::vector<void*> p_e_vec,
+                 std::vector<ContractionDesc<NumDTensor>> contraction_descs,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+            group_count_ = contraction_descs.size();
+
+            if(!(group_count_ == p_a_vec.size() && group_count_ == p_b_vec.size() &&
+                 group_count_ == p_e_vec.size()))
+            {
+                throw std::runtime_error("wrong! group_count_ != a/b/e_vec.size");
+            }
+
+            contraction_multi_d_kernel_args_.reserve(group_count_);
+
+            grid_size_ = 0;
+
+            for(std::size_t i = 0; i < group_count_; i++)
+            {
+                const auto p_a_grid = static_cast<const ADataType*>(p_a_vec[i]);
+                const auto p_b_grid = static_cast<const BDataType*>(p_b_vec[i]);
+                const auto p_e_grid = static_cast<EDataType*>(p_e_vec[i]);
+
+                const auto a_grid_desc_m_k = DeviceOp::MakeAGridDescriptor_M_K(
+                    contraction_descs[i].a_ms_ks_lengths, contraction_descs[i].a_ms_ks_strides);
+                const auto b_grid_desc_n_k = DeviceOp::MakeBGridDescriptor_N_K(
+                    contraction_descs[i].b_ns_ks_lengths, contraction_descs[i].b_ns_ks_strides);
+
+                DsGridDesc_M_N ds_grid_desc_m_n;
+                typename GridwiseGemm::DsGridPointer p_ds_grid;
+
+                // populate pointer, batch stride, desc for Ds
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    using DDataType = remove_cvref_t<tuple_element_t<j.value, DsDataType>>;
+
+                    // D pointer
+                    p_ds_grid(j) = static_cast<const DDataType*>(p_ds_vec[i][j]);
+
+                    // D desc
+                    ds_grid_desc_m_n(j) =
+                        DeviceOp::MakeEGridDescriptor_M_N(contraction_descs[i].ds_ms_ns_lengths[j],
+                                                          contraction_descs[i].ds_ms_ns_strides[j]);
+                });
+
+                const auto e_grid_desc_m_n = DeviceOp::MakeEGridDescriptor_M_N(
+                    contraction_descs[i].e_ms_ns_lengths, contraction_descs[i].e_ms_ns_strides);
+
+                const auto a_grid_desc_ak0_m_ak1 =
+                    GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k);
+                const auto b_grid_desc_bk0_n_bk1 =
+                    GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k);
+
+                const auto ds_grid_desc_mblock_mperblock_nblock_nperblock =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n);
+                const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n);
+
+                const index_t grid_size_grp =
+                    GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n)
+                        .CalculateGridSize(e_grid_desc_m_n);
+
+                const index_t BlockStart = grid_size_;
+                const index_t BlockEnd   = grid_size_ + grid_size_grp;
+
+                grid_size_ += grid_size_grp;
+
+                const auto block_2_etile_map =
+                    GroupedContractionBlock2ETileMap(e_grid_desc_m_n, BlockStart);
+
+                // for sanity check of vector memory access
+                const index_t a_mz_stride = contraction_descs[i].a_ms_ks_strides[NumDimM - 1];
+                const index_t a_kz_stride =
+                    contraction_descs[i].a_ms_ks_strides[NumDimM + NumDimK - 1];
+
+                const index_t b_nz_stride = contraction_descs[i].b_ns_ks_strides[NumDimN - 1];
+                const index_t b_kz_stride =
+                    contraction_descs[i].b_ns_ks_strides[NumDimN + NumDimK - 1];
+
+                std::array<index_t, NumDTensor> ds_nz_stride;
+                for(index_t j = 0; j < NumDTensor; ++j)
+                {
+                    ds_nz_stride[j] =
+                        contraction_descs[i].ds_ms_ns_strides[j][NumDimM + NumDimN - 1];
+                }
+
+                const index_t e_nz_stride =
+                    contraction_descs[i].e_ms_ns_strides[NumDimM + NumDimN - 1];
+
+                if(GridwiseGemm::CheckValidity(a_grid_desc_m_k,
+                                               b_grid_desc_n_k,
+                                               ds_grid_desc_m_n,
+                                               e_grid_desc_m_n,
+                                               block_2_etile_map))
+                {
+                    contraction_multi_d_kernel_args_.push_back(
+                        {p_a_grid,
+                         p_b_grid,
+                         p_ds_grid,
+                         p_e_grid,
+                         a_grid_desc_ak0_m_ak1,
+                         b_grid_desc_bk0_n_bk1,
+                         ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                         e_grid_desc_mblock_mperblock_nblock_nperblock,
+                         block_2_etile_map,
+                         BlockStart,
+                         BlockEnd});
+
+                    contraction_multi_d_device_args_.push_back({a_grid_desc_m_k,
+                                                                b_grid_desc_n_k,
+                                                                ds_grid_desc_m_n,
+                                                                e_grid_desc_m_n,
+                                                                a_mz_stride,
+                                                                a_kz_stride,
+                                                                b_nz_stride,
+                                                                b_kz_stride,
+                                                                ds_nz_stride,
+                                                                e_nz_stride});
+                }
+            }
+        }
+
+        std::vector<ContractionMultiDKernelArg> contraction_multi_d_kernel_args_;
+        std::vector<ContractionMultiDDeviceArg> contraction_multi_d_device_args_;
+
+        std::size_t group_count_;
+        index_t grid_size_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            bool has_main_k_block_loop = true;
+
+            for(std::size_t i = 0; i < arg.group_count_; i++)
+            {
+                const auto K =
+                    arg.contraction_multi_d_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0) *
+                    arg.contraction_multi_d_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+                if(GridwiseGemm::CalculateHasMainKBlockLoop(K) != has_main_k_block_loop)
+                {
+                    throw std::runtime_error("wrong! not all gemm has_main_k_block_loop");
+                }
+            }
+
+            hipGetErrorString(hipMemcpyWithStream(arg.p_workspace_,
+                                                  arg.contraction_multi_d_kernel_args_.data(),
+                                                  arg.contraction_multi_d_kernel_args_.size() *
+                                                      sizeof(ContractionMultiDKernelArg),
+                                                  hipMemcpyHostToDevice,
+                                                  stream_config.stream_id_));
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel =
+                    kernel_grouped_contraction_multiple_d_xdl_cshuffle<GridwiseGemm,
+                                                                       ContractionMultiDKernelArg,
+                                                                       AElementwiseOperation,
+                                                                       BElementwiseOperation,
+                                                                       CDEElementwiseOperation,
+                                                                       has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(arg.grid_size_),
+                    dim3(BlockSize),
+                    0,
+                    cast_pointer_to_constant_address_space(arg.p_workspace_),
+                    arg.group_count_,
+                    arg.a_element_op_,
+                    arg.b_element_op_,
+                    arg.cde_element_op_);
+            };
+
+            if(has_main_k_block_loop)
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        for(std::size_t i = 0; i < arg.group_count_; i++)
+        {
+            const auto a_grid_desc_m_k_ = arg.contraction_multi_d_device_args_[i].a_grid_desc_m_k_;
+            const auto b_grid_desc_n_k_ = arg.contraction_multi_d_device_args_[i].b_grid_desc_n_k_;
+            const auto ds_grid_desc_m_n_ =
+                arg.contraction_multi_d_device_args_[i].ds_grid_desc_m_n_;
+            const auto e_grid_desc_m_n_ = arg.contraction_multi_d_device_args_[i].e_grid_desc_m_n_;
+            const auto a_grid_desc_ak0_m_ak1_ =
+                arg.contraction_multi_d_kernel_args_[i].a_grid_desc_ak0_m_ak1_;
+            const auto b_grid_desc_bk0_n_bk1_ =
+                arg.contraction_multi_d_kernel_args_[i].b_grid_desc_bk0_n_bk1_;
+            const auto ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                arg.contraction_multi_d_kernel_args_[i]
+                    .ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+            const auto e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                arg.contraction_multi_d_kernel_args_[i]
+                    .e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+            const auto block_2_etile_map_ =
+                arg.contraction_multi_d_kernel_args_[i].block_2_etile_map_;
+
+            const auto a_mz_stride_  = arg.contraction_multi_d_device_args_[i].a_mz_stride_;
+            const auto a_kz_stride_  = arg.contraction_multi_d_device_args_[i].a_kz_stride_;
+            const auto b_nz_stride_  = arg.contraction_multi_d_device_args_[i].b_nz_stride_;
+            const auto b_kz_stride_  = arg.contraction_multi_d_device_args_[i].b_kz_stride_;
+            const auto ds_nz_stride_ = arg.contraction_multi_d_device_args_[i].ds_nz_stride_;
+            const auto e_nz_stride_  = arg.contraction_multi_d_device_args_[i].e_nz_stride_;
+
+            if(!GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                            b_grid_desc_n_k_,
+                                            ds_grid_desc_m_n_,
+                                            e_grid_desc_m_n_,
+                                            block_2_etile_map_))
+            {
+                return false;
+            }
+
+            // check vector access
+            static_assert((ABlockTransferSrcVectorDim == 1 || ABlockTransferSrcVectorDim == 2) &&
+                              (BBlockTransferSrcVectorDim == 1 || BBlockTransferSrcVectorDim == 2),
+                          "wrong!");
+
+            // vector memory access of A: could be on M or AK1 dimension
+            if constexpr(ABlockTransferSrcVectorDim == 1)
+            {
+                if(!(a_mz_stride_ == 1 &&
+                     a_grid_desc_ak0_m_ak1_.GetLength(I1) % ABlockTransferSrcScalarPerVector == 0))
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                if(!(a_kz_stride_ == 1 &&
+                     a_grid_desc_ak0_m_ak1_.GetLength(I2) % ABlockTransferSrcScalarPerVector == 0))
+                {
+                    return false;
+                }
+            }
+
+            // vector memory access of B: could be on N or BK1 dimension
+            if constexpr(BBlockTransferSrcVectorDim == 1)
+            {
+                if(!(b_nz_stride_ == 1 &&
+                     b_grid_desc_bk0_n_bk1_.GetLength(I1) % BBlockTransferSrcScalarPerVector == 0))
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                if(!(b_kz_stride_ == 1 &&
+                     b_grid_desc_bk0_n_bk1_.GetLength(I2) % BBlockTransferSrcScalarPerVector == 0))
+                {
+                    return false;
+                }
+            }
+
+            // vector memory access of Ds: always on NPerBlock dimension
+            bool valid_d_access = true;
+
+            static_for<0, NumDTensor, 1>{}([&](auto j) {
+                if(!(ds_nz_stride_[j] == 1 &&
+                     ds_grid_desc_mblock_mperblock_nblock_nperblock_[j].GetLength(I3) %
+                             CDEBlockTransferScalarPerVector_NPerBlock ==
+                         0))
+                {
+                    valid_d_access = false;
+                }
+            });
+
+            if(valid_d_access == false)
+            {
+                return false;
+            }
+
+            // vector memory access of E: always on NPerBlock dimension
+            if(!(e_nz_stride_ == 1 && e_grid_desc_mblock_mperblock_nblock_nperblock_.GetLength(I3) %
+                                              CDEBlockTransferScalarPerVector_NPerBlock ==
+                                          0))
+            {
+                return false;
+            }
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::vector<const void*> p_a_vec,
+                             std::vector<const void*> p_b_vec,
+                             std::vector<std::array<const void*, NumDTensor>> p_ds_vec,
+                             std::vector<void*> p_e_vec,
+                             std::vector<ContractionDesc<NumDTensor>> contraction_descs,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{p_a_vec,
+                        p_b_vec,
+                        p_ds_vec,
+                        p_e_vec,
+                        contraction_descs,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*> p_a_vec,
+                        std::vector<const void*> p_b_vec,
+                        std::vector<std::array<const void*, NumDTensor>> p_ds_vec,
+                        std::vector<void*> p_e_vec,
+                        std::vector<ContractionDesc<NumDTensor>> contraction_descs,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a_vec,
+                                          p_b_vec,
+                                          p_ds_vec,
+                                          p_e_vec,
+                                          contraction_descs,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedContractionMultipleD_Xdl_CShuffle"
+            << "<"
+            << NumDimM << ", "
+            << NumDimN << ", "
+            << NumDimK << ", "
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << ABlockTransferSrcVectorDim << ", "
+            << BBlockTransferSrcVectorDim
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
+    {
+        return dynamic_cast<const Argument*>(p_arg)->group_count_ *
+               sizeof(ContractionMultiDKernelArg);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
new file mode 100644
index 0000000000..c828269acf
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
@@ -0,0 +1,1081 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+template <index_t NumDTensor>
+struct ComputePtrOffsetOfStridedBatch
+{
+    ComputePtrOffsetOfStridedBatch() = default;
+
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                   index_t BatchStrideB,
+                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
+                                   index_t BatchStrideE)
+        : BatchStrideA_(BatchStrideA),
+          BatchStrideB_(BatchStrideB),
+          BatchStrideDs_(BatchStrideDs),
+          BatchStrideE_(BatchStrideE)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumDTensor> ds_offset;
+        static_for<0, NumDTensor, 1>{}(
+            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
+        return ds_offset;
+    }
+
+    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    Array<ck::index_t, NumDTensor> BatchStrideDs_;
+    index_t BatchStrideE_;
+};
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
+ * DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
+ * pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2ETileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_conv_bwd_data_multiple_d_xdl_cshuffle(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const index_t batch_count,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock_,
+            const Block2ETileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    // offset base pointer for each work-group
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_ds_grid_grp,
+                                                  p_e_grid + e_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+#endif
+}
+
+} // namespace
+
+// Conv backward data multiple D:
+//   input : output image A: [G, N, K, Ho, Wo]
+//   input : weight B: [G, K, C, Y, X],
+//   input : D0, D1, ... : [G, N, K, Ho, Wo]
+//   output : input image E: [G, N, C, Hi, Wi]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+template <index_t NDimSpatial,
+          typename ALayout,   // output image
+          typename BLayout,   // weight
+          typename DsLayout,  // bias
+          typename ELayout,   // input image
+          typename ADataType, // output image
+          typename BDataType, // weight
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,       // bias
+          typename EDataType,        // input image
+          typename AElementwiseOp,   // output image
+          typename BElementwiseOp,   // weight
+          typename CDEElementwiseOp, // C, bias, and input image
+          ConvolutionBackwardDataSpecialization ConvBackwardDataSpecialization,
+          bool DoPadGemmM,
+          bool DoPadGemmN,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
+    : public DeviceGroupedConvBwdDataMultipleD<NDimSpatial,
+                                               ALayout,    // output image
+                                               BLayout,    // weight
+                                               DsLayout,   // bias
+                                               ELayout,    // input image
+                                               ADataType,  // output image
+                                               BDataType,  // weight
+                                               DsDataType, // bias
+                                               EDataType,  // input image
+                                               AElementwiseOp,
+                                               BElementwiseOp,
+                                               CDEElementwiseOp>
+{
+    // FIXME
+    static_assert(NDimSpatial == 2 || NDimSpatial == 3,
+                  "wrong! only implemented for 2D and 3D now");
+
+    using DeviceOp = DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    // TODO make A/B datatype different
+    using ABDataType = ADataType;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto transform_conv_to_gemm =
+        TransformConvBwdDataToGemm_v1<NDimSpatial,
+                                      ConvBackwardDataSpecialization,
+                                      AK1,
+                                      BK1,
+                                      MPerBlock,
+                                      NPerBlock,
+                                      KPerBlock,
+                                      DoPadGemmM,
+                                      DoPadGemmN>{};
+
+    static auto GetDummyABDsEGridDescriptor()
+    {
+        const std::array<index_t, NDimSpatial + 3> dummy_tensor_lengths = {1};
+        const std::array<index_t, NDimSpatial + 3> dummy_tensor_strides = {1};
+        const std::array<index_t, NDimSpatial> dummy_spatial_lengths    = {1};
+
+        const auto a_grid_desc_ak0_m_ak1 =
+            transform_conv_to_gemm.template MakeADescriptor_AK0_M_AK1<ALayout>(
+                dummy_tensor_lengths,
+                dummy_tensor_strides,
+                dummy_tensor_lengths,
+                dummy_tensor_strides,
+                dummy_tensor_lengths,
+                dummy_tensor_strides,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths);
+
+        const auto b_grid_desc_bk0_n_bk1 =
+            transform_conv_to_gemm.template MakeBDescriptor_BK0_N_BK1<BLayout>(
+                dummy_tensor_lengths,
+                dummy_tensor_strides,
+                dummy_tensor_lengths,
+                dummy_tensor_strides,
+                dummy_tensor_lengths,
+                dummy_tensor_strides,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths,
+                dummy_spatial_lengths);
+
+        const auto ds_grid_desc_m_n = generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return transform_conv_to_gemm.template MakeCDescriptor_M_N<DLayout>(
+                    dummy_tensor_lengths,
+                    dummy_tensor_strides,
+                    dummy_tensor_lengths,
+                    dummy_tensor_strides,
+                    dummy_tensor_lengths,
+                    dummy_tensor_strides,
+                    dummy_spatial_lengths,
+                    dummy_spatial_lengths,
+                    dummy_spatial_lengths,
+                    dummy_spatial_lengths,
+                    dummy_spatial_lengths);
+            },
+            Number<NumDTensor>{});
+
+        const auto e_grid_desc_m_n =
+            transform_conv_to_gemm.template MakeCDescriptor_M_N<ELayout>(dummy_tensor_lengths,
+                                                                         dummy_tensor_strides,
+                                                                         dummy_tensor_lengths,
+                                                                         dummy_tensor_strides,
+                                                                         dummy_tensor_lengths,
+                                                                         dummy_tensor_strides,
+                                                                         dummy_spatial_lengths,
+                                                                         dummy_spatial_lengths,
+                                                                         dummy_spatial_lengths,
+                                                                         dummy_spatial_lengths,
+                                                                         dummy_spatial_lengths);
+
+        return make_tuple(
+            a_grid_desc_ak0_m_ak1, b_grid_desc_bk0_n_bk1, ds_grid_desc_m_n, e_grid_desc_m_n);
+    }
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ABDataType, // TODO: distinguish A/B datatype
+        ABDataType, // TODO: distinguish A/B datatype
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOp,
+        BElementwiseOp,
+        CDEElementwiseOp,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    template <typename Desc_K0_M_K1>
+    static auto transform_k0_m_k1_to_m_k(const Desc_K0_M_K1& desc_k0_m_k1)
+    {
+        const auto grid_desc_m_k = transform_tensor_descriptor(
+            desc_k0_m_k1,
+            make_tuple(make_pass_through_transform(desc_k0_m_k1.GetLength(I1)),
+                       make_merge_transform(
+                           make_tuple(desc_k0_m_k1.GetLength(I0), desc_k0_m_k1.GetLength(I2)))),
+            make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return grid_desc_m_k;
+    }
+
+    // desc
+    using ABDsEGridDesc = decltype(GetDummyABDsEGridDescriptor());
+
+    using AGridDesc_AK0_M_AK1 = remove_cvref_t<tuple_element_t<0, ABDsEGridDesc>>;
+    using BGridDesc_BK0_N_BK1 = remove_cvref_t<tuple_element_t<1, ABDsEGridDesc>>;
+    using DsGridDesc_M_N      = remove_cvref_t<tuple_element_t<2, ABDsEGridDesc>>;
+    using EGridDesc_M_N       = remove_cvref_t<tuple_element_t<3, ABDsEGridDesc>>;
+
+    using AGridDesc_M_K = decltype(transform_k0_m_k1_to_m_k(AGridDesc_AK0_M_AK1{}));
+    using BGridDesc_N_K = decltype(transform_k0_m_k1_to_m_k(BGridDesc_BK0_N_BK1{}));
+
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}));
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}));
+
+    // block-to-e-tile map
+    using Block2ETileMap =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a,                                 // output image
+                 const void* p_b,                                 // weight
+                 const std::array<const void*, NumDTensor>& p_ds, // bias
+                 void* p_e,                                       // input image
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_c_wis_lengths,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOp& a_element_op,
+                 const BElementwiseOp& b_element_op,
+                 const CDEElementwiseOp& cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a)},
+              p_b_grid_{static_cast<const BDataType*>(p_b)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e)},
+              num_group_{a_g_n_k_wos_lengths[0]},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_g_n_k_wos_lengths_{a_g_n_k_wos_lengths},
+              a_g_n_k_wos_strides_{a_g_n_k_wos_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              ds_g_n_c_wis_lengths_{ds_g_n_c_wis_lengths},
+              ds_g_n_c_wis_strides_{ds_g_n_c_wis_strides},
+              e_g_n_c_wis_lengths_{e_g_n_c_wis_lengths},
+              e_g_n_c_wis_strides_{e_g_n_c_wis_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // populate Ds pointer
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
+            });
+
+            // A/B/Ds/E Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_k_wos_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_c_wis_strides[0];
+
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_c_wis_strides[i][0];
+            });
+
+            static constexpr auto NonSpatialDimsNum = Number<3>{};
+
+            static constexpr auto DIdx = Number<NonSpatialDimsNum>{};
+            static constexpr auto HIdx =
+                NDimSpatial == 2 ? Number<NonSpatialDimsNum>{} : Number<NonSpatialDimsNum + 1>{};
+            static constexpr auto WIdx = NDimSpatial == 2 ? Number<NonSpatialDimsNum + 1>{}
+                                                          : Number<NonSpatialDimsNum + 2>{};
+
+            static constexpr auto ZIdx = Number<NonSpatialDimsNum>{};
+            static constexpr auto YIdx =
+                NDimSpatial == 2 ? Number<NonSpatialDimsNum>{} : Number<NonSpatialDimsNum + 1>{};
+            static constexpr auto XIdx = NDimSpatial == 2 ? Number<NonSpatialDimsNum + 1>{}
+                                                          : Number<NonSpatialDimsNum + 2>{};
+
+            // problem definition
+            const index_t Z = b_g_k_c_xs_lengths[ZIdx];
+            const index_t Y = b_g_k_c_xs_lengths[YIdx];
+            const index_t X = b_g_k_c_xs_lengths[XIdx];
+
+            const index_t ConvStrideD = conv_filter_strides[DIdx - NonSpatialDimsNum];
+            const index_t ConvStrideH = conv_filter_strides[HIdx - NonSpatialDimsNum];
+            const index_t ConvStrideW = conv_filter_strides[WIdx - NonSpatialDimsNum];
+
+            const index_t ConvDilationD = conv_filter_dilations[DIdx - NonSpatialDimsNum];
+            const index_t ConvDilationH = conv_filter_dilations[HIdx - NonSpatialDimsNum];
+            const index_t ConvDilationW = conv_filter_dilations[WIdx - NonSpatialDimsNum];
+
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto ZTilde = NDimSpatial == 3 ? ConvStrideD / GcdStrideDilationD : 1;
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            for(index_t i_ztilde = 0; i_ztilde < ZTilde; ++i_ztilde)
+            {
+
+                for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
+                {
+                    for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
+                    {
+                        // check slice is valid
+                        const auto ZDotSlice =
+                            NDimSpatial == 3 ? math::integer_divide_ceil(Z - i_ztilde, ZTilde) : 1;
+                        const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+                        const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+                        if(YDotSlice * XDotSlice * ZDotSlice <= 0)
+                        {
+                            continue;
+                        }
+
+                        std::array<index_t, NDimSpatial> tildes;
+                        if constexpr(NDimSpatial == 2)
+                        {
+                            tildes = {i_ytilde, i_xtilde};
+                        }
+                        else if constexpr(NDimSpatial == 3)
+                        {
+                            tildes = {i_ztilde, i_ytilde, i_xtilde};
+                        }
+                        else
+                        {
+                            throw std::runtime_error("wrong! only implemented for 2D and 3D now");
+                        }
+
+                        const auto a_grid_desc_ak0_m_ak1 =
+                            transform_conv_to_gemm.template MakeADescriptor_AK0_M_AK1<ALayout>(
+                                a_g_n_k_wos_lengths,
+                                a_g_n_k_wos_strides,
+                                b_g_k_c_xs_lengths,
+                                b_g_k_c_xs_strides,
+                                e_g_n_c_wis_lengths,
+                                e_g_n_c_wis_strides,
+                                conv_filter_strides,
+                                conv_filter_dilations,
+                                input_left_pads,
+                                input_right_pads,
+                                tildes);
+
+                        const auto b_grid_desc_bk0_n_bk1 =
+                            transform_conv_to_gemm.template MakeBDescriptor_BK0_N_BK1<BLayout>(
+                                a_g_n_k_wos_lengths,
+                                a_g_n_k_wos_strides,
+                                b_g_k_c_xs_lengths,
+                                b_g_k_c_xs_strides,
+                                e_g_n_c_wis_lengths,
+                                e_g_n_c_wis_strides,
+                                conv_filter_strides,
+                                conv_filter_dilations,
+                                input_left_pads,
+                                input_right_pads,
+                                tildes);
+
+                        DsGridDesc_M_N ds_grid_desc_m_n;
+
+                        // populate Ds desc
+                        static_for<0, NumDTensor, 1>{}([&](auto i) {
+                            using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                            ds_grid_desc_m_n(i) =
+                                transform_conv_to_gemm.template MakeCDescriptor_M_N<DLayout>(
+                                    a_g_n_k_wos_lengths,
+                                    a_g_n_k_wos_strides,
+                                    b_g_k_c_xs_lengths,
+                                    b_g_k_c_xs_strides,
+                                    ds_g_n_c_wis_lengths[i],
+                                    ds_g_n_c_wis_strides[i],
+                                    conv_filter_strides,
+                                    conv_filter_dilations,
+                                    input_left_pads,
+                                    input_right_pads,
+                                    tildes);
+                        });
+
+                        const auto e_grid_desc_m_n =
+                            transform_conv_to_gemm.template MakeCDescriptor_M_N<ELayout>(
+                                a_g_n_k_wos_lengths,
+                                a_g_n_k_wos_strides,
+                                b_g_k_c_xs_lengths,
+                                b_g_k_c_xs_strides,
+                                e_g_n_c_wis_lengths,
+                                e_g_n_c_wis_strides,
+                                conv_filter_strides,
+                                conv_filter_dilations,
+                                input_left_pads,
+                                input_right_pads,
+                                tildes);
+
+                        // desc for problem definition
+                        const auto a_grid_desc_m_k =
+                            transform_k0_m_k1_to_m_k(a_grid_desc_ak0_m_ak1);
+                        const auto b_grid_desc_n_k =
+                            transform_k0_m_k1_to_m_k(b_grid_desc_bk0_n_bk1);
+
+                        a_grid_desc_m_k_container_.push_back(a_grid_desc_m_k);
+                        b_grid_desc_n_k_container_.push_back(b_grid_desc_n_k);
+                        ds_grid_desc_m_n_container_.push_back(ds_grid_desc_m_n);
+                        e_grid_desc_m_n_container_.push_back(e_grid_desc_m_n);
+
+                        // desc for blockwise copy
+                        a_grid_desc_ak0_m_ak1_container_.push_back(a_grid_desc_ak0_m_ak1);
+                        b_grid_desc_bk0_n_bk1_container_.push_back(b_grid_desc_bk0_n_bk1);
+
+                        // block-to-e-tile-map
+                        auto block_2_etile_map =
+                            GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n);
+
+                        block_2_etile_map_container_.push_back(block_2_etile_map);
+
+                        if(GridwiseGemm::CheckValidity(a_grid_desc_m_k,
+                                                       b_grid_desc_n_k,
+                                                       ds_grid_desc_m_n,
+                                                       e_grid_desc_m_n,
+                                                       block_2_etile_map))
+                        {
+                            ds_grid_desc_mblock_mperblock_nblock_nperblock_container_.push_back(
+                                GridwiseGemm::
+                                    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                                        ds_grid_desc_m_n));
+
+                            e_grid_desc_mblock_mperblock_nblock_nperblock_container_.push_back(
+                                GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                                    e_grid_desc_m_n));
+                        }
+                    }
+                }
+            }
+        }
+
+        void Print() const
+        {
+            for(std::size_t i = 0; i < a_grid_desc_ak0_m_ak1_container_.size(); i++)
+            {
+                std::cout << "a_grid_desc_ak0_m_ak1_container_"
+                          << a_grid_desc_ak0_m_ak1_container_[i] << std::endl;
+
+                std::cout << "b_grid_desc_bk0_n_bk1_container_"
+                          << b_grid_desc_bk0_n_bk1_container_[i] << std::endl;
+
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    std::cout << "ds_grid_desc_mblock_mperblock_nblock_nperblock_container_"
+                              << ds_grid_desc_mblock_mperblock_nblock_nperblock_container_[i][j]
+                              << std::endl;
+                });
+
+                std::cout << "e_grid_desc_mblock_mperblock_nblock_nperblock_container_"
+                          << e_grid_desc_mblock_mperblock_nblock_nperblock_container_[i]
+                          << std::endl;
+            }
+        }
+
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptor for problem definition
+        index_t num_group_;
+        std::vector<AGridDesc_M_K> a_grid_desc_m_k_container_;
+        std::vector<BGridDesc_N_K> b_grid_desc_n_k_container_;
+        std::vector<DsGridDesc_M_N> ds_grid_desc_m_n_container_;
+        std::vector<EGridDesc_M_N> e_grid_desc_m_n_container_;
+
+        // tensor descriptor for block-wise copy
+        std::vector<AGridDesc_AK0_M_AK1> a_grid_desc_ak0_m_ak1_container_;
+        std::vector<BGridDesc_BK0_N_BK1> b_grid_desc_bk0_n_bk1_container_;
+        std::vector<DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_container_;
+        std::vector<EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>
+            e_grid_desc_mblock_mperblock_nblock_nperblock_container_;
+
+        // block-to-e-tile map
+        std::vector<Block2ETileMap> block_2_etile_map_container_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch<NumDTensor> compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOp a_element_op_;
+        BElementwiseOp b_element_op_;
+        CDEElementwiseOp cde_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_c_wis_lengths_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            float ave_time = 0;
+
+            for(std::size_t i = 0; i < arg.a_grid_desc_ak0_m_ak1_container_.size(); i++)
+            {
+                if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_container_[i],
+                                                arg.b_grid_desc_n_k_container_[i],
+                                                arg.ds_grid_desc_m_n_container_[i],
+                                                arg.e_grid_desc_m_n_container_[i],
+                                                arg.block_2_etile_map_container_[i]))
+                {
+                    throw std::runtime_error("wrong! device_op has invalid setting");
+                }
+
+                const index_t grid_size = arg.block_2_etile_map_container_[i].CalculateGridSize(
+                                              arg.e_grid_desc_m_n_container_[i]) *
+                                          arg.num_group_;
+
+                const auto GemmK = arg.a_grid_desc_m_k_container_[i].GetLength(I1);
+
+                auto launch_kernel = [&](auto has_main_k_block_loop) {
+                    constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                    const auto kernel = kernel_grouped_conv_bwd_data_multiple_d_xdl_cshuffle<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        typename GridwiseGemm::DsGridPointer,
+                        EDataType,
+                        AElementwiseOp,
+                        BElementwiseOp,
+                        CDEElementwiseOp,
+                        DeviceOp::AGridDesc_AK0_M_AK1,
+                        DeviceOp::BGridDesc_BK0_N_BK1,
+                        DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                        DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                        Block2ETileMap,
+                        ComputePtrOffsetOfStridedBatch<NumDTensor>,
+                        has_main_loop>;
+
+                    return launch_and_time_kernel(
+                        stream_config,
+                        kernel,
+                        dim3(grid_size),
+                        dim3(BlockSize),
+                        0,
+                        arg.p_a_grid_,
+                        arg.p_b_grid_,
+                        arg.p_ds_grid_,
+                        arg.p_e_grid_,
+                        arg.a_element_op_,
+                        arg.b_element_op_,
+                        arg.cde_element_op_,
+                        arg.a_g_n_k_wos_lengths_[0], // Group count
+                        arg.a_grid_desc_ak0_m_ak1_container_[i],
+                        arg.b_grid_desc_bk0_n_bk1_container_[i],
+                        arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_container_[i],
+                        arg.e_grid_desc_mblock_mperblock_nblock_nperblock_container_[i],
+                        arg.block_2_etile_map_container_[i],
+                        arg.compute_ptr_offset_of_batch_);
+                };
+
+                if(GridwiseGemm::CalculateHasMainKBlockLoop(GemmK))
+                {
+                    ave_time += launch_kernel(integral_constant<bool, true>{});
+                }
+                else
+                {
+                    ave_time += launch_kernel(integral_constant<bool, false>{});
+                }
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        const index_t ConvK = arg.b_g_k_c_xs_lengths_[1];
+        const index_t ConvC = arg.b_g_k_c_xs_lengths_[2];
+
+        // Specifialization
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 pad = 0 conv
+            for(int i = 0; i < NDimSpatial; i++)
+            {
+                if(!(arg.b_g_k_c_xs_lengths_[3 + i] == 1 && arg.conv_filter_strides_[i] == 1 &&
+                     arg.input_left_pads_[i] == 0 && arg.input_right_pads_[i] == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // vector load for A matrix from global memory to LDS
+        if constexpr(is_same_v<ALayout, tensor_layout::convolution::GNHWK> ||
+                     is_same_v<ALayout, tensor_layout::convolution::GNDHWK> ||
+                     is_same_v<ALayout, tensor_layout::convolution::NHWGK> ||
+                     is_same_v<ALayout, tensor_layout::convolution::NDHWGK>)
+        {
+            if(!(ABlockTransferSrcVectorDim == 2 && ConvK % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // vector load for B matrix from global memory to LDS
+        if constexpr(is_same_v<BLayout, tensor_layout::convolution::GKYXC> ||
+                     is_same_v<BLayout, tensor_layout::convolution::GKZYXC>)
+        {
+            if(!(BBlockTransferSrcVectorDim == 1 && ConvC % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // vector store for Ds
+        bool ds_valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+            if constexpr(is_same_v<DLayout, tensor_layout::convolution::GNHWC> ||
+                         is_same_v<DLayout, tensor_layout::convolution::GNDHWC> ||
+                         is_same_v<DLayout, tensor_layout::convolution::NHWGC> ||
+                         is_same_v<DLayout, tensor_layout::convolution::NDHWGC> ||
+                         is_same_v<DLayout, tensor_layout::convolution::G_NHW_C> ||
+                         is_same_v<DLayout, tensor_layout::convolution::GC> ||
+                         is_same_v<DLayout, tensor_layout::convolution::G_C>)
+            {
+                // vector load D matrix from global memory
+                if(!(ConvC % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+                {
+                    ds_valid = false;
+                }
+            }
+            else
+            {
+                ds_valid = false;
+            }
+        });
+
+        if(!ds_valid)
+        {
+            return false;
+        }
+
+        // vector store for E
+        if constexpr(is_same_v<ELayout, tensor_layout::convolution::GNHWC> ||
+                     is_same_v<ELayout, tensor_layout::convolution::GNDHWC> ||
+                     is_same_v<ELayout, tensor_layout::convolution::NHWGC> ||
+                     is_same_v<ELayout, tensor_layout::convolution::NDHWGC>)
+        {
+            // vector store C matrix into global memory
+            if(!(ConvC % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        for(std::size_t i = 0; i < arg.a_grid_desc_ak0_m_ak1_container_.size(); i++)
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_container_[i],
+                                            arg.b_grid_desc_n_k_container_[i],
+                                            arg.ds_grid_desc_m_n_container_[i],
+                                            arg.e_grid_desc_m_n_container_[i],
+                                            arg.block_2_etile_map_container_[i]))
+            {
+                return false;
+            }
+        }
+
+        return true;
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto
+    MakeArgument(const void* p_a,                                                 // output image
+                 const void* p_b,                                                 // weight
+                 const std::array<const void*, NumDTensor>& p_ds,                 // bias
+                 void* p_e,                                                       // input image
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_lengths, // output image
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_strides, // output image
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,  // weight
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,  // weight
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_c_wis_lengths, // bias
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_c_wis_strides,                                        // bias
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_lengths, // input image
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_strides, // input image
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOp& a_element_op,
+                 const BElementwiseOp& b_element_op,
+                 const CDEElementwiseOp& cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_g_n_k_wos_lengths,
+                        a_g_n_k_wos_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        ds_g_n_c_wis_lengths,
+                        ds_g_n_c_wis_strides,
+                        e_g_n_c_wis_lengths,
+                        e_g_n_c_wis_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,                                                 // output image
+        const void* p_b,                                                 // weight
+        const std::array<const void*, NumDTensor>& p_ds,                 // bias
+        void* p_e,                                                       // input image
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_lengths, // output image
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_k_wos_strides, // output image
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,  // weight
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,  // weight
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+            ds_g_n_c_wis_lengths, // bias
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+            ds_g_n_c_wis_strides,                                        // bias
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_lengths, // input image
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_c_wis_strides, // input image
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOp& a_element_op,
+        const BElementwiseOp& b_element_op,
+        const CDEElementwiseOp& cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_g_n_k_wos_lengths,
+                                          a_g_n_k_wos_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          ds_g_n_c_wis_lengths,
+                                          ds_g_n_c_wis_strides,
+                                          e_g_n_c_wis_lengths,
+                                          e_g_n_c_wis_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << getConvBackwardDataSpecializationString(ConvBackwardDataSpecialization) << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle
+            << ">";
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_dl.hpp
new file mode 100644
index 0000000000..198751cdf3
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_dl.hpp
@@ -0,0 +1,1226 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+struct ComputePtrOffsetOfStridedBatch
+{
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideC_);
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    index_t BatchStrideC_;
+};
+
+} // namespace
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_B_K0_M0_M1_K1,
+          typename BGridDesc_B_K0_N0_N1_K1,
+          typename CGridDesc_M0_M10_M11_N0_N10_N11,
+          typename Block2CTileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop,
+          bool HasDoubleTailKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_dlops_bwd_weight(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const index_t batch_count,
+            const AGridDesc_B_K0_M0_M1_K1 a_grid_desc_kbatch_k0_m0_m1_k1,
+            const BGridDesc_B_K0_N0_N1_K1 b_grid_desc_kbatch_k0_n0_n1_k1,
+            const CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11,
+            const Block2CTileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
+
+    __shared__ FloatAB p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB)];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop, HasDoubleTailKBlockLoop>(
+        p_a_grid + a_batch_offset,
+        p_b_grid + b_batch_offset,
+        p_c_grid + c_batch_offset,
+        p_shared,
+        a_grid_desc_kbatch_k0_m0_m1_k1,
+        b_grid_desc_kbatch_k0_n0_n1_k1,
+        c_grid_desc_m0_m10_m11_n0_n10_n11,
+        block_2_ctile_map,
+        integral_constant<bool, HasMainKBlockLoop>{},
+        integral_constant<bool, HasDoubleTailKBlockLoop>{});
+}
+
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ConvolutionBackwardWeightSpecialization ConvBackwardWeightSpecialization,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          index_t M1PerThread,
+          index_t N1PerThread,
+          index_t KPerThread,
+          typename M1N1ThreadClusterM1Xs,
+          typename M1N1ThreadClusterN1Xs,
+          typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+          typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+          typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+          typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+          typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector>
+struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
+    : public DeviceGroupedConvBwdWeight<
+          NDimSpatial,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::GNWC,
+                                        ck::tensor_layout::convolution::GNHWC,
+                                        ck::tensor_layout::convolution::GNDHWC>>,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::GKXC,
+                                        ck::tensor_layout::convolution::GKYXC,
+                                        ck::tensor_layout::convolution::GKZYXC>>,
+          ck::tuple_element_t<NDimSpatial - 1,
+                              ck::Tuple<ck::tensor_layout::convolution::GNWK,
+                                        ck::tensor_layout::convolution::GNHWK,
+                                        ck::tensor_layout::convolution::GNDHWK>>,
+          InDataType,
+          WeiDataType,
+          OutDataType,
+          InElementwiseOperation,
+          WeiElementwiseOperation,
+          OutElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl;
+
+    using ADataType = OutDataType;
+    using BDataType = InDataType;
+    using CDataType = WeiDataType;
+
+    using AElementwiseOperation = OutElementwiseOperation;
+    using BElementwiseOperation = InElementwiseOperation;
+    using CElementwiseOperation = WeiElementwiseOperation;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    // Bytes per 32 lds bank: 32 * 4 bytes
+    static constexpr auto BankLength = 128;
+    static constexpr auto ElePerBank = BankLength / sizeof(ADataType);
+
+    // M1 & M0
+    static constexpr auto ABlockLdsM1PerBlock = ElePerBank / K1;
+    static constexpr auto ABlockLdsM0PerBlock = MPerBlock / ABlockLdsM1PerBlock;
+    static constexpr auto ABlockLdsM1Padding  = 4;
+
+    // N1 & N0
+    static constexpr auto BBlockLdsN1PerBlock = ElePerBank / K1;
+    static constexpr auto BBlockLdsN0PerBlock = NPerBlock / BBlockLdsN1PerBlock;
+    static constexpr auto BBlockLdsN1Padding  = 4;
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+    static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        const ck::index_t N,
+        const ck::index_t K,
+        const ck::index_t C,
+        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+        const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+        const ck::index_t batch_k)
+    {
+        using namespace ck;
+
+        const index_t Wi            = input_spatial_lengths[0];
+        const index_t Wo            = output_spatial_lengths[0];
+        const index_t X             = filter_spatial_lengths[0];
+        const index_t InLeftPadW    = input_left_pads[0];
+        const index_t InRightPadW   = input_right_pads[0];
+        const index_t ConvStrideW   = conv_filter_strides[0];
+        const index_t ConvDilationW = conv_filter_dilations[0];
+
+        const index_t GemmKTotal = N * Wo;
+        const index_t GemmM      = K;
+        const index_t GemmN      = C * X;
+
+        const index_t GemmKBatch = batch_k;
+        const index_t GemmK0 =
+            math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
+            K0PerBlock;
+        const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
+
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmktotal_gemmm_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Wo, K));
+
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_gemmktotal_gemmm_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_gemmktotal_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Wi, C));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // C: weights tensor
+            const auto wei_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, X * C));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto out_gemmktotal_gemmm_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Wo, K));
+            const auto in_n_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
+
+            // A: output tensor
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_gemmktotal_gemmm_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_n_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_n_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_gemmktotal_gemmn_grid_desc =
+                transform_tensor_descriptor(in_n_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(X, C)),
+                                                       make_merge_transform(make_tuple(N, Wo))),
+                                            make_tuple(Sequence<1, 3>{}, Sequence<0, 2>{}),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // C: weight tensor
+            const auto wei_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, X * C));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        const ck::index_t N,
+        const ck::index_t K,
+        const ck::index_t C,
+        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+        const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+        const ck::index_t batch_k)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t GemmKTotal = N * Ho * Wo;
+        const index_t GemmM      = K;
+        const index_t GemmN      = C * X * Y;
+
+        const index_t GemmKBatch = batch_k;
+        const index_t GemmK0 =
+            math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
+            K0PerBlock;
+        const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
+
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmktotal_gemmm_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_gemmktotal_gemmm_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_gemmktotal_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Hi * Wi, C));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // C: weight tensor
+            const auto wei_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto out_gemmktotal_gemmm_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            // A: output tensor
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_gemmktotal_gemmm_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmktotal_gemmn_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // C: weight tensor
+            const auto wei_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        const ck::index_t N,
+        const ck::index_t K,
+        const ck::index_t C,
+        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+        const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+        const ck::index_t batch_k)
+    {
+        using namespace ck;
+
+        const index_t Di = input_spatial_lengths[0];
+        const index_t Hi = input_spatial_lengths[1];
+        const index_t Wi = input_spatial_lengths[2];
+
+        const index_t Do = output_spatial_lengths[0];
+        const index_t Ho = output_spatial_lengths[1];
+        const index_t Wo = output_spatial_lengths[2];
+
+        const index_t Z = filter_spatial_lengths[0];
+        const index_t Y = filter_spatial_lengths[1];
+        const index_t X = filter_spatial_lengths[2];
+
+        const index_t InLeftPadD = input_left_pads[0];
+        const index_t InLeftPadH = input_left_pads[1];
+        const index_t InLeftPadW = input_left_pads[2];
+
+        const index_t InRightPadD = input_right_pads[0];
+        const index_t InRightPadH = input_right_pads[1];
+        const index_t InRightPadW = input_right_pads[2];
+
+        const index_t ConvStrideD = conv_filter_strides[0];
+        const index_t ConvStrideH = conv_filter_strides[1];
+        const index_t ConvStrideW = conv_filter_strides[2];
+
+        const index_t ConvDilationD = conv_filter_dilations[0];
+        const index_t ConvDilationH = conv_filter_dilations[1];
+        const index_t ConvDilationW = conv_filter_dilations[2];
+
+        const index_t GemmKTotal = N * Do * Ho * Wo;
+        const index_t GemmM      = K;
+        const index_t GemmN      = C * Z * X * Y;
+
+        const index_t GemmKBatch = batch_k;
+        const index_t GemmK0 =
+            math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
+            K0PerBlock;
+        const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
+
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmktotal_gemmm_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K));
+
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_gemmktotal_gemmm_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_gemmktotal_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Di * Hi * Wi, C));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // C: weight tensor
+            const auto wei_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Z * Y * X * C));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto out_gemmktotal_gemmm_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K));
+            const auto in_n_di_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
+
+            // A: output tensor
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_gemmktotal_gemmm_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_n_dip_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Di, InLeftPadD, InRightPadD),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_n_z_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_dip_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+            const auto in_gemmktotal_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_z_do_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(Z, Y, X, C)),
+                           make_merge_transform(make_tuple(N, Do, Ho, Wo))),
+                make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // C: weight tensor
+            const auto wei_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, Z * Y * X * C));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_gemmm_gemmn_grid_desc);
+        }
+
+    } // function end
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<1>(
+            1, 1, 1, {1}, {1}, {1}, {1}, {1}, {1}, {1}, 1);
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<2>(
+            1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, 1);
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<3>(1,
+                                                                  1,
+                                                                  1,
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  1);
+    }
+
+    using ABCGridDescs = decltype(GetABCGridDesc<NDimSpatial>());
+
+    using AGridDesc_B_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_B_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N       = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    using GridwiseGemm =
+        GridwiseGemmDl_bkm_bkn_mn_v1r3<BlockSize,
+                                       ADataType,
+                                       AccDataType,
+                                       CDataType,
+                                       InMemoryDataOperationEnum::Set,
+                                       AGridDesc_B_K0_M_K1,
+                                       BGridDesc_B_K0_N_K1,
+                                       CGridDesc_M_N,
+                                       MPerBlock,
+                                       NPerBlock,
+                                       K0PerBlock,
+                                       K1,
+                                       M1PerThread,
+                                       N1PerThread,
+                                       KPerThread,
+                                       M1N1ThreadClusterM1Xs,
+                                       M1N1ThreadClusterN1Xs,
+                                       ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+                                       ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+                                       ABlockTransferThreadClusterArrangeOrder,
+                                       ABlockTransferSrcAccessOrder,
+                                       ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+                                       ABlockTransferSrcVectorTensorContiguousDimOrder,
+                                       ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+                                       BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+                                       BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+                                       BBlockTransferThreadClusterArrangeOrder,
+                                       BBlockTransferSrcAccessOrder,
+                                       BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+                                       BBlockTransferSrcVectorTensorContiguousDimOrder,
+                                       BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+                                       CThreadTransferSrcDstAccessOrder,
+                                       CThreadTransferSrcDstVectorDim,
+                                       CThreadTransferDstScalarPerVector>;
+
+    // Argument
+    using AGridDesc_B_K0_M0_M1_K1 =
+        decltype(GridwiseGemm::MakeAGridDescriptor_B_K0_M0_M1_K1(AGridDesc_B_K0_M_K1{}));
+    using BGridDesc_B_K0_N0_N1_K1 =
+        decltype(GridwiseGemm::MakeBGridDescriptor_B_K0_N0_N1_K1(BGridDesc_B_K0_N_K1{}));
+    using CGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
+    using Block2CTileMap =
+        decltype(GridwiseGemm::MakeCBlockClusterAdaptor(CGridDesc_M_N{}, 1, 1, 1));
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 WeiDataType* p_wei_grid,
+                 const OutDataType* p_out_grid,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+                 const std::array<index_t, NDimSpatial + 3>& /*a_g_n_c_wis_strides*/,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+                 const std::array<index_t, NDimSpatial + 3>& /*b_g_k_c_xs_strides*/,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+                 const std::array<index_t, NDimSpatial + 3>& /*e_g_n_k_wos_strides*/,
+                 const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+                 const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op,
+                 ck::index_t split_k)
+            : p_a_grid_{p_out_grid},
+              p_b_grid_{p_in_grid},
+              p_c_grid_{p_wei_grid},
+              a_grid_desc_kbatch_k0_m_k1_{},
+              b_grid_desc_kbatch_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              block_2_ctile_map_{},
+              compute_ptr_offset_of_batch_{},
+              a_element_op_{out_element_op},
+              b_element_op_{wei_element_op},
+              c_element_op_{in_element_op},
+              Conv_G_{a_g_n_c_wis_lengths[0]},
+              Conv_N_{a_g_n_c_wis_lengths[1]},
+              Conv_K_{b_g_k_c_xs_lengths[1]},
+              Conv_C_{a_g_n_c_wis_lengths[2]},
+              input_spatial_lengths_{},
+              filter_spatial_lengths_{},
+              output_spatial_lengths_{},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads},
+              k_batch_{split_k}
+        {
+            constexpr index_t spatial_offset = 3;
+            std::copy(begin(a_g_n_c_wis_lengths) + spatial_offset,
+                      end(a_g_n_c_wis_lengths),
+                      begin(input_spatial_lengths_));
+            std::copy(begin(b_g_k_c_xs_lengths) + spatial_offset,
+                      end(b_g_k_c_xs_lengths),
+                      begin(filter_spatial_lengths_));
+            std::copy(begin(e_g_n_k_wos_lengths) + spatial_offset,
+                      end(e_g_n_k_wos_lengths),
+                      begin(output_spatial_lengths_));
+
+            const auto descs =
+                DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
+                    Conv_N_,
+                    Conv_K_,
+                    Conv_C_,
+                    input_spatial_lengths_,
+                    filter_spatial_lengths_,
+                    output_spatial_lengths_,
+                    conv_filter_strides,
+                    conv_filter_dilations,
+                    input_left_pads,
+                    input_right_pads,
+                    k_batch_);
+
+            a_grid_desc_kbatch_k0_m_k1_ = descs[I0];
+            b_grid_desc_kbatch_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_            = descs[I2];
+
+            a_grid_desc_kbatch_k0_m0_m1_k1_ =
+                GridwiseGemm::MakeAGridDescriptor_B_K0_M0_M1_K1(a_grid_desc_kbatch_k0_m_k1_);
+            b_grid_desc_kbatch_k0_n0_n1_k1_ =
+                GridwiseGemm::MakeBGridDescriptor_B_K0_N0_N1_K1(b_grid_desc_kbatch_k0_n_k1_);
+            c_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(c_grid_desc_m_n_);
+            ck::index_t M01 = 1;
+            ck::index_t N01 = 1;
+            block_2_ctile_map_ =
+                GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_);
+
+            // A/B/C Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ =
+                Conv_N_ * Conv_K_ *
+                std::accumulate(begin(output_spatial_lengths_),
+                                end(output_spatial_lengths_),
+                                index_t{1},
+                                std::multiplies<>{});
+            compute_ptr_offset_of_batch_.BatchStrideB_ =
+                Conv_N_ * Conv_C_ *
+                std::accumulate(begin(input_spatial_lengths_),
+                                end(input_spatial_lengths_),
+                                index_t{1},
+                                std::multiplies<>{});
+            compute_ptr_offset_of_batch_.BatchStrideC_ =
+                Conv_K_ * Conv_C_ *
+                std::accumulate(begin(filter_spatial_lengths_),
+                                end(filter_spatial_lengths_),
+                                index_t{1},
+                                std::multiplies<>{});
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+
+        AGridDesc_B_K0_M_K1 a_grid_desc_kbatch_k0_m_k1_;
+        BGridDesc_B_K0_N_K1 b_grid_desc_kbatch_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+
+        AGridDesc_B_K0_M0_M1_K1 a_grid_desc_kbatch_k0_m0_m1_k1_;
+        BGridDesc_B_K0_N0_N1_K1 b_grid_desc_kbatch_k0_n0_n1_k1_;
+        CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11_;
+
+        // DefaultBlock2CTileMap block_2_ctile_map_;
+        Block2CTileMap block_2_ctile_map_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        OutElementwiseOperation a_element_op_;
+        WeiElementwiseOperation b_element_op_;
+        InElementwiseOperation c_element_op_;
+
+        // for checking IsSupportedArgument()
+        const index_t Conv_G_;
+        const index_t Conv_N_;
+        const index_t Conv_K_;
+        const index_t Conv_C_;
+
+        std::array<ck::index_t, NDimSpatial> input_spatial_lengths_;
+        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths_;
+        std::array<ck::index_t, NDimSpatial> output_spatial_lengths_;
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides_;
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations_;
+        const std::array<ck::index_t, NDimSpatial>& input_left_pads_;
+        const std::array<ck::index_t, NDimSpatial>& input_right_pads_;
+        index_t k_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        void ShowInfo(const Argument& arg)
+        {
+            std::cout << "arg.a_grid_desc_kbatch_k0_m_k1_{"
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I2) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I3) << "}" << std::endl;
+
+            std::cout << "arg.b_grid_desc_kbatch_k0_n_k1_{"
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I0) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I1) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I2) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I3) << "}" << std::endl;
+
+            std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                      << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+        }
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+
+            ShowInfo(arg);
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
+                                            arg.b_grid_desc_kbatch_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm GridwiseGemmDl_bkm_bkn_mn_v1r3 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.Conv_G_;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop,
+                                     auto has_double_tail_k_block_loop) {
+                constexpr bool has_main_loop   = has_main_k_block_loop.value;
+                constexpr bool has_double_loop = has_double_tail_k_block_loop.value;
+
+                const auto kernel = kernel_batched_gemm_dlops_bwd_weight<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    remove_reference_t<DeviceOp::AGridDesc_B_K0_M0_M1_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_B_K0_N0_N1_K1>,
+                    remove_reference_t<DeviceOp::CGridDesc_M0_M10_M11_N0_N10_N11>,
+                    remove_reference_t<DeviceOp::Block2CTileMap>,
+                    ComputePtrOffsetOfStridedBatch,
+                    has_main_loop,
+                    has_double_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_c_grid_,
+                                              arg.Conv_G_,
+                                              arg.a_grid_desc_kbatch_k0_m0_m1_k1_,
+                                              arg.b_grid_desc_kbatch_k0_n0_n1_k1_,
+                                              arg.c_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                              arg.block_2_ctile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            const auto K0                    = arg.a_grid_desc_kbatch_k0_m0_m1_k1_.GetLength(I1);
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
+            const bool has_double_tail_k_block_loop =
+                GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
+
+            if(has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, true>{});
+            }
+            else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, false>{});
+            }
+            else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, false>{});
+            }
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        // check device
+        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
+             ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
+             ck::get_device_name() == "gfx1102"))
+        {
+            return false;
+        }
+
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 pad = 0 conv
+            for(int i = 0; i < NDimSpatial; i++)
+            {
+                if(!(arg.filter_spatial_lengths_[i] == 1 && arg.conv_filter_strides_[i] == 1 &&
+                     arg.input_left_pads_[i] == 0 && arg.input_right_pads_[i] == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // matrix A
+        {
+            auto srcVectorLengths = ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1{};
+            if(srcVectorLengths[I2] != 1 || srcVectorLengths[I3] != 1)
+            {
+                return false;
+            }
+            if(K1 % srcVectorLengths[I4] != 0 || K0PerBlock % srcVectorLengths[I1] != 0)
+            {
+                return false;
+            }
+
+            const index_t K = arg.Conv_K_;
+
+            if(K % (srcVectorLengths[I1] * srcVectorLengths[I4]) != 0)
+            {
+                return false;
+            }
+        }
+
+        // matrix B
+        {
+            auto srcLoadLenghts   = BBlockTransferThreadSliceLengths_K0_N0_N1_K1{};
+            auto srcVectorLengths = BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1{};
+            if(srcVectorLengths[I1] != 1 || srcVectorLengths[I4] != 1)
+            {
+                return false;
+            }
+            if(srcLoadLenghts[I2] % srcVectorLengths[I2] != 0 ||
+               srcLoadLenghts[I3] % srcVectorLengths[I3] != 0)
+            {
+                return false;
+            }
+
+            const index_t C = arg.Conv_K_;
+
+            if(C % (srcVectorLengths[I2] * srcVectorLengths[I3]) != 0)
+            {
+                return false;
+            }
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_C_ % CThreadTransferDstScalarPerVector == 0))
+        {
+            std::cout << "Not surpport,because: arg.Conv_C_ % CThreadTransferDstScalarPerVector = "
+                      << arg.Conv_C_ % CThreadTransferDstScalarPerVector << std::endl;
+            return false;
+        }
+
+        // Gridwise GEMM size
+        return GridwiseGemm::CheckValidity(
+            arg.a_grid_desc_kbatch_k0_m_k1_, arg.b_grid_desc_kbatch_k0_n_k1_, arg.c_grid_desc_m_n_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto
+    MakeArgument(const InDataType* p_in_grid,
+                 WeiDataType* p_wei_grid,
+                 const OutDataType* p_out_grid,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                 const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+                 const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op,
+                 ck::index_t split_k)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        a_g_n_c_wis_lengths, // input
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths, // weight
+                        b_g_k_c_xs_strides,
+                        e_g_n_k_wos_lengths, // output
+                        e_g_n_k_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op,
+                        split_k};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        void* p_wei_grid,
+                        const void* p_out_grid,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+                        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+                        const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op,
+                        ck::index_t split_k) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<WeiDataType*>(p_wei_grid),
+                                          static_cast<const OutDataType*>(p_out_grid),
+                                          a_g_n_c_wis_lengths, // input
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths, // weight
+                                          b_g_k_c_xs_strides,
+                                          e_g_n_k_wos_lengths, // output
+                                          e_g_n_k_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op,
+                                          split_k);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << getConvBackwardWeightSpecializationString(ConvBackwardWeightSpecialization) << ", "
+            << K1
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..71c528e4c9
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp
@@ -0,0 +1,1524 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+struct ComputePtrOffsetOfStridedBatch
+{
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideC_);
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    index_t BatchStrideC_;
+};
+
+} // namespace
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename AGridDesc_B_K0_M_K1,
+          typename BGridDesc_B_K0_N_K1,
+          typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2CTileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batched_gemm_xdlops_bwd_weight(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const index_t batch_count,
+            const AGridDesc_B_K0_M_K1 a_b_k0_m_k1_grid_desc,
+            const BGridDesc_B_K0_N_K1 b_b_k0_n_k1_grid_desc,
+            const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const Block2CTileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
+
+    __shared__ FloatAB p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB)];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_c_grid + c_batch_offset,
+                                                  p_shared,
+                                                  a_b_k0_m_k1_grid_desc,
+                                                  b_b_k0_n_k1_grid_desc,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = a_b_k0_m_k1_grid_desc;
+    ignore = b_b_k0_n_k1_grid_desc;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = batch_count;
+    ignore = block_2_ctile_map;
+    ignore = compute_ptr_offset_of_batch;
+
+    compute_ptr_offset_of_batch.GetAPtrOffset(0);
+    compute_ptr_offset_of_batch.GetBPtrOffset(0);
+    compute_ptr_offset_of_batch.GetCPtrOffset(0);
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          ConvolutionBackwardWeightSpecialization ConvBackwardWeightSpecialization,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerXdl,
+          ck::index_t NPerXdl,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsAddExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsAddExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXdl>
+struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
+    : public DeviceGroupedConvBwdWeight<NDimSpatial,
+                                        InLayout,
+                                        WeiLayout,
+                                        OutLayout,
+                                        InDataType,
+                                        WeiDataType,
+                                        OutDataType,
+                                        InElementwiseOperation,
+                                        WeiElementwiseOperation,
+                                        OutElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvBwdWeight_Xdl_CShuffle;
+
+    using ADataType = OutDataType;
+    using BDataType = InDataType;
+    using CDataType = WeiDataType;
+
+    using AElementwiseOperation = OutElementwiseOperation;
+    using BElementwiseOperation = InElementwiseOperation;
+    using CElementwiseOperation = WeiElementwiseOperation;
+
+    // TODO make A/B datatype different
+    using ABDataType = InDataType;
+
+    // 1d
+    static constexpr bool is_GNWK_GKXC_GNWC =
+        is_same_v<InLayout, tensor_layout::convolution::GNWC> &&
+        is_same_v<WeiLayout, tensor_layout::convolution::GKXC> &&
+        is_same_v<OutLayout, tensor_layout::convolution::GNWK>;
+    // 2d
+    static constexpr bool is_NHWGK_GKYXC_NHWGC =
+        is_same_v<InLayout, tensor_layout::convolution::NHWGC> &&
+        is_same_v<WeiLayout, tensor_layout::convolution::GKYXC> &&
+        is_same_v<OutLayout, tensor_layout::convolution::NHWGK>;
+    static constexpr bool is_GNHWK_GKYXC_GNHWC =
+        is_same_v<InLayout, tensor_layout::convolution::GNHWC> &&
+        is_same_v<WeiLayout, tensor_layout::convolution::GKYXC> &&
+        is_same_v<OutLayout, tensor_layout::convolution::GNHWK>;
+    // 3d
+    static constexpr bool is_NDHWGK_GKZYXC_NDHWGC =
+        is_same_v<InLayout, tensor_layout::convolution::NDHWGC> &&
+        is_same_v<WeiLayout, tensor_layout::convolution::GKZYXC> &&
+        is_same_v<OutLayout, tensor_layout::convolution::NDHWGK>;
+    static constexpr bool is_GNDHWK_GKZYXC_GNDHWC =
+        is_same_v<InLayout, tensor_layout::convolution::GNDHWC> &&
+        is_same_v<WeiLayout, tensor_layout::convolution::GKZYXC> &&
+        is_same_v<OutLayout, tensor_layout::convolution::GNDHWK>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static constexpr auto K1Number     = Number<K1>{};
+    static constexpr auto GemmK1Number = K1Number;
+
+    // Bytes per 32 lds bank: 32 * 4 bytes
+    static constexpr auto BankLength = 128;
+    static constexpr auto ElePerBank = BankLength / sizeof(ADataType);
+
+    // M1 & M0
+    static constexpr auto ABlockLdsM1PerBlock = ElePerBank / K1;
+    static constexpr auto ABlockLdsM0PerBlock = MPerBlock / ABlockLdsM1PerBlock;
+    static constexpr auto ABlockLdsM1Padding  = 4;
+
+    // N1 & N0
+    static constexpr auto BBlockLdsN1PerBlock = ElePerBank / K1;
+    static constexpr auto BBlockLdsN0PerBlock = NPerBlock / BBlockLdsN1PerBlock;
+    static constexpr auto BBlockLdsN1Padding  = 4;
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    constexpr static auto
+    make_out_grid_desc(const ck::index_t N,
+                       const ck::index_t Ho,
+                       const ck::index_t Wo,
+                       const ck::index_t K,
+                       const std::array<ck::index_t, NDimSpatial + 3>& output_strides)
+    {
+        const index_t WoStride = output_strides[4];
+        const auto KStride     = Number<1>{};
+        return make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, K),
+                                            make_tuple(WoStride, KStride));
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    constexpr static auto
+    make_in_grid_desc(const ck::index_t N,
+                      const ck::index_t Hi,
+                      const ck::index_t Wi,
+                      const ck::index_t C,
+                      const std::array<ck::index_t, NDimSpatial + 3>& input_strides)
+    {
+        const index_t NStride  = input_strides[1];
+        const index_t HiStride = input_strides[3];
+        const index_t WiStride = input_strides[4];
+        const auto CStride     = input_strides[2];
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            return make_naive_tensor_descriptor(make_tuple(N * Hi * Wi, C),
+                                                make_tuple(WiStride, CStride));
+        }
+        else
+        {
+            return make_naive_tensor_descriptor(make_tuple(N, Hi, Wi, C),
+                                                make_tuple(NStride, HiStride, WiStride, CStride));
+        }
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    constexpr static auto
+    make_wei_grid_desc(const ck::index_t K,
+                       const ck::index_t Y,
+                       const ck::index_t X,
+                       const ck::index_t C,
+                       const std::array<ck::index_t, NDimSpatial + 3>& weights_strides)
+    {
+        const auto CStride = Number<1>{};
+        const auto KStride = weights_strides[1];
+        return make_naive_tensor_descriptor(make_tuple(K, Y * X * C), make_tuple(KStride, CStride));
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    constexpr static auto
+    make_out_grid_desc(const ck::index_t N,
+                       const ck::index_t Do,
+                       const ck::index_t Ho,
+                       const ck::index_t Wo,
+                       const ck::index_t K,
+                       const std::array<ck::index_t, NDimSpatial + 3>& output_strides)
+    {
+        const index_t WoStride = output_strides[5];
+        const auto KStride     = Number<1>{};
+        return make_naive_tensor_descriptor(make_tuple(N * Do * Ho * Wo, K),
+                                            make_tuple(WoStride, KStride));
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    constexpr static auto
+    make_in_grid_desc(const ck::index_t N,
+                      const ck::index_t Di,
+                      const ck::index_t Hi,
+                      const ck::index_t Wi,
+                      const ck::index_t C,
+                      const std::array<ck::index_t, NDimSpatial + 3>& input_strides)
+    {
+        const index_t NStride  = input_strides[1];
+        const index_t DiStride = input_strides[3];
+        const index_t HiStride = input_strides[4];
+        const index_t WiStride = input_strides[5];
+        const auto CStride     = input_strides[2];
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            return make_naive_tensor_descriptor(make_tuple(N * Di * Hi * Wi, C),
+                                                make_tuple(WiStride, CStride));
+        }
+        else
+        {
+            return make_naive_tensor_descriptor(
+                make_tuple(N, Di, Hi, Wi, C),
+                make_tuple(NStride, DiStride, HiStride, WiStride, CStride));
+        }
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    constexpr static auto
+    make_wei_grid_desc(const ck::index_t K,
+                       const ck::index_t Z,
+                       const ck::index_t Y,
+                       const ck::index_t X,
+                       const ck::index_t C,
+                       const std::array<ck::index_t, NDimSpatial + 3>& weights_strides)
+    {
+        const auto CStride = Number<1>{};
+        const auto KStride = weights_strides[1];
+        return make_naive_tensor_descriptor(make_tuple(K, Z * Y * X * C),
+                                            make_tuple(KStride, CStride));
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+    static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        const ck::index_t N,
+        const ck::index_t K,
+        const ck::index_t C,
+        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial + 3>& /* input_strides */,
+        const std::array<ck::index_t, NDimSpatial + 3>& /* weights_strides */,
+        const std::array<ck::index_t, NDimSpatial + 3>& /* output_strides */,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+        const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+        const ck::index_t batch_k)
+    {
+        using namespace ck;
+
+        const index_t Wi            = input_spatial_lengths[0];
+        const index_t Wo            = output_spatial_lengths[0];
+        const index_t X             = filter_spatial_lengths[0];
+        const index_t ConvStrideW   = conv_filter_strides[0];
+        const index_t ConvDilationW = conv_filter_dilations[0];
+        const index_t InLeftPadW    = input_left_pads[0];
+        const index_t InRightPadW   = input_right_pads[0];
+
+        const index_t GemmKTotal = N * Wo;
+        const index_t GemmM      = K;
+        const index_t GemmN      = C * X;
+
+        const auto PadGemmM = (MPerBlock - GemmM % MPerBlock) % MPerBlock;
+        const auto PadGemmN = (NPerBlock - GemmN % NPerBlock) % NPerBlock;
+
+        const index_t GemmKBatch = batch_k;
+        const index_t GemmK0 =
+            math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
+            K0PerBlock;
+        const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
+
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmktotal_gemmm_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Wo, K));
+
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_gemmktotal_gemmm_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_gemmktotal_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Wi, C));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // C: weight tensor
+            const auto wei_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, X * C));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_gemmm_gemmn_grid_desc);
+        }
+        else
+        {
+            const auto out_gemmktotal_gemmm_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N * Wo, K));
+            const auto in_n_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
+
+            // A: output tensor
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_gemmktotal_gemmm_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_n_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_n_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_gemmktotal_gemmn_grid_desc =
+                transform_tensor_descriptor(in_n_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(X, C)),
+                                                       make_merge_transform(make_tuple(N, Wo))),
+                                            make_tuple(Sequence<1, 3>{}, Sequence<0, 2>{}),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // C: weight tensor
+            const auto wei_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(K, X * C));
+
+            // Padd
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_pad_grid_desc =
+                transform_tensor_descriptor(
+                    out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                    make_tuple(make_pass_through_transform(GemmKBatch),
+                               make_pass_through_transform(GemmK0),
+                               make_right_pad_transform(GemmM, PadGemmM),
+                               make_pass_through_transform(GemmK1Number)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_pad_grid_desc =
+                transform_tensor_descriptor(
+                    in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                    make_tuple(make_pass_through_transform(GemmKBatch),
+                               make_pass_through_transform(GemmK0),
+                               make_right_pad_transform(GemmN, PadGemmN),
+                               make_pass_through_transform(GemmK1Number)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto wei_gemmm_gemmn_pad_grid_desc =
+                transform_tensor_descriptor(wei_gemmm_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmM, PadGemmM),
+                                                       make_right_pad_transform(GemmN, PadGemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_pad_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_pad_grid_desc,
+                              wei_gemmm_gemmn_pad_grid_desc);
+        }
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        const ck::index_t N,
+        const ck::index_t K,
+        const ck::index_t C,
+        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
+        const std::array<ck::index_t, NDimSpatial + 3>& weights_strides,
+        const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+        const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+        const ck::index_t batch_k)
+    {
+        using namespace ck;
+
+        const index_t Hi = input_spatial_lengths[0];
+        const index_t Wi = input_spatial_lengths[1];
+
+        const index_t Ho = output_spatial_lengths[0];
+        const index_t Wo = output_spatial_lengths[1];
+
+        const index_t Y = filter_spatial_lengths[0];
+        const index_t X = filter_spatial_lengths[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t InRightPadH = input_right_pads[0];
+        const index_t InRightPadW = input_right_pads[1];
+
+        const index_t GemmKTotal = N * Ho * Wo;
+        const index_t GemmM      = K;
+        const index_t GemmN      = C * X * Y;
+
+        const auto PadGemmM = (MPerBlock - GemmM % MPerBlock) % MPerBlock;
+        const auto PadGemmN = (NPerBlock - GemmN % NPerBlock) % NPerBlock;
+
+        const index_t GemmKBatch = batch_k;
+        const index_t GemmK0 =
+            math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
+            K0PerBlock;
+        const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
+
+        const auto out_grid_desc = make_out_grid_desc<NDim>(N, Ho, Wo, K, output_strides);
+        const auto in_grid_desc  = make_in_grid_desc<NDim>(N, Hi, Wi, C, input_strides);
+        const auto wei_grid_desc = make_wei_grid_desc<NDim>(K, Y, X, C, weights_strides);
+
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_grid_desc);
+        }
+        else
+        {
+            // A: output tensor
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmktotal_gemmn_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                       make_merge_transform(make_tuple(N, Ho, Wo))),
+                                            make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                            make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // Padd
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_pad_grid_desc =
+                transform_tensor_descriptor(
+                    out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                    make_tuple(make_pass_through_transform(GemmKBatch),
+                               make_pass_through_transform(GemmK0),
+                               make_right_pad_transform(GemmM, PadGemmM),
+                               make_pass_through_transform(GemmK1Number)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_pad_grid_desc =
+                transform_tensor_descriptor(
+                    in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                    make_tuple(make_pass_through_transform(GemmKBatch),
+                               make_pass_through_transform(GemmK0),
+                               make_right_pad_transform(GemmN, PadGemmN),
+                               make_pass_through_transform(GemmK1Number)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto wei_gemmm_gemmn_pad_grid_desc =
+                transform_tensor_descriptor(wei_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmM, PadGemmM),
+                                                       make_right_pad_transform(GemmN, PadGemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_pad_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_pad_grid_desc,
+                              wei_gemmm_gemmn_pad_grid_desc);
+        }
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        const ck::index_t N,
+        const ck::index_t K,
+        const ck::index_t C,
+        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
+        const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
+        const std::array<ck::index_t, NDimSpatial + 3>& weights_strides,
+        const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+        const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+        const ck::index_t batch_k)
+    {
+        using namespace ck;
+
+        const index_t Di = input_spatial_lengths[0];
+        const index_t Hi = input_spatial_lengths[1];
+        const index_t Wi = input_spatial_lengths[2];
+
+        const index_t Do = output_spatial_lengths[0];
+        const index_t Ho = output_spatial_lengths[1];
+        const index_t Wo = output_spatial_lengths[2];
+
+        const index_t Z = filter_spatial_lengths[0];
+        const index_t Y = filter_spatial_lengths[1];
+        const index_t X = filter_spatial_lengths[2];
+
+        const index_t ConvStrideD = conv_filter_strides[0];
+        const index_t ConvStrideH = conv_filter_strides[1];
+        const index_t ConvStrideW = conv_filter_strides[2];
+
+        const index_t ConvDilationD = conv_filter_dilations[0];
+        const index_t ConvDilationH = conv_filter_dilations[1];
+        const index_t ConvDilationW = conv_filter_dilations[2];
+
+        const index_t InLeftPadD = input_left_pads[0];
+        const index_t InLeftPadH = input_left_pads[1];
+        const index_t InLeftPadW = input_left_pads[2];
+
+        const index_t InRightPadD = input_right_pads[0];
+        const index_t InRightPadH = input_right_pads[1];
+        const index_t InRightPadW = input_right_pads[2];
+
+        const index_t GemmKTotal = N * Do * Ho * Wo;
+        const index_t GemmM      = K;
+        const index_t GemmN      = C * Z * X * Y;
+
+        const auto PadGemmM = (MPerBlock - GemmM % MPerBlock) % MPerBlock;
+        const auto PadGemmN = (NPerBlock - GemmN % NPerBlock) % NPerBlock;
+
+        const index_t GemmKBatch = batch_k;
+        const index_t GemmK0 =
+            math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
+            K0PerBlock;
+        const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
+
+        const auto out_grid_desc = make_out_grid_desc<NDim>(N, Do, Ho, Wo, K, output_strides);
+        const auto in_grid_desc  = make_in_grid_desc<NDim>(N, Di, Hi, Wi, C, input_strides);
+        const auto wei_grid_desc = make_wei_grid_desc<NDim>(K, Z, Y, X, C, weights_strides);
+
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_grid_desc);
+        }
+        else
+        {
+            // A: output tensor
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // B: input tensor
+            const auto in_n_dip_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Di, InLeftPadD, InRightPadD),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_n_z_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_dip_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+            const auto in_gemmktotal_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_z_do_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(Z, Y, X, C)),
+                           make_merge_transform(make_tuple(N, Do, Ho, Wo))),
+                make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+            // Padd
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_pad_grid_desc =
+                transform_tensor_descriptor(
+                    out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                    make_tuple(make_pass_through_transform(GemmKBatch),
+                               make_pass_through_transform(GemmK0),
+                               make_right_pad_transform(GemmM, PadGemmM),
+                               make_pass_through_transform(GemmK1Number)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_pad_grid_desc =
+                transform_tensor_descriptor(
+                    in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                    make_tuple(make_pass_through_transform(GemmKBatch),
+                               make_pass_through_transform(GemmK0),
+                               make_right_pad_transform(GemmN, PadGemmN),
+                               make_pass_through_transform(GemmK1Number)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto wei_gemmm_gemmn_pad_grid_desc =
+                transform_tensor_descriptor(wei_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmM, PadGemmM),
+                                                       make_right_pad_transform(GemmN, PadGemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_pad_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_pad_grid_desc,
+                              wei_gemmm_gemmn_pad_grid_desc);
+        }
+    } // function end
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        const ck::index_t dim   = 1;
+        const ck::index_t batch = 1;
+        const std::array<ck::index_t, NDimSpatial> lengths{1};
+        const std::array<ck::index_t, NDimSpatial + 3> strides{1, 1, 1, 1};
+        const std::array<ck::index_t, NDimSpatial> params{1};
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<1>(dim,
+                                                                  dim,
+                                                                  dim,
+                                                                  lengths,
+                                                                  lengths,
+                                                                  lengths,
+                                                                  strides,
+                                                                  strides,
+                                                                  strides,
+                                                                  params,
+                                                                  params,
+                                                                  params,
+                                                                  params,
+                                                                  batch);
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        const ck::index_t dim   = 1;
+        const ck::index_t batch = 1;
+        const std::array<ck::index_t, NDimSpatial> lengths{1, 1};
+        const std::array<ck::index_t, NDimSpatial + 3> strides{1, 1, 1, 1, 1};
+        const std::array<ck::index_t, NDimSpatial> params{1, 1};
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<2>(dim,
+                                                                  dim,
+                                                                  dim,
+                                                                  lengths,
+                                                                  lengths,
+                                                                  lengths,
+                                                                  strides,
+                                                                  strides,
+                                                                  strides,
+                                                                  params,
+                                                                  params,
+                                                                  params,
+                                                                  params,
+                                                                  batch);
+    }
+
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    static auto GetABCGridDesc()
+    {
+        const ck::index_t dim   = 1;
+        const ck::index_t batch = 1;
+        const std::array<ck::index_t, NDimSpatial> lengths{1, 1, 1};
+        const std::array<ck::index_t, NDimSpatial + 3> strides{1, 1, 1, 1, 1, 1};
+        const std::array<ck::index_t, NDimSpatial> params{1, 1, 1};
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<3>(dim,
+                                                                  dim,
+                                                                  dim,
+                                                                  lengths,
+                                                                  lengths,
+                                                                  lengths,
+                                                                  strides,
+                                                                  strides,
+                                                                  strides,
+                                                                  params,
+                                                                  params,
+                                                                  params,
+                                                                  params,
+                                                                  batch);
+    }
+
+    // type convert descs
+    template <typename Desc_M0>
+    static auto PadDescriptor_M0_1d(Desc_M0 desc_m0, index_t gridSize, index_t blockSize)
+    {
+        const auto m0           = desc_m0.GetLength(I0);
+        const index_t loop_step = gridSize * blockSize * 4;
+        const auto pad          = math::integer_least_multiple(m0, loop_step) - m0;
+        const auto desc_m0_pad =
+            transform_tensor_descriptor(desc_m0,
+                                        make_tuple(make_right_pad_transform(m0, pad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return desc_m0_pad;
+    }
+
+    template <index_t Dim>
+    static auto MakeDescriptor_M0(const std::array<index_t, Dim>& shape,
+                                  const std::array<index_t, Dim>& stride,
+                                  index_t gridSize,
+                                  index_t blockSize)
+    {
+        auto tupleOfShape  = generate_tuple([&](auto I) { return shape[I]; }, Number<Dim>{});
+        auto tupleOfStride = generate_tuple([&](auto I) { return stride[I]; }, Number<Dim>{});
+
+        // nd desc - [s0, s1, s2, ...]
+        const auto desc = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
+
+        // merge nd to 1d desc - [s0 * s1 * ...]
+        if constexpr(Dim > 1)
+        {
+            const auto desc_m0 = transform_tensor_descriptor(
+                desc,
+                make_tuple(make_merge_transform(tupleOfShape)),
+                make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<Dim>{})),
+                make_tuple(Sequence<0>{}));
+
+            return PadDescriptor_M0_1d(desc_m0, gridSize, blockSize);
+        }
+        else
+            return PadDescriptor_M0_1d(desc, gridSize, blockSize);
+    }
+
+    using GridDesc_M0 = decltype(MakeDescriptor_M0<1>({1}, {1}, 1, 1));
+
+    using ABCGridDescs = decltype(GetABCGridDesc<NDimSpatial>());
+
+    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
+    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
+    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
+
+    using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum::AtomicAdd,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXdl,
+        NPerXdl,
+        K1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsAddExtraM,
+        ABlockLdsM1PerBlock,
+        ABlockLdsM0PerBlock,
+        ABlockLdsM1Padding,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsAddExtraN,
+        BBlockLdsN1PerBlock,
+        BBlockLdsN0PerBlock,
+        BBlockLdsN1Padding,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CBlockTransferScalarPerVector_NWaveNPerXdl,
+        CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        true,
+        true>;
+
+    // Argument
+    using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        decltype(GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}));
+
+    using Block2CTileMap =
+        decltype(GridwiseGemm::MakeCBlockClusterAdaptor(CGridDesc_M_N{}, 1, 1, 1));
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 WeiDataType* p_wei_grid,
+                 const OutDataType* p_out_grid,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                 const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+                 const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+                 const ck::index_t M01,
+                 const ck::index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op,
+                 ck::index_t split_k)
+            : p_a_grid_{p_out_grid},
+              p_b_grid_{p_in_grid},
+              p_c_grid_{p_wei_grid},
+              a_grid_desc_kbatch_k0_m_k1_{},
+              b_grid_desc_kbatch_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_ctile_map_{},
+              compute_ptr_offset_of_batch_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{out_element_op},
+              b_element_op_{in_element_op},
+              c_element_op_{wei_element_op},
+              Conv_G_{a_g_n_c_wis_lengths[0]},
+              Conv_N_{a_g_n_c_wis_lengths[1]},
+              Conv_K_{b_g_k_c_xs_lengths[1]},
+              Conv_C_{a_g_n_c_wis_lengths[2]},
+              input_spatial_lengths_{},
+              filter_spatial_lengths_{},
+              output_spatial_lengths_{},
+              conv_filter_strides_{conv_filter_strides},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads},
+              k_batch_{split_k}
+        {
+            constexpr index_t spatial_offset = 3;
+            std::copy(begin(a_g_n_c_wis_lengths) + spatial_offset,
+                      end(a_g_n_c_wis_lengths),
+                      begin(input_spatial_lengths_));
+            std::copy(begin(b_g_k_c_xs_lengths) + spatial_offset,
+                      end(b_g_k_c_xs_lengths),
+                      begin(filter_spatial_lengths_));
+            std::copy(begin(e_g_n_k_wos_lengths) + spatial_offset,
+                      end(e_g_n_k_wos_lengths),
+                      begin(output_spatial_lengths_));
+
+            const auto descs =
+                DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
+                    Conv_N_,
+                    Conv_K_,
+                    Conv_C_,
+                    input_spatial_lengths_,
+                    filter_spatial_lengths_,
+                    output_spatial_lengths_,
+                    a_g_n_c_wis_strides,
+                    b_g_k_c_xs_strides,
+                    e_g_n_k_wos_strides,
+                    conv_filter_strides,
+                    conv_filter_dilations,
+                    input_left_pads,
+                    input_right_pads,
+                    k_batch_);
+
+            a_grid_desc_kbatch_k0_m_k1_ = descs[I0];
+            b_grid_desc_kbatch_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_            = descs[I2];
+
+            block_2_ctile_map_ =
+                GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_);
+
+            // A/B/C Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = e_g_n_k_wos_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = a_g_n_c_wis_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideC_ =
+                Conv_K_ * Conv_C_ *
+                std::accumulate(begin(filter_spatial_lengths_),
+                                end(filter_spatial_lengths_),
+                                index_t{1},
+                                std::multiplies<>{});
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_kbatch_k0_m_k1_,
+                                           b_grid_desc_kbatch_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           block_2_ctile_map_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n_);
+            }
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_kbatch_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_kbatch_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock c_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        Block2CTileMap block_2_ctile_map_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+
+        index_t M01_;
+        index_t N01_;
+
+        InElementwiseOperation a_element_op_;
+        OutElementwiseOperation b_element_op_;
+        WeiElementwiseOperation c_element_op_;
+
+        // for checking IsSupportedArgument()
+        const index_t Conv_G_;
+        const index_t Conv_N_;
+        const index_t Conv_K_;
+        const index_t Conv_C_;
+        std::array<ck::index_t, NDimSpatial> input_spatial_lengths_;
+        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths_;
+        std::array<ck::index_t, NDimSpatial> output_spatial_lengths_;
+        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides_;
+        const std::array<ck::index_t, NDimSpatial>& input_left_pads_;
+        const std::array<ck::index_t, NDimSpatial>& input_right_pads_;
+        const index_t k_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        void ShowInfo(const Argument& arg)
+        {
+            std::cout << "arg.a_grid_desc_kbatch_k0_m_k1_{"
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I2) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I3) << "}" << std::endl;
+
+            std::cout << "arg.b_grid_desc_kbatch_k0_n_k1_{"
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I0) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I1) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I2) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I3) << "}" << std::endl;
+
+            std::cout << "arg.c_grid_desc_m_n_{" << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                      << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+        }
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
+                                            arg.b_grid_desc_kbatch_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.block_2_ctile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_) * arg.Conv_G_;
+
+            const auto K0 = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1);
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_batched_gemm_xdlops_bwd_weight<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    CDataType,
+                    OutElementwiseOperation,
+                    InElementwiseOperation,
+                    WeiElementwiseOperation,
+                    remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                    remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                    remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                    remove_reference_t<DeviceOp::Block2CTileMap>,
+                    ComputePtrOffsetOfStridedBatch,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_c_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.c_element_op_,
+                                              arg.Conv_G_,
+                                              arg.a_grid_desc_kbatch_k0_m_k1_,
+                                              arg.b_grid_desc_kbatch_k0_n_k1_,
+                                              arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_ctile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            if(has_main_k0_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if constexpr(NDimSpatial == 1)
+        {
+            if constexpr(!is_GNWK_GKXC_GNWC)
+            {
+                return false;
+            }
+        }
+        else if constexpr(NDimSpatial == 2)
+        {
+            if constexpr(!(is_NHWGK_GKYXC_NHWGC || is_GNHWK_GKYXC_GNHWC))
+            {
+                return false;
+            }
+        }
+        else if constexpr(NDimSpatial == 3)
+        {
+            if constexpr(!(is_NDHWGK_GKZYXC_NDHWGC || is_GNDHWK_GKZYXC_GNDHWC))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 pad = 0 conv
+            for(int i = 0; i < NDimSpatial; i++)
+            {
+                if(!(arg.filter_spatial_lengths_[i] == 1 && arg.conv_filter_strides_[i] == 1 &&
+                     arg.input_left_pads_[i] == 0 && arg.input_right_pads_[i] == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 2 &&
+             arg.Conv_K_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_C_ % CBlockTransferScalarPerVector_NWaveNPerXdl == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
+                                           arg.b_grid_desc_kbatch_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.block_2_ctile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto
+    MakeArgument(const InDataType* p_in_grid,
+                 WeiDataType* p_wei_grid,
+                 const OutDataType* p_out_grid,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                 const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+                 const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op,
+                 const ck::index_t split_k)
+    {
+        return Argument{p_in_grid,
+                        p_wei_grid,
+                        p_out_grid,
+                        a_g_n_c_wis_lengths, // input
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths, // weight
+                        b_g_k_c_xs_strides,
+                        e_g_n_k_wos_lengths, // output
+                        e_g_n_k_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op,
+                        split_k};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        void* p_wei_grid,
+                        const void* p_out_grid,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+                        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                        const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+                        const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+                        InElementwiseOperation in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op,
+                        const ck::index_t split_k) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
+                                          static_cast<WeiDataType*>(p_wei_grid),
+                                          static_cast<const OutDataType*>(p_out_grid),
+                                          a_g_n_c_wis_lengths, // input
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths, // weight
+                                          b_g_k_c_xs_strides,
+                                          e_g_n_k_wos_lengths, // output
+                                          e_g_n_k_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op,
+                                          split_k);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvBwdWeight_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << getConvBackwardWeightSpecializationString(ConvBackwardWeightSpecialization) << ", "
+            << K1 << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << ABlockTransferDstScalarPerVector_K1 << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferDstScalarPerVector_K1 << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle << ", "
+            << CBlockTransferScalarPerVector_NWaveNPerXdl
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000000..8b22bd2090
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,966 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+template <index_t NumDTensor>
+struct ComputePtrOffsetOfStridedBatch
+{
+    ComputePtrOffsetOfStridedBatch() = default;
+
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                   index_t BatchStrideB,
+                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
+                                   index_t BatchStrideE)
+        : BatchStrideA_(BatchStrideA),
+          BatchStrideB_(BatchStrideB),
+          BatchStrideDs_(BatchStrideDs),
+          BatchStrideE_(BatchStrideE)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumDTensor> ds_offset;
+        static_for<0, NumDTensor, 1>{}(
+            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
+        return ds_offset;
+    }
+
+    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    Array<ck::index_t, NumDTensor> BatchStrideDs_;
+    index_t BatchStrideE_;
+};
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
+ * DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
+ * pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_K0_M0_M1_K1,
+          typename BGridDesc_K0_N0_N1_K1,
+          typename DsGridDesc_M0_M10_M11_N0_N10_N11,
+          typename CGridDesc_M0_M10_M11_N0_N10_N11,
+          typename Block2CTileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop,
+          bool HasDoubleTailKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_conv_fwd_dl_multiple_d(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const index_t batch_count,
+            const AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1,
+            const BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1,
+            const DsGridDesc_M0_M10_M11_N0_N10_N11 ds_grid_desc_m0_m10_m11_n0_n10_n11,
+            const CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11,
+            const Block2CTileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__) ||           \
+    defined(__gfx90a__) || defined(__gfx908__) || defined(__gfx940__) || defined(__gfx1100__) || \
+    defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx941__) || defined(__gfx942__))
+    // offset base pointer for each work-group
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);
+
+    __shared__ ABDataType p_shared[shared_block_size];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor = DsGridDesc_M0_M10_M11_N0_N10_N11::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseGemm::Run(p_a_grid + a_batch_offset,
+                      p_b_grid + b_batch_offset,
+                      p_ds_grid_grp,
+                      p_e_grid + c_batch_offset,
+                      p_shared,
+                      a_element_op,
+                      b_element_op,
+                      cde_element_op,
+                      a_grid_desc_k0_m0_m1_k1,
+                      b_grid_desc_k0_n0_n1_k1,
+                      ds_grid_desc_m0_m10_m11_n0_n10_n11,
+                      e_grid_desc_m0_m10_m11_n0_n10_n11,
+                      block_2_ctile_map,
+                      integral_constant<bool, HasMainKBlockLoop>{},
+                      integral_constant<bool, HasDoubleTailKBlockLoop>{});
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m0_m1_k1;
+    ignore = b_grid_desc_k0_n0_n1_k1;
+    ignore = ds_grid_desc_m0_m10_m11_n0_n10_n11;
+    ignore = e_grid_desc_m0_m10_m11_n0_n10_n11;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+
+    compute_ptr_offset_of_batch.GetAPtrOffset(0);
+    compute_ptr_offset_of_batch.GetBPtrOffset(0);
+    compute_ptr_offset_of_batch.GetEPtrOffset(0);
+#endif
+}
+} // namespace
+
+//
+// @brief      Device Convolution operation.
+//
+// Supports:
+//  @li         Forward convolution with up to 3 spatial dimentions
+//  @li         Input tensor in GNWC data format
+//  @li         Weight tensor in GKXC data format
+//  @li         Output tensor in GNWK data format
+//
+// 1D:
+// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
+// 2D:
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+// 3D:
+// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
+//
+template <index_t NDimSpatial,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ConvolutionForwardSpecialization ConvForwardSpecialization,
+          GemmSpecialization GemmSpec,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t K1,
+          index_t M1PerThread,
+          index_t N1PerThread,
+          index_t KPerThread,
+          typename M1N1ThreadClusterM1Xs,
+          typename M1N1ThreadClusterN1Xs,
+          typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+          typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+          typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+          typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+          typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector>
+struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
+    : public DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                           ALayout,
+                                           BLayout,
+                                           DsLayout,
+                                           ELayout,
+                                           ADataType,
+                                           BDataType,
+                                           DsDataType,
+                                           EDataType,
+                                           AElementwiseOperation,
+                                           BElementwiseOperation,
+                                           CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock};
+
+    template <typename ALay>
+    static auto
+    MakeAGridDescriptor_AK0_M_AK1(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                                  const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                                  const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                                  const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                                  const std::array<index_t, NDimSpatial>& input_left_pads,
+                                  const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
+                                                                        a_g_n_c_wis_strides,
+                                                                        b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides,
+                                                                        e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides,
+                                                                        conv_filter_strides,
+                                                                        conv_filter_dilations,
+                                                                        input_left_pads,
+                                                                        input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+
+        const auto M   = in_gemmm_gemmk_desc.GetLength(I0);
+        const auto K   = in_gemmm_gemmk_desc.GetLength(I1);
+        const auto AK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            in_gemmm_gemmk_desc,
+            make_tuple(make_unmerge_transform(make_tuple(AK0, K1)), make_pass_through_transform(M)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    template <typename BLay>
+    static auto
+    MakeBGridDescriptor_BK0_N_BK1(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const auto wei_gemmnraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides);
+
+        const auto wei_gemmn_gemmk_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
+
+        const auto N = wei_gemmn_gemmk_desc.GetLength(I0);
+        const auto K = wei_gemmn_gemmk_desc.GetLength(I1);
+
+        const auto BK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            wei_gemmn_gemmk_desc,
+            make_tuple(make_unmerge_transform(make_tuple(BK0, K1)), make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    template <typename ELay>
+    static auto
+    MakeEGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides)
+    {
+        const auto out_gemmmraw_gemmnraw_desc =
+            conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>(e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides);
+
+        const auto out_gemmm_gemmn_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(ds_g_n_k_wos_lengths[i],
+                                                                  ds_g_n_k_wos_strides[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // desc for problem definition
+    using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(MakeAGridDescriptor_AK0_M_AK1<ALayout>(
+        {}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(MakeBGridDescriptor_BK0_N_BK1<BLayout>({}, {}))>;
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
+    using EGridDesc_M_N  = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<ELayout>({}, {}))>;
+
+    // GridwiseGemm
+    using GridwiseGemm =
+        GridwiseGemmDlMultipleD_km_kn_mn<BlockSize,
+                                         ADataType,
+                                         AccDataType,
+                                         DsDataType,
+                                         EDataType,
+                                         AElementwiseOperation,
+                                         BElementwiseOperation,
+                                         CDEElementwiseOperation,
+                                         InMemoryDataOperationEnum::Set,
+                                         AGridDesc_AK0_M_AK1,
+                                         BGridDesc_BK0_N_BK1,
+                                         EGridDesc_M_N,
+                                         MPerBlock,
+                                         NPerBlock,
+                                         K0PerBlock,
+                                         K1,
+                                         M1PerThread,
+                                         N1PerThread,
+                                         KPerThread,
+                                         M1N1ThreadClusterM1Xs,
+                                         M1N1ThreadClusterN1Xs,
+                                         ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+                                         ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+                                         ABlockTransferThreadClusterArrangeOrder,
+                                         ABlockTransferSrcAccessOrder,
+                                         ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+                                         ABlockTransferSrcVectorTensorContiguousDimOrder,
+                                         ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+                                         BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+                                         BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+                                         BBlockTransferThreadClusterArrangeOrder,
+                                         BBlockTransferSrcAccessOrder,
+                                         BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+                                         BBlockTransferSrcVectorTensorContiguousDimOrder,
+                                         BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+                                         CThreadTransferSrcDstAccessOrder,
+                                         CThreadTransferSrcDstVectorDim,
+                                         CThreadTransferDstScalarPerVector>;
+
+    using AGridDesc_K0_M0_M1_K1 =
+        decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_AK0_M_AK1{}));
+    using BGridDesc_K0_N0_N1_K1 =
+        decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_BK0_N_BK1{}));
+    using DsGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(DsGridDesc_M_N{}));
+    using CGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(EGridDesc_M_N{}));
+    using DefaultBlock2CTileMap =
+        decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(EGridDesc_M_N{}));
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a,
+                 const void* p_b,
+                 const std::array<const void*, NumDTensor>& p_ds,
+                 void* p_e,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_lengths,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOperation& a_element_op,
+                 const BElementwiseOperation& b_element_op,
+                 const CDEElementwiseOperation& cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a)},
+              p_b_grid_{static_cast<const BDataType*>(p_b)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e)},
+              num_group_{a_g_n_c_wis_lengths[0]},
+              a_grid_desc_ak0_m_ak1_{
+                  DeviceOp::MakeAGridDescriptor_AK0_M_AK1<ALayout>(a_g_n_c_wis_lengths,
+                                                                   a_g_n_c_wis_strides,
+                                                                   b_g_k_c_xs_lengths,
+                                                                   b_g_k_c_xs_strides,
+                                                                   e_g_n_k_wos_lengths,
+                                                                   e_g_n_k_wos_strides,
+                                                                   conv_filter_strides,
+                                                                   conv_filter_dilations,
+                                                                   input_left_pads,
+                                                                   input_right_pads)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1<BLayout>(
+                  b_g_k_c_xs_lengths, b_g_k_c_xs_strides)},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides)},
+              a_grid_desc_k0_m0_m1_k1_{},
+              b_grid_desc_k0_n0_n1_k1_{},
+              ds_grid_desc_m0_m10_m11_n0_n10_n11_{},
+              e_grid_desc_m0_m10_m11_n0_n10_n11_{},
+              block_2_ctile_map_{},
+              compute_ptr_offset_of_batch_{},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
+              a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              e_g_n_k_wos_lengths_{e_g_n_k_wos_lengths},
+              e_g_n_k_wos_strides_{e_g_n_k_wos_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // A/B/E Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_k_wos_strides[0];
+
+            // populate pointer, batch stride, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
+
+                // D batch stride
+                compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
+
+                // D desc
+                ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N<DLayout>(
+                    ds_g_n_k_wos_lengths[i], ds_g_n_k_wos_strides[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_ak0_m_ak1_, b_grid_desc_bk0_n_bk1_, e_grid_desc_m_n_))
+            {
+
+                a_grid_desc_k0_m0_m1_k1_ =
+                    GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(a_grid_desc_ak0_m_ak1_);
+                b_grid_desc_k0_n0_n1_k1_ =
+                    GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(b_grid_desc_bk0_n_bk1_);
+                e_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(e_grid_desc_m_n_);
+
+                ds_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(ds_grid_desc_m_n_);
+
+                block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[K0, M, K1]: " << a_grid_desc_ak0_m_ak1_ << std::endl;
+            std::cout << "B[K0, N, K1]: " << b_grid_desc_bk0_n_bk1_ << std::endl;
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+            std::cout << "num_group: " << num_group_ << std::endl;
+
+            std::cout << "A[k0, m0, m1, k1]: " << a_grid_desc_k0_m0_m1_k1_ << std::endl;
+            std::cout << "B[k0, n0, n1, k1]: " << b_grid_desc_k0_n0_n1_k1_ << std::endl;
+            std::cout << "A[m0, m10, m11, n0, n10, n11]: " << e_grid_desc_m0_m10_m11_n0_n10_n11_
+                      << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        index_t num_group_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1_;
+        BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1_;
+        DsGridDesc_M0_M10_M11_N0_N10_N11 ds_grid_desc_m0_m10_m11_n0_n10_n11_;
+        CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11_;
+
+        // block-to-e-tile map
+        DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch<NumDTensor> compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config)
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(
+                   arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.e_grid_desc_m_n_))
+            {
+                throw std::runtime_error(
+                    "wrong! DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK has invalid setting");
+            }
+
+            const index_t grid_size =
+                GridwiseGemm::CalculateGridSize(arg.e_grid_desc_m_n_.GetLength(I0),
+                                                arg.e_grid_desc_m_n_.GetLength(I1)) *
+                arg.num_group_;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop,
+                                     auto has_double_tail_k_block_loop) {
+                constexpr bool has_main_loop   = has_main_k_block_loop.value;
+                constexpr bool has_double_loop = has_double_tail_k_block_loop;
+
+                const auto kernel = kernel_grouped_conv_fwd_dl_multiple_d<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_K0_M0_M1_K1,
+                    DeviceOp::BGridDesc_K0_N0_N1_K1,
+                    DeviceOp::DsGridDesc_M0_M10_M11_N0_N10_N11,
+                    DeviceOp::CGridDesc_M0_M10_M11_N0_N10_N11,
+                    DefaultBlock2CTileMap,
+                    ComputePtrOffsetOfStridedBatch<NumDTensor>,
+                    has_main_loop,
+                    has_double_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_g_n_c_wis_lengths_[0], // Group count
+                                              arg.a_grid_desc_k0_m0_m1_k1_,
+                                              arg.b_grid_desc_k0_n0_n1_k1_,
+                                              arg.ds_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                              arg.e_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                              arg.block_2_ctile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            const auto K0                    = arg.a_grid_desc_k0_m0_m1_k1_.GetLength(I0);
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
+            const bool has_double_tail_k_block_loop =
+                GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
+
+            if(has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, true>{});
+            }
+            else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, false>{});
+            }
+            else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, false>{});
+            }
+            return 0;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        namespace ctc = tensor_layout::convolution;
+
+        // check device
+        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
+             ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx908" ||
+             ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx1100" ||
+             ck::get_device_name() == "gfx1101" || ck::get_device_name() == "gfx1102" ||
+             ck::get_device_name() == "gfx941" || ck::get_device_name() == "gfx942"))
+        {
+            return false;
+        }
+
+        // check ConvolutionForwardSpecialization
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 3];
+                const index_t ConvStride = arg.conv_filter_strides_[i];
+                const index_t LeftPad    = arg.input_left_pads_[i];
+                const index_t RightPad   = arg.input_right_pads_[i];
+
+                if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    std::cout << "Filter1x1Stride1Pad0 check: XY_index = " << i << " X = " << X
+                              << " ConvStride = " << ConvStride << " LeftPad = " << LeftPad
+                              << " RightPad = " << RightPad << std::endl;
+                    return false;
+                }
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 3];
+                const index_t LeftPad  = arg.input_left_pads_[i];
+                const index_t RightPad = arg.input_right_pads_[i];
+
+                if(!(X == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    std::cout << "Filter1x1Stride1Pad0 check: XY_index = " << i << " X = " << X
+                              << " LeftPad = " << LeftPad << " RightPad = " << RightPad
+                              << std::endl;
+                    return false;
+                }
+            }
+        }
+
+        // check vector access of A
+        // FIXME: layout
+        if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
+                     is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
+                     is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
+                     is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
+                     is_same_v<ALayout, ctc::NDHWGC>)
+        {
+            auto srcVectorLengths = ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1{};
+            if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1)
+            {
+                return false;
+            }
+            if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0)
+            {
+                return false;
+            }
+
+            const index_t C = arg.a_g_n_c_wis_lengths_[2];
+
+            if(C % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of B
+        // FIXME: layout
+        if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
+                     is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
+                     is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
+                     is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
+                     is_same_v<BLayout, ctc::KZYXGC>)
+
+        {
+            auto srcVectorLengths = BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1{};
+            if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1)
+            {
+                return false;
+            }
+            if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0)
+            {
+                return false;
+            }
+
+            const index_t C = arg.b_g_k_c_xs_lengths_[2];
+
+            if(C % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of E
+        if constexpr(is_same_v<ELayout, ctc::G_NW_K> || is_same_v<ELayout, ctc::G_NHW_K> ||
+                     is_same_v<ELayout, ctc::G_NDHW_K> || is_same_v<ELayout, ctc::GNWK> ||
+                     is_same_v<ELayout, ctc::GNHWK> || is_same_v<ELayout, ctc::GNDHWK> ||
+                     is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> ||
+                     is_same_v<ELayout, ctc::NDHWGK>)
+        {
+            const index_t K = arg.e_g_n_k_wos_lengths_[2];
+
+            if(!(K % CThreadTransferDstScalarPerVector == 0 && CThreadTransferSrcDstVectorDim == 5))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check Gridwise GEMM
+        return GridwiseGemm::CheckValidity(
+            arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.e_grid_desc_m_n_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_g_n_c_wis_lengths,
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        ds_g_n_k_wos_lengths,
+                        ds_g_n_k_wos_strides,
+                        e_g_n_k_wos_lengths,
+                        e_g_n_k_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_g_n_c_wis_lengths,
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          ds_g_n_k_wos_lengths,
+                                          ds_g_n_k_wos_strides,
+                                          e_g_n_k_wos_lengths,
+                                          e_g_n_k_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization) << ", "
+            << K1
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
new file mode 100644
index 0000000000..f18fbcfe4b
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
@@ -0,0 +1,852 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+struct ComputePtrOffsetOfStridedBatch
+{
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA, index_t BatchStrideB, index_t BatchStrideC)
+        : BatchStrideA_(BatchStrideA), BatchStrideB_(BatchStrideB), BatchStrideC_(BatchStrideC)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideC_);
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    index_t BatchStrideC_;
+};
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
+ * DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
+ * pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename CDataType,
+          typename AGridDesc_K0_M0_M1_K1,
+          typename BGridDesc_K0_N0_N1_K1,
+          typename CGridDesc_M0_M10_M11_N0_N10_N11,
+          typename Block2CTileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop,
+          bool HasDoubleTailKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_conv_fwd_dl(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            CDataType* __restrict__ p_c_grid,
+            const index_t batch_count,
+            const AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1,
+            const BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1,
+            const CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11,
+            const Block2CTileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__) || \
+    defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__))
+    // offset base pointer for each work-group
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
+
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);
+
+    __shared__ ABDataType p_shared[shared_block_size];
+
+    GridwiseGemm::Run(p_a_grid + a_batch_offset,
+                      p_b_grid + b_batch_offset,
+                      p_c_grid + c_batch_offset,
+                      p_shared,
+                      a_grid_desc_k0_m0_m1_k1,
+                      b_grid_desc_k0_n0_n1_k1,
+                      c_grid_desc_m0_m10_m11_n0_n10_n11,
+                      block_2_ctile_map,
+                      integral_constant<bool, HasMainKBlockLoop>{},
+                      integral_constant<bool, HasDoubleTailKBlockLoop>{});
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m0_m1_k1;
+    ignore = b_grid_desc_k0_n0_n1_k1;
+    ignore = c_grid_desc_m0_m10_m11_n0_n10_n11;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+
+    compute_ptr_offset_of_batch.GetAPtrOffset(0);
+    compute_ptr_offset_of_batch.GetBPtrOffset(0);
+    compute_ptr_offset_of_batch.GetCPtrOffset(0);
+#endif
+}
+
+} // namespace
+
+//
+// @brief      Device Convolution operation.
+//
+// Supports:
+//  @li         Forward convolution with up to 3 spatial dimentions
+//  @li         Input tensor in GNWC data format
+//  @li         Weight tensor in GKXC data format
+//  @li         Output tensor in GNWK data format
+//
+// 1D:
+// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
+// 2D:
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+// 3D:
+// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
+//
+template <
+    index_t NDimSpatial,
+    typename ADataType,
+    typename BDataType,
+    typename CDataType,
+    typename AccDataType,
+    typename ALayout,
+    typename BLayout,
+    typename CLayout,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    ConvolutionForwardSpecialization ConvForwardSpecialization,
+    GemmSpecialization GemmSpec,
+    index_t BlockSize,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t K0PerBlock,
+    index_t K1,
+    index_t M1PerThread,
+    index_t N1PerThread,
+    index_t KPerThread,
+    typename M1N1ThreadClusterM1Xs,
+    typename M1N1ThreadClusterN1Xs,
+    typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+    typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+    typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+    typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+    typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+    typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+    typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+    typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+    typename CThreadTransferSrcDstAccessOrder,
+    index_t CThreadTransferSrcDstVectorDim,
+    index_t CThreadTransferDstScalarPerVector,
+    enable_if_t<
+        is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<CElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+        bool> = false>
+struct DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK : public DeviceGroupedConvFwd<NDimSpatial,
+                                                                           ALayout,
+                                                                           BLayout,
+                                                                           CLayout,
+                                                                           ADataType,
+                                                                           BDataType,
+                                                                           CDataType,
+                                                                           AElementwiseOperation,
+                                                                           BElementwiseOperation,
+                                                                           CElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, K0PerBlock};
+
+    template <typename ALay>
+    static auto
+    MakeAGridDescriptor_AK0_M_AK1(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                                  const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides,
+                                  const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                                  const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                                  const std::array<index_t, NDimSpatial>& input_left_pads,
+                                  const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
+                                                                        a_g_n_c_wis_strides,
+                                                                        b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides,
+                                                                        c_g_n_k_wos_lengths,
+                                                                        c_g_n_k_wos_strides,
+                                                                        conv_filter_strides,
+                                                                        conv_filter_dilations,
+                                                                        input_left_pads,
+                                                                        input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+
+        const auto M = in_gemmm_gemmk_desc.GetLength(I0);
+        const auto K = in_gemmm_gemmk_desc.GetLength(I1);
+
+        const auto AK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            in_gemmm_gemmk_desc,
+            make_tuple(make_unmerge_transform(make_tuple(AK0, K1)), make_pass_through_transform(M)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    template <typename BLay>
+    static auto
+    MakeBGridDescriptor_BK0_N_BK1(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const auto wei_gemmnraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides);
+
+        const auto wei_gemmn_gemmk_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
+
+        const auto N = wei_gemmn_gemmk_desc.GetLength(I0);
+        const auto K = wei_gemmn_gemmk_desc.GetLength(I1);
+
+        const auto BK0 = K / K1;
+
+        return transform_tensor_descriptor(
+            wei_gemmn_gemmk_desc,
+            make_tuple(make_unmerge_transform(make_tuple(BK0, K1)), make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    template <typename CLay>
+    static auto
+    MakeCGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides)
+    {
+        const auto out_gemmmraw_gemmnraw_desc =
+            conv_to_gemm_transformer.template MakeCDescriptor_M_N<CLay>(c_g_n_k_wos_lengths,
+                                                                        c_g_n_k_wos_strides);
+
+        const auto out_gemmm_gemmn_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    // desc for problem definition
+    using AGridDesc_AK0_M_AK1 = remove_cvref_t<decltype(MakeAGridDescriptor_AK0_M_AK1<ALayout>(
+        {}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(MakeBGridDescriptor_BK0_N_BK1<BLayout>({}, {}))>;
+    using CGridDesc_M_N = remove_cvref_t<decltype(MakeCGridDescriptor_M_N<CLayout>({}, {}))>;
+
+    // GridwiseGemm
+    using GridwiseGemm =
+        GridwiseGemmDl_km_kn_mn_v1r3<BlockSize,
+                                     ADataType,
+                                     AccDataType,
+                                     CDataType,
+                                     InMemoryDataOperationEnum::Set,
+                                     AGridDesc_AK0_M_AK1,
+                                     BGridDesc_BK0_N_BK1,
+                                     CGridDesc_M_N,
+                                     MPerBlock,
+                                     NPerBlock,
+                                     K0PerBlock,
+                                     K1,
+                                     M1PerThread,
+                                     N1PerThread,
+                                     KPerThread,
+                                     M1N1ThreadClusterM1Xs,
+                                     M1N1ThreadClusterN1Xs,
+                                     ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+                                     ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+                                     ABlockTransferThreadClusterArrangeOrder,
+                                     ABlockTransferSrcAccessOrder,
+                                     ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+                                     ABlockTransferSrcVectorTensorContiguousDimOrder,
+                                     ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+                                     BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+                                     BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+                                     BBlockTransferThreadClusterArrangeOrder,
+                                     BBlockTransferSrcAccessOrder,
+                                     BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+                                     BBlockTransferSrcVectorTensorContiguousDimOrder,
+                                     BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+                                     CThreadTransferSrcDstAccessOrder,
+                                     CThreadTransferSrcDstVectorDim,
+                                     CThreadTransferDstScalarPerVector>;
+
+    using AGridDesc_K0_M0_M1_K1 =
+        decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_AK0_M_AK1{}));
+    using BGridDesc_K0_N0_N1_K1 =
+        decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_BK0_N_BK1{}));
+    using CGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
+    using DefaultBlock2CTileMap =
+        decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(CGridDesc_M_N{}));
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a,
+                 const void* p_b,
+                 void* p_c,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOperation& a_element_op,
+                 const BElementwiseOperation& b_element_op,
+                 const CElementwiseOperation& c_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a)},
+              p_b_grid_{static_cast<const BDataType*>(p_b)},
+              p_c_grid_{static_cast<CDataType*>(p_c)},
+              num_group_{a_g_n_c_wis_lengths[0]},
+              a_grid_desc_ak0_m_ak1_{
+                  DeviceOp::MakeAGridDescriptor_AK0_M_AK1<ALayout>(a_g_n_c_wis_lengths,
+                                                                   a_g_n_c_wis_strides,
+                                                                   b_g_k_c_xs_lengths,
+                                                                   b_g_k_c_xs_strides,
+                                                                   c_g_n_k_wos_lengths,
+                                                                   c_g_n_k_wos_strides,
+                                                                   conv_filter_strides,
+                                                                   conv_filter_dilations,
+                                                                   input_left_pads,
+                                                                   input_right_pads)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1<BLayout>(
+                  b_g_k_c_xs_lengths, b_g_k_c_xs_strides)},
+              c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N<CLayout>(c_g_n_k_wos_lengths,
+                                                                          c_g_n_k_wos_strides)},
+              a_grid_desc_k0_m0_m1_k1_{},
+              b_grid_desc_k0_n0_n1_k1_{},
+              c_grid_desc_m0_m10_m11_n0_n10_n11_{},
+              block_2_ctile_map_{},
+              compute_ptr_offset_of_batch_{
+                  a_g_n_c_wis_strides[0], b_g_k_c_xs_strides[0], c_g_n_k_wos_strides[0]},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op},
+              a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
+              a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              c_g_n_k_wos_lengths_{c_g_n_k_wos_lengths},
+              c_g_n_k_wos_strides_{c_g_n_k_wos_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // A/B/E Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideC_ = c_g_n_k_wos_strides[0];
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(
+                   a_grid_desc_ak0_m_ak1_, b_grid_desc_bk0_n_bk1_, c_grid_desc_m_n_))
+            {
+
+                a_grid_desc_k0_m0_m1_k1_ =
+                    GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(a_grid_desc_ak0_m_ak1_);
+                b_grid_desc_k0_n0_n1_k1_ =
+                    GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(b_grid_desc_bk0_n_bk1_);
+                c_grid_desc_m0_m10_m11_n0_n10_n11_ =
+                    GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(c_grid_desc_m_n_);
+
+                block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[K0, M, K1]: " << a_grid_desc_ak0_m_ak1_ << std::endl;
+            std::cout << "B[K0, N, K1]: " << b_grid_desc_bk0_n_bk1_ << std::endl;
+            std::cout << "C[M, N]: " << c_grid_desc_m_n_ << std::endl;
+            std::cout << "num_group: " << num_group_ << std::endl;
+
+            std::cout << "A[k0, m0, m1, k1]: " << a_grid_desc_k0_m0_m1_k1_ << std::endl;
+            std::cout << "B[k0, n0, n1, k1]: " << b_grid_desc_k0_n0_n1_k1_ << std::endl;
+            std::cout << "A[m0, m10, m11, n0, n10, n11]: " << c_grid_desc_m0_m10_m11_n0_n10_n11_
+                      << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+
+        // tensor descriptors for problem definiton
+        index_t num_group_;
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1_;
+        BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1_;
+        CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11_;
+
+        // block-to-e-tile map
+        DefaultBlock2CTileMap block_2_ctile_map_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<index_t, NDimSpatial + 3> c_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> c_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            // if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(
+                   arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.c_grid_desc_m_n_))
+            {
+                throw std::runtime_error(
+                    "wrong! DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK has invalid setting");
+            }
+
+            const index_t grid_size =
+                GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_.GetLength(I0),
+                                                arg.c_grid_desc_m_n_.GetLength(I1)) *
+                arg.num_group_;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop,
+                                     auto has_double_tail_k_block_loop) {
+                constexpr bool has_main_loop   = has_main_k_block_loop.value;
+                constexpr bool has_double_loop = has_double_tail_k_block_loop;
+
+                const auto kernel =
+                    kernel_grouped_conv_fwd_dl<GridwiseGemm,
+                                               ADataType, // TODO: distiguish A/B datatype
+                                               CDataType,
+                                               DeviceOp::AGridDesc_K0_M0_M1_K1,
+                                               DeviceOp::BGridDesc_K0_N0_N1_K1,
+                                               DeviceOp::CGridDesc_M0_M10_M11_N0_N10_N11,
+                                               DefaultBlock2CTileMap,
+                                               ComputePtrOffsetOfStridedBatch,
+                                               has_main_loop,
+                                               has_double_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_c_grid_,
+                                              arg.a_g_n_c_wis_lengths_[0], // Group count
+                                              arg.a_grid_desc_k0_m0_m1_k1_,
+                                              arg.b_grid_desc_k0_n0_n1_k1_,
+                                              arg.c_grid_desc_m0_m10_m11_n0_n10_n11_,
+                                              arg.block_2_ctile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            const auto K0                    = arg.a_grid_desc_k0_m0_m1_k1_.GetLength(I0);
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
+            const bool has_double_tail_k_block_loop =
+                GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
+
+            if(has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, true>{});
+            }
+            else if(has_main_k_block_loop && !has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, false>{});
+            }
+            else if(!has_main_k_block_loop && has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, false>{});
+            }
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        namespace ctc = tensor_layout::convolution;
+
+        // check device
+        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
+             ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
+             ck::get_device_name() == "gfx1102"))
+        {
+            return false;
+        }
+
+        // check ConvolutionForwardSpecialization
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 3];
+                const index_t ConvStride = arg.conv_filter_strides_[i];
+                const index_t LeftPad    = arg.input_left_pads_[i];
+                const index_t RightPad   = arg.input_right_pads_[i];
+
+                if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    std::cout << "Filter1x1Stride1Pad0 check: i = " << i << " X = " << X
+                              << " ConvStride = " << ConvStride << " LeftPad = " << LeftPad
+                              << " RightPad = " << RightPad << std::endl;
+                    return false;
+                }
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 3];
+                const index_t LeftPad  = arg.input_left_pads_[i];
+                const index_t RightPad = arg.input_right_pads_[i];
+
+                if(!(X == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    std::cout << "Filter1x1Stride1Pad0 check: i = " << i << " X = " << X
+                              << " LeftPad = " << LeftPad << " RightPad = " << RightPad
+                              << std::endl;
+                    return false;
+                }
+            }
+        }
+
+        // check vector access of A
+        // FIXME: layout
+        if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
+                     is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
+                     is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
+                     is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
+                     is_same_v<ALayout, ctc::NDHWGC>)
+        {
+            auto srcVectorLengths = ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1{};
+            if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1)
+            {
+                return false;
+            }
+            if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0)
+            {
+                return false;
+            }
+
+            const index_t C = arg.a_g_n_c_wis_lengths_[2];
+
+            if(C % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of B
+        // FIXME: layout
+        if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
+                     is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
+                     is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
+                     is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
+                     is_same_v<BLayout, ctc::KZYXGC>)
+
+        {
+            auto srcVectorLengths = BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1{};
+            if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1)
+            {
+                return false;
+            }
+            if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0)
+            {
+                return false;
+            }
+
+            const index_t C = arg.b_g_k_c_xs_lengths_[2];
+
+            if(C % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of C
+        if constexpr(is_same_v<CLayout, ctc::G_NW_K> || is_same_v<CLayout, ctc::G_NHW_K> ||
+                     is_same_v<CLayout, ctc::G_NDHW_K> || is_same_v<CLayout, ctc::GNWK> ||
+                     is_same_v<CLayout, ctc::GNHWK> || is_same_v<CLayout, ctc::GNDHWK> ||
+                     is_same_v<CLayout, ctc::NWGK> || is_same_v<CLayout, ctc::NHWGK> ||
+                     is_same_v<CLayout, ctc::NDHWGK>)
+        {
+            const index_t K = arg.c_g_n_k_wos_lengths_[2];
+
+            if(!(K % CThreadTransferDstScalarPerVector == 0 && CThreadTransferSrcDstVectorDim == 5))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+        // check Gridwise GEMM
+        return GridwiseGemm::CheckValidity(
+            arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.c_grid_desc_m_n_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_a,
+                             const void* p_b,
+                             void* p_c,
+                             const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                             const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                             const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                             const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                             const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                             const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides,
+                             const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                             const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                             const std::array<index_t, NDimSpatial>& input_left_pads,
+                             const std::array<index_t, NDimSpatial>& input_right_pads,
+                             const AElementwiseOperation& a_element_op,
+                             const BElementwiseOperation& b_element_op,
+                             const CElementwiseOperation& c_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_c,
+                        a_g_n_c_wis_lengths,
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        c_g_n_k_wos_lengths,
+                        c_g_n_k_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_c,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads,
+                        const AElementwiseOperation& a_element_op,
+                        const BElementwiseOperation& b_element_op,
+                        const CElementwiseOperation& c_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_c,
+                                          a_g_n_c_wis_lengths,
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          c_g_n_k_wos_lengths,
+                                          c_g_n_k_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization) << ", "
+            << K1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << ABlockTransferDstScalarPerVector_K1 << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferDstScalarPerVector_K1 << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle << ", "
+            << CBlockTransferScalarPerVector_NWaveNPerXdl
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp
new file mode 100644
index 0000000000..face627e1f
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp
@@ -0,0 +1,77 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Grouped Convolution Forward:
+//   input : input image A[G, N, C, Hi, Wi],
+//   input : weight B[G, K, C, Y, X],
+//   input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
+//   output : output image E[G, N, K, Ho, Wo]
+//   output : R0[G, N, Ho, Wo], R1[G, N, Ho, Wo], ...
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+//   Q0 = reduce0(q_op0(E)), Q1 = reduce1(q_op0(E)), ...
+//   R0 = r_op0(Q0), R1 = r_op1(Q1), ...
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename RLayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename RsDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation>
+struct DeviceGroupedConvFwdMultipleDMultipleR : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+    static constexpr index_t NumRTensor = RsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        std::array<void*, NumRTensor> p_rs,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op,
+        const QsElementwiseOperation& qs_element_op,
+        const RsElementwiseOperation& rs_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..caa18b709c
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
@@ -0,0 +1,1117 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+template <index_t NumDTensor, index_t NumRTensor>
+struct ComputePtrOffsetOfStridedBatch
+{
+    ComputePtrOffsetOfStridedBatch() = default;
+
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                   index_t BatchStrideB,
+                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
+                                   index_t BatchStrideE,
+                                   Array<ck::index_t, NumRTensor> BatchStrideRs)
+        : BatchStrideA_(BatchStrideA),
+          BatchStrideB_(BatchStrideB),
+          BatchStrideDs_(BatchStrideDs),
+          BatchStrideE_(BatchStrideE),
+          BatchStrideRs_(BatchStrideRs)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumDTensor> ds_offset;
+        static_for<0, NumDTensor, 1>{}(
+            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
+        return ds_offset;
+    }
+
+    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    __host__ __device__ constexpr auto GetRsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumRTensor> rs_offset;
+        static_for<0, NumRTensor, 1>{}(
+            [&](auto i) { rs_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideRs_[i]); });
+        return rs_offset;
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    Array<ck::index_t, NumDTensor> BatchStrideDs_;
+    index_t BatchStrideE_;
+    Array<ck::index_t, NumRTensor> BatchStrideRs_;
+};
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for
+ * \link DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the
+ * computing of pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename RsPointer,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename RsGridDescriptor_MBlock_MPerBlock,
+          typename Block2ETileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_batch_gemm_multiple_d_xdl_cshuffle(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            RsPointer p_rs_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const QsElementwiseOperation qs_element_op,
+            const RsElementwiseOperation rs_element_op,
+            const index_t batch_count,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock_,
+            const RsGridDescriptor_MBlock_MPerBlock rs_grid_desc_mblock_mperblock,
+            const Block2ETileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+    const auto rs_batch_offset = compute_ptr_offset_of_batch.GetRsPtrOffset(g_idx);
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    RsPointer p_rs_grid_grp;
+
+    static constexpr index_t NumRTensor = RsGridDescriptor_MBlock_MPerBlock::Size();
+
+    static_for<0, NumRTensor, 1>{}(
+        [&](auto i) { p_rs_grid_grp(i) = p_rs_grid[i] + rs_batch_offset[i]; });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_ds_grid_grp,
+                                                  p_e_grid + e_batch_offset,
+                                                  p_rs_grid_grp,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  qs_element_op,
+                                                  rs_element_op,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                                  rs_grid_desc_mblock_mperblock,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = p_rs_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
+    ignore = rs_grid_desc_mblock_mperblock;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = qs_element_op;
+    ignore = rs_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+#endif
+}
+
+} // namespace
+
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename RLayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ReduceAccDataType,
+          typename RsDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation,
+          typename ThreadReduceOperations,
+          typename RsGlobalMemoryDataOperation,
+          ConvolutionForwardSpecialization ConvForwardSpecialization,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDRThreadTransferClusterLengths_MPerBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          index_t RThreadTransferDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
+    : public DeviceGroupedConvFwdMultipleDMultipleR<NDimSpatial,
+                                                    ALayout,
+                                                    BLayout,
+                                                    DELayout,
+                                                    RLayout,
+                                                    ADataType,
+                                                    BDataType,
+                                                    DsDataType,
+                                                    EDataType,
+                                                    RsDataType,
+                                                    AElementwiseOperation,
+                                                    BElementwiseOperation,
+                                                    CDEElementwiseOperation,
+                                                    RsElementwiseOperation,
+                                                    QsElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+    static constexpr index_t NumRTensor = RsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    template <typename ALay>
+    static auto
+    MakeAGridDescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                            const std::array<index_t, NDimSpatial>& input_left_pads,
+                            const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
+                                                                        a_g_n_c_wis_strides,
+                                                                        b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides,
+                                                                        e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides,
+                                                                        conv_filter_strides,
+                                                                        conv_filter_dilations,
+                                                                        input_left_pads,
+                                                                        input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+
+        return in_gemmm_gemmk_desc;
+    }
+
+    template <typename BLay>
+    static auto
+    MakeBGridDescriptor_N_K(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const auto wei_gemmnraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides);
+
+        const auto wei_gemmn_gemmk_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
+
+        return wei_gemmn_gemmk_desc;
+    }
+
+    template <typename ELay>
+    static auto
+    MakeEGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides)
+    {
+        const auto out_gemmmraw_gemmnraw_desc =
+            conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>(e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides);
+
+        const auto out_gemmm_gemmn_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    template <typename Descriptor>
+    static auto GetPaddedRGridDescriptor(Descriptor descriptor, index_t MRaw)
+    {
+        const auto M    = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
+        const auto MPad = M - MRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                     GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M
+            return transform_tensor_descriptor(
+                descriptor,
+                make_tuple(make_right_pad_transform(descriptor, MPad)),
+                make_tuple(Sequence<0>{}),
+                make_tuple(Sequence<0>{}));
+        }
+        else
+        {
+            // not pad M
+            return descriptor;
+        }
+    }
+
+    template <typename RLay,
+              typename std::enable_if<is_same_v<RLay, tensor_layout::convolution::GNW> ||
+                                          is_same_v<RLay, tensor_layout::convolution::GNHW> ||
+                                          is_same_v<RLay, tensor_layout::convolution::GNDHW>,
+                                      bool>::type = false>
+    static auto
+    MakeRGridDescriptor_M(const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+                          const std::array<index_t, NDimSpatial + 2>& /* r_g_n_wos_strides */)
+    {
+        const index_t N = r_g_n_wos_lengths[1];
+
+        const index_t NHoWo =
+            N * ck::accumulate_n<index_t>(
+                    r_g_n_wos_lengths.begin() + 2, NDimSpatial, 1, std::multiplies<>());
+
+        const auto r_grid_desc_mraw = make_naive_tensor_descriptor_packed(make_tuple(NHoWo));
+
+        return GetPaddedRGridDescriptor(r_grid_desc_mraw, NHoWo);
+    }
+
+    template <typename RLay,
+              typename std::enable_if<is_same_v<RLay, tensor_layout::convolution::G_NW> ||
+                                          is_same_v<RLay, tensor_layout::convolution::G_NHW> ||
+                                          is_same_v<RLay, tensor_layout::convolution::G_NDHW> ||
+                                          is_same_v<RLay, tensor_layout::convolution::NWG> ||
+                                          is_same_v<RLay, tensor_layout::convolution::NHWG> ||
+                                          is_same_v<RLay, tensor_layout::convolution::NDHWG>,
+                                      bool>::type = false>
+    static auto MakeRGridDescriptor_M(const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+                                      const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_strides)
+    {
+        const index_t N = r_g_n_wos_lengths[1];
+
+        const index_t WoStride = r_g_n_wos_strides[NDimSpatial + 2];
+
+        const index_t NHoWo =
+            N * ck::accumulate_n<index_t>(
+                    r_g_n_wos_lengths.begin() + 2, NDimSpatial, 1, std::multiplies<>());
+
+        const auto r_grid_desc_mraw =
+            make_naive_tensor_descriptor(make_tuple(NHoWo), make_tuple(WoStride));
+
+        return GetPaddedRGridDescriptor(r_grid_desc_mraw, NHoWo);
+    }
+
+    using AGridDesc_M_K = remove_cvref_t<decltype(MakeAGridDescriptor_M_K<ALayout>(
+        {}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using BGridDesc_N_K = remove_cvref_t<decltype(MakeBGridDescriptor_N_K<BLayout>({}, {}))>;
+    using EGridDesc_M_N = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<DELayout>({}, {}))>;
+    using RGridDesc_M   = remove_cvref_t<decltype(MakeRGridDescriptor_M<RLayout>({}, {}))>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleDMultipleR_k0mk1_k0nk1_mn_xdl_cshuffle_v1<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        ReduceAccDataType,
+        RsDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        QsElementwiseOperation,
+        RsElementwiseOperation,
+        ThreadReduceOperations,
+        InMemoryDataOperationEnum::Set,
+        RsGlobalMemoryDataOperation,
+        AGridDesc_M_K,
+        BGridDesc_N_K,
+        EGridDesc_M_N,
+        RGridDesc_M,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDRThreadTransferClusterLengths_MPerBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        RThreadTransferDstScalarPerVector_MPerBlock,
+        LoopSched>;
+
+    using AGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
+            AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
+            BGridDesc_N_K{}))>;
+
+    using Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a,
+                 const void* p_b,
+                 const std::array<const void*, NumDTensor>& p_ds,
+                 void* p_e,
+                 std::array<void*, NumRTensor> p_rs,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_lengths,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOperation& a_element_op,
+                 const BElementwiseOperation& b_element_op,
+                 const CDEElementwiseOperation& cde_element_op,
+                 const QsElementwiseOperation& qs_element_op,
+                 const RsElementwiseOperation& rs_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a)},
+              p_b_grid_{static_cast<const BDataType*>(p_b)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e)},
+              p_rs_grid_{}, // FIXME
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K<ALayout>(a_g_n_c_wis_lengths,
+                                                                          a_g_n_c_wis_strides,
+                                                                          b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides,
+                                                                          e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K<BLayout>(b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<DELayout>(e_g_n_k_wos_lengths,
+                                                                           e_g_n_k_wos_strides)},
+              r_grid_desc_m_{
+                  DeviceOp::MakeRGridDescriptor_M<RLayout>(r_g_n_wos_lengths, r_g_n_wos_strides)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              rs_grid_desc_mblock_mperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              compute_ptr_offset_of_batch_{},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              qs_element_op_{qs_element_op},
+              rs_element_op_{rs_element_op},
+              a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
+              a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              ds_g_n_k_wos_lengths_{ds_g_n_k_wos_lengths},
+              ds_g_n_k_wos_strides_{ds_g_n_k_wos_strides},
+              e_g_n_k_wos_lengths_{e_g_n_k_wos_lengths},
+              e_g_n_k_wos_strides_{e_g_n_k_wos_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // A/B/E Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_k_wos_strides[0];
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           e_grid_desc_m_n_,
+                                           r_grid_desc_m_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+
+                // populate pointer, batch stride, desc for Ds
+                static_for<0, NumDTensor, 1>{}([&](auto i) {
+                    using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                    // D pointer
+                    p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
+
+                    // D batch stride
+                    compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
+
+                    // D desc
+                    ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N<DELayout>(
+                        ds_g_n_k_wos_lengths[i], ds_g_n_k_wos_strides[i]);
+
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock_(i) =
+                        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                            ds_grid_desc_m_n_(i));
+                });
+
+                // populate pointer for Rs
+                static_for<0, NumRTensor, 1>{}([&](auto i) {
+                    using RDataType = remove_cvref_t<tuple_element_t<i.value, RsDataType>>;
+
+                    // R pointer
+                    p_rs_grid_(i) = static_cast<RDataType*>(p_rs[i]);
+
+                    rs_grid_desc_mblock_mperblock_(i) =
+                        GridwiseGemm::MakeRGridDescriptor_MBlock_MPerBlock(r_grid_desc_m_);
+                });
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+        typename GridwiseGemm::RsGridPointer p_rs_grid_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        EGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+        RGridDesc_M r_grid_desc_m_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        StaticallyIndexedArray<
+            typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+            NumDTensor>
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_; // FIXME: Ds desc may be of different
+                                                             // type from E
+        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        StaticallyIndexedArray<typename GridwiseGemm::RGridDescriptor_MBlock_MPerBlock, NumRTensor>
+            rs_grid_desc_mblock_mperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        ComputePtrOffsetOfStridedBatch<NumDTensor, NumRTensor> compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+        QsElementwiseOperation qs_element_op_;
+        RsElementwiseOperation rs_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.r_grid_desc_m_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemmMultipleD_xdl_cshuffle has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) *
+                arg.a_g_n_c_wis_lengths_[0]; // Group count
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_batch_gemm_multiple_d_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    typename GridwiseGemm::RsGridPointer,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    QsElementwiseOperation,
+                    RsElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    ck::StaticallyIndexedArray<
+                        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                        NumDTensor>,
+                    typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    ck::StaticallyIndexedArray<
+                        typename GridwiseGemm::RGridDescriptor_MBlock_MPerBlock,
+                        NumRTensor>,
+                    Block2ETileMap,
+                    ComputePtrOffsetOfStridedBatch<NumDTensor, NumRTensor>,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.p_rs_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.qs_element_op_,
+                                              arg.rs_element_op_,
+                                              arg.a_g_n_c_wis_lengths_[0], // Group count
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.rs_grid_desc_mblock_mperblock_,
+                                              arg.block_2_etile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        namespace ctc = tensor_layout::convolution;
+
+        // check device
+        if(get_device_name() == "gfx908")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t>))
+            {
+                return false;
+            }
+        }
+        else if(get_device_name() == "gfx90a" || get_device_name() == "gfx940" ||
+                get_device_name() == "gfx941" || get_device_name() == "gfx942")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check ConvolutionForwardSpecialization
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t ConvStride = arg.conv_filter_strides_[i];
+                const index_t LeftPad    = arg.input_left_pads_[i];
+                const index_t RightPad   = arg.input_right_pads_[i];
+
+                if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t LeftPad  = arg.input_left_pads_[i];
+                const index_t RightPad = arg.input_right_pads_[i];
+
+                if(!(X == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // check vector access of A
+        // FIXME: layout
+        if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
+                     is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
+                     is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
+                     is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
+                     is_same_v<ALayout, ctc::NDHWGC>)
+        {
+            const index_t C = arg.a_g_n_c_wis_lengths_[2];
+
+            if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of B
+        // FIXME: layout
+        if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
+                     is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
+                     is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
+                     is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
+                     is_same_v<BLayout, ctc::KZYXGC>)
+
+        {
+            const index_t C = arg.b_g_k_c_xs_lengths_[2];
+
+            if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        //  check vector access of Ds
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            // FIXME: layout
+            if constexpr(is_same_v<DELayout, ctc::G_NW_K> || is_same_v<DELayout, ctc::G_NHW_K> ||
+                         is_same_v<DELayout, ctc::G_NDHW_K> || is_same_v<DELayout, ctc::GNWK> ||
+                         is_same_v<DELayout, ctc::GNHWK> || is_same_v<DELayout, ctc::GNDHWK> ||
+                         is_same_v<DELayout, ctc::NWGK> || is_same_v<DELayout, ctc::NHWGK> ||
+                         is_same_v<DELayout, ctc::NDHWGK>)
+            {
+                const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
+
+                if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+                {
+                    valid = false;
+                }
+            }
+            else
+            {
+                valid = false;
+            }
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check vector access of E
+        if constexpr(is_same_v<DELayout, ctc::G_NW_K> || is_same_v<DELayout, ctc::G_NHW_K> ||
+                     is_same_v<DELayout, ctc::G_NDHW_K> || is_same_v<DELayout, ctc::GNWK> ||
+                     is_same_v<DELayout, ctc::GNHWK> || is_same_v<DELayout, ctc::GNDHWK> ||
+                     is_same_v<DELayout, ctc::NWGK> || is_same_v<DELayout, ctc::NHWGK> ||
+                     is_same_v<DELayout, ctc::NDHWGK>)
+        {
+            const index_t K = arg.e_g_n_k_wos_lengths_[2];
+
+            if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of R
+        if constexpr(!(is_same_v<RLayout, ctc::G_NW> || is_same_v<RLayout, ctc::G_NHW> ||
+                       is_same_v<RLayout, ctc::G_NDHW> || is_same_v<RLayout, ctc::GNW> ||
+                       is_same_v<RLayout, ctc::GNHW> || is_same_v<RLayout, ctc::GNDHW> ||
+                       is_same_v<RLayout, ctc::NWG> || is_same_v<RLayout, ctc::NHWG> ||
+                       is_same_v<RLayout, ctc::NDHWG>))
+        {
+            return false;
+        }
+
+        // check Gridwise GEMM
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           arg.e_grid_desc_m_n_,
+                                           arg.r_grid_desc_m_,
+                                           arg.block_2_etile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        std::array<void*, NumRTensor> p_rs,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op,
+        const QsElementwiseOperation& qs_element_op,
+        const RsElementwiseOperation& rs_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        p_rs,
+                        a_g_n_c_wis_lengths,
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        ds_g_n_k_wos_lengths,
+                        ds_g_n_k_wos_strides,
+                        e_g_n_k_wos_lengths,
+                        e_g_n_k_wos_strides,
+                        r_g_n_wos_lengths,
+                        r_g_n_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op,
+                        qs_element_op,
+                        rs_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        std::array<void*, NumRTensor> p_rs,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_lengths,
+        const std::array<index_t, NDimSpatial + 2>& r_g_n_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op,
+        const QsElementwiseOperation& qs_element_op,
+        const RsElementwiseOperation& rs_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          p_rs,
+                                          a_g_n_c_wis_lengths,
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          ds_g_n_k_wos_lengths,
+                                          ds_g_n_k_wos_strides,
+                                          e_g_n_k_wos_lengths,
+                                          e_g_n_k_wos_strides,
+                                          r_g_n_wos_lengths,
+                                          r_g_n_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op,
+                                          qs_element_op,
+                                          rs_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvFwdMultipleD_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization) << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp
new file mode 100644
index 0000000000..7c726cd851
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp
@@ -0,0 +1,854 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+template <index_t NumDTensor>
+struct ComputePtrOffsetOfStridedBatch
+{
+    ComputePtrOffsetOfStridedBatch() = default;
+
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                   index_t BatchStrideB,
+                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
+                                   index_t BatchStrideE)
+        : BatchStrideA_(BatchStrideA),
+          BatchStrideB_(BatchStrideB),
+          BatchStrideDs_(BatchStrideDs),
+          BatchStrideE_(BatchStrideE)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumDTensor> ds_offset;
+        static_for<0, NumDTensor, 1>{}(
+            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
+        return ds_offset;
+    }
+
+    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    Array<ck::index_t, NumDTensor> BatchStrideDs_;
+    index_t BatchStrideE_;
+};
+
+} // namespace
+
+//
+// @brief      Device Convolution operation.
+//
+// Supports:
+//  @li         Forward convolution with up to 3 spatial dimentions
+//  @li         Input tensor in GNWC data format
+//  @li         Weight tensor in GKXC data format
+//  @li         Output tensor in GNWK data format
+//
+// 1D:
+// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
+// 2D:
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+// 3D:
+// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
+// Assume:
+//  AK1 == BK1
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ConvolutionForwardSpecialization ConvForwardSpecialization,
+          GemmSpecialization GemmSpec,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t K0PerBlock,
+          ck::index_t K1,
+          ck::index_t MPerWMMA,
+          ck::index_t NPerWMMA,
+          ck::index_t MRepeat,
+          ck::index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          typename CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          index_t NumGemmKPrefetchStage   = 1,
+          LoopScheduler LoopSched         = make_default_loop_scheduler(),
+          ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>
+struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
+    : public DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                           ALayout,
+                                           BLayout,
+                                           DsLayout,
+                                           ELayout,
+                                           ADataType,
+                                           BDataType,
+                                           DsDataType,
+                                           EDataType,
+                                           AElementwiseOperation,
+                                           BElementwiseOperation,
+                                           CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD_Wmma_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0           = Number<0>{};
+    static constexpr auto I1           = Number<1>{};
+    static constexpr auto I2           = Number<2>{};
+    static constexpr auto I3           = Number<3>{};
+    static constexpr index_t KPerBlock = K0PerBlock * K1;
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    template <typename ALay>
+    static auto
+    MakeAGridDescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                            const std::array<index_t, NDimSpatial>& input_left_pads,
+                            const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
+                                                                        a_g_n_c_wis_strides,
+                                                                        b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides,
+                                                                        e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides,
+                                                                        conv_filter_strides,
+                                                                        conv_filter_dilations,
+                                                                        input_left_pads,
+                                                                        input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+
+        return in_gemmm_gemmk_desc;
+    }
+
+    template <typename BLay>
+    static auto
+    MakeBGridDescriptor_N_K(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const auto wei_gemmnraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides);
+
+        const auto wei_gemmn_gemmk_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
+
+        return wei_gemmn_gemmk_desc;
+    }
+
+    template <typename ELay>
+    static auto
+    MakeEGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides)
+    {
+        const auto out_gemmmraw_gemmnraw_desc =
+            conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>(e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides);
+
+        const auto out_gemmm_gemmn_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(ds_g_n_k_wos_lengths[i],
+                                                                  ds_g_n_k_wos_strides[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // desc for problem definition
+    using AGridDesc_M_K  = remove_cvref_t<decltype(MakeAGridDescriptor_M_K<ALayout>(
+        {}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using BGridDesc_N_K  = remove_cvref_t<decltype(MakeBGridDescriptor_N_K<BLayout>({}, {}))>;
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
+    using EGridDesc_M_N  = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<ELayout>({}, {}))>;
+
+    // A desc for source in blockwise copy
+    template <typename AGridDesc_M_K>
+    __host__ __device__ static constexpr auto
+    MakeAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK1 = K1;
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B desc for source in blockwise copy
+    template <typename BGridDesc_N_K>
+    __host__ __device__ static constexpr auto
+    MakeBGridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK1 = K1;
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    using AGridDesc_AK0_M_AK1 = decltype(DeviceOp::MakeAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}));
+    using BGridDesc_BK0_N_BK1 = decltype(DeviceOp::MakeBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}));
+
+    // GridwiseOp
+    using GridwiseOp = GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle<
+        // DataType Family
+        ADataType,
+        BDataType,
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        // InMemory Data Descriptor
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        DsGridDesc_M_N,
+        EGridDesc_M_N,
+        // ElementwiseOp Family
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        // Tiling Family
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerWMMA,
+        NPerWMMA,
+        K1,
+        MRepeat,
+        NRepeat,
+        // ThreadCluster Family
+        BlockSize,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMRepeatPerShuffle,
+        CShuffleNRepeatPerShuffle,
+        CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+        NumGemmKPrefetchStage,
+        LoopSched,
+        PipelineVer>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a,
+                 const void* p_b,
+                 const std::array<const void*, NumDTensor>& p_ds,
+                 void* p_e,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_lengths,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 index_t M01,
+                 index_t N01,
+                 const AElementwiseOperation& a_element_op,
+                 const BElementwiseOperation& b_element_op,
+                 const CDEElementwiseOperation& cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a)},
+              p_b_grid_{static_cast<const BDataType*>(p_b)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e)},
+              num_group_{a_g_n_c_wis_lengths[0]},
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K<ALayout>(a_g_n_c_wis_lengths,
+                                                                          a_g_n_c_wis_strides,
+                                                                          b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides,
+                                                                          e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K<BLayout>(b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides)},
+              a_grid_desc_ak0_m_ak1_{DeviceOp::MakeAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{GridwiseOp::MakeDefaultBlock2CTileMap(e_grid_desc_m_n_, M01, N01)},
+              compute_ptr_offset_of_batch_{},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
+              a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              ds_g_n_k_wos_lengths_{ds_g_n_k_wos_lengths},
+              ds_g_n_k_wos_strides_{ds_g_n_k_wos_strides},
+              e_g_n_k_wos_lengths_{e_g_n_k_wos_lengths},
+              e_g_n_k_wos_strides_{e_g_n_k_wos_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // A/B/E Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_k_wos_strides[0];
+
+            // populate pointer, batch stride, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                // using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
+
+                // D batch stride
+                compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
+            });
+
+            // D desc
+            ds_grid_desc_m_n_ =
+                DeviceOp::MakeDsGridDescriptor_M_N(ds_g_n_k_wos_lengths, ds_g_n_k_wos_strides);
+
+            // populate desc for Ds/E
+            e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                GridwiseOp::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n_);
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                GridwiseOp::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                    ds_grid_desc_m_n_);
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseOp::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        index_t num_group_;
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        typename GridwiseOp::DefaultBlock2CTileMap block_2_etile_map_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch<NumDTensor> compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * arg.num_group_;
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_grouped_conv_fwd_multiple_d_wmma_cshuffle<
+                    GridwiseOp,
+                    ADataType,
+                    BDataType,
+                    typename GridwiseOp::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    remove_reference_t<typename GridwiseOp::DefaultBlock2CTileMap>,
+                    ComputePtrOffsetOfStridedBatch<NumDTensor>,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_g_n_c_wis_lengths_[0], // Group count
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_etile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            if(GridwiseOp::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        namespace ctc = tensor_layout::convolution;
+
+        // check device
+        if(get_device_name() == "gfx1100" || get_device_name() == "gfx1101" ||
+           ck::get_device_name() == "gfx1102")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check ConvolutionForwardSpecialization
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t ConvStride = arg.conv_filter_strides_[i];
+                const index_t LeftPad    = arg.input_left_pads_[i];
+                const index_t RightPad   = arg.input_right_pads_[i];
+
+                if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t LeftPad  = arg.input_left_pads_[i];
+                const index_t RightPad = arg.input_right_pads_[i];
+
+                if(!(X == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // check vector access of A
+        // FIXME: layout
+        if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
+                     is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
+                     is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
+                     is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
+                     is_same_v<ALayout, ctc::NDHWGC>)
+        {
+            const index_t C = arg.a_g_n_c_wis_lengths_[2];
+
+            if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of B
+        // FIXME: layout
+        if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
+                     is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
+                     is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
+                     is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
+                     is_same_v<BLayout, ctc::KZYXGC>)
+
+        {
+            const index_t C = arg.b_g_k_c_xs_lengths_[2];
+
+            if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        //  check vector access of Ds
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+            // FIXME: layout
+            if constexpr(is_same_v<DLayout, ctc::G_NW_K> || is_same_v<DLayout, ctc::G_NHW_K> ||
+                         is_same_v<DLayout, ctc::G_NDHW_K> || is_same_v<DLayout, ctc::GNWK> ||
+                         is_same_v<DLayout, ctc::GNHWK> || is_same_v<DLayout, ctc::GNDHWK> ||
+                         is_same_v<DLayout, ctc::NWGK> || is_same_v<DLayout, ctc::NHWGK> ||
+                         is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::GK> ||
+                         is_same_v<DLayout, ctc::G_K>)
+            {
+                const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
+
+                if(!(K % CDEShuffleBlockTransferScalarPerVector_NPerBlock == 0))
+                {
+                    valid = false;
+                }
+            }
+            else
+            {
+                valid = false;
+            }
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check vector access of E
+        if constexpr(is_same_v<ELayout, ctc::G_NW_K> || is_same_v<ELayout, ctc::G_NHW_K> ||
+                     is_same_v<ELayout, ctc::G_NDHW_K> || is_same_v<ELayout, ctc::GNWK> ||
+                     is_same_v<ELayout, ctc::GNHWK> || is_same_v<ELayout, ctc::GNDHWK> ||
+                     is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> ||
+                     is_same_v<ELayout, ctc::NDHWGK>)
+        {
+            const index_t K = arg.e_g_n_k_wos_lengths_[2];
+
+            if(!(K % CDEShuffleBlockTransferScalarPerVector_NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check Gridwise GEMM
+        return GridwiseOp::CheckValidity(arg.a_grid_desc_ak0_m_ak1_,
+                                         arg.b_grid_desc_bk0_n_bk1_,
+                                         arg.ds_grid_desc_m_n_,
+                                         arg.e_grid_desc_m_n_,
+                                         arg.block_2_etile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_g_n_c_wis_lengths,
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        ds_g_n_k_wos_lengths,
+                        ds_g_n_k_wos_strides,
+                        e_g_n_k_wos_lengths,
+                        e_g_n_k_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_g_n_c_wis_lengths,
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          ds_g_n_k_wos_lengths,
+                                          ds_g_n_k_wos_strides,
+                                          e_g_n_k_wos_lengths,
+                                          e_g_n_k_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvFwdMultipleD_Wmma_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization) << ", "
+            << K1 << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..bcef5c3b62
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,970 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+template <index_t NumDTensor>
+struct ComputePtrOffsetOfStridedBatch
+{
+    ComputePtrOffsetOfStridedBatch() = default;
+
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                   index_t BatchStrideB,
+                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
+                                   index_t BatchStrideE)
+        : BatchStrideA_(BatchStrideA),
+          BatchStrideB_(BatchStrideB),
+          BatchStrideDs_(BatchStrideDs),
+          BatchStrideE_(BatchStrideE)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumDTensor> ds_offset;
+        static_for<0, NumDTensor, 1>{}(
+            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
+        return ds_offset;
+    }
+
+    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    index_t BatchStrideA_;
+    index_t BatchStrideB_;
+    Array<ck::index_t, NumDTensor> BatchStrideDs_;
+    index_t BatchStrideE_;
+};
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for
+ * \link DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the
+ * computing of pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename ABDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2ETileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_conv_fwd_multiple_d_xdl_cshuffle(
+            const ABDataType* __restrict__ p_a_grid,
+            const ABDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const index_t batch_count,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock_,
+            const Block2ETileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    // offset base pointer for each work-group
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_ds_grid_grp,
+                                                  p_e_grid + e_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+#endif
+}
+
+} // namespace
+
+//
+// @brief      Device Convolution operation.
+//
+// Supports:
+//  @li         Forward convolution with up to 3 spatial dimentions
+//  @li         Input tensor in GNWC data format
+//  @li         Weight tensor in GKXC data format
+//  @li         Output tensor in GNWK data format
+//
+// 1D:
+// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
+// 2D:
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+// 3D:
+// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
+//
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ConvolutionForwardSpecialization ConvForwardSpecialization,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle
+    : public DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                           ALayout,
+                                           BLayout,
+                                           DsLayout,
+                                           ELayout,
+                                           ADataType,
+                                           BDataType,
+                                           DsDataType,
+                                           EDataType,
+                                           AElementwiseOperation,
+                                           BElementwiseOperation,
+                                           CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    template <typename ALay>
+    static auto
+    MakeAGridDescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                            const std::array<index_t, NDimSpatial>& input_left_pads,
+                            const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
+                                                                        a_g_n_c_wis_strides,
+                                                                        b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides,
+                                                                        e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides,
+                                                                        conv_filter_strides,
+                                                                        conv_filter_dilations,
+                                                                        input_left_pads,
+                                                                        input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+
+        return in_gemmm_gemmk_desc;
+    }
+
+    template <typename BLay>
+    static auto
+    MakeBGridDescriptor_N_K(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const auto wei_gemmnraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides);
+
+        const auto wei_gemmn_gemmk_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
+
+        return wei_gemmn_gemmk_desc;
+    }
+
+    template <typename ELay>
+    static auto
+    MakeEGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides)
+    {
+        const auto out_gemmmraw_gemmnraw_desc =
+            conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>(e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides);
+
+        const auto out_gemmm_gemmn_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(ds_g_n_k_wos_lengths[i],
+                                                                  ds_g_n_k_wos_strides[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // desc for problem definition
+    using AGridDesc_M_K  = remove_cvref_t<decltype(MakeAGridDescriptor_M_K<ALayout>(
+        {}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using BGridDesc_N_K  = remove_cvref_t<decltype(MakeBGridDescriptor_N_K<BLayout>({}, {}))>;
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
+    using EGridDesc_M_N  = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<ELayout>({}, {}))>;
+
+    using ComputeDataType = ADataType;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeDataType,
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // desc for blockwise copy
+    using AGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
+            AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
+            BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
+        decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+
+    // block-to-e-tile map
+    using Block2ETileMap =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a,
+                 const void* p_b,
+                 const std::array<const void*, NumDTensor>& p_ds,
+                 void* p_e,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_lengths,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads,
+                 const AElementwiseOperation& a_element_op,
+                 const BElementwiseOperation& b_element_op,
+                 const CDEElementwiseOperation& cde_element_op)
+            : p_a_grid_{static_cast<const ADataType*>(p_a)},
+              p_b_grid_{static_cast<const BDataType*>(p_b)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e)},
+              num_group_{a_g_n_c_wis_lengths[0]},
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K<ALayout>(a_g_n_c_wis_lengths,
+                                                                          a_g_n_c_wis_strides,
+                                                                          b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides,
+                                                                          e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K<BLayout>(b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              compute_ptr_offset_of_batch_{},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
+              a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              ds_g_n_k_wos_lengths_{ds_g_n_k_wos_lengths},
+              ds_g_n_k_wos_strides_{ds_g_n_k_wos_strides},
+              e_g_n_k_wos_lengths_{e_g_n_k_wos_lengths},
+              e_g_n_k_wos_strides_{e_g_n_k_wos_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // A/B/E Batch Stride
+            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_k_wos_strides[0];
+
+            // populate pointer, batch stride, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
+
+                // D batch stride
+                compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
+
+                // D desc
+                ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N<DLayout>(
+                    ds_g_n_k_wos_lengths[i], ds_g_n_k_wos_strides[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                           b_grid_desc_n_k_,
+                                           ds_grid_desc_m_n_,
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        index_t num_group_;
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch<NumDTensor> compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                            arg.b_grid_desc_n_k_,
+                                            arg.ds_grid_desc_m_n_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemmMultipleD_xdl_cshuffle has invalid setting");
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * arg.num_group_;
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_grouped_conv_fwd_multiple_d_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BK0_N_BK1,
+                    DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Block2ETileMap,
+                    ComputePtrOffsetOfStridedBatch<NumDTensor>,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_g_n_c_wis_lengths_[0], // Group count
+                                              arg.a_grid_desc_ak0_m_ak1_,
+                                              arg.b_grid_desc_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.block_2_etile_map_,
+                                              arg.compute_ptr_offset_of_batch_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        namespace ctc = tensor_layout::convolution;
+
+        // check device
+        if(get_device_name() == "gfx908")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t>))
+            {
+                return false;
+            }
+        }
+        else if(get_device_name() == "gfx90a" || get_device_name() == "gfx940" ||
+                get_device_name() == "gfx941" || get_device_name() == "gfx942")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check ConvolutionForwardSpecialization
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t ConvStride = arg.conv_filter_strides_[i];
+                const index_t LeftPad    = arg.input_left_pads_[i];
+                const index_t RightPad   = arg.input_right_pads_[i];
+
+                if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t LeftPad  = arg.input_left_pads_[i];
+                const index_t RightPad = arg.input_right_pads_[i];
+
+                if(!(X == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // check vector access of A
+        // FIXME: layout
+        if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
+                     is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
+                     is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
+                     is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
+                     is_same_v<ALayout, ctc::NDHWGC>)
+        {
+            const index_t C = arg.a_g_n_c_wis_lengths_[2];
+
+            if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of B
+        // FIXME: layout
+        if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
+                     is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
+                     is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
+                     is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
+                     is_same_v<BLayout, ctc::KZYXGC>)
+
+        {
+            const index_t C = arg.b_g_k_c_xs_lengths_[2];
+
+            if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        //  check vector access of Ds
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+            // FIXME: layout
+            if constexpr(is_same_v<DLayout, ctc::G_NW_K> || is_same_v<DLayout, ctc::G_NHW_K> ||
+                         is_same_v<DLayout, ctc::G_NDHW_K> || is_same_v<DLayout, ctc::GNWK> ||
+                         is_same_v<DLayout, ctc::GNHWK> || is_same_v<DLayout, ctc::GNDHWK> ||
+                         is_same_v<DLayout, ctc::NWGK> || is_same_v<DLayout, ctc::NHWGK> ||
+                         is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::GK> ||
+                         is_same_v<DLayout, ctc::G_K>)
+            {
+                const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
+
+                if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+                {
+                    valid = false;
+                }
+            }
+            else
+            {
+                valid = false;
+            }
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check vector access of E
+        if constexpr(is_same_v<ELayout, ctc::G_NW_K> || is_same_v<ELayout, ctc::G_NHW_K> ||
+                     is_same_v<ELayout, ctc::G_NDHW_K> || is_same_v<ELayout, ctc::GNWK> ||
+                     is_same_v<ELayout, ctc::GNHWK> || is_same_v<ELayout, ctc::GNDHWK> ||
+                     is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> ||
+                     is_same_v<ELayout, ctc::NDHWGK>)
+        {
+            const index_t K = arg.e_g_n_k_wos_lengths_[2];
+
+            if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check Gridwise GEMM
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                           arg.b_grid_desc_n_k_,
+                                           arg.ds_grid_desc_m_n_,
+                                           arg.e_grid_desc_m_n_,
+                                           arg.block_2_etile_map_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_g_n_c_wis_lengths,
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        ds_g_n_k_wos_lengths,
+                        ds_g_n_k_wos_strides,
+                        e_g_n_k_wos_lengths,
+                        e_g_n_k_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const void* p_a,
+        const void* p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_g_n_c_wis_lengths,
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          ds_g_n_k_wos_lengths,
+                                          ds_g_n_k_wos_strides,
+                                          e_g_n_k_wos_lengths,
+                                          e_g_n_k_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvFwdMultipleD_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization) << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp
new file mode 100644
index 0000000000..0190b3cee6
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp
@@ -0,0 +1,766 @@
+#pragma once
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename GemmDesc,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          bool HasMainKBlockLoop,
+          bool HasDoubleTailKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_gemm_multiple_d_dl(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
+                                          const index_t group_count,
+                                          const AElementwiseOperation a_element_op,
+                                          const BElementwiseOperation b_element_op,
+                                          const CDEElementwiseOperation cde_element_op)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||              \
+    defined(__gfx90a__) || defined(__gfx1030__) || defined(__gfx1100__) || defined(__gfx1101__) || \
+    defined(__gfx1102__) || defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t block_id = get_block_1d_id();
+
+    const auto gemm_desc_ptr =
+        reinterpret_cast<const GemmDesc*>(cast_pointer_to_generic_address_space(gemm_descs_const));
+
+    index_t left     = 0;
+    index_t right    = group_count;
+    index_t group_id = index_t((left + right) / 2);
+    while((!(block_id >= gemm_desc_ptr[group_id].BlockStart_ &&
+             block_id < gemm_desc_ptr[group_id].BlockEnd_)) &&
+          left <= right)
+    {
+        if(block_id < gemm_desc_ptr[group_id].BlockStart_)
+        {
+            right = group_id;
+        }
+        else
+        {
+            left = group_id;
+        }
+        group_id = index_t((left + right) / 2);
+    }
+
+    GridwiseGemm::Run(gemm_desc_ptr[group_id].a_ptr_,
+                      gemm_desc_ptr[group_id].b_ptr_,
+                      gemm_desc_ptr[group_id].ds_ptr_,
+                      gemm_desc_ptr[group_id].e_ptr_,
+                      p_shared,
+                      a_element_op,
+                      b_element_op,
+                      cde_element_op,
+                      gemm_desc_ptr[group_id].a_grid_desc_k0_m0_m1_k1_,
+                      gemm_desc_ptr[group_id].b_grid_desc_k0_n0_n1_k1_,
+                      gemm_desc_ptr[group_id].ds_grid_desc_m0_m10_m11_n0_n10_n11_,
+                      gemm_desc_ptr[group_id].e_grid_desc_m0_m10_m11_n0_n10_n11_,
+                      gemm_desc_ptr[group_id].block_2_etile_map_,
+                      integral_constant<bool, HasMainKBlockLoop>{},
+                      integral_constant<bool, HasDoubleTailKBlockLoop>{});
+#else
+    ignore = gemm_descs_const;
+    ignore = group_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+#endif
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t K1,
+          index_t M1PerThread,
+          index_t N1PerThread,
+          index_t KPerThread,
+          typename M1N1ThreadClusterM1Xs,
+          typename M1N1ThreadClusterN1Xs,
+          typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+          typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+          typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+          typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+          typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          enable_if_t<
+              is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+                  is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+              bool> = false>
+struct DeviceGroupedGemmMultipleD_Dl : public DeviceGroupedGemm<ALayout,
+                                                                BLayout,
+                                                                DsLayout,
+                                                                ELayout,
+                                                                ADataType,
+                                                                BDataType,
+                                                                DsDataType,
+                                                                EDataType,
+                                                                AElementwiseOperation,
+                                                                BElementwiseOperation,
+                                                                CDEElementwiseOperation>
+{
+    using DeviceOp                      = DeviceGroupedGemmMultipleD_Dl;
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static constexpr auto K1Number = Number<K1>{};
+
+    static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(M, PadM)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
+    {
+        assert(K % K1 == 0);
+
+        const index_t K0 = K / K1;
+
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    template <typename ELay>
+    static auto MakeEGridDescriptor_M_N(index_t M, index_t N, index_t StrideE)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideE));
+            }
+        }();
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding)
+        {
+            const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+    }
+
+    static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                                         const std::array<index_t, NumDTensor>& NRaws,
+                                         const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
+    using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
+    using DsGridDesc_M_N    = decltype(MakeDsGridDescriptor_M_N({}, {}, {}));
+    using EGridDesc_M_N     = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
+
+    // GridwiseGemm
+    using GridwiseGemm =
+        GridwiseGemmDlMultipleD_km_kn_mn<BlockSize,
+                                         ADataType,
+                                         AccDataType,
+                                         DsDataType,
+                                         EDataType,
+                                         AElementwiseOperation,
+                                         BElementwiseOperation,
+                                         CDEElementwiseOperation,
+                                         InMemoryDataOperationEnum::Set,
+                                         AGridDesc_K0_M_K1,
+                                         BGridDesc_K0_N_K1,
+                                         EGridDesc_M_N,
+                                         MPerBlock,
+                                         NPerBlock,
+                                         K0PerBlock,
+                                         K1,
+                                         M1PerThread,
+                                         N1PerThread,
+                                         KPerThread,
+                                         M1N1ThreadClusterM1Xs,
+                                         M1N1ThreadClusterN1Xs,
+                                         ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+                                         ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+                                         ABlockTransferThreadClusterArrangeOrder,
+                                         ABlockTransferSrcAccessOrder,
+                                         ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+                                         ABlockTransferSrcVectorTensorContiguousDimOrder,
+                                         ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+                                         BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+                                         BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+                                         BBlockTransferThreadClusterArrangeOrder,
+                                         BBlockTransferSrcAccessOrder,
+                                         BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+                                         BBlockTransferSrcVectorTensorContiguousDimOrder,
+                                         BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+                                         CThreadTransferSrcDstAccessOrder,
+                                         CThreadTransferSrcDstVectorDim,
+                                         CThreadTransferDstScalarPerVector>;
+
+    using AGridDesc_K0_M0_M1_K1 =
+        decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_K0_M_K1{}));
+    using BGridDesc_K0_N0_N1_K1 =
+        decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{}));
+    using DsGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(DsGridDesc_M_N{}));
+    using EGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(EGridDesc_M_N{}));
+
+    struct GroupedGemmBlock2ETileMap
+    {
+        using Block2ETileMap =
+            remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(EGridDesc_M_N{}))>;
+
+        GroupedGemmBlock2ETileMap()
+        {
+            block_2_etile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(EGridDesc_M_N{});
+            BlockStart_        = -1;
+        }
+
+        GroupedGemmBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n, ck::index_t BlockStart)
+        {
+            block_2_etile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(e_grid_desc_m_n);
+            BlockStart_        = BlockStart;
+        }
+
+        template <typename TopIdx>
+        __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+        {
+            return block_2_etile_map_.CalculateBottomIndex(
+                make_multi_index(idx_top[I0] - BlockStart_));
+        }
+
+        // it's actually E-Tile
+        template <typename CTileIdx, typename CTileDim>
+        __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                                 const CTileDim& c_tile_dim) const
+        {
+            return block_2_etile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
+        }
+
+        __host__ bool CheckValidity(const EGridDesc_M_N& e_grid_desc_m_n) const
+        {
+            return block_2_etile_map_.CheckValidity(e_grid_desc_m_n);
+        }
+
+        Block2ETileMap block_2_etile_map_;
+        ck::index_t BlockStart_;
+    };
+
+    struct GemmKernelArg
+    {
+        // pointers
+        const ADataType* a_ptr_;
+        const BDataType* b_ptr_;
+        typename GridwiseGemm::DsGridPointer ds_ptr_;
+        EDataType* e_ptr_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1_;
+        BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1_;
+        DsGridDesc_M0_M10_M11_N0_N10_N11 ds_grid_desc_m0_m10_m11_n0_n10_n11_;
+        EGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11_;
+
+        // block-to-e-tile map
+        GroupedGemmBlock2ETileMap block_2_etile_map_;
+        ck::index_t BlockStart_, BlockEnd_;
+    };
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(std::vector<const void*>& p_As,
+                 std::vector<const void*>& p_Bs,
+                 std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                 std::vector<void*>& p_Es,
+                 std::vector<GemmDesc>& gemm_descs,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op)
+            : a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op}
+        {
+            grid_size_ = 0;
+
+            group_count_ = ck::type_convert<ck::index_t>(gemm_descs.size());
+
+            if(!(group_count_ == ck::type_convert<ck::index_t>(p_As.size()) &&
+                 group_count_ == ck::type_convert<ck::index_t>(p_Bs.size()) &&
+                 group_count_ == ck::type_convert<ck::index_t>(p_Es.size())))
+            {
+                throw std::runtime_error("wrong! group_count_ != p_As/b/c.size");
+            }
+
+            gemm_desc_kernel_arg_.reserve(group_count_);
+
+            skipped_group_count_ = 0;
+
+            for(std::size_t i = 0; i < gemm_descs.size(); i++)
+            {
+                const index_t M = gemm_descs[i].M_;
+                const index_t N = gemm_descs[i].N_;
+                const index_t K = gemm_descs[i].K_;
+
+                a_mtx_mraw_kraw_.emplace_back(M, K);
+                b_mtx_nraw_kraw_.emplace_back(N, K);
+
+                if(M == 0)
+                {
+                    skipped_group_count_++;
+                    continue;
+                }
+
+                const index_t StrideA = gemm_descs[i].stride_A_;
+                const index_t StrideB = gemm_descs[i].stride_B_;
+                const index_t StrideE = gemm_descs[i].stride_C_;
+
+                typename GridwiseGemm::DsGridPointer p_ds_grid{};
+                DsGridDesc_M_N ds_grid_desc_m_n;
+
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    using DLayout   = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+                    using DDataType = remove_cvref_t<tuple_element_t<j.value, DsDataType>>;
+
+                    p_ds_grid(j)        = static_cast<const DDataType*>(p_Ds[i][j]);
+                    ds_grid_desc_m_n(j) = DeviceOp::MakeEGridDescriptor_M_N<DLayout>(
+                        M, N, gemm_descs[i].stride_Ds_[j]);
+                });
+
+                // tensor descriptors for problem definiton
+                const auto a_grid_desc_k0_m_k1 =
+                    DeviceOp::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
+                const auto b_grid_desc_k0_n_k1 =
+                    DeviceOp::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
+                const auto e_grid_desc_m_n =
+                    DeviceOp::MakeEGridDescriptor_M_N<ELayout>(M, N, StrideE);
+
+                if(GridwiseGemm::CheckValidity(
+                       a_grid_desc_k0_m_k1, b_grid_desc_k0_n_k1, e_grid_desc_m_n))
+                {
+
+                    const index_t grid_size_grp =
+                        GroupedGemmBlock2ETileMap(e_grid_desc_m_n, 0)
+                            .block_2_etile_map_.CalculateGridSize(e_grid_desc_m_n);
+
+                    const index_t BlockStart = grid_size_;
+                    const index_t BlockEnd   = grid_size_ + grid_size_grp;
+
+                    grid_size_ += grid_size_grp;
+
+                    // block-to-e-tile map
+                    const auto block_2_etile_map =
+                        GroupedGemmBlock2ETileMap(e_grid_desc_m_n, BlockStart);
+
+                    // tensor descriptors for block/thread-wise copy
+                    const auto a_grid_desc_k0_m0_m1_k1 =
+                        GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(a_grid_desc_k0_m_k1);
+                    const auto b_grid_desc_k0_n0_n1_k1 =
+                        GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(b_grid_desc_k0_n_k1);
+                    const auto ds_grid_desc_m0_m10_m11_n0_n10_n11 =
+                        GridwiseGemm::MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(ds_grid_desc_m_n);
+                    const auto e_grid_desc_m0_m10_m11_n0_n10_n11 =
+                        GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(e_grid_desc_m_n);
+
+                    gemm_desc_kernel_arg_.push_back(
+                        GemmKernelArg{static_cast<const ADataType*>(p_As[i]),
+                                      static_cast<const BDataType*>(p_Bs[i]),
+                                      p_ds_grid,
+                                      static_cast<EDataType*>(p_Es[i]),
+                                      a_grid_desc_k0_m_k1,
+                                      b_grid_desc_k0_n_k1,
+                                      ds_grid_desc_m_n,
+                                      e_grid_desc_m_n,
+                                      a_grid_desc_k0_m0_m1_k1,
+                                      b_grid_desc_k0_n0_n1_k1,
+                                      ds_grid_desc_m0_m10_m11_n0_n10_n11,
+                                      e_grid_desc_m0_m10_m11_n0_n10_n11,
+                                      block_2_etile_map,
+                                      BlockStart,
+                                      BlockEnd});
+                }
+            }
+        }
+
+        //  private:
+        index_t group_count_;
+        index_t skipped_group_count_;
+
+        // TODO: A,B element op is unused since gridwise_gemm_dl_v1r3 does NOT support prologue
+        //       for the time being.
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        std::vector<GemmKernelArg> gemm_desc_kernel_arg_;
+        std::vector<Tuple<index_t, index_t>> a_mtx_mraw_kraw_;
+        std::vector<Tuple<index_t, index_t>> b_mtx_nraw_kraw_;
+
+        index_t grid_size_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            auto K0 = arg.gemm_desc_kernel_arg_[0].a_grid_desc_k0_m_k1_.GetLength(I0);
+            bool all_has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0);
+            bool all_has_double_tail_k_block_loop =
+                GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
+
+            for(std::size_t i = 0; i < arg.gemm_desc_kernel_arg_.size(); i++)
+            {
+#if DEBUG_LOG
+                std::cout << "group: " << i << " arg.a_grid_desc_k0_m_k1_{"
+                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_k0_m_k1_.GetLength(I0) << ", "
+                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
+                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_k0_m_k1_.GetLength(I2) << "}"
+                          << std::endl;
+
+                std::cout << ", arg.b_grid_desc_k0_n_k1_{"
+                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_k0_n_k1_.GetLength(I0) << ", "
+                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
+                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_k0_n_k1_.GetLength(I2) << "}"
+                          << std::endl;
+
+                std::cout << ", arg.e_grid_desc_m_n_{ "
+                          << arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_.GetLength(I1) << "}"
+                          << std::endl;
+#endif
+
+                if(!GridwiseGemm::CheckValidity(arg.gemm_desc_kernel_arg_[i].a_grid_desc_k0_m_k1_,
+                                                arg.gemm_desc_kernel_arg_[i].b_grid_desc_k0_n_k1_,
+                                                arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_))
+                {
+                    throw std::runtime_error(
+                        "wrong! GridwiseGemmDlMultipleD_km_kn_mn has invalid setting");
+                }
+
+                K0 = arg.gemm_desc_kernel_arg_[i].a_grid_desc_k0_m0_m1_k1_.GetLength(I0);
+                bool not_all_has_main_k_block_loop_same =
+                    all_has_main_k_block_loop xor GridwiseGemm::CalculateHasMainKBlockLoop(K0);
+                bool not_all_has_double_tail_k_block_loop_same =
+                    all_has_double_tail_k_block_loop xor
+                    GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0);
+
+                if(not_all_has_main_k_block_loop_same or not_all_has_double_tail_k_block_loop_same)
+                {
+                    std::ostringstream err;
+                    err << "Not all gemms have same value for [main|double_tail]_k_block_loop! in "
+                        << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
+                    throw std::runtime_error(err.str());
+                }
+            }
+
+            hipGetErrorString(
+                hipMemcpyWithStream(arg.p_workspace_,
+                                    arg.gemm_desc_kernel_arg_.data(),
+                                    arg.gemm_desc_kernel_arg_.size() * sizeof(GemmKernelArg),
+                                    hipMemcpyHostToDevice,
+                                    stream_config.stream_id_));
+
+            auto launch_kernel = [&](auto has_main_k_block_loop,
+                                     auto has_double_tail_k_block_loop) {
+                constexpr bool has_main_loop   = has_main_k_block_loop.value;
+                constexpr bool has_double_loop = has_double_tail_k_block_loop.value;
+
+                const auto kernel = kernel_grouped_gemm_multiple_d_dl<GridwiseGemm,
+                                                                      GemmKernelArg,
+                                                                      AElementwiseOperation,
+                                                                      BElementwiseOperation,
+                                                                      CDEElementwiseOperation,
+                                                                      has_main_loop,
+                                                                      has_double_loop>;
+
+                return launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(arg.grid_size_),
+                    dim3(BlockSize),
+                    0,
+                    cast_pointer_to_constant_address_space(arg.p_workspace_),
+                    arg.gemm_desc_kernel_arg_.size(),
+                    arg.a_element_op_,
+                    arg.b_element_op_,
+                    arg.cde_element_op_);
+            };
+
+            if(all_has_main_k_block_loop && all_has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, true>{});
+            }
+            else if(all_has_main_k_block_loop && !all_has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, true>{},
+                                     integral_constant<bool, false>{});
+            }
+            else if(!all_has_main_k_block_loop && all_has_double_tail_k_block_loop)
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{},
+                                     integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if((ck::type_convert<ck::index_t>(arg.gemm_desc_kernel_arg_.size()) +
+            arg.skipped_group_count_) != arg.group_count_)
+        {
+            return false;
+        }
+
+        const std::string device_name = ck::get_device_name();
+
+        //  TODO add newer Navi arch
+        if(device_name != "gfx906" and device_name != "gfx908" and device_name != "gfx90a" and
+           device_name != "gfx1030")
+        {
+            return false;
+        }
+
+        for(std::size_t i = 0; i < arg.gemm_desc_kernel_arg_.size(); i++)
+        {
+            if(!GridwiseGemm::CheckValidity(arg.gemm_desc_kernel_arg_[i].a_grid_desc_k0_m_k1_,
+                                            arg.gemm_desc_kernel_arg_[i].b_grid_desc_k0_n_k1_,
+                                            arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_))
+            {
+                return false;
+            }
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::vector<const void*>& p_As,
+                             std::vector<const void*>& p_Bs,
+                             std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                             std::vector<void*>& p_Es,
+                             std::vector<GemmDesc> gemm_descs,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation cde_element_op)
+    {
+        return Argument{
+            p_As, p_Bs, p_Ds, p_Es, gemm_descs, a_element_op, b_element_op, cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*>& p_As,
+                        std::vector<const void*>& p_Bs,
+                        std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                        std::vector<void*>& p_Es,
+                        std::vector<GemmDesc>& gemm_descs,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return std::make_unique<Argument>(
+            p_As, p_Bs, p_Ds, p_Es, gemm_descs, a_element_op, b_element_op, cde_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedGemmMultipleD_Dl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock << ", "
+            << K1 << ", "
+            << M1PerThread << ", "
+            << N1PerThread << ", "
+            << KPerThread
+            << getGemmSpecializationString(GemmSpec)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
+    {
+        return dynamic_cast<const Argument*>(p_arg)->group_count_ * sizeof(GemmKernelArg);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..ddd762af1c
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
@@ -0,0 +1,893 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp"
+#include "ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename GroupKernelArg,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_gemm_softmax_gemm_xdl_cshuffle_v1(
+            const void CK_CONSTANT_ADDRESS_SPACE* group_kernel_args,
+            const index_t group_count,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const AccElementwiseOperation acc_element_op,
+            const B1ElementwiseOperation b1_element_op,
+            const CElementwiseOperation c_element_op)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t block_id = get_block_1d_id();
+
+    const auto arg_ptr = reinterpret_cast<const GroupKernelArg*>(
+        cast_pointer_to_generic_address_space(group_kernel_args));
+
+    index_t left     = 0;
+    index_t right    = group_count;
+    index_t group_id = index_t((left + right) / 2);
+
+    while(
+        (!(block_id >= arg_ptr[group_id].block_start_ && block_id < arg_ptr[group_id].block_end_)))
+    {
+        if(block_id < arg_ptr[group_id].block_start_)
+        {
+            right = group_id;
+        }
+        else
+        {
+            left = group_id;
+        }
+        group_id = index_t((left + right) / 2);
+    }
+
+    // per-group batch offset
+    const index_t num_blocks_per_batch = arg_ptr[group_id].num_blocks_per_batch_;
+    const index_t g_idx                = __builtin_amdgcn_readfirstlane(
+        (block_id - arg_ptr[group_id].block_start_) / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetABasePtr(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetBBasePtr(g_idx)));
+    const long_index_t b1_batch_offset = __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
+        arg_ptr[group_id].compute_base_ptr_of_batch_.GetB1BasePtr(g_idx)));
+    const long_index_t c_batch_offset  = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(arg_ptr[group_id].compute_base_ptr_of_batch_.GetCBasePtr(g_idx)));
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        arg_ptr[group_id].p_a_grid_ + a_batch_offset,
+        arg_ptr[group_id].p_b_grid_ + b_batch_offset,
+        arg_ptr[group_id].p_b1_grid_ + b1_batch_offset,
+        arg_ptr[group_id].p_c_grid_ + c_batch_offset,
+        p_shared,
+        a_element_op,
+        b_element_op,
+        acc_element_op,
+        b1_element_op,
+        c_element_op,
+        arg_ptr[group_id].a_grid_desc_ak0_m_ak1_,
+        arg_ptr[group_id].b_grid_desc_bk0_n_bk1_,
+        arg_ptr[group_id].b1_grid_desc_bk0_n_bk1_,
+        arg_ptr[group_id].c_grid_desc_mblock_mperblock_nblock_nperblock_,
+        arg_ptr[group_id].block_2_ctile_map_,
+        arg_ptr[group_id].c0_matrix_mask_);
+#else
+    ignore = group_kernel_args;
+    ignore = group_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = acc_element_op;
+    ignore = b1_element_op;
+    ignore = c_element_op;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// Computes C = A * B0 * B1
+//              ^^^^^^ (Acc0)
+//              ^^^^^^^^^^^ (Acc1)
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO, // NumDimGemm1N
+          typename ADataType,
+          typename BDataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasDataType,
+          typename Acc1BiasDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization B1Spec,
+          TensorSpecialization CSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock, // Gemm0NPerBlock
+          index_t KPerBlock, // Gemm0KPerBlock
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t B1K1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          MaskingSpecialization MaskingSpec,
+          LoopScheduler LoopSched = LoopScheduler::Default>
+struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
+    : public DeviceGroupedGemmSoftmaxGemmPermute<NumDimG,
+                                                 NumDimM,
+                                                 NumDimN,
+                                                 NumDimK,
+                                                 NumDimO,
+                                                 ADataType,
+                                                 BDataType,
+                                                 B1DataType,
+                                                 CDataType,
+                                                 Acc0BiasDataType,
+                                                 Acc1BiasDataType,
+                                                 AElementwiseOperation,
+                                                 BElementwiseOperation,
+                                                 AccElementwiseOperation,
+                                                 B1ElementwiseOperation,
+                                                 CElementwiseOperation,
+                                                 MaskingSpec>
+{
+    static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0 && NumDimO > 0,
+                  "Number of dimension must be greater than 0");
+
+    static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
+    static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
+
+    // TODO ANT: implement bias combination
+    static_assert(NumAcc0Bias == 0 && NumAcc0Bias == 0, "Bias addition is unimplemented");
+
+#if 0
+    // TODO ANT: use alias
+    static constexpr index_t NumDimGemm0M = NumDimM;
+    static constexpr index_t NumDimGemm0N = NumDimN;
+    static constexpr index_t NumDimGemm0K = NumDimK;
+    static constexpr index_t NumDimGemm1M = NumDimM;
+    static constexpr index_t NumDimGemm1N = NumDimO;
+    static constexpr index_t NumDimGemm1K = NumDimN;
+#endif
+
+    using DeviceOp    = DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle;
+    using ProblemDesc = typename DeviceGroupedGemmSoftmaxGemmPermute<NumDimG,
+                                                                     NumDimM,
+                                                                     NumDimN,
+                                                                     NumDimK,
+                                                                     NumDimO,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     B1DataType,
+                                                                     CDataType,
+                                                                     Acc0BiasDataType,
+                                                                     Acc1BiasDataType,
+                                                                     AElementwiseOperation,
+                                                                     BElementwiseOperation,
+                                                                     AccElementwiseOperation,
+                                                                     B1ElementwiseOperation,
+                                                                     CElementwiseOperation,
+                                                                     MaskingSpec>::ProblemDesc;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    using Transform = TransformBatchedContractionContractionToBatchedGemmGemm<
+        Sequence<NumDimG, NumDimM, NumDimN, NumDimK, NumDimO>,
+        Sequence<MPerBlock, NPerBlock, KPerBlock, Gemm1NPerBlock>,
+        GemmSpec,
+        ASpec,
+        BSpec,
+        B1Spec,
+        CSpec>;
+
+    static auto MakeAGridDescriptor_AK0_M_AK1(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                              const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        return Transform::MakeAGridDescriptor_AK0_M_AK1(
+            Transform::MakeAGridDescriptor_M_K(a_gs_ms_ks_lengths_vec, a_gs_ms_ks_strides_vec),
+            Number<AK1>{});
+    }
+
+    static auto MakeBGridDescriptor_BK0_N_BK1(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
+                                              const std::vector<index_t>& b_gs_ns_ks_strides_vec)
+    {
+        return Transform::MakeB0GridDescriptor_BK0_N_BK1(
+            Transform::MakeB0GridDescriptor_N_K(b_gs_ns_ks_lengths_vec, b_gs_ns_ks_strides_vec),
+            Number<BK1>{});
+    }
+
+    static auto
+    MakeB1GridDescriptor_BK0_N_BK1(const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_lengths_vec,
+                                   const std::vector<index_t>& b1_gs_gemm1ns_gemm1ks_strides_vec)
+    {
+        return Transform::MakeB1GridDescriptor_BK0_N_BK1(
+            Transform::MakeB1GridDescriptor_N_K(b1_gs_gemm1ns_gemm1ks_lengths_vec,
+                                                b1_gs_gemm1ns_gemm1ks_strides_vec),
+            Number<B1K1>{});
+    }
+
+    using AGridDesc_AK0_M_AK1  = decltype(MakeAGridDescriptor_AK0_M_AK1({}, {}));
+    using BGridDesc_BK0_N_BK1  = decltype(MakeBGridDescriptor_BK0_N_BK1({}, {}));
+    using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1({}, {}));
+    using CGridDesc_M_N        = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
+    using AGridDesc_G_M_K      = decltype(Transform::MakeAGridDescriptor_G_M_K({}, {}));
+    using BGridDesc_G_N_K      = decltype(Transform::MakeB0GridDescriptor_G_N_K({}, {}));
+    using B1GridDesc_G_N_K     = decltype(Transform::MakeB1GridDescriptor_G_N_K({}, {}));
+    using CGridDesc_G_M_N      = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
+
+    constexpr static auto make_MaskOutPredicate()
+    {
+        if constexpr(MaskingSpec == MaskingSpecialization::MaskDisabled)
+        {
+            return MaskDisabledPredicate{};
+        }
+        else if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+        {
+            return MaskOutUpperTrianglePredicate{};
+        }
+    }
+    using C0MatrixMask = C0MatrixMask_impl<decltype(make_MaskOutPredicate())>;
+
+    struct ComputeBasePtrOfStridedBatch
+    {
+        ComputeBasePtrOfStridedBatch(const AGridDesc_G_M_K& a_grid_desc_g_m_k,
+                                     const BGridDesc_G_N_K& b_grid_desc_g_n_k,
+                                     const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
+                                     const CGridDesc_G_M_N& c_grid_desc_g_m_n)
+            : a_grid_desc_g_m_k_(a_grid_desc_g_m_k),
+              b_grid_desc_g_n_k_(b_grid_desc_g_n_k),
+              b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
+              c_grid_desc_g_m_n_(c_grid_desc_g_m_n)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetABasePtr(index_t g_idx) const
+        {
+            return a_grid_desc_g_m_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetBBasePtr(index_t g_idx) const
+        {
+            return b_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
+        {
+            return b1_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        __host__ __device__ constexpr long_index_t GetCBasePtr(index_t g_idx) const
+        {
+            return c_grid_desc_g_m_n_.CalculateOffset(make_multi_index(g_idx, 0, 0));
+        }
+
+        private:
+        AGridDesc_G_M_K a_grid_desc_g_m_k_;
+        BGridDesc_G_N_K b_grid_desc_g_n_k_;
+        B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
+        CGridDesc_G_M_N c_grid_desc_g_m_n_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        AccElementwiseOperation,
+        B1ElementwiseOperation,
+        CElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_AK0_M_AK1,
+        BGridDesc_BK0_N_BK1,
+        B1GridDesc_BK0_N_BK1,
+        CGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        Gemm1NPerBlock,
+        Gemm1KPerBlock,
+        AK1,
+        BK1,
+        B1K1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        Gemm1NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        true,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        true,
+        BBlockLdsExtraN,
+        B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+        B1BlockTransferThreadClusterArrangeOrder,
+        B1BlockTransferSrcAccessOrder,
+        B1BlockTransferSrcVectorDim,
+        B1BlockTransferSrcScalarPerVector,
+        B1BlockTransferDstScalarPerVector_BK1,
+        false,
+        B1BlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        LoopSched,
+        Transform::matrix_padder.PadN,
+        MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle>;
+
+    using Block2CTileMap = OffsettedBlockToCTileMap<typename GridwiseGemm::DefaultBlock2CTileMap>;
+
+    struct GroupKernelArg
+    {
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        const B1DataType* p_b1_grid_;
+        CDataType* p_c_grid_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
+        typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            c_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // batch & stride
+        index_t num_blocks_per_batch_;
+        ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch_;
+
+        // check C0 masking and padding
+        C0MatrixMask c0_matrix_mask_;
+
+        // block-to-c-tile map
+        Block2CTileMap block_2_ctile_map_;
+
+        index_t block_start_, block_end_;
+    };
+
+    struct GroupDeviceArg
+    {
+        // lengths for the last dimensions of overall problem for sanity check of vector load/store
+        std::vector<index_t> raw_lengths_mz_nz_kz_gemm1nz_;
+
+        // strides for the last dimensions of each tensor for sanity check of vector load/store
+        std::vector<index_t> a_mz_kz_strides_;
+        std::vector<index_t> b_nz_kz_strides_;
+        std::vector<index_t> b1_nz_kz_strides_;
+        std::vector<index_t> c_mz_gemm1nz_strides_;
+
+        // for gridwise gemm check
+        CGridDesc_M_N c_grid_desc_m_n_;
+    };
+
+    // Argument
+    // FIXME: constness
+    struct Argument : public BaseArgument
+    {
+        Argument(std::vector<const void*> p_a_vec,
+                 std::vector<const void*> p_b_vec,
+                 std::vector<const void*> p_b1_vec,
+                 std::vector<void*> p_c_vec,
+                 std::vector<std::vector<const void*>> p_acc0_biases_vec,
+                 std::vector<std::vector<const void*>> p_acc1_biases_vec,
+                 std::vector<ProblemDesc> problem_desc_vec,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 AccElementwiseOperation acc_element_op,
+                 B1ElementwiseOperation b1_element_op,
+                 CElementwiseOperation c_element_op)
+            : a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              acc_element_op_{acc_element_op},
+              b1_element_op_{b1_element_op},
+              c_element_op_{c_element_op}
+        {
+            // TODO ANT: implement bias addition
+            group_count_ = problem_desc_vec.size();
+
+            if(!(group_count_ == p_a_vec.size() && group_count_ == p_b_vec.size() &&
+                 group_count_ == p_b1_vec.size() && group_count_ == p_c_vec.size()))
+            {
+                throw std::runtime_error("wrong! group_count_ != a/b/b1/c_vec.size");
+            }
+
+            if(!(p_acc0_biases_vec.size() == p_acc1_biases_vec.size()))
+            {
+                throw std::runtime_error("wrong! acc0_bias_vec.size != acc1_bias_vec.size");
+            }
+
+            grid_size_ = 0;
+
+            for(std::size_t i = 0; i < group_count_; i++)
+            {
+                const auto p_a_grid  = static_cast<const ADataType*>(p_a_vec[i]);
+                const auto p_b_grid  = static_cast<const BDataType*>(p_b_vec[i]);
+                const auto p_b1_grid = static_cast<const B1DataType*>(p_b1_vec[i]);
+                const auto p_c_grid  = static_cast<CDataType*>(p_c_vec[i]);
+
+                const auto& problem_desc = problem_desc_vec[i];
+
+                const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
+                    problem_desc.a_gs_ms_ks_lengths, problem_desc.a_gs_ms_ks_strides);
+                const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
+                    problem_desc.b0_gs_ns_ks_lengths, problem_desc.b0_gs_ns_ks_strides);
+                const auto b1_grid_desc_bk0_n_bk1 = MakeB1GridDescriptor_BK0_N_BK1(
+                    problem_desc.b1_gs_os_ns_lengths, problem_desc.b1_gs_os_ns_strides);
+                const auto c_grid_desc_m_n = Transform::MakeCGridDescriptor_M_N(
+                    problem_desc.c_gs_ms_os_lengths, problem_desc.c_gs_ms_os_strides);
+
+                const auto a_grid_desc_g_m_k = Transform::MakeAGridDescriptor_G_M_K(
+                    problem_desc.a_gs_ms_ks_lengths, problem_desc.a_gs_ms_ks_strides);
+                const auto b_grid_desc_g_n_k = Transform::MakeB0GridDescriptor_G_N_K(
+                    problem_desc.b0_gs_ns_ks_lengths, problem_desc.b0_gs_ns_ks_strides);
+                const auto b1_grid_desc_g_n_k = Transform::MakeB1GridDescriptor_G_N_K(
+                    problem_desc.b1_gs_os_ns_lengths, problem_desc.b1_gs_os_ns_strides);
+                const auto c_grid_desc_g_m_n = Transform::MakeCGridDescriptor_G_M_N(
+                    problem_desc.c_gs_ms_os_lengths, problem_desc.c_gs_ms_os_strides);
+
+                const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+                    GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        c_grid_desc_m_n);
+
+                const index_t BlockStart     = grid_size_;
+                const auto block_2_ctile_map = Block2CTileMap(c_grid_desc_m_n, BlockStart);
+                const index_t batch_count    = c_grid_desc_g_m_n.GetLength(I0);
+                const index_t grid_size_grp =
+                    block_2_ctile_map.CalculateGridSize(c_grid_desc_m_n) * batch_count;
+                const index_t BlockEnd = grid_size_ + grid_size_grp;
+
+                // batch stride
+                const auto compute_base_ptr_of_batch = ComputeBasePtrOfStridedBatch(
+                    a_grid_desc_g_m_k, b_grid_desc_g_n_k, b1_grid_desc_g_n_k, c_grid_desc_g_m_n);
+
+                // C0 mask
+                const auto c0_matrix_mask = C0MatrixMask(b_grid_desc_g_n_k.GetLength(I1));
+
+                grid_size_ += grid_size_grp;
+
+                // for each group, make sure acc0_biases_gs_ms_ns_lengths.size() == NumAcc0Bias and
+                // so on
+                if(!(problem_desc.acc0_biases_gs_ms_ns_lengths.size() == NumAcc0Bias &&
+                     problem_desc.acc0_biases_gs_ms_ns_strides.size() == NumAcc0Bias &&
+                     problem_desc.acc1_biases_gs_ms_os_lengths.size() == NumAcc1Bias &&
+                     problem_desc.acc1_biases_gs_ms_os_strides.size() == NumAcc1Bias))
+                {
+                    throw std::runtime_error(
+                        "wrong! number of biases in function argument does not "
+                        "match that in template argument");
+                }
+
+                group_kernel_args_.push_back({p_a_grid,
+                                              p_b_grid,
+                                              p_b1_grid,
+                                              p_c_grid,
+                                              a_grid_desc_ak0_m_ak1,
+                                              b_grid_desc_bk0_n_bk1,
+                                              b1_grid_desc_bk0_n_bk1,
+                                              c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                              block_2_ctile_map.CalculateGridSize(c_grid_desc_m_n),
+                                              compute_base_ptr_of_batch,
+                                              c0_matrix_mask,
+                                              block_2_ctile_map,
+                                              BlockStart,
+                                              BlockEnd});
+
+                group_device_args_.push_back(
+                    {{problem_desc.a_gs_ms_ks_lengths[NumDimG + NumDimM - 1],
+                      problem_desc.b0_gs_ns_ks_lengths[NumDimG + NumDimN - 1],
+                      problem_desc.b0_gs_ns_ks_lengths[NumDimG + NumDimN + NumDimK - 1],
+                      problem_desc.b1_gs_os_ns_lengths[NumDimG + NumDimO - 1]},
+                     {problem_desc.a_gs_ms_ks_strides[NumDimG + NumDimM - 1],
+                      problem_desc.a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1]},
+                     {problem_desc.b0_gs_ns_ks_strides[NumDimG + NumDimN - 1],
+                      problem_desc.b0_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1]},
+                     {problem_desc.b1_gs_os_ns_strides[NumDimG + NumDimO - 1],
+                      problem_desc.b1_gs_os_ns_strides[NumDimG + NumDimO + NumDimN - 1]},
+                     {problem_desc.c_gs_ms_os_strides[NumDimG + NumDimM - 1],
+                      problem_desc.c_gs_ms_os_strides[NumDimG + NumDimM + NumDimO - 1]},
+                     c_grid_desc_m_n});
+            }
+        }
+
+        std::vector<GroupKernelArg> group_kernel_args_;
+        std::vector<GroupDeviceArg> group_device_args_;
+
+        std::size_t group_count_;
+        index_t grid_size_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        AccElementwiseOperation acc_element_op_;
+        B1ElementwiseOperation b1_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!DeviceOp::IsSupportedArgument(arg))
+            {
+                throw std::runtime_error("wrong! unsupported argument");
+            }
+
+            bool all_has_main_k_block_loop  = true;
+            bool some_has_main_k_block_loop = false;
+            for(std::size_t i = 0; i < arg.group_count_; i++)
+            {
+                const auto K = arg.group_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0) *
+                               arg.group_kernel_args_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2);
+                const bool y = GridwiseGemm::CalculateHasMainKBlockLoop(K);
+                all_has_main_k_block_loop &= y;
+                some_has_main_k_block_loop |= y;
+            }
+
+            hipGetErrorString(
+                hipMemcpyWithStream(arg.p_workspace_,
+                                    arg.group_kernel_args_.data(),
+                                    arg.group_kernel_args_.size() * sizeof(GroupKernelArg),
+                                    hipMemcpyHostToDevice,
+                                    stream_config.stream_id_));
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel =
+                    kernel_grouped_gemm_softmax_gemm_xdl_cshuffle_v1<GridwiseGemm,
+                                                                     GroupKernelArg,
+                                                                     AElementwiseOperation,
+                                                                     BElementwiseOperation,
+                                                                     AccElementwiseOperation,
+                                                                     B1ElementwiseOperation,
+                                                                     CElementwiseOperation,
+                                                                     has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(arg.grid_size_),
+                    dim3(BlockSize),
+                    0,
+                    cast_pointer_to_constant_address_space(arg.p_workspace_),
+                    arg.group_count_,
+                    arg.a_element_op_,
+                    arg.b_element_op_,
+                    arg.acc_element_op_,
+                    arg.b1_element_op_,
+                    arg.c_element_op_);
+            };
+
+            // Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
+            // to concern Gemm0's loop
+            if(all_has_main_k_block_loop)
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else if(!some_has_main_k_block_loop)
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+            else
+            {
+                throw std::runtime_error("wrong! all gemm problems have to simultaneously meet "
+                                         "has_main_k_block_loop or no_main_k_block_loop");
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        // TODO ANT: Check if tensor specialization & strides mismatch
+
+        bool all_has_main_k_block_loop  = true;
+        bool some_has_main_k_block_loop = false;
+
+        for(std::size_t i = 0; i < arg.group_count_; i++)
+        {
+            const auto& kernel_arg = arg.group_kernel_args_[i];
+            const auto& device_arg = arg.group_device_args_[i];
+
+            // Check if C permute dimension matches GEMM + GEMM shape
+            const index_t c_m       = device_arg.c_grid_desc_m_n_.GetLength(I0);
+            const index_t c_gemm1n  = device_arg.c_grid_desc_m_n_.GetLength(I1);
+            const index_t a_m       = kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I1);
+            const index_t b1_gemm1n = kernel_arg.b1_grid_desc_bk0_n_bk1_.GetLength(I1);
+            if(!(c_m == a_m && c_gemm1n == b1_gemm1n))
+            {
+                return false;
+            }
+
+            // Check if having main loop
+            const auto K = kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) *
+                           kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+            const bool y = GridwiseGemm::CalculateHasMainKBlockLoop(K);
+            all_has_main_k_block_loop &= y;
+            some_has_main_k_block_loop |= y;
+
+            // Note: we need raw lengths since threadwise copy can not handle vector load when
+            // part of vector is out of bounds
+            const auto MzRaw      = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[0];
+            const auto NzRaw      = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[1];
+            const auto KzRaw      = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[2];
+            const auto Gemm1NzRaw = device_arg.raw_lengths_mz_nz_kz_gemm1nz_[3];
+
+            // Check scalar per vector requirement
+            const auto a_extent_lowest  = ABlockTransferSrcVectorDim == 2 ? KzRaw : MzRaw;
+            const auto b_extent_lowest  = BBlockTransferSrcVectorDim == 2 ? KzRaw : NzRaw;
+            const auto b1_extent_lowest = B1BlockTransferSrcVectorDim == 2 ? NzRaw : Gemm1NzRaw;
+            const auto c_extent_lowest  = Gemm1NzRaw;
+
+            if(!(a_extent_lowest % ABlockTransferSrcScalarPerVector == 0 &&
+                 b_extent_lowest % BBlockTransferSrcScalarPerVector == 0 &&
+                 b1_extent_lowest % B1BlockTransferSrcScalarPerVector == 0 &&
+                 c_extent_lowest % CShuffleBlockTransferScalarPerVector_NPerBlock == 0))
+            {
+                return false;
+            }
+
+            // Check vector load/store requirement
+            const auto a_stride_lowest  = ABlockTransferSrcVectorDim == 2
+                                              ? device_arg.a_mz_kz_strides_[1]
+                                              : device_arg.a_mz_kz_strides_[0];
+            const auto b_stride_lowest  = BBlockTransferSrcVectorDim == 2
+                                              ? device_arg.b_nz_kz_strides_[1]
+                                              : device_arg.b_nz_kz_strides_[0];
+            const auto b1_stride_lowest = B1BlockTransferSrcVectorDim == 2
+                                              ? device_arg.b1_nz_kz_strides_[1]
+                                              : device_arg.b1_nz_kz_strides_[0];
+            const auto c_stride_lowest =
+                device_arg.c_mz_gemm1nz_strides_[1]; // cshuffle assumes lowest dim in Gemm1Ns to be
+                                                     // contiguous
+
+            if(!(a_stride_lowest == 1 || b_stride_lowest == 1 || b1_stride_lowest == 1 ||
+                 c_stride_lowest == 1))
+            {
+                return false;
+            }
+
+            if(!GridwiseGemm::CheckValidity(kernel_arg.a_grid_desc_ak0_m_ak1_,
+                                            kernel_arg.b_grid_desc_bk0_n_bk1_,
+                                            kernel_arg.b1_grid_desc_bk0_n_bk1_,
+                                            device_arg.c_grid_desc_m_n_,
+                                            kernel_arg.block_2_ctile_map_))
+            {
+                return false;
+            }
+        }
+
+        // all gemm problems have to simultaneously meet has_main_k_block_loop or
+        // no_main_k_block_loop
+        if(!(all_has_main_k_block_loop || !some_has_main_k_block_loop))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::vector<const void*> p_a_vec,
+                             std::vector<const void*> p_b_vec,
+                             std::vector<const void*> p_b1_vec,
+                             std::vector<void*> p_c_vec,
+                             std::vector<std::vector<const void*>> p_acc0_biases_vec,
+                             std::vector<std::vector<const void*>> p_acc1_biases_vec,
+                             std::vector<ProblemDesc> problem_desc_vec,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             AccElementwiseOperation acc_element_op,
+                             B1ElementwiseOperation b1_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a_vec,
+                        p_b_vec,
+                        p_b1_vec,
+                        p_c_vec,
+                        p_acc0_biases_vec,
+                        p_acc1_biases_vec,
+                        problem_desc_vec,
+                        a_element_op,
+                        b_element_op,
+                        acc_element_op,
+                        b1_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*> p_a_vec,
+                        std::vector<const void*> p_b_vec,
+                        std::vector<const void*> p_b1_vec,
+                        std::vector<void*> p_c_vec,
+                        std::vector<std::vector<const void*>> p_acc0_biases_vec,
+                        std::vector<std::vector<const void*>> p_acc1_biases_vec,
+                        std::vector<ProblemDesc> problem_desc_vec,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        AccElementwiseOperation acc_element_op,
+                        B1ElementwiseOperation b1_element_op,
+                        CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(p_a_vec,
+                                          p_b_vec,
+                                          p_b1_vec,
+                                          p_c_vec,
+                                          p_acc0_biases_vec,
+                                          p_acc1_biases_vec,
+                                          problem_desc_vec,
+                                          a_element_op,
+                                          b_element_op,
+                                          acc_element_op,
+                                          b1_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerBlock << ", "
+            << Gemm1NPerBlock << ", "
+            << Gemm1KPerBlock << ", "
+            << B1K1 << ", "
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << "ASpec" << getTensorSpecializationString(ASpec) << ", "
+            << "B0Spec" << getTensorSpecializationString(BSpec) << ", "
+            << "B1Spec" << getTensorSpecializationString(B1Spec) << ", "
+            << "CSpec" << getTensorSpecializationString(CSpec) << ", "
+            << getMaskingSpecializationString(MaskingSpec) << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
+    {
+        return dynamic_cast<const Argument*>(p_arg)->group_count_ * sizeof(GroupKernelArg);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp
new file mode 100644
index 0000000000..9290a31554
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp
@@ -0,0 +1,725 @@
+#pragma once
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename GemmDesc,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_gemm_xdl(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
+                                const index_t group_count,
+                                const AElementwiseOperation a_element_op,
+                                const BElementwiseOperation b_element_op,
+                                const CDEElementwiseOperation c_element_op)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t block_id = get_block_1d_id();
+
+    const auto gemm_desc_ptr =
+        reinterpret_cast<const GemmDesc*>(cast_pointer_to_generic_address_space(gemm_descs_const));
+
+    index_t left     = 0;
+    index_t right    = group_count;
+    index_t group_id = index_t((left + right) / 2);
+    while((!(block_id >= gemm_desc_ptr[group_id].BlockStart_ &&
+             block_id < gemm_desc_ptr[group_id].BlockEnd_)) &&
+          left <= right)
+    {
+        if(block_id < gemm_desc_ptr[group_id].BlockStart_)
+        {
+            right = group_id;
+        }
+        else
+        {
+            left = group_id;
+        }
+        group_id = index_t((left + right) / 2);
+    }
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        gemm_desc_ptr[group_id].a_ptr_,
+        gemm_desc_ptr[group_id].b_ptr_,
+        gemm_desc_ptr[group_id].ds_ptr_,
+        gemm_desc_ptr[group_id].e_ptr_,
+        p_shared,
+        a_element_op,
+        b_element_op,
+        c_element_op,
+        gemm_desc_ptr[group_id].a_grid_desc_ak0_m_ak1_,
+        gemm_desc_ptr[group_id].b_grid_desc_bk0_n_bk1_,
+        gemm_desc_ptr[group_id].ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+        gemm_desc_ptr[group_id].e_grid_desc_mblock_mperblock_nblock_nperblock_,
+        gemm_desc_ptr[group_id].block_2_etile_map_);
+#else
+    ignore = gemm_descs_const;
+    ignore = group_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+#endif
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t NumPrefetch,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t KPerBlock,
+          ck::index_t AK1,
+          ck::index_t BK1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout,
+                                                        BLayout,
+                                                        DsLayout,
+                                                        ELayout,
+                                                        ADataType,
+                                                        BDataType,
+                                                        DsDataType,
+                                                        EDataType,
+                                                        AElementwiseOperation,
+                                                        BElementwiseOperation,
+                                                        CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedGemm_Xdl;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    template <typename ELay>
+    static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                                         const std::array<index_t, NumDTensor>& NRaws,
+                                         const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
+
+    using ComputeDataType = ADataType;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeDataType,
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        NumPrefetch, // NumGemmKPrefetchStage
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    using AGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
+            AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
+            BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
+        decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+
+    struct GroupedGemmBlock2ETileMap
+    {
+        using Block2ETileMap =
+            remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+        GroupedGemmBlock2ETileMap()
+        {
+            block_2_etile_map_ = GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{});
+            BlockStart_        = -1;
+        }
+
+        GroupedGemmBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n, ck::index_t BlockStart)
+        {
+            block_2_etile_map_ = GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n);
+            BlockStart_        = BlockStart;
+        }
+
+        template <typename TopIdx>
+        __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+        {
+            return block_2_etile_map_.CalculateBottomIndex(
+                make_multi_index(idx_top[I0] - BlockStart_));
+        }
+
+        // it's actually E-Tile
+        template <typename CTileIdx, typename CTileDim>
+        __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                                 const CTileDim& c_tile_dim) const
+        {
+            return block_2_etile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
+        }
+
+        __host__ bool CheckValidity(const EGridDesc_M_N& e_grid_desc_m_n) const
+        {
+            return block_2_etile_map_.CheckValidity(e_grid_desc_m_n);
+        }
+
+        Block2ETileMap block_2_etile_map_;
+        ck::index_t BlockStart_;
+    };
+
+    struct GemmBiasTransKernelArg
+    {
+        // pointers
+        const ADataType* a_ptr_;
+        const BDataType* b_ptr_;
+        typename GridwiseGemm::DsGridPointer ds_ptr_;
+        EDataType* e_ptr_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        GroupedGemmBlock2ETileMap block_2_etile_map_;
+        ck::index_t BlockStart_, BlockEnd_;
+    };
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(std::vector<const void*>& p_As,
+                 std::vector<const void*>& p_Bs,
+                 std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                 std::vector<void*>& p_Es,
+                 std::vector<GemmDesc>& gemm_descs,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation c_element_op)
+            : a_element_op_{a_element_op}, b_element_op_{b_element_op}, c_element_op_{c_element_op}
+        {
+            grid_size_ = 0;
+
+            group_count_ = ck::type_convert<ck::index_t>(gemm_descs.size());
+
+            if(!(group_count_ == ck::type_convert<ck::index_t>(p_As.size()) &&
+                 group_count_ == ck::type_convert<ck::index_t>(p_Bs.size()) &&
+                 group_count_ == ck::type_convert<ck::index_t>(p_Es.size())))
+            {
+                throw std::runtime_error("wrong! group_count_ != p_As/b/c.size");
+            }
+
+            gemm_desc_kernel_arg_.reserve(group_count_);
+
+            skipped_group_count_ = 0;
+
+            for(std::size_t i = 0; i < gemm_descs.size(); i++)
+            {
+                const index_t M = gemm_descs[i].M_;
+                const index_t N = gemm_descs[i].N_;
+                const index_t K = gemm_descs[i].K_;
+
+                a_mtx_mraw_kraw_.emplace_back(M, K);
+                b_mtx_nraw_kraw_.emplace_back(N, K);
+
+                if(M == 0)
+                {
+                    skipped_group_count_++;
+                    continue;
+                }
+
+                const index_t StrideA = gemm_descs[i].stride_A_;
+                const index_t StrideB = gemm_descs[i].stride_B_;
+                const index_t StrideC = gemm_descs[i].stride_C_;
+
+                // pointer
+                typename GridwiseGemm::DsGridPointer p_ds_grid{};
+
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    using DDataType = remove_cvref_t<tuple_element_t<j.value, DsDataType>>;
+
+                    p_ds_grid(j) = static_cast<const DDataType*>(p_Ds[i][j]);
+                });
+
+                // tensor descriptors for problem definiton
+                const auto a_grid_desc_m_k = DeviceOp::MakeAGridDescriptor_M_K(M, K, StrideA);
+                const auto b_grid_desc_n_k = DeviceOp::MakeBGridDescriptor_N_K(K, N, StrideB);
+
+                DsGridDesc_M_N ds_grid_desc_m_n;
+
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+                    ds_grid_desc_m_n(j) = DeviceOp::MakeEGridDescriptor_M_N<DLayout>(
+                        M, N, gemm_descs[i].stride_Ds_[j]);
+                });
+
+                const auto e_grid_desc_m_n =
+                    DeviceOp::MakeEGridDescriptor_M_N<ELayout>(M, N, StrideC);
+
+                // tensor descriptors for block/thread-wise copy
+                const auto a_grid_desc_ak0_m_ak1 =
+                    GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k);
+
+                const auto b_grid_desc_bk0_n_bk1 =
+                    GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k);
+
+                const index_t grid_size_grp =
+                    GroupedGemmBlock2ETileMap(e_grid_desc_m_n, 0)
+                        .block_2_etile_map_.CalculateGridSize(e_grid_desc_m_n);
+
+                const index_t BlockStart = grid_size_;
+                const index_t BlockEnd   = grid_size_ + grid_size_grp;
+
+                grid_size_ += grid_size_grp;
+
+                // block-to-e-tile map
+                const auto block_2_etile_map =
+                    GroupedGemmBlock2ETileMap(e_grid_desc_m_n, BlockStart);
+
+                if(GridwiseGemm::CheckValidity(a_grid_desc_m_k,
+                                               b_grid_desc_n_k,
+                                               ds_grid_desc_m_n,
+                                               e_grid_desc_m_n,
+                                               block_2_etile_map))
+                {
+                    // tensor descriptors for block/thread-wise copy
+                    DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock;
+
+                    static_for<0, NumDTensor, 1>{}([&](auto j) {
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock(j) =
+                            GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                                ds_grid_desc_m_n[j]);
+                    });
+
+                    const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+                        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                            e_grid_desc_m_n);
+
+                    gemm_desc_kernel_arg_.push_back(
+                        GemmBiasTransKernelArg{static_cast<const ADataType*>(p_As[i]),
+                                               static_cast<const BDataType*>(p_Bs[i]),
+                                               p_ds_grid,
+                                               static_cast<EDataType*>(p_Es[i]),
+                                               a_grid_desc_m_k,
+                                               b_grid_desc_n_k,
+                                               ds_grid_desc_m_n,
+                                               e_grid_desc_m_n,
+                                               a_grid_desc_ak0_m_ak1,
+                                               b_grid_desc_bk0_n_bk1,
+                                               ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                               e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                               block_2_etile_map,
+                                               BlockStart,
+                                               BlockEnd});
+                }
+            }
+        }
+
+        //  private:
+        index_t group_count_;
+        index_t skipped_group_count_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation c_element_op_;
+
+        std::vector<GemmBiasTransKernelArg> gemm_desc_kernel_arg_;
+        std::vector<Tuple<index_t, index_t>> a_mtx_mraw_kraw_;
+        std::vector<Tuple<index_t, index_t>> b_mtx_nraw_kraw_;
+
+        index_t grid_size_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            bool has_main_k_block_loop = true;
+
+            for(std::size_t i = 0; i < arg.gemm_desc_kernel_arg_.size(); i++)
+            {
+#if DEBUG_LOG
+                std::cout << "group: " << i << " arg.a_grid_desc_ak0_m_ak1_{"
+                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I1)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2)
+                          << "}";
+
+                std::cout << ", arg.b_grid_desc_bk0_n_bk1_{"
+                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_bk0_n_bk1_.GetLength(I0)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_bk0_n_bk1_.GetLength(I1)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_bk0_n_bk1_.GetLength(I2)
+                          << "}";
+
+                std::cout << ", arg.e_grid_desc_m_n_{ "
+                          << arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_.GetLength(I0) << ", "
+                          << arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_.GetLength(I1) << "}"
+                          << std::endl;
+#endif
+
+                if(!GridwiseGemm::CheckValidity(arg.gemm_desc_kernel_arg_[i].a_grid_desc_m_k_,
+                                                arg.gemm_desc_kernel_arg_[i].b_grid_desc_n_k_,
+                                                arg.gemm_desc_kernel_arg_[i].ds_grid_desc_m_n_,
+                                                arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_,
+                                                arg.gemm_desc_kernel_arg_[i].block_2_etile_map_))
+                {
+                    throw std::runtime_error(
+                        "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
+                }
+
+                const auto K = arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0) *
+                               arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+                if(GridwiseGemm::CalculateHasMainKBlockLoop(K) != has_main_k_block_loop)
+                {
+                    throw std::runtime_error("wrong! not all gemm has_main_k_block_loop");
+                }
+            }
+
+            hipGetErrorString(hipMemcpyWithStream(arg.p_workspace_,
+                                                  arg.gemm_desc_kernel_arg_.data(),
+                                                  arg.gemm_desc_kernel_arg_.size() *
+                                                      sizeof(GemmBiasTransKernelArg),
+                                                  hipMemcpyHostToDevice,
+                                                  stream_config.stream_id_));
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_) {
+                const auto kernel = kernel_grouped_gemm_xdl<GridwiseGemm,
+                                                            GemmBiasTransKernelArg,
+                                                            AElementwiseOperation,
+                                                            BElementwiseOperation,
+                                                            CDEElementwiseOperation,
+                                                            has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(arg.grid_size_),
+                    dim3(BlockSize),
+                    0,
+                    cast_pointer_to_constant_address_space(arg.p_workspace_),
+                    arg.gemm_desc_kernel_arg_.size(),
+                    arg.a_element_op_,
+                    arg.b_element_op_,
+                    arg.c_element_op_);
+            };
+
+            if(has_main_k_block_loop)
+            {
+                ave_time = launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                ave_time = launch_kernel(integral_constant<bool, false>{});
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if((ck::type_convert<ck::index_t>(arg.gemm_desc_kernel_arg_.size()) +
+            arg.skipped_group_count_) != arg.group_count_)
+        {
+            return false;
+        }
+
+        bool supported = true;
+
+        // If we use padding we do not support vector loads for dimensions not divisible by vector
+        // load size.
+        if constexpr(GemmSpec != GemmSpecialization::Default)
+        {
+            // [A|B]BlockTransferSrcVectorDim value define dimension in the block {K0,M,K1} layout,
+            // thus we have to adapt it to the {M,K} or {N,K} layout.
+            const auto a_raw_vector_dim = ABlockTransferSrcVectorDim != 1 ? 1 : 0;
+            const auto b_raw_vector_dim = BBlockTransferSrcVectorDim != 1 ? 1 : 0;
+
+            for(index_t i = 0; i < arg.group_count_; ++i)
+            {
+                const auto a_vector_dim = arg.a_mtx_mraw_kraw_[i].At(Number<a_raw_vector_dim>{});
+                const auto b_vector_dim = arg.b_mtx_nraw_kraw_[i].At(Number<b_raw_vector_dim>{});
+
+                supported = supported & (a_vector_dim % ABlockTransferSrcScalarPerVector == 0);
+                supported = supported & (b_vector_dim % BBlockTransferSrcScalarPerVector == 0);
+            }
+        }
+
+        return supported;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::vector<const void*>& p_As,
+                             std::vector<const void*>& p_Bs,
+                             std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                             std::vector<void*>& p_Es,
+                             std::vector<GemmDesc> gemm_descs,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation c_element_op)
+    {
+        return Argument{
+            p_As, p_Bs, p_Ds, p_Es, gemm_descs, a_element_op, b_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*>& p_As,
+                        std::vector<const void*>& p_Bs,
+                        std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                        std::vector<void*>& p_Es,
+                        std::vector<GemmDesc>& gemm_descs,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(
+            p_As, p_Bs, p_Ds, p_Es, gemm_descs, a_element_op, b_element_op, c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedGemm_Xdl"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle << ", "
+            << getGemmSpecializationString(GemmSpec)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
+    {
+        return dynamic_cast<const Argument*>(p_arg)->group_count_ * sizeof(GemmBiasTransKernelArg);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp
new file mode 100644
index 0000000000..8cea09ee54
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp
@@ -0,0 +1,836 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename GemmDesc,
+          GemmSpecialization GemmSpec,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename DsDataType,
+          typename Block2ETileMap,
+          typename GroupedGemmBlock2ETileMap,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_gemm_xdl_fixed_nk(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
+                                         uint32_t* barrier_count,
+                                         const index_t barrier_size_grp,
+                                         const index_t group_count,
+                                         const index_t grid_size_grp,
+                                         const index_t KBatch,
+                                         const AElementwiseOperation a_element_op,
+                                         const BElementwiseOperation b_element_op,
+                                         const CDEElementwiseOperation c_element_op)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t block_id = get_block_1d_id();
+
+    const auto gemm_desc_ptr =
+        reinterpret_cast<const GemmDesc*>(cast_pointer_to_generic_address_space(gemm_descs_const));
+
+    const index_t group_id = block_id / grid_size_grp;
+
+    if(group_id >= group_count)
+        return;
+
+    const index_t M = gemm_desc_ptr[group_id].M;
+    const index_t N = gemm_desc_ptr[group_id].N;
+    const index_t K = gemm_desc_ptr[group_id].K;
+
+    if(M * N * K == 0)
+        return;
+
+    const auto StrideA  = gemm_desc_ptr[group_id].StrideA;
+    const auto StrideB  = gemm_desc_ptr[group_id].StrideB;
+    const auto StrideDs = gemm_desc_ptr[group_id].StrideDs;
+    const auto StrideE  = gemm_desc_ptr[group_id].StrideE;
+
+    const auto e_grid_desc_m_n =
+        GridwiseGemm::template MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+    const index_t BlockStart = group_id * grid_size_grp;
+
+    const auto local_b2e_tile_map = Block2ETileMap{e_grid_desc_m_n, KBatch};
+
+    const auto local_grid_size = local_b2e_tile_map.CalculateGridSize(e_grid_desc_m_n);
+
+    constexpr auto NumDTensor = DsDataType::Size();
+
+    using DsGridPointer = decltype(GridwiseGemm::MakeDsGridPointer());
+
+    DsGridPointer p_ds_grid_;
+
+    static_for<0, NumDTensor, 1>{}([&](auto i) {
+        using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+        // D pointer
+        p_ds_grid_(i) = static_cast<const DDataType*>(gemm_desc_ptr[group_id].p_ds_grid[i]);
+    });
+
+    index_t id_off   = 0;
+    index_t id_local = get_block_1d_id() - BlockStart;
+
+    const index_t mn_blocks = local_grid_size / KBatch;
+
+    while(id_local < local_grid_size)
+    {
+        const auto block_2_etile_map =
+            GroupedGemmBlock2ETileMap(local_b2e_tile_map, BlockStart, id_off);
+
+        auto barrier_count_finished =
+            barrier_count + group_id * barrier_size_grp + id_local % mn_blocks;
+
+        GridwiseGemm::template Run<HasMainKBlockLoop,
+                                   EGlobalMemoryDataOperation,
+                                   GemmSpec,
+                                   ALayout,
+                                   BLayout,
+                                   DsLayout,
+                                   ELayout>(gemm_desc_ptr[group_id].p_a_grid,
+                                            gemm_desc_ptr[group_id].p_b_grid,
+                                            p_ds_grid_,
+                                            gemm_desc_ptr[group_id].p_e_grid,
+                                            p_shared,
+                                            barrier_count_finished,
+                                            a_element_op,
+                                            b_element_op,
+                                            c_element_op,
+                                            M,
+                                            N,
+                                            K,
+                                            StrideA,
+                                            StrideB,
+                                            StrideDs,
+                                            StrideE,
+                                            KBatch,
+                                            block_2_etile_map);
+
+        id_off += grid_size_grp;
+        id_local += grid_size_grp;
+    }
+#else
+    ignore = gemm_descs_const;
+    ignore = barrier_count;
+    ignore = barrier_size_grp;
+    ignore = group_count;
+    ignore = grid_size_grp;
+    ignore = KBatch;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+#endif
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t NumPrefetch,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t KPerBlock,
+          ck::index_t AK1,
+          ck::index_t BK1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedGemm_Xdl_Fixed_NK : public DeviceGroupedGemmFixedNK<ALayout,
+                                                                        BLayout,
+                                                                        DsLayout,
+                                                                        ELayout,
+                                                                        ADataType,
+                                                                        BDataType,
+                                                                        DsDataType,
+                                                                        EDataType,
+                                                                        AElementwiseOperation,
+                                                                        BElementwiseOperation,
+                                                                        CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGroupedGemm_Xdl_Fixed_NK;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_splitk_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        NumPrefetch, // NumGemmKPrefetchStage
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    template <typename UnderlyingBlockToCTileMap>
+    struct OffsettedBlockToCTileMapMLoops
+    {
+        using underlying_type = UnderlyingBlockToCTileMap;
+
+        __host__ __device__ OffsettedBlockToCTileMapMLoops(
+            UnderlyingBlockToCTileMap block_to_ctile_map, index_t block_start, index_t id_off = 0)
+        {
+            block_to_ctile_map_ = block_to_ctile_map;
+            block_start_        = block_start;
+            id_off_             = id_off;
+        }
+
+        template <typename TopIdx>
+        __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+        {
+            auto idx_bot = block_to_ctile_map_.CalculateBottomIndex(
+                make_multi_index(idx_top[Number<0>{}] - block_start_ + id_off_));
+
+            return make_tuple(idx_bot[Number<0>{}], idx_bot[Number<1>{}], idx_bot[Number<2>{}]);
+        }
+
+        template <typename CTileIdx, typename CTileDim>
+        __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                                 const CTileDim& c_tile_dim) const
+        {
+            return block_to_ctile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
+        }
+
+        template <typename CGridDesc_M_N>
+        __host__ bool CheckValidity(const CGridDesc_M_N& c_grid_desc_m_n) const
+        {
+            return block_to_ctile_map_.CheckValidity(c_grid_desc_m_n);
+        }
+
+        template <typename CGridDesc_M_N>
+        __host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+        {
+            return block_to_ctile_map_.CalculateGridSize(c_grid_desc_m_n);
+        }
+
+        UnderlyingBlockToCTileMap block_to_ctile_map_;
+        index_t block_start_;
+        index_t id_off_;
+    };
+
+    template <index_t MPerBlock_, index_t NPerBlock_>
+    struct BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops
+    {
+        static constexpr auto I0 = Number<0>{};
+        static constexpr auto I1 = Number<1>{};
+
+        __host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops() = default;
+
+        __host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops(
+            const BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&) = default;
+        __host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops(
+            BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&&) = default;
+        __host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&
+        operator=(const BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&) = default;
+        __host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&
+        operator=(BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops&&) = default;
+
+        __host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops(index_t M,
+                                                                          index_t N,
+                                                                          index_t KBatch,
+                                                                          index_t M01 = 8)
+            : M_(M), N_(N), KBatch_(KBatch), M01_(M01)
+        {
+        }
+
+        template <typename CGridDesc_M_N>
+        __host__ __device__ BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops(
+            const CGridDesc_M_N& c_grid_desc_m_n, index_t KBatch, index_t M01 = 8)
+            : BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops(
+                  c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1), KBatch, M01)
+        {
+        }
+
+        __host__ __device__ constexpr index_t CalculateGridSize(index_t M, index_t N) const
+        {
+            const auto M0 = math::integer_divide_ceil(M, MPerBlock);
+            const auto N0 = math::integer_divide_ceil(N, NPerBlock);
+
+            return M0 * N0 * KBatch_;
+        }
+
+        template <typename CGridDesc_M_N>
+        __host__ __device__ constexpr index_t
+        CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+        {
+            return CalculateGridSize(c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1));
+        }
+
+        template <typename CGridDesc_M_N>
+        __host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const
+        {
+            return true;
+        }
+
+        template <typename TopIdx>
+        __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+        {
+            auto block_1d_id = idx_top[I0];
+
+            const auto M0 = math::integer_divide_ceil(M_, MPerBlock_);
+            const auto N0 = math::integer_divide_ceil(N_, NPerBlock_);
+
+            block_1d_id = block_1d_id % (M0 * N0 * KBatch_); // hide groups
+
+            const index_t idx_ksplit = block_1d_id / (M0 * N0);
+            block_1d_id              = block_1d_id % (M0 * N0);
+
+            index_t idx_N0 = block_1d_id % N0;
+            index_t idx_M0 = block_1d_id / N0;
+
+            const auto M01_adapt = (idx_M0 < M0 - M0 % M01_) ? M01_ : M0 % M01_;
+
+            index_t idx_M00          = idx_M0 / M01_;
+            index_t idx_M01          = idx_M0 % M01_;
+            index_t idx_N0_M01_local = idx_N0 + idx_M01 * N0;
+
+            return make_tuple(idx_ksplit,
+                              idx_N0_M01_local % M01_adapt + idx_M00 * M01_,
+                              idx_N0_M01_local / M01_adapt);
+        }
+
+        template <typename CTileIdx, typename CTileDim>
+        __host__ __device__ bool ValidCTileIndex(const CTileIdx& /* c_tile_idx */,
+                                                 const CTileDim& /* c_tile_dim */) const
+        {
+            return true; // always valid provided that user gets grid size from CalculateGridSize()
+        }
+
+        private:
+        index_t M_;
+        index_t N_;
+        index_t KBatch_;
+        index_t M01_;
+    };
+
+    using Block2ETileMap = BlockToCTileMap_KBatch_M00_N0_M01Adapt_MLoops<MPerBlock, NPerBlock>;
+    using GroupedGemmBlock2ETileMap = OffsettedBlockToCTileMapMLoops<Block2ETileMap>;
+
+    struct GemmBiasTransKernelArg
+    {
+        // pointers
+        const void* a_ptr_;
+        const void* b_ptr_;
+        std::array<const void*, NumDTensor> ds_ptr_;
+        void* e_ptr_;
+
+        index_t M_, N_, K_;
+        index_t StrideA_, StrideB_;
+        std::array<index_t, NumDTensor> StrideDs_;
+        index_t StrideE_;
+    };
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+
+        void UpdateKBatch(index_t k_batch)
+        {
+            k_batch_ = k_batch;
+
+            if(k_batch_ < 1)
+            {
+
+                throw std::runtime_error("wrong! k_batch must be > 0");
+            }
+
+            const index_t AverM = math::integer_divide_ceil(sum_of_m, group_count_);
+
+            const index_t StrideE = gemm_desc_kernel_arg_[0].StrideE_;
+            const index_t N       = gemm_desc_kernel_arg_[0].N_;
+
+            const auto e_grid_desc_m_n =
+                GridwiseGemm::template MakeEGridDescriptor_M_N<ELayout, GemmSpec>(
+                    AverM, N, StrideE);
+
+            const auto local_b2c_tile_map = Block2ETileMap{e_grid_desc_m_n, k_batch_};
+
+            grid_size_grp_ = local_b2c_tile_map.CalculateGridSize(e_grid_desc_m_n);
+
+            grid_size_ = grid_size_grp_ * group_count_;
+        }
+
+        Argument(std::vector<const void*>&,
+                 std::vector<const void*>&,
+                 std::vector<std::array<const void*, NumDTensor>>&,
+                 std::vector<void*>&,
+                 std::vector<GemmDesc>& gemm_descs,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation c_element_op)
+            : a_element_op_{a_element_op}, b_element_op_{b_element_op}, c_element_op_{c_element_op}
+        {
+            grid_size_ = 0;
+
+            k_batch_ = 1;
+
+            grouped_gemm_kernel_args_dev = nullptr;
+
+            group_count_ = ck::type_convert<ck::index_t>(gemm_descs.size());
+
+            gemm_desc_kernel_arg_.reserve(group_count_);
+
+            index_t group_id = 0;
+
+            sum_of_m            = gemm_descs[0].M_;
+            const index_t AverM = math::integer_divide_ceil(sum_of_m, group_count_);
+            const index_t N     = gemm_descs[0].N_;
+            const index_t K     = gemm_descs[0].K_;
+
+            for(std::size_t i = 0; i < gemm_descs.size(); i++)
+            {
+                if(sum_of_m != gemm_descs[i].M_ || N != gemm_descs[i].N_ || K != gemm_descs[i].K_)
+                {
+                    throw std::runtime_error("wrong! M/N/K is not identical");
+                }
+
+                a_mtx_mraw_kraw_.emplace_back(sum_of_m, K);
+                b_mtx_nraw_kraw_.emplace_back(N, K);
+
+                const index_t StrideA = gemm_descs[i].stride_A_;
+                const index_t StrideB = gemm_descs[i].stride_B_;
+                const index_t StrideE = gemm_descs[i].stride_C_;
+
+                // pointer
+                std::array<const void*, NumDTensor> p_ds_grid;
+
+                static_for<0, NumDTensor, 1>{}([&](auto j) { p_ds_grid[j] = nullptr; });
+
+                std::array<index_t, NumDTensor> StrideDs;
+
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    // using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+                    if(gemm_descs[i].stride_Ds_.size() != NumDTensor)
+                    {
+                        throw std::runtime_error(
+                            "wrong! gemm_descs[i].stride_Ds_.size() does not match NumDTensor");
+                    }
+
+                    StrideDs[j] = gemm_descs[i].stride_Ds_[j];
+                });
+
+                const auto e_grid_desc_m_n =
+                    GridwiseGemm::template MakeEGridDescriptor_M_N<ELayout, GemmSpec>(
+                        AverM, N, StrideE);
+
+                // block-to-e-tile map
+                const auto local_b2c_tile_map = Block2ETileMap{e_grid_desc_m_n, k_batch_};
+
+                grid_size_grp_ = local_b2c_tile_map.CalculateGridSize(e_grid_desc_m_n);
+
+                if(group_id * grid_size_grp_ != grid_size_)
+                {
+                    throw std::runtime_error("wrong! grid_size_grp_ is not identical!");
+                }
+
+                grid_size_ += grid_size_grp_;
+
+                // check block-to-E-tile
+                if(!local_b2c_tile_map.CheckValidity(e_grid_desc_m_n))
+                {
+                    throw std::runtime_error("wrong! block_2_etile_map validation failed");
+                }
+
+                if(!GridwiseGemm::
+                       template CheckValidity<ALayout, BLayout, DsLayout, ELayout, GemmSpec>(
+                           AverM, N, K, StrideA, StrideB, StrideDs, StrideE, 1))
+                {
+                    throw std::runtime_error(
+                        "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
+                }
+
+                gemm_desc_kernel_arg_.push_back(GemmBiasTransKernelArg{
+                    nullptr,
+                    nullptr,
+                    p_ds_grid,
+                    nullptr,
+                    AverM,
+                    N,
+                    K,
+                    StrideA,
+                    StrideB,
+                    StrideDs,
+                    StrideE,
+                });
+
+                group_id++;
+            }
+
+            const auto e_grid_desc_sum_m_n =
+                GridwiseGemm::template MakeEGridDescriptor_M_N<ELayout, GemmSpec>(
+                    sum_of_m, gemm_desc_kernel_arg_[0].N_, gemm_desc_kernel_arg_[0].StrideE_);
+
+            const auto local_b2c_tile_map = Block2ETileMap{e_grid_desc_sum_m_n, 1};
+
+            barrier_size_grp_ = local_b2c_tile_map.CalculateGridSize(e_grid_desc_sum_m_n);
+        }
+
+        //  private:
+        index_t group_count_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation c_element_op_;
+
+        std::vector<GemmBiasTransKernelArg> gemm_desc_kernel_arg_;
+        std::vector<Tuple<index_t, index_t>> a_mtx_mraw_kraw_;
+        std::vector<Tuple<index_t, index_t>> b_mtx_nraw_kraw_;
+
+        const void* grouped_gemm_kernel_args_dev;
+
+        index_t grid_size_;
+        index_t grid_size_grp_;
+        index_t barrier_size_grp_;
+        index_t sum_of_m;
+
+        index_t k_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            bool has_main_k_block_loop = true;
+
+            for(std::size_t i = 0; i < arg.gemm_desc_kernel_arg_.size(); i++)
+            {
+                const auto KPad =
+                    GridwiseGemm::CalculateKPadded(arg.gemm_desc_kernel_arg_[i].K_, arg.k_batch_);
+
+                if(GridwiseGemm::CalculateHasMainKBlockLoop(KPad) != has_main_k_block_loop)
+                {
+                    throw std::runtime_error("wrong! not all gemm has_main_k_block_loop");
+                }
+            }
+
+            if(arg.grouped_gemm_kernel_args_dev == nullptr)
+            {
+                throw std::runtime_error("wrong! grouped_gemm_kernel_args_dev is nullpr");
+            }
+
+            float ave_time = 0;
+
+            auto launch_kernel = [&](auto has_main_k_block_loop_, auto e_global_memory_operation_) {
+                const auto kernel =
+                    kernel_grouped_gemm_xdl_fixed_nk<GridwiseGemm,
+                                                     GroupedGemmKernelArgument<NumDTensor>,
+                                                     GemmSpec,
+                                                     ALayout,
+                                                     BLayout,
+                                                     DsLayout,
+                                                     ELayout,
+                                                     DsDataType,
+                                                     Block2ETileMap,
+                                                     GroupedGemmBlock2ETileMap,
+                                                     AElementwiseOperation,
+                                                     BElementwiseOperation,
+                                                     CDEElementwiseOperation,
+                                                     e_global_memory_operation_,
+                                                     has_main_k_block_loop_>;
+
+                return launch_and_time_kernel(
+                    stream_config,
+                    kernel,
+                    dim3(arg.grid_size_),
+                    dim3(BlockSize),
+                    0,
+                    cast_pointer_to_constant_address_space(arg.grouped_gemm_kernel_args_dev),
+                    reinterpret_cast<uint32_t*>(arg.p_workspace_),
+                    arg.barrier_size_grp_,
+                    arg.gemm_desc_kernel_arg_.size(),
+                    arg.grid_size_grp_,
+                    arg.k_batch_,
+                    arg.a_element_op_,
+                    arg.b_element_op_,
+                    arg.c_element_op_);
+            };
+
+            constexpr auto AtomicAdd = InMemoryDataOperationEnum::AtomicAdd;
+            constexpr auto Set       = InMemoryDataOperationEnum::Set;
+
+            if(arg.k_batch_ > 1)
+            {
+                if(has_main_k_block_loop)
+                {
+                    ave_time =
+                        launch_kernel(integral_constant<bool, true>{},
+                                      integral_constant<InMemoryDataOperationEnum, AtomicAdd>{});
+                }
+                else
+                {
+                    ave_time =
+                        launch_kernel(integral_constant<bool, false>{},
+                                      integral_constant<InMemoryDataOperationEnum, AtomicAdd>{});
+                }
+            }
+            else
+            {
+                if(has_main_k_block_loop)
+                {
+                    ave_time = launch_kernel(integral_constant<bool, true>{},
+                                             integral_constant<InMemoryDataOperationEnum, Set>{});
+                }
+                else
+                {
+                    ave_time = launch_kernel(integral_constant<bool, false>{},
+                                             integral_constant<InMemoryDataOperationEnum, Set>{});
+                }
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(ck::type_convert<ck::index_t>(arg.gemm_desc_kernel_arg_.size()) != arg.group_count_)
+        {
+            return false;
+        }
+
+        bool supported = true;
+
+        // If we use padding we do not support vector loads for dimensions not divisible by vector
+        // load size.
+        if constexpr(GemmSpec != GemmSpecialization::Default)
+        {
+            // [A|B]BlockTransferSrcVectorDim value define dimension in the block {K0,M,K1} layout,
+            // thus we have to adapt it to the {M,K} or {N,K} layout.
+            const auto a_raw_vector_dim = ABlockTransferSrcVectorDim != 1 ? 1 : 0;
+            const auto b_raw_vector_dim = BBlockTransferSrcVectorDim != 1 ? 1 : 0;
+
+            for(index_t i = 0; i < arg.group_count_; ++i)
+            {
+                const auto a_vector_dim = arg.a_mtx_mraw_kraw_[i].At(Number<a_raw_vector_dim>{});
+                const auto b_vector_dim = arg.b_mtx_nraw_kraw_[i].At(Number<b_raw_vector_dim>{});
+
+                supported = supported & (a_vector_dim % ABlockTransferSrcScalarPerVector == 0);
+                supported = supported & (b_vector_dim % BBlockTransferSrcScalarPerVector == 0);
+            }
+        }
+
+        return supported;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::vector<const void*>& p_As,
+                             std::vector<const void*>& p_Bs,
+                             std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                             std::vector<void*>& p_Es,
+                             std::vector<GemmDesc> gemm_descs,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CDEElementwiseOperation c_element_op)
+    {
+        return Argument{
+            p_As, p_Bs, p_Ds, p_Es, gemm_descs, a_element_op, b_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*>& p_As,
+                        std::vector<const void*>& p_Bs,
+                        std::vector<std::array<const void*, NumDTensor>>& p_Ds,
+                        std::vector<void*>& p_Es,
+                        std::vector<GemmDesc>& gemm_descs,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(
+            p_As, p_Bs, p_Ds, p_Es, gemm_descs, a_element_op, b_element_op, c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedGemm_Xdl_Fixed_NK"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle << ", "
+            << getGemmSpecializationString(GemmSpec)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    static void SetDeviceKernelArgs(Argument& arg, const void* kernel_args)
+    {
+        arg.grouped_gemm_kernel_args_dev = kernel_args;
+    }
+
+    // polymorphic
+    void SetDeviceKernelArgs(BaseArgument* p_arg, const void* kernel_args) const override
+    {
+        return SetDeviceKernelArgs(*dynamic_cast<Argument*>(p_arg), kernel_args);
+    }
+
+    size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
+    {
+        auto arg = *dynamic_cast<const Argument*>(p_arg);
+
+        return arg.group_count_ * arg.barrier_size_grp_ * sizeof(uint32_t);
+    }
+
+    size_t GetDeviceKernelArgSize(const BaseArgument* p_arg) const override
+    {
+        auto arg = *dynamic_cast<const Argument*>(p_arg);
+
+        return arg.group_count_ * sizeof(GroupedGemmKernelArgument<NumDTensor>);
+    }
+
+    void SetWorkSpacePointer(BaseArgument* p_arg, void* p_workspace) const override
+    {
+        auto p_arg_          = dynamic_cast<Argument*>(p_arg);
+        p_arg_->p_workspace_ = p_workspace;
+
+        hip_check_error(hipMemset(p_workspace, 0, GetWorkSpaceSize(p_arg)));
+    }
+
+    static void SetKBatch(Argument& arg, index_t k_batch) { arg.UpdateKBatch(k_batch); }
+
+    // polymorphic
+    void SetKBatch(BaseArgument* p_arg, index_t k_batch) const override
+    {
+        return SetKBatch(*dynamic_cast<Argument*>(p_arg), k_batch);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp
new file mode 100644
index 0000000000..41445aeaf0
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp
@@ -0,0 +1,632 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/ck.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/hip_check_error.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseGemm,
+          typename GemmDesc,
+          bool HasMainKBlockLoop,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_gemm_xdl_splitk(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
+                                       const index_t group_count)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
+    __shared__ uint8_t p_shared[shared_size];
+
+    const index_t block_id = get_block_1d_id();
+    const auto gemm_desc_ptr =
+        reinterpret_cast<const GemmDesc*>(cast_pointer_to_generic_address_space(gemm_descs_const));
+
+    index_t left     = 0;
+    index_t right    = group_count;
+    index_t group_id = index_t((left + right) / 2);
+    while((!(block_id >= gemm_desc_ptr[group_id].block_start_ &&
+             block_id < gemm_desc_ptr[group_id].block_end_)) &&
+          left <= right)
+    {
+        if(block_id < gemm_desc_ptr[group_id].block_start_)
+        {
+            right = group_id;
+        }
+        else
+        {
+            left = group_id;
+        }
+        group_id = index_t((left + right) / 2);
+    }
+
+    GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation>(
+        gemm_desc_ptr[group_id].karg_,
+        static_cast<void*>(p_shared),
+        gemm_desc_ptr[group_id].block_2_ctile_map_);
+#else
+    ignore = gemm_descs_const;
+    ignore = group_count;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          ck::index_t NumGemmKPrefetchStage,
+          ck::index_t BlockSize,
+          ck::index_t MPerBlock,
+          ck::index_t NPerBlock,
+          ck::index_t KPerBlock,
+          ck::index_t AK1,
+          ck::index_t BK1,
+          ck::index_t MPerXDL,
+          ck::index_t NPerXDL,
+          ck::index_t MXdlPerWave,
+          ck::index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t ABlockTransferDstScalarPerVector_K1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferDstScalarPerVector_K1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          PipelineVersion PipelineVer = PipelineVersion::v1,
+          LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          // Current implementation does not support multiple D fusions.
+          enable_if_t<AK1 == BK1 && is_same_v<DsLayout, ck::Tuple<>> &&
+                          is_same_v<DsDataType, ck::Tuple<>>,
+                      bool> = false>
+struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayout,
+                                                                           BLayout,
+                                                                           DsLayout,
+                                                                           ELayout,
+                                                                           ADataType,
+                                                                           BDataType,
+                                                                           DsDataType,
+                                                                           EDataType,
+                                                                           AElementwiseOperation,
+                                                                           BElementwiseOperation,
+                                                                           CDEElementwiseOperation>
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static_assert(KPerBlock % AK1 == 0);
+    static constexpr index_t K0PerBlock = KPerBlock / AK1;
+
+    using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
+        BlockSize,
+        ADataType,
+        BDataType,
+        AccDataType,
+        EDataType,
+        ALayout,
+        BLayout,
+        ELayout,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        GemmSpec,
+        NumGemmKPrefetchStage,
+        MPerBlock,
+        NPerBlock,
+        K0PerBlock,
+        MPerXDL,
+        NPerXDL,
+        AK1,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_K0_M_K1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_K1,
+        false, // AThreadTransferSrcResetCoordinateAfterRun,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_K0_N_K1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_K1,
+        false, // BThreadTransferSrcResetCoordinateAfterRun,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        LoopSched,
+        PipelineVer>;
+
+    using CGridDesc_M_N = typename GridwiseGemm::CGridDesc_M_N;
+    using Block2ETileMapKSplit =
+        BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>;
+    // Block2CTileMap configuration parameter.
+    static constexpr index_t B2E_M01 = 8;
+    using GroupedGemmBlock2ETileMap  = OffsettedBlockToCTileMap<Block2ETileMapKSplit>;
+    using KernelArgument             = typename GridwiseGemm::Argument;
+
+    struct GemmTransKernelArg
+    {
+        KernelArgument karg_;
+        GroupedGemmBlock2ETileMap block_2_ctile_map_;
+        index_t block_start_, block_end_;
+
+        GemmTransKernelArg() = default;
+        GemmTransKernelArg(KernelArgument&& karg,
+                           GroupedGemmBlock2ETileMap&& b2c_map,
+                           index_t block_start,
+                           index_t block_end)
+            : karg_{karg},
+              block_2_ctile_map_{b2c_map},
+              block_start_{block_start},
+              block_end_{block_end}
+        {
+        }
+    };
+
+    static constexpr index_t DefaultKBatch = 1;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+
+        Argument(std::vector<const void*>& p_As,
+                 std::vector<const void*>& p_Bs,
+                 std::vector<void*>& p_Es,
+                 std::vector<GemmDesc>& gemm_descs)
+            : Argument(p_As, p_Bs, p_Es, gemm_descs, DefaultKBatch)
+        {
+            // TODO: use occupancy api to calculate appropriate batch size.
+        }
+
+        Argument(std::vector<const void*>& p_As,
+                 std::vector<const void*>& p_Bs,
+                 std::vector<void*>& p_Es,
+                 std::vector<GemmDesc>& gemm_descs,
+                 index_t kbatch)
+            : K_BATCH{kbatch}
+        {
+            grid_size_   = 0;
+            group_count_ = ck::type_convert<ck::index_t>(gemm_descs.size());
+
+            if(!(group_count_ == ck::type_convert<ck::index_t>(p_As.size()) &&
+                 group_count_ == ck::type_convert<ck::index_t>(p_Bs.size()) &&
+                 group_count_ == ck::type_convert<ck::index_t>(p_Es.size())))
+            {
+                throw std::runtime_error("wrong! group_count_ != p_As/b/c.size");
+            }
+
+            gemm_kernel_args_.reserve(group_count_);
+
+            skipped_group_count_ = 0;
+
+            for(std::size_t i = 0; i < gemm_descs.size(); ++i)
+            {
+                const index_t M = gemm_descs[i].M_;
+                const index_t N = gemm_descs[i].N_;
+                const index_t K = gemm_descs[i].K_;
+
+                if(M == 0)
+                {
+                    skipped_group_count_++;
+                    continue;
+                }
+
+                const index_t stride_a = gemm_descs[i].stride_A_;
+                const index_t stride_b = gemm_descs[i].stride_B_;
+                const index_t stride_c = gemm_descs[i].stride_C_;
+
+                const index_t m_padded = GridwiseGemm::CalculateMPadded(M);
+                const index_t n_padded = GridwiseGemm::CalculateNPadded(N);
+                const index_t k_padded = GridwiseGemm::CalculateKPadded(K, K_BATCH);
+                const index_t k0       = GridwiseGemm::CalculateK0(K, K_BATCH);
+
+                const auto c_grid_desc_m_n = GridwiseGemm::MakeCGridDescriptor_M_N(M, N, stride_c);
+
+                const auto local_b2c_tile_map =
+                    Block2ETileMapKSplit{c_grid_desc_m_n, B2E_M01, K_BATCH};
+                const index_t grid_size_grp = local_b2c_tile_map.CalculateGridSize(c_grid_desc_m_n);
+
+                const index_t block_start = grid_size_;
+                const index_t block_end   = grid_size_ + grid_size_grp;
+
+                grid_size_ += grid_size_grp;
+
+                // block-to-e-tile map
+                auto grouped_block_2_ctile_map =
+                    GroupedGemmBlock2ETileMap(local_b2c_tile_map, block_start);
+
+                auto karg = KernelArgument{type_convert<const ADataType*>(p_As[i]),
+                                           type_convert<const BDataType*>(p_Bs[i]),
+                                           type_convert<EDataType*>(p_Es[i]),
+                                           M,
+                                           N,
+                                           K,
+                                           stride_a,
+                                           stride_b,
+                                           stride_c,
+                                           m_padded,
+                                           n_padded,
+                                           k_padded,
+                                           k0,
+                                           K_BATCH};
+
+                gemm_kernel_args_.emplace_back(
+                    std::move(karg), std::move(grouped_block_2_ctile_map), block_start, block_end);
+            }
+        }
+
+        /**
+         * @brief      Recalculate group grid size for all gemms and update B2C maps.
+         *
+         * @param[in]  kbatch  The new splitK parameter value.
+         */
+        void UpdateKBatch(index_t kbatch)
+        {
+            K_BATCH    = kbatch;
+            grid_size_ = 0;
+
+            for(std::size_t i = 0; i < gemm_kernel_args_.size(); ++i)
+            {
+
+                auto& karg = gemm_kernel_args_[i].karg_;
+
+                const index_t k_padded = GridwiseGemm::CalculateKPadded(karg.K, K_BATCH);
+                const index_t k0       = GridwiseGemm::CalculateK0(karg.K, K_BATCH);
+
+                const auto c_grid_desc_m_n =
+                    GridwiseGemm::MakeCGridDescriptor_M_N(karg.M, karg.N, karg.StrideC);
+
+                const auto local_b2c_tile_map =
+                    Block2ETileMapKSplit{c_grid_desc_m_n, B2E_M01, K_BATCH};
+                const index_t grid_size_grp = local_b2c_tile_map.CalculateGridSize(c_grid_desc_m_n);
+
+                const index_t block_start = grid_size_;
+                const index_t block_end   = grid_size_ + grid_size_grp;
+
+                grid_size_ += grid_size_grp;
+
+                // block-to-e-tile map
+                auto grouped_block_2_ctile_map =
+                    GroupedGemmBlock2ETileMap(local_b2c_tile_map, block_start);
+
+                karg.KPadded                            = k_padded;
+                karg.K0                                 = k0;
+                karg.k_batch                            = K_BATCH;
+                gemm_kernel_args_[i].block_2_ctile_map_ = grouped_block_2_ctile_map;
+                gemm_kernel_args_[i].block_start_       = block_start;
+                gemm_kernel_args_[i].block_end_         = block_end;
+            }
+        }
+
+        //  private:
+        index_t K_BATCH;
+        index_t group_count_;
+        index_t skipped_group_count_;
+
+        std::vector<GemmTransKernelArg> gemm_kernel_args_;
+        index_t grid_size_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            index_t K0                       = arg.gemm_kernel_args_[0].karg_.K0;
+            bool all_have_kbatch_gt_one      = arg.gemm_kernel_args_[0].karg_.k_batch > 1;
+            bool all_have_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            for(std::size_t i = 0; i < arg.gemm_kernel_args_.size(); ++i)
+            {
+                const auto& karg = arg.gemm_kernel_args_[i].karg_;
+                if(stream_config.log_level_ > 0)
+                {
+                    karg.Print();
+                }
+
+                auto kbatch = karg.k_batch;
+
+                if(!GridwiseGemm::CheckValidity(karg))
+                {
+                    std::ostringstream err;
+                    err << "Group id: " << i << " has invalid GridwiseGemm settings!" << __FILE__
+                        << ":" << __LINE__ << ", in function: " << __func__;
+                    throw std::runtime_error(err.str());
+                }
+
+                K0 = karg.K0;
+                bool not_all_have_main_k0_block_loop_same =
+                    all_have_main_k0_block_loop xor GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+                bool not_all_have_kbatch_value_same = all_have_kbatch_gt_one xor (kbatch > 1);
+
+                if(not_all_have_main_k0_block_loop_same)
+                {
+                    std::ostringstream err;
+                    err << "Not all gemms have same value for main_k0_block_loop! in " << __FILE__
+                        << ":" << __LINE__ << ", in function: " << __func__;
+                    throw std::runtime_error(err.str());
+                }
+
+                if(not_all_have_kbatch_value_same)
+                {
+                    std::ostringstream err;
+                    err << "Not all gemms have same kbatch value (=1 or >1)! "
+                        << "group [" << i << "], kbatch: " << kbatch
+                        << ", group [0], kbatch: " << arg.gemm_kernel_args_[0].karg_.k_batch
+                        << " in " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
+                    throw std::runtime_error(err.str());
+                }
+            }
+
+            hip_check_error(
+                hipMemcpyWithStream(arg.p_workspace_,
+                                    arg.gemm_kernel_args_.data(),
+                                    arg.gemm_kernel_args_.size() * sizeof(GemmTransKernelArg),
+                                    hipMemcpyHostToDevice,
+                                    stream_config.stream_id_));
+
+            float ave_time = 0;
+
+            const auto Run = [&](const auto& kernel) {
+                if(all_have_kbatch_gt_one)
+                {
+                    for(const auto& trans_arg : arg.gemm_kernel_args_)
+                    {
+                        const auto& karg = trans_arg.karg_;
+                        hip_check_error(hipMemsetAsync(karg.p_c_grid,
+                                                       0,
+                                                       karg.M * karg.N * sizeof(EDataType),
+                                                       stream_config.stream_id_));
+                    }
+                }
+
+                ave_time =
+                    launch_and_time_kernel(stream_config,
+                                           kernel,
+                                           dim3(arg.grid_size_),
+                                           dim3(BlockSize),
+                                           0,
+                                           cast_pointer_to_constant_address_space(arg.p_workspace_),
+                                           arg.gemm_kernel_args_.size());
+            };
+
+            if(all_have_main_k0_block_loop)
+            {
+                if(all_have_kbatch_gt_one)
+                {
+                    const auto kernel =
+                        kernel_grouped_gemm_xdl_splitk<GridwiseGemm,
+                                                       GemmTransKernelArg,
+                                                       true,
+                                                       InMemoryDataOperationEnum::AtomicAdd>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel =
+                        kernel_grouped_gemm_xdl_splitk<GridwiseGemm,
+                                                       GemmTransKernelArg,
+                                                       true,
+                                                       InMemoryDataOperationEnum::Set>;
+
+                    Run(kernel);
+                }
+            }
+            else
+            {
+                if(all_have_kbatch_gt_one)
+                {
+                    const auto kernel =
+                        kernel_grouped_gemm_xdl_splitk<GridwiseGemm,
+                                                       GemmTransKernelArg,
+                                                       false,
+                                                       InMemoryDataOperationEnum::AtomicAdd>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel =
+                        kernel_grouped_gemm_xdl_splitk<GridwiseGemm,
+                                                       GemmTransKernelArg,
+                                                       false,
+                                                       InMemoryDataOperationEnum::Set>;
+
+                    Run(kernel);
+                }
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if((ck::type_convert<ck::index_t>(arg.gemm_kernel_args_.size()) +
+            arg.skipped_group_count_) != arg.group_count_)
+        {
+#if DEBUG_LOG
+            std::cout << "The group count is not equal to sum of skipped groups "
+                         "and kernel args size!"
+                      << std::endl;
+#endif // DEBUG_LOG
+            return false;
+        }
+
+        bool supported = true;
+        for(std::size_t i = 0; i < arg.gemm_kernel_args_.size(); ++i)
+        {
+            const auto& a        = arg.gemm_kernel_args_[i].karg_;
+            bool group_arg_valid = GridwiseGemm::CheckValidity(a);
+            if(not group_arg_valid)
+            {
+#if DEBUG_LOG
+                std::cout << "[" << __func__ << "] group id: " << i
+                          << " has invalid GridwiseGemm settings!" << std::endl;
+                a.Print();
+#endif // DEBUG_LOG
+            }
+            supported = supported && group_arg_valid;
+        }
+        return supported;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(std::vector<const void*>& p_As,
+                             std::vector<const void*>& p_Bs,
+                             std::vector<std::array<const void*, NumDTensor>>&,
+                             std::vector<void*>& p_Es,
+                             std::vector<GemmDesc> gemm_descs,
+                             AElementwiseOperation,
+                             BElementwiseOperation,
+                             CDEElementwiseOperation)
+    {
+        return Argument{p_As, p_Bs, p_Es, gemm_descs};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(std::vector<const void*>& p_As,
+                        std::vector<const void*>& p_Bs,
+                        std::vector<std::array<const void*, NumDTensor>>&,
+                        std::vector<void*>& p_Es,
+                        std::vector<GemmDesc>& gemm_descs,
+                        AElementwiseOperation,
+                        BElementwiseOperation,
+                        CDEElementwiseOperation) override
+    {
+        return std::make_unique<Argument>(p_As, p_Bs, p_Es, gemm_descs);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedGemm_XdlSplitK"
+            << "<"
+            << std::string(ALayout::name)[0] << ","
+            << std::string(BLayout::name)[0] << ","
+            << std::string(ELayout::name)[0] << ","
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle << ", "
+            << getGemmSpecializationString(GemmSpec)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+
+    size_t GetWorkSpaceSize(const BaseArgument* p_arg) const override
+    {
+        return dynamic_cast<const Argument*>(p_arg)->gemm_kernel_args_.size() *
+               sizeof(GemmTransKernelArg);
+    }
+
+    static void SetKBatchSize(Argument& arg, index_t kbatch) { arg.UpdateKBatch(kbatch); }
+
+    // polymorphic
+    void SetKBatchSize(BaseArgument* p_arg, index_t kbatch) const override
+    {
+        return SetKBatchSize(*dynamic_cast<Argument*>(p_arg), kbatch);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp
new file mode 100644
index 0000000000..19f126e66f
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp
@@ -0,0 +1,408 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/device_image_to_column.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_image_to_column.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InputGridDesc,
+          typename InputDataType,
+          typename OutputGridDesc,
+          typename OutputDataType,
+          typename Block2ETileMap,
+          typename GridwiseImageToColumnKernel>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_image_to_column(const InputGridDesc in_grid_desc,
+                               const InputDataType* __restrict__ p_in_global,
+                               const OutputGridDesc out_grid_desc,
+                               OutputDataType* __restrict__ p_out_global,
+                               const Block2ETileMap block_2_tile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||             \
+    defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx1030__) || defined(__gfx1100__) || \
+    defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx941__) || defined(__gfx942__))
+    GridwiseImageToColumnKernel::Run(
+        in_grid_desc, p_in_global, out_grid_desc, p_out_global, block_2_tile_map);
+#else
+    ignore = in_grid_desc;
+    ignore = p_in_global;
+    ignore = out_grid_desc;
+    ignore = p_out_global;
+    ignore = block_2_tile_map;
+#endif
+}
+
+// Image to column for input layout NDHWC:
+//   input : input image [N, Di, Hi, Wi, C],
+//   output : output image [N * Do * Ho * Wo, Z * Y * X * C]
+template <index_t NDimSpatial,
+          typename InputLayout,
+          typename InputDataType,
+          typename OutputDataType,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t KPerBlock,
+          typename ThreadClusterLengths,
+          index_t ScalarPerVector>
+struct DeviceImageToColumnImpl
+    : public DeviceImageToColumn<NDimSpatial, InputLayout, InputDataType, OutputDataType>
+{
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvolutionForwardSpecialization::Default>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpecialization::MKPadding, index_t, index_t, index_t>{
+            MPerBlock, 0 /* NPerBlock*/, KPerBlock};
+
+    // Use MakeADescriptor_M_K from grouped convolution forward
+    static auto
+    MakeInputDescriptor_M_K(const ck::index_t N,
+                            const ck::index_t C,
+                            const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+                            const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                            const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                            const std::array<index_t, NDimSpatial>& input_left_pads,
+                            const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{1};
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{1};
+        std::array<index_t, NDimSpatial + 3> c_g_n_k_wos_lengths{1};
+
+        auto copy = [](const auto& x, auto& y, index_t dst_offset) {
+            std::copy(x.begin(), x.end(), y.begin() + dst_offset);
+        };
+
+        constexpr index_t spatial_offset = 3;
+
+        copy(input_spatial_lengths, a_g_n_c_wis_lengths, spatial_offset);
+        copy(filter_spatial_lengths, b_g_k_c_xs_lengths, spatial_offset);
+        copy(output_spatial_lengths, c_g_n_k_wos_lengths, spatial_offset);
+
+        // fill only significant values (C and N)
+        a_g_n_c_wis_lengths[I1] = N;
+        a_g_n_c_wis_lengths[I2] = C;
+        b_g_k_c_xs_lengths[I2]  = C;
+        c_g_n_k_wos_lengths[I1] = N;
+
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<InputLayout>(
+                a_g_n_c_wis_lengths,
+                input_g_n_c_wis_strides,
+                b_g_k_c_xs_lengths,
+                {}, // not needed for A Descriptor
+                c_g_n_k_wos_lengths,
+                {}, // not needed for A Descriptor
+                conv_filter_strides,
+                conv_filter_dilations,
+                input_left_pads,
+                input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+        return in_gemmm_gemmk_desc;
+    }
+
+    static auto
+    MakeOutDescriptor_M_K(const ck::index_t N,
+                          const ck::index_t C,
+                          const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                          const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                          const std::array<index_t, 2>& output_m_k_strides)
+    {
+        const index_t NDoHoWo =
+            N * ck::accumulate_n<index_t>(
+                    output_spatial_lengths.begin(), NDimSpatial, 1, std::multiplies<>());
+        const index_t CZYX =
+            C * ck::accumulate_n<index_t>(
+                    filter_spatial_lengths.begin(), NDimSpatial, 1, std::multiplies<>());
+        const auto desc_mraw_kraw = make_naive_tensor_descriptor(
+            make_tuple(NDoHoWo, CZYX), make_tuple(output_m_k_strides[I0], output_m_k_strides[I1]));
+
+        const auto desc_m_k = matrix_padder.PadADescriptor_M_K(desc_mraw_kraw);
+        return desc_m_k;
+    }
+
+    using InputGridDesc =
+        remove_cvref_t<decltype(MakeInputDescriptor_M_K(1, 1, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using OutputGridDesc = remove_cvref_t<decltype(MakeOutDescriptor_M_K(1, 1, {}, {}, {}))>;
+
+    using Block2ETileMap = remove_cvref_t<
+        decltype(BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, KPerBlock, OutputGridDesc>(
+            OutputGridDesc{}))>;
+
+    using GridwiseImageToColumnKernel = GridwiseImageToColumn<InputGridDesc,
+                                                              InputDataType,
+                                                              OutputGridDesc,
+                                                              OutputDataType,
+                                                              BlockSize,
+                                                              MPerBlock,
+                                                              KPerBlock,
+                                                              ThreadClusterLengths,
+                                                              ScalarPerVector,
+                                                              Block2ETileMap>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_in, // input image
+                 void* p_out,      // output image
+                 const ck::index_t N,
+                 const ck::index_t C,
+                 const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+                 const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                 const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                 const std::array<index_t, 2>& output_m_k_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads)
+            : C_(C),
+              X_(filter_spatial_lengths[NDimSpatial - I1]),
+              p_in_{static_cast<const InputDataType*>(p_in)},
+              p_out_{static_cast<OutputDataType*>(p_out)},
+              input_g_n_c_wis_strides_{input_g_n_c_wis_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+
+            in_grid_desc_m_k_ = MakeInputDescriptor_M_K(N,
+                                                        C,
+                                                        input_spatial_lengths,
+                                                        filter_spatial_lengths,
+                                                        output_spatial_lengths,
+                                                        input_g_n_c_wis_strides,
+
+                                                        conv_filter_strides,
+                                                        conv_filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads);
+
+            out_grid_desc_m_k_ = MakeOutDescriptor_M_K(
+                N, C, filter_spatial_lengths, output_spatial_lengths, output_m_k_strides);
+        }
+
+        void Print() const
+        {
+            std::cout << in_grid_desc_m_k_ << std::endl;
+            std::cout << out_grid_desc_m_k_ << std::endl;
+        }
+
+        const ck::index_t C_;
+        const ck::index_t X_;
+
+        const InputDataType* p_in_;
+        OutputDataType* p_out_;
+
+        const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides_;
+        const std::array<index_t, NDimSpatial>& conv_filter_strides_;
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations_;
+        const std::array<index_t, NDimSpatial>& input_left_pads_;
+        const std::array<index_t, NDimSpatial>& input_right_pads_;
+
+        InputGridDesc in_grid_desc_m_k_;
+        OutputGridDesc out_grid_desc_m_k_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            const auto block_2_tile_map =
+                BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, KPerBlock, OutputGridDesc>(
+                    arg.out_grid_desc_m_k_);
+            const index_t grid_size = block_2_tile_map.CalculateGridSize(arg.out_grid_desc_m_k_);
+            const auto kernel       = kernel_image_to_column<InputGridDesc,
+                                                       InputDataType,
+                                                       OutputGridDesc,
+                                                       OutputDataType,
+                                                       Block2ETileMap,
+                                                       GridwiseImageToColumnKernel>;
+
+            float elapsed_time = launch_and_time_kernel(stream_config,
+                                                        kernel,
+                                                        dim3(grid_size),
+                                                        dim3(BlockSize),
+                                                        0,
+                                                        arg.in_grid_desc_m_k_,
+                                                        arg.p_in_,
+                                                        arg.out_grid_desc_m_k_,
+                                                        arg.p_out_,
+                                                        block_2_tile_map);
+            return elapsed_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    bool IsSupportedArgument(const Argument& arg)
+    {
+        using namespace tensor_layout::convolution;
+        if(!(std::is_same_v<InputLayout, GNWC> || std::is_same_v<InputLayout, GNHWC> ||
+             std::is_same_v<InputLayout, GNDHWC>))
+        {
+            return false;
+        }
+        if(!(NDimSpatial >= 1 && NDimSpatial <= 3))
+        {
+            return false;
+        }
+
+        const auto w_pad_left  = arg.input_left_pads_[NDimSpatial - I1];
+        const auto w_pad_right = arg.input_right_pads_[NDimSpatial - I1];
+        const auto dilation_x  = arg.conv_filter_dilations_[NDimSpatial - I1];
+        const auto stride_x    = arg.conv_filter_strides_[NDimSpatial - I1];
+        bool is_w_packed       = arg.input_g_n_c_wis_strides_[NDimSpatial + I2] == arg.C_;
+        bool is_c_packed       = arg.input_g_n_c_wis_strides_[I2] == 1;
+
+        // check vector acces with c not packed
+        if(!is_c_packed && ScalarPerVector != 1)
+            return false;
+        // check vector access of filter window row (only C if C is not packed)
+        if(!is_w_packed && arg.C_ % ScalarPerVector != 0)
+            return false;
+        // check vector access of filter window row (X * C)
+        if(arg.X_ * arg.C_ % ScalarPerVector != 0)
+            return false;
+        // check vector access of pads (w_pad_left/w_pad_right * C)
+        if(w_pad_left * arg.C_ % ScalarPerVector != 0 ||
+           w_pad_right * arg.C_ % ScalarPerVector != 0)
+            return false;
+        // check vector access of with stride and pad
+        if((w_pad_left != 0 || w_pad_right != 0) && stride_x > 1 && arg.C_ % ScalarPerVector != 0)
+            return false;
+        // check vector access of with dilation
+        if(dilation_x > 1 && arg.C_ % ScalarPerVector != 0)
+            return false;
+
+        return GridwiseImageToColumnKernel::CheckValidity(arg.in_grid_desc_m_k_,
+                                                          arg.out_grid_desc_m_k_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_in, // input image
+                             void* p_out,      // output image
+                             const ck::index_t N,
+                             const ck::index_t C,
+                             const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+                             const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                             const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                             const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                             const std::array<index_t, 2>& output_m_k_strides,
+                             const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                             const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                             const std::array<index_t, NDimSpatial>& input_left_pads,
+                             const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        return Argument{static_cast<const InputDataType*>(p_in),
+                        static_cast<OutputDataType*>(p_out),
+                        N,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        input_g_n_c_wis_strides,
+                        output_m_k_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in, // input image
+                        void* p_out,      // output image
+                        const ck::index_t N,
+                        const ck::index_t C,
+                        const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+                        const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                        const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                        const std::array<index_t, 2>& output_m_k_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads) override
+    {
+        return std::make_unique<Argument>(static_cast<const InputDataType*>(p_in),
+                                          static_cast<OutputDataType*>(p_out),
+                                          N,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          input_g_n_c_wis_strides,
+                                          output_m_k_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceImageToColumn"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << KPerBlock << ", "
+            << ScalarPerVector
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp
new file mode 100644
index 0000000000..e98a85defe
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp
@@ -0,0 +1,325 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/stream_utility.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// output[indices] = input
+template <typename DOutDataType,
+          typename IndexDataType,
+          typename DInDataType,
+          ck::index_t InOutVectorSize>
+struct DeviceMaxPoolBwdImpl : public DeviceMaxPoolBwd<DOutDataType, IndexDataType, DInDataType>
+{
+    using DInDataType_AutomicAddPreCast =
+        conditional_t<is_same_v<DInDataType, float> || is_same_v<DInDataType, double>,
+                      DInDataType,
+                      float>;
+
+    using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+    using UnaryConvert = ck::tensor_operation::element_wise::UnaryConvert;
+
+    static constexpr auto I0 = Number<0>{};
+
+    template <typename Desc_M>
+    static auto PadDescriptor_M_1d(Desc_M desc_m, index_t loop_step)
+    {
+        const auto m   = desc_m.GetLength(I0);
+        const auto pad = math::integer_least_multiple(m, loop_step) - m;
+        const auto desc_m_pad =
+            transform_tensor_descriptor(desc_m,
+                                        make_tuple(make_right_pad_transform(m, pad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return desc_m_pad;
+    }
+
+    static auto MakeDescriptor_M(index_t length, index_t loop_step)
+    {
+        const auto desc_m = make_naive_tensor_descriptor_packed(make_tuple(length));
+        return PadDescriptor_M_1d(desc_m, loop_step);
+    }
+
+    using InOutGrid1dDesc = decltype(MakeDescriptor_M(1, 1));
+
+    using GridwisePutElementSet = GridwisePutElement_1D<InOutGrid1dDesc,
+                                                        DOutDataType,
+                                                        IndexDataType,
+                                                        DInDataType,
+                                                        PassThrough,
+                                                        InMemoryDataOperationEnum::Set,
+                                                        InOutVectorSize>;
+
+    using GridwisePutElementAtomicAdd = GridwisePutElement_1D<InOutGrid1dDesc,
+                                                              DOutDataType,
+                                                              IndexDataType,
+                                                              DInDataType_AutomicAddPreCast,
+                                                              PassThrough,
+                                                              InMemoryDataOperationEnum::AtomicAdd,
+                                                              InOutVectorSize>;
+
+    using GridwiseCasting = GridwiseElementwise_1D<Tuple<InOutGrid1dDesc>,
+                                                   Tuple<InOutGrid1dDesc>,
+                                                   Tuple<const DInDataType_AutomicAddPreCast*>,
+                                                   Tuple<DInDataType*>,
+                                                   UnaryConvert,
+                                                   InOutVectorSize,
+                                                   Sequence<InOutVectorSize>,
+                                                   Sequence<InOutVectorSize>>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const DOutDataType* p_dout,
+                 const IndexDataType* p_indices,
+                 DInDataType* p_din,
+                 index_t dout_length,
+                 index_t din_length,
+                 const std::vector<ck::index_t>& window_lengths,
+                 const std::vector<ck::index_t>& window_strides,
+                 const std::vector<ck::index_t>& window_dilations)
+            : p_dout_{p_dout},
+              p_indices_{p_indices},
+              p_din_{p_din},
+              dout_length_raw_{dout_length},
+              din_length_raw_{din_length},
+              blockSize_{256},
+              windowOverlap_{false}
+        {
+            for(size_t i = 0; i < window_lengths.size(); ++i)
+            {
+                auto eff = (window_lengths.at(i) - 1) * window_dilations.at(i) + 1;
+                windowOverlap_ |= eff > window_strides.at(i);
+            }
+        }
+
+        const DOutDataType* p_dout_;
+        const IndexDataType* p_indices_;
+        DInDataType* p_din_;
+        index_t dout_length_raw_;
+        index_t din_length_raw_;
+        index_t blockSize_;
+        bool windowOverlap_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            index_t gridSize               = getAvailableComputeUnitCount(stream_config);
+            index_t loop_step              = gridSize * arg.blockSize_ * InOutVectorSize;
+            InOutGrid1dDesc din_grid_desc  = MakeDescriptor_M(arg.din_length_raw_, loop_step);
+            InOutGrid1dDesc dout_grid_desc = MakeDescriptor_M(arg.dout_length_raw_, loop_step);
+
+            if constexpr(is_same_v<DInDataType, float> || is_same_v<DInDataType, double>)
+            {
+                hip_check_error(hipMemsetAsync(arg.p_din_,
+                                               0,
+                                               arg.din_length_raw_ * sizeof(DInDataType),
+                                               stream_config.stream_id_));
+
+                if(arg.windowOverlap_)
+                {
+                    const auto put_kernel = kernel_put_element_1d<GridwisePutElementAtomicAdd,
+                                                                  InOutGrid1dDesc,
+                                                                  DOutDataType,
+                                                                  IndexDataType,
+                                                                  DInDataType,
+                                                                  PassThrough>;
+
+                    return launch_and_time_kernel(stream_config,
+                                                  put_kernel,
+                                                  dim3(gridSize),
+                                                  dim3(arg.blockSize_),
+                                                  0,
+                                                  dout_grid_desc,
+                                                  arg.p_dout_,
+                                                  arg.p_indices_,
+                                                  arg.p_din_,
+                                                  PassThrough{});
+                }
+                else
+                {
+                    const auto put_kernel = kernel_put_element_1d<GridwisePutElementSet,
+                                                                  InOutGrid1dDesc,
+                                                                  DOutDataType,
+                                                                  IndexDataType,
+                                                                  DInDataType,
+                                                                  PassThrough>;
+
+                    return launch_and_time_kernel(stream_config,
+                                                  put_kernel,
+                                                  dim3(gridSize),
+                                                  dim3(arg.blockSize_),
+                                                  0,
+                                                  dout_grid_desc,
+                                                  arg.p_dout_,
+                                                  arg.p_indices_,
+                                                  arg.p_din_,
+                                                  PassThrough{});
+                }
+            }
+            else
+            {
+                if(arg.windowOverlap_)
+                {
+                    if(arg.p_workspace_ == nullptr)
+                        throw std::runtime_error("wrong! WorkSpace pointer has not been set");
+
+                    hip_check_error(
+                        hipMemsetAsync(arg.p_workspace_,
+                                       0,
+                                       arg.din_length_raw_ * sizeof(DInDataType_AutomicAddPreCast),
+                                       stream_config.stream_id_));
+
+                    const auto put_kernel = kernel_put_element_1d<GridwisePutElementAtomicAdd,
+                                                                  InOutGrid1dDesc,
+                                                                  DOutDataType,
+                                                                  IndexDataType,
+                                                                  DInDataType_AutomicAddPreCast,
+                                                                  PassThrough>;
+
+                    const auto cast_kernel =
+                        kernel_elementwise_1d<GridwiseCasting,
+                                              Tuple<InOutGrid1dDesc>,
+                                              Tuple<InOutGrid1dDesc>,
+                                              Tuple<const DInDataType_AutomicAddPreCast*>,
+                                              Tuple<DInDataType*>,
+                                              UnaryConvert>;
+
+                    float elapsed_time = launch_and_time_kernel(
+                        stream_config,
+                        put_kernel,
+                        dim3(gridSize),
+                        dim3(arg.blockSize_),
+                        0,
+                        dout_grid_desc,
+                        arg.p_dout_,
+                        arg.p_indices_,
+                        static_cast<DInDataType_AutomicAddPreCast*>(arg.p_workspace_),
+                        PassThrough{});
+
+                    elapsed_time += launch_and_time_kernel(
+                        stream_config,
+                        cast_kernel,
+                        dim3(gridSize),
+                        dim3(arg.blockSize_),
+                        0,
+                        ck::make_tuple(din_grid_desc),
+                        ck::make_tuple(din_grid_desc),
+                        static_cast<DInDataType_AutomicAddPreCast*>(arg.p_workspace_),
+                        arg.p_din_,
+                        UnaryConvert{});
+
+                    return elapsed_time;
+                }
+                else
+                {
+                    hip_check_error(hipMemsetAsync(arg.p_din_,
+                                                   0,
+                                                   arg.din_length_raw_ * sizeof(DInDataType),
+                                                   stream_config.stream_id_));
+
+                    const auto put_kernel = kernel_put_element_1d<GridwisePutElementSet,
+                                                                  InOutGrid1dDesc,
+                                                                  DOutDataType,
+                                                                  IndexDataType,
+                                                                  DInDataType,
+                                                                  PassThrough>;
+
+                    hip_check_error(hipMemsetAsync(arg.p_din_,
+                                                   0,
+                                                   arg.din_length_raw_ * sizeof(DInDataType),
+                                                   stream_config.stream_id_));
+
+                    return launch_and_time_kernel(stream_config,
+                                                  put_kernel,
+                                                  dim3(gridSize),
+                                                  dim3(arg.blockSize_),
+                                                  0,
+                                                  dout_grid_desc,
+                                                  arg.p_dout_,
+                                                  arg.p_indices_,
+                                                  arg.p_din_,
+                                                  PassThrough{});
+                }
+            }
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    size_t GetWorkSpaceSize(const BaseArgument* pArg) const override
+    {
+        const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
+
+        bool needCast = pArg_->windowOverlap_ &&
+                        !(is_same_v<DInDataType, float> || is_same_v<DInDataType, double>);
+
+        if(!needCast)
+            return 0;
+        else
+            return pArg_->din_length_raw_ * sizeof(DInDataType_AutomicAddPreCast);
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+        if(pArg->din_length_raw_ % InOutVectorSize != 0 ||
+           pArg->dout_length_raw_ % InOutVectorSize != 0)
+        {
+            return false;
+        }
+        return true;
+    }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_dout,
+                        const void* p_indices,
+                        void* p_din,
+                        index_t dout_length,
+                        index_t din_length,
+                        std::vector<ck::index_t> window_lengths,
+                        std::vector<ck::index_t> window_strides,
+                        std::vector<ck::index_t> window_dilations) override
+    {
+        // Assume p_dout, p_indices, p_din are packed memory space, dout_length and din_length are
+        // physical size of the packed tensor
+        return std::make_unique<Argument>(static_cast<const DOutDataType*>(p_dout),
+                                          static_cast<const IndexDataType*>(p_indices),
+                                          static_cast<DInDataType*>(p_din),
+                                          dout_length,
+                                          din_length,
+                                          window_lengths,
+                                          window_strides,
+                                          window_dilations);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp b/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp
new file mode 100644
index 0000000000..aec5a65ccf
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp
@@ -0,0 +1,595 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/reduction_operator.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/device_multiple_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp"
+
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumReduction,
+          typename InDataType,
+          typename AccDataType,
+          typename OutDataTypeTuple,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple,
+          InMemoryDataOperationEnum OutMemoryDataOperation,
+          bool PropagateNan,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          typename OutDstVectorSizeSeq>
+struct DeviceMultipleReduceMultiBlock : public DeviceMultipleReduce<Rank,
+                                                                    NumReduceDim,
+                                                                    NumReduction,
+                                                                    InElementwiseOperationTuple,
+                                                                    AccElementwiseOperationTuple>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
+                  "Invalid thread cluster size assignments!");
+
+    static_assert((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                      (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert(NumReduction == OutDataTypeTuple::Size() &&
+                      NumReduction == InElementwiseOperationTuple::Size() &&
+                      NumReduction == AccElementwiseOperationTuple::Size() &&
+                      NumReduction == OutDstVectorSizeSeq::Size(),
+                  "All tuple should have the same size as the number of Reductions!");
+
+    static_assert(sequence_all_of(OutDstVectorSizeSeq{},
+                                  [](auto vectorSize) {
+                                      return (MThreadSliceSize % vectorSize == 0);
+                                  }),
+                  "The OutDstVectorSize should completely divide the MThreadSliceSize!");
+
+    static constexpr bool CheckDataTypeTuple()
+    {
+        bool flag = true;
+
+        static_for<0, NumReduction, 1>{}([&](auto I) {
+            using OutDataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+            flag =
+                flag && ck::reduce::InMemoryDataOperationSupportedOnDataType<OutMemoryDataOperation,
+                                                                             OutDataType>::value;
+        });
+
+        return flag;
+    };
+
+    static_assert(CheckDataTypeTuple(),
+                  "The OutDataType must support the specified OutMemoryDataOperation!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+    static constexpr index_t NumInputDim  = Rank;
+    static constexpr index_t NumOutputDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
+    static constexpr bool reduceAllDim    = (NumInvariantDim == 0);
+
+    // So far, only AtomicAdd is considered, other Atomic Operation like AtomicMax can be added
+    // later
+    static constexpr bool use_multiblock =
+        (OutMemoryDataOperation == InMemoryDataOperationEnum::AtomicAdd);
+
+    static_assert(
+        ReduceOperation::IsCompatibleInMemoryDataOperation(OutMemoryDataOperation),
+        "The reduction accumulation operation must be compatible with the OutMemoryDataOperation!");
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto GenerateOutDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+
+                return static_cast<DataType*>(nullptr);
+            },
+            Number<NumReduction>{});
+    };
+
+    using OutDataTypePointerTuple = decltype(GenerateOutDataTypePointerTuple());
+
+    static auto MakeSrc2dDescriptor(const std::array<index_t, NumInputDim>& inLengths,
+                                    const std::array<index_t, NumInputDim>& inStrides,
+                                    int blkGroupSize,
+                                    int numBlockTileIteration)
+    {
+        const auto tupleSrcLengths =
+            generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInputDim>{});
+        const auto tupleSrcStrides =
+            generate_tuple([&](auto I) { return inStrides[I]; }, Number<NumInputDim>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, NumInputDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths = generate_tuple(
+                    [&](auto I) { return inLengths[NumInvariantDim + I]; }, Number<NumReduceDim>{});
+                const auto invariantDimLengths =
+                    generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInvariantDim>{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    static auto MakeDst1dDescriptor(const std::array<index_t, NumOutputDim>& outLengths,
+                                    const std::array<index_t, NumOutputDim>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumOutputDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumOutputDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumOutputDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto outPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto out_grid_desc_m_padded = transform_tensor_descriptor(
+            out_grid_desc_m,
+            make_tuple(make_right_pad_transform(invariantLength, outPad)),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    static auto GenerateOutGrid1dDescTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                (void)I;
+                return MakeDst1dDescriptor(std::array<index_t, NumOutputDim>{},
+                                           std::array<index_t, NumOutputDim>{});
+            },
+            Number<NumReduction>{});
+    };
+
+    using InGridDesc_M_K      = decltype(MakeSrc2dDescriptor(
+        std::array<index_t, NumInputDim>{}, std::array<index_t, NumInputDim>{}, 1, 1));
+    using OutGridDesc_M_Tuple = decltype(GenerateOutGrid1dDescTuple());
+
+    static auto MakeDst1dDescriptorForBufferSet(const std::array<index_t, NumOutputDim>& outLengths,
+                                                const std::array<index_t, NumOutputDim>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumOutputDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumOutputDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumOutputDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto length = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto pad = math::integer_least_multiple(length, BlockSize) - length;
+
+        auto out_grid_desc_m_padded =
+            transform_tensor_descriptor(out_grid_desc_m,
+                                        make_tuple(make_right_pad_transform(length, pad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    static auto GenerateOutGrid1dDescTuple_2()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                (void)I;
+                return MakeDst1dDescriptorForBufferSet(std::array<index_t, NumOutputDim>{},
+                                                       std::array<index_t, NumOutputDim>{});
+            },
+            Number<NumReduction>{});
+    };
+
+    using OutGridDesc_M_Tuple_2 = decltype(GenerateOutGrid1dDescTuple_2());
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, NumInputDim>& inLengths,
+                 const std::array<index_t, NumInputDim>& inStrides,
+                 const std::array<index_t, NumOutputDim>& outLengths,
+                 const std::array<std::array<index_t, NumOutputDim>, NumReduction>& outStridesArray,
+                 const std::array<int, NumReduceDim>& reduceDims,
+                 const std::array<double, NumReduction>& alphas,
+                 const std::array<double, NumReduction>& betas,
+                 const void* in_dev,
+                 const std::array<void*, NumReduction>& out_dev_buffers,
+                 const InElementwiseOperationTuple in_elementwise_op_tuple,
+                 const AccElementwiseOperationTuple acc_elementwise_op_tuple)
+            : outLengths_{outLengths},
+              outStridesArray_{outStridesArray},
+              in_elementwise_op_tuple_{in_elementwise_op_tuple},
+              acc_elementwise_op_tuple_{acc_elementwise_op_tuple}
+        {
+            inLengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
+            inStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inStrides, reduceDims);
+
+            for(size_t i = 0; i < NumReduction; i++)
+            {
+                alpha_values_(i) = static_cast<AccDataType>(alphas[i]);
+                beta_values_(i)  = static_cast<AccDataType>(betas[i]);
+            };
+
+            in_dev_ = static_cast<const InDataType*>(in_dev);
+
+            out_dev_buffers_ = generate_tuple(
+                [&](auto iR) {
+                    using OutDataTypePointer =
+                        remove_cvref_t<decltype(OutDataTypePointerTuple{}[iR])>;
+                    using OutDataType = remove_cvref_t<remove_pointer_t<OutDataTypePointer>>;
+                    return static_cast<OutDataType*>(out_dev_buffers[iR]);
+                },
+                Number<NumReduction>{});
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(inLengths_);
+
+            if constexpr(use_multiblock)
+            {
+
+                int iterations = 1;
+                while(true)
+                {
+                    int testBlkGroupSize =
+                        (reduce_total_length + (K_BlockTileSize * iterations) - 1) /
+                        (K_BlockTileSize * iterations);
+
+                    // we want the blkGroupSize be not more than 128
+                    if(testBlkGroupSize <= 128)
+                        break;
+
+                    iterations++;
+                };
+
+                blkGroupSize = (reduce_total_length + (K_BlockTileSize * iterations) - 1) /
+                               (K_BlockTileSize * iterations);
+
+                numBlockTileIteration = iterations;
+            }
+            else
+            {
+                blkGroupSize = 1;
+                numBlockTileIteration =
+                    (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
+            };
+
+            in_grid_desc_m_k =
+                MakeSrc2dDescriptor(inLengths_, inStrides_, blkGroupSize, numBlockTileIteration);
+
+            out_grid_desc_m_tuple = generate_tuple(
+                [&](auto I) { return MakeDst1dDescriptor(outLengths, outStridesArray[I]); },
+                Number<NumReduction>{});
+
+            out_grid_desc_m_tuple_2 = generate_tuple(
+                [&](auto I) {
+                    return MakeDst1dDescriptorForBufferSet(outLengths, outStridesArray[I]);
+                },
+                Number<NumReduction>{});
+
+            gridSize = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                       M_BlockTileSize * blkGroupSize;
+
+            gridSize_pre =
+                math::integer_least_multiple(invariant_total_length, BlockSize) / BlockSize;
+        }
+
+        std::array<index_t, NumInputDim> inLengths_;
+        std::array<index_t, NumInputDim> inStrides_;
+
+        std::array<index_t, NumOutputDim> outLengths_;
+        std::array<std::array<index_t, NumOutputDim>, NumReduction> outStridesArray_;
+
+        Array<AccDataType, NumReduction> alpha_values_;
+        Array<AccDataType, NumReduction> beta_values_;
+
+        const InDataType* in_dev_;
+        OutDataTypePointerTuple out_dev_buffers_;
+
+        InGridDesc_M_K in_grid_desc_m_k;
+        OutGridDesc_M_Tuple out_grid_desc_m_tuple;
+        OutGridDesc_M_Tuple_2 out_grid_desc_m_tuple_2;
+
+        InElementwiseOperationTuple in_elementwise_op_tuple_;
+        AccElementwiseOperationTuple acc_elementwise_op_tuple_;
+
+        long_index_t invariant_total_length;
+        long_index_t reduce_total_length;
+
+        int blkGroupSize;
+        int numBlockTileIteration;
+        size_t gridSize;
+
+        size_t gridSize_pre;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            using GridwiseMultipleReduce =
+                GridwiseMultipleReduction_mk_to_m_multiblock<NumReduction,
+                                                             InDataType,
+                                                             OutDataTypePointerTuple,
+                                                             AccDataType,
+                                                             InGridDesc_M_K,
+                                                             OutGridDesc_M_Tuple,
+                                                             ReduceOperation,
+                                                             InElementwiseOperationTuple,
+                                                             AccElementwiseOperationTuple,
+                                                             OutMemoryDataOperation,
+                                                             PropagateNan,
+                                                             BlockSize,
+                                                             MThreadClusterSize,
+                                                             KThreadClusterSize,
+                                                             MThreadSliceSize,
+                                                             KThreadSliceSize,
+                                                             InSrcVectorDim,
+                                                             InSrcVectorSize,
+                                                             OutDstVectorSizeSeq>;
+
+            const auto kernel_main =
+                kernel_multiple_reduce_multiblock<GridwiseMultipleReduce,
+                                                  NumReduction,
+                                                  InDataType,
+                                                  OutDataTypePointerTuple,
+                                                  AccDataType,
+                                                  InGridDesc_M_K,
+                                                  OutGridDesc_M_Tuple,
+                                                  InElementwiseOperationTuple,
+                                                  AccElementwiseOperationTuple>;
+
+            float avg_time = 0;
+
+            if constexpr(use_multiblock)
+            {
+                auto identity_values = generate_tuple(
+                    [&](auto iR) {
+                        using OutDataType = remove_cvref_t<decltype(OutDataTypeTuple{}[iR])>;
+                        return ck::reduce::GetIdentityValueForInMemoryDataOperation<OutDataType>(
+                            OutMemoryDataOperation);
+                    },
+                    Number<NumReduction>{});
+
+                const auto kernel_pre = kernel_multiple_buffer_set_value<OutGridDesc_M_Tuple_2,
+                                                                         NumReduction,
+                                                                         BlockSize,
+                                                                         OutDataTypePointerTuple,
+                                                                         OutDataTypeTuple>;
+
+                avg_time += launch_and_time_kernel(stream_config,
+                                                   kernel_pre,
+                                                   dim3(arg.gridSize_pre),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.out_grid_desc_m_tuple_2,
+                                                   arg.out_dev_buffers_,
+                                                   identity_values);
+            };
+
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.in_grid_desc_m_k,
+                                               arg.out_grid_desc_m_tuple,
+                                               arg.in_elementwise_op_tuple_,
+                                               arg.acc_elementwise_op_tuple_,
+                                               arg.blkGroupSize,
+                                               arg.numBlockTileIteration,
+                                               arg.alpha_values_,
+                                               arg.in_dev_,
+                                               arg.beta_values_,
+                                               arg.out_dev_buffers_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        if constexpr(use_multiblock)
+        {
+            for(size_t i = 0; i < pArg->beta_values_.Size(); i++)
+                if(pArg->beta_values_[i] != 0.0f)
+                    return (false);
+        };
+
+        if constexpr(InSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return (false);
+            }
+            else
+            {
+                if(pArg->inStrides_[NumInvariantDim - 1] != 1 && InSrcVectorSize != 1)
+                    return (false);
+
+                if(pArg->inLengths_[NumInvariantDim - 1] % InSrcVectorSize != 0)
+                    return (false);
+            };
+        }
+        else
+        {
+            if(pArg->inStrides_[Rank - 1] != 1 && InSrcVectorSize != 1)
+                return (false);
+
+            if(pArg->inLengths_[Rank - 1] % InSrcVectorSize != 0)
+                return (false);
+        };
+        // To improve
+        bool valid = true;
+        static_for<0, NumReduction, 1>{}([&](auto I) {
+            if(pArg->outStridesArray_[I.value][NumOutputDim - 1] != 1 &&
+               OutDstVectorSizeSeq::At(I) != 1)
+                valid = false;
+
+            if(pArg->outLengths_[NumOutputDim - 1] % OutDstVectorSizeSeq::At(I) != 0)
+                valid = false;
+        });
+
+        if(!valid)
+            return (false);
+
+        if constexpr(use_multiblock)
+        {
+            // blkGroupSize of 1 should be handled by Blockwise path using
+            // InMemoryDataOperationEnum::Set
+            if(pArg->blkGroupSize == 1)
+                return (false);
+
+            // This is very strong restriction, but needed to avoid some failure
+            if(pArg->outLengths_[NumOutputDim - 1] % M_BlockTileSize != 0)
+                return (false);
+        }
+        else
+        {
+            // cases with very small reduce_total_length should be handled by ThreadWise kernel
+            if(pArg->reduce_total_length / KThreadSliceSize < 2)
+                return (false);
+        };
+
+        return (true);
+    };
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, NumInputDim> inLengths,
+        const std::array<index_t, NumInputDim> inStrides,
+        const std::array<index_t, NumOutputDim> outLengths,
+        const std::array<std::array<index_t, NumOutputDim>, NumReduction> outStridesArray,
+        const std::array<int, NumReduceDim> reduceDims,
+        const std::array<double, NumReduction> alphas,
+        const std::array<double, NumReduction> betas,
+        const void* in_dev,
+        const std::array<void*, NumReduction> out_dev_buffers,
+        const InElementwiseOperationTuple in_elementwise_op_tuple,
+        const AccElementwiseOperationTuple acc_elementwise_op_tuple) override
+    {
+        return std::make_unique<Argument>(inLengths,
+                                          inStrides,
+                                          outLengths,
+                                          outStridesArray,
+                                          reduceDims,
+                                          alphas,
+                                          betas,
+                                          in_dev,
+                                          out_dev_buffers,
+                                          in_elementwise_op_tuple,
+                                          acc_elementwise_op_tuple);
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << (OutMemoryDataOperation == InMemoryDataOperationEnum::Set? "DeviceMultipleReduceBlockWise<" : "DeviceMultipleReduceMultiBlock<") << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "InSrcVectorDim_" << InSrcVectorDim << "_InSrcVectorSize_" << InSrcVectorSize << ",";
+        str << "OutDstVectorSize"; 
+        static_for<0, OutDstVectorSizeSeq::Size(), 1>{}([&](auto I) {str << "_" << OutDstVectorSizeSeq::At(I); }); 
+        str << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp b/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp
new file mode 100644
index 0000000000..6d1d5c8e23
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp
@@ -0,0 +1,422 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/reduction_operator.hpp"
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/device_multiple_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp"
+
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumReduction,
+          typename InDataType,
+          typename AccDataType,
+          typename OutDataTypeTuple,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple,
+          bool PropagateNan,
+          index_t BlockSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          typename OutDstVectorSizeSeq>
+struct DeviceMultipleReduceThreadWise : public DeviceMultipleReduce<Rank,
+                                                                    NumReduceDim,
+                                                                    NumReduction,
+                                                                    InElementwiseOperationTuple,
+                                                                    AccElementwiseOperationTuple>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+
+    static_assert((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                      (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert(NumReduction == OutDataTypeTuple::Size() &&
+                      NumReduction == InElementwiseOperationTuple::Size() &&
+                      NumReduction == AccElementwiseOperationTuple::Size() &&
+                      NumReduction == OutDstVectorSizeSeq::Size(),
+                  "All tuple should have the same size as the number of Reductions!");
+
+    static_assert(sequence_all_of(OutDstVectorSizeSeq{},
+                                  [](auto vectorSize) {
+                                      return (MThreadSliceSize % vectorSize == 0);
+                                  }),
+                  "The OutDstVectorSize should completely divide the MThreadSliceSize!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+    static constexpr index_t NumInputDim  = Rank;
+    static constexpr index_t NumOutputDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
+    static constexpr bool reduceAllDim    = (NumInvariantDim == 0);
+
+    static constexpr index_t M_BlockTileSize = BlockSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = 1 * KThreadSliceSize;
+
+    static auto GenerateOutDataTypePointerTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
+
+                return static_cast<DataType*>(nullptr);
+            },
+            Number<NumReduction>{});
+    };
+
+    using OutDataTypePointerTuple = decltype(GenerateOutDataTypePointerTuple());
+
+    static auto MakeSrc2dDescriptor(const std::array<index_t, NumInputDim>& inLengths,
+                                    const std::array<index_t, NumInputDim>& inStrides)
+    {
+        const auto tupleSrcLengths =
+            generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInputDim>{});
+        const auto tupleSrcStrides =
+            generate_tuple([&](auto I) { return inStrides[I]; }, Number<NumInputDim>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, NumInputDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths = generate_tuple(
+                    [&](auto I) { return inLengths[NumInvariantDim + I]; }, Number<NumReduceDim>{});
+                const auto invariantDimLengths =
+                    generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInvariantDim>{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K =
+            math::integer_least_multiple(reduceLength, K_BlockTileSize) - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    static auto MakeDst1dDescriptor(const std::array<index_t, NumOutputDim>& outLengths,
+                                    const std::array<index_t, NumOutputDim>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumOutputDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumOutputDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumOutputDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto outPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto out_grid_desc_m_padded = transform_tensor_descriptor(
+            out_grid_desc_m,
+            make_tuple(make_right_pad_transform(invariantLength, outPad)),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    static auto GenerateOutGrid1dDescTuple()
+    {
+        return generate_tuple(
+            [&](auto I) {
+                (void)I;
+                return MakeDst1dDescriptor(std::array<index_t, NumOutputDim>{},
+                                           std::array<index_t, NumOutputDim>{});
+            },
+            Number<NumReduction>{});
+    };
+
+    using InGridDesc_M_K      = decltype(MakeSrc2dDescriptor(std::array<index_t, NumInputDim>{},
+                                                        std::array<index_t, NumInputDim>{}));
+    using OutGridDesc_M_Tuple = decltype(GenerateOutGrid1dDescTuple());
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, NumInputDim>& inLengths,
+                 const std::array<index_t, NumInputDim>& inStrides,
+                 const std::array<index_t, NumOutputDim>& outLengths,
+                 const std::array<std::array<index_t, NumOutputDim>, NumReduction>& outStridesArray,
+                 const std::array<int, NumReduceDim>& reduceDims,
+                 const std::array<double, NumReduction>& alphas,
+                 const std::array<double, NumReduction>& betas,
+                 const void* in_dev,
+                 const std::array<void*, NumReduction>& out_dev_buffers,
+                 const InElementwiseOperationTuple in_elementwise_op_tuple,
+                 const AccElementwiseOperationTuple acc_elementwise_op_tuple)
+            : outLengths_{outLengths},
+              outStridesArray_{outStridesArray},
+              in_elementwise_op_tuple_{in_elementwise_op_tuple},
+              acc_elementwise_op_tuple_{acc_elementwise_op_tuple}
+        {
+            inLengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
+            inStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inStrides, reduceDims);
+
+            for(size_t i = 0; i < NumReduction; i++)
+            {
+                alpha_values_(i) = static_cast<AccDataType>(alphas[i]);
+                beta_values_(i)  = static_cast<AccDataType>(betas[i]);
+            };
+
+            in_dev_ = static_cast<const InDataType*>(in_dev);
+
+            out_dev_buffers_ = generate_tuple(
+                [&](auto iR) {
+                    using OutDataTypePointer =
+                        remove_cvref_t<decltype(OutDataTypePointerTuple{}[iR])>;
+                    using OutDataType = remove_cvref_t<remove_pointer_t<OutDataTypePointer>>;
+                    return static_cast<OutDataType*>(out_dev_buffers[iR]);
+                },
+                Number<NumReduction>{});
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(inLengths_);
+
+            in_grid_desc_m_k = MakeSrc2dDescriptor(inLengths_, inStrides_);
+
+            out_grid_desc_m_tuple = generate_tuple(
+                [&](auto I) { return MakeDst1dDescriptor(outLengths, outStridesArray[I]); },
+                Number<NumReduction>{});
+
+            gridSize = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                       M_BlockTileSize;
+        }
+
+        std::array<index_t, NumInputDim> inLengths_;
+        std::array<index_t, NumInputDim> inStrides_;
+
+        std::array<index_t, NumOutputDim> outLengths_;
+        std::array<std::array<index_t, NumOutputDim>, NumReduction> outStridesArray_;
+
+        Array<AccDataType, NumReduction> alpha_values_;
+        Array<AccDataType, NumReduction> beta_values_;
+
+        const InDataType* in_dev_;
+        OutDataTypePointerTuple out_dev_buffers_;
+
+        InGridDesc_M_K in_grid_desc_m_k;
+        OutGridDesc_M_Tuple out_grid_desc_m_tuple;
+
+        InElementwiseOperationTuple in_elementwise_op_tuple_;
+        AccElementwiseOperationTuple acc_elementwise_op_tuple_;
+
+        long_index_t invariant_total_length;
+        long_index_t reduce_total_length;
+
+        size_t gridSize;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            using GridwiseMultipleReduce =
+                GridwiseMultipleReduction_mk_to_m_threadwise<NumReduction,
+                                                             InDataType,
+                                                             OutDataTypePointerTuple,
+                                                             AccDataType,
+                                                             InGridDesc_M_K,
+                                                             OutGridDesc_M_Tuple,
+                                                             ReduceOperation,
+                                                             InElementwiseOperationTuple,
+                                                             AccElementwiseOperationTuple,
+                                                             InMemoryDataOperationEnum::Set,
+                                                             PropagateNan,
+                                                             BlockSize,
+                                                             MThreadSliceSize,
+                                                             KThreadSliceSize,
+                                                             InSrcVectorDim,
+                                                             InSrcVectorSize,
+                                                             OutDstVectorSizeSeq>;
+
+            const auto kernel_main =
+                kernel_multiple_reduce_threadwise<GridwiseMultipleReduce,
+                                                  NumReduction,
+                                                  InDataType,
+                                                  OutDataTypePointerTuple,
+                                                  AccDataType,
+                                                  InGridDesc_M_K,
+                                                  OutGridDesc_M_Tuple,
+                                                  InElementwiseOperationTuple,
+                                                  AccElementwiseOperationTuple>;
+
+            float avg_time = 0;
+
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.in_grid_desc_m_k,
+                                               arg.out_grid_desc_m_tuple,
+                                               arg.in_elementwise_op_tuple_,
+                                               arg.acc_elementwise_op_tuple_,
+                                               arg.alpha_values_,
+                                               arg.in_dev_,
+                                               arg.beta_values_,
+                                               arg.out_dev_buffers_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        if constexpr(InSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return (false);
+            }
+            else
+            {
+                if(pArg->inStrides_[NumInvariantDim - 1] != 1 && InSrcVectorSize != 1)
+                    return (false);
+
+                if(pArg->inLengths_[NumInvariantDim - 1] % InSrcVectorSize != 0)
+                    return (false);
+            };
+        }
+        else
+        {
+            if(pArg->inStrides_[Rank - 1] != 1 && InSrcVectorSize != 1)
+                return (false);
+
+            if(pArg->inLengths_[Rank - 1] % InSrcVectorSize != 0)
+                return (false);
+        };
+
+        // To improve
+        bool valid = true;
+        static_for<0, NumReduction, 1>{}([&](auto I) {
+            if(pArg->outStridesArray_[I.value][NumOutputDim - 1] != 1 &&
+               OutDstVectorSizeSeq::At(I) != 1)
+                valid = false;
+
+            if(pArg->outLengths_[NumOutputDim - 1] % OutDstVectorSizeSeq::At(I) != 0)
+                valid = false;
+        });
+
+        if(!valid)
+            return (false);
+
+        return (true);
+    };
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        const std::array<index_t, NumInputDim> inLengths,
+        const std::array<index_t, NumInputDim> inStrides,
+        const std::array<index_t, NumOutputDim> outLengths,
+        const std::array<std::array<index_t, NumOutputDim>, NumReduction> outStridesArray,
+        const std::array<int, NumReduceDim> reduceDims,
+        const std::array<double, NumReduction> alphas,
+        const std::array<double, NumReduction> betas,
+        const void* in_dev,
+        const std::array<void*, NumReduction> out_dev_buffers,
+        const InElementwiseOperationTuple in_elementwise_op_tuple,
+        const AccElementwiseOperationTuple acc_elementwise_op_tuple) override
+    {
+        return std::make_unique<Argument>(inLengths,
+                                          inStrides,
+                                          outLengths,
+                                          outStridesArray,
+                                          reduceDims,
+                                          alphas,
+                                          betas,
+                                          in_dev,
+                                          out_dev_buffers,
+                                          in_elementwise_op_tuple,
+                                          acc_elementwise_op_tuple);
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceMultipleReduceThreadwise<" << BlockSize << ",";
+        str << "M_C" << BlockSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << 1 << "_S" << KThreadSliceSize << ",";
+        str << "InSrcVectorDim_" << InSrcVectorDim << "_InSrcVectorSize_" << InSrcVectorSize << ",";
+        str << "OutDstVectorSize"; 
+        static_for<0, OutDstVectorSizeSeq::Size(), 1>{}([&](auto I) {str << "_" << OutDstVectorSizeSeq::At(I); }); 
+        str << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp
new file mode 100644
index 0000000000..ea0d805043
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp
@@ -0,0 +1,402 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Y = Normalization(X, Beta, Gamma)
+// M: Invarient length
+// K: Reduce length (Calculate mean and variance along K dimension)
+// eg. Length = [N, C, H, W], reduce dim = [C, H, W]
+// Then, M = N, K = C * H * W
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename ComputeDataType,
+          typename YDataType,
+          typename YElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XYSrcVectorDim,
+          index_t XSrcVectorSize,
+          index_t GammaSrcVectorDim,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorDim,
+          index_t BetaSrcVectorSize,
+          index_t YDstVectorSize,
+          bool UseWelford = true>
+struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
+                                                            GammaDataType,
+                                                            BetaDataType,
+                                                            ComputeDataType,
+                                                            YDataType,
+                                                            YElementwiseOperation,
+                                                            Rank,
+                                                            NumReduceDim>
+{
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize);
+    static_assert(
+        ((GammaSrcVectorDim == 0 && MThreadSliceSize % GammaSrcVectorSize == 0) ||
+         (GammaSrcVectorDim == 1 && KThreadSliceSize % GammaSrcVectorSize == 0)),
+        "Invalid thread slice sizes and/or gamma vector sizes configuration, please check!");
+
+    static_assert(
+        ((BetaSrcVectorDim == 0 && MThreadSliceSize % BetaSrcVectorSize == 0) ||
+         (BetaSrcVectorDim == 1 && KThreadSliceSize % BetaSrcVectorSize == 0)),
+        "Invalid thread slice sizes and/or beta vector sizes configuration, please check!");
+
+    using PassThrough = tensor_operation::element_wise::PassThrough;
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
+                                    const std::vector<index_t>& inStrides,
+                                    int numBlockTileIteration)
+    {
+        constexpr index_t NumInvariantDim  = Rank - NumReduceDim;
+        static constexpr index_t numSrcDim = Rank;
+        static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+        const auto tupleSrcLengths = make_tuple_from_array(inLengths, Number<numSrcDim>{});
+        const auto tupleSrcStrides = make_tuple_from_array(inStrides, Number<numSrcDim>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, numSrcDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths =
+                    make_tuple_from_array_and_index_seq(inLengths, ReduceDims{});
+                const auto invariantDimLengths =
+                    make_tuple_from_array_and_index_seq(inLengths, InvariantDims{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K = K_BlockTileSize * numBlockTileIteration - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1));
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::vector<index_t> lengths,
+                 const std::vector<index_t> xStrides,
+                 const std::vector<index_t> gammaStrides,
+                 const std::vector<index_t> betaStrides,
+                 const std::vector<index_t> yStrides,
+                 const std::vector<index_t> reduceDims,
+                 YElementwiseOperation y_elementwise_op,
+                 double epsilon,
+                 const XDataType* p_x,
+                 const GammaDataType* p_gamma,
+                 const BetaDataType* p_beta,
+                 YDataType* p_y)
+            : p_x_(p_x),
+              p_gamma_(p_gamma),
+              p_beta_(p_beta),
+              p_y_(p_y),
+              y_elementwise_op_(y_elementwise_op)
+        {
+            epsilon_ = static_cast<ComputeDataType>(epsilon);
+
+            Lengths_      = shuffle_tensor_dimensions<Rank, NumReduceDim>(lengths, reduceDims);
+            xStrides_     = shuffle_tensor_dimensions<Rank, NumReduceDim>(xStrides, reduceDims);
+            yStrides_     = shuffle_tensor_dimensions<Rank, NumReduceDim>(yStrides, reduceDims);
+            gammaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(gammaStrides, reduceDims);
+            betaStrides_  = shuffle_tensor_dimensions<Rank, NumReduceDim>(betaStrides, reduceDims);
+
+            long_index_t invariant_length;
+            long_index_t reduce_length;
+
+            std::tie(invariant_length, reduce_length) =
+                get_2d_lengths<Rank, NumReduceDim>(Lengths_);
+
+            numBlockTileIteration_ = math::integer_divide_ceil(reduce_length, K_BlockTileSize);
+
+            gridSize_ = math::integer_divide_ceil(invariant_length, M_BlockTileSize);
+
+            x_grid_desc_m_k_ = MakeSrc2dDescriptor(Lengths_, xStrides_, numBlockTileIteration_);
+            gamma_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, gammaStrides_, numBlockTileIteration_);
+            beta_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, betaStrides_, numBlockTileIteration_);
+            y_grid_desc_m_k_ = MakeSrc2dDescriptor(Lengths_, yStrides_, numBlockTileIteration_);
+
+            isSweeponce_ =
+                x_grid_desc_m_k_.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize;
+        }
+
+        ComputeDataType epsilon_;
+
+        const XDataType* p_x_;
+        const GammaDataType* p_gamma_;
+        const BetaDataType* p_beta_;
+        YDataType* p_y_;
+
+        std::vector<index_t> Lengths_;
+        std::vector<index_t> xStrides_;
+        std::vector<index_t> gammaStrides_;
+        std::vector<index_t> betaStrides_;
+        std::vector<index_t> yStrides_;
+
+        YElementwiseOperation y_elementwise_op_;
+
+        int numBlockTileIteration_;
+        size_t gridSize_;
+
+        GridDesc_M_K x_grid_desc_m_k_;
+        GridDesc_M_K gamma_grid_desc_m_k_;
+        GridDesc_M_K beta_grid_desc_m_k_;
+        GridDesc_M_K y_grid_desc_m_k_;
+        bool isSweeponce_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            auto kernel_main = NormalizationKernelSelector<XDataType,
+                                                           GammaDataType,
+                                                           BetaDataType,
+                                                           YDataType,
+                                                           ComputeDataType,
+                                                           YElementwiseOperation,
+                                                           GridDesc_M_K,
+                                                           BlockSize,
+                                                           MThreadClusterSize,
+                                                           KThreadClusterSize,
+                                                           MThreadSliceSize,
+                                                           KThreadSliceSize,
+                                                           XYSrcVectorDim,
+                                                           XSrcVectorSize,
+                                                           GammaSrcVectorDim,
+                                                           GammaSrcVectorSize,
+                                                           BetaSrcVectorDim,
+                                                           BetaSrcVectorSize,
+                                                           XYSrcVectorDim,
+                                                           YDstVectorSize,
+                                                           UseWelford>(arg.isSweeponce_);
+
+            float avg_time = 0;
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize_),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.x_grid_desc_m_k_,
+                                               arg.gamma_grid_desc_m_k_,
+                                               arg.beta_grid_desc_m_k_,
+                                               arg.y_grid_desc_m_k_,
+                                               arg.numBlockTileIteration_,
+                                               arg.epsilon_,
+                                               arg.p_x_,
+                                               arg.p_gamma_,
+                                               arg.p_beta_,
+                                               arg.p_y_,
+                                               arg.y_elementwise_op_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
+
+        constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+        if constexpr(XYSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return false;
+            }
+            else
+            {
+                if(p_arg_->xStrides_[NumInvariantDim - 1] != 1)
+                    return false;
+
+                if(p_arg_->invariant_lowest_length % XSrcVectorSize != 0)
+                    return false;
+
+                if(p_arg_->invariant_lowest_length % YDstVectorSize != 0)
+                    return false;
+            };
+        }
+        else
+        {
+            if(p_arg_->xStrides_[Rank - 1] != 1)
+                return false;
+
+            if(p_arg_->Lengths_[Rank - 1] % XSrcVectorSize != 0)
+                return false;
+
+            if(p_arg_->Lengths_[Rank - 1] % YDstVectorSize != 0)
+            {
+                return false;
+            }
+        };
+
+        // if fastest dim is not reduced
+        if constexpr(GammaSrcVectorDim == 0)
+        {
+            if(p_arg_->gammaStrides_[NumInvariantDim - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % GammaSrcVectorSize != 0)
+                return (false);
+        }
+        else // if fastest dim is reduced
+        {
+            if(p_arg_->gammaStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % GammaSrcVectorSize != 0)
+                return (false);
+        }
+
+        // if fastest dim is not reduced
+        if constexpr(BetaSrcVectorDim == 0)
+        {
+            if(p_arg_->betaStrides_[NumInvariantDim - 1] != 1)
+                return (false);
+
+            if(p_arg_->invariant_lowest_length % BetaSrcVectorSize != 0)
+                return (false);
+        }
+        else // if fastest dim is reduced
+        {
+            if(p_arg_->betaStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(p_arg_->Lengths_[Rank - 1] % BetaSrcVectorSize != 0)
+                return (false);
+        }
+
+        return true;
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> lengths,
+                        const std::vector<index_t> xStrides,
+                        const std::vector<index_t> gammaStrides,
+                        const std::vector<index_t> betaStrides,
+                        const std::vector<index_t> yStrides,
+                        const std::vector<index_t> reduceDims,
+                        double epsilon,
+                        const void* p_x,
+                        const void* p_gamma,
+                        const void* p_beta,
+                        void* p_y,
+                        void* p_saveMean,
+                        void* p_saveInvVar,
+                        YElementwiseOperation y_elementwise_op) override
+    {
+        // TODO
+        // Optional cache of the intermediate results (mean and InvVariance) during the
+        // forward pass could speedup in the backward
+        ignore = p_saveMean;
+        ignore = p_saveInvVar;
+
+        return std::make_unique<Argument>(lengths,
+                                          xStrides,
+                                          gammaStrides,
+                                          betaStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          y_elementwise_op,
+                                          epsilon,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const GammaDataType*>(p_gamma),
+                                          static_cast<const BetaDataType*>(p_beta),
+                                          static_cast<YDataType*>(p_y));
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceNormalizationImpl<" << BlockSize << ",";
+        str << "Cluster_MK_" << MThreadClusterSize << "_" << KThreadClusterSize << ",";
+        str << "Slice_MK_" << MThreadSliceSize << "_" << KThreadSliceSize << ",";
+        str << "XYSrcVectorDim_" << XYSrcVectorDim  << ",";
+        str << "VectorSize_X" << XSrcVectorSize << "_Gamma" << GammaSrcVectorSize << "_Beta" << BetaSrcVectorSize << "_Y" << YDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp
new file mode 100644
index 0000000000..8b2b3c41bf
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp
@@ -0,0 +1,658 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp"
+#include "ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_2nd.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+template <typename GridwiseWelford,
+          typename XDataType,
+          typename MeanVarDataType,
+          typename ComputeDataType,
+          typename XGridDesc_M_K,
+          typename MeanVarGridDesc_M_KBlock>
+__global__ void
+kernel_normalizationSplitK1st(const XGridDesc_M_K x_grid_desc_m_k,
+                              const MeanVarGridDesc_M_KBlock mean_var_grid_desc_m_kblock,
+                              index_t num_k_block_tile_iteration,
+                              const XDataType* const __restrict__ p_x_global,
+                              MeanVarDataType* const __restrict__ p_welford_mean,
+                              MeanVarDataType* const __restrict__ p_welford_variance,
+                              int32_t* const __restrict__ p_welford_count)
+{
+    GridwiseWelford::Run(x_grid_desc_m_k,
+                         mean_var_grid_desc_m_kblock,
+                         num_k_block_tile_iteration,
+                         p_x_global,
+                         p_welford_mean,
+                         p_welford_variance,
+                         p_welford_count);
+};
+
+template <typename GridwiseWelfordNormalization,
+          typename MeanVarDataType,
+          typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename ComputeDataType,
+          typename YElementwiseOperation,
+          typename MeanVarGridDesc_M_KBlock,
+          typename CountGridDesc_M_KBlock,
+          typename XYGammaBetaGridDesc_M_K>
+__global__ void
+kernel_normalizationSplitK2nd(const MeanVarGridDesc_M_KBlock mean_var_grid_desc_m_kblock,
+                              const CountGridDesc_M_KBlock count_grid_desc_m_kblock,
+                              const XYGammaBetaGridDesc_M_K x_grid_desc_m_k,
+                              const XYGammaBetaGridDesc_M_K gamma_grid_desc_m_k,
+                              const XYGammaBetaGridDesc_M_K beta_grid_desc_m_k,
+                              const XYGammaBetaGridDesc_M_K y_grid_desc_m_k,
+                              index_t num_k_mean_var_count_iteration,
+                              index_t num_k_block_tile_iteration,
+                              index_t k_grid_size,
+                              ComputeDataType epsilon,
+                              const MeanVarDataType* const p_mean_global,
+                              const MeanVarDataType* const p_variance_global,
+                              const int32_t* const p_welford_count_global,
+                              const XDataType* const __restrict__ p_x_global,
+                              const GammaDataType* const __restrict__ p_gamma_global,
+                              const BetaDataType* const __restrict__ p_beta_global,
+                              YDataType* const __restrict__ p_y_global,
+                              const YElementwiseOperation y_elementwise_op)
+{
+    GridwiseWelfordNormalization::Run(mean_var_grid_desc_m_kblock,
+                                      count_grid_desc_m_kblock,
+                                      x_grid_desc_m_k,
+                                      gamma_grid_desc_m_k,
+                                      beta_grid_desc_m_k,
+                                      y_grid_desc_m_k,
+                                      num_k_mean_var_count_iteration,
+                                      num_k_block_tile_iteration,
+                                      k_grid_size,
+                                      epsilon,
+                                      p_mean_global,
+                                      p_variance_global,
+                                      p_welford_count_global,
+                                      p_x_global,
+                                      p_gamma_global,
+                                      p_beta_global,
+                                      p_y_global,
+                                      y_elementwise_op);
+};
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Y = Normalization(X, Beta, Gamma)
+// M: Invarient length
+// K: Reduce length (Calculate mean and variance along K dimension)
+// eg. Length = [N, C, H, W], reduce dim = [C, H, W]
+// Then, M = N, K = C * H * W
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename ComputeDataType,
+          typename YDataType,
+          typename YElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XYVectorDim,
+          index_t XSrcVectorSize,
+          index_t GammaSrcVectorDim,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorDim,
+          index_t BetaSrcVectorSize,
+          index_t YDstVectorSize>
+struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
+                                                                  GammaDataType,
+                                                                  BetaDataType,
+                                                                  ComputeDataType,
+                                                                  YDataType,
+                                                                  YElementwiseOperation,
+                                                                  Rank,
+                                                                  NumReduceDim>
+{
+    using MeanVarDataType = ComputeDataType;
+
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize);
+    static_assert(
+        ((GammaSrcVectorDim == 0 && MThreadSliceSize % GammaSrcVectorSize == 0) ||
+         (GammaSrcVectorDim == 1 && KThreadSliceSize % GammaSrcVectorSize == 0)),
+        "Invalid thread slice sizes and/or gamma vector sizes configuration, please check!");
+
+    static_assert(
+        ((BetaSrcVectorDim == 0 && MThreadSliceSize % BetaSrcVectorSize == 0) ||
+         (BetaSrcVectorDim == 1 && KThreadSliceSize % BetaSrcVectorSize == 0)),
+        "Invalid thread slice sizes and/or beta vector sizes configuration, please check!");
+
+    using PassThrough = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
+                                    const std::vector<index_t>& inStrides,
+                                    int kBlockSize,
+                                    int numBlockTileIteration)
+    {
+        constexpr index_t NumInvariantDim  = Rank - NumReduceDim;
+        static constexpr index_t numSrcDim = Rank;
+        static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+        const auto tupleSrcLengths = make_tuple_from_array(inLengths, Number<numSrcDim>{});
+        const auto tupleSrcStrides = make_tuple_from_array(inStrides, Number<numSrcDim>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, numSrcDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths =
+                    make_tuple_from_array_and_index_seq(inLengths, ReduceDims{});
+                const auto invariantDimLengths =
+                    make_tuple_from_array_and_index_seq(inLengths, InvariantDims{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K = reduceSizePerBlock * kBlockSize - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    template <typename DoPads, index_t MPerTile, index_t KPerTile>
+    static auto MakeMeanVarDescriptor_M_K(index_t M, index_t K)
+    {
+        const auto grid_desc_m_k =
+            make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(K, I1));
+        return PadTensorDescriptor(grid_desc_m_k, make_tuple(MPerTile, KPerTile), DoPads{});
+    }
+
+    template <typename DoPads, index_t MPerTile, index_t KPerTile>
+    static auto MakeCountDescriptor_M_K(index_t M, index_t K)
+    {
+        const auto grid_desc_m_k =
+            make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I0, I1));
+        return PadTensorDescriptor(grid_desc_m_k, make_tuple(MPerTile, KPerTile), DoPads{});
+    }
+
+    using SrcGridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1));
+    using Kernel1MeanVarGridDesc_M_KBlock =
+        decltype(MakeMeanVarDescriptor_M_K<Sequence<true, false>, 1, 1>(1, 1));
+
+    using Kernel2MeanVarGridDesc_M_KBlock =
+        decltype(MakeMeanVarDescriptor_M_K<Sequence<true, true>, 1, 1>(1, 1));
+
+    using Kernel2CountGridDesc_M_KBlock =
+        decltype(MakeCountDescriptor_M_K<Sequence<true, true>, 1, 1>(1, 1));
+
+    using GridwiseWelford = GridwiseNormalizationSplitK1st<XDataType,
+                                                           ComputeDataType,
+                                                           MeanVarDataType,
+                                                           SrcGridDesc_M_K,
+                                                           Kernel1MeanVarGridDesc_M_KBlock,
+                                                           BlockSize,
+                                                           MThreadClusterSize,
+                                                           KThreadClusterSize,
+                                                           MThreadSliceSize,
+                                                           KThreadSliceSize,
+                                                           XYVectorDim,
+                                                           XSrcVectorSize>;
+
+    using GridwiseWelfordNormalization =
+        GridwiseNormalizationSplitK2nd<MeanVarDataType,
+                                       XDataType,
+                                       GammaDataType,
+                                       BetaDataType,
+                                       YDataType,
+                                       ComputeDataType,
+                                       YElementwiseOperation,
+                                       Kernel2MeanVarGridDesc_M_KBlock,
+                                       Kernel2CountGridDesc_M_KBlock,
+                                       SrcGridDesc_M_K,
+                                       BlockSize,
+                                       MThreadClusterSize,
+                                       KThreadClusterSize,
+                                       MThreadSliceSize,
+                                       KThreadSliceSize,
+                                       XYVectorDim,
+                                       XSrcVectorSize,
+                                       GammaSrcVectorDim,
+                                       GammaSrcVectorSize,
+                                       BetaSrcVectorDim,
+                                       BetaSrcVectorSize,
+                                       XYVectorDim,
+                                       YDstVectorSize>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::vector<index_t> lengths,
+                 const std::vector<index_t> xStrides,
+                 const std::vector<index_t> gammaStrides,
+                 const std::vector<index_t> betaStrides,
+                 const std::vector<index_t> yStrides,
+                 const std::vector<index_t> reduceDims,
+                 YElementwiseOperation y_elementwise_op,
+                 double epsilon,
+                 const XDataType* p_x,
+                 const GammaDataType* p_gamma,
+                 const BetaDataType* p_beta,
+                 YDataType* p_y)
+            : p_x_(p_x),
+              p_gamma_(p_gamma),
+              p_beta_(p_beta),
+              p_y_(p_y),
+              p_workspace_mean_{nullptr},
+              p_workspace_var_{nullptr},
+              p_workspace_count_{nullptr},
+              y_elementwise_op_(y_elementwise_op)
+        {
+            epsilon_ = static_cast<ComputeDataType>(epsilon);
+
+            Lengths_      = shuffle_tensor_dimensions<Rank, NumReduceDim>(lengths, reduceDims);
+            xStrides_     = shuffle_tensor_dimensions<Rank, NumReduceDim>(xStrides, reduceDims);
+            yStrides_     = shuffle_tensor_dimensions<Rank, NumReduceDim>(yStrides, reduceDims);
+            gammaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(gammaStrides, reduceDims);
+            betaStrides_  = shuffle_tensor_dimensions<Rank, NumReduceDim>(betaStrides, reduceDims);
+
+            std::tie(MRaw_, KRaw_) = get_2d_lengths<Rank, NumReduceDim>(Lengths_);
+
+            numBlockTileIteration_ = 1;
+            while(true)
+            {
+                int testKGridSize =
+                    math::integer_divide_ceil(KRaw_, K_BlockTileSize * numBlockTileIteration_);
+
+                // we want the kGridSize_ be not more than 128
+                if(testKGridSize <= 128)
+                    break;
+
+                ++numBlockTileIteration_;
+            };
+
+            kGridSize_ = math::integer_divide_ceil(KRaw_, K_BlockTileSize * numBlockTileIteration_);
+            gridSize_  = math::integer_divide_ceil(MRaw_, M_BlockTileSize) * kGridSize_;
+
+            // We do not use vector load for mean, var and count
+            static constexpr index_t K_MeanVarCountBlockTileSize = KThreadClusterSize;
+
+            numMeanVarCountIteration_ =
+                math::integer_divide_ceil(kGridSize_, K_MeanVarCountBlockTileSize);
+
+            x_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, xStrides_, kGridSize_, numBlockTileIteration_);
+            gamma_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, gammaStrides_, kGridSize_, numBlockTileIteration_);
+            beta_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, betaStrides_, kGridSize_, numBlockTileIteration_);
+            y_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(Lengths_, yStrides_, kGridSize_, numBlockTileIteration_);
+
+            // We don't need to pad in K dimension for Welford1. Set KPerTile 1.
+            kernel1_mean_var_grid_desc_m_kblock_ =
+                MakeMeanVarDescriptor_M_K<Sequence<true, false>, M_BlockTileSize, 1>(MRaw_,
+                                                                                     kGridSize_);
+
+            kernel2_mean_var_grid_desc_m_kblock_ =
+                MakeMeanVarDescriptor_M_K<Sequence<true, true>,
+                                          M_BlockTileSize,
+                                          K_MeanVarCountBlockTileSize>(MRaw_, kGridSize_);
+
+            kernel2_count_grid_desc_m_kblock_ =
+                MakeCountDescriptor_M_K<Sequence<true, true>,
+                                        M_BlockTileSize,
+                                        K_MeanVarCountBlockTileSize>(MRaw_, kGridSize_);
+        }
+
+        ComputeDataType epsilon_;
+
+        const XDataType* p_x_;
+        const GammaDataType* p_gamma_;
+        const BetaDataType* p_beta_;
+        YDataType* p_y_;
+        void* p_workspace_mean_;
+        void* p_workspace_var_;
+        void* p_workspace_count_;
+
+        std::vector<index_t> Lengths_;
+        std::vector<index_t> xStrides_;
+        std::vector<index_t> gammaStrides_;
+        std::vector<index_t> betaStrides_;
+        std::vector<index_t> yStrides_;
+
+        YElementwiseOperation y_elementwise_op_;
+
+        int kGridSize_;
+        int numMeanVarCountIteration_;
+        int numBlockTileIteration_;
+        size_t gridSize_;
+
+        SrcGridDesc_M_K x_grid_desc_m_k_;
+        SrcGridDesc_M_K gamma_grid_desc_m_k_;
+        SrcGridDesc_M_K beta_grid_desc_m_k_;
+        SrcGridDesc_M_K y_grid_desc_m_k_;
+
+        Kernel1MeanVarGridDesc_M_KBlock kernel1_mean_var_grid_desc_m_kblock_;
+        Kernel2MeanVarGridDesc_M_KBlock kernel2_mean_var_grid_desc_m_kblock_;
+        Kernel2CountGridDesc_M_KBlock kernel2_count_grid_desc_m_kblock_;
+
+        index_t MRaw_; // invarient length
+        index_t KRaw_; // reduce length
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(arg.p_workspace_mean_ == nullptr || arg.p_workspace_var_ == nullptr ||
+               arg.p_workspace_count_ == nullptr)
+                throw std::runtime_error("wrong! WorkSpace pointer has not been set");
+
+            auto kernel1 = kernel_normalizationSplitK1st<GridwiseWelford,
+                                                         XDataType,
+                                                         MeanVarDataType,
+                                                         ComputeDataType,
+                                                         SrcGridDesc_M_K,
+                                                         Kernel1MeanVarGridDesc_M_KBlock>;
+
+            auto kernel2 = kernel_normalizationSplitK2nd<GridwiseWelfordNormalization,
+                                                         MeanVarDataType,
+                                                         XDataType,
+                                                         GammaDataType,
+                                                         BetaDataType,
+                                                         YDataType,
+                                                         ComputeDataType,
+                                                         YElementwiseOperation,
+                                                         Kernel2MeanVarGridDesc_M_KBlock,
+                                                         Kernel2CountGridDesc_M_KBlock,
+                                                         SrcGridDesc_M_K>;
+
+            float avg_time = 0;
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel1,
+                                               dim3(arg.gridSize_),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.x_grid_desc_m_k_,
+                                               arg.kernel1_mean_var_grid_desc_m_kblock_,
+                                               arg.numBlockTileIteration_,
+                                               arg.p_x_,
+                                               static_cast<MeanVarDataType*>(arg.p_workspace_mean_),
+                                               static_cast<MeanVarDataType*>(arg.p_workspace_var_),
+                                               static_cast<int32_t*>(arg.p_workspace_count_));
+
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel2,
+                                               dim3(arg.gridSize_),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.kernel2_mean_var_grid_desc_m_kblock_,
+                                               arg.kernel2_count_grid_desc_m_kblock_,
+                                               arg.x_grid_desc_m_k_,
+                                               arg.gamma_grid_desc_m_k_,
+                                               arg.beta_grid_desc_m_k_,
+                                               arg.y_grid_desc_m_k_,
+                                               arg.numMeanVarCountIteration_,
+                                               arg.numBlockTileIteration_,
+                                               arg.kGridSize_,
+                                               arg.epsilon_,
+                                               static_cast<MeanVarDataType*>(arg.p_workspace_mean_),
+                                               static_cast<MeanVarDataType*>(arg.p_workspace_var_),
+                                               static_cast<int32_t*>(arg.p_workspace_count_),
+                                               arg.p_x_,
+                                               arg.p_gamma_,
+                                               arg.p_beta_,
+                                               arg.p_y_,
+                                               arg.y_elementwise_op_);
+
+            return avg_time;
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    size_t GetWorkSpaceSize(const BaseArgument* pArg) const override
+    {
+        const Argument* pArg_ = dynamic_cast<const Argument*>(pArg);
+
+        size_t workspace_size = 0;
+
+        int welford_size = pArg_->MRaw_ * pArg_->kGridSize_;
+
+        // workspace for welford intermediate mean
+        workspace_size += welford_size * sizeof(MeanVarDataType) + 64;
+
+        // workspace for welford intermediate variance
+        workspace_size += welford_size * sizeof(MeanVarDataType) + 64;
+
+        // workspace for welford intermediate count
+        workspace_size += pArg_->kGridSize_ * sizeof(int32_t) + 64;
+
+        return (workspace_size);
+    };
+
+    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    {
+        Argument* pArg_ = dynamic_cast<Argument*>(pArg);
+
+        pArg_->p_workspace_ = p_workspace;
+
+        int welford_size = pArg_->MRaw_ * pArg_->kGridSize_;
+
+        // setup buffer used for intermediate welford mean
+        pArg_->p_workspace_mean_ = static_cast<char*>(pArg_->p_workspace_);
+
+        index_t mean_space_sz = welford_size * sizeof(MeanVarDataType);
+        mean_space_sz         = math::integer_least_multiple(mean_space_sz, 64);
+
+        // setup buffer used for intermediate welford varirance
+        pArg_->p_workspace_var_ = reinterpret_cast<char*>(pArg_->p_workspace_mean_) + mean_space_sz;
+
+        index_t variance_space_sz = welford_size * sizeof(MeanVarDataType);
+        variance_space_sz         = math::integer_least_multiple(variance_space_sz, 64);
+
+        // setup buffer used for intermediate welford count
+        pArg_->p_workspace_count_ =
+            reinterpret_cast<char*>(pArg_->p_workspace_var_) + variance_space_sz;
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
+
+        constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+        if constexpr(XYVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return false;
+            }
+            else
+            {
+                if(p_arg_->xStrides_[NumInvariantDim - 1] != 1)
+                    return false;
+
+                if(p_arg_->invariant_lowest_length % XSrcVectorSize != 0)
+                    return false;
+
+                if(p_arg_->invariant_lowest_length % YDstVectorSize != 0)
+                    return false;
+            };
+        }
+        else
+        {
+            if(p_arg_->xStrides_[Rank - 1] != 1)
+                return false;
+
+            if(p_arg_->Lengths_[Rank - 1] % XSrcVectorSize != 0)
+                return false;
+
+            if(p_arg_->Lengths_[Rank - 1] % YDstVectorSize != 0)
+                return false;
+        };
+
+        // if fastest dim is not reduced
+        if constexpr(GammaSrcVectorDim == 0)
+        {
+            if(p_arg_->gammaStrides_[NumInvariantDim - 1] != 1)
+                return false;
+
+            if(p_arg_->Lengths_[Rank - 1] % GammaSrcVectorSize != 0)
+                return false;
+        }
+        else // if fastest dim is reduced
+        {
+            if(p_arg_->gammaStrides_[Rank - 1] != 1)
+                return false;
+
+            if(p_arg_->Lengths_[Rank - 1] % GammaSrcVectorSize != 0)
+                return false;
+        }
+
+        // if fastest dim is not reduced
+        if constexpr(BetaSrcVectorDim == 0)
+        {
+            if(p_arg_->betaStrides_[NumInvariantDim - 1] != 1)
+                return false;
+
+            if(p_arg_->invariant_lowest_length % BetaSrcVectorSize != 0)
+                return false;
+        }
+        else // if fastest dim is reduced
+        {
+            if(p_arg_->betaStrides_[Rank - 1] != 1)
+                return false;
+
+            if(p_arg_->Lengths_[Rank - 1] % BetaSrcVectorSize != 0)
+                return false;
+        }
+
+        if(p_arg_->kGridSize_ <= 1)
+            return false;
+
+        return true;
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::vector<index_t> lengths,
+                        const std::vector<index_t> xStrides,
+                        const std::vector<index_t> gammaStrides,
+                        const std::vector<index_t> betaStrides,
+                        const std::vector<index_t> yStrides,
+                        const std::vector<index_t> reduceDims,
+                        double epsilon,
+                        const void* p_x,
+                        const void* p_gamma,
+                        const void* p_beta,
+                        void* p_y,
+                        void* p_saveMean,
+                        void* p_saveInvVar,
+                        YElementwiseOperation y_elementwise_op) override
+    {
+        // TODO
+        // Optional cache of the intermediate results (mean and InvVariance) during the
+        // forward pass could speedup in the backward
+        ignore = p_saveMean;
+        ignore = p_saveInvVar;
+
+        return std::make_unique<Argument>(lengths,
+                                          xStrides,
+                                          gammaStrides,
+                                          betaStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          y_elementwise_op,
+                                          epsilon,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const GammaDataType*>(p_gamma),
+                                          static_cast<const BetaDataType*>(p_beta),
+                                          static_cast<YDataType*>(p_y));
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceNormalizationSplitKImpl<" << BlockSize << ",";
+        str << "Cluster_MK_" << MThreadClusterSize << "_" << KThreadClusterSize << ",";
+        str << "Slice_MK_" << MThreadSliceSize << "_" << KThreadSliceSize << ",";
+        str << "XYSrcVectorDim_" << XYVectorDim  << ",";
+        str << "VectorSize_X" << XSrcVectorSize << "_Gamma" << GammaSrcVectorSize << "_Beta" << BetaSrcVectorSize << "_Y" << YDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp
new file mode 100644
index 0000000000..17dab08332
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp
@@ -0,0 +1,282 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include <memory>
+#include <utility>
+
+#include "ck/utility/math.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/device_permute.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_permute.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Swap last 2 dimensions
+// input shape: [d[0], d[1], d[2], ..., d[NumDim-3], d[NumDim-2], d[NumDim-1]]
+//                                                                ^^^^^^^^^^^
+// output shape: [d[0], d[1], d[2], ..., d[NumDim-3], d[NumDim-1], d[NumDim-2]]
+//                                                    ^^^^^^^^^^^
+template <index_t NumDim,
+          typename InDataType,
+          typename OutDataType,
+          typename ElementwiseOperation,
+          index_t BlockSize,
+          index_t NPerBlock,
+          index_t HPerBlock,
+          index_t WPerBlock,
+          index_t InBlockLdsExtraW,
+          typename InBlockTransferThreadClusterLengths,
+          typename InBlockTransferThreadClusterArrangeOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector>
+struct DevicePermuteImpl : DevicePermute<NumDim, InDataType, OutDataType, ElementwiseOperation>
+{
+    using BaseType = DevicePermute<NumDim, InDataType, OutDataType, ElementwiseOperation>;
+    using typename BaseType::Lengths;
+    using typename BaseType::Strides;
+
+    static_assert(3 <= NumDim, "Only accept at least 3D dimension tensor");
+    static_assert((NumDim - 2) <= SrcVectorDim && SrcVectorDim < NumDim);
+    static_assert((NumDim - 2) <= DstVectorDim && DstVectorDim < NumDim);
+    static_assert(SrcVectorDim != DstVectorDim);
+
+    template <index_t N = NumDim>
+    static auto ConvertArrayToTuple(const std::array<index_t, NumDim>& array)
+    {
+        static_assert(1 <= N && N <= NumDim);
+
+        return generate_tuple([&](auto I) { return array[I]; }, Number<N>{});
+    }
+
+    static auto MakeDescriptor_N_H_W(const Lengths& lengths, const Strides& stride)
+    {
+        // create nd descriptor, shape: [d[0], d[1], d[2], ..., d[NumDim-3], d[NumDim-2],
+        // d[NumDim-1]]
+        const auto desc =
+            make_naive_tensor_descriptor(ConvertArrayToTuple(lengths), ConvertArrayToTuple(stride));
+
+        // merge nd to 3d descriptor, shape: [(d[0] * d[1] * d[2] * ... * d[NumDim-3]), d[NumDim-2],
+        // d[NumDim-1]]
+        //                                   => [N, H, W]
+        const index_t H       = *std::next(rbegin(lengths));
+        const index_t W       = *rbegin(lengths);
+        const auto desc_n_h_w = transform_tensor_descriptor(
+            desc,
+            make_tuple(make_merge_transform(ConvertArrayToTuple<NumDim - 2>(lengths)),
+                       make_pass_through_transform(H),
+                       make_pass_through_transform(W)),
+            make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<NumDim - 2>{}),
+                       Sequence<NumDim - 2>{},
+                       Sequence<NumDim - 1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        return PadTensorDescriptor(
+            desc_n_h_w, make_tuple(NPerBlock, HPerBlock, WPerBlock), Sequence<true, true, true>{});
+    }
+
+    using InGridDesc  = decltype(MakeDescriptor_N_H_W({1, 1}, {1, 1}));
+    using OutGridDesc = InGridDesc;
+
+    using GridwisePermute = GridwisePermute<
+        InGridDesc,
+        OutGridDesc,
+        InDataType,
+        OutDataType,
+        ElementwiseOperation,
+        BlockSize,
+        NPerBlock,
+        HPerBlock,
+        WPerBlock,
+        InBlockLdsExtraW,
+        InBlockTransferThreadClusterLengths,
+        InBlockTransferThreadClusterArrangeOrder,
+        SrcVectorDim - (NumDim - 3), // calculate new SrcVectorDim for the merged descriptor
+        DstVectorDim - (NumDim - 3), // calculate new DstVectorDim for the merged descriptor
+        SrcScalarPerVector,
+        DstScalarPerVector>;
+
+    using Block2TileMap = typename GridwisePermute::DefaultBlock2TileMap;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const Lengths& in_lengths,
+                 const Strides& in_strides,
+                 const Lengths& out_lengths,
+                 const Strides& out_strides,
+                 const void* in_dev_buffer,
+                 void* out_dev_buffer,
+                 ElementwiseOperation elementwise_op)
+            : in_dev_buffer_(static_cast<const InDataType*>(in_dev_buffer)),
+              out_dev_buffer_(static_cast<OutDataType*>(out_dev_buffer)),
+              in_grid_desc_(MakeDescriptor_N_H_W(in_lengths, in_strides)),
+              out_grid_desc_(MakeDescriptor_N_H_W(out_lengths, out_strides)),
+              in_lengths_(in_lengths),
+              in_strides_(in_strides),
+              out_lengths_(out_lengths),
+              out_strides_(out_strides),
+              elementwise_op_(elementwise_op),
+              block_2_tile_map_(GridwisePermute::MakeDefaultBlock2TileMap(in_grid_desc_))
+        {
+        }
+
+        const InDataType* in_dev_buffer_;
+        OutDataType* out_dev_buffer_;
+        InGridDesc in_grid_desc_;
+        OutGridDesc out_grid_desc_;
+
+        Lengths in_lengths_;
+        Strides in_strides_;
+        Lengths out_lengths_;
+        Strides out_strides_;
+
+        ElementwiseOperation elementwise_op_;
+
+        Block2TileMap block_2_tile_map_;
+    };
+
+    struct Invoker : BaseInvoker
+    {
+        static float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const index_t grid_size = arg.block_2_tile_map_.CalculateGridSize(arg.in_grid_desc_);
+
+            const auto kernel = kernel_nd_permute<GridwisePermute,
+                                                  InGridDesc,
+                                                  OutGridDesc,
+                                                  InDataType,
+                                                  OutDataType,
+                                                  ElementwiseOperation,
+                                                  Block2TileMap>;
+
+            float elapsed_time = launch_and_time_kernel(stream_config,
+                                                        kernel,
+                                                        dim3(grid_size),
+                                                        dim3(BlockSize),
+                                                        0,
+                                                        arg.in_grid_desc_,
+                                                        arg.out_grid_desc_,
+                                                        arg.in_dev_buffer_,
+                                                        arg.out_dev_buffer_,
+                                                        arg.elementwise_op_,
+                                                        arg.block_2_tile_map_);
+            return elapsed_time;
+        }
+
+        float Run(const BaseArgument* arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override final
+        {
+            const auto* const argument = dynamic_cast<const Argument*>(arg);
+            if(!argument)
+            {
+                return NAN;
+            }
+
+            return Run(*argument, stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        constexpr auto GetPaddedLength = [](index_t length, index_t tile_length) {
+            return math::integer_divide_ceil(length, tile_length) * tile_length;
+        };
+
+        constexpr auto IsScalarPerVectorValid =
+            [](index_t length, index_t stride, index_t scalar_per_vector) {
+                if(stride == 1 && length % scalar_per_vector == 0)
+                {
+                    return true;
+                }
+                else if(stride != 1 && scalar_per_vector == 1)
+                {
+                    return true;
+                }
+
+                return false;
+            };
+
+        return IsScalarPerVectorValid(arg.in_lengths_[SrcVectorDim],
+                                      arg.in_strides_[SrcVectorDim],
+                                      SrcScalarPerVector) &&
+               IsScalarPerVectorValid(
+                   GetPaddedLength(arg.in_lengths_[SrcVectorDim],
+                                   (SrcVectorDim == NumDim - 2 ? HPerBlock : WPerBlock)),
+                   arg.in_strides_[SrcVectorDim],
+                   SrcScalarPerVector) &&
+               IsScalarPerVectorValid(arg.out_lengths_[DstVectorDim],
+                                      arg.out_strides_[DstVectorDim],
+                                      DstScalarPerVector) &&
+               IsScalarPerVectorValid(
+                   GetPaddedLength(arg.out_lengths_[DstVectorDim],
+                                   (DstVectorDim == NumDim - 2 ? HPerBlock : WPerBlock)),
+                   arg.in_strides_[DstVectorDim],
+                   DstScalarPerVector) &&
+               GridwisePermute::CheckValidity(arg.in_grid_desc_, arg.out_grid_desc_);
+    };
+
+    // override methods inherited from 'BaseOperator'
+    bool IsSupportedArgument(const BaseArgument* arg) override final
+    {
+        const auto* const argument = dynamic_cast<const Argument*>(arg);
+        if(!argument)
+        {
+            return false;
+        }
+
+        return IsSupportedArgument(*argument);
+    }
+
+    // override methods inherited from 'DevicePermute'
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const Lengths& in_lengths,
+                        const Strides& in_strides,
+                        const Lengths& out_lengths,
+                        const Strides& out_strides,
+                        const void* in_dev_buffer,
+                        void* out_dev_buffer,
+                        ElementwiseOperation elementwise_op) override final
+    {
+        return std::make_unique<Argument>(in_lengths,
+                                          in_strides,
+                                          out_lengths,
+                                          out_strides,
+                                          in_dev_buffer,
+                                          out_dev_buffer,
+                                          elementwise_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override final
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    // other constructor methods
+    template <typename... Args>
+    static std::enable_if_t<std::is_constructible_v<Argument, Args...>, Argument>
+    MakeArgument(Args&&... args) noexcept(std::is_nothrow_constructible_v<Argument, Args...>)
+    {
+        return Argument{std::forward<Args>(args)...};
+    }
+
+    static std::enable_if_t<std::is_default_constructible_v<Invoker>, Invoker>
+    MakeInvoker() noexcept(std::is_nothrow_default_constructible_v<Invoker>)
+    {
+        return Invoker{};
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp b/include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp
new file mode 100644
index 0000000000..c94c568c49
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp
@@ -0,0 +1,113 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataType,
+          typename OutDataType,
+          typename IndexDataType, // enable if OutputIndex == true
+          typename ComputeDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          bool OutputIndex,
+          ck::index_t BlockSize,
+          ck::index_t ReduceMThreadClusterSize,
+          ck::index_t ReduceKThreadClusterSize,
+          ck::index_t ReduceMThreadSliceSize,
+          ck::index_t ReduceKThreadSliceSize,
+          ck::index_t InSrcOutDstVectorSize>
+struct DevicePool2dFwd_NHWC_NHWC : public DevicePool3dFwd_NDHWC_NDHWC<InDataType,
+                                                                      OutDataType,
+                                                                      IndexDataType,
+                                                                      ComputeDataType,
+                                                                      ReduceOpId,
+                                                                      OutputIndex,
+                                                                      BlockSize,
+                                                                      ReduceMThreadClusterSize,
+                                                                      ReduceKThreadClusterSize,
+                                                                      ReduceMThreadSliceSize,
+                                                                      ReduceKThreadSliceSize,
+                                                                      InSrcOutDstVectorSize>
+{
+    using DevicePool3D = DevicePool3dFwd_NDHWC_NDHWC<InDataType,
+                                                     OutDataType,
+                                                     IndexDataType,
+                                                     ComputeDataType,
+                                                     ReduceOpId,
+                                                     OutputIndex,
+                                                     BlockSize,
+                                                     ReduceMThreadClusterSize,
+                                                     ReduceKThreadClusterSize,
+                                                     ReduceMThreadSliceSize,
+                                                     ReduceKThreadSliceSize,
+                                                     InSrcOutDstVectorSize>;
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_dev,
+                        void* p_out_dev,
+                        void* p_out_indices_dev,
+                        std::vector<ck::index_t> input_lengths,
+                        std::vector<ck::index_t> window_lengths,
+                        std::vector<ck::index_t> output_lengths,
+                        std::vector<ck::index_t> input_stride,
+                        std::vector<ck::index_t> output_stride,
+                        std::vector<ck::index_t> indices_stride,
+                        std::vector<ck::index_t> window_strides,
+                        std::vector<ck::index_t> window_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads,
+                        std::vector<ck::index_t> pooling_dims) override
+    {
+        static constexpr index_t InOutRank  = 4;
+        static constexpr index_t WindowRank = 2;
+
+        if(input_lengths.size() != InOutRank || window_lengths.size() != WindowRank ||
+           input_lengths.size() != InOutRank || window_strides.size() != WindowRank ||
+           window_dilations.size() != WindowRank || input_left_pads.size() != WindowRank ||
+           input_right_pads.size() != WindowRank)
+            throw std::runtime_error("dimension is incorrect");
+
+        if(pooling_dims != std::vector<ck::index_t>{2, 3})
+            throw std::runtime_error("pooling_dims only support {2, 3} in pool2d so far");
+
+        // NCHW to NCDHW
+        input_lengths.insert(input_lengths.begin() + 2, 1);
+        output_lengths.insert(output_lengths.begin() + 2, 1);
+        input_stride.insert(input_stride.begin() + 2, 0);
+        output_stride.insert(output_stride.begin() + 2, 0);
+        indices_stride.insert(indices_stride.begin() + 2, 0);
+
+        // YX to ZYX
+        window_lengths.insert(window_lengths.begin(), 1);
+        window_strides.insert(window_strides.begin(), 0);
+        window_dilations.insert(window_dilations.begin(), 0);
+        input_left_pads.insert(input_left_pads.begin(), 0);
+        input_right_pads.insert(input_right_pads.begin(), 0);
+
+        pooling_dims = {2, 3, 4};
+
+        return DevicePool3D::MakeArgumentPointer(p_in_dev,
+                                                 p_out_dev,
+                                                 p_out_indices_dev,
+                                                 input_lengths,
+                                                 window_lengths,
+                                                 output_lengths,
+                                                 input_stride,
+                                                 output_stride,
+                                                 indices_stride,
+                                                 window_strides,
+                                                 window_dilations,
+                                                 input_left_pads,
+                                                 input_right_pads,
+                                                 pooling_dims);
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp b/include/ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp
new file mode 100644
index 0000000000..384805a0a3
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp
@@ -0,0 +1,411 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataType,
+          typename OutDataType,
+          typename IndexDataType, // enable if OutputIndex == true
+          typename ComputeDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          bool OutputIndex,
+          ck::index_t BlockSize,
+          ck::index_t MThreadClusterSize,
+          ck::index_t KThreadClusterSize,
+          ck::index_t MThreadSliceSize,
+          ck::index_t KThreadSliceSize,
+          ck::index_t InSrcOutDstVectorSize>
+struct DevicePool3dFwd_NDHWC_NDHWC : public DevicePoolFwd<5,
+                                                          3,
+                                                          InDataType,
+                                                          OutDataType,
+                                                          IndexDataType,
+                                                          tensor_layout::convolution::NDHWC,
+                                                          tensor_layout::convolution::NDHWC,
+                                                          ReduceOpId,
+                                                          OutputIndex>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    static constexpr index_t InOutRank  = 5;
+    static constexpr index_t WindowRank = 3;
+
+    using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
+
+    using InElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+
+    using AccElementwiseOperation =
+        typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+
+    static constexpr ck::index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr ck::index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto MakeABGridDescriptor_A_M_K_B_M(std::vector<ck::index_t> input_ncdhw_lengths,
+                                               std::vector<ck::index_t> output_ncdhw_lengths,
+                                               std::vector<ck::index_t> input_ncdhw_stride,
+                                               std::vector<ck::index_t> output_ncdhw_stride,
+                                               std::vector<ck::index_t> window_spatial_zyx_lengths,
+                                               std::vector<ck::index_t> window_zyx_strides,
+                                               std::vector<ck::index_t> window_zyx_dilations,
+                                               std::vector<ck::index_t> input_left_dhw_pads,
+                                               std::vector<ck::index_t> input_right_dhw_pads)
+    {
+        const index_t N  = input_ncdhw_lengths[0];
+        const index_t C  = input_ncdhw_lengths[1];
+        const index_t Di = input_ncdhw_lengths[2];
+        const index_t Hi = input_ncdhw_lengths[3];
+        const index_t Wi = input_ncdhw_lengths[4];
+
+        const index_t Do = output_ncdhw_lengths[2];
+        const index_t Ho = output_ncdhw_lengths[3];
+        const index_t Wo = output_ncdhw_lengths[4];
+
+        const index_t Z = window_spatial_zyx_lengths[0];
+        const index_t Y = window_spatial_zyx_lengths[1];
+        const index_t X = window_spatial_zyx_lengths[2];
+
+        const index_t WindowStrideD = window_zyx_strides[0];
+        const index_t WindowStrideH = window_zyx_strides[1];
+        const index_t WindowStrideW = window_zyx_strides[2];
+
+        const index_t WindowDilationD = window_zyx_dilations[0];
+        const index_t WindowDilationH = window_zyx_dilations[1];
+        const index_t WindowDilationW = window_zyx_dilations[2];
+
+        const index_t InLeftPadD = input_left_dhw_pads[0];
+        const index_t InLeftPadH = input_left_dhw_pads[1];
+        const index_t InLeftPadW = input_left_dhw_pads[2];
+
+        const index_t InRightPadD = input_right_dhw_pads[0];
+        const index_t InRightPadH = input_right_dhw_pads[1];
+        const index_t InRightPadW = input_right_dhw_pads[2];
+
+        const index_t MRaw = N * Do * Ho * Wo * C;
+        const index_t MPad = math::integer_least_multiple(MRaw, M_BlockTileSize) - MRaw;
+
+        const index_t KRaw = Z * Y * X;
+        const index_t KPad = math::integer_least_multiple(KRaw, K_BlockTileSize) - KRaw;
+
+        // A[ReduceM, ReduceK]
+        const index_t Ni_stride = input_ncdhw_stride[0];
+        const index_t Ci_stride = input_ncdhw_stride[1];
+        const index_t Di_stride = input_ncdhw_stride[2];
+        const index_t Hi_stride = input_ncdhw_stride[3];
+        const index_t Wi_stride = input_ncdhw_stride[4];
+
+        const auto in_grid_desc_n_di_hi_wi_c = make_naive_tensor_descriptor(
+            make_tuple(N, Di, Hi, Wi, C),
+            make_tuple(Ni_stride, Di_stride, Hi_stride, Wi_stride, Ci_stride));
+
+        const auto in_grid_desc_n_dip_hip_wip_c = transform_tensor_descriptor(
+            in_grid_desc_n_di_hi_wi_c,
+            make_tuple(make_pass_through_transform(N),
+                       make_pad_transform(Di, InLeftPadD, InRightPadD),
+                       make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                       make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+        const auto in_grid_desc_n_z_do_y_ho_x_wo_c = transform_tensor_descriptor(
+            in_grid_desc_n_dip_hip_wip_c,
+            make_tuple(
+                make_pass_through_transform(N),
+                make_embed_transform(make_tuple(Z, Do), make_tuple(WindowDilationD, WindowStrideD)),
+                make_embed_transform(make_tuple(Y, Ho), make_tuple(WindowDilationH, WindowStrideH)),
+                make_embed_transform(make_tuple(X, Wo), make_tuple(WindowDilationW, WindowStrideW)),
+                make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+            make_tuple(Sequence<0>{},
+                       Sequence<1, 2>{},
+                       Sequence<3, 4>{},
+                       Sequence<5, 6>{},
+                       Sequence<7>{}));
+
+        const auto in_grid_desc_reducemraw_reducekraw = transform_tensor_descriptor(
+            in_grid_desc_n_z_do_y_ho_x_wo_c,
+            make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo, C)),
+                       make_merge_transform(make_tuple(Z, Y, X))),
+            make_tuple(Sequence<0, 2, 4, 6, 7>{}, Sequence<1, 3, 5>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto in_grid_desc_reducem_reducek = transform_tensor_descriptor(
+            in_grid_desc_reducemraw_reducekraw,
+            make_tuple(make_right_pad_transform(MRaw, MPad), make_right_pad_transform(KRaw, KPad)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        // B[ReduceM]
+        const index_t No_stride = output_ncdhw_stride[0];
+        const index_t Co_stride = output_ncdhw_stride[1];
+        const index_t Do_stride = output_ncdhw_stride[2];
+        const index_t Ho_stride = output_ncdhw_stride[3];
+        const index_t Wo_stride = output_ncdhw_stride[4];
+
+        const auto out_grid_desc_n_do_ho_wo_c = make_naive_tensor_descriptor(
+            make_tuple(N, Di, Hi, Wi, C),
+            make_tuple(No_stride, Do_stride, Ho_stride, Wo_stride, Co_stride));
+
+        const auto out_grid_desc_reducemraw = transform_tensor_descriptor(
+            out_grid_desc_n_do_ho_wo_c,
+            make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo, C))),
+            make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto out_grid_desc_reducem =
+            transform_tensor_descriptor(out_grid_desc_reducemraw,
+                                        make_tuple(make_right_pad_transform(MRaw, MPad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+
+        return make_tuple(in_grid_desc_reducem_reducek, out_grid_desc_reducem);
+    }
+
+    using ABGridDescs =
+        decltype(MakeABGridDescriptor_A_M_K_B_M({}, {}, {}, {}, {}, {}, {}, {}, {}));
+
+    using AGridDesc_M_K = remove_cvref_t<decltype(ABGridDescs{}[I0])>;
+    using BGridDesc_M   = remove_cvref_t<decltype(ABGridDescs{}[I1])>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_dev,
+                 OutDataType* p_out_dev,
+                 IndexDataType* p_out_indices_dev,
+                 std::vector<ck::index_t>& input_ncdhw_lengths,
+                 std::vector<ck::index_t>& output_ncdhw_lengths,
+                 std::vector<ck::index_t>& input_ncdhw_stride,
+                 std::vector<ck::index_t>& output_ncdhw_stride,
+                 std::vector<ck::index_t>&, // indices_ncdhw_stride
+                 std::vector<ck::index_t>& window_spatial_zyx_lengths,
+                 std::vector<ck::index_t>& window_zyx_strides,
+                 std::vector<ck::index_t>& window_zyx_dilations,
+                 std::vector<ck::index_t>& input_left_dhw_pads,
+                 std::vector<ck::index_t>& input_right_dhw_pads)
+            : p_in_dev_{p_in_dev},
+              p_out_dev_{p_out_dev},
+              p_out_indices_dev_{p_out_indices_dev},
+              a_grid_desc_m_k_{},
+              b_grid_desc_m_{},
+              input_ncdhw_lengths_{input_ncdhw_lengths},
+              output_ncdhw_lengths_{output_ncdhw_lengths},
+              input_ncdhw_stride_{input_ncdhw_stride},
+              output_ncdhw_stride_{output_ncdhw_stride}
+        {
+            const auto descs = MakeABGridDescriptor_A_M_K_B_M(input_ncdhw_lengths,
+                                                              output_ncdhw_lengths,
+                                                              input_ncdhw_stride,
+                                                              output_ncdhw_stride,
+                                                              window_spatial_zyx_lengths,
+                                                              window_zyx_strides,
+                                                              window_zyx_dilations,
+                                                              input_left_dhw_pads,
+                                                              input_right_dhw_pads);
+
+            a_grid_desc_m_k_ = descs[I0];
+            b_grid_desc_m_   = descs[I1];
+
+            int32_t reduceLength = window_spatial_zyx_lengths[0] * window_spatial_zyx_lengths[1] *
+                                   window_spatial_zyx_lengths[2];
+
+            std::tie(in_element_op_, acc_element_op_) =
+                reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(reduceLength);
+        }
+
+        const InDataType* p_in_dev_;
+        OutDataType* p_out_dev_;
+        IndexDataType* p_out_indices_dev_;
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_M b_grid_desc_m_;
+
+        InElementwiseOperation in_element_op_;
+        AccElementwiseOperation acc_element_op_;
+
+        // for checking vector load/store
+        std::vector<ck::index_t> input_ncdhw_lengths_;
+        std::vector<ck::index_t> output_ncdhw_lengths_;
+        std::vector<ck::index_t> input_ncdhw_stride_;
+        std::vector<ck::index_t> output_ncdhw_stride_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            // for NDHWC, the dim C is the fastest dimension, and is not reduced.
+            // Hence, it is in M dimension for reduction kernel.
+            static constexpr index_t InSrcOutDstVectorDim = 0; // 0: M, 1: K
+
+            using gridwise_reduce =
+                GridwiseReduction_mk_to_m_threadwise<InDataType,
+                                                     OutDataType,
+                                                     ComputeDataType,
+                                                     IndexDataType,
+                                                     AGridDesc_M_K,
+                                                     BGridDesc_M,
+                                                     ReduceOperation,
+                                                     InElementwiseOperation,
+                                                     AccElementwiseOperation,
+                                                     InMemoryDataOperationEnum::Set,
+                                                     false, // propagate_nan
+                                                     BlockSize,
+                                                     MThreadSliceSize,
+                                                     KThreadSliceSize,
+                                                     InSrcOutDstVectorDim,
+                                                     InSrcOutDstVectorSize,
+                                                     InSrcOutDstVectorSize>;
+
+            const auto kernel =
+                kernel_reduce_threadwise<gridwise_reduce,
+                                         OutputIndex,
+                                         true,  // pooling need to return global index
+                                         false, // don't have index input
+                                         InDataType,
+                                         OutDataType,
+                                         ComputeDataType,
+                                         IndexDataType,
+                                         AGridDesc_M_K,
+                                         BGridDesc_M,
+                                         InElementwiseOperation,
+                                         AccElementwiseOperation>;
+
+            ck::index_t M = arg.a_grid_desc_m_k_.GetLength(I0);
+
+            const index_t grid_size = (M / M_BlockTileSize);
+
+            return launch_and_time_kernel(stream_config,
+                                          kernel,
+                                          dim3(grid_size),
+                                          dim3(BlockSize),
+                                          0,
+                                          arg.a_grid_desc_m_k_,
+                                          arg.b_grid_desc_m_,
+                                          arg.in_element_op_,
+                                          arg.acc_element_op_,
+                                          float(1),
+                                          arg.p_in_dev_,
+                                          nullptr,
+                                          float(0),
+                                          arg.p_out_dev_,
+                                          arg.p_out_indices_dev_);
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        // C should be fastest dimension
+        if(pArg->input_ncdhw_stride_[1] != 1)
+            return false;
+
+        for(int i = 0; i < InOutRank; ++i)
+        {
+            if(pArg->input_ncdhw_stride_[i] == 1 &&
+               pArg->input_ncdhw_lengths_[i] % InSrcOutDstVectorSize != 0)
+                return false;
+
+            if(pArg->output_ncdhw_stride_[i] == 1 &&
+               pArg->output_ncdhw_lengths_[i] % InSrcOutDstVectorSize != 0)
+                return false;
+        }
+
+        return true;
+    }
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_dev,
+                        void* p_out_dev,
+                        void* p_out_indices_dev,
+                        std::vector<ck::index_t> input_ncdhw_lengths,
+                        std::vector<ck::index_t> window_zyx_lengths,
+                        std::vector<ck::index_t> output_ncdhw_lengths,
+                        std::vector<ck::index_t> input_ncdhw_stride,
+                        std::vector<ck::index_t> output_ncdhw_stride,
+                        std::vector<ck::index_t> indices_ncdhw_stride,
+                        std::vector<ck::index_t> window_zyx_strides,
+                        std::vector<ck::index_t> window_zyx_dilations,
+                        std::vector<ck::index_t> input_left_dhw_pads,
+                        std::vector<ck::index_t> input_right_dhw_pads,
+                        std::vector<ck::index_t> pooling_dims) override
+    {
+        if(input_ncdhw_lengths.size() != InOutRank || window_zyx_lengths.size() != WindowRank ||
+           input_ncdhw_lengths.size() != InOutRank || window_zyx_strides.size() != WindowRank ||
+           window_zyx_dilations.size() != WindowRank || input_left_dhw_pads.size() != WindowRank ||
+           input_right_dhw_pads.size() != WindowRank)
+            throw std::runtime_error("dimension is incorrect");
+
+        if(pooling_dims != std::vector<ck::index_t>{2, 3, 4})
+            throw std::runtime_error("pooling_dims only support {2, 3, 4} in pool3d so far");
+
+        if(output_ncdhw_stride != indices_ncdhw_stride)
+            throw std::runtime_error(
+                "output_ncdhw_stride need to be equal to indices_ncdhw_stride for now");
+
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_dev),
+                                          static_cast<OutDataType*>(p_out_dev),
+                                          static_cast<IndexDataType*>(p_out_indices_dev),
+                                          input_ncdhw_lengths,
+                                          output_ncdhw_lengths,
+                                          input_ncdhw_stride,
+                                          output_ncdhw_stride,
+                                          indices_ncdhw_stride,
+                                          window_zyx_lengths,
+                                          window_zyx_strides,
+                                          window_zyx_dilations,
+                                          input_left_dhw_pads,
+                                          input_right_dhw_pads);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DevicePool3dFwd_NDHWC_NDHWC<" << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str <<"InSrcOutDstVectorSize_" << InSrcOutDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_put_element_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_put_element_impl.hpp
new file mode 100644
index 0000000000..7334da0e35
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_put_element_impl.hpp
@@ -0,0 +1,155 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/device_put_element.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/stream_utility.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// output[indices] = input
+template <typename InDataType,
+          typename IndexDataType,
+          typename OutDataType,
+          typename ElementwiseOperation,
+          InMemoryDataOperationEnum MemOp,
+          ck::index_t InVectorSize>
+struct DevicePutElementImpl
+    : public DevicePutElement<InDataType, IndexDataType, OutDataType, ElementwiseOperation, MemOp>
+{
+    template <typename Desc_M>
+    static auto PadDescriptor_M_1d(Desc_M desc_m, index_t gridSize, index_t blockSize)
+    {
+        constexpr auto I0 = Number<0>{};
+
+        const auto m            = desc_m.GetLength(I0);
+        const index_t loop_step = gridSize * blockSize * InVectorSize;
+        const auto pad          = math::integer_least_multiple(m, loop_step) - m;
+        const auto desc_m_pad =
+            transform_tensor_descriptor(desc_m,
+                                        make_tuple(make_right_pad_transform(m, pad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return desc_m_pad;
+    }
+
+    static auto MakeDescriptor_M(index_t length, index_t gridSize, index_t blockSize)
+    {
+        const auto desc_m = make_naive_tensor_descriptor_packed(make_tuple(length));
+        return PadDescriptor_M_1d(desc_m, gridSize, blockSize);
+    }
+
+    using InGrid1dDesc = decltype(MakeDescriptor_M(1, 1, 1));
+
+    using GridwisePutElement = GridwisePutElement_1D<InGrid1dDesc,
+                                                     InDataType,
+                                                     IndexDataType,
+                                                     OutDataType,
+                                                     ElementwiseOperation,
+                                                     MemOp,
+                                                     InVectorSize>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_input,
+                 const IndexDataType* p_indices,
+                 OutDataType* p_output,
+                 index_t input_length,
+                 ElementwiseOperation elementwise_op)
+            : p_input_{p_input},
+              p_indices_{p_indices},
+              p_output_{p_output},
+              input_length_raw_{input_length},
+              elementwise_op_{elementwise_op},
+              blockSize_{256}
+        {
+        }
+
+        const InDataType* p_input_;
+        const IndexDataType* p_indices_;
+        OutDataType* p_output_;
+        index_t input_length_raw_;
+        ElementwiseOperation elementwise_op_;
+        index_t blockSize_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            index_t gridSize = getAvailableComputeUnitCount(stream_config);
+            InGrid1dDesc in_grid_desc =
+                MakeDescriptor_M(arg.input_length_raw_, gridSize, arg.blockSize_);
+
+            const auto kernel = kernel_put_element_1d<GridwisePutElement,
+                                                      InGrid1dDesc,
+                                                      InDataType,
+                                                      IndexDataType,
+                                                      OutDataType,
+                                                      ElementwiseOperation>;
+
+            float elapsed_time = launch_and_time_kernel(stream_config,
+                                                        kernel,
+                                                        dim3(gridSize),
+                                                        dim3(arg.blockSize_),
+                                                        0,
+                                                        in_grid_desc,
+                                                        arg.p_input_,
+                                                        arg.p_indices_,
+                                                        arg.p_output_,
+                                                        arg.elementwise_op_);
+            return elapsed_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        if(pArg->input_length_raw_ % InVectorSize != 0)
+        {
+            return false;
+        }
+        return true;
+    }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_input,
+                                                      const void* p_indices,
+                                                      void* p_output,
+                                                      index_t input_length,
+                                                      index_t,
+                                                      ElementwiseOperation elementwise_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const InDataType*>(p_input),
+                                          static_cast<const IndexDataType*>(p_indices),
+                                          static_cast<OutDataType*>(p_output),
+                                          input_length,
+                                          elementwise_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp b/include/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp
new file mode 100644
index 0000000000..2481c5c769
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp
@@ -0,0 +1,142 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <cassert>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/utility/reduction_operator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// here, inLengths[] is already shuffled so that lengths of invariant dims are included before those
+// of reduce dims
+template <index_t Rank, int NumReduceDim>
+std::pair<long_index_t, long_index_t> get_2d_lengths(const std::vector<index_t>& inLengths)
+{
+    static_assert(Rank <= 6, "bigger Rank size not supported!");
+
+    long_index_t invariant_total_length = 1;
+    long_index_t reduce_total_length    = 1;
+
+    constexpr int NumInvariantDim = Rank - NumReduceDim;
+
+    for(int i = NumInvariantDim; i < Rank; i++)
+        reduce_total_length *= inLengths[i];
+
+    for(int i = 0; i < NumInvariantDim; i++)
+        invariant_total_length *= inLengths[i];
+
+    return std::make_pair(invariant_total_length, reduce_total_length);
+};
+
+template <index_t Rank, int NumReduceDim>
+std::pair<long_index_t, long_index_t> get_2d_lengths(const std::array<index_t, Rank>& inLengths)
+{
+    static_assert(Rank <= 6, "bigger Rank size not supported!");
+
+    long_index_t invariant_total_length = 1;
+    long_index_t reduce_total_length    = 1;
+
+    constexpr int NumInvariantDim = Rank - NumReduceDim;
+
+    for(int i = NumInvariantDim; i < Rank; i++)
+        reduce_total_length *= inLengths[i];
+
+    for(int i = 0; i < NumInvariantDim; i++)
+        invariant_total_length *= inLengths[i];
+
+    return std::make_pair(invariant_total_length, reduce_total_length);
+};
+
+// helper functions using variadic template arguments
+template <index_t... Ns>
+auto make_tuple_from_array_and_index_seq(const std::vector<index_t>& lengths, Sequence<Ns...>)
+{
+    return make_tuple(static_cast<index_t>(lengths[Ns])...);
+};
+
+template <index_t arraySize>
+auto make_tuple_from_array(const std::vector<index_t>& lengths, Number<arraySize>)
+{
+    static_assert(arraySize >= 1 && arraySize <= 6, "The tensor should have 1 to 6 dimensions");
+
+    constexpr auto index_seq = typename arithmetic_sequence_gen<0, arraySize, 1>::type{};
+
+    return make_tuple_from_array_and_index_seq(lengths, index_seq);
+};
+
+template <index_t Rank, index_t NumReduceDim>
+std::vector<index_t> shuffle_tensor_dimensions(const std::vector<index_t>& origLengthsStrides,
+                                               const std::vector<int>& reduceDims)
+{
+    std::vector<index_t> newLengthsStrides;
+
+    assert(Rank == origLengthsStrides.size() && NumReduceDim == reduceDims.size());
+
+    int reduceFlag = 0;
+
+    // flag the bits for the reduceDims
+    for(int i = 0; i < NumReduceDim; i++)
+    {
+        reduceFlag |= 1 << reduceDims[i];
+    };
+
+    // collect invariant dimensions
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) == 0)
+        {
+            newLengthsStrides.push_back(origLengthsStrides[i]);
+        };
+
+    // collect reduce dimensions
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) > 0)
+        {
+            newLengthsStrides.push_back(origLengthsStrides[i]);
+        };
+
+    return newLengthsStrides;
+};
+
+template <index_t Rank, index_t NumReduceDim>
+std::array<index_t, Rank>
+shuffle_tensor_dimensions(const std::array<index_t, Rank>& origLengthsStrides,
+                          const std::array<int, NumReduceDim>& reduceDims)
+{
+    std::array<index_t, Rank> newLengthsStrides;
+
+    int reduceFlag = 0;
+
+    // flag the bits for the reduceDims
+    for(int i = 0; i < NumReduceDim; i++)
+    {
+        reduceFlag |= 1 << reduceDims[i];
+    };
+
+    // collect invariant dimensions
+    int pos = 0;
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) == 0)
+        {
+            newLengthsStrides[pos++] = origLengthsStrides[i];
+        };
+
+    // collect reduce dimensions
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) > 0)
+        {
+            newLengthsStrides[pos++] = origLengthsStrides[i];
+        };
+
+    return newLengthsStrides;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp b/include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp
new file mode 100644
index 0000000000..bf3deeb57a
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp
@@ -0,0 +1,551 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <array>
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation,
+          InMemoryDataOperationEnum OutMemoryDataOperation,
+          bool PropagateNan,
+          bool OutputIndex,
+          bool HaveIndexInputIfOutputIndex,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          index_t OutDstVectorSize>
+struct DeviceReduceMultiBlock : public DeviceReduce<InDataType,
+                                                    AccDataType,
+                                                    OutDataType,
+                                                    Rank,
+                                                    NumReduceDim,
+                                                    ReduceOperation,
+                                                    InElementwiseOperation,
+                                                    AccElementwiseOperation,
+                                                    PropagateNan,
+                                                    OutputIndex>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
+                  "Invalid thread cluster size assignments!");
+
+    static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                   (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)) &&
+                      (MThreadSliceSize % OutDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    using IndexDataType = int32_t;
+
+    static constexpr bool HaveIndexInput = OutputIndex && HaveIndexInputIfOutputIndex;
+
+    static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+    static constexpr index_t NumSrcDim = Rank;
+    static constexpr index_t NumDstDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
+    static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+    // So far, only AtomicAdd is considered, other Atomic Operation like AtomicMax can be added
+    // later
+    static constexpr bool use_multiblock =
+        (OutMemoryDataOperation == InMemoryDataOperationEnum::AtomicAdd);
+
+    static_assert(ck::reduce::InMemoryDataOperationSupportedOnDataType<OutMemoryDataOperation,
+                                                                       OutDataType>::value,
+                  "The OutDataType must support the specified OutMemoryDataOperation!");
+
+    static_assert(!use_multiblock || (use_multiblock && !OutputIndex),
+                  "MultiBlock reduction can only be used when outputing index is not required");
+
+    static_assert(
+        ReduceOperation::IsCompatibleInMemoryDataOperation(OutMemoryDataOperation),
+        "The reduction accumulation operation must be compatible with the OutMemoryDataOperation!");
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto MakeSrc2dDescriptor(const std::array<index_t, Rank>& inLengths,
+                                    const std::array<index_t, Rank>& inStrides,
+                                    int blkGroupSize,
+                                    int numBlockTileIteration)
+    {
+        const auto tupleSrcLengths =
+            generate_tuple([&](auto I) { return inLengths[I]; }, Number<Rank>{});
+        const auto tupleSrcStrides =
+            generate_tuple([&](auto I) { return inStrides[I]; }, Number<Rank>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, NumSrcDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths = generate_tuple(
+                    [&](auto I) { return inLengths[NumInvariantDim + I]; }, Number<NumReduceDim>{});
+                const auto invariantDimLengths =
+                    generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInvariantDim>{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    static auto MakeDst1dDescriptor(const std::array<index_t, NumDstDim>& outLengths,
+                                    const std::array<index_t, NumDstDim>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumDstDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumDstDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumDstDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto outPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto out_grid_desc_m_padded = transform_tensor_descriptor(
+            out_grid_desc_m,
+            make_tuple(make_right_pad_transform(invariantLength, outPad)),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    static auto MakeDst1dDescriptorForBufferSet(const std::array<index_t, NumDstDim>& outLengths,
+                                                const std::array<index_t, NumDstDim>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumDstDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumDstDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumDstDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto length = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto pad = math::integer_least_multiple(length, BlockSize) - length;
+
+        auto out_grid_desc_m_padded =
+            transform_tensor_descriptor(out_grid_desc_m,
+                                        make_tuple(make_right_pad_transform(length, pad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> inLengths,
+                 const std::array<index_t, Rank> inStrides,
+                 const std::array<index_t, NumDstDim> outLengths,
+                 const std::array<index_t, NumDstDim> outStrides,
+                 const std::array<int, NumReduceDim> reduceDims,
+                 double alpha,
+                 double beta,
+                 const InDataType* in_dev,
+                 const IndexDataType* in_index_dev,
+                 OutDataType* out_dev,
+                 IndexDataType* out_index_dev,
+                 const InElementwiseOperation in_elementwise_op,
+                 const AccElementwiseOperation acc_elementwise_op)
+            : outLengths_{outLengths},
+              outStrides_{outStrides},
+              in_dev_{in_dev},
+              in_index_dev_{in_index_dev},
+              out_dev_{out_dev},
+              out_index_dev_{out_index_dev},
+              in_elementwise_op_{in_elementwise_op},
+              acc_elementwise_op_{acc_elementwise_op}
+        {
+            if(Rank != inLengths.size() || Rank != inStrides.size() ||
+               NumReduceDim != reduceDims.size())
+            {
+                throw std::runtime_error(
+                    "One of inLengths/inStrides/reduceDims has invalid size!"
+                    "\nExpected size inLengths: " +
+                    std::to_string(Rank) + ", inStrides: " + std::to_string(Rank) +
+                    ", reduceDims: " + std::to_string(NumReduceDim) +
+                    "\nBut have inLengths: " + std::to_string(inLengths.size()) +
+                    ", inStrides: " + std::to_string(inStrides.size()) +
+                    ", reduceDims: " + std::to_string(reduceDims.size()));
+            }
+
+            for(std::size_t i = 0; i < reduceDims.size(); ++i)
+            {
+                if(reduceDims[i] < 0 || reduceDims[i] >= Rank)
+                {
+                    throw std::runtime_error("Provided reduce dimension exceed input tensor Rank!"
+                                             "\nHave reduceDims[" +
+                                             std::to_string(i) +
+                                             "]: " + std::to_string(reduceDims[i]));
+                }
+            }
+
+            inLengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
+            inStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inStrides, reduceDims);
+
+            alpha_ = type_convert<AccDataType>(alpha);
+            beta_  = type_convert<AccDataType>(beta);
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(inLengths_);
+
+            if constexpr(NumInvariantDim == 0)
+                invariant_lowest_length = 1;
+            else
+                invariant_lowest_length = inLengths_[NumInvariantDim - 1];
+
+            reduce_lowest_length = inLengths_[Rank - 1];
+
+            if constexpr(use_multiblock)
+            {
+
+                int iterations = 1;
+                while(true)
+                {
+                    int testBlkGroupSize =
+                        (reduce_total_length + (K_BlockTileSize * iterations) - 1) /
+                        (K_BlockTileSize * iterations);
+
+                    // we want the blkGroupSize be not more than 128
+                    if(testBlkGroupSize <= 128)
+                        break;
+
+                    iterations++;
+                };
+
+                blkGroupSize = (reduce_total_length + (K_BlockTileSize * iterations) - 1) /
+                               (K_BlockTileSize * iterations);
+
+                numBlockTileIteration = iterations;
+            }
+            else
+            {
+                blkGroupSize = 1;
+                numBlockTileIteration =
+                    (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
+            };
+
+            gridSize = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                       M_BlockTileSize * blkGroupSize;
+
+            gridSize_pre =
+                math::integer_least_multiple(invariant_total_length, BlockSize) / BlockSize;
+        }
+
+        std::array<index_t, Rank> inLengths_;
+        std::array<index_t, Rank> inStrides_;
+        std::array<index_t, NumDstDim> outLengths_;
+        std::array<index_t, NumDstDim> outStrides_;
+
+        AccDataType alpha_;
+        AccDataType beta_;
+
+        const InDataType* in_dev_;
+        const IndexDataType* in_index_dev_;
+        OutDataType* out_dev_;
+        IndexDataType* out_index_dev_;
+
+        InElementwiseOperation in_elementwise_op_;
+        AccElementwiseOperation acc_elementwise_op_;
+
+        index_t invariant_lowest_length;
+        index_t reduce_lowest_length;
+        long_index_t invariant_total_length;
+        long_index_t reduce_total_length;
+
+        int blkGroupSize;
+        int numBlockTileIteration;
+        size_t gridSize;
+
+        size_t gridSize_pre;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto in_grid_desc_m_k = DeviceReduceMultiBlock::MakeSrc2dDescriptor(
+                arg.inLengths_, arg.inStrides_, arg.blkGroupSize, arg.numBlockTileIteration);
+            const auto out_grid_desc_m =
+                DeviceReduceMultiBlock::MakeDst1dDescriptor(arg.outLengths_, arg.outStrides_);
+            const auto out_grid_desc_m_2 = DeviceReduceMultiBlock::MakeDst1dDescriptorForBufferSet(
+                arg.outLengths_, arg.outStrides_);
+
+            using InGridDesc_M_K  = decltype(in_grid_desc_m_k);
+            using OutGridDesc_M   = decltype(out_grid_desc_m);
+            using OutGridDesc_M_2 = decltype(out_grid_desc_m_2);
+
+            using GridwiseReduce = GridwiseReduction_mk_to_m_multiblock<InDataType,
+                                                                        OutDataType,
+                                                                        AccDataType,
+                                                                        IndexDataType,
+                                                                        InGridDesc_M_K,
+                                                                        OutGridDesc_M,
+                                                                        ReduceOperation,
+                                                                        InElementwiseOperation,
+                                                                        AccElementwiseOperation,
+                                                                        OutMemoryDataOperation,
+                                                                        PropagateNan,
+                                                                        BlockSize,
+                                                                        MThreadClusterSize,
+                                                                        KThreadClusterSize,
+                                                                        MThreadSliceSize,
+                                                                        KThreadSliceSize,
+                                                                        InSrcVectorDim,
+                                                                        InSrcVectorSize,
+                                                                        OutDstVectorSize>;
+
+            const auto kernel_main = kernel_reduce_multiblock<GridwiseReduce,
+                                                              OutputIndex,
+                                                              HaveIndexInput,
+                                                              InDataType,
+                                                              OutDataType,
+                                                              AccDataType,
+                                                              int32_t,
+                                                              InGridDesc_M_K,
+                                                              OutGridDesc_M,
+                                                              InElementwiseOperation,
+                                                              AccElementwiseOperation>;
+
+            float avg_time = 0;
+
+            if constexpr(use_multiblock)
+            {
+                const auto identityVal =
+                    ck::reduce::GetIdentityValueForInMemoryDataOperation<OutDataType>(
+                        OutMemoryDataOperation);
+
+                const auto kernel_pre =
+                    kernel_buffer_set_value<BlockSize, OutDataType, OutGridDesc_M_2>;
+
+                avg_time += launch_and_time_kernel(stream_config,
+                                                   kernel_pre,
+                                                   dim3(arg.gridSize_pre),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   out_grid_desc_m_2,
+                                                   arg.out_dev_,
+                                                   identityVal);
+            };
+
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize),
+                                               dim3(BlockSize),
+                                               0,
+                                               in_grid_desc_m_k,
+                                               out_grid_desc_m,
+                                               arg.in_elementwise_op_,
+                                               arg.acc_elementwise_op_,
+                                               arg.blkGroupSize,
+                                               arg.numBlockTileIteration,
+                                               arg.alpha_,
+                                               arg.in_dev_,
+                                               arg.in_index_dev_,
+                                               arg.beta_,
+                                               arg.out_dev_,
+                                               arg.out_index_dev_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    static bool IsSupportedArgument(const Argument* pArg)
+    {
+        if constexpr(use_multiblock)
+        {
+            if(static_cast<float>(pArg->beta_) != 0.0f)
+                return (false);
+        };
+
+        if constexpr(InSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return (false);
+            }
+            else
+            {
+                if(pArg->inStrides_[NumInvariantDim - 1] != 1)
+                    return (false);
+
+                if(pArg->invariant_lowest_length % InSrcVectorSize != 0)
+                    return (false);
+            };
+        }
+        else
+        {
+            if(pArg->inStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(pArg->reduce_lowest_length % InSrcVectorSize != 0)
+                return (false);
+        };
+
+        // To improve
+        if(pArg->invariant_lowest_length % OutDstVectorSize != 0)
+            return (false);
+
+        if constexpr(use_multiblock)
+        {
+            // blkGroupSize of 1 should be handled by Blockwise path using
+            // InMemoryDataOperationEnum::Set
+            if(pArg->blkGroupSize == 1)
+                return (false);
+
+            // This is very strong restriction, but needed to avoid some failure
+            if(pArg->invariant_lowest_length % M_BlockTileSize != 0)
+                return (false);
+        }
+        else
+        {
+            // cases with very small reduce_total_length should be handled by ThreadWise kernel
+            // if(pArg->reduce_total_length / KThreadSliceSize < 2)
+            //     return (false);
+        };
+
+        return (true);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(dynamic_cast<const Argument*>(p_arg));
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> inLengths,
+                        const std::array<index_t, Rank> inStrides,
+                        const std::array<index_t, NumDstDim> outLengths,
+                        const std::array<index_t, NumDstDim> outStrides,
+                        const std::array<int, NumReduceDim> reduceDims,
+                        double alpha,
+                        double beta,
+                        const void* in_dev,
+                        const void* in_index_dev,
+                        void* out_dev,
+                        void* out_index_dev,
+                        const InElementwiseOperation in_elementwise_op,
+                        const AccElementwiseOperation acc_elementwise_op) override
+    {
+        return std::make_unique<Argument>(inLengths,
+                                          inStrides,
+                                          outLengths,
+                                          outStrides,
+                                          reduceDims,
+                                          alpha,
+                                          beta,
+                                          static_cast<const InDataType*>(in_dev),
+                                          static_cast<const IndexDataType*>(in_index_dev),
+                                          static_cast<OutDataType*>(out_dev),
+                                          static_cast<IndexDataType*>(out_index_dev),
+                                          in_elementwise_op,
+                                          acc_elementwise_op);
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << (OutMemoryDataOperation == InMemoryDataOperationEnum::Set? "DeviceReduceBlockWise<" : "DeviceReduceMultiBlock<") << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str << "InSrcVectorDim_" << InSrcVectorDim << "_InSrcVectorSize_" << InSrcVectorSize << "_OutDstVectorSize_" << OutDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp b/include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp
new file mode 100644
index 0000000000..6c5895b010
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp
@@ -0,0 +1,393 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <array>
+
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation,
+          bool PropagateNan,
+          bool OutputIndex,
+          bool TransformIndexKtoGlobal,
+          bool HaveIndexInputIfOutputIndex,
+          index_t BlockSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          index_t OutDstVectorSize>
+struct DeviceReduceThreadWise : public DeviceReduce<InDataType,
+                                                    AccDataType,
+                                                    OutDataType,
+                                                    Rank,
+                                                    NumReduceDim,
+                                                    ReduceOperation,
+                                                    InElementwiseOperation,
+                                                    AccElementwiseOperation,
+                                                    PropagateNan,
+                                                    OutputIndex>
+
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+
+    static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                   (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)) &&
+                      (MThreadSliceSize % OutDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    using IndexDataType = int32_t;
+
+    static constexpr bool HaveIndexInput = OutputIndex && HaveIndexInputIfOutputIndex;
+
+    static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+    static constexpr index_t NumSrcDim = Rank;
+    static constexpr index_t NumDstDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
+    static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+    static constexpr index_t M_BlockTileSize = BlockSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = 1 * KThreadSliceSize;
+
+    static auto MakeSrc2dDescriptor(const std::array<index_t, Rank>& inLengths,
+                                    const std::array<index_t, Rank>& inStrides)
+    {
+        const auto tupleSrcLengths =
+            generate_tuple([&](auto I) { return inLengths[I]; }, Number<Rank>{});
+        const auto tupleSrcStrides =
+            generate_tuple([&](auto I) { return inStrides[I]; }, Number<Rank>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, NumSrcDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths = generate_tuple(
+                    [&](auto I) { return inLengths[NumInvariantDim + I]; }, Number<NumReduceDim>{});
+                const auto invariantDimLengths =
+                    generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInvariantDim>{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K =
+            math::integer_least_multiple(reduceLength, K_BlockTileSize) - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    static auto MakeDst1dDescriptor(const std::array<index_t, NumDstDim>& outLengths,
+                                    const std::array<index_t, NumDstDim>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumDstDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumDstDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumDstDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto outPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto out_grid_desc_m_padded = transform_tensor_descriptor(
+            out_grid_desc_m,
+            make_tuple(make_right_pad_transform(invariantLength, outPad)),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> inLengths,
+                 const std::array<index_t, Rank> inStrides,
+                 const std::array<index_t, NumDstDim> outLengths,
+                 const std::array<index_t, NumDstDim> outStrides,
+                 const std::array<int, NumReduceDim> reduceDims,
+                 double alpha,
+                 double beta,
+                 const InDataType* in_dev,
+                 OutDataType* out_dev,
+                 IndexDataType* out_index_dev,
+                 const InElementwiseOperation in_elementwise_op,
+                 const AccElementwiseOperation acc_elementwise_op)
+            : outLengths_{outLengths},
+              outStrides_{outStrides},
+              in_dev_{in_dev},
+              out_dev_{out_dev},
+              out_index_dev_{out_index_dev},
+              in_elementwise_op_{in_elementwise_op},
+              acc_elementwise_op_{acc_elementwise_op}
+        {
+            inLengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
+            inStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inStrides, reduceDims);
+
+            alpha_ = type_convert<AccDataType>(alpha);
+            beta_  = type_convert<AccDataType>(beta);
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(inLengths_);
+
+            if constexpr(NumInvariantDim == 0)
+                invariant_lowest_length = 1;
+            else
+                invariant_lowest_length = inLengths_[NumInvariantDim - 1];
+
+            reduce_lowest_length = inLengths_[Rank - 1];
+
+            numBlockTileIteration = (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
+
+            gridSize = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                       M_BlockTileSize;
+        }
+
+        std::array<index_t, Rank> inLengths_;
+        std::array<index_t, Rank> inStrides_;
+        std::array<index_t, NumDstDim> outLengths_;
+        std::array<index_t, NumDstDim> outStrides_;
+
+        AccDataType alpha_;
+        AccDataType beta_;
+
+        const InDataType* in_dev_;
+        OutDataType* out_dev_;
+        IndexDataType* out_index_dev_;
+
+        InElementwiseOperation in_elementwise_op_;
+        AccElementwiseOperation acc_elementwise_op_;
+
+        index_t invariant_lowest_length;
+        index_t reduce_lowest_length;
+        long_index_t invariant_total_length;
+        long_index_t reduce_total_length;
+
+        int numBlockTileIteration;
+        size_t gridSize;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto in_grid_desc_m_k =
+                DeviceReduceThreadWise::MakeSrc2dDescriptor(arg.inLengths_, arg.inStrides_);
+            const auto out_grid_desc_m =
+                DeviceReduceThreadWise::MakeDst1dDescriptor(arg.outLengths_, arg.outStrides_);
+            using InGridDesc_M_K = decltype(in_grid_desc_m_k);
+            using OutGridDesc_M  = decltype(out_grid_desc_m);
+
+            float avg_time = 0;
+
+            using GridwiseReduce =
+                GridwiseReduction_mk_to_m_threadwise<InDataType,
+                                                     OutDataType,
+                                                     AccDataType,
+                                                     IndexDataType,
+                                                     InGridDesc_M_K,
+                                                     OutGridDesc_M,
+                                                     ReduceOperation,
+                                                     InElementwiseOperation,
+                                                     AccElementwiseOperation,
+                                                     InMemoryDataOperationEnum::Set,
+                                                     PropagateNan,
+                                                     BlockSize,
+                                                     MThreadSliceSize,
+                                                     KThreadSliceSize,
+                                                     InSrcVectorDim,
+                                                     InSrcVectorSize,
+                                                     OutDstVectorSize>;
+
+            const auto kernel = kernel_reduce_threadwise<GridwiseReduce,
+                                                         OutputIndex,
+                                                         TransformIndexKtoGlobal,
+                                                         HaveIndexInput,
+                                                         InDataType,
+                                                         OutDataType,
+                                                         AccDataType,
+                                                         IndexDataType,
+                                                         InGridDesc_M_K,
+                                                         OutGridDesc_M,
+                                                         InElementwiseOperation,
+                                                         AccElementwiseOperation>;
+
+            avg_time = launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(arg.gridSize),
+                                              dim3(BlockSize),
+                                              0,
+                                              in_grid_desc_m_k,
+                                              out_grid_desc_m,
+                                              arg.in_elementwise_op_,
+                                              arg.acc_elementwise_op_,
+                                              arg.alpha_,
+                                              arg.in_dev_,
+                                              nullptr,
+                                              arg.beta_,
+                                              arg.out_dev_,
+                                              arg.out_index_dev_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
+
+        if constexpr(InSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return (false);
+            }
+            else
+            {
+                if(pArg->inStrides_[NumInvariantDim - 1] != 1)
+                    return (false);
+
+                if(pArg->invariant_lowest_length % InSrcVectorSize != 0)
+                    return (false);
+            };
+        }
+        else
+        {
+            if(pArg->inStrides_[Rank - 1] != 1)
+                return (false);
+
+            if(pArg->reduce_lowest_length % InSrcVectorSize != 0)
+                return (false);
+        };
+
+        // To improve
+        if(pArg->invariant_lowest_length % OutDstVectorSize != 0)
+            return (false);
+
+        // cases with big reduce_total_length should be handled by Blockwise kernel
+        if(pArg->reduce_total_length / KThreadSliceSize >= 32)
+            return (false);
+
+        return (true);
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> inLengths,
+                        const std::array<index_t, Rank> inStrides,
+                        const std::array<index_t, NumDstDim> outLengths,
+                        const std::array<index_t, NumDstDim> outStrides,
+                        const std::array<int, NumReduceDim> reduceDims,
+                        double alpha,
+                        double beta,
+                        const void* in_dev,
+                        const void* in_index_dev,
+                        void* out_dev,
+                        void* out_index_dev,
+                        const InElementwiseOperation in_elementwise_op,
+                        const AccElementwiseOperation acc_elementwise_op) override
+    {
+        (void)in_index_dev;
+
+        return std::make_unique<Argument>(inLengths,
+                                          inStrides,
+                                          outLengths,
+                                          outStrides,
+                                          reduceDims,
+                                          alpha,
+                                          beta,
+                                          static_cast<const InDataType*>(in_dev),
+                                          static_cast<OutDataType*>(out_dev),
+                                          static_cast<IndexDataType*>(out_index_dev),
+                                          in_elementwise_op,
+                                          acc_elementwise_op);
+    };
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceReduceThreadWise<" << BlockSize << ",";
+        str << "M_C" << BlockSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << 1 << "_S" << KThreadSliceSize << ",";
+        str << "InSrcVectorDim_" << InSrcVectorDim << "_InSrcVectorSize_" << InSrcVectorSize << "_OutDstVectorSize_" << OutDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp
new file mode 100644
index 0000000000..8eff9d2415
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp
@@ -0,0 +1,417 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_softmax.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          typename InElementwiseOp,
+          typename AccElementwiseOp,
+          index_t Rank,
+          index_t NumReduceDim,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          index_t OutDstVectorSize>
+struct DeviceSoftmaxImpl : public DeviceSoftmax<InDataType,
+                                                AccDataType,
+                                                OutDataType,
+                                                InElementwiseOp,
+                                                AccElementwiseOp,
+                                                Rank,
+                                                NumReduceDim>
+{
+    static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+
+    static constexpr index_t NumSrcDim = Rank;
+    static constexpr index_t NumDstDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
+    static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
+                                    const std::vector<index_t>& inStrides,
+                                    int blkGroupSize,
+                                    int numBlockTileIteration)
+    {
+        const auto tupleSrcLengths =
+            generate_tuple([&](auto I) { return inLengths[I]; }, Number<Rank>{});
+        const auto tupleSrcStrides =
+            generate_tuple([&](auto I) { return inStrides[I]; }, Number<Rank>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            if constexpr(reduceAllDim)
+            {
+                const auto one_dim_inDesc = transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(tupleSrcLengths)),
+                    make_tuple(typename arithmetic_sequence_gen<0, NumSrcDim, 1>::type{}),
+                    make_tuple(Sequence<0>{}));
+
+                return transform_tensor_descriptor(one_dim_inDesc,
+                                                   make_tuple(make_unmerge_transform(make_tuple(
+                                                       1, one_dim_inDesc.GetLength(Number<0>{})))),
+                                                   make_tuple(Sequence<0>{}),
+                                                   make_tuple(Sequence<0, 1>{}));
+            }
+            else
+            {
+                using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+                using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+                const auto reduceDimLengths = generate_tuple(
+                    [&](auto I) { return inLengths[NumInvariantDim + I]; }, Number<NumReduceDim>{});
+                const auto invariantDimLengths =
+                    generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInvariantDim>{});
+
+                return transform_tensor_descriptor(
+                    inDesc,
+                    make_tuple(make_merge_transform(invariantDimLengths),
+                               make_merge_transform(reduceDimLengths)),
+                    make_tuple(InvariantDims{}, ReduceDims{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+            }
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return (in_grid_desc_m_k_padded);
+    };
+
+    using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1));
+
+    using GridwiseSoftmaxGeneric = GridwiseSoftmax_mk_to_mk<InDataType,
+                                                            OutDataType,
+                                                            AccDataType,
+                                                            GridDesc_M_K,
+                                                            BlockSize,
+                                                            MThreadClusterSize,
+                                                            KThreadClusterSize,
+                                                            MThreadSliceSize,
+                                                            KThreadSliceSize,
+                                                            InSrcVectorDim,
+                                                            InSrcVectorSize,
+                                                            OutDstVectorSize,
+                                                            false>;
+
+    using GridwiseSoftmaxSweepOnce = GridwiseSoftmax_mk_to_mk<InDataType,
+                                                              OutDataType,
+                                                              AccDataType,
+                                                              GridDesc_M_K,
+                                                              BlockSize,
+                                                              MThreadClusterSize,
+                                                              KThreadClusterSize,
+                                                              MThreadSliceSize,
+                                                              KThreadSliceSize,
+                                                              InSrcVectorDim,
+                                                              InSrcVectorSize,
+                                                              OutDstVectorSize,
+                                                              true>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::vector<index_t> inLengths,
+                 const std::vector<index_t> inStrides,
+                 const std::vector<index_t> reduceDims,
+                 double alpha,
+                 double beta,
+                 const InDataType* in_dev,
+                 OutDataType* out_dev,
+                 InElementwiseOp in_elementwise_op,
+                 AccElementwiseOp acc_elementwise_op)
+            : in_dev_{in_dev},
+              out_dev_{out_dev},
+              in_elementwise_op_{in_elementwise_op},
+              acc_elementwise_op_{acc_elementwise_op}
+        {
+            alpha_ = static_cast<AccDataType>(alpha);
+            beta_  = static_cast<AccDataType>(beta);
+
+            if(Rank != inLengths.size() || Rank != inStrides.size() ||
+               NumReduceDim != reduceDims.size())
+            {
+                throw std::runtime_error(
+                    "One of inLengths/inStrides/reduceDims has invalid size!"
+                    "\nExpected size inLengths: " +
+                    std::to_string(Rank) + ", inStrides: " + std::to_string(Rank) +
+                    ", reduceDims: " + std::to_string(NumReduceDim) +
+                    "\nBut have inLengths: " + std::to_string(inLengths.size()) +
+                    ", inStrides: " + std::to_string(inStrides.size()) +
+                    ", reduceDims: " + std::to_string(reduceDims.size()));
+            }
+
+            for(std::size_t i = 0; i < reduceDims.size(); ++i)
+            {
+                if(reduceDims[i] < 0 || reduceDims[i] >= Rank)
+                {
+                    throw std::runtime_error("Provided reduce dimension exceed input tensor Rank!"
+                                             "\nHave reduceDims[" +
+                                             std::to_string(i) +
+                                             "]: " + std::to_string(reduceDims[i]));
+                }
+            }
+
+            inLengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
+            inStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inStrides, reduceDims);
+
+            long_index_t invariant_total_length;
+            long_index_t reduce_total_length;
+
+            std::tie(invariant_total_length, reduce_total_length) =
+                get_2d_lengths<Rank, NumReduceDim>(inLengths_);
+
+            if constexpr(NumInvariantDim == 0)
+                invariant_lowest_length_ = 1;
+            else
+                invariant_lowest_length_ = inLengths_[NumInvariantDim - 1];
+
+            blkGroupSize          = 1;
+            numBlockTileIteration = (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
+
+            gridSize = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
+                       M_BlockTileSize * blkGroupSize;
+        }
+
+        std::vector<index_t> inLengths_;
+        std::vector<index_t> inStrides_;
+
+        AccDataType alpha_;
+        AccDataType beta_;
+
+        const InDataType* in_dev_;
+        OutDataType* out_dev_;
+
+        InElementwiseOp in_elementwise_op_;
+        AccElementwiseOp acc_elementwise_op_;
+
+        index_t invariant_lowest_length_;
+
+        int blkGroupSize;
+        int numBlockTileIteration;
+        size_t gridSize;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto in_grid_desc_m_k = DeviceSoftmaxImpl::MakeSrc2dDescriptor(
+                arg.inLengths_, arg.inStrides_, arg.blkGroupSize, arg.numBlockTileIteration);
+            const auto out_grid_desc_m_k = DeviceSoftmaxImpl::MakeSrc2dDescriptor(
+                arg.inLengths_, arg.inStrides_, arg.blkGroupSize, arg.numBlockTileIteration);
+
+            bool sweep_once =
+                in_grid_desc_m_k.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize;
+
+            const auto kernel_main = sweep_once ? kernel_softmax<GridwiseSoftmaxSweepOnce,
+                                                                 InDataType,
+                                                                 OutDataType,
+                                                                 AccDataType,
+                                                                 GridDesc_M_K>
+                                                : kernel_softmax<GridwiseSoftmaxGeneric,
+                                                                 InDataType,
+                                                                 OutDataType,
+                                                                 AccDataType,
+                                                                 GridDesc_M_K>;
+
+            float avg_time = 0;
+
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.gridSize),
+                                               dim3(BlockSize),
+                                               0,
+                                               in_grid_desc_m_k,
+                                               out_grid_desc_m_k,
+                                               arg.blkGroupSize,
+                                               arg.numBlockTileIteration,
+                                               arg.alpha_,
+                                               arg.in_dev_,
+                                               arg.beta_,
+                                               arg.out_dev_);
+
+            return (avg_time);
+        };
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if constexpr(InSrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+            {
+                return false;
+            }
+            else
+            {
+                if(arg.inStrides_[NumInvariantDim - 1] != 1 && InSrcVectorSize != 1)
+                {
+                    return false;
+                }
+                if(arg.invariant_lowest_length_ % InSrcVectorSize != 0)
+                {
+                    return false;
+                }
+            }
+        }
+        else
+        {
+            if(arg.inStrides_[Rank - 1] != 1 && InSrcVectorSize != 1)
+            {
+                return false;
+            }
+            if(arg.inLengths_[Rank - 1] % InSrcVectorSize != 0)
+            {
+                return false;
+            }
+        }
+
+        // To improve
+        if(NumInvariantDim > 0 && arg.invariant_lowest_length_ % OutDstVectorSize != 0)
+        {
+            return false;
+        }
+
+        if(arg.inLengths_[Rank - 1] % OutDstVectorSize != 0)
+        {
+            return false;
+        }
+
+        return true;
+    };
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const std::vector<index_t> inLengths,
+                             const std::vector<index_t> inStrides,
+                             const std::vector<int> reduceDims,
+                             double alpha,
+                             double beta,
+                             const InDataType* in_dev,
+                             OutDataType* out_dev,
+                             InElementwiseOp in_elementwise_op,
+                             AccElementwiseOp acc_elementwise_op)
+    {
+        return Argument{inLengths,
+                        inStrides,
+                        reduceDims,
+                        alpha,
+                        beta,
+                        in_dev,
+                        out_dev,
+                        in_elementwise_op,
+                        acc_elementwise_op};
+    };
+
+    //
+    // @brief      Makes a pointer to Argument class.
+    //
+    // @param[in]  inLengths           Input tensor extent(s) from high to low dimension
+    // @param[in]  inStrides           Input tensor stride(s) from high to low dimension
+    // @param[in]  reduceDims          The dimension(s) the normalization operation is applied
+    // @param[in]  alpha               Typeless pointer in host memory storing the alpha scaling
+    //                                 value as type AccDataType
+    // @param[in]  beta                Typeless pointer in host memory storing the beta scaling
+    //                                 value as type AccDataType
+    // @param[in]  in_dev              Typeless const pointer in device memory storing the input
+    //                                 tensor
+    // @param      out_dev             Typeless pointer in device memory storing the output tensor
+    // @param[in]  in_elementwise_op   The input elementwise operation.
+    // @param[in]  acc_elementwise_op  The accumulation elementwise operation.
+    //
+    // @return     Unique pointer to the Argument class.
+    //
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const std::vector<index_t> inLengths,
+                                                      const std::vector<index_t> inStrides,
+                                                      const std::vector<int> reduceDims,
+                                                      double alpha,
+                                                      double beta,
+                                                      const void* in_dev,
+                                                      void* out_dev,
+                                                      InElementwiseOp in_elementwise_op,
+                                                      AccElementwiseOp acc_elementwise_op) override
+    {
+        return std::make_unique<Argument>(inLengths,
+                                          inStrides,
+                                          reduceDims,
+                                          alpha,
+                                          beta,
+                                          static_cast<const InDataType*>(in_dev),
+                                          static_cast<OutDataType*>(out_dev),
+                                          in_elementwise_op,
+                                          acc_elementwise_op);
+    };
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceReduceSoftmax<" 
+            << Rank << "," << NumReduceDim << "," << BlockSize << ","
+            << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ","
+            << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ","
+            << "InSrcVectorDim_" << InSrcVectorDim 
+            << "_InSrcVectorSize_" << InSrcVectorSize 
+            << "_OutDstVectorSize_" << OutDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_sparse_embeddings_forward_layernorm.hpp b/include/ck/tensor_operation/gpu/device/impl/device_sparse_embeddings_forward_layernorm.hpp
new file mode 100644
index 0000000000..7a62ec0465
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_sparse_embeddings_forward_layernorm.hpp
@@ -0,0 +1,193 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_sparse_embeddings_forward_layernorm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename EmbType,
+          typename IndexType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename OutType,
+          typename EmbElementwiseOperation,
+          ck::index_t BlockSize,
+          ck::index_t DimClusterSize,
+          ck::index_t RowClusterSize,
+          ck::index_t DimPerBlock,
+          ck::index_t RowPerBlock,
+          ck::index_t DimThreadSize,
+          ck::index_t RowVectorSize,
+          ck::index_t NumEmbeddings>
+struct DeviceSparseEmbeddingsForwardLayernorm : public BaseOperator
+{
+    static auto MakeOutputDescriptor(const index_t index_length, const index_t rows)
+    {
+        return make_naive_tensor_descriptor_packed(make_tuple(index_length, rows));
+    }
+
+    struct Argument : public BaseArgument
+    {
+        Argument(OutType* p_out,
+                 const ck::Array<EmbType*, NumEmbeddings>& p_embs,
+                 const ck::Array<IndexType*, NumEmbeddings>& p_indexs,
+                 const GammaDataType* p_gamma,
+                 const BetaDataType* p_beta,
+                 const ck::index_t EmbeddingDim,
+                 const ck::index_t IndexLength,
+                 const AccDataType epsilon,
+                 const EmbElementwiseOperation emb_elementwise_op)
+            : p_out_(p_out),
+              p_embs_(p_embs),
+              p_indexs_(p_indexs),
+              p_gamma_(p_gamma),
+              p_beta_(p_beta),
+              EmbeddingDim_(EmbeddingDim),
+              IndexLength_(IndexLength),
+              epsilon_(epsilon),
+              emb_elementwise_op_(emb_elementwise_op)
+        {
+            grid_size_ = (IndexLength + DimClusterSize - 1) / DimClusterSize;
+        }
+
+        OutType* p_out_;
+        ck::Array<EmbType*, NumEmbeddings> p_embs_;
+        ck::Array<IndexType*, NumEmbeddings> p_indexs_;
+        const GammaDataType* p_gamma_;
+        const BetaDataType* p_beta_;
+        ck::index_t EmbeddingDim_;
+        ck::index_t IndexLength_;
+        AccDataType epsilon_;
+        EmbElementwiseOperation emb_elementwise_op_;
+
+        size_t grid_size_;
+    };
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(void* p_out,
+                        const ck::Array<EmbType*, NumEmbeddings>& p_embs,
+                        const ck::Array<IndexType*, NumEmbeddings>& p_indexs,
+                        const void* p_gamma,
+                        const void* p_beta,
+                        ck::index_t EmbeddingDim,
+                        ck::index_t IndexLength,
+                        const AccDataType epsilon,
+                        const EmbElementwiseOperation emb_elementwise_op)
+    {
+        return std::make_unique<Argument>(reinterpret_cast<OutType*>(p_out),
+                                          p_embs,
+                                          p_indexs,
+                                          reinterpret_cast<const GammaDataType*>(p_gamma),
+                                          reinterpret_cast<const BetaDataType*>(p_beta),
+                                          EmbeddingDim,
+                                          IndexLength,
+                                          epsilon,
+                                          emb_elementwise_op);
+    }
+
+    using GridwiseSparseEmbedding =
+        GridwiseSparseEmbeddingsForwardLayernorm<EmbType,
+                                                 IndexType,
+                                                 GammaDataType,
+                                                 BetaDataType,
+                                                 AccDataType,
+                                                 OutType,
+                                                 decltype(MakeOutputDescriptor(1, 1)),
+                                                 EmbElementwiseOperation,
+                                                 BlockSize,
+                                                 DimClusterSize,
+                                                 RowClusterSize,
+                                                 DimPerBlock,
+                                                 RowPerBlock,
+                                                 DimThreadSize,
+                                                 RowVectorSize,
+                                                 NumEmbeddings>;
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            auto out_desc = MakeOutputDescriptor(arg.IndexLength_, arg.EmbeddingDim_);
+            const auto kernel_main =
+                kernel_sparse_embeddings_forward_layernorm<GridwiseSparseEmbedding,
+                                                           EmbType,
+                                                           IndexType,
+                                                           GammaDataType,
+                                                           BetaDataType,
+                                                           AccDataType,
+                                                           OutType,
+                                                           decltype(out_desc),
+                                                           EmbElementwiseOperation,
+                                                           NumEmbeddings>;
+            float avg_time = 0;
+            avg_time += launch_and_time_kernel(stream_config,
+                                               kernel_main,
+                                               dim3(arg.grid_size_),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.p_out_,
+                                               arg.p_embs_,
+                                               arg.p_indexs_,
+                                               arg.p_gamma_,
+                                               arg.p_beta_,
+                                               out_desc,
+                                               arg.epsilon_,
+                                               arg.emb_elementwise_op_);
+
+            return (avg_time);
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    static bool IsSupportedArgument(const Argument* p_arg)
+    {
+        return (RowPerBlock == p_arg->EmbeddingDim_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(dynamic_cast<const Argument*>(p_arg));
+    }
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>();
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceSparseEmbeddingsForwardLayernorm_"<< BlockSize << "_" <<
+            DimClusterSize << "x" << RowClusterSize << "_" <<
+            DimPerBlock << "x" << RowPerBlock << "_" <<
+            DimThreadSize << "x" << RowVectorSize;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..cfab7d29c9
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,1151 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatDsPointer,
+          typename FloatE,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AKB_AK0_M_AK1,
+          typename BGridDesc_BKB_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename ComputePtrOffsetOfBatch,
+          typename Block2ETileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_contraction_multiple_d_xdl_cshuffle(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatDsPointer p_ds_grid,
+            FloatE* __restrict__ p_e_grid,
+            const index_t batch_count,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const AGridDesc_AKB_AK0_M_AK1 a_grid_desc_akb_ak0_m_ak1,
+            const BGridDesc_BKB_BK0_N_BK1 b_grid_desc_bkb_bk0_n_bk1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
+            const Block2ETileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    FloatDsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                  p_b_grid + b_batch_offset,
+                                                  p_ds_grid_grp,
+                                                  p_e_grid + e_batch_offset,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  cde_element_op,
+                                                  a_grid_desc_akb_ak0_m_ak1,
+                                                  b_grid_desc_bkb_bk0_n_bk1,
+                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = a_grid_desc_akb_ak0_m_ak1;
+    ignore = b_grid_desc_bkb_bk0_n_bk1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_etile_map;
+    ignore = compute_ptr_offset_of_batch;
+#endif
+}
+
+} // namespace ck
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Tensor Contraction:
+//   input : A
+//   input : B
+//   input : D0, D1, ...
+//   output : E
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+//   B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+//   D[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+//   E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2, ...]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          TensorSpecialization ASpec,
+          TensorSpecialization BSpec,
+          TensorSpecialization DESpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceSplitKContractionMultipleD_Xdl_CShuffle
+    : public DeviceSplitKContractionMultipleD<NumDimG,
+                                              NumDimM,
+                                              NumDimN,
+                                              NumDimK,
+                                              ADataType,
+                                              BDataType,
+                                              DsDataType,
+                                              EDataType,
+                                              AElementwiseOperation,
+                                              BElementwiseOperation,
+                                              CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceSplitKContractionMultipleD_Xdl_CShuffle;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    // Assume: A[G0, G1, ..., M0, M1, M2, ..., K0, K1, K2, ...]
+    static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        assert(a_gs_ms_ks_lengths_vec.size() == NumDimG + NumDimM + NumDimK &&
+               a_gs_ms_ks_strides_vec.size() == NumDimG + NumDimM + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto a_ms_ks_lengths = to_tuple(
+            a_gs_ms_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimK>{});
+        const auto a_ms_ks_strides = to_tuple(
+            a_gs_ms_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimK>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimK, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(a_ms_ks_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(a_ms_ks_lengths, kDimIds);
+
+        if constexpr(ASpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto a_grid_desc_mraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(M, K),
+                make_tuple(a_ms_ks_strides[Number<NumDimM - 1>{}],
+                           a_ms_ks_strides[Number<NumDimM + NumDimK - 1>{}]));
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+        else
+        {
+            // naive tensor A[M0, M1, M2, ..., K0, K1, K2...]
+            const auto a_grid_desc_ms_ks =
+                make_naive_tensor_descriptor(a_ms_ks_lengths, a_ms_ks_strides);
+
+            // transformed tensor A[MRaw = M0 * M1 * M2 * ... , KRaw = K0 * K1 * K2 * ...]
+            const auto a_grid_desc_mraw_kraw = transform_tensor_descriptor(
+                a_grid_desc_ms_ks,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(kLengths)),
+                make_tuple(mDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        }
+    }
+
+    // Assume: B[G0, G1, ..., N0, N1, N2, ..., K0, K1, K2, ...]
+    static auto MakeBGridDescriptor_N_K(const std::vector<index_t>& b_gs_ns_ks_lengths_vec,
+                                        const std::vector<index_t>& b_gs_ns_ks_strides_vec)
+    {
+        assert(b_gs_ns_ks_lengths_vec.size() == NumDimG + NumDimN + NumDimK &&
+               b_gs_ns_ks_strides_vec.size() == NumDimG + NumDimN + NumDimK);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto b_ns_ks_lengths = to_tuple(
+            b_gs_ns_ks_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimN + NumDimK>{});
+        const auto b_ns_ks_strides = to_tuple(
+            b_gs_ns_ks_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimN + NumDimK>{});
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<0, NumDimN, 1>::type{};
+
+        // dimension Ids for K0, K1, ...
+        constexpr auto kDimIds =
+            typename arithmetic_sequence_gen<NumDimN, NumDimN + NumDimK, 1>::type{};
+
+        // lengths for K0, K1, ...
+        const auto kLengths = get_container_subset(b_ns_ks_lengths, kDimIds);
+
+        // lengths for N0, N1, ...
+        const auto nLengths = get_container_subset(b_ns_ks_lengths, nDimIds);
+
+        if constexpr(BSpec == TensorSpecialization::Packed)
+        {
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            auto K = container_reduce(kLengths, math::multiplies{}, Number<1>{});
+            const auto b_grid_desc_nraw_kraw = make_naive_tensor_descriptor(
+                make_tuple(N, K),
+                make_tuple(b_ns_ks_strides[Number<NumDimN - 1>{}],
+                           b_ns_ks_strides[Number<NumDimN + NumDimK - 1>{}]));
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+        else
+        {
+            // naive tensor B[N0, N1, N2, ..., K0, K1, K2, ...]
+            const auto b_grid_desc_ns_ks =
+                make_naive_tensor_descriptor(b_ns_ks_lengths, b_ns_ks_strides);
+
+            // transformed tensor B[NRaw = N0 * N1 * N2 * ..., KRaw = K0 * K1 * K2 * ...]
+            const auto b_grid_desc_nraw_kraw = transform_tensor_descriptor(
+                b_grid_desc_ns_ks,
+                make_tuple(make_merge_transform(nLengths), make_merge_transform(kLengths)),
+                make_tuple(nDimIds, kDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        }
+    }
+
+    // assume E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
+                                        const std::vector<index_t>& e_gs_ms_ns_strides_vec)
+    {
+        assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               e_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto e_ms_ns_lengths = to_tuple(
+            e_gs_ms_ns_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto e_ms_ns_strides = to_tuple(
+            e_gs_ms_ns_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds =
+            typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimN, 1>::type{};
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(M, N),
+                make_tuple(e_ms_ns_strides[Number<NumDimM - 1>{}],
+                           e_ms_ns_strides[Number<NumDimM + NumDimN - 1>{}]));
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+        else
+        {
+            // naive tensor E[M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_ms_ns =
+                make_naive_tensor_descriptor(e_ms_ns_lengths, e_ms_ns_strides);
+
+            // transformed tensor E[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 * N2 * ...]
+            const auto e_grid_desc_mraw_nraw = transform_tensor_descriptor(
+                e_grid_desc_ms_ns,
+                make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+                make_tuple(mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+        }
+    }
+
+    // assume E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+    static auto MakeEGridDescriptor_G_M_N(const std::vector<index_t>& e_gs_ms_ns_lengths_vec,
+                                          const std::vector<index_t>& e_gs_ms_ns_strides_vec)
+    {
+        assert(e_gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+               e_gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN);
+
+        const auto to_tuple = [&](auto& vec, auto start, auto end) {
+            return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+        };
+
+        const auto e_gs_ms_ns_lengths =
+            to_tuple(e_gs_ms_ns_lengths_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto e_gs_ms_ns_strides =
+            to_tuple(e_gs_ms_ns_strides_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // dimension Ids for G0, G1, ...
+        constexpr auto gDimIds = typename arithmetic_sequence_gen<0, NumDimG, 1>::type{};
+
+        // dimension Ids for M0, M1, ...
+        constexpr auto mDimIds =
+            typename arithmetic_sequence_gen<NumDimG, NumDimG + NumDimM, 1>::type{};
+
+        // dimension Ids for N0, N1, ...
+        constexpr auto nDimIds = typename arithmetic_sequence_gen<NumDimG + NumDimM,
+                                                                  NumDimG + NumDimM + NumDimN,
+                                                                  1>::type{};
+
+        // lengths for G0, G1, ...
+        const auto gLengths = get_container_subset(e_gs_ms_ns_lengths, gDimIds);
+
+        // lengths for M0, M1, ...
+        const auto mLengths = get_container_subset(e_gs_ms_ns_lengths, mDimIds);
+
+        // lengths for K0, K1, ...
+        const auto nLengths = get_container_subset(e_gs_ms_ns_lengths, nDimIds);
+
+        if constexpr(DESpec == TensorSpecialization::Packed)
+        {
+            auto G = container_reduce(gLengths, math::multiplies{}, Number<1>{});
+            auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+            auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+            const auto e_grid_desc_g_mraw_nraw = make_naive_tensor_descriptor(
+                make_tuple(G, M, N),
+                make_tuple(e_gs_ms_ns_strides[Number<NumDimG - 1>{}],
+                           e_gs_ms_ns_strides[Number<NumDimG + NumDimM - 1>{}],
+                           e_gs_ms_ns_strides[Number<NumDimG + NumDimM + NumDimN - 1>{}]));
+            // return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
+            return e_grid_desc_g_mraw_nraw;
+        }
+        else
+        {
+            // naive tensor E[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+            const auto e_grid_desc_gs_ms_ns =
+                make_naive_tensor_descriptor(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
+
+            // transformed tensor E[G = G0 * G1 * ..., MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
+            // N2 * ...]
+            const auto e_grid_desc_g_mraw_nraw = transform_tensor_descriptor(
+                e_grid_desc_gs_ms_ns,
+                make_tuple(make_merge_transform(gLengths),
+                           make_merge_transform(mLengths),
+                           make_merge_transform(nLengths)),
+                make_tuple(gDimIds, mDimIds, nDimIds),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            // return matrix_padder.PadCDescriptor_M_N(e_grid_desc_g_mraw_nraw);
+            return e_grid_desc_g_mraw_nraw;
+        }
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_M_N(ds_gs_ms_ns_lengths_vec[i],
+                                                         ds_gs_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    static auto MakeDsGridDescriptor_G_M_N(
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths_vec,
+        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides_vec)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return DeviceOp::MakeEGridDescriptor_G_M_N(ds_gs_ms_ns_lengths_vec[i],
+                                                           ds_gs_ms_ns_strides_vec[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K({}, {}));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K({}, {}));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({{}}, {{}}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N({}, {}));
+
+    using DsGridDesc_G_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_G_M_N({}, {}))>;
+    using EGridDesc_G_M_N  = decltype(MakeEGridDescriptor_G_M_N({}, {}));
+
+    struct ComputePtrOffsetOfStridedBatch
+    {
+        ComputePtrOffsetOfStridedBatch(index_t batch_stride_A,
+                                       index_t batch_stride_B,
+                                       DsGridDesc_G_M_N ds_grid_desc_g_m_n,
+                                       EGridDesc_G_M_N e_grid_desc_g_m_n)
+            : batch_stride_A_(batch_stride_A),
+              batch_stride_B_(batch_stride_B),
+              ds_grid_desc_g_m_n_(ds_grid_desc_g_m_n),
+              e_grid_desc_g_m_n_(e_grid_desc_g_m_n)
+        {
+        }
+
+        __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(batch_stride_A_);
+        }
+
+        __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+        {
+            return g_idx * static_cast<long_index_t>(batch_stride_B_);
+        }
+
+        __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+        {
+            std::array<long_index_t, NumDTensor> ds_offset;
+
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                ds_offset[i] = static_cast<long_index_t>(g_idx) *
+                               ds_grid_desc_g_m_n_[i].CalculateOffset(make_multi_index(1, 0, 0));
+            });
+
+            return ds_offset;
+        }
+
+        __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+        {
+            return static_cast<long_index_t>(g_idx) *
+                   e_grid_desc_g_m_n_.CalculateOffset(make_multi_index(1, 0, 0));
+        }
+
+        private:
+        index_t batch_stride_A_;
+        index_t batch_stride_B_;
+        DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
+        EGridDesc_G_M_N e_grid_desc_g_m_n_;
+    };
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmSplitKMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_M_K,
+        BGridDesc_N_K,
+        DsGridDesc_M_N,
+        EGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    // GridwiseGemm
+    using GridwiseGemmAtomicAdd = GridwiseGemmSplitKMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::AtomicAdd,
+        AGridDesc_M_K,
+        BGridDesc_N_K,
+        DsGridDesc_M_N,
+        EGridDesc_M_N,
+        NumGemmKPrefetchStage,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CDEBlockTransferScalarPerVector_NPerBlock,
+        LoopSched>;
+
+    using AGridDesc_AKB_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AKB_AK0_M_AK1(
+            AGridDesc_M_K{}, 1))>;
+    using BGridDesc_BKB_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BKB_BK0_N_BK1(
+            BGridDesc_N_K{}, 1))>;
+
+    using Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_a_grid,
+                 const void* p_b_grid,
+                 std::array<const void*, NumDTensor> p_ds_grid,
+                 void* p_e_grid,
+                 const std::vector<index_t>& a_gs_ms_ns_lengths,
+                 const std::vector<index_t>& a_gs_ms_ks_strides,
+                 const std::vector<index_t>& b_gs_ns_ks_lengths,
+                 const std::vector<index_t>& b_gs_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                 const std::vector<index_t>& e_gs_ms_ns_lengths,
+                 const std::vector<index_t>& e_gs_ms_ns_strides,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op,
+                 index_t split_k)
+            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
+              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
+              a_grid_desc_m_k_{
+                  DeviceOp::MakeAGridDescriptor_M_K(a_gs_ms_ns_lengths, a_gs_ms_ks_strides)},
+              b_grid_desc_n_k_{
+                  DeviceOp::MakeBGridDescriptor_N_K(b_gs_ns_ks_lengths, b_gs_ns_ks_strides)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{
+                  DeviceOp::MakeEGridDescriptor_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides)},
+              ds_grid_desc_g_m_n_{
+                  DeviceOp::MakeDsGridDescriptor_G_M_N(ds_gs_ms_ns_lengths, ds_gs_ms_ns_strides)},
+              e_grid_desc_g_m_n_{
+                  DeviceOp::MakeEGridDescriptor_G_M_N(e_gs_ms_ns_lengths, e_gs_ms_ns_strides)},
+              a_grid_desc_akb_ak0_m_ak1_{GridwiseGemm::MakeDefaultAGridDescriptor_AKB_AK0_M_AK1(
+                  a_grid_desc_m_k_, split_k)},
+              b_grid_desc_bkb_bk0_n_bk1_{GridwiseGemm::MakeDefaultBGridDescriptor_BKB_BK0_N_BK1(
+                  b_grid_desc_n_k_, split_k)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{
+                  GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_, split_k)},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_mz_stride_{},
+              a_kz_stride_{},
+              b_nz_stride_{},
+              b_kz_stride_{},
+              ds_nz_stride_{},
+              e_nz_stride_{},
+              a_batch_stride_{a_gs_ms_ks_strides[NumDimG - 1]},
+              b_batch_stride_{b_gs_ns_ks_strides[NumDimG - 1]},
+              compute_ptr_offset_of_batch_{
+                  a_batch_stride_, b_batch_stride_, ds_grid_desc_g_m_n_, e_grid_desc_g_m_n_},
+              split_k_{split_k}
+        {
+            static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0, "");
+
+            // populate pointer, batch stride, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
+
+                // D desc
+                ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N(ds_gs_ms_ns_lengths[i],
+                                                                         ds_gs_ms_ns_strides[i]);
+            });
+
+            // populate desc for Ds/E
+            if(GridwiseGemm::CheckValidity(a_grid_desc_akb_ak0_m_ak1_,
+                                           b_grid_desc_bkb_bk0_n_bk1_,
+                                           ds_grid_desc_m_n_,
+                                           e_grid_desc_m_n_,
+                                           block_2_etile_map_))
+            {
+                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        e_grid_desc_m_n_);
+
+                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                        ds_grid_desc_m_n_);
+            }
+
+            // for sanity check of vector memory access
+            a_mz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM - 1];
+            a_kz_stride_ = a_gs_ms_ks_strides[NumDimG + NumDimM + NumDimK - 1];
+            b_nz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN - 1];
+            b_kz_stride_ = b_gs_ns_ks_strides[NumDimG + NumDimN + NumDimK - 1];
+
+            for(index_t i = 0; i < NumDTensor; ++i)
+            {
+                ds_nz_stride_[i] = ds_gs_ms_ns_strides[i][NumDimG + NumDimM + NumDimN - 1];
+            }
+
+            e_nz_stride_ = e_gs_ms_ns_strides[NumDimG + NumDimM + NumDimN - 1];
+
+            Print();
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_.GetLength(I0) << ", "
+                      << a_grid_desc_m_k_.GetLength(I1) << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_.GetLength(I0) << ", "
+                      << b_grid_desc_n_k_.GetLength(I1) << std::endl;
+
+            std::cout << "A[akb, ak0, m, ak1]: " << a_grid_desc_akb_ak0_m_ak1_.GetLength(I0) << ", "
+                      << a_grid_desc_akb_ak0_m_ak1_.GetLength(I1) << ", "
+                      << a_grid_desc_akb_ak0_m_ak1_.GetLength(I2) << ", "
+                      << a_grid_desc_akb_ak0_m_ak1_.GetLength(I3) << std::endl;
+            std::cout << "B[bkb, bk0, n, bk1]: " << b_grid_desc_bkb_bk0_n_bk1_.GetLength(I0) << ", "
+                      << b_grid_desc_bkb_bk0_n_bk1_.GetLength(I1) << ", "
+                      << b_grid_desc_bkb_bk0_n_bk1_.GetLength(I2) << ", "
+                      << b_grid_desc_bkb_bk0_n_bk1_.GetLength(I3) << std::endl;
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i].GetLength(I0) << ", "
+                          << ds_grid_desc_m_n_[i].GetLength(I1) << std::endl;
+            });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_.GetLength(I0) << ", "
+                      << e_grid_desc_m_n_.GetLength(I1) << std::endl;
+        }
+
+        //  private:
+        // pointers
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        DsGridDesc_G_M_N ds_grid_desc_g_m_n_;
+        EGridDesc_G_M_N e_grid_desc_g_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AKB_AK0_M_AK1 a_grid_desc_akb_ak0_m_ak1_;
+        BGridDesc_BKB_BK0_N_BK1 b_grid_desc_bkb_bk0_n_bk1_;
+        typename GridwiseGemm::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // Strides for the last M/N/K dimensions of A/B/Ds/E
+        //   for sanity check of vector load/store
+        index_t a_mz_stride_;
+        index_t a_kz_stride_;
+        index_t b_nz_stride_;
+        index_t b_kz_stride_;
+        std::array<index_t, NumDTensor> ds_nz_stride_;
+        index_t e_mz_stride_;
+        index_t e_nz_stride_;
+
+        index_t a_batch_stride_;
+        index_t b_batch_stride_;
+
+        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+
+        index_t split_k_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_akb_ak0_m_ak1_,
+                                            arg.b_grid_desc_bkb_bk0_n_bk1_,
+                                            arg.ds_grid_desc_m_n_,
+                                            arg.e_grid_desc_m_n_,
+                                            arg.block_2_etile_map_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemmMultipleD_xdl_cshuffle has invalid setting");
+            }
+
+            const index_t G = arg.e_grid_desc_g_m_n_.GetLength(I0);
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * G;
+
+            const auto K = arg.a_grid_desc_akb_ak0_m_ak1_.GetLength(I1) *
+                           arg.a_grid_desc_akb_ak0_m_ak1_.GetLength(I3);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_contraction_multiple_d_xdl_cshuffle<
+                    GridwiseGemm,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemm::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AKB_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BKB_BK0_N_BK1,
+                    typename GridwiseGemm::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    ComputePtrOffsetOfStridedBatch,
+                    typename GridwiseGemm::DefaultBlock2ETileMap,
+                    has_main_loop>;
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              G,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_akb_ak0_m_ak1_,
+                                              arg.b_grid_desc_bkb_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.compute_ptr_offset_of_batch_,
+                                              arg.block_2_etile_map_);
+            };
+
+            auto launch_kernel_atomic_add = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                const auto kernel = kernel_contraction_multiple_d_xdl_cshuffle<
+                    GridwiseGemmAtomicAdd,
+                    ADataType, // TODO: distiguish A/B datatype
+                    typename GridwiseGemmAtomicAdd::DsGridPointer,
+                    EDataType,
+                    AElementwiseOperation,
+                    BElementwiseOperation,
+                    CDEElementwiseOperation,
+                    DeviceOp::AGridDesc_AKB_AK0_M_AK1,
+                    DeviceOp::BGridDesc_BKB_BK0_N_BK1,
+                    typename GridwiseGemmAtomicAdd::
+                        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemmAtomicAdd::
+                        EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    ComputePtrOffsetOfStridedBatch,
+                    typename GridwiseGemmAtomicAdd::DefaultBlock2ETileMap,
+                    has_main_loop>;
+
+                hipGetErrorString(hipMemsetAsync(
+                    arg.p_e_grid_,
+                    0,
+                    arg.e_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
+                        sizeof(EDataType),
+                    stream_config.stream_id_));
+
+                return launch_and_time_kernel(stream_config,
+                                              kernel,
+                                              dim3(grid_size),
+                                              dim3(BlockSize),
+                                              0,
+                                              arg.p_a_grid_,
+                                              arg.p_b_grid_,
+                                              arg.p_ds_grid_,
+                                              arg.p_e_grid_,
+                                              G,
+                                              arg.a_element_op_,
+                                              arg.b_element_op_,
+                                              arg.cde_element_op_,
+                                              arg.a_grid_desc_akb_ak0_m_ak1_,
+                                              arg.b_grid_desc_bkb_bk0_n_bk1_,
+                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                              arg.compute_ptr_offset_of_batch_,
+                                              arg.block_2_etile_map_);
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                if(arg.split_k_ <= 1)
+                    return launch_kernel(integral_constant<bool, true>{});
+                else
+                    return launch_kernel_atomic_add(integral_constant<bool, true>{});
+            }
+            else
+            {
+                if(arg.split_k_ <= 1)
+                    return launch_kernel(integral_constant<bool, false>{});
+                else
+                    return launch_kernel_atomic_add(integral_constant<bool, false>{});
+            }
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_akb_ak0_m_ak1_,
+                                        arg.b_grid_desc_bkb_bk0_n_bk1_,
+                                        arg.ds_grid_desc_m_n_,
+                                        arg.e_grid_desc_m_n_,
+                                        arg.block_2_etile_map_))
+        {
+            return false;
+        }
+
+        // check vector access
+        static_assert((ABlockTransferSrcVectorDim == 2 || ABlockTransferSrcVectorDim == 3) &&
+                          (BBlockTransferSrcVectorDim == 2 || BBlockTransferSrcVectorDim == 3),
+                      "wrong!");
+
+        // vector memory access of A: could be on M or AK1 dimension
+        if constexpr(ABlockTransferSrcVectorDim == 2)
+        {
+            if(!(arg.a_mz_stride_ == 1 &&
+                 arg.a_grid_desc_akb_ak0_m_ak1_.GetLength(I2) % ABlockTransferSrcScalarPerVector ==
+                     0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.a_kz_stride_ == 1 &&
+                 arg.a_grid_desc_akb_ak0_m_ak1_.GetLength(I3) % ABlockTransferSrcScalarPerVector ==
+                     0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of B: could be on N or BK1 dimension
+        if constexpr(BBlockTransferSrcVectorDim == 2)
+        {
+            if(!(arg.b_nz_stride_ == 1 &&
+                 arg.b_grid_desc_bkb_bk0_n_bk1_.GetLength(I2) % BBlockTransferSrcScalarPerVector ==
+                     0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(arg.b_kz_stride_ == 1 &&
+                 arg.b_grid_desc_bkb_bk0_n_bk1_.GetLength(I3) % BBlockTransferSrcScalarPerVector ==
+                     0))
+            {
+                return false;
+            }
+        }
+
+        // vector memory access of Ds: always on NPerBlock dimension
+        bool valid_d_access = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            if(!(arg.ds_nz_stride_[i] == 1 &&
+                 arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_[i].GetLength(I3) %
+                         CDEBlockTransferScalarPerVector_NPerBlock ==
+                     0))
+            {
+                valid_d_access = false;
+            }
+        });
+
+        if(valid_d_access == false)
+        {
+            return false;
+        }
+
+        // vector memory access of E: always on NPerBlock dimension
+        if(!((arg.e_nz_stride_ == 1 &&
+              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_.GetLength(I3) %
+                      CDEBlockTransferScalarPerVector_NPerBlock ==
+                  0) ||
+             CDEBlockTransferScalarPerVector_NPerBlock == 1))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto
+    MakeArgument(const void* p_a,
+                 const void* p_b,
+                 std::array<const void*, NumDTensor> p_ds,
+                 void* p_e,
+                 const std::vector<index_t>& a_gs_ms_ns_lengths,
+                 const std::vector<index_t>& a_gs_ms_ks_strides,
+                 const std::vector<index_t>& b_gs_ns_ks_lengths,
+                 const std::vector<index_t>& b_gs_ns_ks_strides,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                 const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                 const std::vector<index_t>& e_gs_ms_ns_lengths,
+                 const std::vector<index_t>& e_gs_ms_ns_strides,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CDEElementwiseOperation cde_element_op,
+                 index_t split_k)
+    {
+        return Argument{p_a,
+                        p_b,
+                        p_ds,
+                        p_e,
+                        a_gs_ms_ns_lengths,
+                        a_gs_ms_ks_strides,
+                        b_gs_ns_ks_lengths,
+                        b_gs_ns_ks_strides,
+                        ds_gs_ms_ns_lengths,
+                        ds_gs_ms_ns_strides,
+                        e_gs_ms_ns_lengths,
+                        e_gs_ms_ns_strides,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op,
+                        split_k};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        const std::vector<index_t>& a_gs_ms_ns_lengths,
+                        const std::vector<index_t>& a_gs_ms_ks_strides,
+                        const std::vector<index_t>& b_gs_ns_ks_lengths,
+                        const std::vector<index_t>& b_gs_ns_ks_strides,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
+                        const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
+                        const std::vector<index_t>& e_gs_ms_ns_lengths,
+                        const std::vector<index_t>& e_gs_ms_ns_strides,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op,
+                        index_t split_k) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_gs_ms_ns_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b_gs_ns_ks_lengths,
+                                          b_gs_ns_ks_strides,
+                                          ds_gs_ms_ns_lengths,
+                                          ds_gs_ms_ns_strides,
+                                          e_gs_ms_ns_lengths,
+                                          e_gs_ms_ns_strides,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op,
+                                          split_k);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceSplitKContractionMultipleD_Xdl_CShuffle"
+            << "<"
+            << NumDimG << ", "
+            << NumDimM << ", "
+            << NumDimN << ", "
+            << NumDimK << ", "
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << AK1 << ", "
+            << BK1 << ", "
+            << ABlockTransferSrcVectorDim << ", "
+            << BBlockTransferSrcVectorDim
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/masking_specialization.hpp b/include/ck/tensor_operation/gpu/device/masking_specialization.hpp
new file mode 100644
index 0000000000..d6d6f74abd
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/masking_specialization.hpp
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum struct MaskingSpecialization
+{
+    MaskDisabled,
+    MaskOutUpperTriangle
+};
+
+inline std::string getMaskingSpecializationString(const MaskingSpecialization& s)
+{
+    switch(s)
+    {
+    case MaskingSpecialization::MaskDisabled: return "MaskDisabled";
+    case MaskingSpecialization::MaskOutUpperTriangle: return "MaskOutUpperTriangle";
+    default: return "Unrecognized specialization!";
+    }
+}
+
+struct MaskDisabledPredicate
+{
+    __host__ __device__ constexpr bool operator()(index_t /*m*/, index_t /*n*/) const
+    {
+        return false;
+    };
+
+    __host__ __device__ constexpr bool
+        IsTileSkippable(index_t /*m*/, index_t /*n*/, index_t /*m_tile*/, index_t /*n_tile*/) const
+    {
+        return false;
+    }
+};
+
+struct MaskOutUpperTrianglePredicate
+{
+    __host__ __device__ constexpr bool operator()(index_t m, index_t n) const { return n > m; }
+
+    __host__ __device__ constexpr bool
+    IsTileSkippable(index_t m, index_t n, index_t m_tile, index_t /*n_tile*/) const
+    {
+        return operator()(m + m_tile - 1, n);
+    }
+};
+
+// to track the points which need to be set to -inf on C0
+// Note: no need to reset M padding value, because they will not be stored out.
+template <typename MaskOutPredicate>
+struct C0MatrixMask_impl
+{
+    C0MatrixMask_impl(index_t NRaw) : NRaw_(NRaw), predicate_(MaskOutPredicate{}) {}
+
+    __host__ __device__ constexpr bool IsNOutOfBound(/*index_t m, */ index_t n) const
+    {
+        return n >= NRaw_;
+    }
+
+    __host__ __device__ constexpr bool IsMaskedElement(index_t m, index_t n) const
+    {
+        return predicate_(m, n) || IsNOutOfBound(n);
+    }
+
+    __host__ __device__ constexpr bool
+    IsTileSkippable(index_t m, index_t n, index_t m_tile, index_t n_tile) const
+    {
+        return predicate_.IsTileSkippable(m, n, m_tile, n_tile);
+    }
+
+    private:
+    // index_t MRaw_;
+    index_t NRaw_;
+    MaskOutPredicate predicate_;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/matrix_padder.hpp b/include/ck/tensor_operation/gpu/device/matrix_padder.hpp
new file mode 100644
index 0000000000..c66d2fc516
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/matrix_padder.hpp
@@ -0,0 +1,382 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename TensorDesc,
+          typename TileLengths, // Tuple<...>
+          typename DoPads>      // Sequence<bool, bool, ...>
+__host__ __device__ constexpr auto
+PadTensorDescriptor(const TensorDesc& desc, const TileLengths& tile_lengths, DoPads)
+{
+    constexpr index_t num_dim = DoPads::Size();
+
+    static_assert(num_dim == TileLengths::Size() && num_dim == TensorDesc::GetNumOfDimension(),
+                  "wrong! inconsistent # of dimensions");
+
+    // transforms
+    const auto transforms = generate_tuple(
+        [&](auto idim) {
+            const auto MRaw = desc.GetLength(idim);
+
+            const auto MPerTile = tile_lengths[idim];
+
+            const auto M = math::integer_divide_ceil(MRaw, MPerTile) * MPerTile;
+
+            const auto MPad = M - MRaw;
+
+            const bool DoPadM = DoPads::At(idim);
+
+            const auto MTransform = conditional_expr<DoPadM>(make_right_pad_transform(MRaw, MPad),
+                                                             make_pass_through_transform(MRaw));
+
+            return MTransform;
+        },
+        Number<num_dim>{});
+
+    // lower dimension Id
+    const auto lower_dimss =
+        generate_tuple([&](auto idim) { return Sequence<idim.value>{}; }, Number<num_dim>{});
+
+    // upper dimension Id
+    const auto upper_dimss = lower_dimss;
+
+    return transform_tensor_descriptor(desc, transforms, lower_dimss, upper_dimss);
+}
+
+// M/N/K/OPerTileType could be index_t or Number<>
+template <GemmSpecialization GemmSpec,
+          typename MPerTileType,
+          typename NPerTileType,
+          typename KPerTileType,
+          typename OPerTileType>
+struct GemmGemmPadder
+{
+    // TODO: hard to scale; use mask instead
+    static constexpr bool PadM =
+        GemmSpec == GemmSpecialization::MPadding || GemmSpec == GemmSpecialization::MNPadding ||
+        GemmSpec == GemmSpecialization::MKPadding || GemmSpec == GemmSpecialization::MNKPadding ||
+        GemmSpec == GemmSpecialization::MOPadding || GemmSpec == GemmSpecialization::MNOPadding ||
+        GemmSpec == GemmSpecialization::MKOPadding || GemmSpec == GemmSpecialization::MNKOPadding;
+    static constexpr bool PadN =
+        GemmSpec == GemmSpecialization::NPadding || GemmSpec == GemmSpecialization::MNPadding ||
+        GemmSpec == GemmSpecialization::NKPadding || GemmSpec == GemmSpecialization::MNKPadding ||
+        GemmSpec == GemmSpecialization::NOPadding || GemmSpec == GemmSpecialization::MNOPadding ||
+        GemmSpec == GemmSpecialization::NKOPadding || GemmSpec == GemmSpecialization::MNKOPadding;
+    static constexpr bool PadK =
+        GemmSpec == GemmSpecialization::KPadding || GemmSpec == GemmSpecialization::MKPadding ||
+        GemmSpec == GemmSpecialization::NKPadding || GemmSpec == GemmSpecialization::MNKPadding ||
+        GemmSpec == GemmSpecialization::KOPadding || GemmSpec == GemmSpecialization::MKOPadding ||
+        GemmSpec == GemmSpecialization::NKOPadding || GemmSpec == GemmSpecialization::MNKOPadding;
+    static constexpr bool PadO =
+        GemmSpec == GemmSpecialization::OPadding || GemmSpec == GemmSpecialization::MOPadding ||
+        GemmSpec == GemmSpecialization::NOPadding || GemmSpec == GemmSpecialization::KOPadding ||
+        GemmSpec == GemmSpecialization::MNOPadding || GemmSpec == GemmSpecialization::MKOPadding ||
+        GemmSpec == GemmSpecialization::NKOPadding || GemmSpec == GemmSpecialization::MNKOPadding;
+
+    // A[M, K]
+    template <typename ADesc_MRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadADescriptor_M_K(const ADesc_MRaw_KRaw& a_desc_mraw_kraw) const
+    {
+        return PadTensorDescriptor(
+            a_desc_mraw_kraw, make_tuple(MPerTile_, KPerTile_), Sequence<PadM, PadK>{});
+    }
+
+    // B[K, N]
+    template <typename BDesc_NRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadBDescriptor_N_K(const BDesc_NRaw_KRaw& b_desc_nraw_kraw) const
+    {
+        return PadTensorDescriptor(
+            b_desc_nraw_kraw, make_tuple(NPerTile_, KPerTile_), Sequence<PadN, PadK>{});
+    }
+
+    // B1[Gemm1N, Gemm1K] = B1[O, N]
+    template <typename B1Desc_NRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadB1Descriptor_N_K(const B1Desc_NRaw_KRaw& b1_desc_nraw_kraw) const
+    {
+        return PadTensorDescriptor(
+            b1_desc_nraw_kraw, make_tuple(OPerTile_, NPerTile_), Sequence<PadO, PadN>{});
+    }
+
+    // C[M, Gemm1N] = C[M, O]
+    template <typename CDesc_MRaw_NRaw>
+    __host__ __device__ constexpr auto
+    PadCDescriptor_M_N(const CDesc_MRaw_NRaw& c_desc_mraw_nraw) const
+    {
+        return PadTensorDescriptor(
+            c_desc_mraw_nraw, make_tuple(MPerTile_, OPerTile_), Sequence<PadM, PadO>{});
+    }
+
+    MPerTileType MPerTile_;
+    NPerTileType NPerTile_;
+    KPerTileType KPerTile_;
+    OPerTileType OPerTile_;
+};
+
+// M/N/KPerTileType could be index_t or Number<>
+template <GemmSpecialization GemmSpec,
+          typename MPerTileType,
+          typename NPerTileType,
+          typename KPerTileType>
+struct GemmPadder
+{
+    static constexpr bool PadM =
+        (GemmSpec == GemmSpecialization::MPadding || GemmSpec == GemmSpecialization::MNPadding ||
+         GemmSpec == GemmSpecialization::MKPadding || GemmSpec == GemmSpecialization::MNKPadding);
+    static constexpr bool PadN =
+        (GemmSpec == GemmSpecialization::NPadding || GemmSpec == GemmSpecialization::MNPadding ||
+         GemmSpec == GemmSpecialization::NKPadding || GemmSpec == GemmSpecialization::MNKPadding);
+    static constexpr bool PadK =
+        (GemmSpec == GemmSpecialization::KPadding || GemmSpec == GemmSpecialization::MKPadding ||
+         GemmSpec == GemmSpecialization::NKPadding || GemmSpec == GemmSpecialization::MNKPadding);
+
+    template <typename ADesc_MRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadADescriptor_M_K(const ADesc_MRaw_KRaw& a_desc_mraw_kraw) const
+    {
+        return PadTensorDescriptor(
+            a_desc_mraw_kraw, make_tuple(MPerTile_, KPerTile_), Sequence<PadM, PadK>{});
+    }
+
+    template <typename BDesc_NRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadBDescriptor_N_K(const BDesc_NRaw_KRaw& b_desc_nraw_kraw) const
+    {
+        return PadTensorDescriptor(
+            b_desc_nraw_kraw, make_tuple(NPerTile_, KPerTile_), Sequence<PadN, PadK>{});
+    }
+
+    template <typename CDesc_MRaw_NRaw>
+    __host__ __device__ constexpr auto
+    PadCDescriptor_M_N(const CDesc_MRaw_NRaw& c_desc_mraw_nraw) const
+    {
+        return PadTensorDescriptor(
+            c_desc_mraw_nraw, make_tuple(MPerTile_, NPerTile_), Sequence<PadM, PadN>{});
+    }
+
+    MPerTileType MPerTile_;
+    NPerTileType NPerTile_;
+    KPerTileType KPerTile_;
+};
+
+// Alias of GemmPadder; to deprecate
+template <GemmSpecialization GemmSpec,
+          typename MPerTileType,
+          typename NPerTileType,
+          typename KPerTileType>
+struct MatrixPadder : public GemmPadder<GemmSpec, MPerTileType, NPerTileType, KPerTileType>
+{
+};
+
+// M/N/KPerTileType could be index_t or Number<>
+template <bool PadM,
+          bool PadN,
+          bool PadK,
+          typename MPerTileType,
+          typename NPerTileType,
+          typename KPerTileType>
+struct GemmPadder_v2
+{
+    template <typename ADesc_MRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadADescriptor_M_K(const ADesc_MRaw_KRaw& a_desc_mraw_kraw) const
+    {
+        return PadTensorDescriptor(
+            a_desc_mraw_kraw, make_tuple(MPerTile_, KPerTile_), Sequence<PadM, PadK>{});
+    }
+
+    template <typename BDesc_NRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadBDescriptor_N_K(const BDesc_NRaw_KRaw& b_desc_nraw_kraw) const
+    {
+        return PadTensorDescriptor(
+            b_desc_nraw_kraw, make_tuple(NPerTile_, KPerTile_), Sequence<PadN, PadK>{});
+    }
+
+    template <typename CDesc_MRaw_NRaw>
+    __host__ __device__ constexpr auto
+    PadCDescriptor_M_N(const CDesc_MRaw_NRaw& c_desc_mraw_nraw) const
+    {
+        return PadTensorDescriptor(
+            c_desc_mraw_nraw, make_tuple(MPerTile_, NPerTile_), Sequence<PadM, PadN>{});
+    }
+
+    MPerTileType MPerTile_;
+    NPerTileType NPerTile_;
+    KPerTileType KPerTile_;
+};
+
+// M/N/KPerTileType could be index_t or Number<>
+template <bool PadM,
+          bool PadN,
+          bool PadK,
+          typename MPerTileType,
+          typename NPerTileType,
+          typename KPerTileType>
+struct MatrixPadder_v2
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    template <typename ADesc_MRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadADescriptor_M_K(const ADesc_MRaw_KRaw& a_desc_mraw_kraw) const
+    {
+        const auto MRaw = a_desc_mraw_kraw.GetLength(I0);
+        const auto KRaw = a_desc_mraw_kraw.GetLength(I1);
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerTile_) * MPerTile_;
+        const auto K = math::integer_divide_ceil(KRaw, KPerTile_) * KPerTile_;
+
+        const auto MPad = M - MRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(PadM && PadK)
+        {
+            // pad both M and K
+            return transform_tensor_descriptor(a_desc_mraw_kraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(KRaw, KPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(PadM && (!PadK))
+        {
+            // pad M, but not K
+            return transform_tensor_descriptor(
+                a_desc_mraw_kraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(KRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr((!PadM) && PadK)
+        {
+            // pad K, but not M
+            return transform_tensor_descriptor(
+                a_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or K
+            return a_desc_mraw_kraw;
+        }
+    }
+
+    template <typename BDesc_NRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadBDescriptor_N_K(const BDesc_NRaw_KRaw& b_desc_nraw_kraw) const
+    {
+        const auto NRaw = b_desc_nraw_kraw.GetLength(I0);
+        const auto KRaw = b_desc_nraw_kraw.GetLength(I1);
+
+        const auto N = math::integer_divide_ceil(NRaw, NPerTile_) * NPerTile_;
+        const auto K = math::integer_divide_ceil(KRaw, KPerTile_) * KPerTile_;
+
+        const auto NPad = N - NRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(PadN && PadK)
+        {
+            // pad both N and K
+            return transform_tensor_descriptor(b_desc_nraw_kraw,
+                                               make_tuple(make_right_pad_transform(NRaw, NPad),
+                                                          make_right_pad_transform(KRaw, KPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(PadN && (!PadK))
+        {
+            // pad N, but not K
+            return transform_tensor_descriptor(
+                b_desc_nraw_kraw,
+                make_tuple(make_right_pad_transform(NRaw, NPad), make_pass_through_transform(KRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr((!PadN) && PadK)
+        {
+            // pad K, but not N
+            return transform_tensor_descriptor(
+                b_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad N or K
+            return b_desc_nraw_kraw;
+        }
+    }
+
+    template <typename CDesc_MRaw_NRaw>
+    __host__ __device__ constexpr auto
+    PadCDescriptor_M_N(const CDesc_MRaw_NRaw& c_desc_mraw_nraw) const
+    {
+        const auto MRaw = c_desc_mraw_nraw.GetLength(I0);
+        const auto NRaw = c_desc_mraw_nraw.GetLength(I1);
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerTile_) * MPerTile_;
+        const auto N = math::integer_divide_ceil(NRaw, NPerTile_) * NPerTile_;
+
+        const auto MPad = M - MRaw;
+        const auto NPad = N - NRaw;
+
+        if constexpr(PadM && PadN)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(NRaw, NPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(PadM && (!PadN))
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr((!PadM) && PadN)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_desc_mraw_nraw;
+        }
+    }
+
+    MPerTileType MPerTile_;
+    NPerTileType NPerTile_;
+    KPerTileType KPerTile_;
+};
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp b/include/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp
new file mode 100644
index 0000000000..5351d4ef24
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp
@@ -0,0 +1,186 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+// FIXME: can it be replaced with ck::Tuple?
+#include <tuple>
+
+namespace ck {
+
+// The templated struct reduce_binary_operator maps the enum Ids of binary operators to their
+// respective functor classes.
+// The boolean member "indexable" are also provided in reduce_binary_operactor for
+// easier checking by the upper-layer codes in the kernels.
+
+template <ReduceTensorOp Op>
+struct reduce_binary_operator;
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::ADD>
+{
+    using opType = reduce::Add;
+
+    static constexpr bool indexable = false;
+};
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::MUL>
+{
+    using opType = reduce::Mul;
+
+    static constexpr bool indexable = false;
+};
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::MIN>
+{
+    using opType = reduce::Min;
+
+    static constexpr bool indexable = true;
+};
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::MAX>
+{
+    using opType = reduce::Max;
+
+    static constexpr bool indexable = true;
+};
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::AMAX>
+{
+    using opType = reduce::AMax;
+
+    static constexpr bool indexable = true;
+};
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::AVG>
+{
+    using opType = reduce::Add;
+
+    static constexpr bool indexable = false;
+};
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::NORM1>
+{
+    using opType = reduce::Add;
+
+    static constexpr bool indexable = false;
+};
+
+template <>
+struct reduce_binary_operator<ReduceTensorOp::NORM2>
+{
+    using opType = reduce::Add;
+
+    static constexpr bool indexable = false;
+};
+
+// The templated struct reduce_unary_operator maps the enum Ids of Reduce operators to two unary
+// functor classes.
+// The two unary functors are called before and afer the Reduction is executed respectively
+template <ReduceTensorOp Op, bool IsFirstReduce, bool IsLastReduce>
+struct reduce_unary_operator
+{
+    using InElementwiseOperation  = tensor_operation::element_wise::PassThrough;
+    using AccElementwiseOperation = tensor_operation::element_wise::PassThrough;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
+};
+
+template <bool IsFirstReduce>
+struct reduce_unary_operator<ReduceTensorOp::AVG, IsFirstReduce, true>
+{
+    using InElementwiseOperation  = tensor_operation::element_wise::PassThrough;
+    using AccElementwiseOperation = tensor_operation::element_wise::UnaryDivide;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{reduceLength});
+    };
+};
+
+template <bool IsLastReduce>
+struct reduce_unary_operator<ReduceTensorOp::NORM1, true, IsLastReduce>
+{
+    using InElementwiseOperation  = tensor_operation::element_wise::UnaryAbs;
+    using AccElementwiseOperation = tensor_operation::element_wise::PassThrough;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
+};
+
+template <bool IsLastReduce>
+struct reduce_unary_operator<ReduceTensorOp::AMAX, true, IsLastReduce>
+{
+    using InElementwiseOperation  = tensor_operation::element_wise::UnaryAbs;
+    using AccElementwiseOperation = tensor_operation::element_wise::PassThrough;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
+};
+
+template <>
+struct reduce_unary_operator<ReduceTensorOp::NORM2, true, false>
+{
+    using InElementwiseOperation  = tensor_operation::element_wise::UnarySquare;
+    using AccElementwiseOperation = tensor_operation::element_wise::PassThrough;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
+};
+
+template <>
+struct reduce_unary_operator<ReduceTensorOp::NORM2, true, true>
+{
+    using InElementwiseOperation  = tensor_operation::element_wise::UnarySquare;
+    using AccElementwiseOperation = tensor_operation::element_wise::UnarySqrt;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
+};
+
+template <>
+struct reduce_unary_operator<ReduceTensorOp::NORM2, false, true>
+{
+    using InElementwiseOperation  = tensor_operation::element_wise::PassThrough;
+    using AccElementwiseOperation = tensor_operation::element_wise::UnarySqrt;
+
+    static std::tuple<InElementwiseOperation, AccElementwiseOperation>
+    GetElementwiseOperator(int32_t reduceLength)
+    {
+        (void)reduceLength;
+        return std::make_tuple(InElementwiseOperation{}, AccElementwiseOperation{});
+    };
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/tensor_layout.hpp b/include/ck/tensor_operation/gpu/device/tensor_layout.hpp
new file mode 100644
index 0000000000..b2d141fd61
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/tensor_layout.hpp
@@ -0,0 +1,417 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_layout {
+
+struct BaseTensorLayout
+{
+};
+
+namespace gemm {
+
+struct RowMajor : public BaseTensorLayout
+{
+    static constexpr const char* name = "RowMajor";
+};
+
+struct ColumnMajor : public BaseTensorLayout
+{
+    static constexpr const char* name = "ColumnMajor";
+};
+} // namespace gemm
+
+namespace convolution {
+
+// input tensor
+// packed NCW/NCHW/NCDHW
+struct NCW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NCW";
+};
+
+struct NCHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NCHW";
+};
+
+struct NCDHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NCDHW";
+};
+
+// packed GNCW/GNCHW/GNCDHW
+struct GNCW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNCW";
+};
+
+struct GNCHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNCHW";
+};
+
+struct GNCDHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNCDHW";
+};
+
+// input tensor
+// packed NWC/NHWC/NDHWC
+struct NWC : public BaseTensorLayout
+{
+    static constexpr const char* name = "NWC";
+};
+
+struct NHWC : public BaseTensorLayout
+{
+    static constexpr const char* name = "NHWC";
+};
+
+struct NDHWC : public BaseTensorLayout
+{
+    static constexpr const char* name = "NDHWC";
+};
+
+// input tensor
+// packed GNWC/GNHWC/GNDHWC
+struct GNWC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNWC";
+};
+
+struct GNHWC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNHWC";
+};
+
+struct GNDHWC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNDHWC";
+};
+
+// for input bias
+struct GC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GC";
+};
+
+// input tensor
+// packed NWGC/NHWGC/NDHWGC
+struct NWGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "NWGC";
+};
+
+struct NHWGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "NHWGC";
+};
+
+struct NDHWGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "NDHWGC";
+};
+
+// input tensor
+// strided layout
+struct G_NW_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NW_C";
+};
+
+struct G_NHW_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NHW_C";
+};
+
+struct G_NDHW_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NDHW_C";
+};
+
+// for input bias
+struct G_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_C";
+};
+
+// weight tensor
+// packed KCX/KCYX/KCZYX
+struct KCX : public BaseTensorLayout
+{
+    static constexpr const char* name = "KCX";
+};
+
+struct KCYX : public BaseTensorLayout
+{
+    static constexpr const char* name = "KCYX";
+};
+
+struct KCZYX : public BaseTensorLayout
+{
+    static constexpr const char* name = "KCZYX";
+};
+
+// weight tensor
+// packed KCX/KCYX/KCZYX
+struct GKCX : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKCX";
+};
+
+struct GKCYX : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKCYX";
+};
+
+struct GKCZYX : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKCZYX";
+};
+
+// weight tensor
+// packed KXC/KYXC/KZYXC
+struct KXC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KXC";
+};
+
+struct KYXC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KYXC";
+};
+
+struct KZYXC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KZYXC";
+};
+
+// weight tensor
+// packed GKXC/GKYXC/GKZYXC
+struct GKXC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKXC";
+};
+
+struct GKYXC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKYXC";
+};
+
+struct GKZYXC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKZYXC";
+};
+
+// weight tensor
+// packed KXGC/KYXGC/KZYXGC
+struct KXGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KXGC";
+};
+
+struct KYXGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KYXGC";
+};
+
+struct KZYXGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KZYXGC";
+};
+
+// weight tensor
+// strided
+struct G_K_X_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_K_X_C";
+};
+
+struct G_K_YX_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_K_YX_C";
+};
+
+struct G_K_ZYX_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_K_ZYX_C";
+};
+
+// output tensor
+// packed NKW/NKHW/NKDHW
+struct NKW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NKW";
+};
+
+struct NKHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NKHW";
+};
+
+struct NKDHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NKDHW";
+};
+
+// output tensor
+// packed GNKW/GNKHW/GNKDHW
+struct GNKW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNKW";
+};
+
+struct GNKHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNKHW";
+};
+
+struct GNKDHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNKDHW";
+};
+
+// output tensor
+// packed NWK/NHWK/NDHWK
+struct NWK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NWK";
+};
+
+struct NHWK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NHWK";
+};
+
+struct NDHWK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NDHWK";
+};
+
+// output tensor
+// packed GNWK/GNHWK/GNDHWK
+struct GNWK : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNWK";
+};
+
+struct GNHWK : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNHWK";
+};
+
+struct GNDHWK : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNDHWK";
+};
+
+// for output bias
+struct GK : public BaseTensorLayout
+{
+    static constexpr const char* name = "GK";
+};
+
+// output tensor
+// packed NWGK/NHWGK/NDHWGK
+struct NWGK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NWGK";
+};
+
+struct NHWGK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NHWGK";
+};
+
+struct NDHWGK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NDHWGK";
+};
+
+// output tensor
+// strided layout
+struct G_NW_K : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NW_K";
+};
+
+struct G_NHW_K : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NHW_K";
+};
+
+struct G_NDHW_K : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NDHW_K";
+};
+
+// for output bias
+struct G_K : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_K";
+};
+
+// K-reduced output tensor (packed)
+struct GNW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNW";
+};
+
+struct GNHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNHW";
+};
+
+struct GNDHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNDHW";
+};
+
+// K-reduced output tensor (packed)
+struct NWG : public BaseTensorLayout
+{
+    static constexpr const char* name = "NWG";
+};
+
+struct NHWG : public BaseTensorLayout
+{
+    static constexpr const char* name = "NHWG";
+};
+
+struct NDHWG : public BaseTensorLayout
+{
+    static constexpr const char* name = "NDHWG";
+};
+
+// K-reduced output tensor (strided)
+struct G_NW : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NW";
+};
+
+struct G_NHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NHW";
+};
+
+struct G_NDHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NDHW";
+};
+
+} // namespace convolution
+
+template <
+    typename Layout,
+    typename std::enable_if<std::is_base_of<BaseTensorLayout, Layout>::value, bool>::type = false>
+std::ostream& operator<<(std::ostream& os, const Layout&)
+{
+    os << Layout::name;
+    return os;
+}
+
+} // namespace tensor_layout
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/tensor_specialization.hpp b/include/ck/tensor_operation/gpu/device/tensor_specialization.hpp
new file mode 100644
index 0000000000..713fc93ebb
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/tensor_specialization.hpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum struct TensorSpecialization
+{
+    Default,
+    Packed
+};
+
+inline std::string getTensorSpecializationString(const TensorSpecialization& s)
+{
+    switch(s)
+    {
+    case TensorSpecialization::Default: return "Default";
+    case TensorSpecialization::Packed: return "Packed";
+    default: return "Unrecognized specialization!";
+    }
+}
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/device/welford_helper.hpp b/include/ck/tensor_operation/gpu/device/welford_helper.hpp
new file mode 100644
index 0000000000..d7772d8764
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/device/welford_helper.hpp
@@ -0,0 +1,89 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t K_BlockTileSize, index_t KThreadSliceSize>
+struct GetReduceCountPerThreadForBlockwiseWelford
+{
+    GetReduceCountPerThreadForBlockwiseWelford(index_t numBlockTileIteration,
+                                               long_index_t reduce_length)
+        : numBlockTileIteration_{numBlockTileIteration}
+    {
+        count_in_last_tile_ = reduce_length % K_BlockTileSize;
+    };
+
+    __device__ index_t operator()(index_t thread_k_cluster_id) const
+    {
+        if(count_in_last_tile_ == 0)
+            return (KThreadSliceSize * numBlockTileIteration_);
+        else
+        {
+            index_t num_complete_slice  = count_in_last_tile_ / KThreadSliceSize;
+            index_t count_in_last_slice = count_in_last_tile_ % KThreadSliceSize;
+
+            if(thread_k_cluster_id < num_complete_slice)
+                return (KThreadSliceSize * numBlockTileIteration_);
+            else if(thread_k_cluster_id == num_complete_slice)
+                return (KThreadSliceSize * (numBlockTileIteration_ - 1) + count_in_last_slice);
+            else
+                return (KThreadSliceSize * (numBlockTileIteration_ - 1));
+        };
+    };
+
+    index_t numBlockTileIteration_;
+    index_t count_in_last_tile_;
+};
+
+template <index_t K_BlockTileSize, index_t KThreadSliceSize>
+struct GetReduceCountPerThreadForMultiblockWelford
+{
+    GetReduceCountPerThreadForMultiblockWelford(index_t blkGroupSize,
+                                                index_t numBlockTileIteration,
+                                                long_index_t reduce_length)
+        : blkGroupSize_(blkGroupSize), numBlockTileIteration_{numBlockTileIteration}
+    {
+        last_block_reduce_length_ =
+            reduce_length - K_BlockTileSize * numBlockTileIteration_ * (blkGroupSize_ - 1);
+        numBlockTileIterationByLastBlock_ =
+            (last_block_reduce_length_ + K_BlockTileSize - 1) / K_BlockTileSize;
+    };
+
+    __device__ index_t operator()(index_t block_local_id, index_t thread_k_cluster_id) const
+    {
+        if(last_block_reduce_length_ == K_BlockTileSize * numBlockTileIteration_ ||
+           block_local_id < blkGroupSize_ - 1)
+            return (KThreadSliceSize * numBlockTileIteration_);
+
+        index_t count_in_last_tile = last_block_reduce_length_ % K_BlockTileSize;
+
+        if(count_in_last_tile == 0)
+            return (KThreadSliceSize * numBlockTileIterationByLastBlock_);
+        else
+        {
+            index_t num_complete_slice = count_in_last_tile / KThreadSliceSize;
+
+            if(thread_k_cluster_id < num_complete_slice)
+                return (KThreadSliceSize * numBlockTileIterationByLastBlock_);
+            else if(thread_k_cluster_id == num_complete_slice)
+                return (KThreadSliceSize * (numBlockTileIterationByLastBlock_ - 1) +
+                        count_in_last_tile);
+            else
+                return (KThreadSliceSize * (numBlockTileIterationByLastBlock_ - 1));
+        };
+    };
+
+    index_t blkGroupSize_;
+    index_t numBlockTileIteration_;
+
+    index_t last_block_reduce_length_;
+    index_t numBlockTileIterationByLastBlock_;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp b/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
new file mode 100644
index 0000000000..9ade02391f
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
@@ -0,0 +1,332 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace element_wise {
+
+struct Add
+{
+    template <typename Y, typename X0, typename X1>
+    __host__ __device__ constexpr void operator()(Y& y, const X0& x0, const X1& x1) const;
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float>(float& y, const float& x0, const float& x1) const
+    {
+        y = x0 + x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<double>(double& y, const double& x0, const double& x1) const
+    {
+        y = x0 + x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float>(float& y, const float& x0, const half_t& x1) const
+    {
+        y = x0 + type_convert<half_t>(x1);
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t>(half_t& y, const float& x0, const float& x1) const
+    {
+        y = type_convert<half_t>(x0 + x1);
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t>(half_t& y, const float& x0, const half_t& x1) const
+    {
+        y = type_convert<half_t>(x0) + x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    {
+        y = x0 + x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float>(float& y, const float& x0, const bhalf_t& x1) const
+    {
+        const float x1_tmp = ck::type_convert<float>(x1);
+        y                  = x0 + x1_tmp;
+    }
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<bhalf_t>(bhalf_t& y, const bhalf_t& x0, const bhalf_t& x1) const
+    {
+        const float x1_tmp = ck::type_convert<float>(x0);
+        const float x2_tmp = ck::type_convert<float>(x1);
+        const float y_tmp  = x1_tmp + x2_tmp;
+        y                  = ck::type_convert<bhalf_t>(y_tmp);
+    }
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<int8_t>(int8_t& y, const int8_t& x0, const int8_t& x1) const
+    {
+        y = x0 + x1;
+    };
+};
+
+struct ScaleAdd
+{
+    __host__ __device__ ScaleAdd(float scale) : scale_(scale) {}
+
+    template <typename Y, typename X0, typename X1>
+    __host__ __device__ constexpr void operator()(Y& y, const X0& x0, const X1& x1) const;
+
+    template <>
+    __host__ __device__ void
+    operator()<float, float, half_t>(float& y, const float& x0, const half_t& x1) const
+    {
+        y = scale_ * x0 + ck::type_convert<float>(x1);
+    };
+
+    template <>
+    __host__ __device__ void
+    operator()<float, float, bhalf_t>(float& y, const float& x0, const bhalf_t& x1) const
+    {
+        y = scale_ * x0 + ck::type_convert<float>(x1);
+    };
+
+    float scale_;
+};
+
+struct Subtract
+{
+    template <typename T>
+    __host__ __device__ constexpr void operator()(T& y, const T& x0, const T& x1) const;
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float>(float& y, const float& x0, const float& x1) const
+    {
+        y = x0 - x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<double>(double& y, const double& x0, const double& x1) const
+    {
+        y = x0 - x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    {
+        y = x0 - x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<bhalf_t>(bhalf_t& y, const bhalf_t& x0, const bhalf_t& x1) const
+    {
+        const float x1_tmp = ck::type_convert<float>(x0);
+        const float x2_tmp = ck::type_convert<float>(x1);
+        const float y_tmp  = x1_tmp - x2_tmp;
+        y                  = ck::type_convert<bhalf_t>(y_tmp);
+    }
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<int8_t>(int8_t& y, const int8_t& x0, const int8_t& x1) const
+    {
+        y = x0 - x1;
+    };
+};
+
+struct Bilinear
+{
+    Bilinear(float alpha, float beta) : alpha_(alpha), beta_(beta){};
+
+    template <typename Y, typename X0, typename X1>
+    __host__ __device__ constexpr void operator()(Y&, const X0&, const X1&) const;
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<double, double, double>(double& y, const double& x0, const double& x1) const
+    {
+        y = alpha_ * x0 + beta_ * x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float, float, float>(float& y, const float& x0, const float& x1) const
+    {
+        y = alpha_ * x0 + beta_ * x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, half_t, half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    {
+        y = type_convert<half_t>(alpha_) * x0 + type_convert<half_t>(beta_) * x1;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, float, half_t>(half_t& y, const float& x0, const half_t& x1) const
+    {
+        y = type_convert<half_t>(alpha_ * x0 + beta_ * ck::type_convert<float>(x1));
+    };
+
+    float alpha_;
+    float beta_;
+};
+
+struct AddRelu
+{
+    template <typename Y, typename X0, typename X1>
+    __host__ __device__ constexpr void operator()(Y& y, const X0& x0, const X1& x1) const;
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float, float, float>(float& y, const float& x0, const float& x1) const
+    {
+        const float a = x0 + x1;
+        y             = a > 0.0f ? a : 0.0f;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<double, double, double>(double& y, const double& x0, const double& x1) const
+    {
+        const double a = x0 + x1;
+        y              = a > 0.0 ? a : 0.0;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, half_t, half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    {
+        const half_t a = x0 + x1;
+        y              = a > type_convert<half_t>(0.0f) ? a : type_convert<half_t>(0.0f);
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, float, half_t>(half_t& y, const float& x0, const half_t& x1) const
+    {
+        const float a = x0 + x1;
+        y             = a > type_convert<half_t>(0.0f) ? a : type_convert<half_t>(0.0f);
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float, float, half_t>(float& y, const float& x0, const half_t& x1) const
+    {
+        const float a = x0 + type_convert<float>(x1);
+        y             = a > 0.0f ? a : 0.0f;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<int, int, int8_t>(int& y, const int& x0, const int8_t& x1) const
+    {
+        const int8_t a = x0 + x1;
+        y              = a > 0 ? a : 0;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<int8_t, int8_t, int8_t>(int8_t& y, const int8_t& x0, const int8_t& x1) const
+    {
+        const int8_t a = x0 + x1;
+        y              = a > 0 ? a : 0;
+    };
+};
+
+struct AddHardswish
+{
+    template <typename T>
+    __host__ __device__ constexpr void operator()(T& y, const T& x0, const T& x1) const;
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float>(float& y, const float& x0, const float& x1) const
+    {
+        float a = x0 + x1;
+        float b = a + float{3};
+        float c = (b > 0) * (b > 6.0f ? 6.0f : b) * a * 0.166667f;
+        y       = c;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<double>(double& y, const double& x0, const double& x1) const
+    {
+        double a = x0 + x1;
+        double b = a + 3.0;
+        double c = (b > 0) * (b > 6.0 ? 6.0 : b) * a * 0.166667;
+        y        = c;
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    {
+        float a = x0 + x1;
+        float b = a + 3.0f;
+        float c = (b > 0) * (b > 6.0f ? 6.0f : b) * a * 0.166667f;
+        y       = c;
+    };
+};
+
+// E = FastGelu(C + D)
+struct AddFastGelu
+{
+    template <typename E, typename C, typename D>
+    __host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<float, float, float>(float& e, const float& c, const float& d) const
+    {
+        const float x = c + d;
+
+        FastGelu{}.template operator()<float, float>(e, x);
+    }
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, half_t, half_t>(half_t& e, const half_t& c, const half_t& d) const
+    {
+        const half_t x = c + d;
+
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<half_t, half_t>(e, x);
+    }
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, float, half_t>(half_t& e, const float& c, const half_t& d) const
+    {
+        const float x0_f = c + d;
+
+        float x1_f = 0;
+
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
+                                                                                         x0_f);
+
+        e = type_convert<half_t>(x1_f);
+    }
+};
+
+} // namespace element_wise
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/element/element_wise_operation.hpp b/include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
new file mode 100644
index 0000000000..9f5ed6adea
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
@@ -0,0 +1,439 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math_v2.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/quantization_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace element_wise {
+
+// Need to ensure compiler will fail if there is no matching candidate, instead of compiler
+// siliently do implicit type conversion
+//
+// Example:
+//
+// struct ExampleElementwiseOp
+// {
+//     template<typename Y, typename X>
+//     __host__ __device__ constexpr void
+//     operator()(Y&, const X) const;
+//
+//     template<>
+//     __host__ __device__ constexpr void
+//     operator()<half_t, half_t>(half_t& y, const half_t& x) const
+//     {
+//     }
+// };
+
+struct AddReluAdd
+{
+    template <typename Y, typename X0, typename X1, typename X2>
+    __host__ __device__ constexpr void operator()(Y&, const X0&, const X1&, const X2&) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, half_t, half_t, half_t>(
+        half_t& y, const half_t& x0, const half_t& x1, const half_t& x2) const
+    {
+        half_t a = x0 + x1;
+        half_t b = a > 0 ? a : 0;
+        y        = b + x2;
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<float, float, float, float>(float& y,
+                                                                              const float& x0,
+                                                                              const float& x1,
+                                                                              const float& x2) const
+    {
+        float a = x0 + x1;
+        float b = a > 0 ? a : 0;
+        float c = b + x2;
+        y       = c;
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, float, half_t, half_t>(
+        half_t& y, const float& x0, const half_t& x1, const half_t& x2) const
+    {
+        float a = x0 + x1;
+        float b = a > 0 ? a : 0;
+        float c = b + x2;
+        y       = c;
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<bhalf_t, float, bhalf_t, bhalf_t>(
+        bhalf_t& y, const float& x0, const bhalf_t& x1, const bhalf_t& x2) const
+    {
+        float a = x0 + x1;
+        float b = a > 0 ? a : 0;
+        float c = b + x2;
+        y       = c;
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<int8_t, int8_t, int8_t, int8_t>(
+        int8_t& y, const int8_t& x0, const int8_t& x1, const int8_t& x2) const
+    {
+        int32_t a = x0 + x1;
+        int32_t b = a > 0 ? a : 0;
+        int32_t c = b + x2;
+        y         = c;
+    }
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    template <>
+    __host__ __device__ constexpr void operator()<int4_t, int8_t, int4_t, int4_t>(
+        int4_t& y, const int8_t& x0, const int4_t& x1, const int4_t& x2) const
+    {
+        int32_t a = x0 + x1;
+        int32_t b = a > 0 ? a : 0;
+        int32_t c = b + x2;
+        y         = c;
+    }
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+};
+
+struct AddHardswishAdd
+{
+    template <typename Y, typename X0, typename X1, typename X2>
+    __host__ __device__ constexpr void operator()(Y&, const X0&, const X1&, const X2&) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<float, float, float, float>(float& y,
+                                                                              const float& x0,
+                                                                              const float& x1,
+                                                                              const float& x2) const
+    {
+        float a = x0 + x1;
+        float b = a + float{3};
+        float c = (b > 0) * (b > float{6} ? float{6} : b) * a * float{0.166667};
+        float d = c + x2;
+        y       = d;
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, half_t, half_t, half_t>(
+        half_t& y, const half_t& x0, const half_t& x1, const half_t& x2) const
+    {
+        float a = x0 + x1;
+        float b = a + float{3};
+        float c = (b > 0) * (b > float{6} ? float{6} : b) * a * float{0.166667};
+        float d = c + x2;
+        y       = d;
+    }
+};
+
+// C = A * B
+// E = C + D0 + D1
+struct AddAdd
+{
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ void operator()(E& e, const C& c, const D0& d0, const D1& d1) const
+    {
+        // Only support floating so far
+        static_assert(is_same<E, half_t>::value || is_same<E, float>::value ||
+                          is_same<E, double>::value,
+                      "Data type is not supported by this operation!");
+
+        static_assert(is_same<C, half_t>::value || is_same<C, float>::value ||
+                          is_same<C, double>::value,
+                      "Data type is not supported by this operation!");
+
+        static_assert(is_same<D0, half_t>::value || is_same<D0, float>::value ||
+                          is_same<D0, double>::value,
+                      "Data type is not supported by this operation!");
+
+        static_assert(is_same<D1, half_t>::value || is_same<D1, float>::value ||
+                          is_same<D1, double>::value,
+                      "Data type is not supported by this operation!");
+
+        const C y = c + type_convert<C>(d0) + type_convert<C>(d1);
+        e         = type_convert<E>(y);
+    }
+};
+
+// C = A * B
+// E = (C + D0) x D1
+struct AddMultiply
+{
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ void operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
+
+    template <>
+    __host__ __device__ void operator()<half_t, half_t, half_t, half_t>(half_t& e,
+                                                                        const half_t& c,
+                                                                        const half_t& d0,
+                                                                        const half_t& d1) const
+    {
+        const half_t y = (c + d0) * d1;
+        e              = y;
+    }
+    template <>
+    __host__ __device__ void operator()<half_t, float, half_t, half_t>(half_t& e,
+                                                                       const float& c,
+                                                                       const half_t& d0,
+                                                                       const half_t& d1) const
+    {
+        const half_t y = (type_convert<half_t>(c) + d0) * d1;
+        e              = y;
+    }
+    template <>
+    __host__ __device__ void operator()<float, float, half_t, half_t>(float& e,
+                                                                      const float& c,
+                                                                      const half_t& d0,
+                                                                      const half_t& d1) const
+    {
+        const float y = (c + d0) * d1;
+        e             = y;
+    }
+};
+
+// C = A * B
+// E = C x D0 + D1
+struct MultiplyAdd
+{
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ void operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
+
+    template <>
+    __host__ __device__ void operator()<half_t, half_t, half_t, half_t>(half_t& e,
+                                                                        const half_t& c,
+                                                                        const half_t& d0,
+                                                                        const half_t& d1) const
+    {
+        const half_t y = (c * d0) + d1;
+        e              = y;
+    }
+    template <>
+    __host__ __device__ void operator()<half_t, float, half_t, half_t>(half_t& e,
+                                                                       const float& c,
+                                                                       const half_t& d0,
+                                                                       const half_t& d1) const
+    {
+        const half_t y = type_convert<half_t>(c) * d0 + d1;
+        e              = y;
+    }
+    template <>
+    __host__ __device__ void operator()<float, float, half_t, half_t>(float& e,
+                                                                      const float& c,
+                                                                      const half_t& d0,
+                                                                      const half_t& d1) const
+    {
+        const float y = c * d0 + d1;
+        e             = y;
+    }
+    template <>
+    __host__ __device__ void operator()<half_t, float, float, float>(half_t& e,
+                                                                     const float& c,
+                                                                     const float& d0,
+                                                                     const float& d1) const
+    {
+        const float y = c * d0 + d1;
+        e             = y;
+    }
+};
+
+// E = FastGelu(C + D0 + D1)
+struct AddAddFastGelu
+{
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<float, float, float, float>(float& e,
+                                                                              const float& c,
+                                                                              const float& d0,
+                                                                              const float& d1) const
+    {
+        const float x = c + d0 + d1;
+
+        FastGelu{}.template operator()<float, float>(e, x);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, half_t, half_t, half_t>(
+        half_t& e, const half_t& c, const half_t& d0, const half_t& d1) const
+    {
+        const half_t x = c + d0 + d1;
+
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<half_t, half_t>(e, x);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, float, half_t, half_t>(
+        half_t& e, const float& c, const half_t& d0, const half_t& d1) const
+    {
+        const float x0_f = c + d0 + d1;
+
+        float x1_f = 0;
+
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
+                                                                                         x0_f);
+
+        e = type_convert<half_t>(x1_f);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<bhalf_t, float, bhalf_t, bhalf_t>(
+        bhalf_t& e, const float& c, const bhalf_t& d0, const bhalf_t& d1) const
+    {
+        const float x0_f = c + type_convert<float>(d0) + type_convert<float>(d1);
+
+        float x1_f = 0;
+
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
+                                                                                         x0_f);
+
+        e = type_convert<bhalf_t>(x1_f);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<int8_t, int32_t, int8_t, int8_t>(
+        int8_t& e, const int32_t& c, const int8_t& d0, const int8_t& d1) const
+    {
+        const float x0_f =
+            type_convert<float>(c) + type_convert<float>(d0) + type_convert<float>(d1);
+
+        float x1_f = 0;
+
+        ck::tensor_operation::element_wise::FastGelu{}.template operator()<float, float>(x1_f,
+                                                                                         x0_f);
+
+        e = type_convert<int8_t>(x1_f);
+    }
+};
+
+struct Normalize
+{
+    // FIXME: is double absolutely necessary?
+    Normalize(double epsilon = 1e-4) : epsilon_(epsilon) {}
+
+    template <typename T1, typename T2, typename T3>
+    __host__ __device__ constexpr void operator()(T1& y,
+                                                  const T1& x,
+                                                  const T2& mean,
+                                                  const T2& mean_square,
+                                                  const T3& gamma,
+                                                  const T3& beta) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, float, half_t>(half_t& y,
+                                                                         const half_t& x,
+                                                                         const float& mean,
+                                                                         const float& mean_square,
+                                                                         const half_t& gamma,
+                                                                         const half_t& beta) const
+    {
+        using ck::math::sqrt;
+
+        float variance = mean_square - (mean * mean);
+
+        float tmp_x     = type_convert<float>(x);
+        float tmp_gamma = type_convert<float>(gamma);
+        float tmp_beta  = type_convert<float>(beta);
+
+        float tmp_y =
+            ((tmp_x - mean) / sqrt(variance + type_convert<float>(epsilon_))) * tmp_gamma +
+            tmp_beta;
+
+        y = type_convert<half_t>(tmp_y);
+    };
+
+    template <>
+    __host__ __device__ constexpr void operator()<float, float, float>(float& y,
+                                                                       const float& x,
+                                                                       const float& mean,
+                                                                       const float& mean_square,
+                                                                       const float& gamma,
+                                                                       const float& beta) const
+    {
+        using ck::math::sqrt;
+
+        float variance = mean_square - (mean * mean);
+        y = ((x - mean) / sqrt(variance + type_convert<float>(epsilon_))) * gamma + beta;
+    };
+
+    template <>
+    __host__ __device__ constexpr void operator()<double, double, double>(double& y,
+                                                                          const double& x,
+                                                                          const double& mean,
+                                                                          const double& mean_square,
+                                                                          const double& gamma,
+                                                                          const double& beta) const
+    {
+        using ck::math::sqrt;
+
+        double variance = mean_square - (mean * mean);
+        y               = ((x - mean) / sqrt(variance + epsilon_)) * gamma + beta;
+    };
+
+    // FIXME: is double absolutely necessary?
+    double epsilon_;
+};
+
+// used by BatchNorm inference
+// y = gamma * (x-mean) / sqrt(epsilon+variance) + beta
+// The data type of mean and variance is used as AccDataType
+struct NormalizeInInfer
+{
+    NormalizeInInfer(double epsilon = 1e-4) : epsilon_(epsilon) {}
+
+    template <typename T1, typename T2, typename T3, typename T4>
+    __host__ __device__ constexpr void operator()(T1& y,
+                                                  const T1& x,
+                                                  const T2& mean,
+                                                  const T2& variance,
+                                                  const T3& gamma,
+                                                  const T4& beta) const
+    {
+        static_assert(std::is_same<T2, float>::value || std::is_same<T2, double>::value,
+                      "Data type is not supported by this operation!");
+
+        using ck::type_convert;
+        using ck::math::sqrt;
+
+        T2 tmp_x, tmp_y;
+
+        tmp_x = type_convert<T2>(x);
+
+        tmp_y = ((tmp_x - mean) / sqrt(variance + type_convert<T2>(epsilon_))) *
+                    type_convert<T2>(gamma) +
+                type_convert<T2>(beta);
+        y = type_convert<T1>(tmp_y);
+    };
+
+    double epsilon_;
+};
+
+template <typename Y, typename X>
+struct UnaryTypeConvert;
+
+template <>
+struct UnaryTypeConvert<float, ck::bhalf_t>
+{
+    __host__ __device__ void operator()(float& y, ck::bhalf_t& x) const
+    {
+        y = ck::type_convert<float, ck::bhalf_t>(x);
+    }
+};
+
+template <>
+struct UnaryTypeConvert<ck::bhalf_t, float>
+{
+    __host__ __device__ void operator()(ck::bhalf_t& y, float& x) const
+    {
+        y = ck::type_convert<ck::bhalf_t, float>(x);
+    }
+};
+
+} // namespace element_wise
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/element/quantization_operation.hpp b/include/ck/tensor_operation/gpu/element/quantization_operation.hpp
new file mode 100644
index 0000000000..fefa6c793f
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/element/quantization_operation.hpp
@@ -0,0 +1,286 @@
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+// #include "ck/utility/get_id.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace element_wise {
+
+// Y = Sy * Qy
+// W = Sw * Qw
+// X = Sx * Qx
+// B = Sb * Qb = Sw * Sx * Qb
+// Where X, W, Y are float32, Qx, Qw, Qy are int8
+// Sx, Sw, Sy are scale of x, w, y (float32), which is calculated from quantization range
+// Qb is int32, scale of B is Sw * Sx for convenient
+
+// Y = W @ X, where @ is convolution or matrix multiplication
+// Sy * Qy = Sw * Qw @ Sx * Qx
+// Qy = [(Sw*Sx)/Sy] * Qw @ Qx
+
+// For Activation function which is piecewise linear function, such as relu, leaky relu ...etc
+// Activation(Sy * Qy) = Sy * Activation(Qy)
+template <typename Activation>
+struct Activation_Mul_Clamp
+{
+    // Convolution + Activation (piecewise linear function)
+    // If an activation is piecewise linear function, then Activation(Sy * Qy) = Sy * Activation(Qy)
+    // Z = Activation(Y) = Activation(W @ X)
+    // Sz * Qz = Activation(Sy * Qy)
+    // Qz = Sy / Sz * Activation(Qy) = (Sw * Sx / Sz) * Activation(Qw @ Qx)
+
+    // requantScale_ = Sw * Sx / Sz
+    Activation_Mul_Clamp(float requantScale, Activation activationOp)
+        : requantScale_(requantScale), activationOp_(activationOp)
+    {
+    }
+
+    __host__ __device__ constexpr void operator()(int8_t& y, const int32_t& x) const
+    {
+        float y_fp32 = ck::type_convert<float>(x);
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(requantScale_ * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int8_t>(y_fp32);
+    }
+
+    __device__ constexpr void operator()(int32_t& y, const int32_t& x) const
+    {
+        // CAUSION - We might type_convert to int8 in threadwise copy
+        // eg. GridwiseGemmDlMultipleD_km_kn_mn
+        float y_fp32 = ck::type_convert<float>(x);
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(requantScale_ * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int32_t>(y_fp32);
+    }
+
+    __host__ constexpr void operator()(float& y, const float& x) const
+    {
+        // CAUSION - We might float in & float out in reference code
+        activationOp_(y, x);
+        y = math::clamp(requantScale_ * y, -128.f, 127.f);
+    }
+
+    float requantScale_;
+    Activation activationOp_;
+};
+
+// For Activation function which is non piecewise linear function, such as TanH, Sigmoid ...etc
+// If an activation is not piecewise linear function
+// then Activation(Sy * Qy) != Sy * Activation(Qy)
+template <typename Activation>
+struct Mul_Activation_Mul_Clamp
+{
+    // Convolution + Activation (non piecewise linear function)
+    // Z = Activation(Y) = Activation(W @ X)
+    // Sz * Qz = Activation(Sy * Qy)
+    // Qz = S1 * Activation[Sacc * (Qw @ Qx)]
+    // Where S1 = 1 / Sz, Sacc = Sw * Sx
+    Mul_Activation_Mul_Clamp(float scale_z_inv, float scaleAcc, Activation activationOp)
+        : scale_z_inv_(scale_z_inv), scaleAcc_(scaleAcc), activationOp_(activationOp)
+    {
+    }
+
+    __host__ __device__ constexpr void operator()(int8_t& y, const int32_t& x) const
+    {
+        float y_fp32 = ck::type_convert<float>(x);
+        y_fp32       = scaleAcc_ * y_fp32;
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(scale_z_inv_ * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int8_t>(y_fp32);
+    }
+
+    float scale_z_inv_;
+    float scaleAcc_;
+    Activation activationOp_;
+};
+
+// Conv Perchannel quantization + Activation function which is piecewise linear function, such as
+// relu, leaky relu ...etc
+// Activation(Sy * Qy) = Sy * Activation(Qy)
+template <typename Activation>
+struct Activation_Mul2_Clamp
+{
+    Activation_Mul2_Clamp(Activation activationOp) : activationOp_(activationOp) {}
+
+    __host__ __device__ constexpr void
+    operator()(int8_t& y, const int32_t& x, const float& requantScale) const
+    {
+        float y_fp32 = ck::type_convert<float>(x);
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(requantScale * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int8_t>(y_fp32);
+    }
+
+    __device__ constexpr void
+    operator()(int32_t& y, const int32_t& x, const float& requantScale) const
+    {
+        // CAUSION - We might type_convert to int8 in threadwise copy
+        // eg. GridwiseGemmDlMultipleD_km_kn_mn
+        float y_fp32 = ck::type_convert<float>(x);
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(requantScale * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int32_t>(y_fp32);
+    }
+
+    Activation activationOp_;
+};
+
+// For Activation function which is piecewise linear function, such as relu, leaky relu ...etc
+// Activation(Sy * Qy) = Sy * Activation(Qy)
+template <typename Activation>
+struct Add_Activation_Mul_Clamp
+{
+    // Convolution + bias
+    // Let Bias = B = Sw * Sx * Qb
+    // Where Qb is int32
+    // Y = W @ X + B
+    // Sy * Qy = Sw * Qw @ Sx * Qx + Sw * Sx * Qb
+    // Qy = [(Sw*Sx)/Sy] * (Qw @ Qx + Qb)
+
+    // For activation, Z = Activaiton(Y)
+    // Sz * Qz = Activation(Sy * Qy)
+    // Qz = Sy / Sz * Activation(Qy) = [(Sw*Sx)/Sz] * Activation(Qw @ Qx + Qb)
+    Add_Activation_Mul_Clamp(float requantScale, Activation activationOp)
+        : requantScale_(requantScale), activationOp_(activationOp)
+    {
+    }
+
+    __host__ __device__ constexpr void
+    operator()(int8_t& y, const int32_t& x, const int32_t& bias) const
+    {
+        float y_fp32 = ck::type_convert<float>(x + bias);
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(requantScale_ * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int8_t>(y_fp32);
+    }
+
+    __host__ __device__ constexpr void
+    operator()(int32_t& y, const int32_t& x, const int32_t& bias) const
+    {
+        // CAUSION - We might type_convert to int8 in threadwise copy
+        // eg. GridwiseGemmDlMultipleD_km_kn_mn
+        float y_fp32 = ck::type_convert<float>(x + bias);
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(requantScale_ * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int32_t>(y_fp32);
+    }
+
+    float requantScale_;
+    Activation activationOp_;
+};
+
+// Conv Perchannel quantization + Activation function which is piecewise linear function, such as
+// relu, leaky relu ...etc
+template <typename Activation>
+struct Add_Activation_Mul2_Clamp
+{
+    Add_Activation_Mul2_Clamp(Activation activationOp) : activationOp_(activationOp) {}
+
+    __host__ __device__ constexpr void
+    operator()(int8_t& y, const int32_t& x, const int32_t& bias, const float& requantScale) const
+    {
+        float y_fp32 = ck::type_convert<float>(x + bias);
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(requantScale * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int8_t>(y_fp32);
+    }
+
+    __host__ __device__ constexpr void
+    operator()(int32_t& y, const int32_t& x, const int32_t& bias, const float& requantScale) const
+    {
+        // CAUSION - We might type_convert to int8 in threadwise copy
+        // eg. GridwiseGemmDlMultipleD_km_kn_mn
+        float y_fp32 = ck::type_convert<float>(x + bias);
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(requantScale * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int32_t>(y_fp32);
+    }
+
+    Activation activationOp_;
+};
+
+// For Activation function which is non piecewise linear function, such as TanH, Sigmoid ...etc
+// If an activation is not piecewise linear function
+// then Activation(Sy * Qy) != Sy * Activation(Qy)
+template <typename Activation>
+struct Add_Mul_Activation_Mul_Clamp
+{
+    // Convolution + Activation (non piecewise linear function)
+    // Z = Activation(Y) = Activation(W @ X + B)
+    // Sz * Qz = Activation(Sy * Qy)
+    // Qz = S1 * Activation[Sacc * (Qw @ Qx + Qb)]
+    // Where S1 = 1 / Sz, Sacc = Sw * Sx
+    Add_Mul_Activation_Mul_Clamp(float scale_z_inv, float scaleAcc, Activation activationOp)
+        : scale_z_inv_(scale_z_inv), scaleAcc_(scaleAcc), activationOp_(activationOp)
+    {
+    }
+
+    __host__ __device__ constexpr void
+    operator()(int8_t& y, const int32_t& x, const int32_t& bias) const
+    {
+        float y_fp32 = ck::type_convert<float>(x + bias);
+        y_fp32       = scaleAcc_ * y_fp32;
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(scale_z_inv_ * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int8_t>(y_fp32);
+    }
+
+    __host__ __device__ constexpr void
+    operator()(int32_t& y, const int32_t& x, const int32_t& bias) const
+    {
+        // CAUSION - We might type_convert to int8 in threadwise copy
+        // eg. GridwiseGemmDlMultipleD_km_kn_mn
+        float y_fp32 = ck::type_convert<float>(x + bias);
+        y_fp32       = scaleAcc_ * y_fp32;
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(scale_z_inv_ * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int32_t>(y_fp32);
+    }
+
+    float scale_z_inv_;
+    float scaleAcc_;
+    Activation activationOp_;
+};
+
+// Conv Perchannel quantization + Activation function which is non piecewise linear function,
+// such as TanH, Sigmoid ...etc
+// If an activation is not piecewise linear function
+// then Activation(Sy *Qy) != Sy * Activation(Qy)
+template <typename Activation>
+struct Add_Mul2_Activation_Mul_Clamp
+{
+    Add_Mul2_Activation_Mul_Clamp(float scale_z_inv, Activation activationOp)
+        : scale_z_inv_(scale_z_inv), activationOp_(activationOp)
+    {
+    }
+
+    __host__ __device__ constexpr void
+    operator()(int8_t& y, const int32_t& x, const int32_t& bias, const float& scaleAcc) const
+    {
+        float y_fp32 = ck::type_convert<float>(x + bias);
+        y_fp32       = scaleAcc * y_fp32;
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(scale_z_inv_ * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int8_t>(y_fp32);
+    }
+
+    __host__ __device__ constexpr void
+    operator()(int32_t& y, const int32_t& x, const int32_t& bias, const float& scaleAcc) const
+    {
+        // CAUSION - We might type_convert to int8 in threadwise copy
+        // eg. GridwiseGemmDlMultipleD_km_kn_mn
+        float y_fp32 = ck::type_convert<float>(x + bias);
+        y_fp32       = scaleAcc * y_fp32;
+        activationOp_(y_fp32, y_fp32);
+        y_fp32 = math::clamp(scale_z_inv_ * y_fp32, -128.f, 127.f);
+        y      = ck::type_convert<int32_t>(y_fp32);
+    }
+
+    float scale_z_inv_;
+    Activation activationOp_;
+};
+
+} // namespace element_wise
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp b/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
new file mode 100644
index 0000000000..905908a1c3
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
@@ -0,0 +1,430 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math.hpp"
+#include "ck/utility/math_v2.hpp"
+#include "ck/utility/type_convert.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace element_wise {
+
+#if CK_WORKAROUND_SWDEV_383542
+extern "C" __device__ float __ocml_native_recip_f32(float);
+#endif
+
+struct PassThrough
+{
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const;
+
+    template <>
+    __host__ __device__ void operator()<double, double>(double& y, const double& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<float, float>(float& y, const float& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<half_t, half_t>(half_t& y, const half_t& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<half_t, float>(half_t& y, const float& x) const
+    {
+        y = type_convert<half_t>(x);
+    }
+
+    template <>
+    __host__ __device__ void operator()<bhalf_t, bhalf_t>(bhalf_t& y, const bhalf_t& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<int32_t, int32_t>(int32_t& y, const int32_t& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<bhalf_t, float>(bhalf_t& y, const float& x) const
+    {
+        y = type_convert<bhalf_t>(x);
+    }
+
+    template <>
+    __host__ __device__ void operator()<bhalf_t, half_t>(bhalf_t& y, const half_t& x) const
+    {
+        y = type_convert<bhalf_t>(x);
+    }
+
+    template <>
+    __host__ __device__ void operator()<int8_t, int8_t>(int8_t& y, const int8_t& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<int8_t, int32_t>(int8_t& y, const int32_t& x) const
+    {
+        y = type_convert<int8_t>(x);
+    }
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    template <>
+    __host__ __device__ void operator()<int4_t, int4_t>(int4_t& y, const int4_t& x) const
+    {
+        y = x;
+    }
+#endif
+
+    template <>
+    __host__ __device__ void operator()<f8_t, f8_t>(f8_t& y, const f8_t& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<float, f8_t>(float& y, const f8_t& x) const
+    {
+        y = type_convert<float>(x);
+    }
+
+    template <>
+    __host__ __device__ void operator()<f8_t, float>(f8_t& y, const float& x) const
+    {
+        y = type_convert<f8_t>(x);
+    }
+
+    template <>
+    __host__ __device__ void operator()<half_t, f8_t>(half_t& y, const f8_t& x) const
+    {
+        y = type_convert<half_t>(x);
+    }
+
+    template <>
+    __host__ __device__ void operator()<f8_t, half_t>(f8_t& y, const half_t& x) const
+    {
+        y = type_convert<f8_t>(x);
+    }
+};
+
+struct UnaryConvert
+{
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const
+    {
+        y = type_convert<Y>(x);
+    }
+};
+
+struct ConvertBF16RTN
+{
+    // convert to bf16 using round to nearest (rtn)
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const
+    {
+        // check Y datatype
+        static_assert(is_same<Y, bhalf_t>::value, "Data type is not supported by this operation!");
+
+        // check X datatype
+        static_assert(is_same<X, float>::value || is_same<X, half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = bf16_convert_rtn<Y>(x);
+    }
+};
+
+struct ConvertF8SR
+{
+    // convert to fp8 using stochastic rounding (SR)
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const
+    {
+        // check Y datatype
+        static_assert(is_same<Y, f8_t>::value, "Data type is not supported by this operation!");
+
+        // check X datatype
+        static_assert(is_same<X, float>::value || is_same<X, half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = f8_convert_sr<Y>(x);
+    }
+};
+
+struct Scale
+{
+    __host__ __device__ Scale(float scale) : scale_(scale) {}
+
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const;
+
+    template <>
+    __host__ __device__ void operator()<float, float>(float& y, const float& x) const
+    {
+        y = scale_ * x;
+    };
+
+    template <>
+    __host__ __device__ void operator()<double, double>(double& y, const double& x) const
+    {
+        y = scale_ * x;
+    };
+
+    float scale_;
+};
+
+struct ScaleAndResetNaNToMinusInfinity
+{
+    __host__ __device__ ScaleAndResetNaNToMinusInfinity(float scale) : scale_(scale) {}
+
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const;
+
+    template <>
+    __host__ __device__ void operator()<float, float>(float& y, const float& x) const
+    {
+        y = ck::math::isnan(x) ? -ck::NumericLimits<float>::Infinity() : scale_ * x;
+    };
+
+    float scale_;
+};
+
+struct UnaryDivide
+{
+    __host__ __device__ UnaryDivide(const int32_t divider = 1) : divider_(divider) {}
+
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, int32_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = x / type_convert<T>(divider_);
+    };
+
+    int32_t divider_ = 1;
+};
+
+struct UnarySquare
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same_v<T, float> || is_same_v<T, double> || is_same_v<T, int32_t> ||
+                          is_same_v<T, int8_t>
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+                          || is_same_v<T, int4_t>
+#endif
+                      ,
+                      "Data type is not supported by this operation!");
+        y = x * x;
+    };
+};
+
+struct UnaryAbs
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = ck::math::abs(x);
+    };
+};
+
+struct UnarySqrt
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value,
+                      "Data type is not supported by this operation!");
+
+        y = ck::math::sqrt(x);
+    };
+};
+
+struct Relu
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "Data type is not supported by this operation!");
+        y = x > 0 ? x : 0;
+    }
+
+    template <>
+    __host__ __device__ void operator()(bhalf_t& y, const bhalf_t& x) const
+    {
+        float x_f32 = ck::type_convert<float>(x);
+        float y_f32 = x_f32 > 0 ? x_f32 : 0;
+        y           = ck::type_convert<bhalf_t>(y_f32);
+    }
+};
+
+// Fast GeLU
+// https://paperswithcode.com/method/gelu
+// y = 0.5*x*(1+tanh(sqrt(2/pi)*(x+0.044715*x^3)))
+// host code use higher accuracy "exp" and "div"
+// gpu code use lower accuracy "__expf" and "rcp" function
+struct FastGelu
+{
+    template <typename Y, typename X>
+    __host__ void operator()(Y& y, const X& x) const;
+
+    template <typename Y, typename X>
+    __device__ void operator()(Y& y, const X& x) const;
+
+    template <>
+    __host__ void operator()<float, float>(float& y, const float& x) const
+    {
+        const float u   = 2.f * x * (0.035677f * x * x + 0.797885f);
+        const float emu = exp(-u);
+        const float cdf = 0.5f + 0.5f * (2.f / (1.f + emu) - 1.f);
+
+        y = x * cdf;
+    }
+
+    // device code, use lower precision "__expf" and "rcp"
+    template <>
+    __device__ void operator()<float, float>(float& y, const float& x) const
+    {
+        const float u   = 2.f * x * (0.035677f * x * x + 0.797885f);
+        const float emu = __expf(-u);
+
+#if !CK_WORKAROUND_SWDEV_383542
+        const float cdf = 0.5f + 0.5f * (2.f * __frcp_rn(1.f + emu) - 1.f);
+#else
+        const float cdf = 0.5f + 0.5f * (2.f * __ocml_native_recip_f32(1.f + emu) - 1.f);
+#endif
+
+        y = x * cdf;
+    }
+
+    template <>
+    __host__ void operator()<half_t, half_t>(half_t& y, const half_t& x) const
+    {
+        float y_f;
+
+        this->operator()<float, float>(y_f, type_convert<float>(x));
+
+        y = type_convert<half_t>(y_f);
+    }
+
+    template <>
+    __device__ void operator()<half_t, half_t>(half_t& y, const half_t& x) const
+    {
+        float y_f;
+
+        this->operator()<float, float>(y_f, type_convert<float>(x));
+
+        y = type_convert<half_t>(y_f);
+    }
+
+    template <>
+    __host__ void operator()<half_t, float>(half_t& y, const float& x) const
+    {
+        float y_f;
+
+        this->operator()<float, float>(y_f, x);
+
+        y = type_convert<half_t>(y_f);
+    }
+
+    template <>
+    __device__ void operator()<half_t, float>(half_t& y, const float& x) const
+    {
+        float y_f;
+
+        this->operator()<float, float>(y_f, x);
+
+        y = type_convert<half_t>(y_f);
+    }
+};
+
+// https://paperswithcode.com/method/gelu
+// y = 0.5*x*(1+erf(x/sqrt(2)))
+struct Gelu
+{
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const;
+
+    template <>
+    __host__ __device__ void operator()<float, float>(float& y, const float& x) const
+    {
+        y = 0.5f * x * (1.f + erf(float(0.70710678118f * x)));
+    }
+
+    template <>
+    __host__ __device__ void operator()<ck::half_t, ck::half_t>(ck::half_t& y,
+                                                                const ck::half_t& x) const
+    {
+        y = ck::half_t(0.5) * x * (ck::half_t(1) + ck::half_t(erf(float(0.70710678118f * x))));
+    }
+};
+
+struct Sigmoid
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, ck::half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = 1 / (ck::type_convert<T>(1) + exp(-x));
+    };
+};
+
+struct TanH
+{
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, ck::half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = ck::math::tanh(x);
+    };
+};
+
+struct Swish
+{
+    Swish(float beta = 1.0f) : beta_(beta) {}
+
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, ck::half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = x / (ck::type_convert<T>(1) + ck::math::exp(-beta_ * x));
+    };
+
+    float beta_ = 1.0f;
+};
+
+} // namespace element_wise
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_batchnorm_forward.hpp b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_batchnorm_forward.hpp
new file mode 100644
index 0000000000..47573107cf
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_batchnorm_forward.hpp
@@ -0,0 +1,704 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math_v2.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/utility/workgroup_synchronization.hpp"
+
+namespace ck {
+
+template <typename GridwiseMultiblockBatchNormForward_,
+          typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename MeanVarCountGridDesc_M_G,
+          typename MeanVarCountGridDesc_M_K,
+          typename ScaleBiasGridDesc_M,
+          typename MeanVarGridDesc_M,
+          typename GetReduceCountPerThreadFunctor>
+__global__ void kernel_multiblock_batchnorm_forward(
+    const XYGridDesc_M_K x_grid_desc_m_k,
+    const XYGridDesc_M_K y_grid_desc_m_k,
+    const MeanVarCountGridDesc_M_G mean_var_count_grid_desc_m_g,
+    const MeanVarCountGridDesc_M_K mean_var_count_grid_desc_m_k,
+    const ScaleBiasGridDesc_M scale_grid_desc_m,
+    const ScaleBiasGridDesc_M bias_grid_desc_m,
+    const MeanVarGridDesc_M mean_var_grid_desc_m,
+    const GetReduceCountPerThreadFunctor get_reduce_count_per_thread,
+    index_t num_k_block_tile_iteration,
+    AccDataType epsilon,
+    const XDataType* const __restrict__ p_x,
+    MeanVarDataType* const __restrict__ p_welford_mean,
+    MeanVarDataType* const __restrict__ p_welford_variance,
+    int32_t* const __restrict__ p_welford_count,
+    int32_t* const __restrict__ p_control,
+    const ScaleDataType* const __restrict__ p_scale,
+    const BiasDataType* const __restrict__ p_bias,
+    const YElementwiseOp y_elementwise_op,
+    YDataType* const __restrict__ p_y,
+    bool updateMovingAverage,
+    AccDataType averageFactor,
+    MeanVarDataType* const __restrict__ resultRunningMean,
+    MeanVarDataType* const __restrict__ resultRunningVariance,
+    bool saveMeanInvVariance,
+    MeanVarDataType* const __restrict__ resultSaveMean,
+    MeanVarDataType* const __restrict__ resultSaveInvVariance)
+{
+    GridwiseMultiblockBatchNormForward_::Run(x_grid_desc_m_k,
+                                             y_grid_desc_m_k,
+                                             mean_var_count_grid_desc_m_g,
+                                             mean_var_count_grid_desc_m_k,
+                                             scale_grid_desc_m,
+                                             bias_grid_desc_m,
+                                             mean_var_grid_desc_m,
+                                             get_reduce_count_per_thread,
+                                             num_k_block_tile_iteration,
+                                             epsilon,
+                                             p_x,
+                                             p_welford_mean,
+                                             p_welford_variance,
+                                             p_welford_count,
+                                             p_control,
+                                             p_scale,
+                                             p_bias,
+                                             y_elementwise_op,
+                                             p_y,
+                                             updateMovingAverage,
+                                             averageFactor,
+                                             resultRunningMean,
+                                             resultRunningVariance,
+                                             saveMeanInvVariance,
+                                             resultSaveMean,
+                                             resultSaveInvVariance);
+};
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename MeanVarCountGridDesc_M_G,
+          typename MeanVarCountGridDesc_M_K,
+          typename ScaleBiasGridDesc_M,
+          typename MeanVarGridDesc_M,
+          typename GetReduceCountPerThreadFunctor,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcYDstVectorDim,
+          index_t XSrcVectorSize,
+          index_t YDstVectorSize,
+          index_t ScaleSrcVectorSize,
+          index_t BiasSrcVectorSize,
+          index_t MeanVarSrcDstVectorSize>
+struct GridwiseMultiblockBatchNormForward
+{
+    static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcYDstVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % YDstVectorSize == 0) ||
+                      (XSrcYDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XSrcYDstVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadReduceSrcDesc_M_1 = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
+
+    using ThreadwiseWelford1 =
+        ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
+
+    using ThreadwiseWelford2 =
+        ThreadwiseWelfordMerge<AccDataType, ThreadReduceSrcDesc_M_1, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford1 = BlockwiseWelford<AccDataType,
+                                               BlockSize,
+                                               ThreadClusterLengths_M_K,
+                                               ThreadClusterArrangeOrder,
+                                               false>;
+
+    using BlockwiseWelford2 = BlockwiseWelford<AccDataType,
+                                               BlockSize,
+                                               ThreadClusterLengths_M_K,
+                                               ThreadClusterArrangeOrder,
+                                               true>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    __device__ static void Run(const XYGridDesc_M_K& x_grid_desc_m_k,
+                               const XYGridDesc_M_K& y_grid_desc_m_k,
+                               const MeanVarCountGridDesc_M_G& mean_var_count_grid_desc_m_g,
+                               const MeanVarCountGridDesc_M_K& mean_var_count_grid_desc_m_k,
+                               const ScaleBiasGridDesc_M& scale_grid_desc_m,
+                               const ScaleBiasGridDesc_M& bias_grid_desc_m,
+                               const MeanVarGridDesc_M& mean_var_grid_desc_m,
+                               const GetReduceCountPerThreadFunctor& get_reduce_count_per_thread,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType epsilon,
+                               const XDataType* const __restrict__ p_x,
+                               MeanVarDataType* const __restrict__ p_welford_mean,
+                               MeanVarDataType* const __restrict__ p_welford_variance,
+                               int32_t* const __restrict__ p_welford_count,
+                               int32_t* const __restrict__ p_control,
+                               const ScaleDataType* const __restrict__ p_scale,
+                               const BiasDataType* const __restrict__ p_bias,
+                               const YElementwiseOp y_elementwise_op,
+                               YDataType* const __restrict__ p_y,
+                               bool updateMovingAverage,
+                               AccDataType averageFactor,
+                               MeanVarDataType* const __restrict__ resultRunningMean,
+                               MeanVarDataType* const __restrict__ resultRunningVariance,
+                               bool saveMeanInvVariance,
+                               MeanVarDataType* const __restrict__ resultSaveMean,
+                               MeanVarDataType* const __restrict__ resultSaveInvVariance)
+    {
+        using ck::math::sqrt;
+
+        const index_t blkgroup_size = mean_var_count_grid_desc_m_g.GetLength(I1);
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / blkgroup_size;
+        const index_t block_local_id  = block_global_id % blkgroup_size;
+
+        if(block_local_id == 0)
+            gms_init(BlockSize / warpSize * blkgroup_size, &p_control[blkgroup_id * 2]);
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        using ThreadBufferLengths_M   = Sequence<MThreadSliceSize>;
+        using ThreadBufferLengths_M_1 = Sequence<MThreadSliceSize, 1>;
+
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m_1 = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<1>{}));
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            x_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true> count_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            tmp_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            tmp_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true> tmp_count_thread_buf;
+
+        const index_t reduceSizePerBlock = K_BlockTileSize * num_k_block_tile_iteration;
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  XYGridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcYDstVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             block_local_id * reduceSizePerBlock +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        constexpr auto xy_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileSize);
+
+        const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x, x_grid_desc_m_k.GetElementSpaceSize());
+
+        // Step 1: each workgroup does local welford reduction
+
+        auto threadwise_welford_1 = ThreadwiseWelford1();
+        threadwise_welford_1.max_count_ =
+            get_reduce_count_per_thread(block_local_id, thread_k_cluster_id);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
+            var_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+        });
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_val_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, xy_copy_fwd_step_m_k);
+            threadwise_welford_1.Run(x_thread_buf, mean_thread_buf, var_thread_buf);
+        }
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            count_thread_buf(I) = threadwise_welford_1.cur_count_;
+            BlockwiseWelford1::Run(mean_thread_buf(I), var_thread_buf(I), count_thread_buf(I));
+        });
+
+        // Step 2: each workgroup writes its local welford result to workspace memory
+
+        auto mean_global_val_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Global, AmdBufferCoherenceEnum::GLC>(
+                p_welford_mean, mean_var_count_grid_desc_m_g.GetElementSpaceSize());
+
+        auto var_global_val_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Global, AmdBufferCoherenceEnum::GLC>(
+                p_welford_variance, mean_var_count_grid_desc_m_g.GetElementSpaceSize());
+
+        auto count_global_val_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Global, AmdBufferCoherenceEnum::GLC>(
+                p_welford_count, mean_var_count_grid_desc_m_g.GetElementSpaceSize());
+
+        auto threadwise_mean_var_store_m_g =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               MeanVarDataType,
+                                               decltype(thread_buffer_desc_m_1),
+                                               MeanVarCountGridDesc_M_G,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M_1,
+                                               Sequence<0, 1>,
+                                               0,
+                                               1,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                mean_var_count_grid_desc_m_g,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 block_local_id),
+                PassThroughOp{});
+
+        auto threadwise_count_store_m_g =
+            ThreadwiseTensorSliceTransfer_v1r3<int32_t,
+                                               int32_t,
+                                               decltype(thread_buffer_desc_m_1),
+                                               MeanVarCountGridDesc_M_G,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M_1,
+                                               Sequence<0, 1>,
+                                               0,
+                                               1,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                mean_var_count_grid_desc_m_g,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 block_local_id),
+                PassThroughOp{});
+
+        if(thread_k_cluster_id == 0)
+        {
+            threadwise_mean_var_store_m_g.Run(thread_buffer_desc_m_1,
+                                              make_tuple(I0, I0),
+                                              mean_thread_buf,
+                                              mean_var_count_grid_desc_m_g,
+                                              mean_global_val_buf);
+
+            threadwise_mean_var_store_m_g.Run(thread_buffer_desc_m_1,
+                                              make_tuple(I0, I0),
+                                              var_thread_buf,
+                                              mean_var_count_grid_desc_m_g,
+                                              var_global_val_buf);
+
+            threadwise_count_store_m_g.Run(thread_buffer_desc_m_1,
+                                           make_tuple(I0, I0),
+                                           count_thread_buf,
+                                           mean_var_count_grid_desc_m_g,
+                                           count_global_val_buf);
+        };
+
+        gms_barrier(&p_control[blkgroup_id * 2]);
+
+        if(block_local_id == 0)
+            gms_reset(&p_control[blkgroup_id * 2]);
+
+        // Step 3: each workgroup reads welford results from workspace memory and does final welford
+        // reduction
+
+        auto threadwise_mean_var_load_m_k =
+            ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
+                                             AccDataType,
+                                             MeanVarCountGridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             Sequence<0, 1>,
+                                             0,
+                                             1,
+                                             1,
+                                             true>(
+                mean_var_count_grid_desc_m_k,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * 1));
+
+        auto threadwise_count_load_m_k =
+            ThreadwiseTensorSliceTransfer_v2<int32_t,
+                                             int32_t,
+                                             MeanVarCountGridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             Sequence<0, 1>,
+                                             0,
+                                             1,
+                                             1,
+                                             true>(
+                mean_var_count_grid_desc_m_k,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * 1));
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            mean_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+            var_thread_buf(I)   = type_convert<AccDataType>(0.0f);
+            count_thread_buf(I) = 0;
+        });
+
+        constexpr auto mean_var_count_read_fwd_step_m_k = make_multi_index(0, KThreadClusterSize);
+
+        int32_t reducedSize = 0;
+        while(reducedSize < blkgroup_size)
+        {
+            threadwise_mean_var_load_m_k.Run(mean_var_count_grid_desc_m_k,
+                                             mean_global_val_buf,
+                                             thread_buffer_desc_m_1,
+                                             make_tuple(I0, I0),
+                                             tmp_mean_thread_buf);
+
+            threadwise_mean_var_load_m_k.Run(mean_var_count_grid_desc_m_k,
+                                             var_global_val_buf,
+                                             thread_buffer_desc_m_1,
+                                             make_tuple(I0, I0),
+                                             tmp_var_thread_buf);
+
+            threadwise_count_load_m_k.Run(mean_var_count_grid_desc_m_k,
+                                          count_global_val_buf,
+                                          thread_buffer_desc_m_1,
+                                          make_tuple(I0, I0),
+                                          tmp_count_thread_buf);
+
+            ThreadwiseWelford2::Run(tmp_mean_thread_buf,
+                                    tmp_var_thread_buf,
+                                    tmp_count_thread_buf,
+                                    mean_thread_buf,
+                                    var_thread_buf,
+                                    count_thread_buf);
+
+            reducedSize += KThreadClusterSize;
+
+            threadwise_mean_var_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k,
+                                                            mean_var_count_read_fwd_step_m_k);
+            threadwise_count_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k,
+                                                         mean_var_count_read_fwd_step_m_k);
+        };
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            BlockwiseWelford2::Run(mean_thread_buf(I), var_thread_buf(I), count_thread_buf(I));
+        });
+
+        // Step 4: do normalization using the mean/variance
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> scale_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> bias_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            y_thread_buf;
+
+        auto threadwise_y_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               YDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               XYGridDesc_M_K,
+                                               YElementwiseOp,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               XSrcYDstVectorDim,
+                                               YDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                y_grid_desc_m_k,
+                make_multi_index(
+                    blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                    block_local_id * reduceSizePerBlock + thread_k_cluster_id * KThreadSliceSize),
+                y_elementwise_op);
+
+        auto threadwise_scale_load =
+            ThreadwiseTensorSliceTransfer_v2<ScaleDataType,
+                                             AccDataType,
+                                             ScaleBiasGridDesc_M,
+                                             decltype(thread_buffer_desc_m),
+                                             ThreadBufferLengths_M,
+                                             Sequence<0>,
+                                             0,
+                                             ScaleSrcVectorSize,
+                                             1,
+                                             true>(
+                scale_grid_desc_m,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize));
+
+        auto threadwise_bias_load = ThreadwiseTensorSliceTransfer_v2<BiasDataType,
+                                                                     AccDataType,
+                                                                     ScaleBiasGridDesc_M,
+                                                                     decltype(thread_buffer_desc_m),
+                                                                     ThreadBufferLengths_M,
+                                                                     Sequence<0>,
+                                                                     0,
+                                                                     BiasSrcVectorSize,
+                                                                     1,
+                                                                     true>(
+            bias_grid_desc_m,
+            make_multi_index(blkgroup_id * M_BlockTileSize +
+                             thread_m_cluster_id * MThreadSliceSize));
+
+        const auto scale_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_scale, scale_grid_desc_m.GetElementSpaceSize());
+
+        const auto bias_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_bias, bias_grid_desc_m.GetElementSpaceSize());
+
+        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_y, y_grid_desc_m_k.GetElementSpaceSize());
+
+        threadwise_scale_load.Run(scale_grid_desc_m,
+                                  scale_global_val_buf,
+                                  thread_buffer_desc_m,
+                                  make_tuple(I0),
+                                  scale_thread_buf);
+
+        threadwise_bias_load.Run(bias_grid_desc_m,
+                                 bias_global_val_buf,
+                                 thread_buffer_desc_m,
+                                 make_tuple(I0),
+                                 bias_thread_buf);
+
+        threadwise_x_load.SetSrcSliceOrigin(
+            x_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             block_local_id * reduceSizePerBlock +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_val_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                AccDataType multiplier =
+                    scale_thread_buf[Number<iM>{}] / sqrt(var_thread_buf[iM] + epsilon);
+
+                AccDataType fused_mean_bias =
+                    bias_thread_buf[Number<iM>{}] - mean_thread_buf[iM] * multiplier;
+
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                    // normalize
+                    y_thread_buf(Number<offset>{}) =
+                        x_thread_buf[Number<offset>{}] * multiplier + fused_mean_bias;
+                });
+            });
+
+            threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                   make_tuple(I0, I0),
+                                   y_thread_buf,
+                                   y_grid_desc_m_k,
+                                   y_global_val_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, xy_copy_fwd_step_m_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, xy_copy_fwd_step_m_k);
+        }
+
+        // Step 5: update the moving average of mean and variance (optional)
+
+        if(updateMovingAverage && block_local_id == 0 && thread_k_cluster_id == 0)
+        {
+            StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+                running_mean_thread_buf;
+            StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+                running_var_thread_buf;
+
+            auto running_mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultRunningMean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto running_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultRunningVariance, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto threadwise_mean_var_load =
+                ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
+                                                 AccDataType,
+                                                 MeanVarGridDesc_M,
+                                                 decltype(thread_buffer_desc_m),
+                                                 ThreadBufferLengths_M,
+                                                 Sequence<0>,
+                                                 0,
+                                                 MeanVarSrcDstVectorSize,
+                                                 1,
+                                                 true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize));
+
+            threadwise_mean_var_load.Run(mean_var_grid_desc_m,
+                                         running_mean_global_buf,
+                                         thread_buffer_desc_m,
+                                         make_tuple(I0),
+                                         running_mean_thread_buf);
+
+            threadwise_mean_var_load.Run(mean_var_grid_desc_m,
+                                         running_var_global_buf,
+                                         thread_buffer_desc_m,
+                                         make_tuple(I0),
+                                         running_var_thread_buf);
+
+            AccDataType oneMinusAverageFactor = type_convert<AccDataType>(1.0) - averageFactor;
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                running_mean_thread_buf(I) = running_mean_thread_buf[I] * oneMinusAverageFactor +
+                                             mean_thread_buf[I] * averageFactor;
+                running_var_thread_buf(I) = running_var_thread_buf[I] * oneMinusAverageFactor +
+                                            var_thread_buf[I] * averageFactor;
+            });
+
+            auto threadwise_mean_var_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   MeanVarDataType,
+                                                   decltype(thread_buffer_desc_m),
+                                                   MeanVarGridDesc_M,
+                                                   PassThroughOp,
+                                                   ThreadBufferLengths_M,
+                                                   Sequence<0>,
+                                                   0,
+                                                   MeanVarSrcDstVectorSize,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_mean_var_store.Run(thread_buffer_desc_m,
+                                          make_tuple(I0),
+                                          running_mean_thread_buf,
+                                          mean_var_grid_desc_m,
+                                          running_mean_global_buf);
+
+            threadwise_mean_var_store.Run(thread_buffer_desc_m,
+                                          make_tuple(I0),
+                                          running_var_thread_buf,
+                                          mean_var_grid_desc_m,
+                                          running_var_global_buf);
+        };
+
+        // Step 6: save mean and inv-variance (optional)
+
+        if(saveMeanInvVariance && block_local_id == 0 && thread_k_cluster_id == 0)
+        {
+            auto result_mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultSaveMean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto result_inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultSaveInvVariance, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            // calculate inv-variance as 1/sqrt(epsilon+variance), stored in place of variance
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                var_thread_buf(I) =
+                    type_convert<AccDataType>(1.0f) / sqrt(epsilon + var_thread_buf[I]);
+            });
+
+            auto threadwise_mean_inv_var_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   MeanVarDataType,
+                                                   decltype(thread_buffer_desc_m),
+                                                   MeanVarGridDesc_M,
+                                                   PassThroughOp,
+                                                   ThreadBufferLengths_M,
+                                                   Sequence<0>,
+                                                   0,
+                                                   MeanVarSrcDstVectorSize,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              mean_thread_buf,
+                                              mean_var_grid_desc_m,
+                                              result_mean_global_buf);
+
+            threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              var_thread_buf,
+                                              mean_var_grid_desc_m,
+                                              result_inv_var_global_buf);
+        };
+    }
+}; // namespace ck
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp
new file mode 100644
index 0000000000..4e182ec29d
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp
@@ -0,0 +1,498 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseReduceSecondHalfBatchNormBackwardFinal_,
+          typename XDataType,
+          typename DyDataType,
+          typename DxDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename DscaleDbiasGridDesc_M_K,
+          typename MeanVarGridDesc_M,
+          typename ScaleBiasGridDesc_M>
+__global__ void kernel_reduce_second_half_batchnorm_backward_final(
+    const XYGridDesc_M_K x_grid_desc_m_k,
+    const XYGridDesc_M_K dy_grid_desc_m_k,
+    const XYGridDesc_M_K dx_grid_desc_m_k,
+    const DscaleDbiasGridDesc_M_K dscale_dbias_grid_desc_m_k,
+    const MeanVarGridDesc_M mean_var_grid_desc_m,
+    const ScaleBiasGridDesc_M scale_grid_desc_m,
+    const ScaleBiasGridDesc_M bias_grid_desc_m,
+    index_t blkgroup_size,
+    long_index_t reduce_size,
+    index_t num_xy_k_block_tile_iteration,
+    index_t num_dscale_dbias_k_block_tile_iteration,
+    const DscaleDbiasDataType* const __restrict__ p_reduce_dscale,
+    const DscaleDbiasDataType* const __restrict__ p_reduce_dbias,
+    const MeanVarDataType* const __restrict__ p_mean,
+    const MeanVarDataType* const __restrict__ p_inv_var,
+    const XDataType* const __restrict__ p_x,
+    const DyDataType* const __restrict__ p_dy,
+    const ScaleDataType* const __restrict__ p_scale,
+    const DyElementwiseOp dy_elementwise_op,
+    DxDataType* const __restrict__ p_dx,
+    DscaleDbiasDataType* const __restrict__ p_dscale,
+    DscaleDbiasDataType* const __restrict__ p_dbias)
+{
+    GridwiseReduceSecondHalfBatchNormBackwardFinal_::Run(x_grid_desc_m_k,
+                                                         dy_grid_desc_m_k,
+                                                         dx_grid_desc_m_k,
+                                                         dscale_dbias_grid_desc_m_k,
+                                                         mean_var_grid_desc_m,
+                                                         scale_grid_desc_m,
+                                                         bias_grid_desc_m,
+                                                         blkgroup_size,
+                                                         reduce_size,
+                                                         num_xy_k_block_tile_iteration,
+                                                         num_dscale_dbias_k_block_tile_iteration,
+                                                         p_reduce_dscale,
+                                                         p_reduce_dbias,
+                                                         p_mean,
+                                                         p_inv_var,
+                                                         p_x,
+                                                         p_dy,
+                                                         p_scale,
+                                                         dy_elementwise_op,
+                                                         p_dx,
+                                                         p_dscale,
+                                                         p_dbias);
+};
+
+template <typename XDataType,
+          typename DyDataType,
+          typename DxDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename DscaleDbiasGridDesc_M_K,
+          typename MeanVarGridDesc_M,
+          typename ScaleBiasGridDesc_M,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XDyDxVectorDim,
+          index_t XSrcVectorSize,
+          index_t DySrcVectorSize,
+          index_t DxDstVectorSize,
+          index_t ScaleSrcVectorSize,
+          index_t DscaleDbiasDstVectorSize,
+          index_t MeanVarSrcVectorSize>
+struct GridwiseReduceSecondHalfBatchNormBackwardFinal
+{
+    static_assert((XDyDxVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0 &&
+                   MThreadSliceSize % DySrcVectorSize == 0 &&
+                   MThreadSliceSize % DxDstVectorSize == 0) ||
+                      (XDyDxVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0 &&
+                       KThreadSliceSize % DySrcVectorSize == 0 &&
+                       KThreadSliceSize % DxDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XDyDxVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_1 = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using BlockwiseReduce = PartitionedBlockwiseReduction<AccDataType,
+                                                          BlockSize,
+                                                          ThreadClusterLengths_M_K,
+                                                          ThreadClusterArrangeOrder,
+                                                          ck::reduce::Add,
+                                                          false>;
+
+    using ThreadwiseReduce = ThreadwiseReduction<AccDataType,
+                                                 ThreadReduceSrcDesc_M_1,
+                                                 ThreadReduceDstDesc_M,
+                                                 ck::reduce::Add,
+                                                 false>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    // clang-format off
+    // Two of the steps of Multiblock BatchNorm Backward
+    // Step 1: Second half of Reduction: dbias = sum(dy), dscale = sum(dy * (x-mean) * inv-variance)
+    // Step 2: calculating dx = 1/reduce_size * inv-variance * scale * (reduce_size * dy - dbias - dscale * (x - mean) * inv-variance)) elementwise-ly
+    // clang-format on
+    __device__ static void Run(const XYGridDesc_M_K& x_grid_desc_m_k,
+                               const XYGridDesc_M_K& dy_grid_desc_m_k,
+                               const XYGridDesc_M_K& dx_grid_desc_m_k,
+                               const DscaleDbiasGridDesc_M_K& dscale_dbias_grid_desc_m_k,
+                               const MeanVarGridDesc_M& mean_var_grid_desc_m,
+                               const ScaleBiasGridDesc_M& scale_grid_desc_m,
+                               const ScaleBiasGridDesc_M& dscale_dbias_grid_desc_m,
+                               index_t blkgroup_size,
+                               long_index_t reduce_size,
+                               index_t num_xy_k_block_tile_iteration,
+                               index_t num_dscale_dbias_k_block_tile_iteration,
+                               const DscaleDbiasDataType* const __restrict__ p_reduce_dscale,
+                               const DscaleDbiasDataType* const __restrict__ p_reduce_dbias,
+                               const MeanVarDataType* const __restrict__ p_mean,
+                               const MeanVarDataType* const __restrict__ p_inv_var,
+                               const XDataType* const __restrict__ p_x,
+                               const DyDataType* const __restrict__ p_dy,
+                               const ScaleDataType* const __restrict__ p_scale,
+                               const DyElementwiseOp dy_elementwise_op,
+                               DxDataType* const __restrict__ p_dx,
+                               DscaleDbiasDataType* const __restrict__ p_dscale,
+                               DscaleDbiasDataType* const __restrict__ p_dbias)
+    {
+        __shared__ AccDataType p_reduce_work_buffer[BlockSize];
+
+        auto reduce_work_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * 1, true>
+            reduce_dscale_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * 1, true>
+            reduce_dbias_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> dscale_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> dbias_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            x_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            dy_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            dx_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            inv_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> scale_thread_buf;
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / blkgroup_size;
+        const index_t block_local_id  = block_global_id % blkgroup_size;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        using ThreadBufferLengths_M           = Sequence<MThreadSliceSize>;
+        using ThreadBufferLengths_M_1         = Sequence<MThreadSliceSize, 1>;
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m_1 = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<1>{}));
+
+        // clang-format off
+        // Step 1: do final reduction of dbias = sum(dy), dscale = sum(dy * (x-mean) * inv-variance)
+        // clang-format on
+
+        auto threadwise_dscale_dbias_load_m_k =
+            ThreadwiseTensorSliceTransfer_v2<DscaleDbiasDataType,
+                                             AccDataType,
+                                             DscaleDbiasGridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             Sequence<0, 1>,
+                                             1,
+                                             1,
+                                             1,
+                                             true>(
+                dscale_dbias_grid_desc_m_k,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * 1));
+
+        auto threadwise_dscale_dbias_store_m =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               DscaleDbiasDataType,
+                                               decltype(thread_buffer_desc_m),
+                                               ScaleBiasGridDesc_M,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M,
+                                               Sequence<0>,
+                                               0,
+                                               DscaleDbiasDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                dscale_dbias_grid_desc_m,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize),
+                PassThroughOp{});
+
+        const auto reduce_dscale_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_reduce_dscale, dscale_dbias_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto reduce_dbias_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_reduce_dbias, dscale_dbias_grid_desc_m_k.GetElementSpaceSize());
+
+        auto dscale_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dscale, dscale_dbias_grid_desc_m.GetElementSpaceSize());
+
+        auto dbias_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dbias, dscale_dbias_grid_desc_m.GetElementSpaceSize());
+
+        constexpr auto dscale_dbias_thread_copy_step_m_k =
+            make_multi_index(0, KThreadClusterSize * 1);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            dscale_thread_buf(I) = type_convert<AccDataType>(0.0f);
+            dbias_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+        });
+
+        for(index_t reducedTiles = 0; reducedTiles < num_dscale_dbias_k_block_tile_iteration;
+            ++reducedTiles)
+        {
+            threadwise_dscale_dbias_load_m_k.Run(dscale_dbias_grid_desc_m_k,
+                                                 reduce_dscale_global_buf,
+                                                 thread_buffer_desc_m_1,
+                                                 make_tuple(I0, I0),
+                                                 reduce_dscale_thread_buf);
+
+            threadwise_dscale_dbias_load_m_k.Run(dscale_dbias_grid_desc_m_k,
+                                                 reduce_dbias_global_buf,
+                                                 thread_buffer_desc_m_1,
+                                                 make_tuple(I0, I0),
+                                                 reduce_dbias_thread_buf);
+
+            ThreadwiseReduce::Reduce(reduce_dscale_thread_buf, dscale_thread_buf);
+            ThreadwiseReduce::Reduce(reduce_dbias_thread_buf, dbias_thread_buf);
+
+            threadwise_dscale_dbias_load_m_k.MoveSrcSliceWindow(dscale_dbias_grid_desc_m_k,
+                                                                dscale_dbias_thread_copy_step_m_k);
+        }
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            BlockwiseReduce::Reduce(reduce_work_buf, dscale_thread_buf(I));
+            block_sync_lds();
+            BlockwiseReduce::Reduce(reduce_work_buf, dbias_thread_buf(I));
+        });
+
+        threadwise_dscale_dbias_store_m.Run(thread_buffer_desc_m,
+                                            make_tuple(I0),
+                                            dscale_thread_buf,
+                                            dscale_dbias_grid_desc_m,
+                                            dscale_global_buf);
+
+        threadwise_dscale_dbias_store_m.Run(thread_buffer_desc_m,
+                                            make_tuple(I0),
+                                            dbias_thread_buf,
+                                            dscale_dbias_grid_desc_m,
+                                            dbias_global_buf);
+
+        // clang-format off
+        // Step 2: calculate dx = 1/N * inv-variance * scale * (N * dy - dbias - dscale * (x - mean) * inv-variance)
+        // clang-format on
+
+        const index_t workSizePerBlock = K_BlockTileSize * num_xy_k_block_tile_iteration;
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  XYGridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XDyDxVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             workSizePerBlock * block_local_id +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_dy_load = ThreadwiseTensorSliceTransfer_v2<DyDataType,
+                                                                   AccDataType,
+                                                                   XYGridDesc_M_K,
+                                                                   decltype(thread_buffer_desc_m_k),
+                                                                   ThreadBufferLengths_M_K,
+                                                                   ThreadBufferDimAccessOrder,
+                                                                   XDyDxVectorDim,
+                                                                   DySrcVectorSize,
+                                                                   1,
+                                                                   true>(
+            dy_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             workSizePerBlock * block_local_id +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_dx_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               DxDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               XYGridDesc_M_K,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               XDyDxVectorDim,
+                                               DxDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                dx_grid_desc_m_k,
+                make_multi_index(
+                    blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                    workSizePerBlock * block_local_id + thread_k_cluster_id * KThreadSliceSize),
+                PassThroughOp{});
+
+        auto threadwise_scale_load =
+            ThreadwiseTensorSliceTransfer_v2<ScaleDataType,
+                                             AccDataType,
+                                             ScaleBiasGridDesc_M,
+                                             decltype(thread_buffer_desc_m),
+                                             ThreadBufferLengths_M,
+                                             Sequence<0>,
+                                             0,
+                                             ScaleSrcVectorSize,
+                                             1,
+                                             true>(
+                scale_grid_desc_m,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize));
+
+        auto threadwise_mean_var_load =
+            ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
+                                             AccDataType,
+                                             MeanVarGridDesc_M,
+                                             decltype(thread_buffer_desc_m),
+                                             ThreadBufferLengths_M,
+                                             Sequence<0>,
+                                             0,
+                                             MeanVarSrcVectorSize,
+                                             1,
+                                             true>(
+                mean_var_grid_desc_m,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize));
+
+        const auto x_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto dy_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dy, dy_grid_desc_m_k.GetElementSpaceSize());
+
+        auto dx_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dx, dx_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto scale_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_scale, scale_grid_desc_m.GetElementSpaceSize());
+
+        const auto mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_mean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+        const auto inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_inv_var, mean_var_grid_desc_m.GetElementSpaceSize());
+
+        threadwise_scale_load.Run(scale_grid_desc_m,
+                                  scale_global_buf,
+                                  thread_buffer_desc_m,
+                                  make_tuple(I0),
+                                  scale_thread_buf);
+
+        threadwise_mean_var_load.Run(mean_var_grid_desc_m,
+                                     mean_global_buf,
+                                     thread_buffer_desc_m,
+                                     make_tuple(I0),
+                                     mean_thread_buf);
+
+        threadwise_mean_var_load.Run(mean_var_grid_desc_m,
+                                     inv_var_global_buf,
+                                     thread_buffer_desc_m,
+                                     make_tuple(I0),
+                                     inv_var_thread_buf);
+
+        constexpr auto xy_thread_copy_step_m_k = make_multi_index(0, K_BlockTileSize);
+
+        AccDataType inv_reduce_size =
+            type_convert<AccDataType>(1.0) / type_convert<AccDataType>(reduce_size);
+
+        for(index_t reducedTiles = 0; reducedTiles < num_xy_k_block_tile_iteration; ++reducedTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            threadwise_dy_load.Run(dy_grid_desc_m_k,
+                                   dy_global_buf,
+                                   thread_buffer_desc_m_k,
+                                   make_tuple(I0, I0),
+                                   dy_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                AccDataType multiplier =
+                    inv_reduce_size * inv_var_thread_buf[iM] * scale_thread_buf[iM];
+
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                    dy_elementwise_op(dy_thread_buf(Number<offset>{}),
+                                      dy_thread_buf[Number<offset>{}]);
+
+                    AccDataType norm_x = (x_thread_buf[Number<offset>{}] - mean_thread_buf[iM]) *
+                                         inv_var_thread_buf[iM];
+
+                    AccDataType tmpVal = norm_x * dscale_thread_buf[iM];
+
+                    dx_thread_buf(Number<offset>{}) =
+                        multiplier *
+                        (type_convert<AccDataType>(reduce_size) * dy_thread_buf[Number<offset>{}] -
+                         dbias_thread_buf[iM] - tmpVal);
+                });
+            });
+
+            threadwise_dx_store.Run(thread_buffer_desc_m_k,
+                                    make_tuple(I0, I0),
+                                    dx_thread_buf,
+                                    dx_grid_desc_m_k,
+                                    dx_global_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, xy_thread_copy_step_m_k);
+            threadwise_dy_load.MoveSrcSliceWindow(dy_grid_desc_m_k, xy_thread_copy_step_m_k);
+            threadwise_dx_store.MoveDstSliceWindow(dx_grid_desc_m_k, xy_thread_copy_step_m_k);
+        }
+    };
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp
new file mode 100644
index 0000000000..a82a173500
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp
@@ -0,0 +1,261 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseMultiblockWelfordFirstHalf_,
+          typename XDataType,
+          typename MeanVarDataType,
+          typename XGridDesc_M_K,
+          typename MeanVarCountGridDesc_M_G,
+          typename GetReduceCountPerThreadFunctor>
+__global__ void kernel_multiblock_welford_first_half(
+    const XGridDesc_M_K x_grid_desc_m_k,
+    const MeanVarCountGridDesc_M_G mean_var_count_grid_desc_m_g,
+    const GetReduceCountPerThreadFunctor get_reduce_count_per_thread,
+    index_t num_k_block_tile_iteration,
+    const XDataType* const __restrict__ p_x,
+    MeanVarDataType* const p_welford_mean,
+    MeanVarDataType* const p_welford_variance,
+    int32_t* const p_welford_count)
+{
+    GridwiseMultiblockWelfordFirstHalf_::Run(x_grid_desc_m_k,
+                                             mean_var_count_grid_desc_m_g,
+                                             get_reduce_count_per_thread,
+                                             num_k_block_tile_iteration,
+                                             p_x,
+                                             p_welford_mean,
+                                             p_welford_variance,
+                                             p_welford_count);
+};
+
+template <typename XDataType,
+          typename AccDataType,
+          typename MeanVarDataType,
+          typename XGridDesc_M_K,
+          typename MeanVarCountGridDesc_M_G,
+          typename GetReduceCountPerThreadFunctor,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcCountSrcVectorDim,
+          index_t XSrcCountSrcVectorSize>
+struct GridwiseMultiblockWelfordFirstHalf
+{
+    static_assert((XSrcCountSrcVectorDim == 0 && MThreadSliceSize % XSrcCountSrcVectorSize == 0) ||
+                      (XSrcCountSrcVectorDim == 1 &&
+                       KThreadSliceSize % XSrcCountSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XSrcCountSrcVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder,
+                                              false>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    // clang-format off
+    // First half of the Multiblock Welford method to calculate mean and variance, used by both batchnorm-forward and batchnorm-backward.
+    // clang-format on
+    __device__ static void Run(const XGridDesc_M_K& x_grid_desc_m_k,
+                               const MeanVarCountGridDesc_M_G& mean_var_count_grid_desc_m_g,
+                               const GetReduceCountPerThreadFunctor& get_reduce_count_per_thread,
+                               index_t num_k_block_tile_iteration,
+                               const XDataType* const __restrict__ p_x,
+                               MeanVarDataType* const p_welford_mean,
+                               MeanVarDataType* const p_welford_variance,
+                               int32_t* const p_welford_count)
+    {
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            x_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            welford_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true>
+            welford_count_thread_buf;
+
+        const index_t blkgroup_size = mean_var_count_grid_desc_m_g.GetLength(I1);
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / blkgroup_size;
+        const index_t block_local_id  = block_global_id % blkgroup_size;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        using ThreadBufferLengths_M_1 = Sequence<MThreadSliceSize, 1>;
+
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m_1 = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<1>{}));
+
+        const index_t reduceSizePerBlock = K_BlockTileSize * num_k_block_tile_iteration;
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  XGridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcCountSrcVectorDim,
+                                                                  XSrcCountSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             block_local_id * reduceSizePerBlock +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_welford_mean_var_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               MeanVarDataType,
+                                               decltype(thread_buffer_desc_m_1),
+                                               MeanVarCountGridDesc_M_G,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M_1,
+                                               Sequence<0, 1>,
+                                               0,
+                                               1,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                mean_var_count_grid_desc_m_g,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 block_local_id),
+                PassThroughOp{});
+
+        auto threadwise_welford_count_store =
+            ThreadwiseTensorSliceTransfer_v1r3<int32_t,
+                                               int32_t,
+                                               decltype(thread_buffer_desc_m_1),
+                                               MeanVarCountGridDesc_M_G,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M_1,
+                                               Sequence<0, 1>,
+                                               0,
+                                               1,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                mean_var_count_grid_desc_m_g,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 block_local_id),
+                PassThroughOp{});
+
+        constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileSize);
+
+        const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x, x_grid_desc_m_k.GetElementSpaceSize());
+
+        auto welford_mean_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_welford_mean, mean_var_count_grid_desc_m_g.GetElementSpaceSize());
+
+        auto welford_var_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_welford_variance, mean_var_count_grid_desc_m_g.GetElementSpaceSize());
+
+        auto welford_count_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_welford_count, mean_var_count_grid_desc_m_g.GetElementSpaceSize());
+
+        auto threadwise_welford = ThreadwiseWelford();
+        threadwise_welford.max_count_ =
+            get_reduce_count_per_thread(block_local_id, thread_k_cluster_id);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            welford_mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
+            welford_var_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+        });
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_val_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+            threadwise_welford.Run(x_thread_buf, welford_mean_thread_buf, welford_var_thread_buf);
+        }
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            welford_count_thread_buf(I) = threadwise_welford.cur_count_;
+            BlockwiseWelford::Run(
+                welford_mean_thread_buf(I), welford_var_thread_buf(I), welford_count_thread_buf(I));
+        });
+
+        if(thread_k_cluster_id == 0)
+        {
+            threadwise_welford_mean_var_store.Run(thread_buffer_desc_m_1,
+                                                  make_tuple(I0, I0),
+                                                  welford_mean_thread_buf,
+                                                  mean_var_count_grid_desc_m_g,
+                                                  welford_mean_global_val_buf);
+
+            threadwise_welford_mean_var_store.Run(thread_buffer_desc_m_1,
+                                                  make_tuple(I0, I0),
+                                                  welford_var_thread_buf,
+                                                  mean_var_count_grid_desc_m_g,
+                                                  welford_var_global_val_buf);
+
+            threadwise_welford_count_store.Run(thread_buffer_desc_m_1,
+                                               make_tuple(I0, I0),
+                                               welford_count_thread_buf,
+                                               mean_var_count_grid_desc_m_g,
+                                               welford_count_global_val_buf);
+        };
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp
new file mode 100644
index 0000000000..672be91a79
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp
@@ -0,0 +1,570 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math_v2.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseWelfordSecondHalfBatchNormForwardFinal_,
+          typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename MeanVarCountGridDesc_M_K,
+          typename ScaleBiasGridDesc_M,
+          typename MeanVarGridDesc_M>
+__global__ void kernel_welford_second_half_batchnorm_forward_final(
+    const XYGridDesc_M_K x_grid_desc_m_k,
+    const XYGridDesc_M_K y_grid_desc_m_k,
+    const MeanVarCountGridDesc_M_K mean_var_count_grid_desc_m_k,
+    const ScaleBiasGridDesc_M scale_grid_desc_m,
+    const ScaleBiasGridDesc_M bias_grid_desc_m,
+    const MeanVarGridDesc_M mean_var_grid_desc_m,
+    index_t blkgroup_size,
+    index_t num_xy_k_block_tile_iteration,
+    AccDataType epsilon,
+    const MeanVarDataType* const __restrict__ p_in_welford_mean,
+    const MeanVarDataType* const __restrict__ p_in_welford_variance,
+    const int32_t* const __restrict__ p_in_welford_count,
+    const XDataType* const __restrict__ p_x,
+    const ScaleDataType* const __restrict__ p_scale,
+    const BiasDataType* const __restrict__ p_bias,
+    const YElementwiseOp y_elementwise_op,
+    YDataType* const __restrict__ p_y,
+    bool updateMovingAverage,
+    AccDataType averageFactor,
+    MeanVarDataType* const __restrict__ resultRunningMean,
+    MeanVarDataType* const __restrict__ resultRunningVariance,
+    bool saveMeanInvVariance,
+    MeanVarDataType* const __restrict__ resultSaveMean,
+    MeanVarDataType* const __restrict__ resultSaveInvVariance)
+{
+    GridwiseWelfordSecondHalfBatchNormForwardFinal_::Run(x_grid_desc_m_k,
+                                                         y_grid_desc_m_k,
+                                                         mean_var_count_grid_desc_m_k,
+                                                         scale_grid_desc_m,
+                                                         bias_grid_desc_m,
+                                                         mean_var_grid_desc_m,
+                                                         blkgroup_size,
+                                                         num_xy_k_block_tile_iteration,
+                                                         epsilon,
+                                                         p_in_welford_mean,
+                                                         p_in_welford_variance,
+                                                         p_in_welford_count,
+                                                         p_x,
+                                                         p_scale,
+                                                         p_bias,
+                                                         y_elementwise_op,
+                                                         p_y,
+                                                         updateMovingAverage,
+                                                         averageFactor,
+                                                         resultRunningMean,
+                                                         resultRunningVariance,
+                                                         saveMeanInvVariance,
+                                                         resultSaveMean,
+                                                         resultSaveInvVariance);
+};
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename MeanVarCountGridDesc_M_K,
+          typename ScaleBiasGridDesc_M,
+          typename MeanVarGridDesc_M,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcYDstVectorDim,
+          index_t XSrcVectorSize,
+          index_t YDstVectorSize,
+          index_t ScaleSrcVectorSize,
+          index_t BiasSrcVectorSize,
+          index_t MeanVarSrcDstVectorSize>
+struct GridwiseWelfordSecondHalfBatchNormForwardFinal
+{
+    static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcYDstVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % YDstVectorSize == 0) ||
+                      (XSrcYDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XSrcYDstVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_1 = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelfordMerge<AccDataType, ThreadReduceSrcDesc_M_1, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    __device__ static void Run(const XYGridDesc_M_K& x_grid_desc_m_k,
+                               const XYGridDesc_M_K& y_grid_desc_m_k,
+                               const MeanVarCountGridDesc_M_K& mean_var_count_grid_desc_m_k,
+                               const ScaleBiasGridDesc_M& scale_grid_desc_m,
+                               const ScaleBiasGridDesc_M& bias_grid_desc_m,
+                               const MeanVarGridDesc_M& mean_var_grid_desc_m,
+                               index_t blkgroup_size,
+                               index_t num_xy_k_block_tile_iteration,
+                               AccDataType epsilon,
+                               const MeanVarDataType* const __restrict__ p_in_welford_mean,
+                               const MeanVarDataType* const __restrict__ p_in_welford_variance,
+                               const int32_t* const __restrict__ p_in_welford_count,
+                               const XDataType* const __restrict__ p_x,
+                               const ScaleDataType* const __restrict__ p_scale,
+                               const BiasDataType* const __restrict__ p_bias,
+                               const YElementwiseOp y_elementwise_op,
+                               YDataType* const __restrict__ p_y,
+                               bool updateMovingAverage,
+                               AccDataType averageFactor,
+                               MeanVarDataType* const __restrict__ resultRunningMean,
+                               MeanVarDataType* const __restrict__ resultRunningVariance,
+                               bool saveMeanInvVariance,
+                               MeanVarDataType* const __restrict__ resultSaveMean,
+                               MeanVarDataType* const __restrict__ resultSaveInvVariance)
+
+    {
+        using ck::math::sqrt;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * 1, true>
+            in_welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * 1, true>
+            in_welford_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize * 1, true>
+            in_welford_count_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            welford_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true>
+            welford_count_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            x_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            y_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> scale_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> bias_thread_buf;
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / blkgroup_size;
+        const index_t block_local_id  = block_global_id % blkgroup_size;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        using ThreadBufferLengths_M           = Sequence<MThreadSliceSize>;
+        using ThreadBufferLengths_M_1         = Sequence<MThreadSliceSize, 1>;
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m_1 = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<1>{}));
+
+        auto threadwise_mean_var_load_m_k =
+            ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
+                                             AccDataType,
+                                             MeanVarCountGridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             Sequence<0, 1>,
+                                             0,
+                                             1,
+                                             1,
+                                             true>(
+                mean_var_count_grid_desc_m_k,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * 1));
+
+        auto threadwise_count_load_m_k =
+            ThreadwiseTensorSliceTransfer_v2<int32_t,
+                                             int32_t,
+                                             MeanVarCountGridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             Sequence<0, 1>,
+                                             0,
+                                             1,
+                                             1,
+                                             true>(
+                mean_var_count_grid_desc_m_k,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * 1));
+
+        const auto welford_mean_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_welford_mean, mean_var_count_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto welford_var_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_welford_variance, mean_var_count_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto welford_count_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_welford_count, mean_var_count_grid_desc_m_k.GetElementSpaceSize());
+
+        // Step 1: do final welford reduction to get mean and variance
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            welford_mean_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+            welford_var_thread_buf(I)   = type_convert<AccDataType>(0.0f);
+            welford_count_thread_buf(I) = 0;
+        });
+
+        constexpr auto mean_var_count_thread_copy_step_m_k =
+            make_multi_index(0, KThreadClusterSize);
+
+        int32_t reducedSize = 0;
+        while(reducedSize < blkgroup_size)
+        {
+            threadwise_mean_var_load_m_k.Run(mean_var_count_grid_desc_m_k,
+                                             welford_mean_global_val_buf,
+                                             thread_buffer_desc_m_1,
+                                             make_tuple(I0, I0),
+                                             in_welford_mean_thread_buf);
+
+            threadwise_mean_var_load_m_k.Run(mean_var_count_grid_desc_m_k,
+                                             welford_var_global_val_buf,
+                                             thread_buffer_desc_m_1,
+                                             make_tuple(I0, I0),
+                                             in_welford_var_thread_buf);
+
+            threadwise_count_load_m_k.Run(mean_var_count_grid_desc_m_k,
+                                          welford_count_global_val_buf,
+                                          thread_buffer_desc_m_1,
+                                          make_tuple(I0, I0),
+                                          in_welford_count_thread_buf);
+
+            ThreadwiseWelford::Run(in_welford_mean_thread_buf,
+                                   in_welford_var_thread_buf,
+                                   in_welford_count_thread_buf,
+                                   welford_mean_thread_buf,
+                                   welford_var_thread_buf,
+                                   welford_count_thread_buf);
+
+            reducedSize += KThreadClusterSize;
+
+            threadwise_mean_var_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k,
+                                                            mean_var_count_thread_copy_step_m_k);
+            threadwise_count_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k,
+                                                         mean_var_count_thread_copy_step_m_k);
+        }
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            BlockwiseWelford::Run(
+                welford_mean_thread_buf(I), welford_var_thread_buf(I), welford_count_thread_buf(I));
+        });
+
+        // Step 2: do normalization and output y
+
+        const index_t workSizePerBlock = K_BlockTileSize * num_xy_k_block_tile_iteration;
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  XYGridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcYDstVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             workSizePerBlock * block_local_id +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_y_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               YDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               XYGridDesc_M_K,
+                                               YElementwiseOp,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               XSrcYDstVectorDim,
+                                               YDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                y_grid_desc_m_k,
+                make_multi_index(
+                    blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                    workSizePerBlock * block_local_id + thread_k_cluster_id * KThreadSliceSize),
+                y_elementwise_op);
+
+        auto threadwise_scale_load =
+            ThreadwiseTensorSliceTransfer_v2<ScaleDataType,
+                                             AccDataType,
+                                             ScaleBiasGridDesc_M,
+                                             decltype(thread_buffer_desc_m),
+                                             ThreadBufferLengths_M,
+                                             Sequence<0>,
+                                             0,
+                                             ScaleSrcVectorSize,
+                                             1,
+                                             true>(
+                scale_grid_desc_m,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize));
+
+        auto threadwise_bias_load = ThreadwiseTensorSliceTransfer_v2<BiasDataType,
+                                                                     AccDataType,
+                                                                     ScaleBiasGridDesc_M,
+                                                                     decltype(thread_buffer_desc_m),
+                                                                     ThreadBufferLengths_M,
+                                                                     Sequence<0>,
+                                                                     0,
+                                                                     BiasSrcVectorSize,
+                                                                     1,
+                                                                     true>(
+            bias_grid_desc_m,
+            make_multi_index(blkgroup_id * M_BlockTileSize +
+                             thread_m_cluster_id * MThreadSliceSize));
+
+        const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto scale_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_scale, scale_grid_desc_m.GetElementSpaceSize());
+
+        const auto bias_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_bias, bias_grid_desc_m.GetElementSpaceSize());
+
+        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_y, y_grid_desc_m_k.GetElementSpaceSize());
+
+        threadwise_scale_load.Run(scale_grid_desc_m,
+                                  scale_global_val_buf,
+                                  thread_buffer_desc_m,
+                                  make_tuple(I0),
+                                  scale_thread_buf);
+
+        threadwise_bias_load.Run(bias_grid_desc_m,
+                                 bias_global_val_buf,
+                                 thread_buffer_desc_m,
+                                 make_tuple(I0),
+                                 bias_thread_buf);
+
+        constexpr auto xy_thread_copy_step_m_k = make_multi_index(0, K_BlockTileSize);
+
+        for(index_t workTiles = 0; workTiles < num_xy_k_block_tile_iteration; ++workTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_val_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                AccDataType multiplier =
+                    scale_thread_buf[iM] / sqrt(welford_var_thread_buf[iM] + epsilon);
+
+                AccDataType fused_mean_bias =
+                    bias_thread_buf[iM] - welford_mean_thread_buf[iM] * multiplier;
+
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                    y_thread_buf(Number<offset>{}) =
+                        x_thread_buf[Number<offset>{}] * multiplier + fused_mean_bias;
+                });
+            });
+
+            threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                   make_tuple(I0, I0),
+                                   y_thread_buf,
+                                   y_grid_desc_m_k,
+                                   y_global_val_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, xy_thread_copy_step_m_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, xy_thread_copy_step_m_k);
+        }
+
+        // Step 3: update the moving average of mean and variance (optional)
+
+        if(updateMovingAverage && block_local_id == 0 && thread_k_cluster_id == 0)
+        {
+            StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+                running_mean_thread_buf;
+            StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+                running_var_thread_buf;
+
+            auto running_mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultRunningMean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto running_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultRunningVariance, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto threadwise_mean_var_load_m =
+                ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
+                                                 AccDataType,
+                                                 MeanVarGridDesc_M,
+                                                 decltype(thread_buffer_desc_m),
+                                                 ThreadBufferLengths_M,
+                                                 Sequence<0>,
+                                                 0,
+                                                 MeanVarSrcDstVectorSize,
+                                                 1,
+                                                 true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize));
+
+            threadwise_mean_var_load_m.Run(mean_var_grid_desc_m,
+                                           running_mean_global_buf,
+                                           thread_buffer_desc_m,
+                                           make_tuple(I0),
+                                           running_mean_thread_buf);
+
+            threadwise_mean_var_load_m.Run(mean_var_grid_desc_m,
+                                           running_var_global_buf,
+                                           thread_buffer_desc_m,
+                                           make_tuple(I0),
+                                           running_var_thread_buf);
+
+            AccDataType oneMinusAverageFactor = type_convert<AccDataType>(1.0) - averageFactor;
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                running_mean_thread_buf(I) = running_mean_thread_buf[I] * oneMinusAverageFactor +
+                                             welford_mean_thread_buf[I] * averageFactor;
+                running_var_thread_buf(I) = running_var_thread_buf[I] * oneMinusAverageFactor +
+                                            welford_var_thread_buf[I] * averageFactor;
+            });
+
+            auto threadwise_mean_var_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   MeanVarDataType,
+                                                   decltype(thread_buffer_desc_m),
+                                                   MeanVarGridDesc_M,
+                                                   PassThroughOp,
+                                                   ThreadBufferLengths_M,
+                                                   Sequence<0>,
+                                                   0,
+                                                   MeanVarSrcDstVectorSize,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_mean_var_store.Run(thread_buffer_desc_m,
+                                          make_tuple(I0),
+                                          running_mean_thread_buf,
+                                          mean_var_grid_desc_m,
+                                          running_mean_global_buf);
+
+            threadwise_mean_var_store.Run(thread_buffer_desc_m,
+                                          make_tuple(I0),
+                                          running_var_thread_buf,
+                                          mean_var_grid_desc_m,
+                                          running_var_global_buf);
+        };
+
+        // Step 4: save mean and inv-variance (optional)
+
+        if(saveMeanInvVariance && block_local_id == 0 && thread_k_cluster_id == 0)
+        {
+            auto result_mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultSaveMean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto result_inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultSaveInvVariance, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            // calculate inv-variance as 1/sqrt(epsilon+variance)
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                welford_var_thread_buf(I) =
+                    type_convert<AccDataType>(1.0f) / sqrt(epsilon + welford_var_thread_buf[I]);
+            });
+
+            auto threadwise_mean_inv_var_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   MeanVarDataType,
+                                                   decltype(thread_buffer_desc_m),
+                                                   MeanVarGridDesc_M,
+                                                   PassThroughOp,
+                                                   ThreadBufferLengths_M,
+                                                   Sequence<0>,
+                                                   0,
+                                                   MeanVarSrcDstVectorSize,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              welford_mean_thread_buf,
+                                              mean_var_grid_desc_m,
+                                              result_mean_global_buf);
+
+            threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              welford_var_thread_buf,
+                                              mean_var_grid_desc_m,
+                                              result_inv_var_global_buf);
+        };
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp
new file mode 100644
index 0000000000..2d5dc90bfb
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp
@@ -0,0 +1,556 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseWelfordSecondHalfReduceFirstHalf_,
+          typename XDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename MeanVarGridDesc_M,
+          typename MeanVarCountGridDesc_M_K,
+          typename DscaleDbiasGridDesc_M_G>
+__global__ void kernel_welford_second_half_reduce_first_half(
+    const XYGridDesc_M_K x_grid_desc_m_k,
+    const XYGridDesc_M_K dy_grid_desc_m_k,
+    const MeanVarGridDesc_M mean_var_grid_desc_m,
+    const MeanVarCountGridDesc_M_K mean_var_count_grid_desc_m_k,
+    const DscaleDbiasGridDesc_M_G dscale_dbias_grid_desc_m_g,
+    index_t blkgroup_size,
+    index_t num_xy_k_block_tile_iteration,
+    index_t num_mean_var_count_k_block_tile_iteration,
+    AccDataType epsilon,
+    bool haveSavedMeanInvVar,
+    const MeanVarDataType* const __restrict__ p_savedMean,
+    const MeanVarDataType* const __restrict__ p_savedInvVar,
+    const MeanVarDataType* const __restrict__ p_in_welford_mean,
+    const MeanVarDataType* const __restrict__ p_in_welford_variance,
+    const int32_t* const __restrict__ p_in_welford_count,
+    const DyElementwiseOp dy_elementwise_op,
+    MeanVarDataType* const __restrict__ p_out_welford_mean,
+    MeanVarDataType* const __restrict__ p_out_welford_inv_variance,
+    const XDataType* const __restrict__ p_x,
+    const DyDataType* const __restrict__ p_dy,
+    DscaleDbiasDataType* const __restrict__ p_reduce_dscale,
+    DscaleDbiasDataType* const __restrict__ p_reduce_dbias)
+{
+    GridwiseWelfordSecondHalfReduceFirstHalf_::Run(x_grid_desc_m_k,
+                                                   dy_grid_desc_m_k,
+                                                   mean_var_grid_desc_m,
+                                                   mean_var_count_grid_desc_m_k,
+                                                   dscale_dbias_grid_desc_m_g,
+                                                   blkgroup_size,
+                                                   num_xy_k_block_tile_iteration,
+                                                   num_mean_var_count_k_block_tile_iteration,
+                                                   epsilon,
+                                                   haveSavedMeanInvVar,
+                                                   p_savedMean,
+                                                   p_savedInvVar,
+                                                   p_in_welford_mean,
+                                                   p_in_welford_variance,
+                                                   p_in_welford_count,
+                                                   dy_elementwise_op,
+                                                   p_out_welford_mean,
+                                                   p_out_welford_inv_variance,
+                                                   p_x,
+                                                   p_dy,
+                                                   p_reduce_dscale,
+                                                   p_reduce_dbias);
+};
+
+template <typename XDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename MeanVarGridDesc_M,
+          typename MeanVarCountGridDesc_M_K,
+          typename DscaleDbiasGridDesc_M_G,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XDyVectorDim,
+          index_t XSrcVectorSize,
+          index_t DySrcVectorSize,
+          index_t MeanVarSrcVectorSize>
+struct GridwiseWelfordSecondHalfReduceFirstHalf
+{
+    static_assert((XDyVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0 &&
+                   MThreadSliceSize % DySrcVectorSize == 0) ||
+                      (XDyVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0 &&
+                       KThreadSliceSize % DySrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XDyVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceSrcDesc_M_1 = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<1>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelfordMerge<AccDataType, ThreadReduceSrcDesc_M_1, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder>;
+
+    using BlockwiseReduce = PartitionedBlockwiseReduction<AccDataType,
+                                                          BlockSize,
+                                                          ThreadClusterLengths_M_K,
+                                                          ThreadClusterArrangeOrder,
+                                                          ck::reduce::Add,
+                                                          false>;
+
+    using ThreadwiseReduce = ThreadwiseReduction<AccDataType,
+                                                 ThreadReduceSrcDesc_M_K,
+                                                 ThreadReduceDstDesc_M,
+                                                 ck::reduce::Add,
+                                                 false>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    // clang-format off
+    // Two of the steps of Multiblock BatchNorm Backward
+    // Step 1: Second half of Welford method to calculate mean and variance, as well as getting inv-variance = 1/sqrt(epsilon+variance) 
+    // Step 2: First half of Reduction: dbias = sum(dy), dscale = sum(dy * (x-mean) * inv-variance)
+    // clang-format on
+    __device__ static void Run(const XYGridDesc_M_K& x_grid_desc_m_k,
+                               const XYGridDesc_M_K& dy_grid_desc_m_k,
+                               const MeanVarGridDesc_M& mean_var_grid_desc_m,
+                               const MeanVarCountGridDesc_M_K& mean_var_count_grid_desc_m_k,
+                               const DscaleDbiasGridDesc_M_G& dscale_dbias_grid_desc_m_g,
+                               index_t blkgroup_size,
+                               index_t num_xy_k_block_tile_iteration,
+                               index_t num_mean_var_count_k_block_tile_iteration,
+                               AccDataType epsilon,
+                               bool haveSavedMeanInvVar,
+                               const MeanVarDataType* const __restrict__ p_savedMean,
+                               const MeanVarDataType* const __restrict__ p_savedInvVar,
+                               const MeanVarDataType* const __restrict__ p_in_welford_mean,
+                               const MeanVarDataType* const __restrict__ p_in_welford_variance,
+                               const int32_t* const __restrict__ p_in_welford_count,
+                               const DyElementwiseOp dy_elementwise_op,
+                               MeanVarDataType* const __restrict__ p_out_welford_mean,
+                               MeanVarDataType* const __restrict__ p_out_welford_inv_variance,
+                               const XDataType* const __restrict__ p_x,
+                               const DyDataType* const __restrict__ p_dy,
+                               DscaleDbiasDataType* const __restrict__ p_reduce_dscale,
+                               DscaleDbiasDataType* const __restrict__ p_reduce_dbias)
+    {
+        __shared__ AccDataType p_reduce_work_buffer[BlockSize];
+
+        auto reduce_work_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * 1, true>
+            in_welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * 1, true>
+            in_welford_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize * 1, true>
+            in_welford_count_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            welford_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true>
+            welford_count_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>& mean_thread_buf =
+            welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>&
+            inv_var_thread_buf = welford_var_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            x_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            dy_thread_buf;
+
+        // buffer of values of dy * (x-mean) * inv-variance, used as input of Blockwise reduction
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            tmp1_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            reduce_dscale_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+            reduce_dbias_thread_buf;
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / blkgroup_size;
+        const index_t block_local_id  = block_global_id % blkgroup_size;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        using ThreadBufferLengths_M           = Sequence<MThreadSliceSize>;
+        using ThreadBufferLengths_M_1         = Sequence<MThreadSliceSize, 1>;
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m_1 = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<1>{}));
+
+        // clang-format off
+        // Step 1: load existing mean and inv-variance, or do final welford reduction on mean and variance as well as get inv-variance = 1/sqrt(epsilon+variance)
+        // clang-format on
+
+        if(haveSavedMeanInvVar)
+        {
+            const auto mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_savedMean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            const auto inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_savedInvVar, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto threadwise_mean_inv_var_load =
+                ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
+                                                 AccDataType,
+                                                 MeanVarGridDesc_M,
+                                                 decltype(thread_buffer_desc_m),
+                                                 ThreadBufferLengths_M,
+                                                 Sequence<0>,
+                                                 0,
+                                                 MeanVarSrcVectorSize,
+                                                 1,
+                                                 true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize));
+
+            threadwise_mean_inv_var_load.Run(mean_var_grid_desc_m,
+                                             mean_global_buf,
+                                             thread_buffer_desc_m,
+                                             make_tuple(I0),
+                                             mean_thread_buf);
+
+            threadwise_mean_inv_var_load.Run(mean_var_grid_desc_m,
+                                             inv_var_global_buf,
+                                             thread_buffer_desc_m,
+                                             make_tuple(I0),
+                                             inv_var_thread_buf);
+        }
+        else
+        {
+            const auto welford_mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_in_welford_mean, mean_var_count_grid_desc_m_k.GetElementSpaceSize());
+
+            const auto welford_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_in_welford_variance, mean_var_count_grid_desc_m_k.GetElementSpaceSize());
+
+            const auto welford_count_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_in_welford_count, mean_var_count_grid_desc_m_k.GetElementSpaceSize());
+
+            auto threadwise_mean_var_load_m_k =
+                ThreadwiseTensorSliceTransfer_v2<AccDataType,
+                                                 AccDataType,
+                                                 MeanVarCountGridDesc_M_K,
+                                                 decltype(thread_buffer_desc_m_1),
+                                                 ThreadBufferLengths_M_1,
+                                                 Sequence<0, 1>,
+                                                 1,
+                                                 1,
+                                                 1,
+                                                 true>(
+                    mean_var_count_grid_desc_m_k,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize,
+                                     thread_k_cluster_id * 1));
+
+            auto threadwise_count_load_m_k =
+                ThreadwiseTensorSliceTransfer_v2<int32_t,
+                                                 int32_t,
+                                                 MeanVarCountGridDesc_M_K,
+                                                 decltype(thread_buffer_desc_m_1),
+                                                 ThreadBufferLengths_M_1,
+                                                 Sequence<0, 1>,
+                                                 1,
+                                                 1,
+                                                 1,
+                                                 true>(
+                    mean_var_count_grid_desc_m_k,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize,
+                                     thread_k_cluster_id * 1));
+
+            constexpr auto mean_var_count_thread_copy_step_m_k =
+                make_multi_index(0, KThreadClusterSize * 1);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                welford_mean_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+                welford_var_thread_buf(I)   = type_convert<AccDataType>(0.0f);
+                welford_count_thread_buf(I) = 0;
+            });
+
+            for(index_t reducedTiles = 0; reducedTiles < num_mean_var_count_k_block_tile_iteration;
+                ++reducedTiles)
+            {
+                threadwise_mean_var_load_m_k.Run(mean_var_count_grid_desc_m_k,
+                                                 welford_mean_global_buf,
+                                                 thread_buffer_desc_m_1,
+                                                 make_tuple(I0, I0),
+                                                 in_welford_mean_thread_buf);
+
+                threadwise_mean_var_load_m_k.Run(mean_var_count_grid_desc_m_k,
+                                                 welford_var_global_buf,
+                                                 thread_buffer_desc_m_1,
+                                                 make_tuple(I0, I0),
+                                                 in_welford_var_thread_buf);
+
+                threadwise_count_load_m_k.Run(mean_var_count_grid_desc_m_k,
+                                              welford_count_global_buf,
+                                              thread_buffer_desc_m_1,
+                                              make_tuple(I0, I0),
+                                              in_welford_count_thread_buf);
+
+                ThreadwiseWelford::Run(in_welford_mean_thread_buf,
+                                       in_welford_var_thread_buf,
+                                       in_welford_count_thread_buf,
+                                       welford_mean_thread_buf,
+                                       welford_var_thread_buf,
+                                       welford_count_thread_buf);
+
+                threadwise_mean_var_load_m_k.MoveSrcSliceWindow(
+                    mean_var_count_grid_desc_m_k, mean_var_count_thread_copy_step_m_k);
+                threadwise_count_load_m_k.MoveSrcSliceWindow(mean_var_count_grid_desc_m_k,
+                                                             mean_var_count_thread_copy_step_m_k);
+            }
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+
+                BlockwiseWelford::Run(welford_mean_thread_buf(I),
+                                      welford_var_thread_buf(I),
+                                      welford_count_thread_buf(I));
+            });
+
+            // calculate inv-variance as 1/sqrt(epsilon+variance), stored in place of variance
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                welford_var_thread_buf(I) =
+                    type_convert<AccDataType>(1.0) / sqrt(welford_var_thread_buf[I] + epsilon);
+            });
+
+            if(block_local_id == 0 && thread_k_cluster_id == 0)
+            {
+
+                auto threadwise_mean_inv_var_store =
+                    ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                       MeanVarDataType,
+                                                       decltype(thread_buffer_desc_m),
+                                                       MeanVarGridDesc_M,
+                                                       PassThroughOp,
+                                                       ThreadBufferLengths_M,
+                                                       Sequence<0>,
+                                                       0,
+                                                       1,
+                                                       InMemoryDataOperationEnum::Set,
+                                                       1,
+                                                       true>(
+                        mean_var_grid_desc_m,
+                        make_multi_index(blkgroup_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize),
+                        PassThroughOp{});
+
+                auto mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_out_welford_mean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+                auto inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_out_welford_inv_variance, mean_var_grid_desc_m.GetElementSpaceSize());
+
+                threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
+                                                  make_tuple(I0),
+                                                  mean_thread_buf,
+                                                  mean_var_grid_desc_m,
+                                                  mean_global_buf);
+
+                threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
+                                                  make_tuple(I0),
+                                                  inv_var_thread_buf,
+                                                  mean_var_grid_desc_m,
+                                                  inv_var_global_buf);
+            };
+        };
+
+        const index_t workSizePerBlock = K_BlockTileSize * num_xy_k_block_tile_iteration;
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  XYGridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XDyVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             workSizePerBlock * block_local_id +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_dy_load = ThreadwiseTensorSliceTransfer_v2<DyDataType,
+                                                                   AccDataType,
+                                                                   XYGridDesc_M_K,
+                                                                   decltype(thread_buffer_desc_m_k),
+                                                                   ThreadBufferLengths_M_K,
+                                                                   ThreadBufferDimAccessOrder,
+                                                                   XDyVectorDim,
+                                                                   DySrcVectorSize,
+                                                                   1,
+                                                                   true>(
+            dy_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             workSizePerBlock * block_local_id +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        const auto x_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto dy_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dy, dy_grid_desc_m_k.GetElementSpaceSize());
+
+        constexpr auto xy_thread_copy_step_m_k = make_multi_index(0, K_BlockTileSize);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            reduce_dscale_thread_buf(I) = type_convert<AccDataType>(0);
+            reduce_dbias_thread_buf(I)  = type_convert<AccDataType>(0);
+        });
+
+        // clang-format off
+        // Step 2: first-half of reduction: dbias = sum(dy), dscale = sum(dy * (x-mean) * inv-variance)
+        // clang-format on
+
+        for(index_t reducedTiles = 0; reducedTiles < num_xy_k_block_tile_iteration; ++reducedTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            threadwise_dy_load.Run(dy_grid_desc_m_k,
+                                   dy_global_buf,
+                                   thread_buffer_desc_m_k,
+                                   make_tuple(I0, I0),
+                                   dy_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                    dy_elementwise_op(dy_thread_buf(Number<offset>{}),
+                                      dy_thread_buf[Number<offset>{}]);
+
+                    AccDataType norm_x = (x_thread_buf[Number<offset>{}] - mean_thread_buf[iM]) *
+                                         inv_var_thread_buf[iM];
+
+                    tmp1_thread_buf(Number<offset>{}) = norm_x * dy_thread_buf[Number<offset>{}];
+                });
+            });
+
+            ThreadwiseReduce::Reduce(tmp1_thread_buf, reduce_dscale_thread_buf);
+            ThreadwiseReduce::Reduce(dy_thread_buf, reduce_dbias_thread_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, xy_thread_copy_step_m_k);
+            threadwise_dy_load.MoveSrcSliceWindow(dy_grid_desc_m_k, xy_thread_copy_step_m_k);
+        };
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            BlockwiseReduce::Reduce(reduce_work_buf, reduce_dscale_thread_buf(I));
+            block_sync_lds();
+            BlockwiseReduce::Reduce(reduce_work_buf, reduce_dbias_thread_buf(I));
+        });
+
+        auto threadwise_dscale_dbias_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               DscaleDbiasDataType,
+                                               decltype(thread_buffer_desc_m_1),
+                                               DscaleDbiasGridDesc_M_G,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M_1,
+                                               Sequence<0, 1>,
+                                               1,
+                                               1,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                dscale_dbias_grid_desc_m_g,
+                make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 block_local_id),
+                PassThroughOp{});
+
+        auto reduce_dscale_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_reduce_dscale, dscale_dbias_grid_desc_m_g.GetElementSpaceSize());
+
+        auto reduce_dbias_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_reduce_dbias, dscale_dbias_grid_desc_m_g.GetElementSpaceSize());
+
+        if(thread_k_cluster_id == 0)
+        {
+            threadwise_dscale_dbias_store.Run(thread_buffer_desc_m_1,
+                                              make_tuple(I0, I0),
+                                              reduce_dscale_thread_buf,
+                                              dscale_dbias_grid_desc_m_g,
+                                              reduce_dscale_global_buf);
+
+            threadwise_dscale_dbias_store.Run(thread_buffer_desc_m_1,
+                                              make_tuple(I0, I0),
+                                              reduce_dbias_thread_buf,
+                                              dscale_dbias_grid_desc_m_g,
+                                              reduce_dbias_global_buf);
+        };
+    };
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp b/include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
new file mode 100644
index 0000000000..2105885e18
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
@@ -0,0 +1,1078 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/math.hpp"
+#include "ck/utility/number.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include <limits>
+#include <stdlib.h>
+
+namespace ck {
+
+// Rows of column-vectors
+template <index_t MPerBlock,
+          index_t NPerBlock,
+          typename CGridDesc_M_N,
+          bool DeviceCTileIndexCheck = false>
+struct BlockToCTileMap_M00_N0_M01
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    __host__ __device__ BlockToCTileMap_M00_N0_M01() = default;
+
+    __host__ __device__ BlockToCTileMap_M00_N0_M01(const CGridDesc_M_N& c_grid_desc_m_n,
+                                                   index_t M01 = 1)
+        : M01_(M01), underlying_map_(GetBlockToCTileMap(c_grid_desc_m_n, M01))
+    {
+    }
+
+    __host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+
+        const auto M00 = math::integer_divide_ceil(M0, M01_);
+
+        const index_t grid_size = M00 * M01_ * N0;
+
+        return grid_size;
+    }
+
+    template <typename TopIdx>
+    __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+    {
+        return underlying_map_.CalculateBottomIndex(idx_top);
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                             const CTileDim& c_tile_dim) const
+    {
+        if constexpr(DeviceCTileIndexCheck)
+            return DefaultValidCTileIndex(c_tile_idx, c_tile_dim);
+        else
+            return true;
+    }
+
+    __host__ bool CheckValidity(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        if constexpr(DeviceCTileIndexCheck)
+            return true; // validity check moved to kernel
+
+        const index_t M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        if(M0 % M01_ == 0)
+        {
+            return true;
+        }
+        else
+        {
+            return false;
+        }
+    }
+
+    private:
+    __host__ __device__ static constexpr auto
+    GetBlockToCTileMap(const CGridDesc_M_N& c_grid_desc_m_n, index_t M01)
+    {
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+
+        const auto M00 = math::integer_divide_ceil(M0, M01);
+
+        const auto m00_n0_m01_to_m0_n0_block_cluster_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_replicate_transform(make_tuple(1)),
+                       make_unmerge_transform(make_tuple(M00, M01)),
+                       make_pass_through_transform(make_tuple(N0))),
+            make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1, 3>{}, Sequence<2>{}));
+
+        const auto cblockid_to_m00_n0_m01_block_cluster_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(1, M00, N0, M01))),
+            make_tuple(Sequence<0, 1, 2, 3>{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto cblockid_to_m0_n0_block_cluster_adaptor =
+            chain_tensor_adaptors(m00_n0_m01_to_m0_n0_block_cluster_adaptor,
+                                  cblockid_to_m00_n0_m01_block_cluster_adaptor);
+
+        return cblockid_to_m0_n0_block_cluster_adaptor;
+    }
+
+    index_t M01_;
+    using UnderlyingMap = decltype(GetBlockToCTileMap(CGridDesc_M_N{}, 1));
+    UnderlyingMap underlying_map_;
+};
+
+// Rows of column-vectors
+// This C-tile map dynamically adjusts M01 when C-tile index is out of range
+template <index_t MPerBlock, index_t NPerBlock, typename CGridDesc_M_N = void>
+struct BlockToCTileMap_M00_N0_M01Adapt;
+
+template <index_t MPerBlock, index_t NPerBlock>
+struct BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, void>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    __host__ __device__ BlockToCTileMap_M00_N0_M01Adapt() = default;
+
+    __host__ __device__ BlockToCTileMap_M00_N0_M01Adapt(const BlockToCTileMap_M00_N0_M01Adapt&) =
+        default;
+    __host__ __device__ BlockToCTileMap_M00_N0_M01Adapt(BlockToCTileMap_M00_N0_M01Adapt&&) =
+        default;
+    __host__ __device__ BlockToCTileMap_M00_N0_M01Adapt&
+    operator=(const BlockToCTileMap_M00_N0_M01Adapt&) = default;
+    __host__ __device__ BlockToCTileMap_M00_N0_M01Adapt&
+    operator=(BlockToCTileMap_M00_N0_M01Adapt&&) = default;
+
+    __host__ __device__ BlockToCTileMap_M00_N0_M01Adapt(index_t M, index_t N, index_t M01 = 8)
+        : M_(M), N_(N), M01_(M01)
+    {
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ __device__ BlockToCTileMap_M00_N0_M01Adapt(const CGridDesc_M_N& c_grid_desc_m_n,
+                                                        index_t M01 = 8)
+        : BlockToCTileMap_M00_N0_M01Adapt(
+              c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1), M01)
+    {
+    }
+
+    __host__ static constexpr index_t CalculateGridSize(index_t M, index_t N)
+    {
+        const auto M0 = math::integer_divide_ceil(M, MPerBlock);
+        const auto N0 = math::integer_divide_ceil(N, NPerBlock);
+
+        return M0 * N0;
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ static constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return CalculateGridSize(c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1));
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const
+    {
+        return true;
+    }
+
+    template <typename TopIdx>
+    __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+    {
+        auto block_1d_id = idx_top[I0];
+
+        const auto M0 = math::integer_divide_ceil(M_, MPerBlock);
+        const auto N0 = math::integer_divide_ceil(N_, NPerBlock);
+
+        block_1d_id = block_1d_id % (M0 * N0); // swallow batch index
+
+        index_t idx_N0 = block_1d_id % N0;
+        index_t idx_M0 = block_1d_id / N0;
+
+        const auto M01_adapt = (idx_M0 < M0 - M0 % M01_) ? M01_ : M0 % M01_;
+
+        index_t idx_M00          = idx_M0 / M01_;
+        index_t idx_M01          = idx_M0 % M01_;
+        index_t idx_N0_M01_local = idx_N0 + idx_M01 * N0;
+
+        /**
+         *                        idxN0
+         *
+         *           |<               mtx   N                 >|
+         *
+         *             NPerBlock   NPerBlock   NPerBlock   NPerBlock
+         *                N_0         N_1        N_2         N_3
+         *       -   |-----------|-----------|-----------|-----|-----|-
+         *       ^   | -   -  0  |/---->  2  |           |     |     |
+         *           | |   |     /     |     |           |     |     |  M_0  MPerBlock
+         *           | M   |    /|     |     |           |     |     |
+         *           |-0---|---/-|-----|-----|-----------|-----|-----|-
+         *           | 1   |  /  |     |     |  blockid  |     |     |
+         * idxM0     | |   | /   |     V     |     5     |     |     |  M_1  MPerBlock
+         *           | -   V   1 |     -  3  |           |     |     |
+         *           |-----------|-----------|-----------|-----|-----|-
+         *    mtx M  |           |           |           |     |     |
+         *           |           |           |           |     |     |  M_2  MPerBlock
+         *           |           |           |           |     |     |
+         *           |-----------|-----------|-----------|-----|-----|-
+         *           |           |           |           |     |     |
+         *           |           |           |           |     |     |  M_3  MPerBlock
+         *           |           |           |           |     |     |
+         *           |-----------|-----------|-----------|-----|-----|-
+         *       V   |           |           |           |     |     |
+         *       -   |-----------|-----------|-----------|-----|-----|- M_4  MPerBlock
+         *           |           |           |           |     |     |
+         *           |-----------|-----------|-----------|-----|-----|-
+         *  Example:
+         *   assume:
+         *      M0 = 5
+         *      N0 = 4
+         *      block_1d_id = 5
+         *      M01 = 2
+         *
+         *   idx_N0 = 1
+         *   idx_M0 = 1
+         *   M01_adapt = 2
+         *   idx_M00 = 0
+         *   idx_M01 = 1
+         *   idx_N0_M01_local = 5
+         *   output {1, 2}
+         */
+
+        return make_tuple(idx_N0_M01_local % M01_adapt + idx_M00 * M01_,
+                          idx_N0_M01_local / M01_adapt);
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& /* c_tile_idx */,
+                                             const CTileDim& /* c_tile_dim */) const
+    {
+        return true; // always valid provided that user gets grid size from CalculateGridSize()
+    }
+
+    private:
+    index_t M_;
+    index_t N_;
+    index_t M01_;
+};
+
+// keep the redundant type argument for backward compatibility
+template <index_t MPerBlock, index_t NPerBlock, typename CGridDesc_M_N>
+struct BlockToCTileMap_M00_N0_M01Adapt : BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, void>
+{
+    using BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, void>::
+        BlockToCTileMap_M00_N0_M01Adapt;
+};
+
+// 2D slices of column-vectors in 3D space
+// This C-tile map dynamically adjusts M01 when C-tile index is out of range
+template <index_t MPerBlock, index_t NPerBlock, typename CGridDesc_M_N>
+struct BlockToCTileMap_KSplit_M00_N0_M01Adapt
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    __host__ __device__ BlockToCTileMap_KSplit_M00_N0_M01Adapt() = default;
+
+    __host__ __device__ BlockToCTileMap_KSplit_M00_N0_M01Adapt(const CGridDesc_M_N& c_grid_desc_m_n,
+                                                               index_t M01    = 8,
+                                                               index_t KSplit = 1)
+        : M01_(M01), KSplit_(KSplit), c_grid_desc_m_n_(c_grid_desc_m_n)
+    {
+    }
+
+    __host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+
+        const index_t grid_size = M0 * N0 * KSplit_;
+
+        return grid_size;
+    }
+
+    template <typename TopIdx>
+    __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+    {
+        auto block_1d_id = idx_top[I0];
+
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I1), NPerBlock);
+
+        block_1d_id = block_1d_id % (M0 * N0 * KSplit_); // hide groups
+
+        const index_t idx_ksplit = block_1d_id / (M0 * N0);
+        block_1d_id              = block_1d_id % (M0 * N0);
+
+        index_t idx_N0 = block_1d_id % N0;
+        index_t idx_M0 = block_1d_id / N0;
+
+        const auto M01_adapt = (idx_M0 < M0 - M0 % M01_) ? M01_ : M0 % M01_;
+
+        index_t idx_M00          = idx_M0 / M01_;
+        index_t idx_M01          = idx_M0 % M01_;
+        index_t idx_N0_M01_local = idx_N0 + idx_M01 * N0;
+
+        return make_tuple(idx_ksplit,
+                          idx_N0_M01_local % M01_adapt + idx_M00 * M01_,
+                          idx_N0_M01_local / M01_adapt);
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& /* c_tile_idx */,
+                                             const CTileDim& /* c_tile_dim */) const
+    {
+        return true; // always valid provided that user gets grid size from CalculateGridSize()
+    }
+
+    __host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const { return true; }
+
+    private:
+    index_t M01_;
+    index_t KSplit_;
+    CGridDesc_M_N c_grid_desc_m_n_;
+};
+
+// Blocks of row-vectors
+template <index_t MPerBlock,
+          index_t NPerBlock,
+          typename CGridDesc_M_N,
+          bool DeviceCTileIndexCheck = false>
+struct BlockToCTileMap_M00_N00_M01_N01
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    __host__ __device__ BlockToCTileMap_M00_N00_M01_N01() = default;
+
+    __host__ __device__ BlockToCTileMap_M00_N00_M01_N01(const CGridDesc_M_N& c_grid_desc_m_n,
+                                                        index_t M01 = 1,
+                                                        index_t N01 = 1)
+        : M01_(M01), N01_(N01), underlying_map_(GetBlockToCTileMap(c_grid_desc_m_n, M01, N01))
+    {
+    }
+
+    __host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+
+        const auto M00 = math::integer_divide_ceil(M0, M01_);
+        const auto N00 = math::integer_divide_ceil(N0, N01_);
+
+        const index_t grid_size = M00 * M01_ * N00 * N01_;
+
+        return grid_size;
+    }
+
+    template <typename TopIdx>
+    __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+    {
+        return underlying_map_.CalculateBottomIndex(idx_top);
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                             const CTileDim& c_tile_dim) const
+    {
+        if constexpr(DeviceCTileIndexCheck)
+            return DefaultValidCTileIndex(c_tile_idx, c_tile_dim);
+        else
+            return true;
+    }
+
+    __host__ bool CheckValidity(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        if constexpr(DeviceCTileIndexCheck)
+            return true; // validity check moved to kernel
+
+        const index_t M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const index_t N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+        if(M0 % M01_ == 0 && N0 % N01_ == 0)
+        {
+            return true;
+        }
+        else
+        {
+            return false;
+        }
+    }
+
+    private:
+    __host__ __device__ static constexpr auto
+    GetBlockToCTileMap(const CGridDesc_M_N& c_grid_desc_m_n, index_t M01, index_t N01)
+    {
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+
+        const auto M00 = math::integer_divide_ceil(M0, M01);
+        const auto N00 = math::integer_divide_ceil(N0, N01);
+
+        const auto m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_replicate_transform(
+                               make_tuple(1)), // swallow the carry from lower dimensions
+                           make_unmerge_transform(make_tuple(M00, M01)),
+                           make_unmerge_transform(make_tuple(N00, N01))),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 3>{}, Sequence<2, 4>{}));
+
+        const auto cblockid_to_m00_m01_n00_n01_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(1, M00, N00, M01, N01))),
+                make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                make_tuple(Sequence<0>{}));
+
+        const auto cblockid_to_m0_n0_block_cluster_adaptor =
+            chain_tensor_adaptors(m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor,
+                                  cblockid_to_m00_m01_n00_n01_block_cluster_adaptor);
+
+        return cblockid_to_m0_n0_block_cluster_adaptor;
+    }
+
+    index_t M01_, N01_;
+    using UnderlyingMap = decltype(GetBlockToCTileMap(CGridDesc_M_N{}, 1, 1));
+    UnderlyingMap underlying_map_;
+};
+
+// 2D slices of row-vectors in 3D space
+template <index_t MPerBlock,
+          index_t NPerBlock,
+          typename CGridDesc_M_N,
+          bool DeviceCTileIndexCheck = false>
+struct BlockToCTileMap_KSplit_M00_N00_M01_N01
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    __host__ BlockToCTileMap_KSplit_M00_N00_M01_N01() = default;
+
+    __host__ BlockToCTileMap_KSplit_M00_N00_M01_N01(const CGridDesc_M_N& c_grid_desc_m_n,
+                                                    index_t M01    = 1,
+                                                    index_t N01    = 1,
+                                                    index_t KSplit = 1)
+        : c_grid_desc_m_n_(c_grid_desc_m_n),
+          M01_(M01),
+          N01_(N01),
+          KSplit_(KSplit),
+          underlying_map_(GetBlockToCTileMap(c_grid_desc_m_n, M01, N01, KSplit))
+    {
+    }
+
+    __host__ __device__ constexpr index_t
+    CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+
+        const auto M00 = math::integer_divide_ceil(M0, M01_);
+        const auto N00 = math::integer_divide_ceil(N0, N01_);
+
+        const index_t grid_size = M00 * M01_ * N00 * N01_ * KSplit_;
+
+        return grid_size;
+    }
+
+    template <typename TopIdx>
+    __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+    {
+        static_assert(TopIdx::Size() == 1);
+
+        return underlying_map_.CalculateBottomIndex(
+            make_multi_index(idx_top[I0] % CalculateGridSize()));
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                             const CTileDim& c_tile_dim) const
+    {
+        if constexpr(DeviceCTileIndexCheck)
+            return DefaultValidCTileIndex(c_tile_idx, c_tile_dim);
+        else
+            return true;
+    }
+
+    __host__ bool CheckValidity(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        if constexpr(DeviceCTileIndexCheck)
+            return true; // validity check moved to kernel
+
+        const index_t M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const index_t N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+        if(M0 % M01_ == 0 && N0 % N01_ == 0)
+        {
+            return true;
+        }
+        else
+        {
+            return false;
+        }
+    }
+
+    private:
+    __device__ constexpr index_t CalculateGridSize() const
+    {
+        return CalculateGridSize(c_grid_desc_m_n_);
+    }
+
+    __host__ static constexpr auto GetBlockToCTileMap(const CGridDesc_M_N& c_grid_desc_m_n,
+                                                      index_t M01,
+                                                      index_t N01,
+                                                      index_t KSplit)
+    {
+        const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
+        const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
+
+        const auto M00 = math::integer_divide_ceil(M0, M01);
+        const auto N00 = math::integer_divide_ceil(N0, N01);
+
+        const auto ksplit_m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_pass_through_transform(KSplit),
+                           make_unmerge_transform(make_tuple(M00, M01)),
+                           make_unmerge_transform(make_tuple(N00, N01))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 3>{}, Sequence<2, 4>{}));
+
+        const auto c_blockid_to_ksplit_m00_m01_n00_n01_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(KSplit, M00, N00, M01, N01))),
+                make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                make_tuple(Sequence<0>{}));
+
+        const auto c_blockid_to_ksplit_m0_n0_block_cluster_adaptor =
+            chain_tensor_adaptors(ksplit_m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor,
+                                  c_blockid_to_ksplit_m00_m01_n00_n01_block_cluster_adaptor);
+
+        return c_blockid_to_ksplit_m0_n0_block_cluster_adaptor;
+    }
+
+    CGridDesc_M_N c_grid_desc_m_n_;
+    index_t M01_, N01_, KSplit_;
+    using UnderlyingMap = decltype(GetBlockToCTileMap(CGridDesc_M_N{}, 1, 1, 1));
+    UnderlyingMap underlying_map_;
+};
+
+template <typename CTileIdx, typename CTileDim>
+__host__ __device__ bool DefaultValidCTileIndex(const CTileIdx& c_tile_idx,
+                                                const CTileDim& c_tile_dim)
+{
+    bool is_valid = false;
+
+    const index_t m_block = c_tile_dim[Number<0>{}];
+    const index_t n_block = c_tile_dim[Number<1>{}];
+
+    if constexpr(CTileIdx::Size() == 2)
+    {
+        const index_t m_block_idx = c_tile_idx[Number<0>{}];
+        const index_t n_block_idx = c_tile_idx[Number<1>{}];
+        if(0 <= m_block_idx && m_block_idx < m_block && 0 <= n_block_idx && n_block_idx < n_block)
+        {
+            is_valid = true;
+        }
+    }
+    else if constexpr(CTileIdx::Size() == 3)
+    {
+        const index_t ksplit_idx  = c_tile_idx[Number<0>{}];
+        const index_t m_block_idx = c_tile_idx[Number<1>{}];
+        const index_t n_block_idx = c_tile_idx[Number<2>{}];
+        if(0 <= m_block_idx && m_block_idx < m_block && 0 <= n_block_idx && n_block_idx < n_block)
+        {
+            is_valid = true;
+        }
+        ignore = ksplit_idx;
+    }
+
+    return is_valid;
+}
+
+// This wrapper class is for grouped gemm where it subtracts blockIdx by a value so that the
+// workgroups assigned to a given gemm problem have top index offsetted to range [0,
+// grid_size_per_gemm]
+template <typename UnderlyingBlockToCTileMap>
+struct OffsettedBlockToCTileMap
+{
+    using underlying_type = UnderlyingBlockToCTileMap;
+
+    __host__ __device__ OffsettedBlockToCTileMap(UnderlyingBlockToCTileMap block_to_ctile_map,
+                                                 index_t block_start)
+    {
+        block_to_ctile_map_ = block_to_ctile_map;
+        block_start_        = block_start;
+    }
+
+    template <typename TopIdx>
+    __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+    {
+        return block_to_ctile_map_.CalculateBottomIndex(
+            make_multi_index(idx_top[Number<0>{}] - block_start_));
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
+                                             const CTileDim& c_tile_dim) const
+    {
+        return block_to_ctile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ bool CheckValidity(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        return block_to_ctile_map_.CheckValidity(c_grid_desc_m_n);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
+    {
+        return block_to_ctile_map_.CalculateGridSize(c_grid_desc_m_n);
+    }
+
+    UnderlyingBlockToCTileMap block_to_ctile_map_;
+    index_t block_start_;
+};
+
+/**
+ * @brief      Simple tile mapping which creates 3D grid of block of threads.
+ *
+ * @paragraph  Description
+ *             This Block-to-C-tile-map creates a 3D grid (n_blocks, m_blocks, z_blocks) of thread
+ *             blocks. The first 2D are regular 2D tiles created by division of output GEMM
+ *             dimenions by corresponding tile size. The third dimension (Z) is a k-split dimension,
+ *             which denotes the number of blocks we use to divide work on GEMM K dimension onto.
+ *
+ * @tparam     MPerBlock  Output block tile size in M dimension.
+ * @tparam     NPerBlock  Output block tile size in N dimension.
+ */
+template <index_t MPerBlock, index_t NPerBlock>
+struct BlockToCTileMap_3DGrid_KSplit
+{
+
+    __host__ __device__ BlockToCTileMap_3DGrid_KSplit() = default;
+
+    __host__ __device__ constexpr auto
+    CalculateGridSize(index_t M, index_t N, index_t k_split) const
+    {
+        // Create 3D grid
+        const auto M0 = math::integer_divide_ceil(M, MPerBlock);
+        const auto N0 = math::integer_divide_ceil(N, NPerBlock);
+
+        return std::make_tuple(N0, M0, k_split);
+    }
+
+    template <typename TopIdx>
+    __device__ constexpr auto CalculateBottomIndex(const TopIdx&) const
+    {
+        return make_tuple(blockIdx.z, blockIdx.y, blockIdx.x);
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& /* c_tile_idx */,
+                                             const CTileDim& /* c_tile_dim */) const
+    {
+        return true; // always valid provided that user gets grid size from CalculateGridSize()
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const
+    {
+        return true;
+    }
+};
+
+enum StreamKReductionStrategy
+{
+    Atomic = 0, // sk block use atomic to do reduction
+    Reduction,  // let some workgroup responsible for doing the reduction operation
+};
+
+template <uint32_t MPerBlock_,
+          uint32_t NPerBlock_,
+          uint32_t KPerBlock_,
+          StreamKReductionStrategy ReductionStrategy_ = StreamKReductionStrategy::Atomic,
+          uint32_t TileSwizzleSubM_                   = 8>
+struct BlockToCTileMap_GemmStreamK
+{
+    static constexpr uint32_t min_k_iters_per_sk_block          = 2;
+    static constexpr uint32_t MPerBlock                         = MPerBlock_;
+    static constexpr uint32_t NPerBlock                         = NPerBlock_;
+    static constexpr uint32_t KPerBlock                         = KPerBlock_;
+    static constexpr StreamKReductionStrategy ReductionStrategy = ReductionStrategy_;
+    static constexpr uint32_t tile_swizzle_sub_m                = TileSwizzleSubM_;
+
+    //--------------------------------------
+    // pass to device
+    uint32_t sk_num_blocks;
+    uint32_t sk_num_big_blocks;
+    uint32_t dp_start_block_idx;
+    uint32_t reduction_start_block_idx;
+    uint32_t k_iters_per_big_block;
+    MDiv2 n_tiles;
+    MDiv k_iters_per_tile;
+    MDiv eqav_tiles_big;    // for reduction
+    MDiv eqav_tiles_little; // for reduction
+
+    // MDiv tile_swizzle_sub_m_rem;
+    //--------------------------------------
+
+    // prefer construct on host
+    BlockToCTileMap_GemmStreamK(uint32_t m,
+                                uint32_t n,
+                                uint32_t k,
+                                uint32_t num_cu,
+                                uint32_t occupancy,
+                                uint32_t sk_blocks = 0xffffffff)
+    {
+        uint32_t num_tiles =
+            math::integer_divide_ceil(m, MPerBlock) * math::integer_divide_ceil(n, NPerBlock);
+        k_iters_per_tile = MDiv(math::integer_divide_ceil(k, KPerBlock));
+
+        // one cu can hold one wg at one time, from the whole chip's point of view
+        // if number of wg is same as num_cu, we call it 1 dispatch
+        // if number of wg is 2x num_cu, we call it 2 dispatches.
+        // one dispatch can deliver wg same as num_cu (full dispatch), or less than num_cu (partial
+        // dispatch)
+        //
+        uint32_t full_dispatches         = num_tiles / num_cu;
+        uint32_t full_dispatch_tiles     = full_dispatches * num_cu;
+        uint32_t partial_dispatche_tiles = num_tiles - full_dispatch_tiles;
+
+        uint32_t sk_occupancy = occupancy;
+        uint32_t dp_tiles     = full_dispatch_tiles;
+        uint32_t sk_tiles     = partial_dispatche_tiles;
+
+        if(full_dispatches < occupancy)
+        {
+            // in this case, we allocate all blocks as sk blocks
+            // sk_occupancy = occupancy - full_dispatches;
+            sk_occupancy = 1; // TODO: single occ seems better
+            dp_tiles     = full_dispatch_tiles;
+            sk_tiles     = partial_dispatche_tiles;
+        }
+        else if((occupancy > 1) && (full_dispatches % occupancy == occupancy - 1))
+        {
+            // e.g. occupancy = 2, full_dispatches = 3, 5, 7 ...
+            //      occupancy = 3, full_dispatches = 5, 8, 11 ...
+            //      occupancy = 4, full_dispatches = 7, 11 ...
+            sk_occupancy = 1; // left 1 slot for sk occupancy
+            dp_tiles     = full_dispatch_tiles;
+            sk_tiles     = partial_dispatche_tiles;
+        }
+        else
+        {
+            // others, we reduce 1 dispatch from dp, together with partial dispatch,
+            // to construct sk dispatch
+            sk_occupancy = occupancy - ((full_dispatches - 1) % occupancy);
+            dp_tiles     = full_dispatch_tiles - num_cu;
+            sk_tiles     = partial_dispatche_tiles + num_cu;
+        }
+
+        // uint32_t dp_iters_per_block = k_iters_per_tile.get();
+        uint32_t sk_total_iters = k_iters_per_tile.get() * sk_tiles;
+        uint32_t dp_num_blocks  = 0;
+
+        {
+            uint32_t min_sk_tiles = (sk_tiles >= num_cu) ? num_cu : (sk_tiles + 1);
+            uint32_t max_sk_tiles =
+                (sk_tiles >= num_cu) ? num_cu * sk_occupancy
+                                     : math::min(num_cu, sk_total_iters / min_k_iters_per_sk_block);
+
+            // if use dp for sk-block, how many iters do we need
+            uint32_t dp_for_sk_iters = k_iters_per_tile.get();
+
+            uint32_t best_sk_score =
+                std::numeric_limits<int>::max(); // we need to find the smallest sk iters
+            for(uint32_t tentative_sk_blocks = min_sk_tiles; tentative_sk_blocks < max_sk_tiles;
+                tentative_sk_blocks++)
+            {
+                uint32_t tentative_sk_iters_per_block =
+                    (sk_total_iters + tentative_sk_blocks - 1) / tentative_sk_blocks;
+                uint32_t tentative_sk_iters = tentative_sk_iters_per_block;
+                uint32_t sk_blocks_per_tile = (tentative_sk_blocks + sk_tiles - 1) / sk_tiles;
+
+                // TODO: carefully adjust this parameter
+                //       the more sk_blocks_per_tile, the worse the overhead
+                uint32_t cross_sk_blocks_overhead = sk_blocks_per_tile;
+                if(tentative_sk_blocks % sk_tiles != 0)
+                {
+                    // penalty for uneven divide
+                    cross_sk_blocks_overhead +=
+                        sk_blocks_per_tile * tentative_sk_iters_per_block / 50;
+                }
+
+                uint32_t tentative_sk_score = tentative_sk_iters + cross_sk_blocks_overhead;
+
+                if(tentative_sk_score < best_sk_score)
+                {
+                    best_sk_score = tentative_sk_score;
+                    sk_num_blocks = tentative_sk_blocks;
+                }
+            }
+
+            if(best_sk_score >= dp_for_sk_iters)
+            {
+                sk_num_blocks = 0;
+            }
+
+            // give a chance to control num of sk blocks
+            sk_num_blocks = sk_blocks != 0xffffffff ? sk_blocks : sk_num_blocks;
+
+            if(sk_num_blocks == 0)
+            {
+                sk_num_big_blocks     = 0;
+                k_iters_per_big_block = 0;
+
+                dp_num_blocks      = num_tiles; // all tile to be dp block
+                dp_start_block_idx = 0;
+                sk_total_iters     = 0; // clear this tiles
+            }
+            else
+            {
+                // k_iters_per_sk_block is the floor of avg each ck block loop over tiles.
+                // we need to decide how many iters for each sk block
+                // let m = k_iters_per_sk_block
+                // some of the sk block (little) will cover m iters, some (big) will cover m+1
+                // we have
+                // 1) l + b = sk_blocks
+                // 2) l * m + b * (m + 1) = sk_total_iters
+                //      => (l + b) * m + b = sk_total_iters
+                //      => sk_blocks * m + b = sk_total_iters
+                //      => b = sk_total_iters - m * sk_blocks
+                //      NOTE: big could be zero
+                uint32_t k_iters_per_sk_block = sk_total_iters / sk_num_blocks;
+                sk_num_big_blocks     = sk_total_iters - k_iters_per_sk_block * sk_num_blocks;
+                k_iters_per_big_block = k_iters_per_sk_block + 1;
+
+                dp_num_blocks      = dp_tiles;
+                dp_start_block_idx = (sk_num_blocks + num_cu - 1) / num_cu * num_cu;
+            }
+        }
+        n_tiles                   = MDiv2(math::integer_divide_ceil(n, NPerBlock));
+        reduction_start_block_idx = dp_start_block_idx + dp_num_blocks;
+
+        if constexpr(ReductionStrategy == StreamKReductionStrategy::Reduction)
+        {
+            uint32_t upper_big    = math::lcm(k_iters_per_big_block, k_iters_per_tile.get());
+            uint32_t upper_little = math::lcm(k_iters_per_big_block - 1, k_iters_per_tile.get());
+            eqav_tiles_big        = MDiv(upper_big / k_iters_per_tile.get());
+            eqav_tiles_little     = MDiv(upper_little / k_iters_per_tile.get());
+        }
+
+#if 0
+        printf("cu:%d, occupancy:%d, grids:%d, num_tiles:%d, dp_tiles:%d, sk_num_big_blocks:%d, "
+               "sk_num_blocks:%d, "
+               "sk_total_iters:%d, dp_start_block_idx:%d, dp_iters_per_block:%d, dp_num_blocks:%d, "
+               "k_iters_per_tile:%d, k_iters_per_big_block:%d, reduction_start_block_idx:%u, "
+               "sk_tiles:%u, workspace(acc float):%u\n",
+               num_cu,
+               occupancy,
+               get_grid_dims().x,
+               num_tiles,
+               dp_tiles,
+               sk_num_big_blocks,
+               sk_num_blocks,
+               sk_total_iters,
+               dp_start_block_idx,
+               dp_iters_per_block,
+               dp_num_blocks,
+               k_iters_per_tile.get(),
+               k_iters_per_big_block,
+               reduction_start_block_idx,
+               get_sk_tiles(),
+               get_workspace_size(sizeof(float)));
+#endif
+    }
+
+    __host__ __device__ uint32_t get_sk_total_iters() const
+    {
+        uint32_t sk_total_iters = sk_num_big_blocks * k_iters_per_big_block +
+                                  (sk_num_blocks - sk_num_big_blocks) * (k_iters_per_big_block - 1);
+        return sk_total_iters;
+    }
+
+    __host__ __device__ uint32_t get_sk_tiles() const
+    {
+        // tiles for sk
+        uint32_t sk_total_iters = get_sk_total_iters();
+        return k_iters_per_tile.div(sk_total_iters);
+    }
+
+    __host__ __device__ dim3 get_grid_dims() const
+    {
+        if constexpr(ReductionStrategy == StreamKReductionStrategy::Reduction)
+        {
+            return dim3(reduction_start_block_idx + get_sk_tiles(), 1, 1);
+        }
+        else
+            return dim3(reduction_start_block_idx, 1, 1);
+    }
+
+    __device__ uint32_t get_block_idx() const
+    {
+        // TODO: swizzle block index for better locality
+        return __builtin_amdgcn_readfirstlane(blockIdx.x);
+    }
+
+    __device__ void
+    get_block_itr(uint32_t block_idx, uint32_t& iter_start, uint32_t& iter_end) const
+    {
+        if(block_idx < sk_num_big_blocks)
+        {
+            iter_start = block_idx * k_iters_per_big_block;
+            iter_end   = iter_start + k_iters_per_big_block;
+        }
+        else if(block_idx < sk_num_blocks)
+        {
+            iter_start = (sk_num_big_blocks * k_iters_per_big_block) +
+                         (block_idx - sk_num_big_blocks) * (k_iters_per_big_block - 1);
+            iter_end = iter_start + (k_iters_per_big_block - 1);
+        }
+        else if(block_idx >= dp_start_block_idx)
+        {
+            uint32_t sk_total_iters     = get_sk_total_iters();
+            uint32_t dp_iters_per_block = k_iters_per_tile.get();
+            iter_start = sk_total_iters + (block_idx - dp_start_block_idx) * dp_iters_per_block;
+            iter_end   = iter_start + dp_iters_per_block;
+        }
+    }
+
+    __device__ uint32_t get_current_iter_length(uint32_t iter_start,
+                                                uint32_t iter_end,
+                                                uint32_t total_iter_length) const
+    {
+        uint32_t iter_length_mod, iter_length_quo /*unused*/;
+        k_iters_per_tile.divmod(iter_end, iter_length_quo, iter_length_mod);
+        uint32_t current_iter_length = math::min(
+            iter_length_mod == 0 ? (iter_end - iter_start) : iter_length_mod, total_iter_length);
+        return current_iter_length;
+    }
+
+    __device__ uint32_t get_tile_idx(uint32_t iter) const { return k_iters_per_tile.div(iter); }
+
+    __device__ void
+    get_tile_idx_with_offset(uint32_t iter, uint32_t& tile_idx, uint32_t& iter_offset) const
+    {
+        k_iters_per_tile.divmod(iter, tile_idx, iter_offset);
+    }
+
+    __device__ auto tile_to_spatial(uint32_t tile_idx, uint32_t m, uint32_t n) const
+    {
+        uint32_t m_tile_idx, n_tile_idx;
+        uint32_t n_tiles_value = math::integer_divide_ceil(n, NPerBlock);
+        n_tiles.divmod(tile_idx, n_tiles_value, m_tile_idx, n_tile_idx);
+
+        // swizzle tile
+        uint32_t m_tiles = math::integer_divide_ceil(m, MPerBlock);
+
+        uint32_t tile_swizzle_sub_m_rem = m_tiles % tile_swizzle_sub_m;
+
+        const auto sub_m_adapt = (m_tile_idx < (m_tiles - tile_swizzle_sub_m_rem))
+                                     ? tile_swizzle_sub_m
+                                     : tile_swizzle_sub_m_rem;
+
+        uint32_t m_tile_idx_sub0, m_tile_idx_sub1;
+        m_tile_idx_sub0 = m_tile_idx / tile_swizzle_sub_m;
+        m_tile_idx_sub1 = m_tile_idx % tile_swizzle_sub_m;
+
+        uint32_t tile_idx_local = n_tile_idx + m_tile_idx_sub1 * n_tiles_value;
+
+        uint32_t m_tile_idx_with_adapt, n_tile_idx_with_adapt;
+
+        n_tile_idx_with_adapt = tile_idx_local / sub_m_adapt;
+        m_tile_idx_with_adapt = tile_idx_local % sub_m_adapt;
+        return make_tuple(m_tile_idx_with_adapt + m_tile_idx_sub0 * tile_swizzle_sub_m,
+                          n_tile_idx_with_adapt);
+    }
+
+    __host__ __device__ uint32_t get_workspace_size_for_acc(uint32_t acc_element_bytes) const
+    {
+        static constexpr uint32_t alignment = 128;
+        uint32_t acc_buffer_bytes =
+            MPerBlock * NPerBlock * get_total_acc_buffers() * acc_element_bytes;
+        return (acc_buffer_bytes + alignment - 1) / alignment * alignment;
+    }
+
+    __host__ __device__ uint32_t get_workspace_size_for_semaphore() const
+    {
+        return get_sk_tiles() * sizeof(uint32_t);
+    }
+
+    __host__ __device__ uint32_t get_workspace_size(uint32_t acc_element_bytes) const
+    {
+        return get_workspace_size_for_acc(acc_element_bytes) + get_workspace_size_for_semaphore();
+    }
+
+    __host__ __device__ uint32_t get_tile_intersections(uint32_t tiles_,
+                                                        const MDiv& eqav_tiles_) const
+    {
+        uint32_t tile_idx_       = tiles_ == 0 ? 0 : (tiles_ - 1);
+        uint32_t max_eqav_tiles_ = eqav_tiles_.get() - 1;
+        uint32_t quo_, rem_;
+        eqav_tiles_.divmod(tile_idx_, quo_, rem_);
+        return quo_ * max_eqav_tiles_ + rem_;
+    }
+
+    __host__ __device__ uint32_t get_tiles_cover_sk_block(uint32_t num_sk_blocks_,
+                                                          uint32_t iters_per_sk_block_) const
+    {
+        return k_iters_per_tile.div(num_sk_blocks_ * iters_per_sk_block_ + k_iters_per_tile.get() -
+                                    1);
+    }
+
+    __host__ __device__ uint32_t get_total_acc_buffers() const
+    {
+        uint32_t tiles_cover_big_blocks =
+            get_tiles_cover_sk_block(sk_num_big_blocks, k_iters_per_big_block);
+        uint32_t tiles_cover_little_blocks =
+            get_tiles_cover_sk_block(sk_num_blocks - sk_num_big_blocks, k_iters_per_big_block - 1);
+
+        uint32_t total_intersec_big =
+            get_tile_intersections(tiles_cover_big_blocks, eqav_tiles_big);
+        uint32_t total_intersec_little =
+            get_tile_intersections(tiles_cover_little_blocks, eqav_tiles_little);
+
+        return sk_num_blocks + total_intersec_big + total_intersec_little;
+    }
+
+    __device__ uint32_t get_acc_buffer_offset_from_tile(uint32_t tile_idx_) const
+    {
+        // TODO: from big to little
+        uint32_t tiles_cover_big_blocks =
+            get_tiles_cover_sk_block(sk_num_big_blocks, k_iters_per_big_block);
+        if(tile_idx_ < tiles_cover_big_blocks)
+        {
+            uint32_t touched_sk_blocks =
+                (tile_idx_ * k_iters_per_tile.get() + k_iters_per_big_block - 1) /
+                k_iters_per_big_block;
+            uint32_t current_intersec = get_tile_intersections(tile_idx_, eqav_tiles_big);
+            return touched_sk_blocks + current_intersec;
+        }
+        else
+        {
+            uint32_t iters_per_little_sk_block = k_iters_per_big_block - 1;
+            uint32_t tile_idx_little_reverse   = get_sk_tiles() - tile_idx_;
+            uint32_t touched_sk_blocks =
+                (tile_idx_little_reverse * k_iters_per_tile.get() + iters_per_little_sk_block - 1) /
+                iters_per_little_sk_block;
+            uint32_t current_intersec =
+                get_tile_intersections(tile_idx_little_reverse, eqav_tiles_little);
+            return get_total_acc_buffers() - (touched_sk_blocks + current_intersec);
+        }
+    }
+
+    __device__ uint32_t get_acc_buffer_offset_from_block(uint32_t block_idx_) const
+    {
+        uint32_t iters_per_big_sk_block    = k_iters_per_big_block;
+        uint32_t iters_per_little_sk_block = k_iters_per_big_block - 1;
+        if(block_idx_ < sk_num_big_blocks)
+        {
+            uint32_t touched_tiles    = k_iters_per_tile.div(block_idx_ * iters_per_big_sk_block +
+                                                          k_iters_per_tile.get() - 1);
+            uint32_t current_intersec = get_tile_intersections(touched_tiles, eqav_tiles_big);
+            return block_idx_ + current_intersec;
+        }
+        else
+        {
+            uint32_t block_idx_little_reverse = sk_num_blocks - block_idx_;
+            uint32_t touched_tiles            = k_iters_per_tile.div(
+                block_idx_little_reverse * iters_per_little_sk_block + k_iters_per_tile.get() - 1);
+            uint32_t current_intersec = get_tile_intersections(touched_tiles, eqav_tiles_little);
+            return get_total_acc_buffers() - (block_idx_little_reverse + current_intersec);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..b25f136a37
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp
@@ -0,0 +1,1115 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+
+namespace ck {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   output : F[M, N0], where N0 is number of blocks along N dimension
+//   output : G[M, N0], where N0 is number of blocks along N dimension
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+//   F, G = welford(E)
+// Assume:
+//   D0, D1, ... and E have the same layout
+//   Calculate mean & variance along N dimension for E
+template <typename ABDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EMeanVarDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+          typename AGridDesc_M_K,
+          typename BGridDesc_N_K,
+          typename DsGridDesc_M_N,
+          typename EGridDesc_M_N,
+          typename MeanVarGridDesc_M_NBlock,
+          typename CountGridDesc_M_NBlock,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename PostShuffleThreadClusterSize_M_N,
+          index_t PostShuffleScalarPerVector,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmMultipleDWelfordFirstHalf_xdl_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(ABDataType),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    // A desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultBGridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDescriptor_M_N>
+    __host__ __device__ static constexpr auto MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const EGridDescriptor_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDescriptor_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const DsGridDescriptor_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    template <typename GridDescriptor_M_N>
+    __host__ __device__ static constexpr auto
+    MakeMeanVarCountGridDescriptor_MBlock_MPerBlock_NBlock(const GridDescriptor_M_N& grid_desc_m_n)
+    {
+        const auto M      = grid_desc_m_n.GetLength(I0);
+        const auto NBlock = grid_desc_m_n.GetLength(I1);
+        const auto MBlock = M / MPerBlock;
+
+        const auto grid_desc_mblock_mperblock_nblock = transform_tensor_descriptor(
+            grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_pass_through_transform(NBlock)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2>{}));
+
+        return grid_desc_mblock_mperblock_nblock;
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2ETileMap>
+    __host__ __device__ static constexpr bool CheckValidity(const AGridDesc_M_K& a_grid_desc_m_k,
+                                                            const BGridDesc_N_K& b_grid_desc_n_k,
+                                                            const DsGridDesc_M_N& ds_grid_desc_m_n,
+                                                            const EGridDesc_M_N& e_grid_desc_m_n,
+                                                            const Block2ETileMap& block_2_etile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        // check consistency of desc
+        if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
+                              N == ds_grid_desc_m_n[i].GetLength(I1));
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check tile size
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // check block-to-E-tile
+        if(!block_2_etile_map.CheckValidity(e_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        // check tensor size: cannot be larger than 2GB each
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_m_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             b_grid_desc_n_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EMeanVarDataType) <= TwoGB))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    using DefaultAGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using DefaultBGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+    using MeanVarGridDescriptor_MBlock_MPerBlock_NBlock =
+        remove_cvref_t<decltype(MakeMeanVarCountGridDescriptor_MBlock_MPerBlock_NBlock(
+            MeanVarGridDesc_M_NBlock{}))>;
+    using CountGridDescriptor_MBlock_MPerBlock_NBlock =
+        remove_cvref_t<decltype(MakeMeanVarCountGridDescriptor_MBlock_MPerBlock_NBlock(
+            CountGridDesc_M_NBlock{}))>;
+    using DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
+
+    using DefaultBlock2ETileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AK0_M_AK1,
+              typename BGridDesc_BK0_N_BK1,
+              typename Block2ETileMap>
+    __device__ static void
+    Run(const ABDataType* __restrict__ p_a_grid,
+        const ABDataType* __restrict__ p_b_grid,
+        DsGridPointer p_ds_grid,
+        EMeanVarDataType* __restrict__ p_e_grid,
+        EMeanVarDataType* __restrict__ p_welford_mean_grid,
+        EMeanVarDataType* __restrict__ p_welford_var_grid,
+        int32_t* __restrict__ p_welford_count,
+        void* __restrict__ p_shared,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op,
+        const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+        const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+        const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            ds_grid_desc_mblock_mperblock_nblock_nperblock,
+        const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            e_grid_desc_mblock_mperblock_nblock_nperblock,
+        const MeanVarGridDescriptor_MBlock_MPerBlock_NBlock&
+            mean_var_grid_desc_mblock_mperblock_nblock,
+        const CountGridDescriptor_MBlock_MPerBlock_NBlock& count_grid_desc_mblock_mperblock_nblock,
+        const Block2ETileMap& block_2_etile_map,
+        index_t NRaw)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        auto mean_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_welford_mean_grid, mean_var_grid_desc_mblock_mperblock_nblock.GetElementSpaceSize());
+
+        auto var_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_welford_var_grid, mean_var_grid_desc_mblock_mperblock_nblock.GetElementSpaceSize());
+
+        auto welford_count_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_welford_count, count_grid_desc_mblock_mperblock_nblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_etile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<ABDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ABDataType,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C, Welford and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>,
+                                  false>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_der_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>,
+                                  false>{};
+
+            // LDS c_shuffle_block_desc_mperblock_nperblock
+            constexpr auto c_shuffle_block_desc_mperblock_nperblock = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1)),
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I3))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<>{}, Sequence<1>{}));
+
+            static_assert(PostShuffleThreadClusterSize_M_N::At(I0) *
+                                  PostShuffleThreadClusterSize_M_N::At(I1) ==
+                              BlockSize,
+                          "wrong!");
+
+            static_assert((CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) %
+                                      PostShuffleThreadClusterSize_M_N::At(I0) ==
+                                  0 &&
+                              (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) %
+                                      PostShuffleThreadClusterSize_M_N::At(I1) ==
+                                  0,
+                          "wrong!");
+
+            constexpr index_t PostShuffleThreadSliceSize_M =
+                (CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) /
+                PostShuffleThreadClusterSize_M_N::At(I0);
+
+            constexpr index_t PostShuffleThreadSliceSize_N =
+                (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) /
+                PostShuffleThreadClusterSize_M_N::At(I1);
+
+            constexpr auto PostShuffleThreadSliceSize_M_N =
+                Sequence<PostShuffleThreadSliceSize_M, PostShuffleThreadSliceSize_N>{};
+
+            // VGPR post_shuffle_thread_desc_m_n
+            constexpr auto post_shuffle_thread_desc_m_n = make_naive_tensor_descriptor_packed(
+                make_tuple(Number<PostShuffleThreadSliceSize_M>{},
+                           Number<PostShuffleThreadSliceSize_N>{}));
+
+            auto e_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
+                post_shuffle_thread_desc_m_n.GetElementSpaceSize());
+
+            // To apply D0, D1, ... and Welford.
+            // threadwise copy from LDS to VGPR
+            constexpr auto post_shuffle_thread_cluster_desc =
+                make_cluster_descriptor(PostShuffleThreadClusterSize_M_N{}, Sequence<0, 1>{});
+
+            const auto post_shuffle_thread_cluster_idx =
+                post_shuffle_thread_cluster_desc.CalculateBottomIndex(
+                    make_multi_index(get_thread_local_1d_id()));
+
+            const auto post_shuffle_thread_data_idx_begin =
+                post_shuffle_thread_cluster_idx * PostShuffleThreadSliceSize_M_N;
+
+            // To apply D0, D1, ... and Welford.
+            // Copy c shuffle from LDS back to VGPR
+            auto post_shuffle_thread_copy_lds_to_vgpr =
+                ThreadwiseTensorSliceTransfer_v2<CShuffleDataType,
+                                                 AccDataType,
+                                                 decltype(c_shuffle_block_desc_mperblock_nperblock),
+                                                 decltype(post_shuffle_thread_desc_m_n),
+                                                 decltype(PostShuffleThreadSliceSize_M_N),
+                                                 Sequence<0, 1>,
+                                                 1,
+                                                 PostShuffleScalarPerVector,
+                                                 1,
+                                                 true>{c_shuffle_block_desc_mperblock_nperblock,
+                                                       post_shuffle_thread_data_idx_begin};
+
+            // D0, D1, ..., Dn
+            constexpr auto post_shuffle_thread_desc_I1_mperblock_I1_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1,
+                               Number<PostShuffleThreadSliceSize_M>{},
+                               I1,
+                               Number<PostShuffleThreadSliceSize_N>{}));
+
+            // FIXME: Decrease usage of VGPR
+            // Apply pointwise lambda function from multi-source (Global and LDS) into VGPR
+            auto ds_thread_buf = generate_tuple(
+                [&](auto) {
+                    return make_static_buffer<AddressSpaceEnum::Vgpr, CShuffleDataType>(
+                        post_shuffle_thread_desc_I1_mperblock_I1_nperblock.GetElementSpaceSize());
+                },
+                Number<NumDTensor>{});
+
+            // Copy D0, D1, ..., Dn from global to VGPR
+            auto ds_thread_copy_global_to_vgpr = generate_tuple(
+                [&](auto I) {
+                    using DDataType = remove_cvref_t<tuple_element_t<I.value, DsDataType>>;
+                    return ThreadwiseTensorSliceTransfer_v2<
+                        DDataType,
+                        AccDataType,
+                        decltype(ds_grid_desc_mblock_mperblock_nblock_nperblock[I]),
+                        decltype(post_shuffle_thread_desc_I1_mperblock_I1_nperblock),
+                        Sequence<I1,
+                                 PostShuffleThreadSliceSize_M,
+                                 I1,
+                                 PostShuffleThreadSliceSize_N>,
+                        Sequence<0, 1, 2, 3>,
+                        3,
+                        PostShuffleScalarPerVector,
+                        1,
+                        true>(
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock[I],
+                        make_multi_index(
+                            I0,
+                            m_block_data_idx_on_grid + post_shuffle_thread_data_idx_begin[I0],
+                            I0,
+                            n_block_data_idx_on_grid + post_shuffle_thread_data_idx_begin[I1]));
+                },
+                Number<NumDTensor>{});
+
+            auto e_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
+                AccDataType,
+                EMeanVarDataType,
+                decltype(post_shuffle_thread_desc_I1_mperblock_I1_nperblock),
+                decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                tensor_operation::element_wise::PassThrough,
+                Sequence<I1,
+                         PostShuffleThreadSliceSize_M,
+                         I1,
+                         PostShuffleThreadSliceSize_N>, // SliceLengths
+                Sequence<0, 1, 2, 3>,                   // DimAccessOrder
+                3,                                      // DstVectorDim
+                PostShuffleScalarPerVector,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                make_multi_index(I0,
+                                 m_block_data_idx_on_grid + post_shuffle_thread_data_idx_begin[I0],
+                                 I0,
+                                 n_block_data_idx_on_grid + post_shuffle_thread_data_idx_begin[I1]),
+                tensor_operation::element_wise::PassThrough{}};
+
+            // Welford
+            constexpr auto thread_welford_src_desc_m_k = make_naive_tensor_descriptor_packed(
+                make_tuple(Number<PostShuffleThreadSliceSize_M>{},
+                           Number<PostShuffleThreadSliceSize_N>{}));
+
+            constexpr auto thread_welford_dst_desc_m = make_naive_tensor_descriptor_packed(
+                make_tuple(Number<PostShuffleThreadSliceSize_M>{}));
+
+            using ThreadwiseWelford = ThreadwiseWelford<AccDataType,
+                                                        decltype(thread_welford_src_desc_m_k),
+                                                        decltype(thread_welford_dst_desc_m)>;
+
+            using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                                      BlockSize,
+                                                      PostShuffleThreadClusterSize_M_N,
+                                                      Sequence<0, 1>,
+                                                      false>;
+
+            constexpr int num_shuffleM =
+                MPerBlock / (CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl);
+
+            constexpr int num_shuffleN =
+                NPerBlock / (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl);
+
+            using mean_var_vgpr_type =
+                decltype(make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
+                    thread_welford_dst_desc_m.GetElementSpaceSize()));
+
+            using welford_count_vgpr_type =
+                decltype(make_static_buffer<AddressSpaceEnum::Vgpr, int32_t>(
+                    thread_welford_dst_desc_m.GetElementSpaceSize()));
+
+            Array<ThreadwiseWelford, num_shuffleM> threadwise_welfords;
+            Array<mean_var_vgpr_type, num_shuffleM> mean_thread_bufs;
+            Array<mean_var_vgpr_type, num_shuffleM> var_thread_bufs;
+            Array<welford_count_vgpr_type, num_shuffleM> welford_count_thread_bufs;
+
+            int max_count     = PostShuffleThreadSliceSize_N * num_shuffleN;
+            const auto nblock = mean_var_grid_desc_mblock_mperblock_nblock.GetLength(I2);
+
+            // tail block
+            if(block_work_idx[I1] % nblock == nblock - 1)
+            {
+                constexpr index_t NPerShuffleBlock =
+                    CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl;
+
+                int NPerBlockTail = NRaw - NPerBlock * (nblock - 1);
+                int thread_max_len =
+                    PostShuffleThreadSliceSize_N * (post_shuffle_thread_cluster_idx[I1] + 1);
+                int shuffle_step = 0;
+                while(thread_max_len <= NPerBlockTail && shuffle_step < num_shuffleN)
+                {
+                    ++shuffle_step;
+                    thread_max_len += NPerShuffleBlock;
+                }
+
+                int delta = 0;
+                if(thread_max_len - NPerBlockTail > PostShuffleThreadSliceSize_N)
+                    delta = 0;
+                else if(NPerBlockTail > thread_max_len)
+                    delta = PostShuffleThreadSliceSize_N;
+                else
+                    delta = PostShuffleThreadSliceSize_N - thread_max_len + NPerBlockTail;
+
+                max_count = shuffle_step * PostShuffleThreadSliceSize_N + delta;
+            }
+
+            static_for<0, num_shuffleM, 1>{}([&](auto i) {
+                threadwise_welfords(i).max_count_ = max_count;
+                mean_thread_bufs(i) = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
+                    thread_welford_dst_desc_m.GetElementSpaceSize());
+
+                var_thread_bufs(i) = make_static_buffer<AddressSpaceEnum::Vgpr, AccDataType>(
+                    thread_welford_dst_desc_m.GetElementSpaceSize());
+
+                welford_count_thread_bufs(i) = make_static_buffer<AddressSpaceEnum::Vgpr, int32_t>(
+                    thread_welford_dst_desc_m.GetElementSpaceSize());
+
+                static_for<0, PostShuffleThreadSliceSize_M, 1>{}([&](auto j) {
+                    mean_thread_bufs(i)(j)          = type_convert<AccDataType>(0.0f);
+                    var_thread_bufs(i)(j)           = type_convert<AccDataType>(0.0f);
+                    welford_count_thread_bufs(i)(j) = 0;
+                });
+            });
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_der_global.GetNumOfAccess(), "wrong!");
+
+            int shuffleM_index = __builtin_amdgcn_readfirstlane(0);
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each thread shuffle data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // Get shuffle data from LDS to VGPR
+                post_shuffle_thread_copy_lds_to_vgpr.Run(c_shuffle_block_desc_mperblock_nperblock,
+                                                         c_shuffle_block_buf,
+                                                         post_shuffle_thread_desc_m_n,
+                                                         make_tuple(I0, I0),
+                                                         e_thread_buf);
+
+                // Global read D0, D1, ...
+                static_for<0, NumDTensor, 1>{}([&](auto Id) {
+                    auto& d_thread_copy_global_to_vgpr = ds_thread_copy_global_to_vgpr(Id);
+                    d_thread_copy_global_to_vgpr.Run(
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock[Id],
+                        ds_grid_buf[Id],
+                        post_shuffle_thread_desc_I1_mperblock_I1_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        ds_thread_buf(Id));
+
+                    if constexpr(access_id < num_access - 1)
+                    {
+                        // move on D0, D1, ...
+                        constexpr auto de_global_step = sfc_der_global.GetForwardStep(access_id);
+                        d_thread_copy_global_to_vgpr.MoveSrcSliceWindow(
+                            ds_grid_desc_mblock_mperblock_nblock_nperblock[Id], de_global_step);
+                    }
+                });
+
+                // cde_element_op(e, c, d0, d1, ...);
+                static_for<0, post_shuffle_thread_desc_m_n.GetElementSize(), 1>{}([&](auto i) {
+                    const auto c_ds_src_data_refs = concat_tuple_of_reference(
+                        tie(e_thread_buf[i]),
+                        generate_tie(
+                            [&](auto Id) -> const auto& { return ds_thread_buf[Id][i]; },
+                            Number<NumDTensor>{}));
+                    auto e_dst_data_refs = tie(e_thread_buf(i));
+                    unpack2(cde_element_op, e_dst_data_refs, c_ds_src_data_refs);
+                });
+
+                // Global write E
+                e_thread_copy_vgpr_to_global.Run(post_shuffle_thread_desc_I1_mperblock_I1_nperblock,
+                                                 make_tuple(I0, I0, I0, I0),
+                                                 e_thread_buf,
+                                                 e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                 e_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    // move on E
+                    constexpr auto de_global_step = sfc_der_global.GetForwardStep(access_id);
+                    e_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                        e_grid_desc_mblock_mperblock_nblock_nperblock, de_global_step);
+                }
+
+                // Threadwise welford
+                auto& threadwise_welford = threadwise_welfords(shuffleM_index);
+                auto& mean_thread_buf    = mean_thread_bufs(shuffleM_index);
+                auto& var_thread_buf     = var_thread_bufs(shuffleM_index);
+
+                threadwise_welford.Run(e_thread_buf, mean_thread_buf, var_thread_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto de_global_step = sfc_der_global.GetForwardStep(access_id);
+                    constexpr int shuffleMInc =
+                        de_global_step[I1] /
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1);
+                    shuffleM_index = __builtin_amdgcn_readfirstlane(shuffleM_index + shuffleMInc);
+                }
+            }); // copy c, d, e + welford
+
+            // Blockwise welford and write out
+            static_for<0, num_shuffleM, 1>{}([&](auto i) {
+                auto& mean_thread_buf  = mean_thread_bufs(i);
+                auto& var_thread_buf   = var_thread_bufs(i);
+                auto& count_thread_buf = welford_count_thread_bufs(i);
+
+                static_for<0, PostShuffleThreadSliceSize_M, 1>{}([&](auto j) {
+                    block_sync_lds();
+                    count_thread_buf(j) = threadwise_welfords(i).cur_count_;
+                    BlockwiseWelford::Run(
+                        mean_thread_buf(j), var_thread_buf(j), count_thread_buf(j));
+                });
+
+                if(post_shuffle_thread_cluster_idx[I1] == 0)
+                {
+                    constexpr auto thread_welford_desc_I_m_I = make_naive_tensor_descriptor_packed(
+                        make_tuple(I1, Number<PostShuffleThreadSliceSize_M>{}, I1));
+
+                    constexpr int shuffleMPerBlock =
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1);
+
+                    auto mean_var_count_thread_copy_index = make_multi_index(
+                        block_work_idx[I0],                                            // mblock
+                        shuffleMPerBlock * i + post_shuffle_thread_data_idx_begin[I0], // mperblock
+                        block_work_idx[I1]);                                           // nblock
+
+                    auto mean_var_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
+                        AccDataType,
+                        EMeanVarDataType,
+                        decltype(thread_welford_desc_I_m_I),
+                        decltype(mean_var_grid_desc_mblock_mperblock_nblock),
+                        tensor_operation::element_wise::PassThrough,
+                        Sequence<1, PostShuffleThreadSliceSize_M, 1>,
+                        Sequence<0, 1, 2>,
+                        1,
+                        1,
+                        InMemoryDataOperationEnum::Set,
+                        1,
+                        true>{mean_var_grid_desc_mblock_mperblock_nblock,
+                              mean_var_count_thread_copy_index,
+                              tensor_operation::element_wise::PassThrough{}};
+
+                    mean_var_thread_copy_vgpr_to_global.Run(
+                        thread_welford_desc_I_m_I,
+                        make_tuple(I0, I0, I0),
+                        mean_thread_buf,
+                        mean_var_grid_desc_mblock_mperblock_nblock,
+                        mean_grid_buf); // write mean
+
+                    mean_var_thread_copy_vgpr_to_global.Run(
+                        thread_welford_desc_I_m_I,
+                        make_tuple(I0, I0, I0),
+                        var_thread_buf,
+                        mean_var_grid_desc_mblock_mperblock_nblock,
+                        var_grid_buf); // write variance
+
+                    // Stride of count is [0, 1]. Only the first row in count[0, 0:nblock] need
+                    // to be written.
+                    if(i == 0 && block_work_idx[I0] == 0 &&
+                       post_shuffle_thread_cluster_idx[I0] == 0)
+                    {
+                        auto count_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
+                            int32_t,
+                            int32_t,
+                            decltype(thread_welford_desc_I_m_I),
+                            decltype(count_grid_desc_mblock_mperblock_nblock),
+                            tensor_operation::element_wise::PassThrough,
+                            Sequence<1, PostShuffleThreadSliceSize_M, 1>,
+                            Sequence<0, 1, 2>,
+                            1,
+                            1,
+                            InMemoryDataOperationEnum::Set,
+                            1,
+                            false>{count_grid_desc_mblock_mperblock_nblock,
+                                   mean_var_count_thread_copy_index,
+                                   tensor_operation::element_wise::PassThrough{}};
+
+                        count_thread_copy_vgpr_to_global.Run(
+                            thread_welford_desc_I_m_I,
+                            make_tuple(I0, I0, I0),
+                            count_thread_buf,
+                            count_grid_desc_mblock_mperblock_nblock,
+                            welford_count_grid_buf); // write count
+                    }
+                }
+            });
+
+        } // shuffle C + Ds + welford + write out
+    }     // run
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_welford_second_half_layernorm2d.hpp b/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_welford_second_half_layernorm2d.hpp
new file mode 100644
index 0000000000..69468c25be
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_welford_second_half_layernorm2d.hpp
@@ -0,0 +1,394 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+
+namespace ck {
+
+template <typename EMeanVarDataType,
+          typename HDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename ComputeDataType,
+          typename EHGridDesc_M_N,
+          typename MeanVarGridDesc_M_NBlock,
+          typename CountGridDesc_M_NBlock,
+          typename GammaBetaGridDesc_N,
+          typename HElementwiseOperation,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t NThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t NThreadSliceSize,
+          index_t ESrcVectorSize,
+          index_t HDstVectorSize,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorSize>
+struct GridwiseWelfordSecondHalfLayernorm2d
+{
+    static_assert(NThreadSliceSize % ESrcVectorSize == 0 &&
+                      NThreadSliceSize % GammaSrcVectorSize == 0 &&
+                      NThreadSliceSize % BetaSrcVectorSize == 0,
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert(NThreadSliceSize % HDstVectorSize == 0,
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    using ThreadClusterLengths_M_N   = Sequence<MThreadClusterSize, NThreadClusterSize>;
+    using ThreadBufferDimAccessOrder = Sequence<0, 1>;
+    using ThreadClusterArrangeOrder  = Sequence<0, 1>;
+
+    static constexpr auto thread_cluster_desc_m_n =
+        make_cluster_descriptor(ThreadClusterLengths_M_N{}, ThreadClusterArrangeOrder{});
+
+    using ThreadBufferLengths_M_N                = Sequence<MThreadSliceSize, NThreadSliceSize>;
+    static constexpr auto thread_buffer_desc_m_n = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<NThreadSliceSize>{}));
+
+    using ThreadBufferLengths_M_1 = Sequence<MThreadSliceSize, 1>;
+    static constexpr auto thread_buffer_desc_m_1 =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}, Number<1>{}));
+
+    using ThreadBufferLengths_N = Sequence<NThreadSliceSize>;
+    static constexpr auto thread_buffer_desc_n =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<NThreadSliceSize>{}));
+
+    using ThreadWelfordSrcDesc_M_1 = decltype(thread_buffer_desc_m_1);
+    using ThreadWelfordDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelfordMerge<ComputeDataType, ThreadWelfordSrcDesc_M_1, ThreadWelfordDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<ComputeDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_N,
+                                              ThreadClusterArrangeOrder>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t N_BlockTileSize = NThreadClusterSize * NThreadSliceSize;
+
+    __device__ static void Run(const EMeanVarDataType* __restrict__ p_e_grid,
+                               const EMeanVarDataType* __restrict__ p_in_welford_mean_grid,
+                               const EMeanVarDataType* __restrict__ p_in_welford_var_grid,
+                               const int32_t* __restrict__ p_in_welford_count_grid,
+                               const GammaDataType* __restrict__ p_gamma_grid,
+                               const BetaDataType* __restrict__ p_beta_grid,
+                               HDataType* __restrict__ p_h_grid,
+                               const EHGridDesc_M_N& e_grid_desc_m_n,
+                               const EHGridDesc_M_N& h_grid_desc_m_n,
+                               const MeanVarGridDesc_M_NBlock& mean_var_grid_desc_m_nblock,
+                               const CountGridDesc_M_NBlock& count_grid_desc_m_nblock,
+                               const GammaBetaGridDesc_N& gamma_grid_desc_n,
+                               const GammaBetaGridDesc_N& beta_grid_desc_n,
+                               index_t numMeanVarCountBlockTileIteration_N,
+                               index_t NBlockClusterLength,
+                               ComputeDataType epsilon,
+                               HElementwiseOperation h_element_op)
+    {
+        // Thread/Block id
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const auto block_work_idx     = make_tuple(block_global_id / NBlockClusterLength,
+                                               block_global_id % NBlockClusterLength);
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc_m_n.CalculateBottomIndex(make_multi_index(thread_local_id));
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_n_cluster_id = thread_cluster_idx[I1];
+
+        // Global Memory
+        const auto e_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_m_n.GetElementSpaceSize());
+
+        const auto welford_mean_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_welford_mean_grid, mean_var_grid_desc_m_nblock.GetElementSpaceSize());
+
+        const auto welford_var_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_welford_var_grid, mean_var_grid_desc_m_nblock.GetElementSpaceSize());
+
+        const auto welford_count_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_welford_count_grid, count_grid_desc_m_nblock.GetElementSpaceSize());
+
+        const auto gamma_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_gamma_grid, gamma_grid_desc_n.GetElementSpaceSize());
+
+        const auto beta_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_beta_grid, beta_grid_desc_n.GetElementSpaceSize());
+
+        auto h_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_h_grid, h_grid_desc_m_n.GetElementSpaceSize());
+
+        // VGPR
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            in_welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            in_welford_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true>
+            in_welford_count_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            welford_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            welford_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true>
+            welford_count_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     ComputeDataType,
+                     MThreadSliceSize * NThreadSliceSize,
+                     true>
+            e_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     ComputeDataType,
+                     MThreadSliceSize * NThreadSliceSize,
+                     true>
+            gamma_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     ComputeDataType,
+                     MThreadSliceSize * NThreadSliceSize,
+                     true>
+            beta_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     ComputeDataType,
+                     MThreadSliceSize * NThreadSliceSize,
+                     true>
+            h_thread_buf;
+
+        // IO
+        auto threadwise_mean_load_m_nblock =
+            ThreadwiseTensorSliceTransfer_v2<EMeanVarDataType,
+                                             ComputeDataType,
+                                             MeanVarGridDesc_M_NBlock,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             ThreadBufferDimAccessOrder,
+                                             1,
+                                             1,
+                                             1,
+                                             true>(
+                mean_var_grid_desc_m_nblock,
+                make_multi_index(block_work_idx[I0] * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_n_cluster_id));
+
+        auto threadwise_var_load_m_nblock =
+            ThreadwiseTensorSliceTransfer_v2<EMeanVarDataType,
+                                             ComputeDataType,
+                                             MeanVarGridDesc_M_NBlock,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             ThreadBufferDimAccessOrder,
+                                             1,
+                                             1,
+                                             1,
+                                             true>(
+                mean_var_grid_desc_m_nblock,
+                make_multi_index(block_work_idx[I0] * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_n_cluster_id));
+
+        auto threadwise_count_load_m_nblock =
+            ThreadwiseTensorSliceTransfer_v2<int32_t,
+                                             int32_t,
+                                             CountGridDesc_M_NBlock,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             ThreadBufferDimAccessOrder,
+                                             1,
+                                             1,
+                                             1,
+                                             true>(
+                count_grid_desc_m_nblock,
+                make_multi_index(block_work_idx[I0] * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_n_cluster_id));
+
+        auto threadwise_e_load_m_n =
+            ThreadwiseTensorSliceTransfer_v2<EMeanVarDataType,
+                                             ComputeDataType,
+                                             decltype(e_grid_desc_m_n),
+                                             decltype(thread_buffer_desc_m_n),
+                                             ThreadBufferLengths_M_N,
+                                             ThreadBufferDimAccessOrder,
+                                             1, // SrcVectorDim
+                                             ESrcVectorSize,
+                                             1,
+                                             true>(
+                e_grid_desc_m_n,
+                make_multi_index(
+                    block_work_idx[I0] * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                    block_work_idx[I1] * N_BlockTileSize + thread_n_cluster_id * NThreadSliceSize));
+
+        auto threadwise_gamma_load_n =
+            ThreadwiseTensorSliceTransfer_v2<GammaDataType,
+                                             ComputeDataType,
+                                             decltype(gamma_grid_desc_n),
+                                             decltype(thread_buffer_desc_n),
+                                             ThreadBufferLengths_N,
+                                             Sequence<0>, // DimAccessOrder,
+                                             0,           // SrcVectorDim,
+                                             GammaSrcVectorSize,
+                                             1,
+                                             true>(
+                gamma_grid_desc_n,
+                make_multi_index(block_work_idx[I1] * N_BlockTileSize +
+                                 thread_n_cluster_id * NThreadSliceSize));
+
+        auto threadwise_beta_load_n =
+            ThreadwiseTensorSliceTransfer_v2<BetaDataType,
+                                             ComputeDataType,
+                                             decltype(beta_grid_desc_n),
+                                             decltype(thread_buffer_desc_n),
+                                             ThreadBufferLengths_N,
+                                             Sequence<0>, // DimAccessOrder,
+                                             0,           // SrcVectorDim,
+                                             BetaSrcVectorSize,
+                                             1,
+                                             true>(
+                beta_grid_desc_n,
+                make_multi_index(block_work_idx[I1] * N_BlockTileSize +
+                                 thread_n_cluster_id * NThreadSliceSize));
+
+        auto threadwise_h_store_m_n =
+            ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
+                                               HDataType,
+                                               decltype(thread_buffer_desc_m_n),
+                                               decltype(h_grid_desc_m_n),
+                                               HElementwiseOperation,
+                                               ThreadBufferLengths_M_N,
+                                               ThreadBufferDimAccessOrder,
+                                               1, // DstVectorDim
+                                               HDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                h_grid_desc_m_n,
+                make_multi_index(
+                    block_work_idx[I0] * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                    block_work_idx[I1] * N_BlockTileSize + thread_n_cluster_id * NThreadSliceSize),
+                h_element_op);
+
+        // step1: Merge mean and variance
+        constexpr auto mean_var_count_thread_copy_step_I0_n =
+            make_multi_index(I0, NThreadClusterSize);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            welford_mean_thread_buf(I)  = type_convert<ComputeDataType>(0.0f);
+            welford_var_thread_buf(I)   = type_convert<ComputeDataType>(0.0f);
+            welford_count_thread_buf(I) = 0;
+        });
+
+        for(index_t n = 0; n < numMeanVarCountBlockTileIteration_N; ++n)
+        {
+            threadwise_mean_load_m_nblock.Run(mean_var_grid_desc_m_nblock,
+                                              welford_mean_global_val_buf,
+                                              thread_buffer_desc_m_1,
+                                              make_tuple(I0, I0),
+                                              in_welford_mean_thread_buf);
+
+            threadwise_var_load_m_nblock.Run(mean_var_grid_desc_m_nblock,
+                                             welford_var_global_val_buf,
+                                             thread_buffer_desc_m_1,
+                                             make_tuple(I0, I0),
+                                             in_welford_var_thread_buf);
+
+            threadwise_count_load_m_nblock.Run(count_grid_desc_m_nblock,
+                                               welford_count_global_val_buf,
+                                               thread_buffer_desc_m_1,
+                                               make_tuple(I0, I0),
+                                               in_welford_count_thread_buf);
+
+            ThreadwiseWelford::Run(in_welford_mean_thread_buf,
+                                   in_welford_var_thread_buf,
+                                   in_welford_count_thread_buf,
+                                   welford_mean_thread_buf,
+                                   welford_var_thread_buf,
+                                   welford_count_thread_buf);
+
+            threadwise_mean_load_m_nblock.MoveSrcSliceWindow(mean_var_grid_desc_m_nblock,
+                                                             mean_var_count_thread_copy_step_I0_n);
+            threadwise_var_load_m_nblock.MoveSrcSliceWindow(mean_var_grid_desc_m_nblock,
+                                                            mean_var_count_thread_copy_step_I0_n);
+            threadwise_count_load_m_nblock.MoveSrcSliceWindow(count_grid_desc_m_nblock,
+                                                              mean_var_count_thread_copy_step_I0_n);
+        }
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            BlockwiseWelford::Run(
+                welford_mean_thread_buf(I), welford_var_thread_buf(I), welford_count_thread_buf(I));
+        });
+
+        // step2: normalization
+        // h[m, n] = [(e[m, n] - mean[m]) / sqrt(var[m] + eps)] * gamma[n] + beta[n]
+        threadwise_e_load_m_n.Run(e_grid_desc_m_n,
+                                  e_global_val_buf,
+                                  thread_buffer_desc_m_n,
+                                  make_tuple(I0, I0),
+                                  e_thread_buf);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto m) {
+            auto divisor = 1 / ck::math::sqrt(welford_var_thread_buf(m) + epsilon);
+            static_for<0, NThreadSliceSize, 1>{}([&](auto n) {
+                constexpr auto m_n = thread_buffer_desc_m_n.CalculateOffset(make_tuple(m, n));
+                h_thread_buf(Number<m_n>{}) =
+                    (e_thread_buf(Number<m_n>{}) - welford_mean_thread_buf(m)) * divisor;
+            });
+        });
+
+        threadwise_gamma_load_n.Run(gamma_grid_desc_n,
+                                    gamma_global_val_buf,
+                                    thread_buffer_desc_n,
+                                    make_tuple(I0),
+                                    gamma_thread_buf);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto m) {
+            static_for<0, NThreadSliceSize, 1>{}([&](auto n) {
+                constexpr auto m_n = thread_buffer_desc_m_n.CalculateOffset(make_tuple(m, n));
+                h_thread_buf(Number<m_n>{}) = h_thread_buf(Number<m_n>{}) * gamma_thread_buf(n);
+            });
+        });
+
+        threadwise_beta_load_n.Run(beta_grid_desc_n,
+                                   beta_global_val_buf,
+                                   thread_buffer_desc_n,
+                                   make_tuple(I0),
+                                   beta_thread_buf);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto m) {
+            static_for<0, NThreadSliceSize, 1>{}([&](auto n) {
+                constexpr auto m_n = thread_buffer_desc_m_n.CalculateOffset(make_tuple(m, n));
+                h_thread_buf(Number<m_n>{}) = h_thread_buf(Number<m_n>{}) + beta_thread_buf(n);
+            });
+        });
+
+        threadwise_h_store_m_n.Run(thread_buffer_desc_m_n,
+                                   make_tuple(I0, I0),
+                                   h_thread_buf,
+                                   h_grid_desc_m_n,
+                                   h_global_val_buf);
+
+    } // run
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp
new file mode 100644
index 0000000000..bd1e0585fc
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp
@@ -0,0 +1,321 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseMultipleReduction,
+          index_t NumReduction,
+          typename InDataType,
+          typename OutDataTypePointerTuple,
+          typename AccDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M_Tuple,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple>
+__global__ void
+kernel_multiple_reduce_multiblock(const InGridDesc_M_K in_grid_desc_m_k,
+                                  const OutGridDesc_M_Tuple out_grid_desc_m_tuple,
+                                  const InElementwiseOperationTuple in_elementwise_op_tuple,
+                                  const AccElementwiseOperationTuple acc_elementwise_op_tuple,
+                                  index_t block_group_size,
+                                  index_t num_k_block_tile_iteration,
+                                  Array<AccDataType, NumReduction> alpha_values,
+                                  const InDataType* const __restrict__ p_in_value_global,
+                                  Array<AccDataType, NumReduction> beta_values,
+                                  OutDataTypePointerTuple p_out_value_global_tuple)
+{
+    GridwiseMultipleReduction::Run(in_grid_desc_m_k,
+                                   out_grid_desc_m_tuple,
+                                   in_elementwise_op_tuple,
+                                   acc_elementwise_op_tuple,
+                                   block_group_size,
+                                   num_k_block_tile_iteration,
+                                   alpha_values,
+                                   p_in_value_global,
+                                   beta_values,
+                                   p_out_value_global_tuple);
+};
+
+template <index_t NumReduction,
+          typename InDataType,
+          typename OutDataTypePointerTuple,
+          typename AccDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M_Tuple,
+          typename ReduceOperation,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple,
+          InMemoryDataOperationEnum OutMemoryDataOperation,
+          bool PropagateNan,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          typename OutDstVectorSizeSeq>
+struct GridwiseMultipleReduction_mk_to_m_multiblock
+{
+    static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                   (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert(NumReduction == OutDataTypePointerTuple::Size() &&
+                      NumReduction == OutGridDesc_M_Tuple::Size() &&
+                      NumReduction == OutDstVectorSizeSeq::Size() &&
+                      NumReduction == InElementwiseOperationTuple::Size() &&
+                      NumReduction == AccElementwiseOperationTuple::Size(),
+                  "All tuple should have the same size as the number of Reductions!");
+
+    static constexpr bool reorder_thread_cluster = (InSrcVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using BlockwiseReduce = PartitionedBlockwiseReduction<AccDataType,
+                                                          BlockSize,
+                                                          ThreadClusterLengths_M_K,
+                                                          ThreadClusterArrangeOrder,
+                                                          ReduceOperation,
+                                                          PropagateNan>;
+
+    using ThreadwiseReduce = ThreadwiseReduction<AccDataType,
+                                                 ThreadReduceSrcDesc_M_K,
+                                                 ThreadReduceDstDesc_M,
+                                                 ReduceOperation,
+                                                 PropagateNan>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    using Accumulation = detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
+
+    __device__ static void Run(const InGridDesc_M_K& in_grid_desc_m_k,
+                               const OutGridDesc_M_Tuple& out_grid_desc_m_tuple,
+                               const InElementwiseOperationTuple& in_elementwise_op_tuple,
+                               const AccElementwiseOperationTuple& acc_elementwise_op_tuple,
+                               index_t block_group_size,
+                               index_t num_k_block_tile_iteration,
+                               Array<AccDataType, NumReduction> alpha_values,
+                               const InDataType* const __restrict__ p_in_value_global,
+                               Array<AccDataType, NumReduction> beta_values,
+                               OutDataTypePointerTuple p_out_value_global_tuple)
+    {
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+        // LDS,  reused by all reductions
+        __shared__ AccDataType p_reduce_work_buffer[BlockSize];
+
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
+        auto out_global_val_buf_tuple = generate_tuple(
+            [&](auto iR) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_out_value_global_tuple[iR], out_grid_desc_m_tuple[iR].GetElementSpaceSize());
+            },
+            Number<NumReduction>{});
+
+        auto reduce_work_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            in_thread_buf;
+
+        auto in_thread_buf_tuple = generate_tuple(
+            [&](auto iR) {
+                (void)iR;
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * KThreadSliceSize,
+                                    true>{};
+            },
+            Number<NumReduction>{});
+
+        auto accu_value_buf_tuple = generate_tuple(
+            [&](auto iR) {
+                (void)iR;
+                return StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>{};
+            },
+            Number<NumReduction>{});
+
+        static_for<0, NumReduction, 1>{}([&](auto iR) {
+            static_for<0, MThreadSliceSize, 1>{}(
+                [&](auto J) { accu_value_buf_tuple(iR)(J) = identityVal; });
+        });
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / block_group_size;
+        const index_t block_local_id  = block_global_id % block_group_size;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        const index_t reduceSizePerBlock = K_BlockTileSize * num_k_block_tile_iteration;
+
+        using ThreadBufferLengths         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        constexpr auto thread_buffer_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+
+        auto threadwise_src_load = ThreadwiseTensorSliceTransfer_v2<InDataType,
+                                                                    AccDataType,
+                                                                    InGridDesc_M_K,
+                                                                    decltype(thread_buffer_desc),
+                                                                    ThreadBufferLengths,
+                                                                    ThreadBufferDimAccessOrder,
+                                                                    InSrcVectorDim,
+                                                                    InSrcVectorSize,
+                                                                    1,
+                                                                    false>(
+            in_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             block_local_id * reduceSizePerBlock +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        constexpr auto in_thread_copy_step = make_multi_index(0, K_BlockTileSize);
+
+        index_t reducedTiles = 0;
+        do
+        {
+            threadwise_src_load.Run(in_grid_desc_m_k,
+                                    in_global_val_buf,
+                                    thread_buffer_desc,
+                                    make_tuple(I0, I0),
+                                    in_thread_buf);
+
+            static_for<0, NumReduction, 1>{}([&](auto iR) {
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    // do element-wise pre-reduction operation
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+                        in_elementwise_op_tuple[iR](in_thread_buf_tuple(iR)(Number<offset>{}),
+                                                    in_thread_buf(Number<offset>{}));
+                    });
+                });
+
+                ThreadwiseReduce::Reduce(in_thread_buf_tuple(iR), accu_value_buf_tuple(iR));
+            });
+
+            threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+            reducedTiles++;
+        } while(reducedTiles < num_k_block_tile_iteration);
+
+        constexpr auto reduced_data_desc = ThreadReduceDstDesc_M{};
+
+        static_for<0, NumReduction, 1>{}([&](auto iR) {
+            using OutDataTypePointer = remove_cvref_t<decltype(OutDataTypePointerTuple{}[iR])>;
+            using OutDataType        = remove_cvref_t<remove_pointer_t<OutDataTypePointer>>;
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                BlockwiseReduce::Reduce(reduce_work_buf, accu_value_buf_tuple(iR)(I));
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if(thread_k_cluster_id == 0)
+                {
+                    acc_elementwise_op_tuple[iR](accu_value_buf_tuple(iR)(I),
+                                                 accu_value_buf_tuple(iR)(I));
+
+                    accu_value_buf_tuple(iR)(I) *= alpha_values[iR];
+                }
+            });
+
+            if(thread_k_cluster_id == 0)
+            {
+                if(block_group_size == 0 && !float_equal_zero{}(beta_values[iR]))
+                {
+                    StaticBuffer<AddressSpaceEnum::Vgpr, OutDataType, MThreadSliceSize, true>
+                        priorDstValueBuf;
+
+                    auto threadwise_dst_load =
+                        ThreadwiseTensorSliceTransfer_v2<OutDataType,
+                                                         OutDataType,
+                                                         decltype(out_grid_desc_m_tuple[iR]),
+                                                         decltype(reduced_data_desc),
+                                                         Sequence<MThreadSliceSize>,
+                                                         Sequence<0>,
+                                                         0,
+                                                         OutDstVectorSizeSeq::At(iR),
+                                                         1,
+                                                         false>(
+                            out_grid_desc_m_tuple[iR],
+                            make_multi_index(blkgroup_id * M_BlockTileSize +
+                                             thread_m_cluster_id * MThreadSliceSize));
+
+                    threadwise_dst_load.Run(out_grid_desc_m_tuple[iR],
+                                            out_global_val_buf_tuple(iR),
+                                            reduced_data_desc,
+                                            make_tuple(I0),
+                                            priorDstValueBuf);
+
+                    static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                        accu_value_buf_tuple(iR)(I) +=
+                            type_convert<AccDataType>(priorDstValueBuf[I]) * beta_values[iR];
+                    });
+                };
+
+                auto threadwise_dst_store =
+                    ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                       OutDataType,
+                                                       decltype(reduced_data_desc),
+                                                       decltype(out_grid_desc_m_tuple[iR]),
+                                                       PassThroughOp,
+                                                       Sequence<MThreadSliceSize>,
+                                                       Sequence<0>,
+                                                       0,
+                                                       OutDstVectorSizeSeq::At(iR),
+                                                       OutMemoryDataOperation,
+                                                       1,
+                                                       true>(
+                        out_grid_desc_m_tuple[iR],
+                        make_multi_index(blkgroup_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize),
+                        PassThroughOp{});
+
+                threadwise_dst_store.Run(reduced_data_desc,
+                                         make_tuple(I0),
+                                         accu_value_buf_tuple[iR],
+                                         out_grid_desc_m_tuple[iR],
+                                         out_global_val_buf_tuple(iR));
+            };
+        });
+    };
+}; // namespace ck
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp
new file mode 100644
index 0000000000..fc4f27e33b
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp
@@ -0,0 +1,264 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseMultipleReduction,
+          index_t NumReduction,
+          typename InDataType,
+          typename OutDataTypePointerTuple,
+          typename AccDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M_Tuple,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple>
+__global__ void
+kernel_multiple_reduce_threadwise(const InGridDesc_M_K in_grid_desc_m_k,
+                                  const OutGridDesc_M_Tuple out_grid_desc_m_tuple,
+                                  const InElementwiseOperationTuple in_elementwise_op_tuple,
+                                  const AccElementwiseOperationTuple acc_elementwise_op_tuple,
+                                  Array<AccDataType, NumReduction> alpha_values,
+                                  const InDataType* const __restrict__ p_in_value_global,
+                                  Array<AccDataType, NumReduction> beta_values,
+                                  OutDataTypePointerTuple p_out_value_global_tuple)
+{
+    GridwiseMultipleReduction::Run(in_grid_desc_m_k,
+                                   out_grid_desc_m_tuple,
+                                   in_elementwise_op_tuple,
+                                   acc_elementwise_op_tuple,
+                                   alpha_values,
+                                   p_in_value_global,
+                                   beta_values,
+                                   p_out_value_global_tuple);
+};
+
+template <index_t NumReduction,
+          typename InDataType,
+          typename OutDataTypePointerTuple,
+          typename AccDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M_Tuple,
+          typename ReduceOperation,
+          typename InElementwiseOperationTuple,
+          typename AccElementwiseOperationTuple,
+          InMemoryDataOperationEnum OutMemoryDataOperation,
+          bool PropagateNan,
+          index_t BlockSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          typename OutDstVectorSizeSeq>
+struct GridwiseMultipleReduction_mk_to_m_threadwise
+{
+    static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                   (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert(NumReduction == OutDataTypePointerTuple::Size() &&
+                      NumReduction == OutGridDesc_M_Tuple::Size() &&
+                      NumReduction == OutDstVectorSizeSeq::Size() &&
+                      NumReduction == InElementwiseOperationTuple::Size() &&
+                      NumReduction == AccElementwiseOperationTuple::Size(),
+                  "All tuple should have the same size as the number of Reductions!");
+
+    static constexpr bool reorder_thread_cluster = (InSrcVectorDim == 0);
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseReduce = ThreadwiseReduction<AccDataType,
+                                                 ThreadReduceSrcDesc_M_K,
+                                                 ThreadReduceDstDesc_M,
+                                                 ReduceOperation,
+                                                 PropagateNan>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+
+    using Accumulation = detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
+
+    __device__ static void Run(const InGridDesc_M_K& in_grid_desc_m_k,
+                               const OutGridDesc_M_Tuple& out_grid_desc_m_tuple,
+                               const InElementwiseOperationTuple& in_elementwise_op_tuple,
+                               const AccElementwiseOperationTuple& acc_elementwise_op_tuple,
+                               Array<AccDataType, NumReduction> alpha_values,
+                               const InDataType* const __restrict__ p_in_value_global,
+                               Array<AccDataType, NumReduction> beta_values,
+                               OutDataTypePointerTuple p_out_value_global_tuple)
+    {
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
+        auto out_global_val_buf_tuple = generate_tuple(
+            [&](auto iR) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_out_value_global_tuple[iR], out_grid_desc_m_tuple[iR].GetElementSpaceSize());
+            },
+            Number<NumReduction>{});
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            in_thread_buf;
+
+        auto in_thread_buf_tuple = generate_tuple(
+            [&](auto iR) {
+                (void)iR;
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * KThreadSliceSize,
+                                    true>{};
+            },
+            Number<NumReduction>{});
+
+        auto accu_value_buf_tuple = generate_tuple(
+            [&](auto iR) {
+                (void)iR;
+                return StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>{};
+            },
+            Number<NumReduction>{});
+
+        static_for<0, NumReduction, 1>{}([&](auto iR) {
+            static_for<0, MThreadSliceSize, 1>{}(
+                [&](auto J) { accu_value_buf_tuple(iR)(J) = identityVal; });
+        });
+
+        const index_t thread_global_1d_id = get_thread_global_1d_id();
+
+        const auto toReduceLength = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        using ThreadBufferLengths         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        constexpr auto thread_buffer_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+
+        auto threadwise_src_load = ThreadwiseTensorSliceTransfer_v2<InDataType,
+                                                                    AccDataType,
+                                                                    InGridDesc_M_K,
+                                                                    decltype(thread_buffer_desc),
+                                                                    ThreadBufferLengths,
+                                                                    ThreadBufferDimAccessOrder,
+                                                                    InSrcVectorDim,
+                                                                    InSrcVectorSize,
+                                                                    1,
+                                                                    false>(
+            in_grid_desc_m_k, make_multi_index(thread_global_1d_id * MThreadSliceSize, 0));
+
+        constexpr auto in_thread_copy_step = make_multi_index(0, KThreadSliceSize);
+
+        index_t reducedLength = 0;
+        do
+        {
+            threadwise_src_load.Run(in_grid_desc_m_k,
+                                    in_global_val_buf,
+                                    thread_buffer_desc,
+                                    make_tuple(I0, I0),
+                                    in_thread_buf);
+
+            static_for<0, NumReduction, 1>{}([&](auto iR) {
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    // do element-wise pre-reduction operation
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+                        in_elementwise_op_tuple[iR](in_thread_buf_tuple(iR)(Number<offset>{}),
+                                                    in_thread_buf(Number<offset>{}));
+                    });
+                });
+
+                ThreadwiseReduce::Reduce(in_thread_buf_tuple(iR), accu_value_buf_tuple(iR));
+            });
+
+            threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+            reducedLength += KThreadSliceSize;
+        } while(reducedLength < toReduceLength);
+
+        constexpr auto reduced_data_desc = ThreadReduceDstDesc_M{};
+
+        static_for<0, NumReduction, 1>{}([&](auto iR) {
+            using OutDataTypePointer = remove_cvref_t<decltype(OutDataTypePointerTuple{}[iR])>;
+            using OutDataType        = remove_cvref_t<remove_pointer_t<OutDataTypePointer>>;
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                acc_elementwise_op_tuple[iR](accu_value_buf_tuple(iR)(I),
+                                             accu_value_buf_tuple(iR)(I));
+
+                accu_value_buf_tuple(iR)(I) *= alpha_values[iR];
+            });
+
+            if(!float_equal_zero{}(beta_values[iR]))
+            {
+                StaticBuffer<AddressSpaceEnum::Vgpr, OutDataType, MThreadSliceSize, true>
+                    priorDstValueBuf;
+
+                auto threadwise_dst_load =
+                    ThreadwiseTensorSliceTransfer_v2<OutDataType,
+                                                     OutDataType,
+                                                     decltype(out_grid_desc_m_tuple[iR]),
+                                                     decltype(reduced_data_desc),
+                                                     Sequence<MThreadSliceSize>,
+                                                     Sequence<0>,
+                                                     0,
+                                                     OutDstVectorSizeSeq::At(iR),
+                                                     1,
+                                                     false>(
+                        out_grid_desc_m_tuple[iR],
+                        make_multi_index(thread_global_1d_id * MThreadSliceSize));
+
+                threadwise_dst_load.Run(out_grid_desc_m_tuple[iR],
+                                        out_global_val_buf_tuple(iR),
+                                        reduced_data_desc,
+                                        make_tuple(I0),
+                                        priorDstValueBuf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                    accu_value_buf_tuple(iR)(I) +=
+                        type_convert<AccDataType>(priorDstValueBuf[I]) * beta_values[iR];
+                });
+            };
+
+            auto threadwise_dst_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   OutDataType,
+                                                   decltype(reduced_data_desc),
+                                                   decltype(out_grid_desc_m_tuple[iR]),
+                                                   PassThroughOp,
+                                                   Sequence<MThreadSliceSize>,
+                                                   Sequence<0>,
+                                                   0,
+                                                   OutDstVectorSizeSeq::At(iR),
+                                                   OutMemoryDataOperation,
+                                                   1,
+                                                   true>(
+                    out_grid_desc_m_tuple[iR],
+                    make_multi_index(thread_global_1d_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_dst_store.Run(reduced_data_desc,
+                                     make_tuple(I0),
+                                     accu_value_buf_tuple[iR],
+                                     out_grid_desc_m_tuple[iR],
+                                     out_global_val_buf_tuple(iR));
+        });
+    };
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp
new file mode 100644
index 0000000000..203be3c42d
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp
@@ -0,0 +1,613 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseReduction,
+          bool OutputIndex,
+          bool HaveIndexInput,
+          typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename IndexDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation>
+__global__ void kernel_reduce_multiblock(const InGridDesc_M_K in_grid_desc_m_k,
+                                         const OutGridDesc_M out_grid_desc_m,
+                                         const InElementwiseOperation in_elementwise_op,
+                                         const AccElementwiseOperation acc_elementwise_op,
+                                         index_t block_group_size,
+                                         index_t num_k_block_tile_iteration,
+                                         AccDataType alpha,
+                                         const InDataType* const __restrict__ p_in_value_global,
+                                         const IndexDataType* const __restrict__ p_in_index_global,
+                                         AccDataType beta,
+                                         OutDataType* const __restrict__ p_out_value_global,
+                                         IndexDataType* const __restrict__ p_out_index_global)
+{
+    if constexpr(!OutputIndex)
+    {
+        (void)p_in_index_global;
+        (void)p_out_index_global;
+
+        GridwiseReduction::Run(in_grid_desc_m_k,
+                               out_grid_desc_m,
+                               in_elementwise_op,
+                               acc_elementwise_op,
+                               block_group_size,
+                               num_k_block_tile_iteration,
+                               alpha,
+                               p_in_value_global,
+                               beta,
+                               p_out_value_global);
+    }
+    else
+    {
+        GridwiseReduction::template RunWithIndex<HaveIndexInput>(in_grid_desc_m_k,
+                                                                 out_grid_desc_m,
+                                                                 in_elementwise_op,
+                                                                 acc_elementwise_op,
+                                                                 num_k_block_tile_iteration,
+                                                                 alpha,
+                                                                 p_in_value_global,
+                                                                 p_in_index_global,
+                                                                 beta,
+                                                                 p_out_value_global,
+                                                                 p_out_index_global);
+    };
+};
+
+template <typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename IndexDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M,
+          typename ReduceOperation,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation,
+          InMemoryDataOperationEnum OutMemoryDataOperation,
+          bool PropagateNan,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          index_t OutDstVectorSize>
+struct GridwiseReduction_mk_to_m_multiblock
+{
+    static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                   (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)) &&
+                      (MThreadSliceSize % OutDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (InSrcVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using BlockwiseReduce = PartitionedBlockwiseReduction<AccDataType,
+                                                          BlockSize,
+                                                          ThreadClusterLengths_M_K,
+                                                          ThreadClusterArrangeOrder,
+                                                          ReduceOperation,
+                                                          PropagateNan>;
+
+    using ThreadwiseReduce = ThreadwiseReduction<AccDataType,
+                                                 ThreadReduceSrcDesc_M_K,
+                                                 ThreadReduceDstDesc_M,
+                                                 ReduceOperation,
+                                                 PropagateNan>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    using Accumulation = detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
+
+    __device__ static void Run(const InGridDesc_M_K& in_grid_desc_m_k,
+                               const OutGridDesc_M& out_grid_desc_m,
+                               const InElementwiseOperation& in_elementwise_op,
+                               const AccElementwiseOperation& acc_elementwise_op,
+                               index_t block_group_size,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType alpha,
+                               const InDataType* const __restrict__ p_in_value_global,
+                               AccDataType beta,
+                               OutDataType* const __restrict__ p_out_value_global)
+    {
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+        // LDS
+        __shared__ AccDataType p_reduce_work_buffer[BlockSize];
+
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
+        auto out_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_value_global, out_grid_desc_m.GetElementSpaceSize());
+
+        auto reduce_work_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            in_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> accu_value_buf;
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) { accu_value_buf(I) = identityVal; });
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / block_group_size;
+        const index_t block_local_id  = block_global_id % block_group_size;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        const index_t reduceSizePerBlock = K_BlockTileSize * num_k_block_tile_iteration;
+
+        using ThreadBufferLengths         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        constexpr auto thread_buffer_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+
+        auto threadwise_src_load = ThreadwiseTensorSliceTransfer_v2<InDataType,
+                                                                    AccDataType,
+                                                                    InGridDesc_M_K,
+                                                                    decltype(thread_buffer_desc),
+                                                                    ThreadBufferLengths,
+                                                                    ThreadBufferDimAccessOrder,
+                                                                    InSrcVectorDim,
+                                                                    InSrcVectorSize,
+                                                                    1,
+                                                                    false>(
+            in_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             block_local_id * reduceSizePerBlock +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        constexpr auto in_thread_copy_step = make_multi_index(0, K_BlockTileSize);
+
+        index_t reducedTiles = 0;
+        do
+        {
+            threadwise_src_load.Run(in_grid_desc_m_k,
+                                    in_global_val_buf,
+                                    thread_buffer_desc,
+                                    make_tuple(I0, I0),
+                                    in_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                // do element-wise pre-reduction operation
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset = thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+                    in_elementwise_op(in_thread_buf(Number<offset>{}),
+                                      in_thread_buf(Number<offset>{}));
+                });
+            });
+
+            ThreadwiseReduce::Reduce(in_thread_buf, accu_value_buf);
+
+            threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+            reducedTiles++;
+        } while(reducedTiles < num_k_block_tile_iteration);
+
+        constexpr auto reduced_data_desc = ThreadReduceDstDesc_M{};
+
+        static_for<0, MThreadSliceSize, 1>{}(
+            [&](auto I) { BlockwiseReduce::Reduce(reduce_work_buf, accu_value_buf(I)); });
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if(thread_k_cluster_id == 0)
+            {
+                acc_elementwise_op(accu_value_buf(I), accu_value_buf(I));
+
+                accu_value_buf(I) *= alpha;
+            }
+        });
+
+        if(thread_k_cluster_id == 0)
+        {
+            if(block_group_size == 0 && !float_equal_zero{}(beta))
+            {
+                StaticBuffer<AddressSpaceEnum::Vgpr, OutDataType, MThreadSliceSize, true>
+                    priorDstValueBuf;
+
+                auto threadwise_dst_load =
+                    ThreadwiseTensorSliceTransfer_v2<OutDataType,
+                                                     OutDataType,
+                                                     OutGridDesc_M,
+                                                     decltype(reduced_data_desc),
+                                                     Sequence<MThreadSliceSize>,
+                                                     Sequence<0>,
+                                                     0,
+                                                     OutDstVectorSize,
+                                                     1,
+                                                     false>(
+                        out_grid_desc_m,
+                        make_multi_index(blkgroup_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize));
+
+                threadwise_dst_load.Run(out_grid_desc_m,
+                                        out_global_val_buf,
+                                        reduced_data_desc,
+                                        make_tuple(I0),
+                                        priorDstValueBuf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                    accu_value_buf(I) += type_convert<AccDataType>(priorDstValueBuf[I]) * beta;
+                });
+            };
+
+            auto threadwise_dst_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   OutDataType,
+                                                   decltype(reduced_data_desc),
+                                                   OutGridDesc_M,
+                                                   PassThroughOp,
+                                                   Sequence<MThreadSliceSize>,
+                                                   Sequence<0>,
+                                                   0,
+                                                   OutDstVectorSize,
+                                                   OutMemoryDataOperation,
+                                                   1,
+                                                   true>(
+                    out_grid_desc_m,
+                    make_multi_index(blkgroup_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_dst_store.Run(reduced_data_desc,
+                                     make_tuple(I0),
+                                     accu_value_buf,
+                                     out_grid_desc_m,
+                                     out_global_val_buf);
+        }
+    };
+
+    template <bool HaveIndexInput>
+    __device__ static void RunWithIndex(const InGridDesc_M_K& in_grid_desc_m_k,
+                                        const OutGridDesc_M& out_grid_desc_m,
+                                        const InElementwiseOperation in_elementwise_op,
+                                        const AccElementwiseOperation acc_elementwise_op,
+                                        index_t num_k_block_tile_iteration,
+                                        AccDataType alpha,
+                                        const InDataType* const __restrict__ p_in_value_global,
+                                        const IndexDataType* const __restrict__ p_in_index_global,
+                                        AccDataType beta,
+                                        OutDataType* const __restrict__ p_out_value_global,
+                                        IndexDataType* const __restrict__ p_out_index_global)
+    {
+        using BlockwiseReduceWithIndex =
+            PartitionedBlockwiseReductionWithIndex<AccDataType,
+                                                   IndexDataType,
+                                                   BlockSize,
+                                                   Sequence<MThreadClusterSize, KThreadClusterSize>,
+                                                   ThreadClusterArrangeOrder,
+                                                   ReduceOperation,
+                                                   PropagateNan>;
+
+        using AccumulationWithIndex = detail::AccumulateWithIndexAndNanCheck<PropagateNan,
+                                                                             ReduceOperation,
+                                                                             AccDataType,
+                                                                             IndexDataType>;
+
+        (void)in_elementwise_op;
+
+        // LDS
+        __shared__ AccDataType p_reduce_work_val_buffer[BlockSize];
+        __shared__ IndexDataType p_reduce_work_idx_buffer[BlockSize];
+
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
+        const auto in_global_idx_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_index_global, in_grid_desc_m_k.GetElementSpaceSize());
+        auto out_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_value_global, out_grid_desc_m.GetElementSpaceSize());
+        auto out_global_idx_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_index_global, out_grid_desc_m.GetElementSpaceSize());
+
+        auto reduce_work_val_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_val_buffer, BlockSize);
+        auto reduce_work_idx_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_idx_buffer, BlockSize);
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            in_thread_val_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     IndexDataType,
+                     MThreadSliceSize * KThreadSliceSize,
+                     true>
+            in_thread_idx_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> accu_value_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, IndexDataType, MThreadSliceSize, true> accu_index_buf;
+
+        const index_t thread_local_id    = get_thread_local_1d_id();
+        const index_t block_global_1d_id = get_block_1d_id();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        constexpr auto thread_buffer_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+
+        auto threadwise_src_val_load =
+            ThreadwiseTensorSliceTransfer_v2<InDataType,
+                                             AccDataType,
+                                             InGridDesc_M_K,
+                                             decltype(thread_buffer_desc),
+                                             ThreadBufferLengths,
+                                             ThreadBufferDimAccessOrder,
+                                             InSrcVectorDim,
+                                             InSrcVectorSize,
+                                             1,
+                                             false>(
+                in_grid_desc_m_k,
+                make_multi_index(block_global_1d_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            accu_value_buf(I) = identityVal;
+            accu_index_buf(I) = 0;
+        });
+
+        constexpr auto in_thread_copy_step = make_multi_index(0, K_BlockTileSize);
+
+        index_t reducedTiles = 0;
+
+        if constexpr(HaveIndexInput)
+        {
+            auto threadwise_src_idx_load =
+                ThreadwiseTensorSliceTransfer_v2<IndexDataType,
+                                                 IndexDataType,
+                                                 InGridDesc_M_K,
+                                                 decltype(thread_buffer_desc),
+                                                 ThreadBufferLengths,
+                                                 ThreadBufferDimAccessOrder,
+                                                 InSrcVectorDim,
+                                                 InSrcVectorSize,
+                                                 1,
+                                                 false>(
+                    in_grid_desc_m_k,
+                    make_multi_index(block_global_1d_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize,
+                                     thread_k_cluster_id * KThreadSliceSize));
+
+            do
+            {
+                // load the thread slice
+                threadwise_src_val_load.Run(in_grid_desc_m_k,
+                                            in_global_val_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_val_buf);
+                threadwise_src_idx_load.Run(in_grid_desc_m_k,
+                                            in_global_idx_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_idx_buf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    AccDataType tmpValue   = identityVal;
+                    IndexDataType tmpIndex = 0;
+
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+
+                        AccumulationWithIndex::Calculate(tmpValue,
+                                                         in_thread_val_buf[Number<offset>{}],
+                                                         tmpIndex,
+                                                         in_thread_idx_buf[Number<offset>{}]);
+                    });
+
+                    BlockwiseReduceWithIndex::Reduce(
+                        reduce_work_val_buf, reduce_work_idx_buf, tmpValue, tmpIndex);
+
+                    AccumulationWithIndex::Calculate(
+                        accu_value_buf(iM), tmpValue, accu_index_buf(iM), tmpIndex);
+                });
+
+                threadwise_src_val_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+                threadwise_src_idx_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+                reducedTiles++;
+            } while(reducedTiles < num_k_block_tile_iteration);
+        }
+        else
+        {
+            index_t indexOffset = 0;
+
+            do
+            {
+                // load the thread slice
+                threadwise_src_val_load.Run(in_grid_desc_m_k,
+                                            in_global_val_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_val_buf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+
+                        // initialize the indices for the per-thread to-reduce values
+                        in_thread_idx_buf(Number<offset>{}) =
+                            indexOffset + thread_k_cluster_id * KThreadSliceSize + iK();
+
+                        // do element-wise pre-reduction operation
+                        in_elementwise_op(in_thread_val_buf(Number<offset>{}),
+                                          in_thread_val_buf(Number<offset>{}));
+                    });
+
+                    AccDataType tmpValue   = identityVal;
+                    IndexDataType tmpIndex = 0;
+
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+
+                        AccumulationWithIndex::Calculate(tmpValue,
+                                                         in_thread_val_buf[Number<offset>{}],
+                                                         tmpIndex,
+                                                         in_thread_idx_buf[Number<offset>{}]);
+                    });
+
+                    BlockwiseReduceWithIndex::Reduce(
+                        reduce_work_val_buf, reduce_work_idx_buf, tmpValue, tmpIndex);
+
+                    AccumulationWithIndex::Calculate(
+                        accu_value_buf(iM), tmpValue, accu_index_buf(iM), tmpIndex);
+                });
+
+                threadwise_src_val_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+                indexOffset += K_BlockTileSize;
+                reducedTiles++;
+            } while(reducedTiles < num_k_block_tile_iteration);
+        };
+
+        constexpr auto reduced_data_desc = ThreadReduceDstDesc_M{};
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if(thread_k_cluster_id == 0)
+            {
+                // for indiced operation, acc_elementwise_op shoud do nothing
+                acc_elementwise_op(accu_value_buf(I), accu_value_buf(I));
+
+                accu_value_buf(I) *= alpha;
+            }
+        });
+
+        if(thread_k_cluster_id == 0)
+        {
+            if(!float_equal_zero{}(beta))
+            {
+                StaticBuffer<AddressSpaceEnum::Vgpr, OutDataType, MThreadSliceSize, true>
+                    priorDstValueBuf;
+
+                auto threadwise_dst_load =
+                    ThreadwiseTensorSliceTransfer_v2<OutDataType,
+                                                     OutDataType,
+                                                     OutGridDesc_M,
+                                                     decltype(reduced_data_desc),
+                                                     Sequence<MThreadSliceSize>,
+                                                     Sequence<0>,
+                                                     0,
+                                                     OutDstVectorSize,
+                                                     1,
+                                                     true>(
+                        out_grid_desc_m,
+                        make_multi_index(block_global_1d_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize));
+
+                threadwise_dst_load.Run(out_grid_desc_m,
+                                        out_global_val_buf,
+                                        reduced_data_desc,
+                                        make_tuple(I0),
+                                        priorDstValueBuf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                    accu_value_buf(I) += type_convert<AccDataType>(priorDstValueBuf[I]) * beta;
+                });
+            };
+
+            auto threadwise_dst_val_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   OutDataType,
+                                                   decltype(reduced_data_desc),
+                                                   OutGridDesc_M,
+                                                   PassThroughOp,
+                                                   Sequence<MThreadSliceSize>,
+                                                   Sequence<0>,
+                                                   0,
+                                                   OutDstVectorSize,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>(
+                    out_grid_desc_m,
+                    make_multi_index(block_global_1d_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            auto threadwise_dst_idx_store =
+                ThreadwiseTensorSliceTransfer_v1r3<IndexDataType,
+                                                   IndexDataType,
+                                                   decltype(reduced_data_desc),
+                                                   OutGridDesc_M,
+                                                   PassThroughOp,
+                                                   Sequence<MThreadSliceSize>,
+                                                   Sequence<0>,
+                                                   0,
+                                                   OutDstVectorSize,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>(
+                    out_grid_desc_m,
+                    make_multi_index(block_global_1d_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_dst_val_store.Run(reduced_data_desc,
+                                         make_tuple(I0),
+                                         accu_value_buf,
+                                         out_grid_desc_m,
+                                         out_global_val_buf);
+            threadwise_dst_idx_store.Run(reduced_data_desc,
+                                         make_tuple(I0),
+                                         accu_index_buf,
+                                         out_grid_desc_m,
+                                         out_global_idx_buf);
+        }
+    };
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp
new file mode 100644
index 0000000000..910c926c7e
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp
@@ -0,0 +1,488 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseReduction,
+          bool OutputIndex,
+          bool TransformIndexKtoGlobal,
+          bool HaveIndexInput,
+          typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename IndexDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation>
+__global__ void kernel_reduce_threadwise(const InGridDesc_M_K in_grid_desc_m_k,
+                                         const OutGridDesc_M out_grid_desc_m,
+                                         const InElementwiseOperation in_elementwise_op,
+                                         const AccElementwiseOperation acc_elementwise_op,
+                                         AccDataType alpha,
+                                         const InDataType* const __restrict__ p_in_value_global,
+                                         const IndexDataType* const __restrict__ p_in_index_global,
+                                         AccDataType beta,
+                                         OutDataType* const __restrict__ p_out_value_global,
+                                         IndexDataType* const __restrict__ p_out_index_global)
+{
+    if constexpr(!OutputIndex)
+    {
+        GridwiseReduction::Run(in_grid_desc_m_k,
+                               out_grid_desc_m,
+                               in_elementwise_op,
+                               acc_elementwise_op,
+                               alpha,
+                               p_in_value_global,
+                               beta,
+                               p_out_value_global);
+    }
+    else
+    {
+        GridwiseReduction::template RunWithIndex<TransformIndexKtoGlobal, HaveIndexInput>(
+            in_grid_desc_m_k,
+            out_grid_desc_m,
+            in_elementwise_op,
+            acc_elementwise_op,
+            alpha,
+            p_in_value_global,
+            p_in_index_global,
+            beta,
+            p_out_value_global,
+            p_out_index_global);
+    };
+};
+
+template <typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename IndexDataType,
+          typename InGridDesc_M_K,
+          typename OutGridDesc_M,
+          typename ReduceOperation,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation,
+          InMemoryDataOperationEnum OutMemoryDataOperation,
+          bool PropagateNan,
+          index_t BlockSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          index_t OutDstVectorSize>
+struct GridwiseReduction_mk_to_m_threadwise
+{
+    static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                   (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)) &&
+                      (MThreadSliceSize % OutDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<InSrcVectorDim == 0, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ static void Run(const InGridDesc_M_K& in_grid_desc_m_k,
+                               const OutGridDesc_M& out_grid_desc_m,
+                               const InElementwiseOperation& in_elementwise_op,
+                               const AccElementwiseOperation& acc_elementwise_op,
+                               AccDataType alpha,
+                               const InDataType* const __restrict__ p_in_value_global,
+                               AccDataType beta,
+                               OutDataType* const __restrict__ p_out_value_global)
+    {
+        using ThreadwiseReduce = ThreadwiseReduction<AccDataType,
+                                                     ThreadReduceSrcDesc_M_K,
+                                                     ThreadReduceDstDesc_M,
+                                                     ReduceOperation,
+                                                     PropagateNan>;
+
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
+        auto dst_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_value_global, out_grid_desc_m.GetElementSpaceSize());
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            in_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> accu_value_buf;
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) { accu_value_buf(I) = identityVal; });
+
+        const auto toReduceLength = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        using ThreadBufferLengths         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        constexpr auto thread_buffer_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+
+        index_t thread_global_1d_id = get_block_1d_id() * BlockSize + get_thread_local_1d_id();
+
+        auto threadwise_src_val_load =
+            ThreadwiseTensorSliceTransfer_v2<InDataType,
+                                             AccDataType,
+                                             InGridDesc_M_K,
+                                             decltype(thread_buffer_desc),
+                                             ThreadBufferLengths,
+                                             ThreadBufferDimAccessOrder,
+                                             InSrcVectorDim,
+                                             InSrcVectorSize,
+                                             1,
+                                             false>(
+                in_grid_desc_m_k, make_multi_index(thread_global_1d_id * MThreadSliceSize, 0));
+
+        constexpr auto in_thread_copy_step = make_multi_index(0, KThreadSliceSize);
+
+        index_t reducedLength = 0;
+        do
+        {
+            threadwise_src_val_load.Run(in_grid_desc_m_k,
+                                        in_global_val_buf,
+                                        thread_buffer_desc,
+                                        make_tuple(I0, I0),
+                                        in_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                // do element-wise pre-reduction operation
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset = thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+                    in_elementwise_op(in_thread_buf(Number<offset>{}),
+                                      in_thread_buf(Number<offset>{}));
+                });
+            });
+
+            ThreadwiseReduce::Reduce(in_thread_buf, accu_value_buf);
+
+            threadwise_src_val_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+            reducedLength += KThreadSliceSize;
+        } while(reducedLength < toReduceLength);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            acc_elementwise_op(accu_value_buf(I), accu_value_buf(I));
+
+            accu_value_buf(I) *= alpha;
+        });
+
+        constexpr auto reduced_data_desc = ThreadReduceDstDesc_M{};
+
+        if(!float_equal_zero{}(beta))
+        {
+            auto threadwise_dst_load = ThreadwiseTensorSliceTransfer_v2<OutDataType,
+                                                                        OutDataType,
+                                                                        OutGridDesc_M,
+                                                                        decltype(reduced_data_desc),
+                                                                        Sequence<MThreadSliceSize>,
+                                                                        Sequence<0>,
+                                                                        0,
+                                                                        1,
+                                                                        1,
+                                                                        true>(
+                out_grid_desc_m, make_multi_index(thread_global_1d_id * MThreadSliceSize));
+
+            StaticBuffer<AddressSpaceEnum::Vgpr, OutDataType, MThreadSliceSize, true>
+                priorDstValue_buf;
+
+            threadwise_dst_load.Run(out_grid_desc_m,
+                                    dst_global_buf,
+                                    reduced_data_desc,
+                                    make_tuple(I0),
+                                    priorDstValue_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                accu_value_buf(I) += type_convert<AccDataType>(priorDstValue_buf[I]) * beta;
+            });
+        };
+
+        auto threadwise_dst_store = ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                                       OutDataType,
+                                                                       decltype(reduced_data_desc),
+                                                                       OutGridDesc_M,
+                                                                       PassThroughOp,
+                                                                       Sequence<MThreadSliceSize>,
+                                                                       Sequence<0>,
+                                                                       0,
+                                                                       OutDstVectorSize,
+                                                                       OutMemoryDataOperation,
+                                                                       1,
+                                                                       false>(
+            out_grid_desc_m,
+            make_multi_index(thread_global_1d_id * MThreadSliceSize),
+            PassThroughOp{});
+
+        threadwise_dst_store.Run(
+            reduced_data_desc, make_tuple(I0), accu_value_buf, out_grid_desc_m, dst_global_buf);
+    };
+
+    template <bool TransformIndexKtoGlobal, bool HaveIndexInput>
+    __device__ static void RunWithIndex(const InGridDesc_M_K& in_grid_desc_m_k,
+                                        const OutGridDesc_M& out_grid_desc_m,
+                                        const InElementwiseOperation& in_elementwise_op,
+                                        const AccElementwiseOperation& acc_elementwise_op,
+                                        AccDataType alpha,
+                                        const InDataType* const __restrict__ p_in_value_global,
+                                        const IndexDataType* const __restrict__ p_in_index_global,
+                                        AccDataType beta,
+                                        OutDataType* const __restrict__ p_out_value_global,
+                                        IndexDataType* const __restrict__ p_out_index_global)
+    {
+        using ThreadwiseReduceWithIndex = ThreadwiseReductionWithIndex<AccDataType,
+                                                                       IndexDataType,
+                                                                       ThreadReduceSrcDesc_M_K,
+                                                                       ThreadReduceDstDesc_M,
+                                                                       ReduceOperation,
+                                                                       PropagateNan>;
+
+        (void)acc_elementwise_op;
+
+        const auto identityVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global,
+            in_grid_desc_m_k.GetElementSpaceSize(),
+            ReduceOperation::template GetIdentityValue<InDataType>());
+        const auto in_global_idx_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_index_global, in_grid_desc_m_k.GetElementSpaceSize());
+
+        auto out_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_value_global, out_grid_desc_m.GetElementSpaceSize());
+        auto out_global_idx_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_index_global, out_grid_desc_m.GetElementSpaceSize());
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            in_thread_val_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     IndexDataType,
+                     MThreadSliceSize * KThreadSliceSize,
+                     true>
+            in_thread_idx_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> accu_value_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, IndexDataType, MThreadSliceSize, true> accu_index_buf;
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            accu_value_buf(I) = identityVal;
+            accu_index_buf(I) = 0;
+        });
+
+        const auto toReduceLength = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        using ThreadBufferLengths         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        constexpr auto thread_buffer_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+
+        index_t thread_global_1d_id = get_block_1d_id() * BlockSize + get_thread_local_1d_id();
+
+        auto threadwise_src_val_load =
+            ThreadwiseTensorSliceTransfer_v2<InDataType,
+                                             AccDataType,
+                                             InGridDesc_M_K,
+                                             decltype(thread_buffer_desc),
+                                             ThreadBufferLengths,
+                                             ThreadBufferDimAccessOrder,
+                                             InSrcVectorDim,
+                                             InSrcVectorSize,
+                                             1,
+                                             false>(
+                in_grid_desc_m_k, make_multi_index(thread_global_1d_id * MThreadSliceSize, 0));
+
+        constexpr auto in_thread_copy_step = make_multi_index(0, KThreadSliceSize);
+
+        index_t indexStart    = 0;
+        index_t reducedLength = 0;
+        if constexpr(HaveIndexInput)
+        {
+            auto threadwise_src_idx_load =
+                ThreadwiseTensorSliceTransfer_v2<IndexDataType,
+                                                 IndexDataType,
+                                                 InGridDesc_M_K,
+                                                 decltype(thread_buffer_desc),
+                                                 ThreadBufferLengths,
+                                                 ThreadBufferDimAccessOrder,
+                                                 InSrcVectorDim,
+                                                 InSrcVectorSize,
+                                                 1,
+                                                 false>(
+                    in_grid_desc_m_k, make_multi_index(thread_global_1d_id * MThreadSliceSize, 0));
+
+            do
+            {
+                threadwise_src_val_load.Run(in_grid_desc_m_k,
+                                            in_global_val_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_val_buf);
+
+                threadwise_src_idx_load.Run(in_grid_desc_m_k,
+                                            in_global_idx_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_idx_buf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    // do element-wise pre-reduction operation
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+
+                        in_elementwise_op(in_thread_val_buf(Number<offset>{}),
+                                          in_thread_val_buf(Number<offset>{}));
+                    });
+                });
+
+                ThreadwiseReduceWithIndex::Reduce(
+                    in_thread_val_buf, in_thread_idx_buf, accu_value_buf, accu_index_buf);
+
+                threadwise_src_val_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+                threadwise_src_idx_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+                indexStart += KThreadSliceSize;
+                reducedLength += KThreadSliceSize;
+            } while(reducedLength < toReduceLength);
+        }
+        else
+        {
+            do
+            {
+                threadwise_src_val_load.Run(in_grid_desc_m_k,
+                                            in_global_val_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_val_buf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    // do element-wise pre-reduction operation
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+
+                        in_thread_idx_buf(Number<offset>{}) = indexStart + iK();
+
+                        in_elementwise_op(in_thread_val_buf(Number<offset>{}),
+                                          in_thread_val_buf(Number<offset>{}));
+                    });
+                });
+
+                ThreadwiseReduceWithIndex::Reduce(
+                    in_thread_val_buf, in_thread_idx_buf, accu_value_buf, accu_index_buf);
+
+                threadwise_src_val_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_step);
+
+                indexStart += KThreadSliceSize;
+                reducedLength += KThreadSliceSize;
+            } while(reducedLength < toReduceLength);
+
+            if constexpr(TransformIndexKtoGlobal)
+            {
+                static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                    const auto coord = make_tensor_coordinate(
+                        in_grid_desc_m_k,
+                        make_multi_index(thread_global_1d_id * MThreadSliceSize + I,
+                                         accu_index_buf(I)));
+
+                    accu_index_buf(I) = coord.GetOffset();
+                });
+            }
+        };
+
+        // for indiced operation, acc_elementwise_op shoud do nothing
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            acc_elementwise_op(accu_value_buf(I), accu_value_buf(I));
+
+            accu_value_buf(I) *= alpha;
+        });
+
+        constexpr auto reduced_data_desc = ThreadReduceDstDesc_M{};
+
+        if(!float_equal_zero{}(beta))
+        {
+            auto threadwise_dst_load = ThreadwiseTensorSliceTransfer_v2<OutDataType,
+                                                                        OutDataType,
+                                                                        OutGridDesc_M,
+                                                                        decltype(reduced_data_desc),
+                                                                        Sequence<MThreadSliceSize>,
+                                                                        Sequence<0>,
+                                                                        0,
+                                                                        1,
+                                                                        1,
+                                                                        false>(
+                out_grid_desc_m, make_multi_index(thread_global_1d_id * MThreadSliceSize));
+
+            StaticBuffer<AddressSpaceEnum::Vgpr, OutDataType, MThreadSliceSize, true>
+                priorDstValue_buf;
+
+            threadwise_dst_load.Run(out_grid_desc_m,
+                                    out_global_val_buf,
+                                    reduced_data_desc,
+                                    make_tuple(I0),
+                                    priorDstValue_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                accu_value_buf(I) += type_convert<AccDataType>(priorDstValue_buf[I]) * beta;
+            });
+        };
+
+        auto threadwise_dst_val_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               OutDataType,
+                                               decltype(reduced_data_desc),
+                                               OutGridDesc_M,
+                                               PassThroughOp,
+                                               Sequence<MThreadSliceSize>,
+                                               Sequence<0>,
+                                               0,
+                                               OutDstVectorSize,
+                                               OutMemoryDataOperation,
+                                               1,
+                                               false>(
+                out_grid_desc_m,
+                make_multi_index(thread_global_1d_id * MThreadSliceSize),
+                PassThroughOp{});
+
+        auto threadwise_dst_idx_store =
+            ThreadwiseTensorSliceTransfer_v1r3<IndexDataType,
+                                               IndexDataType,
+                                               decltype(reduced_data_desc),
+                                               OutGridDesc_M,
+                                               PassThroughOp,
+                                               Sequence<MThreadSliceSize>,
+                                               Sequence<0>,
+                                               0,
+                                               OutDstVectorSize,
+                                               OutMemoryDataOperation,
+                                               1,
+                                               false>(
+                out_grid_desc_m,
+                make_multi_index(thread_global_1d_id * MThreadSliceSize),
+                PassThroughOp{});
+
+        threadwise_dst_val_store.Run(
+            reduced_data_desc, make_tuple(I0), accu_value_buf, out_grid_desc_m, out_global_val_buf);
+
+        threadwise_dst_idx_store.Run(
+            reduced_data_desc, make_tuple(I0), accu_index_buf, out_grid_desc_m, out_global_idx_buf);
+    };
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp
new file mode 100644
index 0000000000..52dd8ea916
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp
@@ -0,0 +1,932 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename B1GridDesc_BK0_N_BK1,
+          typename CGridDesc_M_N,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t B1K1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1ThreadTransferSrcResetCoordinateAfterRun,
+          index_t B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseBatchedGemmGemm_Xdl_CShuffle
+{
+    static_assert(LoopSched == LoopScheduler::Default,
+                  "Non-default loop scheduler is currently not supported");
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    // Gemm0
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+    // Gemm1
+    static constexpr auto B1K0 = Number<Gemm1KPerBlock / B1K1Value>{};
+    static constexpr auto B1K1 = Number<B1K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, MWaves, MPerXdl>(
+            ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<NXdlPerWave, NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, 1, 1>(ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t Gemm1NWaves = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm1NXdlPerWave, Gemm1NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B1 matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(B1K0, Number<Gemm1NPerBlock>{}, B1K1),
+            make_tuple(Number<Gemm1NPerBlock + B1BlockLdsExtraN>{} * B1K1, B1K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        const index_t gemm0_bytes_end = (SharedMemTrait::a_block_space_size_aligned +
+                                         SharedMemTrait::b_block_space_size_aligned) *
+                                        sizeof(FloatAB);
+        const index_t gemm1_bytes_end =
+            (SharedMemTrait::b1_block_space_offset + SharedMemTrait::b1_block_space_size_aligned) *
+            sizeof(FloatAB);
+        const index_t c_block_bytes_end =
+            SharedMemTrait::c_block_space_size * sizeof(FloatCShuffle);
+
+        return math::max(gemm0_bytes_end, gemm1_bytes_end, c_block_bytes_end);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+        const auto Gemm1N = b1_grid_desc_bk0_n_bk1.GetLength(I1);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && Gemm1N == c_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0 &&
+             Gemm1N % Gemm1NPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gemm0 gridwise gemm pipeline
+        const auto num_gemm0_k_loop = K / KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm0_k_loop))
+        {
+            return false;
+        }
+
+        // check gemm1 gridwise gemm pipeline
+        if(!(NPerBlock % Gemm1KPerBlock == 0))
+        {
+            return false;
+        }
+
+        const auto num_gemm1_k_inner_loop = NPerBlock / Gemm1KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm1_k_inner_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / Gemm1NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<Gemm1NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, Gemm1NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            CGridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
+
+    struct SharedMemTrait
+    {
+        // LDS allocation for A and B: be careful of alignment
+        static constexpr auto a_block_desc_ak0_m_ak1 =
+            GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        static constexpr auto b_block_desc_bk0_n_bk1 =
+            GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        static constexpr auto b1_block_desc_bk0_n_bk1 =
+            GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        static constexpr auto max_lds_align = math::lcm(math::lcm(AK1, BK1), B1K1);
+
+        static constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b1_block_space_size_aligned = math::integer_least_multiple(
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        static constexpr auto a_block_space_offset  = 0;
+        static constexpr auto b_block_space_offset  = a_block_space_size_aligned.value;
+        static constexpr auto b1_block_space_offset = 0;
+
+        // LDS allocation for C shuffle in LDS
+        static constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+        static constexpr auto c_block_space_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+    };
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap>
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               const FloatAB* __restrict__ p_b1_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const AccElementwiseOperation& acc_element_op,
+                               const B1ElementwiseOperation& b1_element_op,
+                               const CElementwiseOperation& c_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                               const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        const auto b1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b1_grid, b1_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * Gemm1NPerBlock);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        //
+        // set up Gemm0
+        //
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0), // will loop over GemmN dimension
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // Fused Gemm+Gemm pipeline
+        // for n in N0:
+        //   for k in K0:
+        //     acc[m][n] += A[m][k] * B0[k][n]
+        //   acc1[m][o] += acc[m][n] * B1[n][o]
+
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            decltype(MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(a_block_desc_ak0_m_ak1)),
+            decltype(MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(b_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            NPerBlock,
+            KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            true>{}; // TransposeC
+
+        auto acc_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::a_block_space_offset,
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::b_block_space_offset,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+        const auto a_block_reset_copy_step =
+            make_multi_index(-a_grid_desc_ak0_m_ak1.GetLength(I0), 0, 0);
+        const auto b_block_reset_copy_step =
+            make_multi_index(-b_grid_desc_bk0_n_bk1.GetLength(I0), NPerBlock, 0);
+
+        // gridwise GEMM pipeline
+        // Only supports LoopScheduler::Default
+        const auto gridwise_gemm_pipeline = GridwiseGemmPipeline_Selector<PipelineVer,
+                                                                          NumGemmKPrefetchStage,
+                                                                          LoopScheduler::Default>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        //
+        // set up Gemm1
+        //
+
+        // Acc matrix threadwise copy: AccVGPR to VGPR and downcast to XDL input data type
+        constexpr auto acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+        constexpr auto m0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+        constexpr auto n0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+        constexpr auto m1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+        constexpr auto n1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+        constexpr auto m2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+        constexpr auto n2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+        constexpr auto n3 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+        constexpr auto n4 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+
+        constexpr auto b1_block_slice_copy_step = make_multi_index(Gemm1KPerBlock / B1K1, 0, 0);
+
+        // acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 to acc_thread_desc_k0_m_k1
+        // n0_n1_n2_n3 -> k0
+        // m0_m1_m2 -> m
+        // n4 -> k1
+        // NOTE: had to use merge_v3 or will spit out compilation errors
+        constexpr auto acc_thread_desc_k0_m_k1 = transform_tensor_descriptor(
+            acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(n0, n1, n2, n3)),
+                       make_merge_transform_v3_division_mod(make_tuple(m0, m1, m2)),
+                       make_pass_through_transform(n4)),
+            make_tuple(Sequence<1, 3, 5, 6>{}, Sequence<0, 2, 4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        // A1 matrix in AccVGPR
+        // N2 num_groups_per_blk, N3 num_input_blks, N4 group_size
+        constexpr auto AccN3 =
+            blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLength(I6);
+
+        constexpr auto A1ThreadSlice_K0_M_K1 =
+            make_tuple(Number<Gemm1KPerBlock / n4 / AccN3>{}, Number<m0 * m1 * m2>{}, Number<n4>{});
+
+        constexpr auto A1ThreadSliceK0        = A1ThreadSlice_K0_M_K1[I0];
+        constexpr auto A1ThreadSliceM         = A1ThreadSlice_K0_M_K1[I1];
+        constexpr auto A1ThreadSliceK1        = A1ThreadSlice_K0_M_K1[I2];
+        constexpr auto a1_thread_desc_k0_m_k1 = make_naive_tensor_descriptor(
+            A1ThreadSlice_K0_M_K1,
+            make_tuple(A1ThreadSliceM * A1ThreadSliceK1, A1ThreadSliceK1, I1));
+
+        // B1 matrix in LDS memory, dst of blockwise copy
+        constexpr auto b1_block_desc_bk0_n_bk1 = GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A1 matrix blockwise copy
+        auto a1_blockwise_copy = ThreadwiseTensorSliceTransfer_StaticToStatic<
+            FloatGemmAcc,
+            FloatAB,
+            decltype(acc_thread_desc_k0_m_k1),
+            decltype(a1_thread_desc_k0_m_k1),
+            decltype(acc_element_op),
+            Sequence<A1ThreadSliceK0, A1ThreadSliceM, A1ThreadSliceK1>,
+            Sequence<1, 0, 2>,
+            2,
+            n4>{acc_element_op};
+
+        // B1 matrix blockwise copy
+        auto b1_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<B1K0, Gemm1NPerBlock, B1K1>,
+                                                B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                B1BlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b1_grid_desc_bk0_n_bk1),
+                                                decltype(b1_block_desc_bk0_n_bk1),
+                                                B1BlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                B1BlockTransferSrcVectorDim,
+                                                2,
+                                                B1BlockTransferSrcScalarPerVector,
+                                                B1BlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                B1ThreadTransferSrcResetCoordinateAfterRun,
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b1_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b1_element_op,
+                b1_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        auto a1_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a1_thread_desc_k0_m_k1.GetElementSpaceSize());
+
+        // reuse LDS space for gemm0's b_block_buf
+        auto b1_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::b1_block_space_offset,
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        // selected_mfma.group_size or B1K1 <= Gemm1KPack <= selected_mfma.group_size
+        // selected_mfma.k_per_blk <= Gemm1KPack
+        //
+        // Following similar rationale behind Gemm0KPack, let Gemm1KPack be the lowest common
+        // multiples of A1K1 (predetermined by selected_mfma.group_size) and B1K1. But in this case
+        // Gemm1KPack can't be higher than A1K1 itself because A1 matrix is distributed in VGPRs
+        // with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
+        // cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
+        // therefore we may just as well assign Gemm1KPack = group_size
+        constexpr index_t Gemm1KPack =
+            MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size;
+
+        auto gemm1_blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a1_thread_desc_k0_m_k1),
+            decltype(b1_block_desc_bk0_n_bk1),
+            decltype(MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(a1_thread_desc_k0_m_k1)),
+            decltype(MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(b1_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            Gemm1NPerBlock,
+            Gemm1KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            Gemm1NXdlPerWave,
+            Gemm1KPack,
+            false,      // TransposeC
+            Gemm1KPack, // AMmaKStride
+            Gemm1KPack * XdlopsGemm<FloatAB, MPerXdl, NPerXdl, Gemm1KPack, false>{}.K0PerXdlops>{
+            // BMmaKStride
+            make_multi_index(0, 0, 0, 0)}; // A_origin
+
+        auto c_thread_buf = gemm1_blockwise_gemm.GetCThreadBuffer();
+
+        const index_t num_gemm1_k_block_outer_loop =
+            b_grid_desc_bk0_n_bk1.GetLength(I1) / NPerBlock;
+        constexpr index_t num_gemm1_k_block_inner_loop = NPerBlock / Gemm1KPerBlock;
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // gemm1 K loop
+        index_t gemm1_k_block_outer_index = 0;
+        do
+        {
+            // gemm0
+            gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                                   a_block_desc_ak0_m_ak1,
+                                                                   a_blockwise_copy,
+                                                                   a_grid_buf,
+                                                                   a_block_buf,
+                                                                   a_block_slice_copy_step,
+                                                                   b_grid_desc_bk0_n_bk1,
+                                                                   b_block_desc_bk0_n_bk1,
+                                                                   b_blockwise_copy,
+                                                                   b_grid_buf,
+                                                                   b_block_buf,
+                                                                   b_block_slice_copy_step,
+                                                                   blockwise_gemm,
+                                                                   acc_thread_buf,
+                                                                   num_k_block_main_loop);
+            // gemm1
+            {
+                // TODO: explore using dynamic buffer for a1 thread buffer
+                // For a1_blockwise_copy, the goal is to satisfy pipeline requirements RunRead(),
+                // RunWrite(), and MoveSliceWindow(). But it is impossible to implement given that
+                // the A1 source buffer is static buffer holding the output of first GEMM and
+                // requires constexpr offset by design. Therefore, we pass tensor coordinate offset
+                // explicitly in Run() below.
+
+                // preload data into LDS
+                b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                     b1_block_slice_copy_step);
+
+                block_sync_lds(); // wait for gemm0 LDS read
+
+                b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+
+                // main body
+                if constexpr(num_gemm1_k_block_inner_loop > 1)
+                {
+                    static_for<0, num_gemm1_k_block_inner_loop - 1, 1>{}([&](auto i) {
+                        a1_blockwise_copy.Run(acc_thread_desc_k0_m_k1,
+                                              make_tuple(Number<i * A1ThreadSliceK0>{}, I0, I0),
+                                              acc_thread_buf,
+                                              a1_thread_desc_k0_m_k1,
+                                              make_tuple(I0, I0, I0),
+                                              a1_thread_buf);
+
+                        b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                        block_sync_lds();
+
+                        gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, c_thread_buf);
+
+                        block_sync_lds();
+
+                        b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                             b1_block_slice_copy_step);
+
+                        b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+                    });
+                }
+                // tail
+                {
+                    a1_blockwise_copy.Run(
+                        acc_thread_desc_k0_m_k1,
+                        make_tuple(
+                            Number<(num_gemm1_k_block_inner_loop - 1) * A1ThreadSliceK0>{}, I0, I0),
+                        acc_thread_buf,
+                        a1_thread_desc_k0_m_k1,
+                        make_tuple(I0, I0, I0),
+                        a1_thread_buf);
+
+                    block_sync_lds();
+
+                    gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, c_thread_buf);
+                }
+            } // end gemm1
+
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_ak0_m_ak1,
+                                                a_block_reset_copy_step); // rewind K
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_bk0_n_bk1,
+                                                b_block_reset_copy_step); // rewind K and step N
+
+            block_sync_lds(); // wait for gemm1 LDS read
+        } while(++gemm1_k_block_outer_index < num_gemm1_k_block_outer_loop); // end j loop
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              Gemm1NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                gemm1_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                gemm1_blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, Gemm1NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, Gemm1NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp
new file mode 100644
index 0000000000..d2920570e4
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp
@@ -0,0 +1,1259 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename A0B0B1DataType, // FIXME: don't assume A0/B0/B1 have same datatype
+          typename Acc0DataType,
+          typename D0sDataType,
+          typename Acc1DataType,
+          typename C1ShuffleDataType,
+          typename D1sDataType,
+          typename E1DataType,
+          typename A0ElementwiseOperation,
+          typename B0ElementwiseOperation,
+          typename CDE0ElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CDE1ElementwiseOperation,
+          InMemoryDataOperationEnum E1GlobalMemoryDataOperation,
+          typename A0GridDesc_M_K,
+          typename B0GridDesc_N_K,
+          typename D0sGridDesc_M_N,
+          typename B1GridDesc_N_K,
+          typename D1sGridDesc_M_N,
+          typename E1GridDesc_M_N,
+          index_t NumGemm0KPrefetchStage,
+          index_t BlockSize,
+          index_t Gemm0MPerBlock,
+          index_t Gemm0NPerBlock,
+          index_t Gemm0KPerBlock,
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t A0K1Value,
+          index_t B0K1Value,
+          index_t B1K1Value,
+          index_t Gemm0MPerXdl,
+          index_t Gemm0NPerXdl,
+          index_t Gemm0MXdlPerWave,
+          index_t Gemm0NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename A0BlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename A0BlockTransferThreadClusterArrangeOrder,
+          typename A0BlockTransferSrcAccessOrder,
+          index_t A0BlockTransferSrcVectorDim,
+          index_t A0BlockTransferSrcScalarPerVector,
+          index_t A0BlockTransferDstScalarPerVector_AK1,
+          bool A0ThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t A0BlockLdsExtraM,
+          typename B0BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B0BlockTransferThreadClusterArrangeOrder,
+          typename B0BlockTransferSrcAccessOrder,
+          index_t B0BlockTransferSrcVectorDim,
+          index_t B0BlockTransferSrcScalarPerVector,
+          index_t B0BlockTransferDstScalarPerVector_BK1,
+          bool B0ThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t B0BlockLdsExtraN,
+          index_t CDE0BlockTransferSrcVectorDim,
+          index_t CDE0BlockTransferSrcScalarPerVector,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1ThreadTransferSrcResetCoordinateAfterRun,
+          index_t B1BlockLdsExtraN,
+          index_t C1ShuffleGemm0MXdlPerWavePerShuffle,
+          index_t C1ShuffleGemm0NXdlPerWavePerShuffle,
+          typename CDE1ShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDE1ShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched>
+struct GridwiseBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
+{
+    static_assert(LoopSched == LoopScheduler::Default,
+                  "Non-default loop scheduler is currently not supported");
+
+    static constexpr index_t NumD0Tensor = D0sDataType::Size();
+    static constexpr index_t NumD1Tensor = D1sDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    static constexpr auto WaveSize = 64;
+    // K1 should be Number<...>
+    // Gemm0
+    static constexpr auto A0K1 = Number<A0K1Value>{};
+    static constexpr auto B0K1 = Number<B0K1Value>{};
+
+    static constexpr auto A0K0PerBlock = Number<Gemm0KPerBlock / A0K1Value>{};
+    static constexpr auto B0K0PerBlock = Number<Gemm0KPerBlock / B0K1Value>{};
+
+    static constexpr auto Gemm0MWaves = Gemm0MPerBlock / (Gemm0MPerXdl * Gemm0MXdlPerWave);
+    static constexpr auto Gemm0NWaves = Gemm0NPerBlock / (Gemm0NPerXdl * Gemm0NXdlPerWave);
+    // Gemm1
+    static constexpr auto B1K1         = Number<B1K1Value>{};
+    static constexpr auto B1K0PerBlock = Number<Gemm1KPerBlock / B1K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = GridwiseGemmPipeline_v1<NumGemm0KPrefetchStage>;
+
+    // ck::Tuple<const D0DataType1*, const D0DataType2*, ...>
+    static constexpr auto MakeD0sGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
+
+                return static_cast<const D0DataType*>(nullptr);
+            },
+            Number<NumD0Tensor>{});
+    }
+
+    // ck::Tuple<const D1DataType1*, const D1DataType2*, ...>
+    static constexpr auto MakeD1sGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using D1DataType = remove_cvref_t<tuple_element_t<i.value, D1sDataType>>;
+
+                return static_cast<const D1DataType*>(nullptr);
+            },
+            Number<NumD1Tensor>{});
+    }
+
+    __device__ static auto GetGemm0WaveIdx()
+    {
+        const index_t thread_id = get_thread_local_1d_id();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(Gemm0MWaves, Gemm0NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto GetGemm0WaveMNIdx(const index_t thread_id)
+    {
+        constexpr auto wave_threadid_to_mn_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(WaveSize / Gemm0NPerXdl, Gemm0NPerXdl))),
+            make_tuple(Sequence<0, 1>{}),
+            make_tuple(Sequence<0>{}));
+
+        return wave_threadid_to_mn_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    template <typename A0BlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(const A0BlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = Gemm0MPerBlock / (Gemm0MXdlPerWave * Gemm0MPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm0MXdlPerWave, MWaves, Gemm0MPerXdl>(
+            A0BlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t NWaves = Gemm0NPerBlock / (Gemm0NXdlPerWave * Gemm0NPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm0NXdlPerWave, NWaves, Gemm0NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    template <typename A0BlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(const A0BlockDesc_AK0_M_AK1&)
+    {
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm0MXdlPerWave, 1, 1>(
+            A0BlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t Gemm1NWaves = Gemm1NPerBlock / (Gemm1NXdlPerWave * Gemm0NPerXdl);
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm1NXdlPerWave, Gemm1NWaves, Gemm0NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    __host__ __device__ static constexpr auto GetA0BlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A0 matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(A0K0PerBlock, Number<Gemm0MPerBlock>{}, A0K1),
+            make_tuple(Number<Gemm0MPerBlock + A0BlockLdsExtraM>{} * A0K1, A0K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetB0BlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B0 matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(B0K0PerBlock, Number<Gemm0NPerBlock>{}, B0K1),
+            make_tuple(Number<Gemm0NPerBlock + B0BlockLdsExtraN>{} * B0K1, B0K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B1 matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(B1K0PerBlock, Number<Gemm1NPerBlock>{}, B1K1),
+            make_tuple(Number<Gemm1NPerBlock + B1BlockLdsExtraN>{} * B1K1, B1K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetC1ShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = Gemm0MPerBlock / (Gemm0MXdlPerWave * Gemm0MPerXdl);
+        constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * Gemm0NPerXdl);
+
+        constexpr auto c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<C1ShuffleGemm0MXdlPerWavePerShuffle * MWave * Gemm0MPerXdl>{},
+                           I1,
+                           Number<C1ShuffleGemm0NXdlPerWavePerShuffle * NWave * Gemm0NPerXdl>{}));
+
+        return c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        const index_t gemm0_bytes_end = (SharedMemTrait::a0_block_space_size_aligned +
+                                         SharedMemTrait::b0_block_space_size_aligned) *
+                                        sizeof(A0B0B1DataType);
+        const index_t gemm1_bytes_end =
+            (SharedMemTrait::b1_block_space_offset + SharedMemTrait::b1_block_space_size_aligned) *
+            sizeof(A0B0B1DataType);
+        const index_t c1_block_bytes_end =
+            SharedMemTrait::c1_block_space_size * sizeof(C1ShuffleDataType);
+
+        return math::max(gemm0_bytes_end, gemm1_bytes_end, c1_block_bytes_end);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2E1TileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const A0GridDesc_M_K& a0_grid_desc_m_k,
+                  const B0GridDesc_N_K& b0_grid_desc_n_k,
+                  const B1GridDesc_N_K& b1_grid_desc_n_k,
+                  const E1GridDesc_M_N& e1_grid_desc_m_n,
+                  const Block2E1TileMap& block_2_e1tile_map)
+    {
+        static_assert((Gemm0MPerBlock % (Gemm0MPerXdl * Gemm0MXdlPerWave) == 0) &&
+                          (Gemm0NPerBlock % (Gemm0NXdlPerWave * Gemm0NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M      = a0_grid_desc_m_k.GetLength(I0);
+        const auto N      = b0_grid_desc_n_k.GetLength(I0);
+        const auto K      = a0_grid_desc_m_k.GetLength(I1);
+        const auto Gemm1N = b1_grid_desc_n_k.GetLength(I0);
+
+        if(!(M == e1_grid_desc_m_n.GetLength(I0) && Gemm1N == e1_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        if(!(M % Gemm0MPerBlock == 0 && N % Gemm0NPerBlock == 0 && K % Gemm0KPerBlock == 0 &&
+             Gemm1N % Gemm1NPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gemm0 gridwise gemm pipeline
+        const auto num_gemm0_k_loop = K / Gemm0KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm0_k_loop))
+        {
+            return false;
+        }
+
+        // check gemm1 gridwise gemm pipeline
+        if(!(Gemm0NPerBlock % Gemm1KPerBlock == 0))
+        {
+            return false;
+        }
+
+        const auto num_gemm1_k_inner_loop = Gemm0NPerBlock / Gemm1KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm1_k_inner_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_e1tile_map.CheckValidity(e1_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / Gemm0KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    // A0 desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultA0GridDescriptor_AK0_M_AK1(const A0GridDesc_M_K& a0_grid_desc_m_k)
+    {
+        const auto M = a0_grid_desc_m_k.GetLength(I0);
+        const auto K = a0_grid_desc_m_k.GetLength(I1);
+
+        const auto A0K0 = K / A0K1;
+
+        return transform_tensor_descriptor(
+            a0_grid_desc_m_k,
+            make_tuple(make_unmerge_transform(make_tuple(A0K0, A0K1)),
+                       make_pass_through_transform(M)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B0 desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultB0GridDescriptor_BK0_N_BK1(const B0GridDesc_N_K& b0_grid_desc_n_k)
+    {
+        const auto N = b0_grid_desc_n_k.GetLength(I0);
+        const auto K = b0_grid_desc_n_k.GetLength(I1);
+
+        const auto B0K0 = K / B0K1;
+
+        return transform_tensor_descriptor(
+            b0_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(B0K0, B0K1)),
+                       make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // D0 desc for source in blockwise copy
+    template <typename D0GridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeGemm0D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const D0GridDesc_M_N& d0_grid_desc_m_n)
+    {
+        const auto M = d0_grid_desc_m_n.GetLength(I0);
+        const auto N = d0_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto mfma =
+            MfmaSelector<A0B0B1DataType, Gemm0MPerXdl, Gemm0NPerXdl>::selected_mfma;
+        constexpr auto N3 = mfma.num_groups_per_blk;
+        constexpr auto N5 = mfma.group_size;
+        return transform_tensor_descriptor(
+            d0_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(
+                           M / Gemm0MPerBlock, Gemm0MXdlPerWave, Gemm0MWaves, Gemm0MPerXdl)),
+                       make_unmerge_transform(make_tuple(N / Gemm0NPerBlock,
+                                                         Gemm0NXdlPerWave,
+                                                         Gemm0NWaves,
+                                                         N3,
+                                                         WaveSize / Gemm0NPerXdl,
+                                                         N5))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7, 8, 9>{}));
+    }
+
+    // B1 desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultB1GridDescriptor_BK0_N_BK1(const B1GridDesc_N_K& b1_grid_desc_n_k)
+    {
+        const auto N = b1_grid_desc_n_k.GetLength(I0);
+        const auto K = b1_grid_desc_n_k.GetLength(I1);
+
+        const auto B1K0 = K / B1K1;
+
+        return transform_tensor_descriptor(
+            b1_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
+                       make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // C1 desc for destination in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeE1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const E1GridDesc_M_N& e1_grid_desc_m_n)
+    {
+        const auto M = e1_grid_desc_m_n.GetLength(I0);
+        const auto N = e1_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / Gemm0MPerBlock;
+        const auto NBlock = N / Gemm1NPerBlock;
+
+        const auto e1_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e1_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<Gemm0MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<Gemm1NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e1_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+    // D0s desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const D0sGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeGemm0D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(ds_grid_desc_m_n[i]);
+            },
+            Number<NumD0Tensor>{});
+    }
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDescriptor_M_N>
+    __host__ __device__ static constexpr auto
+    MakeD1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const DsGridDescriptor_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeE1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumD1Tensor>{});
+    }
+
+    // return block_id to C1 matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2E1TileMap(const E1GridDesc_M_N& e1_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<Gemm0MPerBlock, Gemm1NPerBlock, E1GridDesc_M_N>(
+            e1_grid_desc_m_n);
+    }
+
+    using E1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeE1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            E1GridDesc_M_N{}))>;
+
+    using D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 =
+        remove_cvref_t<decltype(MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(
+            D0sGridDesc_M_N{}))>;
+
+    using D1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeD1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            D1sGridDesc_M_N{}))>;
+
+    using DefaultBlock2E1TileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2E1TileMap(E1GridDesc_M_N{}))>;
+
+    struct SharedMemTrait
+    {
+        // LDS allocation for A0 and B0: be careful of alignment
+        static constexpr auto a0_block_desc_ak0_m_ak1 =
+            GetA0BlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        static constexpr auto b0_block_desc_bk0_n_bk1 =
+            GetB0BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        static constexpr auto b1_block_desc_bk0_n_bk1 =
+            GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        static constexpr auto max_lds_align = math::lcm(math::lcm(A0K1, B0K1), B1K1);
+
+        static constexpr auto a0_block_space_size_aligned = math::integer_least_multiple(
+            a0_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b0_block_space_size_aligned = math::integer_least_multiple(
+            b0_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b1_block_space_size_aligned = math::integer_least_multiple(
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        static constexpr auto a0_block_space_offset = 0;
+        static constexpr auto b0_block_space_offset = a0_block_space_size_aligned.value;
+        static constexpr auto b1_block_space_offset = 0;
+
+        // LDS allocation for C1 shuffle in LDS
+        static constexpr auto c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetC1ShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+        static constexpr auto c1_block_space_size =
+            c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+    };
+
+    using D0sGridPointer = decltype(MakeD0sGridPointer());
+    using D1sGridPointer = decltype(MakeD1sGridPointer());
+
+    template <bool HasMainKBlockLoop,
+              typename A0GridDesc_AK0_M_AK1,
+              typename B0GridDesc_BK0_N_BK1,
+              typename B1GridDesc_BK0_N_BK1,
+              typename Block2E1TileMap>
+    __device__ static void Run(const A0B0B1DataType* __restrict__ p_a0_grid,
+                               const A0B0B1DataType* __restrict__ p_b0_grid,
+                               D0sGridPointer p_d0s_grid,
+                               const A0B0B1DataType* __restrict__ p_b1_grid,
+                               D1sGridPointer p_d1s_grid,
+                               E1DataType* __restrict__ p_e1_grid,
+                               void* __restrict__ p_shared,
+                               const A0ElementwiseOperation& a0_element_op,
+                               const B0ElementwiseOperation& b0_element_op,
+                               const CDE0ElementwiseOperation& cde0_element_op,
+                               const B1ElementwiseOperation& b1_element_op,
+                               const CDE1ElementwiseOperation& cde1_element_op,
+                               const A0GridDesc_AK0_M_AK1& a0_grid_desc_ak0_m_ak1,
+                               const B0GridDesc_BK0_N_BK1& b0_grid_desc_bk0_n_bk1,
+                               const D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5&
+                                   d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                               const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                               const D1sGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   d1s_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const E1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e1_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2E1TileMap& block_2_e1tile_map)
+    {
+        const auto a0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a0_grid, a0_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b0_grid, b0_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        const auto b1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b1_grid, b1_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto e1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e1_grid, e1_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+        const auto d0s_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_d0s_grid[i],
+                    d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i].GetElementSpaceSize());
+            },
+            Number<NumD0Tensor>{});
+        const auto d1s_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_d1s_grid[i],
+                    d1s_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumD1Tensor>{});
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_e1tile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_e1tile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(e1_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e1_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * Gemm0MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * Gemm1NPerBlock);
+
+        // A0 matrix in LDS memory, dst of blockwise copy
+        constexpr auto a0_block_desc_ak0_m_ak1 = GetA0BlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B0 matrix in LDS memory, dst of blockwise copy
+        constexpr auto b0_block_desc_bk0_n_bk1 = GetB0BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        //
+        // set up Gemm0
+        //
+
+        // A0 matrix blockwise copy
+        auto a0_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                A0ElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<A0K0PerBlock, Gemm0MPerBlock, A0K1>,
+                                                A0BlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                A0BlockTransferThreadClusterArrangeOrder,
+                                                A0B0B1DataType,
+                                                A0B0B1DataType,
+                                                decltype(a0_grid_desc_ak0_m_ak1),
+                                                decltype(a0_block_desc_ak0_m_ak1),
+                                                A0BlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                A0BlockTransferSrcVectorDim,
+                                                2,
+                                                A0BlockTransferSrcScalarPerVector,
+                                                A0BlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemm0KPrefetchStage>(
+                a0_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a0_element_op,
+                a0_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // B0 matrix blockwise copy
+        auto b0_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                B0ElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<B0K0PerBlock, Gemm0NPerBlock, B0K1>,
+                                                B0BlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                B0BlockTransferThreadClusterArrangeOrder,
+                                                A0B0B1DataType,
+                                                A0B0B1DataType,
+                                                decltype(b0_grid_desc_bk0_n_bk1),
+                                                decltype(b0_block_desc_bk0_n_bk1),
+                                                B0BlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                B0BlockTransferSrcVectorDim,
+                                                2,
+                                                B0BlockTransferSrcScalarPerVector,
+                                                B0BlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemm0KPrefetchStage>(
+                b0_grid_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0), // will loop over GemmN dimension
+                b0_element_op,
+                b0_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // Fused Gemm+Gemm pipeline
+        // for n in N0:
+        //   for k in K0:
+        //     acc[m][n] += A[m][k] * B0[k][n]
+        //   acc1[m][o] += acc[m][n] * B1[n][o]
+
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(A0K1, B0K1),
+            MfmaSelector<A0B0B1DataType, Gemm0MPerXdl, Gemm0NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm0 = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            A0B0B1DataType,
+            Acc0DataType,
+            decltype(a0_block_desc_ak0_m_ak1),
+            decltype(b0_block_desc_bk0_n_bk1),
+            decltype(MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(a0_block_desc_ak0_m_ak1)),
+            decltype(MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(b0_block_desc_bk0_n_bk1)),
+            Gemm0MPerBlock,
+            Gemm0NPerBlock,
+            Gemm0KPerBlock,
+            Gemm0MPerXdl,
+            Gemm0NPerXdl,
+            Gemm0MXdlPerWave,
+            Gemm0NXdlPerWave,
+            KPack,
+            true>{}; // TransposeC
+
+        auto acc0_thread_buf = blockwise_gemm0.GetCThreadBuffer();
+
+        // LDS allocation for A0 and B0: be careful of alignment
+        auto a0_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<A0B0B1DataType*>(p_shared) + SharedMemTrait::a0_block_space_offset,
+            a0_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b0_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<A0B0B1DataType*>(p_shared) + SharedMemTrait::b0_block_space_offset,
+            b0_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a0_block_slice_copy_step = make_multi_index(Gemm0KPerBlock / A0K1, 0, 0);
+        constexpr auto b0_block_slice_copy_step = make_multi_index(Gemm0KPerBlock / B0K1, 0, 0);
+        const auto a0_block_reset_copy_step =
+            make_multi_index(-a0_grid_desc_ak0_m_ak1.GetLength(I0), 0, 0);
+        const auto b0_block_reset_copy_step =
+            make_multi_index(-b0_grid_desc_bk0_n_bk1.GetLength(I0), Gemm0NPerBlock, 0);
+
+        // gridwise GEMM pipeline
+        // Only supports LoopScheduler::Default
+        const auto gridwise_gemm0_pipeline =
+            GridwiseGemmPipeline_v1_Selector<NumGemm0KPrefetchStage, LoopScheduler::Default>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a0_grid_desc_ak0_m_ak1.GetLength(I0) * a0_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            Gemm0KPerBlock);
+
+        //
+        // set up Gemm1
+        //
+
+        // Acc0 matrix threadwise copy: AccVGPR to VGPR and downcast to XDL input data type
+        constexpr auto acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm0.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+        constexpr auto m0 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+        constexpr auto n0 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+        constexpr auto m1 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+        constexpr auto n1 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+        constexpr auto m2 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+        constexpr auto n2 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+        constexpr auto n3 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+        constexpr auto n4 = acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+
+        constexpr auto b1_block_slice_copy_step = make_multi_index(Gemm1KPerBlock / B1K1, 0, 0);
+
+        // d0 matrix threadwise copy
+        constexpr auto d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,   // MBlockId
+                                                           I1,   // NBlockID
+                                                           m0,   // MRepeat
+                                                           n0,   // NRepeat
+                                                           m1,   // MWaveId
+                                                           n1,   // NWaveId
+                                                           m2,   // MPerXdl
+                                                           n2,   // NGroupNum
+                                                           n3,   // NInputNum
+                                                           n4)); // registerNum
+
+        auto d0s_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<
+                    AddressSpaceEnum::Vgpr,
+                    A0B0B1DataType,
+                    d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize(),
+                    true>{};
+            },
+            Number<NumD0Tensor>{});
+
+        const auto wave_id     = GetGemm0WaveIdx();
+        const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63
+
+        static_assert(CDE0BlockTransferSrcScalarPerVector <= n4,
+                      "vector load must be not greater than n4");
+        static_assert(n4 % CDE0BlockTransferSrcScalarPerVector == 0);
+
+        auto d0s_threadwise_copy = generate_tuple(
+            [&](auto i) {
+                return ThreadwiseTensorSliceTransfer_v2<
+                    A0B0B1DataType,
+                    A0B0B1DataType,
+                    decltype(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i]),
+                    decltype(d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
+                    Sequence<I1, // MBlockId
+                             I1, // NBlockID
+                             m0, // MRepeat
+                             n0, // NRepeat
+                             m1, // MWaveId
+                             n1, // NWaveId
+                             m2, // MPerXdl
+                             n2, // NGroupNum
+                             n3, // NInputNum
+                             n4>,
+                    Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>,
+                    9, // CDE0BlockTransferSrcVectorDim
+                    CDE0BlockTransferSrcScalarPerVector,
+                    1,
+                    false>(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                           make_multi_index(block_work_idx[I0], // MBlockId
+                                            0,                  // NBlockId
+                                            0,                  // mrepeat
+                                            0,                  // nrepeat
+                                            wave_id[I0],        // MWaveId
+                                            wave_id[I1],        // NWaveId
+                                            wave_m_n_id[I1],    // MPerXdl
+                                            0,                  // group
+                                            wave_m_n_id[I0],    // NInputIndex
+                                            0));                // register number
+            },
+            Number<NumD0Tensor>{});
+        // acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 to acc0_thread_desc_k0_m_k1
+        // n0_n1_n2_n3 -> k0
+        // m0_m1_m2 -> m
+        // n4 -> k1
+        // NOTE: had to use merge_v3 or will spit out compilation errors
+        constexpr auto acc0_thread_desc_k0_m_k1 = transform_tensor_descriptor(
+            acc0_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(n0, n1, n2, n3)),
+                       make_merge_transform_v3_division_mod(make_tuple(m0, m1, m2)),
+                       make_pass_through_transform(n4)),
+            make_tuple(Sequence<1, 3, 5, 6>{}, Sequence<0, 2, 4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        // A1 matrix in AccVGPR
+        // N2 num_groups_per_blk, N3 num_input_blks, N4 group_size
+        constexpr auto Acc0N3 =
+            blockwise_gemm0.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLength(I6);
+
+        constexpr auto A1ThreadSlice_K0_M_K1 = make_tuple(
+            Number<Gemm1KPerBlock / n4 / Acc0N3>{}, Number<m0 * m1 * m2>{}, Number<n4>{});
+
+        constexpr auto A1ThreadSliceK0        = A1ThreadSlice_K0_M_K1[I0];
+        constexpr auto A1ThreadSliceM         = A1ThreadSlice_K0_M_K1[I1];
+        constexpr auto A1ThreadSliceK1        = A1ThreadSlice_K0_M_K1[I2];
+        constexpr auto a1_thread_desc_k0_m_k1 = make_naive_tensor_descriptor(
+            A1ThreadSlice_K0_M_K1,
+            make_tuple(A1ThreadSliceM * A1ThreadSliceK1, A1ThreadSliceK1, I1));
+
+        // B1 matrix in LDS memory, dst of blockwise copy
+        constexpr auto b1_block_desc_bk0_n_bk1 = GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A1 matrix blockwise copy
+        auto a1_blockwise_copy = ThreadwiseTensorSliceTransfer_StaticToStatic<
+            Acc0DataType,
+            A0B0B1DataType,
+            decltype(acc0_thread_desc_k0_m_k1),
+            decltype(a1_thread_desc_k0_m_k1),
+            tensor_operation::element_wise::PassThrough,
+            Sequence<A1ThreadSliceK0, A1ThreadSliceM, A1ThreadSliceK1>,
+            Sequence<1, 0, 2>,
+            2,
+            n4>{tensor_operation::element_wise::PassThrough{}};
+
+        // B1 matrix blockwise copy
+        auto b1_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                B0ElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<B1K0PerBlock, Gemm1NPerBlock, B1K1>,
+                                                B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                B1BlockTransferThreadClusterArrangeOrder,
+                                                A0B0B1DataType,
+                                                A0B0B1DataType,
+                                                decltype(b1_grid_desc_bk0_n_bk1),
+                                                decltype(b1_block_desc_bk0_n_bk1),
+                                                B1BlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                B1BlockTransferSrcVectorDim,
+                                                2,
+                                                B1BlockTransferSrcScalarPerVector,
+                                                B1BlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                B1ThreadTransferSrcResetCoordinateAfterRun,
+                                                true, // DstResetCoord
+                                                1>(b1_grid_desc_bk0_n_bk1,
+                                                   make_multi_index(0, n_block_data_idx_on_grid, 0),
+                                                   b1_element_op,
+                                                   b1_block_desc_bk0_n_bk1,
+                                                   make_multi_index(0, 0, 0),
+                                                   tensor_operation::element_wise::PassThrough{});
+
+        auto a1_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, A0B0B1DataType>(
+            a1_thread_desc_k0_m_k1.GetElementSpaceSize());
+
+        // reuse LDS space for gemm0's b0_block_buf
+        auto b1_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<A0B0B1DataType*>(p_shared) + SharedMemTrait::b1_block_space_offset,
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr index_t Gemm1KPack = math::max(
+            math::lcm(
+                MfmaSelector<A0B0B1DataType, Gemm0MPerXdl, Gemm0NPerXdl>::selected_mfma.group_size,
+                B1K1),
+            MfmaSelector<A0B0B1DataType, Gemm0MPerXdl, Gemm0NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm1 = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            A0B0B1DataType,
+            Acc1DataType,
+            decltype(a1_thread_desc_k0_m_k1),
+            decltype(b1_block_desc_bk0_n_bk1),
+            decltype(MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(a1_thread_desc_k0_m_k1)),
+            decltype(MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(b1_block_desc_bk0_n_bk1)),
+            Gemm0MPerBlock,
+            Gemm1NPerBlock,
+            Gemm1KPerBlock,
+            Gemm0MPerXdl,
+            Gemm0NPerXdl,
+            Gemm0MXdlPerWave,
+            Gemm1NXdlPerWave,
+            Gemm1KPack,
+            false,      // TransposeC
+            Gemm1KPack, // AMmaKStride
+            Gemm1KPack * XdlopsGemm<A0B0B1DataType, Gemm0MPerXdl, Gemm0NPerXdl, Gemm1KPack, false>{}
+                             .K0PerXdlops>{                         // BMmaKStride
+                                           make_tuple(0, 0, 0, 0)}; // A_origin
+
+        auto c1_thread_buf = blockwise_gemm1.GetCThreadBuffer();
+
+        const index_t num_gemm1_k_block_outer_loop =
+            b0_grid_desc_bk0_n_bk1.GetLength(I1) / Gemm0NPerBlock;
+        constexpr index_t num_gemm1_k_block_inner_loop = Gemm0NPerBlock / Gemm1KPerBlock;
+
+        // Initialize C1
+        c1_thread_buf.Clear();
+
+        // gemm1 K loop
+        index_t gemm1_k_block_outer_index = 0;
+        do
+        {
+            // gemm0
+            gridwise_gemm0_pipeline.template Run<HasMainKBlockLoop>(a0_grid_desc_ak0_m_ak1,
+                                                                    a0_block_desc_ak0_m_ak1,
+                                                                    a0_blockwise_copy,
+                                                                    a0_grid_buf,
+                                                                    a0_block_buf,
+                                                                    a0_block_slice_copy_step,
+                                                                    b0_grid_desc_bk0_n_bk1,
+                                                                    b0_block_desc_bk0_n_bk1,
+                                                                    b0_blockwise_copy,
+                                                                    b0_grid_buf,
+                                                                    b0_block_buf,
+                                                                    b0_block_slice_copy_step,
+                                                                    blockwise_gemm0,
+                                                                    acc0_thread_buf,
+                                                                    num_k_block_main_loop);
+            // multiple d
+            if constexpr(NumD0Tensor)
+            {
+                static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                    d0s_threadwise_copy(i).Run(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                                               d0s_grid_buf[i],
+                                               d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                                               make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                                               d0s_thread_buf(i));
+                });
+                static_for<0, m0 * n0 * n2 * n4, 1>{}([&](auto i) {
+                    // get reference to src data
+                    const auto src_data_refs = generate_tie(
+                        // return type should be lvalue
+                        [&](auto iSrc) -> const auto& { return d0s_thread_buf[iSrc][i]; },
+                        Number<NumD0Tensor>{});
+
+                    // get reference to dst data
+                    auto dst_data_refs = generate_tie(
+                        // return type should be lvalue
+                        [&](auto) -> auto& { return acc0_thread_buf(i); },
+                        Number<2>{});
+
+                    unpack2(cde0_element_op, dst_data_refs, src_data_refs);
+                });
+                static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                    d0s_threadwise_copy(i).MoveSrcSliceWindow(
+                        d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                        make_multi_index(0, 1, 0, 0, 0, 0, 0, 0, 0, 0));
+                });
+            }
+            else
+            {
+                static_for<0, acc0_thread_buf.Size(), 1>{}(
+                    [&](auto i) { cde0_element_op(acc_thread_buf(i), acc0_thread_buf[i]); });
+            }
+            // gemm1
+            {
+                // TODO: explore using dynamic buffer for a1 thread buffer
+                // For a1_blockwise_copy, the goal is to satisfy pipeline requirements RunRead(),
+                // RunWrite(), and MoveSliceWindow(). But it is impossible to implement given that
+                // the A1 source buffer is static buffer holding the output of first GEMM and
+                // requires constexpr offset by design. Therefore, we pass tensor coordinate offset
+                // explicitly in Run() below.
+
+                // preload data into LDS
+                b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                     b1_block_slice_copy_step);
+
+                block_sync_lds(); // wait for gemm0 LDS read
+
+                b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+
+                // main body
+                if constexpr(num_gemm1_k_block_inner_loop > 1)
+                {
+                    static_for<0, num_gemm1_k_block_inner_loop - 1, 1>{}([&](auto i) {
+                        a1_blockwise_copy.Run(acc0_thread_desc_k0_m_k1,
+                                              make_tuple(Number<i * A1ThreadSliceK0>{}, I0, I0),
+                                              acc0_thread_buf,
+                                              a1_thread_desc_k0_m_k1,
+                                              make_tuple(I0, I0, I0),
+                                              a1_thread_buf);
+
+                        b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                        block_sync_lds();
+
+                        blockwise_gemm1.Run(a1_thread_buf, b1_block_buf, c1_thread_buf);
+
+                        block_sync_lds();
+
+                        b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                             b1_block_slice_copy_step);
+
+                        b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+                    });
+                }
+                // tail
+                {
+                    a1_blockwise_copy.Run(
+                        acc0_thread_desc_k0_m_k1,
+                        make_tuple(
+                            Number<(num_gemm1_k_block_inner_loop - 1) * A1ThreadSliceK0>{}, I0, I0),
+                        acc0_thread_buf,
+                        a1_thread_desc_k0_m_k1,
+                        make_tuple(I0, I0, I0),
+                        a1_thread_buf);
+
+                    block_sync_lds();
+
+                    blockwise_gemm1.Run(a1_thread_buf, b1_block_buf, c1_thread_buf);
+                }
+            } // end gemm1
+
+            a0_blockwise_copy.MoveSrcSliceWindow(a0_grid_desc_ak0_m_ak1,
+                                                 a0_block_reset_copy_step); // rewind K
+            b0_blockwise_copy.MoveSrcSliceWindow(b0_grid_desc_bk0_n_bk1,
+                                                 b0_block_reset_copy_step); // rewind K and step N
+
+            block_sync_lds(); // wait for gemm1 LDS read
+        } while(++gemm1_k_block_outer_index < num_gemm1_k_block_outer_loop); // end j loop
+
+        // shuffle C1 and write out
+        {
+            static_assert(Gemm0MXdlPerWave % C1ShuffleGemm0MXdlPerWavePerShuffle == 0 &&
+                              Gemm1NXdlPerWave % C1ShuffleGemm0NXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = Gemm0MPerBlock / (Gemm0MXdlPerWave * Gemm0MPerXdl);
+            constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * Gemm0NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c1_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm1.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm1.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetC1ShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c1_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<C1ShuffleDataType*>(p_shared),
+                c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<C1ShuffleGemm0MXdlPerWavePerShuffle>{}, // M0 (Gemm0MXdlPerWave) per
+                                                                       // shuffle
+                        M1,                                            // M1 = MWave
+                        M2, // M2 * M3 * M4 = Gemm0MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<C1ShuffleGemm0NXdlPerWavePerShuffle>{}, // N0 (Gemm0NXdlPerWave) per
+                                                                       // shuffle
+                        N1,                                            // N1 = NWave
+                        N2))),                                         // N2 = Gemm0NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM C1 matrix starting index
+            const auto c1_thread_mtx_on_block =
+                blockwise_gemm1.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c1_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c1_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c1_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<Acc1DataType,
+                                                   C1ShuffleDataType,
+                                                   decltype(c1_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   tensor_operation::element_wise::PassThrough,
+                                                   Sequence<C1ShuffleGemm0MXdlPerWavePerShuffle,
+                                                            C1ShuffleGemm0NXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    tensor_operation::element_wise::PassThrough{}};
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c1_d1s_desc_refs = concat_tuple_of_reference(
+                tie(c1_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return d1s_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumD1Tensor>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c1_d1s_buf_refs = concat_tuple_of_reference(
+                tie(c1_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return d1s_grid_buf[i]; },
+                    Number<NumD1Tensor>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c1_d1s_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
+                    },
+                    Number<NumD1Tensor>{}));
+
+            // shuffle: blockwise copy C from LDS to global
+            auto cde1_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v7<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(C1ShuffleDataType{}), D1sDataType{})),
+                Tuple<E1DataType>,
+                decltype(c1_d1s_desc_refs),
+                decltype(tie(e1_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CDE1ElementwiseOperation,
+                Sequence<static_cast<index_t>(E1GlobalMemoryDataOperation)>, // FIXME: make Sequence
+                                                                             // support arbitray
+                                                                             // type
+                Sequence<1,
+                         C1ShuffleGemm0MXdlPerWavePerShuffle * MWave * Gemm0MPerXdl,
+                         1,
+                         C1ShuffleGemm0NXdlPerWavePerShuffle * NWave *
+                             Gemm0NPerXdl>, // BlockSliceLengths,
+                CDE1ShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
+                3,                    // index_t VectorDim,
+                CDE1ShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumD1Tensor,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c1_d1s_desc_refs,
+                 idx_c1_d1s_block_begin,
+                 tie(e1_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0)),
+                 cde1_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c1_vgpr =
+                SpaceFillingCurve<Sequence<Gemm0MXdlPerWave, Gemm1NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<C1ShuffleGemm0MXdlPerWavePerShuffle,
+                                           C1ShuffleGemm0NXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_e1_global = SpaceFillingCurve<
+                Sequence<1, Gemm0MPerBlock, 1, Gemm1NPerBlock>,
+                Sequence<0, 2, 1, 3>,
+                Sequence<1,
+                         C1ShuffleGemm0MXdlPerWavePerShuffle * MWave * Gemm0MPerXdl,
+                         1,
+                         C1ShuffleGemm0NXdlPerWavePerShuffle * NWave * Gemm0NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c1_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_e1_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c1_thread_copy_vgpr_to_lds.Run(c1_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                               sfc_c1_vgpr.GetIndexTupleOfNumber(access_id),
+                                               c1_thread_buf,
+                                               c1_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                               c1_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde1_shuffle_block_copy_lds_to_global.Run(
+                    c1_d1s_desc_refs,
+                    c1_d1s_buf_refs,
+                    tie(e1_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(e1_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto e1_global_step = sfc_e1_global.GetForwardStep(access_id);
+
+                    // move on D1s
+                    static_for<0, NumD1Tensor, 1>{}([&](auto i) {
+                        cde1_shuffle_block_copy_lds_to_global.MoveSrcSliceWindow(
+                            c1_d1s_desc_refs, i + I1, e1_global_step);
+                    });
+
+                    // move on C
+                    cde1_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        tie(e1_grid_desc_mblock_mperblock_nblock_nperblock), I0, e1_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp
new file mode 100644
index 0000000000..6fd49c8752
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp
@@ -0,0 +1,1303 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_softmax.hpp"
+
+namespace ck {
+
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename D0sDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename C0DEElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename C1DEElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename B1GridDesc_BK0_N_BK1,
+          typename C1GridDesc_M_N,
+          typename D0sGridDesc_M_N,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t B1K1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1ThreadTransferSrcResetCoordinateAfterRun,
+          index_t B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          bool PadN,
+          bool MaskOutUpperTriangle,
+          int D0sTransferSrcScalarPerVector = 4,
+          PipelineVersion PipelineVer       = PipelineVersion::v1>
+struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle
+{
+    static_assert(LoopSched == LoopScheduler::Default,
+                  "Non-default loop scheduler is currently not supported");
+
+    static constexpr index_t NumD0Tensor = D0sDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    // Gemm0
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    static constexpr auto Gemm0MWaves = MPerBlock / (MPerXdl * MXdlPerWave);
+    static constexpr auto Gemm0NWaves = NPerBlock / (NPerXdl * NXdlPerWave);
+
+    // Gemm1
+    static constexpr auto B1K0 = Number<Gemm1KPerBlock / B1K1Value>{};
+    static constexpr auto B1K1 = Number<B1K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, MWaves, MPerXdl>(
+            ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<NXdlPerWave, NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, 1, 1>(ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t Gemm1NWaves = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm1NXdlPerWave, Gemm1NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B1 matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(B1K0, Number<Gemm1NPerBlock>{}, B1K1),
+            make_tuple(Number<Gemm1NPerBlock + B1BlockLdsExtraN>{} * B1K1, B1K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        const index_t gemm0_bytes_end = (SharedMemTrait::a_block_space_size_aligned +
+                                         SharedMemTrait::b_block_space_size_aligned) *
+                                        sizeof(FloatAB);
+        const index_t gemm1_bytes_end =
+            (SharedMemTrait::b1_block_space_offset + SharedMemTrait::b1_block_space_size_aligned) *
+            sizeof(FloatAB);
+        const index_t softmax_bytes_end = (SharedMemTrait::reduction_space_offset +
+                                           SharedMemTrait::reduction_space_size_aligned) *
+                                          sizeof(FloatGemmAcc);
+        const index_t c_block_bytes_end =
+            SharedMemTrait::c_block_space_size * sizeof(FloatCShuffle);
+
+        return math::max(gemm0_bytes_end, gemm1_bytes_end, softmax_bytes_end, c_block_bytes_end);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                  const C1GridDesc_M_N& c1_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+        const auto Gemm1N = b1_grid_desc_bk0_n_bk1.GetLength(I1);
+
+        if(!(M == c1_grid_desc_m_n.GetLength(I0) && Gemm1N == c1_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0 &&
+             Gemm1N % Gemm1NPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gemm0 gridwise gemm pipeline
+        const auto num_gemm0_k_loop = K / KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm0_k_loop))
+        {
+            return false;
+        }
+
+        // check gemm1 gridwise gemm pipeline
+        if(!(NPerBlock % Gemm1KPerBlock == 0))
+        {
+            return false;
+        }
+
+        const auto num_gemm1_k_inner_loop = NPerBlock / Gemm1KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm1_k_inner_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c1_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeC1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const C1GridDesc_M_N& c1_grid_desc_m_n)
+    {
+        const auto M = c1_grid_desc_m_n.GetLength(I0);
+        const auto N = c1_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / Gemm1NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c1_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<Gemm1NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const C1GridDesc_M_N& c1_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, Gemm1NPerBlock, C1GridDesc_M_N>(
+            c1_grid_desc_m_n);
+    }
+
+    __device__ static auto GetGemm0WaveIdx()
+    {
+        const index_t thread_id = get_thread_local_1d_id();
+        constexpr auto WaveSize = MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.wave_size;
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(Gemm0MWaves, Gemm0NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto GetGemm0WaveMNIdx(const index_t thread_id)
+    {
+        constexpr auto WaveSize = MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.wave_size;
+        constexpr auto wave_threadid_to_mn_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(WaveSize / MPerXdl, MPerXdl))),
+            make_tuple(Sequence<0, 1>{}),
+            make_tuple(Sequence<0>{}));
+
+        return wave_threadid_to_mn_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    static constexpr auto MakeD0sGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
+
+                return static_cast<const D0DataType*>(nullptr);
+            },
+            Number<NumD0Tensor>{});
+    }
+    // D0 desc for source in blockwise copy
+    template <typename D0GridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeGemm0D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const D0GridDesc_M_N& d0_grid_desc_m_n)
+    {
+        const auto M = d0_grid_desc_m_n.GetLength(I0);
+        const auto N = d0_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto mfma = MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma;
+        constexpr auto N3   = mfma.num_groups_per_blk;
+        constexpr auto N4   = mfma.num_input_blks;
+        constexpr auto N5   = mfma.group_size;
+        return transform_tensor_descriptor(
+            d0_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(
+                           make_tuple(M / MPerBlock, MXdlPerWave, Gemm0MWaves, MPerXdl)),
+                       make_unmerge_transform(
+                           make_tuple(N / NPerBlock, NXdlPerWave, Gemm0NWaves, N3, N4, N5))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7, 8, 9>{}));
+    }
+
+    // D0s desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const D0sGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeGemm0D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(ds_grid_desc_m_n[i]);
+            },
+            Number<NumD0Tensor>{});
+    }
+
+    using D0sGridPointer = decltype(MakeD0sGridPointer());
+    using D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 =
+        remove_cvref_t<decltype(MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(
+            D0sGridDesc_M_N{}))>;
+
+    using C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeC1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            C1GridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(C1GridDesc_M_N{}))>;
+
+    struct SharedMemTrait
+    {
+        // LDS allocation for A and B: be careful of alignment
+        static constexpr auto a_block_desc_ak0_m_ak1 =
+            GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        static constexpr auto b_block_desc_bk0_n_bk1 =
+            GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        static constexpr auto b1_block_desc_bk0_n_bk1 =
+            GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        static constexpr auto max_lds_align = math::lcm(math::lcm(AK1, BK1), B1K1);
+
+        static constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b1_block_space_size_aligned = math::integer_least_multiple(
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        static constexpr auto a_block_space_offset  = 0;
+        static constexpr auto b_block_space_offset  = a_block_space_size_aligned.value;
+        static constexpr auto b1_block_space_offset = 0;
+
+        // LDS allocation for reduction
+        static constexpr index_t reduction_space_size_aligned =
+            math::integer_least_multiple(BlockSize, max_lds_align);
+
+        static constexpr auto reduction_space_offset = 0;
+
+        // LDS allocation for C shuffle in LDS
+        static constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+        static constexpr auto c_block_space_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+    };
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap, typename C0MatrixMask>
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               const FloatAB* __restrict__ p_b1_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               D0sGridPointer p_d0s_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const C0DEElementwiseOperation& c0de_element_op,
+                               const B1ElementwiseOperation& b1_element_op,
+                               const C1DEElementwiseOperation& c1de_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                               const C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5&
+                                   d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                               const Block2CTileMap& block_2_ctile_map,
+                               const C0MatrixMask& c0_matrix_mask)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        const auto b1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b1_grid, b1_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+        const auto d0s_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_d0s_grid[i],
+                    d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i].GetElementSpaceSize());
+            },
+            Number<NumD0Tensor>{});
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/gemm1_n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t gemm1_n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * Gemm1NPerBlock);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        //
+        // set up Gemm0
+        //
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0), // will loop over GemmN dimension
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // Fused Gemm+Gemm pipeline
+        // for n in N0:
+        //   for k in K0:
+        //     acc[m][n] += A[m][k] * B0[k][n]
+        //   acc1[m][o] += acc[m][n] * B1[n][o]
+
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            decltype(MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(a_block_desc_ak0_m_ak1)),
+            decltype(MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(b_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            NPerBlock,
+            KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            true>{}; // TransposeC
+
+        auto acc_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::a_block_space_offset,
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::b_block_space_offset,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+        const auto a_block_reset_copy_step =
+            make_multi_index(-a_grid_desc_ak0_m_ak1.GetLength(I0), 0, 0);
+        const auto b_block_reset_copy_step =
+            make_multi_index(-b_grid_desc_bk0_n_bk1.GetLength(I0), NPerBlock, 0);
+
+        // gridwise GEMM pipeline
+        // Only supports LoopScheduler::Default
+        const auto gridwise_gemm_pipeline = GridwiseGemmPipeline_Selector<PipelineVer,
+                                                                          NumGemmKPrefetchStage,
+                                                                          LoopScheduler::Default>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        //
+        // set up Gemm1
+        //
+
+        // Acc matrix threadwise copy: AccVGPR to VGPR and downcast to XDL input data type
+        constexpr auto acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+        constexpr auto m0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+        constexpr auto n0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+        constexpr auto m1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+        constexpr auto n1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+        constexpr auto m2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+        constexpr auto n2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+        constexpr auto n3 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+        constexpr auto n4 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+
+        constexpr auto b1_block_slice_copy_step = make_multi_index(Gemm1KPerBlock / B1K1, 0, 0);
+
+        // d0 matrix threadwise copy
+        constexpr auto d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,   // MBlockId
+                                                           I1,   // NBlockID
+                                                           m0,   // MRepeat
+                                                           n0,   // NRepeat
+                                                           m1,   // MWaveId
+                                                           n1,   // NWaveId
+                                                           m2,   // MPerXdl
+                                                           n2,   // NGroupNum
+                                                           n3,   // NInputNum
+                                                           n4)); // registerNum
+
+        auto d0s_thread_buf = generate_tuple(
+            [&](auto i) {
+                using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
+                return StaticBuffer<
+                    AddressSpaceEnum::Vgpr,
+                    D0DataType,
+                    d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize(),
+                    true>{};
+            },
+            Number<NumD0Tensor>{});
+
+        const auto wave_id     = GetGemm0WaveIdx();
+        const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63
+
+        auto d0s_threadwise_copy = generate_tuple(
+            [&](auto i) {
+                using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
+                return ThreadwiseTensorSliceTransfer_v2<
+                    D0DataType,
+                    D0DataType,
+                    decltype(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i]),
+                    decltype(d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
+                    Sequence<I1, // MBlockId
+                             I1, // NBlockID
+                             m0, // MRepeat
+                             n0, // NRepeat
+                             m1, // MWaveId
+                             n1, // NWaveId
+                             m2, // MPerXdl
+                             n2, // NGroupNum
+                             n3, // NInputNum
+                             n4>,
+                    Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>,
+                    9,
+                    D0sTransferSrcScalarPerVector,
+                    1,
+                    false>(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                           make_multi_index(block_work_idx[I0], // MBlockId
+                                            0,                  // NBlockId
+                                            0,                  // mrepeat
+                                            0,                  // nrepeat
+                                            wave_id[I0],        // MWaveId
+                                            wave_id[I1],        // NWaveId
+                                            wave_m_n_id[I1],    // MPerXdl
+                                            0,                  // group
+                                            wave_m_n_id[I0],    // NInputIndex
+                                            0));                // register number
+            },
+            Number<NumD0Tensor>{});
+        // acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 to acc_thread_desc_k0_m_k1
+        // n0_n1_n2_n3 -> k0
+        // m0_m1_m2 -> m
+        // n4 -> k1
+        // NOTE: had to use merge_v3 or will spit out compilation errors
+        constexpr auto acc_thread_desc_k0_m_k1 = transform_tensor_descriptor(
+            acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(n0, n1, n2, n3)),
+                       make_merge_transform_v3_division_mod(make_tuple(m0, m1, m2)),
+                       make_pass_through_transform(n4)),
+            make_tuple(Sequence<1, 3, 5, 6>{}, Sequence<0, 2, 4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        // A1 matrix in AccVGPR
+        // N2 num_groups_per_blk, N3 num_input_blks, N4 group_size
+        constexpr auto AccN3 =
+            blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLength(I6);
+
+        constexpr auto A1ThreadSlice_K0_M_K1 =
+            make_tuple(Number<Gemm1KPerBlock / n4 / AccN3>{}, Number<m0 * m1 * m2>{}, Number<n4>{});
+
+        constexpr auto A1ThreadSliceK0        = A1ThreadSlice_K0_M_K1[I0];
+        constexpr auto A1ThreadSliceM         = A1ThreadSlice_K0_M_K1[I1];
+        constexpr auto A1ThreadSliceK1        = A1ThreadSlice_K0_M_K1[I2];
+        constexpr auto a1_thread_desc_k0_m_k1 = make_naive_tensor_descriptor(
+            A1ThreadSlice_K0_M_K1,
+            make_tuple(A1ThreadSliceM * A1ThreadSliceK1, A1ThreadSliceK1, I1));
+
+        // B1 matrix in LDS memory, dst of blockwise copy
+        constexpr auto b1_block_desc_bk0_n_bk1 = GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A1 matrix blockwise copy
+        auto a1_blockwise_copy = ThreadwiseTensorSliceTransfer_StaticToStatic<
+            FloatGemmAcc,
+            FloatAB,
+            decltype(acc_thread_desc_k0_m_k1),
+            decltype(a1_thread_desc_k0_m_k1),
+            tensor_operation::element_wise::PassThrough,
+            Sequence<A1ThreadSliceK0, A1ThreadSliceM, A1ThreadSliceK1>,
+            Sequence<1, 0, 2>,
+            2,
+            n4>{tensor_operation::element_wise::PassThrough{}};
+
+        // B1 matrix blockwise copy
+        auto b1_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<B1K0, Gemm1NPerBlock, B1K1>,
+                                                B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                B1BlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b1_grid_desc_bk0_n_bk1),
+                                                decltype(b1_block_desc_bk0_n_bk1),
+                                                B1BlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                B1BlockTransferSrcVectorDim,
+                                                2,
+                                                B1BlockTransferSrcScalarPerVector,
+                                                B1BlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                B1ThreadTransferSrcResetCoordinateAfterRun,
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b1_grid_desc_bk0_n_bk1,
+                make_multi_index(0, gemm1_n_block_data_idx_on_grid, 0),
+                b1_element_op,
+                b1_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        auto a1_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a1_thread_desc_k0_m_k1.GetElementSpaceSize());
+
+        // reuse LDS space for gemm0's b_block_buf
+        auto b1_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::b1_block_space_offset,
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        // selected_mfma.group_size or B1K1 <= Gemm1KPack <= selected_mfma.group_size
+        // selected_mfma.k_per_blk <= Gemm1KPack
+        //
+        // Following similar rationale behind Gemm0KPack, let Gemm1KPack be the lowest common
+        // multiples of A1K1 (predetermined by selected_mfma.group_size) and B1K1. But in this case
+        // Gemm1KPack can't be higher than A1K1 itself because A1 matrix is distributed in VGPRs
+        // with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
+        // cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
+        // therefore we may just as well assign Gemm1KPack = group_size
+        constexpr index_t Gemm1KPack =
+            MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size;
+
+        auto gemm1_blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a1_thread_desc_k0_m_k1),
+            decltype(b1_block_desc_bk0_n_bk1),
+            decltype(MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(a1_thread_desc_k0_m_k1)),
+            decltype(MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(b1_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            Gemm1NPerBlock,
+            Gemm1KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            Gemm1NXdlPerWave,
+            Gemm1KPack,
+            true,       // TransposeC
+            Gemm1KPack, // AMmaKStride
+            Gemm1KPack * XdlopsGemm<FloatAB, MPerXdl, NPerXdl, Gemm1KPack, false>{}.K0PerXdlops>{
+            // BMmaKStride
+            make_multi_index(0, 0, 0, 0)}; // A_origin
+
+        auto acc1_thread_buf = gemm1_blockwise_gemm.GetCThreadBuffer();
+
+        //
+        // Blockwise softmax
+        //
+        auto workspace_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatGemmAcc*>(p_shared) + SharedMemTrait::reduction_space_offset,
+            SharedMemTrait::reduction_space_size_aligned);
+
+        // get acc0 8D thread cluster
+        constexpr auto thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths() /
+            blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+        constexpr auto tm0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I0);
+        constexpr auto tn0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I1);
+        constexpr auto tm1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I2);
+        constexpr auto tn1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I3);
+        constexpr auto tm2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I4);
+        constexpr auto tn2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I5);
+        constexpr auto tn3 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I6);
+        constexpr auto tn4 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I7);
+
+        // get acc0 thread map
+        constexpr auto m0_n_m1_to_m_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(tm0 * tm1, tm2)),
+                       make_pass_through_transform(I1)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        constexpr auto threadid_to_m0_n_m1_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(
+                make_merge_transform(make_tuple(tm0 * tm1, tn0 * tn1 * tn2 * tn3 * tn4, tm2))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+        const auto threadid_to_m_n_thread_cluster_adaptor =
+            chain_tensor_adaptors(m0_n_m1_to_m_n_adaptor, threadid_to_m0_n_m1_adaptor);
+
+        // get acc0 2D thread cluster & 2D thread slice
+        constexpr auto thread_cluster_desc_m_n = make_naive_tensor_descriptor_packed(
+            make_tuple(tm0 * tm1 * tm2, tn0 * tn1 * tn2 * tn3 * tn4));
+        constexpr auto thread_slice_desc_m_n =
+            make_naive_tensor_descriptor_packed(make_tuple(m0 * m1 * m2, n0 * n1 * n2 * n3 * n4));
+
+        auto blockwise_softmax = BlockwiseSoftmax<BlockSize,
+                                                  FloatGemmAcc,
+                                                  decltype(threadid_to_m_n_thread_cluster_adaptor),
+                                                  decltype(thread_cluster_desc_m_n),
+                                                  decltype(thread_slice_desc_m_n)>{};
+
+        const index_t num_gemm1_k_block_outer_loop =
+            b_grid_desc_bk0_n_bk1.GetLength(I1) / NPerBlock;
+        constexpr index_t num_gemm1_k_block_inner_loop = NPerBlock / Gemm1KPerBlock;
+
+        // Initialize C
+        StaticBuffer<AddressSpaceEnum::Vgpr, FloatGemmAcc, acc1_thread_buf.Size(), true>
+            c_thread_buf;
+        c_thread_buf.Clear();
+
+        // Initialize running sum and max of exponentiating row vectors
+        using SoftmaxBuf = typename decltype(blockwise_softmax)::BufferType;
+        SoftmaxBuf running_sum, running_sum_new, running_max, running_max_new;
+        running_sum     = 0;
+        running_sum_new = 0;
+        running_max     = NumericLimits<FloatGemmAcc>::Lowest();
+        running_max_new = NumericLimits<FloatGemmAcc>::Lowest();
+
+        // gemm1 K loop
+        index_t gemm1_k_block_outer_index = 0;
+        do
+        {
+            auto n_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(gemm1_k_block_outer_index * NPerBlock);
+            if(c0_matrix_mask.IsTileSkippable(
+                   m_block_data_idx_on_grid, n_block_data_idx_on_grid, MPerBlock, NPerBlock))
+            {
+                continue;
+            }
+            // gemm0
+            gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                                   a_block_desc_ak0_m_ak1,
+                                                                   a_blockwise_copy,
+                                                                   a_grid_buf,
+                                                                   a_block_buf,
+                                                                   a_block_slice_copy_step,
+                                                                   b_grid_desc_bk0_n_bk1,
+                                                                   b_block_desc_bk0_n_bk1,
+                                                                   b_blockwise_copy,
+                                                                   b_grid_buf,
+                                                                   b_block_buf,
+                                                                   b_block_slice_copy_step,
+                                                                   blockwise_gemm,
+                                                                   acc_thread_buf,
+                                                                   num_k_block_main_loop);
+            // multiple d
+            if constexpr(NumD0Tensor)
+            {
+                static_assert(NXdlPerWave == n0);
+                static_assert(MXdlPerWave == m0);
+
+                static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                    d0s_threadwise_copy(i).Run(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                                               d0s_grid_buf[i],
+                                               d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
+                                               make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
+                                               d0s_thread_buf(i));
+                });
+                static_for<0, m0 * n0 * n2 * n4, 1>{}([&](auto i) {
+                    // get reference to src data
+                    const auto src_data_refs = generate_tie(
+                        // return type should be lvalue
+                        [&](auto iSrc) -> const auto& { return d0s_thread_buf[iSrc][i]; },
+                        Number<NumD0Tensor>{});
+
+                    // get reference to dst data
+                    auto dst_data_refs = generate_tie(
+                        // return type should be lvalue
+                        [&](auto) -> auto& { return acc_thread_buf(i); },
+                        Number<2>{});
+
+                    unpack2(c0de_element_op, dst_data_refs, src_data_refs);
+                });
+                static_for<0, NumD0Tensor, 1>{}([&](auto i) {
+                    d0s_threadwise_copy(i).MoveSrcSliceWindow(
+                        d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
+                        make_multi_index(0, 1, 0, 0, 0, 0, 0, 0, 0, 0));
+                });
+            }
+            else
+            {
+                static_for<0, acc_thread_buf.Size(), 1>{}(
+                    [&](auto i) { c0de_element_op(acc_thread_buf(i), acc_thread_buf[i]); });
+            }
+
+            // do MNK padding or upper triangular masking
+            if constexpr(MaskOutUpperTriangle || PadN)
+            {
+                // 8d thread_desc in thread scope
+                constexpr auto c_thread_lengths =
+                    blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+
+                // 8d block_desc in block scope
+                constexpr auto c_block_lengths =
+                    blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+
+                constexpr auto M0 = c_block_lengths[I0];
+                constexpr auto N0 = c_block_lengths[I1];
+                constexpr auto M1 = c_block_lengths[I2];
+                constexpr auto N1 = c_block_lengths[I3];
+                constexpr auto M2 = c_block_lengths[I4];
+                constexpr auto N2 = c_block_lengths[I5];
+                constexpr auto N3 = c_block_lengths[I6];
+                constexpr auto N4 = c_block_lengths[I7];
+
+                // works like multi-dimension static_for (static_ford), but provides both the linear
+                // index as well as n-d index
+                using Acc0TileIterator = SpaceFillingCurve<
+                    decltype(to_sequence(c_thread_lengths)),
+                    typename arithmetic_sequence_gen<0, c_thread_lengths.Size(), 1>::type,
+                    typename uniform_sequence_gen<c_thread_lengths.Size(), 1>::type,
+                    false>; // SnakeCurved
+
+                auto acc0_thread_origin = blockwise_gemm.CalculateCThreadOriginDataIndex8D(
+                    Number<0>{}, Number<0>{}, Number<0>{}, Number<0>{});
+
+                constexpr auto block_idx_to_m_n_adaptor = make_single_stage_tensor_adaptor(
+                    make_tuple(make_unmerge_transform(make_tuple(M0, M1, M2)),
+                               make_unmerge_transform(make_tuple(N0, N1, N2, N3, N4))),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5, 6, 7>{}));
+
+                static_for<0, Acc0TileIterator::GetNumOfAccess(), 1>{}([&](auto i) {
+                    auto acc0_thread_idx = Acc0TileIterator::GetIndex(i) + acc0_thread_origin;
+                    auto m_local =
+                        block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I0];
+                    auto n_local =
+                        block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I1];
+                    auto m_global = m_local + m_block_data_idx_on_grid;
+                    auto n_global = n_local + n_block_data_idx_on_grid;
+                    if(c0_matrix_mask.IsMaskedElement(m_global, n_global))
+                    {
+                        acc_thread_buf(i) = -ck::NumericLimits<float>::Infinity();
+                    }
+                });
+            }
+
+            block_sync_lds(); // wait for lds read in gemm0 blockwise gemm
+
+            // softmax
+            SoftmaxBuf& max = blockwise_softmax.max_value_buf;
+            SoftmaxBuf& sum = blockwise_softmax.sum_value_buf;
+
+            blockwise_softmax.Run(acc_thread_buf, workspace_buf);
+
+            // TODO: may convert to log domain
+            running_max_new = mathext::max(max, running_max);
+            running_sum_new = mathext::exp(running_max - running_max_new) * running_sum +
+                              mathext::exp(max - running_max_new) * sum;
+
+            // gemm1
+            {
+                // TODO: explore using dynamic buffer for a1 thread buffer
+                // For a1_blockwise_copy, the goal is to satisfy pipeline requirements RunRead(),
+                // RunWrite(), and MoveSliceWindow(). But it is impossible to implement given that
+                // the A1 source buffer is static buffer holding the output of first GEMM and
+                // requires constexpr offset by design. Therefore, we pass tensor coordinate offset
+                // explicitly in Run() below.
+
+                // Initialize acc1
+                acc1_thread_buf.Clear();
+
+                // preload data into LDS
+                b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                     b1_block_slice_copy_step);
+
+                block_sync_lds(); // wait for reduction LDS read
+
+                b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+
+                // main body
+                if constexpr(num_gemm1_k_block_inner_loop > 1)
+                {
+                    static_for<0, num_gemm1_k_block_inner_loop - 1, 1>{}([&](auto i) {
+                        a1_blockwise_copy.Run(acc_thread_desc_k0_m_k1,
+                                              make_tuple(Number<i * A1ThreadSliceK0>{}, I0, I0),
+                                              acc_thread_buf,
+                                              a1_thread_desc_k0_m_k1,
+                                              make_tuple(I0, I0, I0),
+                                              a1_thread_buf);
+
+                        b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                        block_sync_lds();
+
+                        gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, acc1_thread_buf);
+
+                        block_sync_lds();
+
+                        b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                             b1_block_slice_copy_step);
+
+                        b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+                    });
+                }
+                // tail
+                {
+                    a1_blockwise_copy.Run(
+                        acc_thread_desc_k0_m_k1,
+                        make_tuple(
+                            Number<(num_gemm1_k_block_inner_loop - 1) * A1ThreadSliceK0>{}, I0, I0),
+                        acc_thread_buf,
+                        a1_thread_desc_k0_m_k1,
+                        make_tuple(I0, I0, I0),
+                        a1_thread_buf);
+
+                    block_sync_lds();
+
+                    gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, acc1_thread_buf);
+                }
+            } // end gemm1
+
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+                gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+            constexpr auto cm0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+            constexpr auto cn0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+            constexpr auto cm1 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+            constexpr auto cn1 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+            constexpr auto cm2 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+            constexpr auto cn2 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+            constexpr auto cn3 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+            constexpr auto cn4 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+            constexpr auto c_thread_slice_desc_m_n = make_naive_tensor_descriptor_packed(
+                make_tuple(cm0 * cm1 * cm2, cn0 * cn1 * cn2 * cn3 * cn4));
+            constexpr auto c_thread_buf_slice_m = c_thread_slice_desc_m_n.GetLength(I0);
+            constexpr auto c_thread_buf_slice_n = c_thread_slice_desc_m_n.GetLength(I1);
+
+            static_for<0, c_thread_buf_slice_m, 1>{}([&](auto iM) {
+                static_for<0, c_thread_buf_slice_n, 1>{}([&](auto iN) {
+                    auto I = Number<c_thread_slice_desc_m_n.CalculateOffset(make_tuple(iM, iN))>{};
+                    FloatGemmAcc acc1 = acc1_thread_buf[I]; // P*V
+                    FloatGemmAcc c    = c_thread_buf[I];    // O
+                    FloatGemmAcc c_new =
+                        (running_sum[iM] * math::exp(running_max[iM] - running_max_new[iM]) * c +
+                         math::exp(max[iM] - running_max_new[iM]) * acc1) /
+                        running_sum_new[iM]; // Formula by Dao et al.,
+                                             // https://arxiv.org/pdf/2205.14135v2.pdf section 3.1
+
+                    c_thread_buf(I) = c_new; // O_new
+                });
+            });
+
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_ak0_m_ak1,
+                                                a_block_reset_copy_step); // rewind K
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_bk0_n_bk1,
+                                                b_block_reset_copy_step); // rewind K and step N
+
+            // update before next j iteration
+            running_max = running_max_new;
+            running_sum = running_sum_new;
+
+            block_sync_lds(); // wait for gemm1 LDS read
+        } while(++gemm1_k_block_outer_index < num_gemm1_k_block_outer_loop); // end j loop
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              Gemm1NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+                gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp =
+                gemm1_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I4);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I5);
+            constexpr auto N3 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I6);
+            constexpr auto N4 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2)),                                    // M2 = MPerXdl
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2,                                      // N2 * N3 * N4 = NPerXdl
+                        N3,
+                        N4))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4>{}, Sequence<>{}, Sequence<1, 3, 5, 6, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                gemm1_blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2, N3, N4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+                                                   tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            I1,
+                                                            N2,
+                                                            I1,
+                                                            N4>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     n_thread_data_on_block_idx[I2],
+                                     n_thread_data_on_block_idx[I3],
+                                     n_thread_data_on_block_idx[I4]),
+                    tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                C1DEElementwiseOperation,   // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c1de_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, Gemm1NXdlPerWave, 1, 1, 1, N2, 1, N4>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           1,
+                                           N2,
+                                           1,
+                                           N4>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, Gemm1NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
new file mode 100644
index 0000000000..e28b74b2a9
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
@@ -0,0 +1,1128 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_softmax.hpp"
+
+namespace ck {
+
+/**
+ * @brief Gridwise gemm + softmax + gemm fusion
+ *
+ */
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename B1ElementwiseOperation,
+          typename CElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename B1GridDesc_BK0_N_BK1,
+          typename CGridDesc_M_N,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t Gemm1NPerBlock,
+          index_t Gemm1KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t B1K1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          index_t Gemm1NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun, // ignored
+          index_t BBlockLdsExtraN,
+          typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename B1BlockTransferThreadClusterArrangeOrder,
+          typename B1BlockTransferSrcAccessOrder,
+          index_t B1BlockTransferSrcVectorDim,
+          index_t B1BlockTransferSrcScalarPerVector,
+          index_t B1BlockTransferDstScalarPerVector_BK1,
+          bool B1ThreadTransferSrcResetCoordinateAfterRun,
+          index_t B1BlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          bool PadN,
+          bool MaskOutUpperTriangle,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
+{
+    static_assert(LoopSched == LoopScheduler::Default,
+                  "Non-default loop scheduler is currently not supported");
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    // Gemm0
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    static constexpr auto Gemm0MWaves = MPerBlock / (MPerXdl * MXdlPerWave);
+    static constexpr auto Gemm0NWaves = NPerBlock / (NPerXdl * NXdlPerWave);
+
+    // Gemm1
+    static constexpr auto B1K0 = Number<Gemm1KPerBlock / B1K1Value>{};
+    static constexpr auto B1K1 = Number<B1K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, MWaves, MPerXdl>(
+            ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<NXdlPerWave, NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    template <typename ABlockDesc_AK0_M_AK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
+    {
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, 1, 1>(ABlockDesc_AK0_M_AK1{});
+    }
+
+    template <typename BBlockDesc_BK0_N_BK1>
+    __host__ __device__ static constexpr auto
+    MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
+    {
+        constexpr index_t Gemm1NWaves = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+        return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm1NXdlPerWave, Gemm1NWaves, NPerXdl>(
+            BBlockDesc_BK0_N_BK1{});
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B1 matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(B1K0, Number<Gemm1NPerBlock>{}, B1K1),
+            make_tuple(Number<Gemm1NPerBlock + B1BlockLdsExtraN>{} * B1K1, B1K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        const index_t gemm0_bytes_end = (SharedMemTrait::a_block_space_size_aligned +
+                                         SharedMemTrait::b_block_space_size_aligned) *
+                                        sizeof(FloatAB);
+        const index_t gemm1_bytes_end =
+            (SharedMemTrait::b1_block_space_offset + SharedMemTrait::b1_block_space_size_aligned) *
+            sizeof(FloatAB);
+        const index_t softmax_bytes_end = (SharedMemTrait::reduction_space_offset +
+                                           SharedMemTrait::reduction_space_size_aligned) *
+                                          sizeof(FloatGemmAcc);
+        const index_t c_block_bytes_end =
+            SharedMemTrait::c_block_space_size * sizeof(FloatCShuffle);
+
+        return math::max(gemm0_bytes_end, gemm1_bytes_end, softmax_bytes_end, c_block_bytes_end);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+        const auto Gemm1N = b1_grid_desc_bk0_n_bk1.GetLength(I1);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && Gemm1N == c_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0 &&
+             Gemm1N % Gemm1NPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gemm0 gridwise gemm pipeline
+        const auto num_gemm0_k_loop = K / KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm0_k_loop))
+        {
+            return false;
+        }
+
+        // check gemm1 gridwise gemm pipeline
+        if(!(NPerBlock % Gemm1KPerBlock == 0))
+        {
+            return false;
+        }
+
+        const auto num_gemm1_k_inner_loop = NPerBlock / Gemm1KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_gemm1_k_inner_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / Gemm1NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<Gemm1NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, Gemm1NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            CGridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
+
+    struct SharedMemTrait
+    {
+        // LDS allocation for A and B: be careful of alignment
+        static constexpr auto a_block_desc_ak0_m_ak1 =
+            GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        static constexpr auto b_block_desc_bk0_n_bk1 =
+            GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        static constexpr auto b1_block_desc_bk0_n_bk1 =
+            GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        static constexpr auto max_lds_align = math::lcm(math::lcm(AK1, BK1), B1K1);
+
+        static constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+        static constexpr auto b1_block_space_size_aligned = math::integer_least_multiple(
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        static constexpr auto a_block_space_offset  = 0;
+        static constexpr auto b_block_space_offset  = a_block_space_size_aligned.value;
+        static constexpr auto b1_block_space_offset = 0;
+
+        // LDS allocation for reduction
+        static constexpr index_t reduction_space_size_aligned =
+            math::integer_least_multiple(BlockSize, max_lds_align);
+
+        static constexpr auto reduction_space_offset = 0;
+
+        // LDS allocation for C shuffle in LDS
+        static constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+        static constexpr auto c_block_space_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+    };
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap, typename C0MatrixMask>
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               const FloatAB* __restrict__ p_b1_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const AccElementwiseOperation& acc_element_op,
+                               const B1ElementwiseOperation& b1_element_op,
+                               const CElementwiseOperation& c_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
+                               const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2CTileMap& block_2_ctile_map,
+                               const C0MatrixMask& c0_matrix_mask)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        const auto b1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b1_grid, b1_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/gemm1_n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t gemm1_n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * Gemm1NPerBlock);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        //
+        // set up Gemm0
+        //
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                true, // SrcResetCoord
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0), // will loop over GemmN dimension
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        // Fused Gemm+Gemm pipeline
+        // for n in N0:
+        //   for k in K0:
+        //     acc[m][n] += A[m][k] * B0[k][n]
+        //   acc1[m][o] += acc[m][n] * B1[n][o]
+
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            decltype(MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(a_block_desc_ak0_m_ak1)),
+            decltype(MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(b_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            NPerBlock,
+            KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            true>{}; // TransposeC
+
+        auto acc_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::a_block_space_offset,
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::b_block_space_offset,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+        const auto a_block_reset_copy_step =
+            make_multi_index(-a_grid_desc_ak0_m_ak1.GetLength(I0), 0, 0);
+        const auto b_block_reset_copy_step =
+            make_multi_index(-b_grid_desc_bk0_n_bk1.GetLength(I0), NPerBlock, 0);
+
+        // gridwise GEMM pipeline
+        // Only supports LoopScheduler::Default
+        const auto gridwise_gemm_pipeline = GridwiseGemmPipeline_Selector<PipelineVer,
+                                                                          NumGemmKPrefetchStage,
+                                                                          LoopScheduler::Default>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        //
+        // set up Gemm1
+        //
+
+        // Acc matrix threadwise copy: AccVGPR to VGPR and downcast to XDL input data type
+        constexpr auto acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+        constexpr auto m0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+        constexpr auto n0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+        constexpr auto m1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+        constexpr auto n1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+        constexpr auto m2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+        constexpr auto n2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+        constexpr auto n3 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+        constexpr auto n4 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+
+        constexpr auto b1_block_slice_copy_step = make_multi_index(Gemm1KPerBlock / B1K1, 0, 0);
+
+        // acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 to acc_thread_desc_k0_m_k1
+        // n0_n1_n2_n3 -> k0
+        // m0_m1_m2 -> m
+        // n4 -> k1
+        // NOTE: had to use merge_v3 or will spit out compilation errors
+        constexpr auto acc_thread_desc_k0_m_k1 = transform_tensor_descriptor(
+            acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+            make_tuple(make_merge_transform_v3_division_mod(make_tuple(n0, n1, n2, n3)),
+                       make_merge_transform_v3_division_mod(make_tuple(m0, m1, m2)),
+                       make_pass_through_transform(n4)),
+            make_tuple(Sequence<1, 3, 5, 6>{}, Sequence<0, 2, 4>{}, Sequence<7>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        // A1 matrix in AccVGPR
+        // N2 num_groups_per_blk, N3 num_input_blks, N4 group_size
+        constexpr auto AccN3 =
+            blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLength(I6);
+
+        constexpr auto A1ThreadSlice_K0_M_K1 =
+            make_tuple(Number<Gemm1KPerBlock / n4 / AccN3>{}, Number<m0 * m1 * m2>{}, Number<n4>{});
+
+        constexpr auto A1ThreadSliceK0        = A1ThreadSlice_K0_M_K1[I0];
+        constexpr auto A1ThreadSliceM         = A1ThreadSlice_K0_M_K1[I1];
+        constexpr auto A1ThreadSliceK1        = A1ThreadSlice_K0_M_K1[I2];
+        constexpr auto a1_thread_desc_k0_m_k1 = make_naive_tensor_descriptor(
+            A1ThreadSlice_K0_M_K1,
+            make_tuple(A1ThreadSliceM * A1ThreadSliceK1, A1ThreadSliceK1, I1));
+
+        // B1 matrix in LDS memory, dst of blockwise copy
+        constexpr auto b1_block_desc_bk0_n_bk1 = GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A1 matrix blockwise copy
+        auto a1_blockwise_copy = ThreadwiseTensorSliceTransfer_StaticToStatic<
+            FloatGemmAcc,
+            FloatAB,
+            decltype(acc_thread_desc_k0_m_k1),
+            decltype(a1_thread_desc_k0_m_k1),
+            tensor_operation::element_wise::PassThrough,
+            Sequence<A1ThreadSliceK0, A1ThreadSliceM, A1ThreadSliceK1>,
+            Sequence<1, 0, 2>,
+            2,
+            n4>{tensor_operation::element_wise::PassThrough{}};
+
+        // B1 matrix blockwise copy
+        auto b1_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<B1K0, Gemm1NPerBlock, B1K1>,
+                                                B1BlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                B1BlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b1_grid_desc_bk0_n_bk1),
+                                                decltype(b1_block_desc_bk0_n_bk1),
+                                                B1BlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                B1BlockTransferSrcVectorDim,
+                                                2,
+                                                B1BlockTransferSrcScalarPerVector,
+                                                B1BlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                B1ThreadTransferSrcResetCoordinateAfterRun,
+                                                true, // DstResetCoord
+                                                NumGemmKPrefetchStage>(
+                b1_grid_desc_bk0_n_bk1,
+                make_multi_index(0, gemm1_n_block_data_idx_on_grid, 0),
+                b1_element_op,
+                b1_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                tensor_operation::element_wise::PassThrough{});
+
+        auto a1_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
+            a1_thread_desc_k0_m_k1.GetElementSpaceSize());
+
+        // reuse LDS space for gemm0's b_block_buf
+        auto b1_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + SharedMemTrait::b1_block_space_offset,
+            b1_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        // selected_mfma.group_size or B1K1 <= Gemm1KPack <= selected_mfma.group_size
+        // selected_mfma.k_per_blk <= Gemm1KPack
+        //
+        // Following similar rationale behind Gemm0KPack, let Gemm1KPack be the lowest common
+        // multiples of A1K1 (predetermined by selected_mfma.group_size) and B1K1. But in this case
+        // Gemm1KPack can't be higher than A1K1 itself because A1 matrix is distributed in VGPRs
+        // with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
+        // cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
+        // therefore we may just as well assign Gemm1KPack = group_size
+        constexpr index_t Gemm1KPack =
+            MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size;
+
+        auto gemm1_blockwise_gemm = BlockwiseGemmXdlops_v2<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a1_thread_desc_k0_m_k1),
+            decltype(b1_block_desc_bk0_n_bk1),
+            decltype(MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(a1_thread_desc_k0_m_k1)),
+            decltype(MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(b1_block_desc_bk0_n_bk1)),
+            MPerBlock,
+            Gemm1NPerBlock,
+            Gemm1KPerBlock,
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            Gemm1NXdlPerWave,
+            Gemm1KPack,
+            true,       // TransposeC
+            Gemm1KPack, // AMmaKStride
+            Gemm1KPack * XdlopsGemm<FloatAB, MPerXdl, NPerXdl, Gemm1KPack, false>{}.K0PerXdlops>{
+            // BMmaKStride
+            make_multi_index(0, 0, 0, 0)}; // A_origin
+
+        auto acc1_thread_buf = gemm1_blockwise_gemm.GetCThreadBuffer();
+
+        //
+        // Blockwise softmax
+        //
+        auto workspace_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatGemmAcc*>(p_shared) + SharedMemTrait::reduction_space_offset,
+            SharedMemTrait::reduction_space_size_aligned);
+
+        // get acc0 8D thread cluster
+        constexpr auto thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4 =
+            blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths() /
+            blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+        constexpr auto tm0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I0);
+        constexpr auto tn0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I1);
+        constexpr auto tm1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I2);
+        constexpr auto tn1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I3);
+        constexpr auto tm2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I4);
+        constexpr auto tn2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I5);
+        constexpr auto tn3 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I6);
+        constexpr auto tn4 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I7);
+
+        // get acc0 thread map
+        constexpr auto m0_n_m1_to_m_n_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_unmerge_transform(make_tuple(tm0 * tm1, tm2)),
+                       make_pass_through_transform(I1)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        constexpr auto threadid_to_m0_n_m1_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(
+                make_merge_transform(make_tuple(tm0 * tm1, tn0 * tn1 * tn2 * tn3 * tn4, tm2))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+        const auto threadid_to_m_n_thread_cluster_adaptor =
+            chain_tensor_adaptors(m0_n_m1_to_m_n_adaptor, threadid_to_m0_n_m1_adaptor);
+
+        // get acc0 2D thread cluster & 2D thread slice
+        constexpr auto thread_cluster_desc_m_n = make_naive_tensor_descriptor_packed(
+            make_tuple(tm0 * tm1 * tm2, tn0 * tn1 * tn2 * tn3 * tn4));
+        constexpr auto thread_slice_desc_m_n =
+            make_naive_tensor_descriptor_packed(make_tuple(m0 * m1 * m2, n0 * n1 * n2 * n3 * n4));
+
+        auto blockwise_softmax = BlockwiseSoftmax<BlockSize,
+                                                  FloatGemmAcc,
+                                                  decltype(threadid_to_m_n_thread_cluster_adaptor),
+                                                  decltype(thread_cluster_desc_m_n),
+                                                  decltype(thread_slice_desc_m_n)>{};
+
+        const index_t num_gemm1_k_block_outer_loop =
+            b_grid_desc_bk0_n_bk1.GetLength(I1) / NPerBlock;
+        constexpr index_t num_gemm1_k_block_inner_loop = NPerBlock / Gemm1KPerBlock;
+
+        // Initialize C
+        StaticBuffer<AddressSpaceEnum::Vgpr, FloatGemmAcc, acc1_thread_buf.Size(), true>
+            c_thread_buf;
+        c_thread_buf.Clear();
+
+        // Initialize running sum and max of exponentiating row vectors
+        using SoftmaxBuf = typename decltype(blockwise_softmax)::BufferType;
+        SoftmaxBuf running_sum, running_sum_new, running_max, running_max_new;
+        running_sum     = 0;
+        running_sum_new = 0;
+        running_max     = NumericLimits<FloatGemmAcc>::Lowest();
+        running_max_new = NumericLimits<FloatGemmAcc>::Lowest();
+
+        // gemm1 K loop
+        index_t gemm1_k_block_outer_index = 0;
+        do
+        {
+            auto n_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(gemm1_k_block_outer_index * NPerBlock);
+            if(c0_matrix_mask.IsTileSkippable(
+                   m_block_data_idx_on_grid, n_block_data_idx_on_grid, MPerBlock, NPerBlock))
+            {
+                continue;
+            }
+            // gemm0
+            gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                                   a_block_desc_ak0_m_ak1,
+                                                                   a_blockwise_copy,
+                                                                   a_grid_buf,
+                                                                   a_block_buf,
+                                                                   a_block_slice_copy_step,
+                                                                   b_grid_desc_bk0_n_bk1,
+                                                                   b_block_desc_bk0_n_bk1,
+                                                                   b_blockwise_copy,
+                                                                   b_grid_buf,
+                                                                   b_block_buf,
+                                                                   b_block_slice_copy_step,
+                                                                   blockwise_gemm,
+                                                                   acc_thread_buf,
+                                                                   num_k_block_main_loop);
+
+            // do MNK padding or upper triangular masking
+            if constexpr(MaskOutUpperTriangle || PadN)
+            {
+                // 8d thread_desc in thread scope
+                constexpr auto c_thread_lengths =
+                    blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+
+                // 8d block_desc in block scope
+                constexpr auto c_block_lengths =
+                    blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
+
+                constexpr auto M0 = c_block_lengths[I0];
+                constexpr auto N0 = c_block_lengths[I1];
+                constexpr auto M1 = c_block_lengths[I2];
+                constexpr auto N1 = c_block_lengths[I3];
+                constexpr auto M2 = c_block_lengths[I4];
+                constexpr auto N2 = c_block_lengths[I5];
+                constexpr auto N3 = c_block_lengths[I6];
+                constexpr auto N4 = c_block_lengths[I7];
+
+                // works like multi-dimension static_for (static_ford), but provides both the linear
+                // index as well as n-d index
+                using Acc0TileIterator = SpaceFillingCurve<
+                    decltype(to_sequence(c_thread_lengths)),
+                    typename arithmetic_sequence_gen<0, c_thread_lengths.Size(), 1>::type,
+                    typename uniform_sequence_gen<c_thread_lengths.Size(), 1>::type,
+                    false>; // SnakeCurved
+
+                auto acc0_thread_origin = blockwise_gemm.CalculateCThreadOriginDataIndex8D(
+                    Number<0>{}, Number<0>{}, Number<0>{}, Number<0>{});
+
+                constexpr auto block_idx_to_m_n_adaptor = make_single_stage_tensor_adaptor(
+                    make_tuple(make_unmerge_transform(make_tuple(M0, M1, M2)),
+                               make_unmerge_transform(make_tuple(N0, N1, N2, N3, N4))),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5, 6, 7>{}));
+
+                static_for<0, Acc0TileIterator::GetNumOfAccess(), 1>{}([&](auto i) {
+                    auto acc0_thread_idx = Acc0TileIterator::GetIndex(i) + acc0_thread_origin;
+                    auto m_local =
+                        block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I0];
+                    auto n_local =
+                        block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I1];
+                    auto m_global = m_local + m_block_data_idx_on_grid;
+                    auto n_global = n_local + n_block_data_idx_on_grid;
+                    if(c0_matrix_mask.IsMaskedElement(m_global, n_global))
+                    {
+                        acc_thread_buf(i) = -ck::NumericLimits<float>::Infinity();
+                    }
+                    else
+                    {
+                        acc_element_op(acc_thread_buf(i), acc_thread_buf[i]);
+                    }
+                });
+            }
+            else
+            {
+                static_for<0, acc_thread_buf.Size(), 1>{}(
+                    [&](auto i) { acc_element_op(acc_thread_buf(i), acc_thread_buf[i]); });
+            }
+
+            block_sync_lds(); // wait for lds read in gemm0 blockwise gemm
+
+            // softmax
+            SoftmaxBuf& max = blockwise_softmax.max_value_buf;
+            SoftmaxBuf& sum = blockwise_softmax.sum_value_buf;
+
+            blockwise_softmax.Run(acc_thread_buf, workspace_buf);
+
+            // TODO: may convert to log domain
+            running_max_new = mathext::max(max, running_max);
+            running_sum_new = mathext::exp(running_max - running_max_new) * running_sum +
+                              mathext::exp(max - running_max_new) * sum;
+
+            // gemm1
+            {
+                // TODO: explore using dynamic buffer for a1 thread buffer
+                // For a1_blockwise_copy, the goal is to satisfy pipeline requirements RunRead(),
+                // RunWrite(), and MoveSliceWindow(). But it is impossible to implement given that
+                // the A1 source buffer is static buffer holding the output of first GEMM and
+                // requires constexpr offset by design. Therefore, we pass tensor coordinate offset
+                // explicitly in Run() below.
+
+                // Initialize acc1
+                acc1_thread_buf.Clear();
+
+                // preload data into LDS
+                b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                     b1_block_slice_copy_step);
+
+                block_sync_lds(); // wait for reduction LDS read
+
+                b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+
+                // main body
+                if constexpr(num_gemm1_k_block_inner_loop > 1)
+                {
+                    static_for<0, num_gemm1_k_block_inner_loop - 1, 1>{}([&](auto i) {
+                        a1_blockwise_copy.Run(acc_thread_desc_k0_m_k1,
+                                              make_tuple(Number<i * A1ThreadSliceK0>{}, I0, I0),
+                                              acc_thread_buf,
+                                              a1_thread_desc_k0_m_k1,
+                                              make_tuple(I0, I0, I0),
+                                              a1_thread_buf);
+
+                        b1_blockwise_copy.RunRead(b1_grid_desc_bk0_n_bk1, b1_grid_buf);
+
+                        block_sync_lds();
+
+                        gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, acc1_thread_buf);
+
+                        block_sync_lds();
+
+                        b1_blockwise_copy.MoveSrcSliceWindow(b1_grid_desc_bk0_n_bk1,
+                                                             b1_block_slice_copy_step);
+
+                        b1_blockwise_copy.RunWrite(b1_block_desc_bk0_n_bk1, b1_block_buf);
+                    });
+                }
+                // tail
+                {
+                    a1_blockwise_copy.Run(
+                        acc_thread_desc_k0_m_k1,
+                        make_tuple(
+                            Number<(num_gemm1_k_block_inner_loop - 1) * A1ThreadSliceK0>{}, I0, I0),
+                        acc_thread_buf,
+                        a1_thread_desc_k0_m_k1,
+                        make_tuple(I0, I0, I0),
+                        a1_thread_buf);
+
+                    block_sync_lds();
+
+                    gemm1_blockwise_gemm.Run(a1_thread_buf, b1_block_buf, acc1_thread_buf);
+                }
+            } // end gemm1
+
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+                gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+            constexpr auto cm0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
+            constexpr auto cn0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
+            constexpr auto cm1 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
+            constexpr auto cn1 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
+            constexpr auto cm2 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
+            constexpr auto cn2 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
+            constexpr auto cn3 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
+            constexpr auto cn4 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
+            constexpr auto c_thread_slice_desc_m_n = make_naive_tensor_descriptor_packed(
+                make_tuple(cm0 * cm1 * cm2, cn0 * cn1 * cn2 * cn3 * cn4));
+            constexpr auto c_thread_buf_slice_m = c_thread_slice_desc_m_n.GetLength(I0);
+            constexpr auto c_thread_buf_slice_n = c_thread_slice_desc_m_n.GetLength(I1);
+
+            static_for<0, c_thread_buf_slice_m, 1>{}([&](auto iM) {
+                static_for<0, c_thread_buf_slice_n, 1>{}([&](auto iN) {
+                    auto I = Number<c_thread_slice_desc_m_n.CalculateOffset(make_tuple(iM, iN))>{};
+                    FloatGemmAcc acc1 = acc1_thread_buf[I]; // P*V
+                    FloatGemmAcc c    = c_thread_buf[I];    // O
+                    FloatGemmAcc c_new =
+                        (running_sum[iM] * math::exp(running_max[iM] - running_max_new[iM]) * c +
+                         math::exp(max[iM] - running_max_new[iM]) * acc1) /
+                        running_sum_new[iM]; // Formula by Dao et al.,
+                                             // https://arxiv.org/pdf/2205.14135v2.pdf section 3.1
+
+                    c_thread_buf(I) = c_new; // O_new
+                });
+            });
+
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_ak0_m_ak1,
+                                                a_block_reset_copy_step); // rewind K
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_bk0_n_bk1,
+                                                b_block_reset_copy_step); // rewind K and step N
+
+            // update before next j iteration
+            running_max = running_max_new;
+            running_sum = running_sum_new;
+
+            block_sync_lds(); // wait for gemm1 LDS read
+        } while(++gemm1_k_block_outer_index < num_gemm1_k_block_outer_loop); // end j loop
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              Gemm1NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
+                gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp =
+                gemm1_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I4);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I5);
+            constexpr auto N3 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I6);
+            constexpr auto N4 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2)),                                    // M2 = MPerXdl
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2,                                      // N2 * N3 * N4 = NPerXdl
+                        N3,
+                        N4))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4>{}, Sequence<>{}, Sequence<1, 3, 5, 6, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                gemm1_blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2, N3, N4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4),
+                                                   tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            I1,
+                                                            N2,
+                                                            I1,
+                                                            N4>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     n_thread_data_on_block_idx[I2],
+                                     n_thread_data_on_block_idx[I3],
+                                     n_thread_data_on_block_idx[I4]),
+                    tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, Gemm1NXdlPerWave, 1, 1, 1, N2, 1, N4>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           1,
+                                           N2,
+                                           1,
+                                           N4>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, Gemm1NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp
new file mode 100644
index 0000000000..ed1ffdd857
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp
@@ -0,0 +1,554 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math_v2.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseBatchrNormBackwardWithBlockwiseWelford_,
+          typename XDataType,
+          typename DyDataType,
+          typename DxDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename ScaleBiasGridDesc_M,
+          typename MeanVarGridDesc_M,
+          typename GetReduceCountPerThreadFunctor>
+__global__ void kernel_batchnorm_backward_with_blockwise_welford(
+    const XYGridDesc_M_K x_grid_desc_m_k,
+    const XYGridDesc_M_K dy_grid_desc_m_k,
+    const XYGridDesc_M_K dx_grid_desc_m_k,
+    const ScaleBiasGridDesc_M scale_grid_desc_m,
+    const ScaleBiasGridDesc_M dscale_dbias_grid_desc_m,
+    const MeanVarGridDesc_M mean_var_grid_desc_m,
+    const GetReduceCountPerThreadFunctor get_reduce_count_per_thread,
+    long_index_t reduce_size,
+    index_t num_k_block_tile_iteration,
+    AccDataType epsilon,
+    const XDataType* const __restrict__ p_x,
+    const DyDataType* const __restrict__ p_dy,
+    const ScaleDataType* const __restrict__ p_scale,
+    bool haveSavedMeanInvVar,
+    const MeanVarDataType* const __restrict__ p_savedMean,
+    const MeanVarDataType* const __restrict__ p_savedInvVar,
+    const DyElementwiseOp dy_elementwise_op,
+    DxDataType* const __restrict__ p_dx,
+    DscaleDbiasDataType* const __restrict__ p_dscale,
+    DscaleDbiasDataType* const __restrict__ p_dbias)
+{
+    GridwiseBatchrNormBackwardWithBlockwiseWelford_::Run(x_grid_desc_m_k,
+                                                         dy_grid_desc_m_k,
+                                                         dx_grid_desc_m_k,
+                                                         scale_grid_desc_m,
+                                                         dscale_dbias_grid_desc_m,
+                                                         mean_var_grid_desc_m,
+                                                         get_reduce_count_per_thread,
+                                                         reduce_size,
+                                                         num_k_block_tile_iteration,
+                                                         epsilon,
+                                                         p_x,
+                                                         p_dy,
+                                                         p_scale,
+                                                         haveSavedMeanInvVar,
+                                                         p_savedMean,
+                                                         p_savedInvVar,
+                                                         dy_elementwise_op,
+                                                         p_dx,
+                                                         p_dscale,
+                                                         p_dbias);
+};
+
+template <typename XDataType,
+          typename DyDataType,
+          typename DxDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename ScaleBiasGridDesc_M,
+          typename MeanVarGridDesc_M,
+          typename GetReduceCountPerThreadFunctor,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XDyDxVectorDim,
+          index_t XSrcVectorSize,
+          index_t DySrcVectorSize,
+          index_t DxDstVectorSize,
+          index_t ScaleSrcVectorSize,
+          index_t DscaleDbiasDstVectorSize,
+          index_t MeanVarSrcVectorSize>
+struct GridwiseBatchNormBackwardWithBlockwiseWelford
+{
+    static_assert((XDyDxVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0 &&
+                   MThreadSliceSize % DySrcVectorSize == 0 &&
+                   MThreadSliceSize % DxDstVectorSize == 0) ||
+                      (XDyDxVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0 &&
+                       KThreadSliceSize % DySrcVectorSize == 0 &&
+                       KThreadSliceSize % DxDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XDyDxVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder>;
+
+    using BlockwiseReduce = PartitionedBlockwiseReduction<AccDataType,
+                                                          BlockSize,
+                                                          ThreadClusterLengths_M_K,
+                                                          ThreadClusterArrangeOrder,
+                                                          ck::reduce::Add,
+                                                          false>;
+
+    using ThreadwiseReduce = ThreadwiseReduction<AccDataType,
+                                                 ThreadReduceSrcDesc_M_K,
+                                                 ThreadReduceDstDesc_M,
+                                                 ck::reduce::Add,
+                                                 false>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    // clang-format off
+    // Blockwise BatchNorm Backward
+    // Input: x, dy, scale, savedMean and savedInvVar (optional), reduce_size
+    // Output: dx, dscale, dbias
+    // Step 1: calculating mean and inv-variance using welford method (if savedMean/savedInvVar not available), where inv-variance = 1/sqrt(epsilon+variance)
+    // Step 2: reduction: dbias = sum(dy),  dscale = sum(dy *(x-mean) * inv-variance)
+    // Step 3: calculating dx = 1/reduce_size * inv-variance * scale * (reduce_size * dy - dbias - dscale * (x - mean) * inv-variance)) elementwise-ly
+    // clang-format on
+    __device__ static void Run(const XYGridDesc_M_K x_grid_desc_m_k,
+                               const XYGridDesc_M_K dy_grid_desc_m_k,
+                               const XYGridDesc_M_K dx_grid_desc_m_k,
+                               const ScaleBiasGridDesc_M scale_grid_desc_m,
+                               const ScaleBiasGridDesc_M dscale_dbias_grid_desc_m,
+                               const MeanVarGridDesc_M mean_var_grid_desc_m,
+                               const GetReduceCountPerThreadFunctor get_reduce_count_per_thread,
+                               long_index_t reduce_size,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType epsilon,
+                               const XDataType* const __restrict__ p_x,
+                               const DyDataType* const __restrict__ p_dy,
+                               const ScaleDataType* const __restrict__ p_scale,
+                               bool haveSavedMeanInvVar,
+                               const MeanVarDataType* const __restrict__ p_savedMean,
+                               const MeanVarDataType* const __restrict__ p_savedInvVar,
+                               const DyElementwiseOp dy_elementwise_op,
+                               DxDataType* const __restrict__ p_dx,
+                               DscaleDbiasDataType* const __restrict__ p_dscale,
+                               DscaleDbiasDataType* const __restrict__ p_dbias)
+    {
+        using ck::math::sqrt;
+
+        __shared__ AccDataType p_reduce_work_buffer[BlockSize];
+
+        auto reduce_work_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            x_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            dy_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            dx_thread_buf;
+
+        // buffer of values of dy * (x-mean) * invVariance, used as input of Blockwise reduction
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            tmp1_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> scale_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>&
+            inv_var_thread_buf = var_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> dscale_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> dbias_thread_buf;
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        using ThreadBufferLengths_M           = Sequence<MThreadSliceSize>;
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  XYGridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XDyDxVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize,
+                             thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_dy_load = ThreadwiseTensorSliceTransfer_v2<DyDataType,
+                                                                   AccDataType,
+                                                                   XYGridDesc_M_K,
+                                                                   decltype(thread_buffer_desc_m_k),
+                                                                   ThreadBufferLengths_M_K,
+                                                                   ThreadBufferDimAccessOrder,
+                                                                   XDyDxVectorDim,
+                                                                   XSrcVectorSize,
+                                                                   1,
+                                                                   true>(
+            dy_grid_desc_m_k,
+            make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize,
+                             thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_dx_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               DxDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               XYGridDesc_M_K,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               XDyDxVectorDim,
+                                               DxDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                dx_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * KThreadSliceSize),
+                PassThroughOp{});
+
+        auto threadwise_scale_load =
+            ThreadwiseTensorSliceTransfer_v2<ScaleDataType,
+                                             AccDataType,
+                                             ScaleBiasGridDesc_M,
+                                             decltype(thread_buffer_desc_m),
+                                             ThreadBufferLengths_M,
+                                             Sequence<0>,
+                                             0,
+                                             ScaleSrcVectorSize,
+                                             1,
+                                             true>(
+                scale_grid_desc_m,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize));
+
+        auto threadwise_dscale_dbias_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               DscaleDbiasDataType,
+                                               decltype(thread_buffer_desc_m),
+                                               ScaleBiasGridDesc_M,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M,
+                                               Sequence<0>,
+                                               0,
+                                               DscaleDbiasDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                dscale_dbias_grid_desc_m,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize),
+                PassThroughOp{});
+
+        constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileSize);
+        constexpr auto thread_copy_bwd_step_m_k = make_multi_index(0, -K_BlockTileSize);
+
+        const auto x_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto dy_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dy, dy_grid_desc_m_k.GetElementSpaceSize());
+
+        auto dx_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dx, dx_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto scale_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_scale, scale_grid_desc_m.GetElementSpaceSize());
+
+        auto dscale_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dscale, dscale_dbias_grid_desc_m.GetElementSpaceSize());
+
+        auto dbias_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_dbias, dscale_dbias_grid_desc_m.GetElementSpaceSize());
+
+        // clang-format off
+        // Step 1: calculating mean and inv-variance using welford method (if savedMean/savedInvVar not available), where inv-variance = 1/sqrt(epsilon+variance)
+        // clang-format on
+
+        if(haveSavedMeanInvVar)
+        {
+            const auto mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_savedMean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            const auto inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_savedInvVar, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto threadwise_mean_inv_var_load =
+                ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
+                                                 AccDataType,
+                                                 MeanVarGridDesc_M,
+                                                 decltype(thread_buffer_desc_m),
+                                                 ThreadBufferLengths_M,
+                                                 Sequence<0>,
+                                                 0,
+                                                 MeanVarSrcVectorSize,
+                                                 1,
+                                                 true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize));
+
+            threadwise_mean_inv_var_load.Run(mean_var_grid_desc_m,
+                                             mean_global_buf,
+                                             thread_buffer_desc_m,
+                                             make_tuple(I0),
+                                             mean_thread_buf);
+
+            threadwise_mean_inv_var_load.Run(mean_var_grid_desc_m,
+                                             inv_var_global_buf,
+                                             thread_buffer_desc_m,
+                                             make_tuple(I0),
+                                             inv_var_thread_buf);
+        }
+        else
+        {
+            auto threadwise_welford       = ThreadwiseWelford();
+            threadwise_welford.max_count_ = get_reduce_count_per_thread(thread_k_cluster_id);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
+                var_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+            });
+
+            for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+            {
+
+                threadwise_x_load.Run(x_grid_desc_m_k,
+                                      x_global_buf,
+                                      thread_buffer_desc_m_k,
+                                      make_tuple(I0, I0),
+                                      x_thread_buf);
+
+                threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                threadwise_welford.Run(x_thread_buf, mean_thread_buf, var_thread_buf);
+            }
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+
+                int count = threadwise_welford.cur_count_;
+                BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
+            });
+
+            // calculate inv-variance as 1/sqrt(epsilon+variance)
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                inv_var_thread_buf(I) =
+                    type_convert<AccDataType>(1.0) / sqrt(var_thread_buf[I] + epsilon);
+            });
+
+            threadwise_x_load.SetSrcSliceOrigin(
+                x_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * KThreadSliceSize));
+        };
+
+        // clang-format off
+        // Step 2: reduction: dbias = sum(dy),  dscale = sum(dy *(x-mean) * inv-variance)
+        // clang-format on
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            dscale_thread_buf(I) = type_convert<AccDataType>(0);
+            dbias_thread_buf(I)  = type_convert<AccDataType>(0);
+        });
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            threadwise_dy_load.Run(dx_grid_desc_m_k,
+                                   dy_global_buf,
+                                   thread_buffer_desc_m_k,
+                                   make_tuple(I0, I0),
+                                   dy_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                    dy_elementwise_op(dy_thread_buf(Number<offset>{}),
+                                      dy_thread_buf[Number<offset>{}]);
+
+                    AccDataType norm_x = (x_thread_buf[Number<offset>{}] - mean_thread_buf[iM]) *
+                                         inv_var_thread_buf[iM];
+
+                    tmp1_thread_buf(Number<offset>{}) = norm_x * dy_thread_buf[Number<offset>{}];
+                });
+            });
+
+            ThreadwiseReduce::Reduce(tmp1_thread_buf, dscale_thread_buf);
+            ThreadwiseReduce::Reduce(dy_thread_buf, dbias_thread_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+            threadwise_dy_load.MoveSrcSliceWindow(dy_grid_desc_m_k, thread_copy_fwd_step_m_k);
+        };
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+            BlockwiseReduce::Reduce(reduce_work_buf, dscale_thread_buf(I));
+            block_sync_lds();
+            BlockwiseReduce::Reduce(reduce_work_buf, dbias_thread_buf(I));
+        });
+
+        if(thread_k_cluster_id == 0)
+        {
+            threadwise_dscale_dbias_store.Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              dscale_thread_buf,
+                                              dscale_dbias_grid_desc_m,
+                                              dscale_global_buf);
+
+            threadwise_dscale_dbias_store.Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              dbias_thread_buf,
+                                              dscale_dbias_grid_desc_m,
+                                              dbias_global_buf);
+        };
+
+        // clang-format off
+        // Step 3: calculating dx = 1/reduce_size * inv-variance * scale * (reduce_size * dy - dbias - dscale * (x - mean) * inv-variance)) elementwise-ly
+        // clang-format on
+
+        threadwise_scale_load.Run(scale_grid_desc_m,
+                                  scale_global_buf,
+                                  thread_buffer_desc_m,
+                                  make_tuple(I0),
+                                  scale_thread_buf);
+
+        auto thread_copy_tail_m_k = (num_k_block_tile_iteration - 1) * thread_copy_fwd_step_m_k;
+
+        threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+        threadwise_dy_load.MoveSrcSliceWindow(dy_grid_desc_m_k, thread_copy_bwd_step_m_k);
+        threadwise_dx_store.MoveDstSliceWindow(dx_grid_desc_m_k, thread_copy_tail_m_k);
+
+        AccDataType inv_reduce_size =
+            type_convert<AccDataType>(1.0) / type_convert<AccDataType>(reduce_size);
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            threadwise_dy_load.Run(dy_grid_desc_m_k,
+                                   dy_global_buf,
+                                   thread_buffer_desc_m_k,
+                                   make_tuple(I0, I0),
+                                   dy_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                AccDataType multiplier =
+                    inv_reduce_size * inv_var_thread_buf[iM] * scale_thread_buf[iM];
+
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                    dy_elementwise_op(dy_thread_buf(Number<offset>{}),
+                                      dy_thread_buf[Number<offset>{}]);
+
+                    AccDataType norm_x = (x_thread_buf[Number<offset>{}] - mean_thread_buf[iM]) *
+                                         inv_var_thread_buf[iM];
+
+                    AccDataType tmpVal = norm_x * dscale_thread_buf[iM];
+
+                    dx_thread_buf(Number<offset>{}) =
+                        multiplier *
+                        (type_convert<AccDataType>(reduce_size) * dy_thread_buf[Number<offset>{}] -
+                         dbias_thread_buf[iM] - tmpVal);
+                });
+            });
+
+            threadwise_dx_store.Run(thread_buffer_desc_m_k,
+                                    make_tuple(I0, I0),
+                                    dx_thread_buf,
+                                    dx_grid_desc_m_k,
+                                    dx_global_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_dy_load.MoveSrcSliceWindow(dy_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_dx_store.MoveDstSliceWindow(dx_grid_desc_m_k, thread_copy_bwd_step_m_k);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp
new file mode 100644
index 0000000000..b6c83af13a
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp
@@ -0,0 +1,483 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math_v2.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseBatchrNormForwardWithBlockwiseWelford_,
+          typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename ScaleBiasGridDesc_M,
+          typename MeanVarGridDesc_M,
+          typename GetReduceCountPerThreadFunctor>
+__global__ void kernel_batchnorm_forward_with_blockwise_welford(
+    const XYGridDesc_M_K x_grid_desc_m_k,
+    const XYGridDesc_M_K y_grid_desc_m_k,
+    const ScaleBiasGridDesc_M scale_grid_desc_m,
+    const ScaleBiasGridDesc_M bias_grid_desc_m,
+    const MeanVarGridDesc_M mean_var_grid_desc_m,
+    const GetReduceCountPerThreadFunctor get_reduce_count_per_thread,
+    index_t num_k_block_tile_iteration,
+    AccDataType epsilon,
+    const XDataType* const __restrict__ p_x,
+    const ScaleDataType* const __restrict__ p_scale,
+    const BiasDataType* const __restrict__ p_bias,
+    const YElementwiseOp y_elementwise_op,
+    YDataType* const __restrict__ p_y,
+    bool updateMovingAverage,
+    AccDataType averageFactor,
+    MeanVarDataType* const __restrict__ resultRunningMean,
+    MeanVarDataType* const __restrict__ resultRunningVariance,
+    bool saveMeanInvVariance,
+    MeanVarDataType* const __restrict__ resultSaveMean,
+    MeanVarDataType* const __restrict__ resultSaveInvVariance)
+{
+    GridwiseBatchrNormForwardWithBlockwiseWelford_::Run(x_grid_desc_m_k,
+                                                        y_grid_desc_m_k,
+                                                        scale_grid_desc_m,
+                                                        bias_grid_desc_m,
+                                                        mean_var_grid_desc_m,
+                                                        get_reduce_count_per_thread,
+                                                        num_k_block_tile_iteration,
+                                                        epsilon,
+                                                        p_x,
+                                                        p_scale,
+                                                        p_bias,
+                                                        y_elementwise_op,
+                                                        p_y,
+                                                        updateMovingAverage,
+                                                        averageFactor,
+                                                        resultRunningMean,
+                                                        resultRunningVariance,
+                                                        saveMeanInvVariance,
+                                                        resultSaveMean,
+                                                        resultSaveInvVariance);
+};
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          typename XYGridDesc_M_K,
+          typename ScaleBiasGridDesc_M,
+          typename MeanVarGridDesc_M,
+          typename GetReduceCountPerThreadFunctor,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcYDstVectorDim,
+          index_t XSrcVectorSize,
+          index_t YDstVectorSize,
+          index_t ScaleSrcVectorSize,
+          index_t BiasSrcVectorSize,
+          index_t MeanVarSrcDstVectorSize>
+struct GridwiseBatchNormForwardWithBlockwiseWelford
+{
+    static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcYDstVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert((XSrcYDstVectorDim == 0 && MThreadSliceSize % YDstVectorSize == 0) ||
+                      (XSrcYDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XSrcYDstVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    __device__ static void Run(const XYGridDesc_M_K& x_grid_desc_m_k,
+                               const XYGridDesc_M_K& y_grid_desc_m_k,
+                               const ScaleBiasGridDesc_M& scale_grid_desc_m,
+                               const ScaleBiasGridDesc_M& bias_grid_desc_m,
+                               const MeanVarGridDesc_M& mean_var_grid_desc_m,
+                               const GetReduceCountPerThreadFunctor& get_reduce_count_per_thread,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType epsilon,
+                               const XDataType* const __restrict__ p_x,
+                               const ScaleDataType* const __restrict__ p_scale,
+                               const BiasDataType* const __restrict__ p_bias,
+                               const YElementwiseOp y_elementwise_op,
+                               YDataType* const __restrict__ p_y,
+                               bool updateMovingAverage,
+                               AccDataType averageFactor,
+                               MeanVarDataType* const __restrict__ resultRunningMean,
+                               MeanVarDataType* const __restrict__ resultRunningVariance,
+                               bool saveMeanInvVariance,
+                               MeanVarDataType* const __restrict__ resultSaveMean,
+                               MeanVarDataType* const __restrict__ resultSaveInvVariance)
+    {
+        using ck::math::sqrt;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            x_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> scale_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> bias_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            y_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> var_thread_buf;
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        using ThreadBufferLengths_M           = Sequence<MThreadSliceSize>;
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+        constexpr auto thread_buffer_desc_m =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}));
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  XYGridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcYDstVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize,
+                             thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_y_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               YDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               XYGridDesc_M_K,
+                                               YElementwiseOp,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               XSrcYDstVectorDim,
+                                               YDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                y_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * KThreadSliceSize),
+                y_elementwise_op);
+
+        auto threadwise_scale_load =
+            ThreadwiseTensorSliceTransfer_v2<ScaleDataType,
+                                             AccDataType,
+                                             ScaleBiasGridDesc_M,
+                                             decltype(thread_buffer_desc_m),
+                                             ThreadBufferLengths_M,
+                                             Sequence<0>,
+                                             0,
+                                             ScaleSrcVectorSize,
+                                             1,
+                                             true>(
+                scale_grid_desc_m,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize));
+
+        auto threadwise_bias_load = ThreadwiseTensorSliceTransfer_v2<BiasDataType,
+                                                                     AccDataType,
+                                                                     ScaleBiasGridDesc_M,
+                                                                     decltype(thread_buffer_desc_m),
+                                                                     ThreadBufferLengths_M,
+                                                                     Sequence<0>,
+                                                                     0,
+                                                                     BiasSrcVectorSize,
+                                                                     1,
+                                                                     true>(
+            bias_grid_desc_m,
+            make_multi_index(block_global_id * M_BlockTileSize +
+                             thread_m_cluster_id * MThreadSliceSize));
+
+        constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileSize);
+        constexpr auto thread_copy_bwd_step_m_k = make_multi_index(0, -K_BlockTileSize);
+
+        const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto scale_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_scale, scale_grid_desc_m.GetElementSpaceSize());
+
+        const auto bias_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_bias, bias_grid_desc_m.GetElementSpaceSize());
+
+        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_y, y_grid_desc_m_k.GetElementSpaceSize());
+
+        // Step 1:  do welford reduction to get mean and variance
+
+        auto threadwise_welford       = ThreadwiseWelford();
+        threadwise_welford.max_count_ = get_reduce_count_per_thread(thread_k_cluster_id);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
+            var_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+        });
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_val_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+            threadwise_welford.Run(x_thread_buf, mean_thread_buf, var_thread_buf);
+        }
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            int count = threadwise_welford.cur_count_;
+            BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
+        });
+
+        // Step 2: do normalization and output y
+
+        threadwise_scale_load.Run(scale_grid_desc_m,
+                                  scale_global_val_buf,
+                                  thread_buffer_desc_m,
+                                  make_tuple(I0),
+                                  scale_thread_buf);
+
+        threadwise_bias_load.Run(bias_grid_desc_m,
+                                 bias_global_val_buf,
+                                 thread_buffer_desc_m,
+                                 make_tuple(I0),
+                                 bias_thread_buf);
+
+        auto thread_copy_tail_m_k = (num_k_block_tile_iteration - 1) * thread_copy_fwd_step_m_k;
+
+        threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+        threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            threadwise_x_load.Run(x_grid_desc_m_k,
+                                  x_global_val_buf,
+                                  thread_buffer_desc_m_k,
+                                  make_tuple(I0, I0),
+                                  x_thread_buf);
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                AccDataType multiplier =
+                    scale_thread_buf[Number<iM>{}] / sqrt(var_thread_buf[iM] + epsilon);
+
+                AccDataType fused_mean_bias =
+                    bias_thread_buf[Number<iM>{}] - mean_thread_buf[iM] * multiplier;
+
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset =
+                        thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                    // normalize
+                    y_thread_buf(Number<offset>{}) =
+                        x_thread_buf[Number<offset>{}] * multiplier + fused_mean_bias;
+                });
+            });
+
+            threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                   make_tuple(I0, I0),
+                                   y_thread_buf,
+                                   y_grid_desc_m_k,
+                                   y_global_val_buf);
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_bwd_step_m_k);
+        }
+
+        // Step 3: update the moving average of mean and variance (optional)
+
+        if(updateMovingAverage && thread_k_cluster_id == 0)
+        {
+            StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+                running_mean_thread_buf;
+            StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true>
+                running_var_thread_buf;
+
+            auto running_mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultRunningMean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto running_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultRunningVariance, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto threadwise_mean_var_load =
+                ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
+                                                 AccDataType,
+                                                 MeanVarGridDesc_M,
+                                                 decltype(thread_buffer_desc_m),
+                                                 ThreadBufferLengths_M,
+                                                 Sequence<0>,
+                                                 0,
+                                                 MeanVarSrcDstVectorSize,
+                                                 1,
+                                                 true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize));
+
+            threadwise_mean_var_load.Run(mean_var_grid_desc_m,
+                                         running_mean_global_buf,
+                                         thread_buffer_desc_m,
+                                         make_tuple(I0),
+                                         running_mean_thread_buf);
+
+            threadwise_mean_var_load.Run(mean_var_grid_desc_m,
+                                         running_var_global_buf,
+                                         thread_buffer_desc_m,
+                                         make_tuple(I0),
+                                         running_var_thread_buf);
+
+            AccDataType oneMinusAverageFactor = type_convert<AccDataType>(1.0) - averageFactor;
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                running_mean_thread_buf(I) = running_mean_thread_buf[I] * oneMinusAverageFactor +
+                                             mean_thread_buf[I] * averageFactor;
+                running_var_thread_buf(I) = running_var_thread_buf[I] * oneMinusAverageFactor +
+                                            var_thread_buf[I] * averageFactor;
+            });
+
+            auto threadwise_mean_var_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   MeanVarDataType,
+                                                   decltype(thread_buffer_desc_m),
+                                                   MeanVarGridDesc_M,
+                                                   PassThroughOp,
+                                                   ThreadBufferLengths_M,
+                                                   Sequence<0>,
+                                                   0,
+                                                   MeanVarSrcDstVectorSize,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_mean_var_store.Run(thread_buffer_desc_m,
+                                          make_tuple(I0),
+                                          running_mean_thread_buf,
+                                          mean_var_grid_desc_m,
+                                          running_mean_global_buf);
+
+            threadwise_mean_var_store.Run(thread_buffer_desc_m,
+                                          make_tuple(I0),
+                                          running_var_thread_buf,
+                                          mean_var_grid_desc_m,
+                                          running_var_global_buf);
+        };
+
+        // Step 4: save mean and inv-variance (optional)
+
+        if(saveMeanInvVariance && thread_k_cluster_id == 0)
+        {
+            auto result_mean_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultSaveMean, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            auto result_inv_var_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                resultSaveInvVariance, mean_var_grid_desc_m.GetElementSpaceSize());
+
+            // calculate inv-variance as 1/sqrt(epsilon+variance), stored in place of variance
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                var_thread_buf(I) =
+                    type_convert<AccDataType>(1.0f) / sqrt(epsilon + var_thread_buf[I]);
+            });
+
+            auto threadwise_mean_inv_var_store =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   MeanVarDataType,
+                                                   decltype(thread_buffer_desc_m),
+                                                   MeanVarGridDesc_M,
+                                                   PassThroughOp,
+                                                   ThreadBufferLengths_M,
+                                                   Sequence<0>,
+                                                   0,
+                                                   MeanVarSrcDstVectorSize,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>(
+                    mean_var_grid_desc_m,
+                    make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize),
+                    PassThroughOp{});
+
+            threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              mean_thread_buf,
+                                              mean_var_grid_desc_m,
+                                              result_mean_global_buf);
+
+            threadwise_mean_inv_var_store.Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              var_thread_buf,
+                                              mean_var_grid_desc_m,
+                                              result_inv_var_global_buf);
+        };
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp
new file mode 100644
index 0000000000..d686c14b35
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp
@@ -0,0 +1,195 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseElementwise1dFunctor,
+          typename InGrid1dDescTuple,
+          typename OutGrid1dDescTuple,
+          typename InDataTypePointerTuple,
+          typename OutDataTypePointerTuple,
+          typename ElementwiseOperation>
+__global__ void kernel_elementwise_1d(const InGrid1dDescTuple in_grid_1d_desc_tuple,
+                                      const OutGrid1dDescTuple out_grid_1d_desc_tuple,
+                                      const InDataTypePointerTuple p_in_global_tuple,
+                                      const OutDataTypePointerTuple p_out_global_tuple,
+                                      const ElementwiseOperation elementwise_op)
+{
+    GridwiseElementwise1dFunctor::Run(in_grid_1d_desc_tuple,
+                                      out_grid_1d_desc_tuple,
+                                      p_in_global_tuple,
+                                      p_out_global_tuple,
+                                      elementwise_op);
+}
+
+template <typename InGrid1dDescTuple,
+          typename OutGrid1dDescTuple,
+          typename InDataTypePointerTuple,
+          typename OutDataTypePointerTuple,
+          typename ElementwiseOperation,
+          index_t MPerThread,
+          typename InScalarPerVectorSeq,
+          typename OutScalarPerVectorSeq>
+struct GridwiseElementwise_1D
+{
+    static constexpr index_t NumInput  = InDataTypePointerTuple::Size();
+    static constexpr index_t NumOutput = OutDataTypePointerTuple::Size();
+
+    static_assert(NumInput == InScalarPerVectorSeq::Size() &&
+                      NumOutput == OutScalarPerVectorSeq::Size() &&
+                      NumInput == InGrid1dDescTuple::Size() &&
+                      NumOutput == OutGrid1dDescTuple::Size(),
+                  "Tuple size is inconsistent with the number of in/out!");
+
+    static constexpr auto I0 = Number<0>{};
+
+    static constexpr auto thread_buffer_desc_m =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<MPerThread>{}));
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    __device__ static void Run(const InGrid1dDescTuple in_grid_1d_desc_tuple,
+                               const OutGrid1dDescTuple out_grid_1d_desc_tuple,
+                               const InDataTypePointerTuple p_in_global_tuple,
+                               const OutDataTypePointerTuple p_out_global_tuple,
+                               const ElementwiseOperation elementwise_op)
+    {
+        const index_t thread_global_id = get_thread_global_1d_id();
+
+        auto in_thread_buf_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(InDataTypePointerTuple{}[I])>;
+                using DataType        = remove_cv_t<remove_pointer_t<DataTypePointer>>;
+
+                return StaticBuffer<AddressSpaceEnum::Vgpr, DataType, MPerThread, true>{};
+            },
+            Number<NumInput>{});
+
+        auto out_thread_buf_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(OutDataTypePointerTuple{}[I])>;
+                using DataType        = remove_pointer_t<DataTypePointer>;
+
+                return StaticBuffer<AddressSpaceEnum::Vgpr, DataType, MPerThread, true>{};
+            },
+            Number<NumOutput>{});
+
+        auto in_global_buf_tuple = generate_tuple(
+            [&](auto I) {
+                static_assert(in_grid_1d_desc_tuple[I].GetNumOfDimension() == 1);
+
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_in_global_tuple[I], in_grid_1d_desc_tuple[I].GetElementSpaceSize());
+            },
+            Number<NumInput>{});
+
+        auto out_global_buf_tuple = generate_tuple(
+            [&](auto I) {
+                static_assert(out_grid_1d_desc_tuple[I].GetNumOfDimension() == 1);
+
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_out_global_tuple[I], out_grid_1d_desc_tuple[I].GetElementSpaceSize());
+            },
+            Number<NumOutput>{});
+
+        const auto thread_global_offset = make_multi_index(thread_global_id * MPerThread);
+
+        const index_t blockSize    = get_block_size();
+        const index_t blockPerGrid = get_grid_size();
+        const auto M               = in_grid_1d_desc_tuple[I0].GetLength(I0);
+        const index_t loop_step    = blockPerGrid * blockSize * MPerThread;
+        const auto loop_step_index = make_multi_index(loop_step);
+
+        auto in_global_load_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(InDataTypePointerTuple{}[I])>;
+                using DataType        = remove_cv_t<remove_pointer_t<DataTypePointer>>;
+
+                return ThreadwiseTensorSliceTransfer_v2<DataType,
+                                                        DataType,
+                                                        decltype(in_grid_1d_desc_tuple[I]),
+                                                        decltype(thread_buffer_desc_m),
+                                                        Sequence<MPerThread>, // SliceLengths
+                                                        Sequence<0>,          // DimAccessOrder
+                                                        0,                    // SrcVectorDim
+                                                        InScalarPerVectorSeq::At(
+                                                            I), // ScalarPerVector
+                                                        1,      // SrcScalarStrideInVector
+                                                        false>{in_grid_1d_desc_tuple[I],
+                                                               thread_global_offset};
+            },
+            Number<NumInput>{});
+
+        auto out_global_store_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(OutDataTypePointerTuple{}[I])>;
+                using DataType        = remove_pointer_t<DataTypePointer>;
+
+                return ThreadwiseTensorSliceTransfer_v1r3<DataType,
+                                                          DataType,
+                                                          decltype(thread_buffer_desc_m),
+                                                          decltype(out_grid_1d_desc_tuple[I]),
+                                                          PassThroughOp,
+                                                          Sequence<MPerThread>, // SliceLengths
+                                                          Sequence<0>,          // DimAccessOrder
+                                                          0,                    // SrcVectorDim
+                                                          OutScalarPerVectorSeq::At(I),
+                                                          InMemoryDataOperationEnum::Set,
+                                                          1,
+                                                          false>(
+                    out_grid_1d_desc_tuple[I], thread_global_offset, PassThroughOp{});
+            },
+            Number<NumOutput>{});
+
+        index_t num_iter = M / (loop_step);
+        do
+        {
+            static_for<0, NumInput, 1>{}([&](auto I) {
+                in_global_load_tuple(I).Run(in_grid_1d_desc_tuple[I],
+                                            in_global_buf_tuple[I],
+                                            thread_buffer_desc_m,
+                                            make_tuple(I0),
+                                            in_thread_buf_tuple(I));
+
+                in_global_load_tuple(I).MoveSrcSliceWindow(in_grid_1d_desc_tuple[I],
+                                                           loop_step_index);
+            });
+
+            static_for<0, MPerThread, 1>{}([&](auto iM) {
+                // get reference to in data
+                const auto in_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto I) -> const auto& { return in_thread_buf_tuple(I)(iM); },
+                    Number<NumInput>{});
+
+                // get reference to dst data
+                auto out_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto I) -> auto& { return out_thread_buf_tuple(I)(iM); },
+                    Number<NumOutput>{});
+
+                unpack2(elementwise_op, out_data_refs, in_data_refs);
+            });
+
+            static_for<0, NumOutput, 1>{}([&](auto I) {
+                out_global_store_tuple(I).Run(thread_buffer_desc_m,
+                                              make_tuple(I0),
+                                              out_thread_buf_tuple[I],
+                                              out_grid_1d_desc_tuple[I],
+                                              out_global_buf_tuple(I));
+
+                out_global_store_tuple(I).MoveDstSliceWindow(out_grid_1d_desc_tuple[I],
+                                                             loop_step_index);
+            });
+        } while(--num_iter);
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp
new file mode 100644
index 0000000000..bf0e8c186c
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp
@@ -0,0 +1,230 @@
+// SPDX-License-Identifier: MIT
+// // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+//
+#pragma once
+
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseElementwise2dFunctor,
+          typename InGrid2dDescTuple,
+          typename OutGrid2dDescTuple,
+          typename InDataTypePointerTuple,
+          typename OutDataTypePointerTuple,
+          typename ElementwiseOperation>
+__global__ void kernel_elementwise_2d(const InGrid2dDescTuple in_grid_2d_desc_tuple,
+                                      const OutGrid2dDescTuple out_grid_2d_desc_tuple,
+                                      const InDataTypePointerTuple p_in_global_tuple,
+                                      const OutDataTypePointerTuple p_out_global_tuple,
+                                      const ElementwiseOperation elementwise_op,
+                                      const index_t num_threads_m,
+                                      const index_t num_threads_n)
+{
+    GridwiseElementwise2dFunctor::Run(in_grid_2d_desc_tuple,
+                                      out_grid_2d_desc_tuple,
+                                      p_in_global_tuple,
+                                      p_out_global_tuple,
+                                      elementwise_op,
+                                      num_threads_m,
+                                      num_threads_n);
+}
+
+template <typename InGrid2dDescTuple,
+          typename OutGrid2dDescTuple,
+          typename InDataTypePointerTuple,
+          typename OutDataTypePointerTuple,
+          typename ElementwiseOperation,
+          index_t MPerThread,
+          index_t NPerThread,
+          typename InScalarPerVectorSeq,
+          typename OutScalarPerVectorSeq>
+struct GridwiseElementwise_2D
+{
+    static constexpr index_t NumInput  = InDataTypePointerTuple::Size();
+    static constexpr index_t NumOutput = OutDataTypePointerTuple::Size();
+
+    static_assert(NumInput == InScalarPerVectorSeq::Size() &&
+                      NumOutput == OutScalarPerVectorSeq::Size() &&
+                      NumInput == InGrid2dDescTuple::Size() &&
+                      NumOutput == OutGrid2dDescTuple::Size(),
+                  "Tuple size is inconsistent with the number of in/out!");
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr auto thread_buffer_desc_mn =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<MPerThread>{}, Number<NPerThread>{}));
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    __device__ static void Run(const InGrid2dDescTuple in_grid_2d_desc_tuple,
+                               const OutGrid2dDescTuple out_grid_2d_desc_tuple,
+                               const InDataTypePointerTuple p_in_global_tuple,
+                               const OutDataTypePointerTuple p_out_global_tuple,
+                               const ElementwiseOperation elementwise_op,
+                               const index_t num_threads_m,
+                               const index_t num_threads_n)
+    {
+        auto in_thread_buf_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(InDataTypePointerTuple{}[I])>;
+                using DataType        = remove_cv_t<remove_pointer_t<DataTypePointer>>;
+
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    DataType,
+                                    MPerThread * NPerThread,
+                                    true>{};
+            },
+            Number<NumInput>{});
+
+        auto out_thread_buf_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(OutDataTypePointerTuple{}[I])>;
+                using DataType        = remove_pointer_t<DataTypePointer>;
+
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    DataType,
+                                    MPerThread * NPerThread,
+                                    true>{};
+            },
+            Number<NumOutput>{});
+
+        auto in_global_buf_tuple = generate_tuple(
+            [&](auto I) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_in_global_tuple[I], in_grid_2d_desc_tuple[I].GetElementSpaceSize());
+            },
+            Number<NumInput>{});
+
+        auto out_global_buf_tuple = generate_tuple(
+            [&](auto I) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_out_global_tuple[I], out_grid_2d_desc_tuple[I].GetElementSpaceSize());
+            },
+            Number<NumOutput>{});
+
+        const auto M = in_grid_2d_desc_tuple[I0].GetLength(I0);
+        const auto N = in_grid_2d_desc_tuple[I0].GetLength(I1);
+
+        const index_t loop_step_m = num_threads_m * MPerThread;
+        const index_t loop_step_n = num_threads_n * NPerThread;
+
+        const index_t thread_1d_id = get_thread_global_1d_id();
+        index_t tid_m              = thread_1d_id / num_threads_n;
+        index_t tid_n              = thread_1d_id % num_threads_n;
+
+        const auto thread_global_offset = make_multi_index(tid_m * MPerThread, tid_n * NPerThread);
+
+        auto in_global_load_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(InDataTypePointerTuple{}[I])>;
+                using DataType        = remove_cv_t<remove_pointer_t<DataTypePointer>>;
+
+                return ThreadwiseTensorSliceTransfer_v2<
+                    DataType,
+                    DataType,
+                    decltype(in_grid_2d_desc_tuple[I]),
+                    decltype(thread_buffer_desc_mn),
+                    Sequence<MPerThread, NPerThread>, // SliceLengths
+                    Sequence<0, 1>,                   // DimAccessOrder
+                    0,                                // SrcVectorDim
+                    InScalarPerVectorSeq::At(I),      // ScalarPerVector
+                    1,                                // SrcScalarStrideInVector
+                    true>{in_grid_2d_desc_tuple[I], thread_global_offset};
+            },
+            Number<NumInput>{});
+
+        auto out_global_store_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(OutDataTypePointerTuple{}[I])>;
+                using DataType        = remove_pointer_t<DataTypePointer>;
+
+                return ThreadwiseTensorSliceTransfer_v1r3<
+                    DataType,
+                    DataType,
+                    decltype(thread_buffer_desc_mn),
+                    decltype(out_grid_2d_desc_tuple[I]),
+                    PassThroughOp,
+                    Sequence<MPerThread, NPerThread>, // SliceLengths
+                    Sequence<0, 1>,                   // DimAccessOrder
+                    1,                                // SrcVectorDim
+                    1,                                // OutScalarPerVectorSeq::At(I),
+                    InMemoryDataOperationEnum::Set,
+                    1,
+                    true>(out_grid_2d_desc_tuple[I], thread_global_offset, PassThroughOp{});
+            },
+            Number<NumOutput>{});
+
+        index_t num_iter_m = M / (loop_step_m);
+        do
+        {
+            index_t num_iter_n = N / (loop_step_n);
+            do
+            {
+                static_for<0, NumInput, 1>{}([&](auto I) {
+                    in_global_load_tuple(I).Run(in_grid_2d_desc_tuple[I],
+                                                in_global_buf_tuple[I],
+                                                thread_buffer_desc_mn,
+                                                make_tuple(I0, I0),
+                                                in_thread_buf_tuple(I));
+
+                    in_global_load_tuple(I).MoveSrcSliceWindow(in_grid_2d_desc_tuple[I],
+                                                               make_multi_index(0, loop_step_n));
+                });
+
+                static_for<0, MPerThread, 1>{}([&](auto iM) {
+                    static_for<0, NPerThread, 1>{}([&](auto iN) {
+                        constexpr auto offset =
+                            thread_buffer_desc_mn.CalculateOffset(make_tuple(iM, iN));
+                        // get reference to in data
+                        const auto in_data_refs = generate_tie(
+                            // return type should be lvalue
+                            [&](auto I) -> const auto& {
+                                return in_thread_buf_tuple(I)(Number<offset>{});
+                            },
+                            Number<NumInput>{});
+
+                        // get referenec to dst data
+                        auto out_data_refs = generate_tie(
+                            // return type should be lvalue
+                            [&](auto I) -> auto& {
+                                return out_thread_buf_tuple(I)(Number<offset>{});
+                            },
+                            Number<NumOutput>{});
+                        unpack2(elementwise_op, out_data_refs, in_data_refs);
+                    });
+                });
+
+                static_for<0, NumOutput, 1>{}([&](auto I) {
+                    out_global_store_tuple(I).Run(thread_buffer_desc_mn,
+                                                  make_tuple(I0, I0),
+                                                  out_thread_buf_tuple[I],
+                                                  out_grid_2d_desc_tuple[I],
+                                                  out_global_buf_tuple(I));
+
+                    out_global_store_tuple(I).MoveDstSliceWindow(out_grid_2d_desc_tuple[I],
+                                                                 make_multi_index(0, loop_step_n));
+                });
+
+            } while(--num_iter_n);
+
+            static_for<0, NumInput, 1>{}([&](auto I) {
+                in_global_load_tuple(I).MoveSrcSliceWindow(
+                    in_grid_2d_desc_tuple[I],
+                    make_multi_index(loop_step_m, -(N / loop_step_n) * loop_step_n));
+            });
+
+            static_for<0, NumOutput, 1>{}([&](auto I) {
+                out_global_store_tuple(I).MoveDstSliceWindow(
+                    out_grid_2d_desc_tuple[I],
+                    make_multi_index(loop_step_m, -(N / loop_step_n) * loop_step_n));
+            });
+        } while(--num_iter_m);
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp
new file mode 100644
index 0000000000..3ea72b8534
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp
@@ -0,0 +1,500 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+// X = Elementwise(input1, input2, input3, ...)
+// Y = Normalization(X, beta, gamma)
+template <typename InDataTypePointerTuple,
+          typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename XElementwiseOperation,
+          typename YElementwiseOperation,
+          typename InGrid2dDescTuple,
+          typename GridDesc_M_K,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcVectorDim,
+          index_t XSrcVectorSize,
+          index_t GammaSrcVectorDim,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorDim,
+          index_t BetaSrcVectorSize,
+          index_t YDstVectorDim,
+          index_t YDstVectorSize,
+          bool SweepOnce>
+struct GridwiseElementwiseLayernormWelfordVariance_mk_to_mk
+{
+    static_assert((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert((YDstVectorDim == 0 && MThreadSliceSize % YDstVectorSize == 0) ||
+                      (YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr index_t NumInput = InDataTypePointerTuple::Size();
+
+    static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelford<AccDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<AccDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr index_t M_BlockTileSize     = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize     = KThreadClusterSize * KThreadSliceSize;
+    static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
+
+    static constexpr auto XThreadBufferNumber     = Number<KThreadSliceSize / XSrcVectorSize>{};
+    static constexpr auto GammaThreadBufferNumber = Number<KThreadSliceSize / GammaSrcVectorSize>{};
+    static constexpr auto BetaThreadBufferNumber  = Number<KThreadSliceSize / BetaSrcVectorSize>{};
+    static constexpr auto YThreadBufferNumber     = Number<KThreadSliceSize / YDstVectorSize>{};
+
+    __device__ static int GetKPerThread(const GridDesc_M_K& x_grid_desc_m_k,
+                                        int thread_k_cluster_id)
+    {
+        int kPerBlock = x_grid_desc_m_k.GetTransforms()[I2].GetUpperLengths()[I0];
+        int kPerThread =
+            kPerBlock < K_BlockTileSize ? 0 : KThreadSliceSize * (kPerBlock / K_BlockTileSize);
+        int kPerBlockTail = kPerBlock - kPerThread * KThreadClusterSize;
+
+        if(kPerBlockTail > 0)
+        {
+            static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
+                int thread_max_len =
+                    (thread_k_cluster_id + 1) * XSrcVectorSize + K_BlockTileStepSize * i;
+                int delta = thread_max_len - kPerBlockTail;
+                delta     = math::clamp(thread_max_len - kPerBlockTail, 0, XSrcVectorSize);
+                kPerThread += XSrcVectorSize - delta;
+            });
+        }
+
+        return kPerThread;
+    }
+
+    __device__ static void Run(const InGrid2dDescTuple in_grid_2d_desc_tuple,
+                               const GridDesc_M_K& x_grid_desc_m_k,
+                               const GridDesc_M_K& gamma_grid_desc_m_k,
+                               const GridDesc_M_K& beta_grid_desc_m_k,
+                               const GridDesc_M_K& y_grid_desc_m_k,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType epsilon,
+                               const InDataTypePointerTuple p_in_global_tuple,
+                               XDataType* const __restrict__ p_x_lds,
+                               const GammaDataType* const __restrict__ p_gamma_global,
+                               const BetaDataType* const __restrict__ p_beta_global,
+                               YDataType* const __restrict__ p_y_global,
+                               const XElementwiseOperation x_elementwise_op,
+                               const YElementwiseOperation y_elementwise_op)
+    {
+        if constexpr(SweepOnce)
+        {
+            num_k_block_tile_iteration = 1;
+        }
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t grid_size       = get_grid_size();
+
+        auto in_global_buf_tuple = generate_tuple(
+            [&](auto I) {
+                static_assert(in_grid_2d_desc_tuple[I].GetNumOfDimension() ==
+                              2); // matrix dimension
+
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_in_global_tuple[I], in_grid_2d_desc_tuple[I].GetElementSpaceSize());
+            },
+            Number<NumInput>{});
+
+        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
+
+        auto x_lds_val_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_x_lds, x_grid_desc_m_k.GetElementSpaceSize() / grid_size);
+
+        auto in_thread_buf_tuple = generate_tuple(
+            [&](auto) {
+                return generate_tuple(
+                    [&](auto) {
+                        return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                            AccDataType,
+                                            MThreadSliceSize * XSrcVectorSize,
+                                            true>{};
+                    },
+                    Number<NumInput>{});
+            },
+            Number<XThreadBufferNumber>{});
+
+        auto x_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * XSrcVectorSize,
+                                    true>{};
+            },
+            Number<XThreadBufferNumber>{});
+
+        auto gamma_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * GammaSrcVectorSize,
+                                    true>{};
+            },
+            Number<GammaThreadBufferNumber>{});
+
+        auto beta_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * BetaSrcVectorSize,
+                                    true>{};
+            },
+            Number<BetaThreadBufferNumber>{});
+
+        auto y_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    AccDataType,
+                                    MThreadSliceSize * YDstVectorSize,
+                                    true>{};
+            },
+            Number<YThreadBufferNumber>{});
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> var_thread_buf;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, XSrcVectorSize>;
+
+        constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
+
+        auto in_global_load_tuple = generate_tuple(
+            [&](auto I) {
+                using DataTypePointer = remove_cvref_t<decltype(InDataTypePointerTuple{}[I])>;
+                using DataType        = remove_cv_t<remove_pointer_t<DataTypePointer>>;
+
+                return ThreadwiseTensorSliceTransfer_v2<DataType,
+                                                        AccDataType,
+                                                        decltype(in_grid_2d_desc_tuple[I]),
+                                                        decltype(thread_buffer_desc_m_k),
+                                                        ThreadBufferLengths_M_K,
+                                                        ThreadBufferDimAccessOrder,
+                                                        XSrcVectorDim,
+                                                        XSrcVectorSize,
+                                                        1,
+                                                        false>{
+                    in_grid_2d_desc_tuple[I],
+                    make_multi_index(block_global_id * M_BlockTileSize +
+                                         thread_m_cluster_id * MThreadSliceSize,
+                                     thread_k_cluster_id * XSrcVectorSize)};
+            },
+            Number<NumInput>{});
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  AccDataType,
+                                                                  GridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(thread_m_cluster_id * MThreadSliceSize,
+                             thread_k_cluster_id * XSrcVectorSize));
+
+        auto threadwise_gamma_load =
+            ThreadwiseTensorSliceTransfer_v2<GammaDataType,
+                                             AccDataType,
+                                             GridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_k),
+                                             ThreadBufferLengths_M_K,
+                                             ThreadBufferDimAccessOrder,
+                                             GammaSrcVectorDim,
+                                             GammaSrcVectorSize,
+                                             1,
+                                             true>(
+                gamma_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * GammaSrcVectorSize));
+
+        auto threadwise_beta_load =
+            ThreadwiseTensorSliceTransfer_v2<BetaDataType,
+                                             AccDataType,
+                                             GridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_k),
+                                             ThreadBufferLengths_M_K,
+                                             ThreadBufferDimAccessOrder,
+                                             BetaSrcVectorDim,
+                                             BetaSrcVectorSize,
+                                             1,
+                                             true>(
+                beta_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * BetaSrcVectorSize));
+
+        using PassThrough = tensor_operation::element_wise::PassThrough;
+        PassThrough pass_through_op;
+        auto threadwise_x_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               XDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               GridDesc_M_K,
+                                               PassThrough,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               XSrcVectorDim,
+                                               XSrcVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                x_grid_desc_m_k,
+                make_multi_index(thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * XSrcVectorSize),
+                pass_through_op);
+
+        auto threadwise_y_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               YDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               GridDesc_M_K,
+                                               YElementwiseOperation,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               YDstVectorDim,
+                                               YDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                y_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * YDstVectorSize),
+                y_elementwise_op);
+
+        // Copy x from Cache
+        // one pass: fwd, second pass: bwd
+        constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
+        constexpr auto thread_copy_bwd_step_m_k =
+            make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize);
+
+        const auto gamma_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_gamma_global, gamma_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto beta_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_beta_global, beta_grid_desc_m_k.GetElementSpaceSize());
+
+        auto threadwise_welford       = ThreadwiseWelford();
+        threadwise_welford.max_count_ = GetKPerThread(x_grid_desc_m_k, thread_k_cluster_id);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
+            var_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+        });
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
+                static_for<0, NumInput, 1>{}([&](auto I) { // input load loop
+                    in_global_load_tuple(I).Run(in_grid_2d_desc_tuple[I],
+                                                in_global_buf_tuple[I],
+                                                thread_buffer_desc_m_k,
+                                                make_tuple(I0, I0),
+                                                in_thread_buf_tuple(iK0)(I));
+
+                    in_global_load_tuple(I).MoveSrcSliceWindow(in_grid_2d_desc_tuple[I],
+                                                               thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) { // input add loop
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // get reference to in data
+                        const auto in_data_refs = generate_tie(
+                            // return type should be lvalue
+                            [&](auto I) -> const auto& {
+                                return in_thread_buf_tuple(iK0)(I)(Number<offset_m_k>{});
+                            },
+                            Number<NumInput>{});
+
+                        // get reference to dst data
+                        auto out_data_refs = generate_tie(
+                            // return type should be lvalue
+                            [&](auto) -> auto& { return x_thread_buf(iK0)(Number<offset_m_k>{}); },
+                            I1);
+
+                        unpack2(x_elementwise_op, out_data_refs, in_data_refs);
+                    });
+                });
+                threadwise_welford.Run(x_thread_buf[iK0], mean_thread_buf, var_thread_buf);
+
+                if constexpr(!SweepOnce)
+                {
+                    threadwise_x_store.Run(thread_buffer_desc_m_k,
+                                           make_tuple(I0, I0),
+                                           x_thread_buf(iK0),
+                                           x_grid_desc_m_k,
+                                           x_lds_val_buf);
+                    threadwise_x_store.MoveDstSliceWindow(x_grid_desc_m_k,
+                                                          thread_copy_fwd_step_m_k);
+                }
+            });
+        }
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            int count = threadwise_welford.cur_count_;
+            BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
+        });
+
+        auto thread_copy_tail_m_k =
+            (num_k_block_tile_iteration - 1) * XThreadBufferNumber * thread_copy_fwd_step_m_k;
+
+        if constexpr(!SweepOnce)
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_tail_m_k);
+        threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
+        threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
+        threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
+
+        for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+        {
+            if constexpr(!SweepOnce)
+            {
+                static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_x_load.Run(x_grid_desc_m_k,
+                                          x_lds_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          x_thread_buf(i));
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                });
+            }
+
+            static_for<0, GammaThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_gamma_load.Run(gamma_grid_desc_m_k,
+                                          gamma_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          gamma_thread_buf(i));
+                threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                         thread_copy_fwd_step_m_k);
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
+                static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // normalize
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
+                            divisor;
+
+                        // gamma
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) *
+                            gamma_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
+
+            static_for<0, BetaThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_beta_load.Run(beta_grid_desc_m_k,
+                                         beta_global_val_buf,
+                                         thread_buffer_desc_m_k,
+                                         make_tuple(I0, I0),
+                                         beta_thread_buf(i));
+                threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                        thread_copy_fwd_step_m_k);
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // beta
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) +
+                            beta_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
+
+            static_for<0, YThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                       make_tuple(I0, I0),
+                                       y_thread_buf(i),
+                                       y_grid_desc_m_k,
+                                       y_global_val_buf);
+                threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_fwd_step_m_k);
+            });
+
+            if constexpr(!SweepOnce)
+                threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
+            threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                     2 * thread_copy_bwd_step_m_k);
+            threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                    2 * thread_copy_bwd_step_m_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp
new file mode 100644
index 0000000000..3ced4b9ad6
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp
@@ -0,0 +1,1001 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename FloatC0,
+          typename FloatC1,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename C1ElementwiseOperation,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename ReduceGridDescriptor_MBlock_MPerBlock,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_bias_add_reduce_xdl_cshuffle_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const FloatC0* __restrict__ p_bias_grid,
+            const FloatC1* __restrict__ p_d0_grid,
+            ReducePtrsGlobal p_reduces_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const C1ElementwiseOperation c1_element_op,
+            const ReduceInElementwiseOperations reduce_in_element_ops,
+            const ReduceAccElementwiseOperations reduce_out_element_ops,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c0_grid_desc_mblock_mperblock_nblock_nperblock,
+            const C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c1_grid_desc_mblock_mperblock_nblock_nperblock,
+            const ReduceGridDescriptor_MBlock_MPerBlock reduce_grid_desc_mblock_mperblock,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_bias_grid,
+                                                  p_d0_grid,
+                                                  p_reduces_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  c1_element_op,
+                                                  reduce_in_element_ops,
+                                                  reduce_out_element_ops,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  c1_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  reduce_grid_desc_mblock_mperblock,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_bias_grid;
+    ignore = p_d0_grid;
+    ignore = p_reduces_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = c1_element_op;
+    ignore = reduce_in_element_ops;
+    ignore = reduce_out_element_ops;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = c0_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = c1_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = reduce_grid_desc_mblock_mperblock;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename FloatC0,
+          typename FloatC1,
+          typename FloatReduceAcc,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename C1ElementwiseOperation,
+          typename ReduceOperations,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename ReduceGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDesc_M_N,
+          typename C0GridDesc_M_N,
+          typename C1GridDesc_M_N,
+          typename ReduceGridDesc_M,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+          index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+          index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmBiasAddReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        // static_assert(is_known_at_compile_time<remove_cv_t<decltype(AK1)>>::value &&
+        //               is_known_at_compile_time<remove_cv_t<decltype(BK1)>>::value,
+        //               "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    template <typename CGridDesc_M_N_>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N_& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeReduceGridDescriptor_MBlock_MPerBlock(const ReduceGridDesc_M& d_grid_desc_m)
+    {
+        const auto M      = d_grid_desc_m.GetLength(I0);
+        const auto MBlock = M / MPerBlock;
+
+        const auto reduce_grid_desc_mblock_mperblock = transform_tensor_descriptor(
+            d_grid_desc_m,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{}))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1>{}));
+
+        return reduce_grid_desc_mblock_mperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            CGridDesc_M_N{}))>;
+
+    using C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            C0GridDesc_M_N{}))>;
+
+    using C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            C1GridDesc_M_N{}))>;
+
+    using ReduceGridDescriptor_MBlock_MPerBlock =
+        remove_cvref_t<decltype(MakeReduceGridDescriptor_MBlock_MPerBlock(ReduceGridDesc_M{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        const FloatC0* __restrict__ p_bias_grid,
+        const FloatC1* __restrict__ p_d0_grid,
+        ReducePtrsGlobal p_reduces_grid,
+        void* __restrict__ p_shared,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const C1ElementwiseOperation& c1_element_op,
+        const ReduceInElementwiseOperations& reduce_in_element_ops,
+        const ReduceAccElementwiseOperations& reduce_out_element_ops,
+        const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+        const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+        const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            c_grid_desc_mblock_mperblock_nblock_nperblock,
+        const C0GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            c0_grid_desc_mblock_mperblock_nblock_nperblock,
+        const C1GridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            c1_grid_desc_mblock_mperblock_nblock_nperblock,
+        const ReduceGridDescriptor_MBlock_MPerBlock& reduce_grid_desc_mblock_mperblock,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+        auto c0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_bias_grid, c0_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+        auto c1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_d0_grid, c1_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C + reduction + write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            // TODO: this should be implemented as a blockwise reduction
+            // LDS c_reduce_block_desc_mperblock_nperblock
+            constexpr auto c_reduce_block_desc_mperblock_nperblock = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1)),
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I3))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<>{}, Sequence<1>{}));
+
+            static_assert(CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) *
+                                  CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                              BlockSize,
+                          "wrong!");
+
+            static_assert((CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) %
+                                      CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) ==
+                                  0 &&
+                              (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) %
+                                      CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                                  0,
+                          "wrong!");
+
+            constexpr index_t mreduce_per_thread =
+                (CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) /
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0);
+
+            constexpr index_t nreduce_per_thread =
+                (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) /
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1);
+
+            constexpr auto c_reduce_thread_lengths_mperblock_nperblock =
+                Sequence<mreduce_per_thread, nreduce_per_thread>{};
+
+            // VGPR c_reduce_thread_desc_mperblock_nperblock
+            constexpr auto c_reduce_thread_desc_mperblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(Number<mreduce_per_thread>{}, Number<nreduce_per_thread>{}));
+
+            // VGPR reduce_thread_desc_mperblock
+            constexpr auto reduce_thread_desc_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(Number<mreduce_per_thread>{}));
+
+            // VGPR reduce_thread_desc_mblock_mperblock
+            constexpr auto reduce_thread_desc_mblock_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(I1, Number<mreduce_per_thread>{}));
+
+            auto c_reduce_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                c_reduce_thread_desc_mperblock_nperblock.GetElementSpaceSize());
+
+            // reduce: threadwise copy from LDS to VGPR
+            constexpr auto c_reduce_thread_cluster_desc = make_cluster_descriptor(
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock{}, Sequence<1, 0>{});
+
+            const auto c_reduce_thread_cluster_idx =
+                c_reduce_thread_cluster_desc.CalculateBottomIndex(
+                    make_multi_index(get_thread_local_1d_id()));
+
+            const auto c_reduce_thread_data_idx_begin =
+                c_reduce_thread_cluster_idx * c_reduce_thread_lengths_mperblock_nperblock;
+
+            auto c_reduce_thread_copy_lds_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatCShuffle,
+                FloatReduceAcc,
+                decltype(c_reduce_block_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_lengths_mperblock_nperblock),
+                Sequence<0, 1>,
+                1,
+                CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>{c_reduce_block_desc_mperblock_nperblock, c_reduce_thread_data_idx_begin};
+
+            auto reduce_tuple_thread_copy_vgpr_to_global = generate_tuple(
+                [&](auto I) {
+                    auto p_reduce_grid         = p_reduces_grid[I];
+                    auto reduce_acc_element_op = reduce_out_element_ops[I];
+
+                    return ThreadwiseTensorSliceTransfer_v1r3<
+                        FloatReduceAcc,
+                        remove_pointer_t<decltype(p_reduce_grid)>,
+                        decltype(reduce_thread_desc_mblock_mperblock),
+                        decltype(reduce_grid_desc_mblock_mperblock),
+                        decltype(reduce_acc_element_op),
+                        Sequence<1, mreduce_per_thread>,
+                        Sequence<0, 1>,
+                        1,
+                        CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+                        ReduceGlobalMemoryDataOperation::At(I),
+                        1,
+                        false>{reduce_grid_desc_mblock_mperblock,
+                               make_multi_index(block_work_idx[I0],                  // mblock
+                                                c_reduce_thread_data_idx_begin[I0]), // mperblock
+                               reduce_acc_element_op};
+                },
+                Number<p_reduces_grid.Size()>{});
+
+            // c0 and c1
+            constexpr auto c0_reduce_thread_desc_mblock_mperblock_nblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mreduce_per_thread>{}, I1, Number<nreduce_per_thread>{}));
+
+            constexpr auto c1_reduce_thread_desc_mblock_mperblock_nblock_nperblock =
+                c0_reduce_thread_desc_mblock_mperblock_nblock_nperblock;
+
+            auto c01_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                c0_reduce_thread_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            auto c0_thread_copy_global_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatC0,
+                FloatReduceAcc,
+                decltype(c0_grid_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c0_reduce_thread_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>(
+                c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                make_multi_index(I0,
+                                 m_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I0],
+                                 I0,
+                                 n_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I1]));
+
+            auto c1_thread_copy_global_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatC1,
+                FloatReduceAcc,
+                decltype(c1_grid_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c1_reduce_thread_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>(
+                c1_grid_desc_mblock_mperblock_nblock_nperblock,
+                make_multi_index(I0,
+                                 m_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I0],
+                                 I0,
+                                 n_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I1]));
+
+            constexpr auto c_reduce_thread_desc_mblock_mperblock_nblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mreduce_per_thread>{}, I1, Number<nreduce_per_thread>{}));
+
+            auto c_reduce_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
+                FloatReduceAcc,
+                FloatC,
+                decltype(c_reduce_thread_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                tensor_operation::element_wise::PassThrough,
+                Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>, // SliceLengths
+                Sequence<0, 1, 2, 3>,                                     // DimAccessOrder
+                3,                                                        // DstVectorDim
+                CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+                make_multi_index(I0,
+                                 m_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I0],
+                                 I0,
+                                 n_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I1]),
+                tensor_operation::element_wise::PassThrough{}};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+                {
+                    c_reduce_thread_copy_lds_to_vgpr.Run(c_reduce_block_desc_mperblock_nperblock,
+                                                         c_shuffle_block_buf,
+                                                         c_reduce_thread_desc_mperblock_nperblock,
+                                                         make_tuple(I0, I0),
+                                                         c_reduce_thread_buf);
+
+                    c0_thread_copy_global_to_vgpr.Run(
+                        c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                        c0_grid_buf,
+                        c0_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        c01_thread_buf);
+
+                    // c = activation(c + bias)
+                    static_for<0, c_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                        [&](auto i) {
+                            FloatReduceAcc out;
+                            c_element_op(out, c_reduce_thread_buf(i) + c01_thread_buf(i));
+                            c_reduce_thread_buf(i) = out;
+                        });
+
+                    c1_thread_copy_global_to_vgpr.Run(
+                        c1_grid_desc_mblock_mperblock_nblock_nperblock,
+                        c1_grid_buf,
+                        c1_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        c01_thread_buf);
+
+                    // c = c + c1_functior(c1)
+                    static_for<0, c_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                        [&](auto i) {
+                            c1_element_op(c01_thread_buf(i), c01_thread_buf(i));
+                            c_reduce_thread_buf(i) += c01_thread_buf(i);
+                        });
+
+                    c_reduce_thread_copy_vgpr_to_global.Run(
+                        c_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        c_reduce_thread_buf,
+                        c_grid_desc_mblock_mperblock_nblock_nperblock,
+                        c_grid_buf);
+
+                    static_for<0, p_reduces_grid.Size(), 1>{}([&](auto In) {
+                        auto& p_reduce_grid = p_reduces_grid[In];
+
+                        auto reduce_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                            p_reduce_grid, reduce_grid_desc_mblock_mperblock.GetElementSpaceSize());
+
+                        auto reduce_thread_buf =
+                            make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                                reduce_thread_desc_mperblock.GetElementSpaceSize());
+
+                        auto& reduce_in_element_op = reduce_in_element_ops[In];
+
+                        auto& reduce_thread_copy_vgpr_to_global =
+                            reduce_tuple_thread_copy_vgpr_to_global(In);
+
+                        using ReduceOperation = remove_cvref_t<decltype(ReduceOperations{}[In])>;
+                        using ThreadwiseReduce =
+                            ThreadwiseReduction<FloatReduceAcc,
+                                                decltype(c_reduce_thread_desc_mperblock_nperblock),
+                                                decltype(reduce_thread_desc_mperblock),
+                                                ReduceOperation,
+                                                false>;
+
+                        // Global write Gemm shuffle + reduction
+                        const auto reduce_identityVal =
+                            ReduceOperation::template GetIdentityValue<FloatReduceAcc>();
+
+                        static_for<0, mreduce_per_thread, 1>{}(
+                            [&](auto I) { reduce_thread_buf(I) = reduce_identityVal; });
+
+                        // reduce in VGPR
+                        static_for<0, mreduce_per_thread, 1>{}([&](auto im) {
+                            static_for<0, nreduce_per_thread, 1>{}([&](auto in) {
+                                constexpr auto offset =
+                                    Number<c_reduce_thread_desc_mperblock_nperblock.CalculateOffset(
+                                        make_tuple(im, in))>{};
+
+                                reduce_in_element_op(c_reduce_thread_buf(offset),
+                                                     c_reduce_thread_buf(offset));
+                            });
+                        });
+
+                        ThreadwiseReduce::Reduce(c_reduce_thread_buf, reduce_thread_buf);
+
+                        // copy from VGPR to Global
+                        reduce_thread_copy_vgpr_to_global.Run(reduce_thread_desc_mblock_mperblock,
+                                                              make_tuple(I0, I0),
+                                                              reduce_thread_buf,
+                                                              reduce_grid_desc_mblock_mperblock,
+                                                              reduce_grid_buf);
+
+                        if constexpr(access_id < num_access - 1)
+                        {
+                            constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+                            reduce_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                                reduce_grid_desc_mblock_mperblock,
+                                make_tuple(c_global_step[I0], c_global_step[I1]));
+                        }
+                    });
+                }
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_reduce_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+
+                    // move on C0
+                    c0_thread_copy_global_to_vgpr.MoveSrcSliceWindow(
+                        c0_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+
+                    // move on C1
+                    c1_thread_copy_global_to_vgpr.MoveSrcSliceWindow(
+                        c1_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        } // Reduction
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp
new file mode 100644
index 0000000000..27f48a84ba
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp
@@ -0,0 +1,681 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename DsDataType,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M_N,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t K1Value,
+          index_t M1PerThreadM111,
+          index_t N1PerThreadN111,
+          index_t KPerThread,
+          typename M11N11ThreadClusterM110Xs,
+          typename M11N11ThreadClusterN110Xs,
+          typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+          typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+          typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+          typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+          typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector>
+struct GridwiseGemmDlMultipleD_km_kn_mn
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // TODO: change this. I think it needs multi-dimensional alignment
+        constexpr auto max_lds_align = K1;
+
+        // TODO: check alignment
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k_m = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k_n = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_aligned_space_size =
+            math::integer_least_multiple(a_block_desc_k_m.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_aligned_space_size =
+            math::integer_least_multiple(b_block_desc_k_n.GetElementSpaceSize(), max_lds_align);
+
+        return 2 * (a_block_aligned_space_size + b_block_aligned_space_size) * sizeof(FloatAB);
+    }
+
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+
+        return (M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+                K0 == b_grid_desc_k0_n_k1.GetLength(I0) &&
+                K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+                K1 == b_grid_desc_k0_n_k1.GetLength(I2)) &&
+               (M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0);
+    }
+
+    __host__ __device__ static constexpr index_t CalculateGridSize(index_t M, index_t N)
+    {
+        const index_t grid_size = (M / MPerBlock) * (N / NPerBlock);
+
+        return grid_size;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K0)
+    {
+        const bool has_main_k_block_loop = (K0 + K0PerBlock) / (2 * K0PerBlock) > 1;
+
+        return has_main_k_block_loop;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasDoubleTailKBlockLoop(index_t K0)
+    {
+        const bool has_double_tail_k_block_loop = (K0 / K0PerBlock) % 2 == 0;
+
+        return has_double_tail_k_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeAGridDescriptor_K0_M0_M1_K1(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1)
+    {
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+
+        const auto M1 = Number<MPerBlock>{};
+        const auto M0 = M / M1;
+
+        const auto a_grid_desc_k0_m0_m1_k1 =
+            transform_tensor_descriptor(a_grid_desc_k0_m_k1,
+                                        make_tuple(make_pass_through_transform(K0),
+                                                   make_unmerge_transform(make_tuple(M0, M1)),
+                                                   make_pass_through_transform(K1)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+        return a_grid_desc_k0_m0_m1_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeBGridDescriptor_K0_N0_N1_K1(const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1)
+    {
+        const auto K0 = b_grid_desc_k0_n_k1.GetLength(I0);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+
+        const auto N1 = Number<NPerBlock>{};
+        const auto N0 = N / N1;
+
+        const auto b_grid_desc_k0_n0_n1_k1 =
+            transform_tensor_descriptor(b_grid_desc_k0_n_k1,
+                                        make_tuple(make_pass_through_transform(K0),
+                                                   make_unmerge_transform(make_tuple(N0, N1)),
+                                                   make_pass_through_transform(K1)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+        return b_grid_desc_k0_n0_n1_k1;
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename CGridDesc_M_N_>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(const CGridDesc_M_N_& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        constexpr auto M11 =
+            Number<container_reduce(M11N11ThreadClusterM110Xs{}, math::multiplies{}, I1) *
+                   M1PerThreadM111>{};
+        constexpr auto N11 =
+            Number<container_reduce(M11N11ThreadClusterN110Xs{}, math::multiplies{}, I1) *
+                   N1PerThreadN111>{};
+
+        constexpr auto M10 = M1 / M11;
+        constexpr auto N10 = N1 / N11;
+
+        const auto c_grid_desc_m0_m10_m11_n0_n10_n11 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M0, M10, M11)),
+                       make_unmerge_transform(make_tuple(N0, N10, N11))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
+
+        return c_grid_desc_m0_m10_m11_n0_n10_n11;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(const DsGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) { return MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(ds_grid_desc_m_n[i]); },
+            Number<NumDTensor>{});
+    }
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N00_M01_N01<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using AGridDesc_K0_M0_M1_K1 = decltype(MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_K0_M_K1{}));
+    using BGridDesc_K0_N0_N1_K1 = decltype(MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{}));
+    using CGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <typename DsGridDesc_M0_M10_M11_N0_N10_N11,
+              bool HasMainKBlockLoop,
+              bool HasDoubleTailKBlockLoop,
+              typename Block2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        DsGridPointer p_ds_grid,
+        FloatC* __restrict__ p_c_grid,
+        void* __restrict__ p_shared_block,
+        const AElementwiseOperation&,
+        const BElementwiseOperation&,
+        const CDEElementwiseOperation& cde_element_op,
+        const AGridDesc_K0_M0_M1_K1& a_grid_desc_k0_m0_m1_k1,
+        const BGridDesc_K0_N0_N1_K1& b_grid_desc_k0_n0_n1_k1,
+        const DsGridDesc_M0_M10_M11_N0_N10_N11& ds_grid_desc_m0_m10_m11_n0_n10_n11,
+        const CGridDesc_M0_M10_M11_N0_N10_N11& c_grid_desc_m0_m10_m11_n0_n10_n11,
+        const Block2CTileMap& block_2_ctile_map,
+        integral_constant<bool, HasMainKBlockLoop>,
+        integral_constant<bool, HasDoubleTailKBlockLoop>)
+    {
+        const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m0_m1_k1.GetElementSpaceSize());
+        const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_n0_n1_k1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_m0_m10_m11_n0_n10_n11.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto c_m0_n0_block_cluster_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        // HACK: this force index data into SGPR
+        const index_t im0 = __builtin_amdgcn_readfirstlane(c_m0_n0_block_cluster_idx[I0]);
+        const index_t in0 = __builtin_amdgcn_readfirstlane(c_m0_n0_block_cluster_idx[I1]);
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               make_tuple(im0, in0),
+               make_tuple(c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I0),
+                          c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I3))))
+        {
+            return;
+        }
+
+        // TODO: change this. I think it needs multi-dimensional alignment
+        constexpr auto max_lds_align = K1;
+
+        // TODO: check alignment
+        // A matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto a_block_desc_k0_m0_m1_k1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, I1, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto b_block_desc_k0_n0_n1_k1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, I1, Number<NPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // A matrix in LDS memory, for blockwise GEMM
+        constexpr auto a_k0_m_k1_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, for blockwise GEMM
+        constexpr auto b_k0_n_k1_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+
+        static_assert(a_block_desc_k0_m0_m1_k1.GetElementSpaceSize() ==
+                          a_k0_m_k1_block_desc.GetElementSpaceSize() &&
+                      b_block_desc_k0_n0_n1_k1.GetElementSpaceSize() ==
+                          b_k0_n_k1_block_desc.GetElementSpaceSize() &&
+                      "wrong!");
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
+            BlockSize,
+            InMemoryDataOperationEnum::Set,
+            Sequence<K0PerBlock, 1, MPerBlock, K1.value>,
+            ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+            ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+            ABlockTransferThreadClusterArrangeOrder,
+            FloatAB,
+            FloatAB,
+            remove_reference_t<decltype(a_grid_desc_k0_m0_m1_k1)>,
+            decltype(a_block_desc_k0_m0_m1_k1),
+            ABlockTransferSrcAccessOrder,
+            Sequence<0, 1, 2, 3>,
+            ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1, // SrcVectorTensorLengths
+            ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1, // DstVectorTensorLengths
+            ABlockTransferSrcVectorTensorContiguousDimOrder,  // SrcVectorTensorContiguousDimOrder
+            Sequence<0, 1, 2, 3>,                             // DstVectorTensorContiguousDimOrder
+            false,
+            true>(a_grid_desc_k0_m0_m1_k1,
+                  make_multi_index(0, im0, 0, 0),
+                  a_block_desc_k0_m0_m1_k1,
+                  make_multi_index(0, 0, 0, 0));
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
+            BlockSize,
+            InMemoryDataOperationEnum::Set,
+            Sequence<K0PerBlock, 1, NPerBlock, K1.value>,
+            BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+            BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+            BBlockTransferThreadClusterArrangeOrder,
+            FloatAB,
+            FloatAB,
+            remove_reference_t<decltype(b_grid_desc_k0_n0_n1_k1)>,
+            decltype(b_block_desc_k0_n0_n1_k1),
+            BBlockTransferSrcAccessOrder,
+            Sequence<0, 1, 2, 3>,
+            BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1, // SrcVectorTensorLengths
+            BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1, // DstVectorTensorLengths
+            BBlockTransferSrcVectorTensorContiguousDimOrder,  // SrcVectorTensorContiguousDimOrder
+            Sequence<0, 1, 2, 3>,                             // DstVectorTensorContiguousDimOrder
+            false,
+            true>(b_grid_desc_k0_n0_n1_k1,
+                  make_multi_index(0, in0, 0, 0),
+                  b_block_desc_k0_n0_n1_k1,
+                  make_multi_index(0, 0, 0, 0));
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[KPerBlocl, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        const auto blockwise_gemm =
+            BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2<
+                BlockSize,
+                FloatAB,
+                FloatAB,
+                FloatAcc,
+                decltype(a_k0_m_k1_block_desc),
+                decltype(b_k0_n_k1_block_desc),
+                M1PerThreadM111,
+                N1PerThreadN111,
+                KPerThread,
+                M11N11ThreadClusterM110Xs,
+                M11N11ThreadClusterN110Xs,
+                M1PerThreadM111,
+                N1PerThreadN111>{};
+
+        constexpr auto c_m10_m11_n10_n11_thread_tensor_lengths =
+            decltype(blockwise_gemm)::GetCThreadTensorLengths_BM0_BM1_BN0_BN1();
+
+        constexpr auto c_thread_desc_m10_m11_n10_n11 = make_naive_tensor_descriptor_packed(
+            sequence_to_tuple_of_number(c_m10_m11_n10_n11_thread_tensor_lengths));
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_aligned_space_size = math::integer_least_multiple(
+            a_block_desc_k0_m0_m1_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_aligned_space_size = math::integer_least_multiple(
+            b_block_desc_k0_n0_n1_k1.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block_double = static_cast<FloatAB*>(p_shared_block);
+        FloatAB* p_b_block_double =
+            static_cast<FloatAB*>(p_shared_block) + 2 * a_block_aligned_space_size;
+
+        // register allocation for output
+        auto c_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAcc>(
+            c_thread_desc_m10_m11_n10_n11.GetElementSpaceSize());
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0, 0);
+
+        auto a_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block_double, a_block_desc_k0_m0_m1_k1.GetElementSpaceSize());
+        auto b_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block_double, b_block_desc_k0_n0_n1_k1.GetElementSpaceSize());
+
+        auto a_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block_double + a_block_aligned_space_size,
+            a_block_desc_k0_m0_m1_k1.GetElementSpaceSize());
+        auto b_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block_double + b_block_aligned_space_size,
+            b_block_desc_k0_n0_n1_k1.GetElementSpaceSize());
+
+        // LDS double buffer: preload data into LDS
+        {
+            a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+            b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+            a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_even_buf);
+            b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_even_buf);
+        }
+
+        if constexpr(HasMainKBlockLoop)
+        {
+            const auto K0 = a_grid_desc_k0_m0_m1_k1.GetLength(I0);
+
+            index_t k_block_data_begin = 0;
+
+            // LDS double buffer: main body
+            // use Do-While loop instead of For loop to simplify control flow
+            do
+            {
+                // even iteration
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m0_m1_k1,
+                                                    a_block_slice_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n0_n1_k1,
+                                                    b_block_slice_copy_step);
+
+                // LDS doubel buffer: load next data from device mem
+                a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+                b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+                block_sync_lds();
+
+                // LDS double buffer: GEMM on current data
+                blockwise_gemm.Run(c_thread_desc_m10_m11_n10_n11,
+                                   a_block_even_buf,
+                                   b_block_even_buf,
+                                   c_thread_buf);
+
+                // LDS double buffer: store next data to LDS
+                a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_odd_buf);
+                b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_odd_buf);
+
+                // odd iteration
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m0_m1_k1,
+                                                    a_block_slice_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n0_n1_k1,
+                                                    b_block_slice_copy_step);
+
+                // LDS doubel buffer: load next data from device mem
+                a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+                b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+                block_sync_lds();
+
+                // LDS double buffer: GEMM on current data
+                blockwise_gemm.Run(
+                    c_thread_desc_m10_m11_n10_n11, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
+
+                // LDS double buffer: store next data to LDS
+                a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_even_buf);
+                b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_even_buf);
+
+                k_block_data_begin += 2 * K0PerBlock;
+            } while(k_block_data_begin < K0 - 2 * K0PerBlock);
+        }
+
+        // LDS double buffer: tail
+        if constexpr(HasDoubleTailKBlockLoop) // if has 2 iteration left
+        {
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m0_m1_k1, a_block_slice_copy_step);
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n0_n1_k1, b_block_slice_copy_step);
+
+            block_sync_lds();
+
+            // LDS double buffer: load last data from device mem
+            a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+            b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+            // LDS double buffer: GEMM on 2nd-last data
+            blockwise_gemm.Run(
+                c_thread_desc_m10_m11_n10_n11, a_block_even_buf, b_block_even_buf, c_thread_buf);
+
+            // LDS double buffer: store last data to LDS
+            a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_odd_buf);
+            b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_odd_buf);
+
+            block_sync_lds();
+
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(
+                c_thread_desc_m10_m11_n10_n11, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
+        }
+        else // if has 1 iteration left
+        {
+            __syncthreads();
+
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(
+                c_thread_desc_m10_m11_n10_n11, a_block_even_buf, b_block_even_buf, c_thread_buf);
+        }
+
+        // output: register to global memory
+        {
+            constexpr auto c_thread_desc_m0_m10_m11_n0_n10_n11 =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1,
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I0]>{},
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I1]>{},
+                               I1,
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I2]>{},
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I3]>{}));
+
+            const auto c_m10_m11_n10_n11_thread_origin_idx_on_block =
+                blockwise_gemm.CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
+                    get_thread_local_1d_id());
+
+            const auto ds_grid_buf = generate_tuple(
+                [&](auto i) {
+                    return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                        p_ds_grid[i], ds_grid_desc_m0_m10_m11_n0_n10_n11[i].GetElementSpaceSize());
+                },
+                Number<NumDTensor>{});
+
+            auto ds_thread_buf = generate_tuple(
+                [&](auto i) {
+                    using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                    return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                        DDataType,
+                                        c_m10_m11_n10_n11_thread_tensor_lengths[I3],
+                                        true>{};
+                },
+                Number<NumDTensor>{});
+
+            auto ds_threadwise_copy = generate_tuple(
+                [&](auto i) {
+                    using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                    return ThreadwiseTensorSliceTransfer_v2<
+                        DDataType,
+                        DDataType,
+                        decltype(ds_grid_desc_m0_m10_m11_n0_n10_n11[i]),
+                        decltype(c_thread_desc_m0_m10_m11_n0_n10_n11),
+                        Sequence<I1,
+                                 I1,
+                                 I1,
+                                 I1,
+                                 I1,
+                                 Number<c_m10_m11_n10_n11_thread_tensor_lengths[I3]>{}>,
+                        CThreadTransferSrcDstAccessOrder,
+                        CThreadTransferSrcDstVectorDim,
+                        CThreadTransferDstScalarPerVector,
+                        1,
+                        false>(ds_grid_desc_m0_m10_m11_n0_n10_n11[i],
+                               make_multi_index(im0,
+                                                c_m10_m11_n10_n11_thread_origin_idx_on_block[I0],
+                                                c_m10_m11_n10_n11_thread_origin_idx_on_block[I1],
+                                                in0,
+                                                c_m10_m11_n10_n11_thread_origin_idx_on_block[I2],
+                                                c_m10_m11_n10_n11_thread_origin_idx_on_block[I3]));
+                },
+                Number<NumDTensor>{});
+
+            static_for<0, c_m10_m11_n10_n11_thread_tensor_lengths[I0], 1>{}([&](auto m10) {
+                static_for<0, c_m10_m11_n10_n11_thread_tensor_lengths[I1], 1>{}([&](auto m11) {
+                    static_for<0, c_m10_m11_n10_n11_thread_tensor_lengths[I2], 1>{}([&](auto n10) {
+                        // load d matrix data
+                        static_for<0, NumDTensor, 1>{}([&](auto i) {
+                            ds_threadwise_copy(i).Run(ds_grid_desc_m0_m10_m11_n0_n10_n11[i],
+                                                      ds_grid_buf[i],
+                                                      c_thread_desc_m0_m10_m11_n0_n10_n11,
+                                                      make_tuple(I0, I0, I0, I0, I0, I0),
+                                                      ds_thread_buf(i));
+                        });
+                        // cal element op
+                        static_for<0, c_m10_m11_n10_n11_thread_tensor_lengths[I3], 1>{}(
+                            [&](auto i) {
+                                // get reference to src data
+                                const auto src_data_refs = generate_tie(
+                                    // return type should be lvalue
+                                    [&](auto iSrc) -> const auto& {
+                                        return ds_thread_buf[iSrc][i];
+                                    },
+                                    Number<NumDTensor>{});
+
+                                // get reference to dst data
+                                constexpr index_t c_offset =
+                                    c_thread_desc_m0_m10_m11_n0_n10_n11.CalculateOffset(
+                                        make_tuple(0, m10, m11, 0, n10, i));
+                                auto dst_data_refs = generate_tie(
+                                    // return type should be lvalue
+                                    [&](auto) -> auto& { return c_thread_buf(Number<c_offset>{}); },
+                                    Number<2>{});
+
+                                unpack2(cde_element_op, dst_data_refs, src_data_refs);
+                            });
+
+                        static_for<0, NumDTensor, 1>{}([&](auto i) {
+                            ds_threadwise_copy(i).MoveSrcSliceWindow(
+                                ds_grid_desc_m0_m10_m11_n0_n10_n11[i],
+                                make_multi_index(0, 0, 0, 0, 1, 0));
+                        });
+                    });
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        ds_threadwise_copy(i).MoveSrcSliceWindow(
+                            ds_grid_desc_m0_m10_m11_n0_n10_n11[i],
+                            make_multi_index(
+                                0, 0, 1, 0, -c_m10_m11_n10_n11_thread_tensor_lengths[I2], 0));
+                    });
+                });
+                static_for<0, NumDTensor, 1>{}([&](auto i) {
+                    ds_threadwise_copy(i).MoveSrcSliceWindow(
+                        ds_grid_desc_m0_m10_m11_n0_n10_n11[i],
+                        make_multi_index(
+                            0, 1, -c_m10_m11_n10_n11_thread_tensor_lengths[I1], 0, 0, 0));
+                });
+            });
+
+            ThreadwiseTensorSliceTransfer_v1r3<
+                FloatAcc,
+                FloatC,
+                decltype(c_thread_desc_m0_m10_m11_n0_n10_n11),
+                decltype(c_grid_desc_m0_m10_m11_n0_n10_n11),
+                ck::tensor_operation::element_wise::PassThrough,
+                Sequence<1,
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I0],
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I1],
+                         1,
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I2],
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I3]>,
+                CThreadTransferSrcDstAccessOrder,
+                CThreadTransferSrcDstVectorDim,
+                CThreadTransferDstScalarPerVector,
+                CGlobalMemoryDataOperation,
+                1,
+                true>{c_grid_desc_m0_m10_m11_n0_n10_n11,
+                      make_multi_index(im0,
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I0],
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I1],
+                                       in0,
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I2],
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I3]),
+                      ck::tensor_operation::element_wise::PassThrough{}}
+                .Run(c_thread_desc_m0_m10_m11_n0_n10_n11,
+                     make_tuple(I0, I0, I0, I0, I0, I0),
+                     c_thread_buf,
+                     c_grid_desc_m0_m10_m11_n0_n10_n11,
+                     c_grid_buf);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp
new file mode 100644
index 0000000000..1d1bb6ed2d
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp
@@ -0,0 +1,1158 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_dl_algorithm.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_dl_dpp8.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+using GemmDlAlgorithm = tensor_operation::device::GemmDlAlgorithm;
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M0_M1_K1,
+          typename BGridDesc_K0_N0_N1_K1,
+          typename CGridDesc_M0_M10_M11_N0_N10_N11,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop,
+          bool HasDoubleTailKBlockLoop,
+          GemmDlAlgorithm GemmDlAlg = GemmDlAlgorithm::Default>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_dl_v1r3(const FloatAB* __restrict__ p_a_grid,
+                            const FloatAB* __restrict__ p_b_grid,
+                            FloatC* __restrict__ p_c_grid,
+                            const AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1,
+                            const BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1,
+                            const CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11,
+                            const Block2CTileMap block_2_ctile_map)
+{
+// DPP8 is currently only supported on gfx1030
+#if !defined(__gfx1030__)
+    if(GemmDlAlg == GemmDlAlgorithm::Dpp8)
+    {
+        return;
+    }
+#endif
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+    __shared__ FloatAB p_shared_block[shared_block_size];
+
+    GridwiseGemm::Run(p_a_grid,
+                      p_b_grid,
+                      p_c_grid,
+                      p_shared_block,
+                      a_grid_desc_k0_m0_m1_k1,
+                      b_grid_desc_k0_n0_n1_k1,
+                      c_grid_desc_m0_m10_m11_n0_n10_n11,
+                      block_2_ctile_map,
+                      integral_constant<bool, HasMainKBlockLoop>{},
+                      integral_constant<bool, HasDoubleTailKBlockLoop>{});
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M_N,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t K1Value,
+          index_t M1PerThreadM111,
+          index_t N1PerThreadN111,
+          index_t KPerThread,
+          typename M11N11ThreadClusterM110Xs,
+          typename M11N11ThreadClusterN110Xs,
+          typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+          typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+          typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+          typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+          typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          GemmDlAlgorithm GemmDlAlg = GemmDlAlgorithm::Default>
+struct GridwiseGemmDl_km_kn_mn_v1r3
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // TODO: change this. I think it needs multi-dimensional alignment
+        constexpr auto max_lds_align = K1;
+
+        // TODO: check alignment
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k_m = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k_n = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_aligned_space_size =
+            math::integer_least_multiple(a_block_desc_k_m.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_aligned_space_size =
+            math::integer_least_multiple(b_block_desc_k_n.GetElementSpaceSize(), max_lds_align);
+
+        return 2 * (a_block_aligned_space_size + b_block_aligned_space_size) * sizeof(FloatAB);
+    }
+
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+
+        return (M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+                K0 == b_grid_desc_k0_n_k1.GetLength(I0) &&
+                K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+                K1 == b_grid_desc_k0_n_k1.GetLength(I2)) &&
+               (M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0);
+    }
+
+    __host__ __device__ static constexpr index_t CalculateGridSize(index_t M, index_t N)
+    {
+        const index_t grid_size = (M / MPerBlock) * (N / NPerBlock);
+
+        return grid_size;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K0)
+    {
+        const bool has_main_k_block_loop = (K0 + K0PerBlock) / (2 * K0PerBlock) > 1;
+
+        return has_main_k_block_loop;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasDoubleTailKBlockLoop(index_t K0)
+    {
+        const bool has_double_tail_k_block_loop = (K0 / K0PerBlock) % 2 == 0;
+
+        return has_double_tail_k_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeAGridDescriptor_K0_M0_M1_K1(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1)
+    {
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+
+        const auto M1 = Number<MPerBlock>{};
+        const auto M0 = M / M1;
+
+        const auto a_grid_desc_k0_m0_m1_k1 =
+            transform_tensor_descriptor(a_grid_desc_k0_m_k1,
+                                        make_tuple(make_pass_through_transform(K0),
+                                                   make_unmerge_transform(make_tuple(M0, M1)),
+                                                   make_pass_through_transform(K1)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+        return a_grid_desc_k0_m0_m1_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeBGridDescriptor_K0_N0_N1_K1(const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1)
+    {
+        const auto K0 = b_grid_desc_k0_n_k1.GetLength(I0);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+
+        const auto N1 = Number<NPerBlock>{};
+        const auto N0 = N / N1;
+
+        const auto b_grid_desc_k0_n0_n1_k1 =
+            transform_tensor_descriptor(b_grid_desc_k0_n_k1,
+                                        make_tuple(make_pass_through_transform(K0),
+                                                   make_unmerge_transform(make_tuple(N0, N1)),
+                                                   make_pass_through_transform(K1)),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+        return b_grid_desc_k0_n0_n1_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        constexpr auto M11 =
+            Number<container_reduce(M11N11ThreadClusterM110Xs{}, math::multiplies{}, I1) *
+                   M1PerThreadM111>{};
+        constexpr auto N11 =
+            Number<container_reduce(M11N11ThreadClusterN110Xs{}, math::multiplies{}, I1) *
+                   N1PerThreadN111>{};
+
+        constexpr auto M10 = M1 / M11;
+        constexpr auto N10 = N1 / N11;
+
+        const auto c_grid_desc_m0_m10_m11_n0_n10_n11 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M0, M10, M11)),
+                       make_unmerge_transform(make_tuple(N0, N10, N11))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
+
+        return c_grid_desc_m0_m10_m11_n0_n10_n11;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N00_M01_N01<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    template <typename ABlockDesc_BK0_BM_BK1, typename BBlockDesc_BK0_BN_BK1>
+    __host__ __device__ static constexpr auto GetBlockwiseGemm()
+    {
+        if constexpr(GemmDlAlg == GemmDlAlgorithm::Dpp8)
+        {
+            return BlockwiseGemmDlDpp8_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_loop_BM0_BN0<
+                BlockSize,
+                FloatAB,
+                FloatAB,
+                FloatAcc,
+                ABlockDesc_BK0_BM_BK1,
+                BBlockDesc_BK0_BN_BK1,
+                M1PerThreadM111,
+                N1PerThreadN111,
+                KPerThread,
+                M11N11ThreadClusterM110Xs,
+                M11N11ThreadClusterN110Xs,
+                M1PerThreadM111,
+                N1PerThreadN111>{};
+        }
+        else
+        {
+            return BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2<
+                BlockSize,
+                FloatAB,
+                FloatAB,
+                FloatAcc,
+                ABlockDesc_BK0_BM_BK1,
+                BBlockDesc_BK0_BN_BK1,
+                M1PerThreadM111,
+                N1PerThreadN111,
+                KPerThread,
+                M11N11ThreadClusterM110Xs,
+                M11N11ThreadClusterN110Xs,
+                M1PerThreadM111,
+                N1PerThreadN111>{};
+        }
+    }
+
+    using AGridDesc_K0_M0_M1_K1 = decltype(MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_K0_M_K1{}));
+    using BGridDesc_K0_N0_N1_K1 = decltype(MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{}));
+    using CGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
+    using Block2CTileMap = decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}));
+
+    template <bool HasMainKBlockLoop, bool HasDoubleTailKBlockLoop>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        FloatAB* __restrict__ p_shared_block,
+        const AGridDesc_K0_M0_M1_K1& a_grid_desc_k0_m0_m1_k1,
+        const BGridDesc_K0_N0_N1_K1& b_grid_desc_k0_n0_n1_k1,
+        const CGridDesc_M0_M10_M11_N0_N10_N11& c_grid_desc_m0_m10_m11_n0_n10_n11,
+        const Block2CTileMap& block_2_ctile_map,
+        integral_constant<bool, HasMainKBlockLoop>,
+        integral_constant<bool, HasDoubleTailKBlockLoop>)
+    {
+        const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m0_m1_k1.GetElementSpaceSize());
+        const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_n0_n1_k1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_m0_m10_m11_n0_n10_n11.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto c_m0_n0_block_cluster_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        // HACK: this forces index data into SGPR
+        const index_t im0 = __builtin_amdgcn_readfirstlane(c_m0_n0_block_cluster_idx[I0]);
+        const index_t in0 = __builtin_amdgcn_readfirstlane(c_m0_n0_block_cluster_idx[I1]);
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               make_tuple(im0, in0),
+               make_tuple(c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I0),
+                          c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I3))))
+        {
+            return;
+        }
+
+        // TODO: change this. I think it needs multi-dimensional alignment
+        constexpr auto max_lds_align = K1;
+
+        // TODO: check alignment
+        // A matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto a_block_desc_k0_m0_m1_k1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, I1, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto b_block_desc_k0_n0_n1_k1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, I1, Number<NPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // A matrix in LDS memory, for blockwise GEMM
+        constexpr auto a_k0_m_k1_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, for blockwise GEMM
+        constexpr auto b_k0_n_k1_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+
+        static_assert(a_block_desc_k0_m0_m1_k1.GetElementSpaceSize() ==
+                          a_k0_m_k1_block_desc.GetElementSpaceSize() &&
+                      b_block_desc_k0_n0_n1_k1.GetElementSpaceSize() ==
+                          b_k0_n_k1_block_desc.GetElementSpaceSize() &&
+                      "wrong!");
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
+            BlockSize,
+            InMemoryDataOperationEnum::Set,
+            Sequence<K0PerBlock, 1, MPerBlock, K1.value>,
+            ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+            ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+            ABlockTransferThreadClusterArrangeOrder,
+            FloatAB,
+            FloatAB,
+            remove_reference_t<decltype(a_grid_desc_k0_m0_m1_k1)>,
+            decltype(a_block_desc_k0_m0_m1_k1),
+            ABlockTransferSrcAccessOrder,
+            Sequence<0, 1, 2, 3>,
+            ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1, // SrcVectorTensorLengths
+            ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1, // DstVectorTensorLengths
+            ABlockTransferSrcVectorTensorContiguousDimOrder,  // SrcVectorTensorContiguousDimOrder
+            Sequence<0, 1, 2, 3>,                             // DstVectorTensorContiguousDimOrder
+            false,
+            true>(a_grid_desc_k0_m0_m1_k1,
+                  make_multi_index(0, im0, 0, 0),
+                  a_block_desc_k0_m0_m1_k1,
+                  make_multi_index(0, 0, 0, 0));
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
+            BlockSize,
+            InMemoryDataOperationEnum::Set,
+            Sequence<K0PerBlock, 1, NPerBlock, K1.value>,
+            BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+            BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+            BBlockTransferThreadClusterArrangeOrder,
+            FloatAB,
+            FloatAB,
+            remove_reference_t<decltype(b_grid_desc_k0_n0_n1_k1)>,
+            decltype(b_block_desc_k0_n0_n1_k1),
+            BBlockTransferSrcAccessOrder,
+            Sequence<0, 1, 2, 3>,
+            BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1, // SrcVectorTensorLengths
+            BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1, // DstVectorTensorLengths
+            BBlockTransferSrcVectorTensorContiguousDimOrder,  // SrcVectorTensorContiguousDimOrder
+            Sequence<0, 1, 2, 3>,                             // DstVectorTensorContiguousDimOrder
+            false,
+            true>(b_grid_desc_k0_n0_n1_k1,
+                  make_multi_index(0, in0, 0, 0),
+                  b_block_desc_k0_n0_n1_k1,
+                  make_multi_index(0, 0, 0, 0));
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[KPerBlocl, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        const auto blockwise_gemm =
+            GetBlockwiseGemm<decltype(a_k0_m_k1_block_desc), decltype(b_k0_n_k1_block_desc)>();
+
+        constexpr auto c_m10_m11_n10_n11_thread_tensor_lengths =
+            decltype(blockwise_gemm)::GetCThreadTensorLengths_BM0_BM1_BN0_BN1();
+
+        constexpr auto c_thread_desc_m10_m11_n10_n11 = make_naive_tensor_descriptor_packed(
+            sequence_to_tuple_of_number(c_m10_m11_n10_n11_thread_tensor_lengths));
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_aligned_space_size = math::integer_least_multiple(
+            a_block_desc_k0_m0_m1_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_aligned_space_size = math::integer_least_multiple(
+            b_block_desc_k0_n0_n1_k1.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block_double = p_shared_block;
+        FloatAB* p_b_block_double = p_shared_block + 2 * a_block_aligned_space_size;
+
+        // register allocation for output
+        auto c_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAcc>(
+            c_thread_desc_m10_m11_n10_n11.GetElementSpaceSize());
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0, 0);
+
+        auto a_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block_double, a_block_desc_k0_m0_m1_k1.GetElementSpaceSize());
+        auto b_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block_double, b_block_desc_k0_n0_n1_k1.GetElementSpaceSize());
+
+        auto a_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block_double + a_block_aligned_space_size,
+            a_block_desc_k0_m0_m1_k1.GetElementSpaceSize());
+        auto b_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block_double + b_block_aligned_space_size,
+            b_block_desc_k0_n0_n1_k1.GetElementSpaceSize());
+
+        // LDS double buffer: preload data into LDS
+        {
+            a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+            b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+            a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_even_buf);
+            b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_even_buf);
+        }
+
+        if constexpr(HasMainKBlockLoop)
+        {
+            const auto K0 = a_grid_desc_k0_m0_m1_k1.GetLength(I0);
+
+            index_t k_block_data_begin = 0;
+
+            // LDS double buffer: main body
+            // use Do-While loop instead of For loop to simplify control flow
+            do
+            {
+                // even iteration
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m0_m1_k1,
+                                                    a_block_slice_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n0_n1_k1,
+                                                    b_block_slice_copy_step);
+
+                // LDS doubel buffer: load next data from device mem
+                a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+                b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+                block_sync_lds();
+
+                // LDS double buffer: GEMM on current data
+                blockwise_gemm.Run(c_thread_desc_m10_m11_n10_n11,
+                                   a_block_even_buf,
+                                   b_block_even_buf,
+                                   c_thread_buf);
+
+                // LDS double buffer: store next data to LDS
+                a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_odd_buf);
+                b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_odd_buf);
+
+                // odd iteration
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m0_m1_k1,
+                                                    a_block_slice_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n0_n1_k1,
+                                                    b_block_slice_copy_step);
+
+                // LDS double buffer: load next data from device mem
+                a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+                b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+                block_sync_lds();
+
+                // LDS double buffer: GEMM on current data
+                blockwise_gemm.Run(
+                    c_thread_desc_m10_m11_n10_n11, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
+
+                // LDS double buffer: store next data to LDS
+                a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_even_buf);
+                b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_even_buf);
+
+                k_block_data_begin += 2 * K0PerBlock;
+            } while(k_block_data_begin < K0 - 2 * K0PerBlock);
+        }
+
+        // LDS double buffer: tail
+        if constexpr(HasDoubleTailKBlockLoop) // if has 2 iteration left
+        {
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m0_m1_k1, a_block_slice_copy_step);
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n0_n1_k1, b_block_slice_copy_step);
+
+            block_sync_lds();
+
+            // LDS double buffer: load last data from device mem
+            a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
+            b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
+
+            // LDS double buffer: GEMM on 2nd-last data
+            blockwise_gemm.Run(
+                c_thread_desc_m10_m11_n10_n11, a_block_even_buf, b_block_even_buf, c_thread_buf);
+
+            // LDS double buffer: store last data to LDS
+            a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_odd_buf);
+            b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_odd_buf);
+
+            block_sync_lds();
+
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(
+                c_thread_desc_m10_m11_n10_n11, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
+        }
+        else // if has 1 iteration left
+        {
+            __syncthreads();
+
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(
+                c_thread_desc_m10_m11_n10_n11, a_block_even_buf, b_block_even_buf, c_thread_buf);
+        }
+
+        // output: register to global memory
+        {
+            constexpr auto c_thread_desc_m0_m10_m11_n0_n10_n11 =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1,
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I0]>{},
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I1]>{},
+                               I1,
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I2]>{},
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I3]>{}));
+
+            const auto c_m10_m11_n10_n11_thread_origin_idx_on_block =
+                blockwise_gemm.CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
+                    get_thread_local_1d_id());
+
+            ThreadwiseTensorSliceTransfer_v1r3<
+                FloatAcc,
+                FloatC,
+                decltype(c_thread_desc_m0_m10_m11_n0_n10_n11),
+                decltype(c_grid_desc_m0_m10_m11_n0_n10_n11),
+                ck::tensor_operation::element_wise::PassThrough,
+                Sequence<1,
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I0],
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I1],
+                         1,
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I2],
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I3]>,
+                CThreadTransferSrcDstAccessOrder,
+                CThreadTransferSrcDstVectorDim,
+                CThreadTransferDstScalarPerVector,
+                CGlobalMemoryDataOperation,
+                1,
+                true>{c_grid_desc_m0_m10_m11_n0_n10_n11,
+                      make_multi_index(im0,
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I0],
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I1],
+                                       in0,
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I2],
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I3]),
+                      ck::tensor_operation::element_wise::PassThrough{}}
+                .Run(c_thread_desc_m0_m10_m11_n0_n10_n11,
+                     make_tuple(I0, I0, I0, I0, I0, I0),
+                     c_thread_buf,
+                     c_grid_desc_m0_m10_m11_n0_n10_n11,
+                     c_grid_buf);
+        }
+    }
+};
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_B_K0_M_K1,
+          typename BGridDesc_B_K0_N_K1,
+          typename CGridDesc_M_N,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t K1Value,
+          index_t M1PerThreadM111,
+          index_t N1PerThreadN111,
+          index_t KPerThread,
+          typename M11N11ThreadClusterM110Xs,
+          typename M11N11ThreadClusterN110Xs,
+          typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
+          typename ABlockTransferSrcVectorTensorContiguousDimOrder,
+          typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
+          typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
+          typename BBlockTransferSrcVectorTensorContiguousDimOrder,
+          typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector>
+struct GridwiseGemmDl_bkm_bkn_mn_v1r3
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // TODO: change this. I think it needs multi-dimensional alignment
+        constexpr auto max_lds_align = K1;
+
+        // TODO: check alignment
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_b_k0_m_k1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_b_k0_n_k1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_aligned_space_size = math::integer_least_multiple(
+            a_block_desc_b_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_aligned_space_size = math::integer_least_multiple(
+            b_block_desc_b_k0_n_k1.GetElementSpaceSize(), max_lds_align);
+
+        return 2 * (a_block_aligned_space_size + b_block_aligned_space_size) * sizeof(FloatAB);
+    }
+
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_B_K0_M_K1& a_grid_desc_b_k0_m_k1,
+                  const BGridDesc_B_K0_N_K1& b_grid_desc_b_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M      = a_grid_desc_b_k0_m_k1.GetLength(I2);
+        const auto N      = b_grid_desc_b_k0_n_k1.GetLength(I2);
+        const auto K0     = a_grid_desc_b_k0_m_k1.GetLength(I1);
+        const auto KBatch = a_grid_desc_b_k0_m_k1.GetLength(I0);
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+
+        return (M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+                K0 == b_grid_desc_b_k0_n_k1.GetLength(I1) &&
+                K1 == a_grid_desc_b_k0_m_k1.GetLength(I3) &&
+                K1 == b_grid_desc_b_k0_n_k1.GetLength(I3)) &&
+               KBatch == b_grid_desc_b_k0_n_k1.GetLength(I0) &&
+               (M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0);
+    }
+
+    __host__ __device__ static constexpr index_t CalculateGridSize(index_t M, index_t N)
+    {
+        const index_t grid_size = (M / MPerBlock) * (N / NPerBlock);
+
+        return grid_size;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K0)
+    {
+        const bool has_main_k_block_loop = (K0 + K0PerBlock) / (2 * K0PerBlock) > 1;
+
+        return has_main_k_block_loop;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasDoubleTailKBlockLoop(index_t K0)
+    {
+        const bool has_double_tail_k_block_loop = (K0 / K0PerBlock) % 2 == 0;
+
+        return has_double_tail_k_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeAGridDescriptor_B_K0_M0_M1_K1(const AGridDesc_B_K0_M_K1& a_grid_desc_b_k0_m_k1)
+    {
+        const auto KBatch = a_grid_desc_b_k0_m_k1.GetLength(I0);
+        const auto K0     = a_grid_desc_b_k0_m_k1.GetLength(I1);
+        const auto M      = a_grid_desc_b_k0_m_k1.GetLength(I2);
+
+        const auto M1 = Number<MPerBlock>{};
+        const auto M0 = M / M1;
+
+        const auto a_grid_desc_b_k0_m0_m1_k1 = transform_tensor_descriptor(
+            a_grid_desc_b_k0_m_k1,
+            make_tuple(make_pass_through_transform(KBatch),
+                       make_pass_through_transform(K0),
+                       make_unmerge_transform(make_tuple(M0, M1)),
+                       make_pass_through_transform(K1)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}));
+
+        return a_grid_desc_b_k0_m0_m1_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeBGridDescriptor_B_K0_N0_N1_K1(const BGridDesc_B_K0_N_K1& b_grid_desc_b_k0_n_k1)
+    {
+        const auto KBatch = b_grid_desc_b_k0_n_k1.GetLength(I0);
+        const auto K0     = b_grid_desc_b_k0_n_k1.GetLength(I1);
+        const auto N      = b_grid_desc_b_k0_n_k1.GetLength(I2);
+
+        const auto N1 = Number<NPerBlock>{};
+        const auto N0 = N / N1;
+
+        const auto b_grid_desc_b_k0_n0_n1_k1 = transform_tensor_descriptor(
+            b_grid_desc_b_k0_n_k1,
+            make_tuple(make_pass_through_transform(KBatch),
+                       make_pass_through_transform(K0),
+                       make_unmerge_transform(make_tuple(N0, N1)),
+                       make_pass_through_transform(K1)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}));
+
+        return b_grid_desc_b_k0_n0_n1_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        constexpr auto M11 =
+            Number<container_reduce(M11N11ThreadClusterM110Xs{}, math::multiplies{}, I1) *
+                   M1PerThreadM111>{};
+        constexpr auto N11 =
+            Number<container_reduce(M11N11ThreadClusterN110Xs{}, math::multiplies{}, I1) *
+                   N1PerThreadN111>{};
+
+        constexpr auto M10 = M1 / M11;
+        constexpr auto N10 = N1 / N11;
+
+        const auto c_grid_desc_m0_m10_m11_n0_n10_n11 = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(M0, M10, M11)),
+                       make_unmerge_transform(make_tuple(N0, N10, N11))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
+
+        return c_grid_desc_m0_m10_m11_n0_n10_n11;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
+        const CGridDesc_M_N& c_m_n_grid_desc, index_t M01, index_t N01, index_t KBatch)
+    {
+        return BlockToCTileMap_KSplit_M00_N00_M01_N01<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_m_n_grid_desc, M01, N01, KBatch);
+    }
+
+    using AGridDesc_B_K0_M0_M1_K1 =
+        decltype(MakeAGridDescriptor_B_K0_M0_M1_K1(AGridDesc_B_K0_M_K1{}));
+    using BGridDesc_B_K0_N0_N1_K1 =
+        decltype(MakeBGridDescriptor_B_K0_N0_N1_K1(BGridDesc_B_K0_N_K1{}));
+    using CGridDesc_M0_M10_M11_N0_N10_N11 =
+        decltype(MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
+    using CBlockClusterAdaptor = decltype(MakeCBlockClusterAdaptor(CGridDesc_M_N{}, 1, 1, 1));
+
+    template <bool HasMainKBlockLoop, bool HasDoubleTailKBlockLoop>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        FloatAB* __restrict__ p_shared_block,
+        const AGridDesc_B_K0_M0_M1_K1& a_grid_desc_b_k0_m0_m1_k1,
+        const BGridDesc_B_K0_N0_N1_K1& b_grid_desc_b_k0_n0_n1_k1,
+        const CGridDesc_M0_M10_M11_N0_N10_N11& c_grid_desc_m0_m10_m11_n0_n10_n11,
+        const CBlockClusterAdaptor& c_block_cluster_adaptor,
+        integral_constant<bool, HasMainKBlockLoop>,
+        integral_constant<bool, HasDoubleTailKBlockLoop>)
+    {
+        const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_b_k0_m0_m1_k1.GetElementSpaceSize());
+        const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_b_k0_n0_n1_k1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_m0_m10_m11_n0_n10_n11.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            c_block_cluster_adaptor.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t k_batch_id = block_work_idx[I0];
+
+        if(!c_block_cluster_adaptor.ValidCTileIndex(
+               make_tuple(block_work_idx[I1], block_work_idx[I2]),
+               make_tuple(c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I0),
+                          c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I3))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
+
+        const index_t n_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[I2]);
+
+        // TODO: change this. I think it needs multi-dimensional alignment
+        constexpr auto max_lds_align = K1;
+
+        // TODO: check alignment
+        // A matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto a_block_desc_b_k0_m0_m1_k1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(I1, Number<K0PerBlock>{}, I1, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto b_block_desc_b_k0_n0_n1_k1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(I1, Number<K0PerBlock>{}, I1, Number<NPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // A matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto a_block_desc_k0_m0_m1_k1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, I1, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, dst of blockwise copy
+        //   be careful of LDS alignment
+        constexpr auto b_block_desc_k0_n0_n1_k1 = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, I1, Number<NPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // A matrix in LDS memory, for blockwise GEMM
+        constexpr auto a_k0_m_k1_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+
+        // TODO: check alignment
+        // B matrix in LDS memory, for blockwise GEMM
+        constexpr auto b_k0_n_k1_block_desc = make_naive_tensor_descriptor_aligned(
+            make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+
+        static_assert(a_block_desc_k0_m0_m1_k1.GetElementSpaceSize() ==
+                          a_k0_m_k1_block_desc.GetElementSpaceSize() &&
+                      b_block_desc_k0_n0_n1_k1.GetElementSpaceSize() ==
+                          b_k0_n_k1_block_desc.GetElementSpaceSize() &&
+                      "wrong!");
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
+            BlockSize,
+            InMemoryDataOperationEnum::Set,
+            Sequence<1, K0PerBlock, 1, MPerBlock, K1.value>,
+            ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
+            ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
+            ABlockTransferThreadClusterArrangeOrder,
+            FloatAB,
+            FloatAB,
+            remove_reference_t<decltype(a_grid_desc_b_k0_m0_m1_k1)>,
+            decltype(a_block_desc_b_k0_m0_m1_k1),
+            ABlockTransferSrcAccessOrder,
+            Sequence<0, 1, 2, 3, 4>,
+            ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1, // SrcVectorTensorLengths
+            ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1, // DstVectorTensorLengths
+            ABlockTransferSrcVectorTensorContiguousDimOrder,  // SrcVectorTensorContiguousDimOrder
+            Sequence<0, 1, 2, 3, 4>,                          // DstVectorTensorContiguousDimOrder
+            false,
+            true>(a_grid_desc_b_k0_m0_m1_k1,
+                  make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0, 0),
+                  a_block_desc_b_k0_m0_m1_k1,
+                  make_multi_index(0, 0, 0, 0, 0));
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
+            BlockSize,
+            InMemoryDataOperationEnum::Set,
+            Sequence<1, K0PerBlock, 1, NPerBlock, K1.value>,
+            BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
+            BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
+            BBlockTransferThreadClusterArrangeOrder,
+            FloatAB,
+            FloatAB,
+            remove_reference_t<decltype(b_grid_desc_b_k0_n0_n1_k1)>,
+            decltype(b_block_desc_b_k0_n0_n1_k1),
+            BBlockTransferSrcAccessOrder,
+            Sequence<0, 1, 2, 3, 4>,
+            BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1, // SrcVectorTensorLengths
+            BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1, // DstVectorTensorLengths
+            BBlockTransferSrcVectorTensorContiguousDimOrder,  // SrcVectorTensorContiguousDimOrder
+            Sequence<0, 1, 2, 3, 4>,                          // DstVectorTensorContiguousDimOrder
+            false,
+            true>(b_grid_desc_b_k0_n0_n1_k1,
+                  make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0, 0),
+                  b_block_desc_b_k0_n0_n1_k1,
+                  make_multi_index(0, 0, 0, 0, 0));
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[KPerBlocl, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        const auto blockwise_gemm =
+            BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2<
+                BlockSize,
+                FloatAB,
+                FloatAB,
+                FloatAcc,
+                decltype(a_k0_m_k1_block_desc),
+                decltype(b_k0_n_k1_block_desc),
+                M1PerThreadM111,
+                N1PerThreadN111,
+                KPerThread,
+                M11N11ThreadClusterM110Xs,
+                M11N11ThreadClusterN110Xs,
+                M1PerThreadM111,
+                N1PerThreadN111>{};
+
+        constexpr auto c_m10_m11_n10_n11_thread_tensor_lengths =
+            decltype(blockwise_gemm)::GetCThreadTensorLengths_BM0_BM1_BN0_BN1();
+
+        constexpr auto c_thread_desc_m10_m11_n10_n11 = make_naive_tensor_descriptor_packed(
+            sequence_to_tuple_of_number(c_m10_m11_n10_n11_thread_tensor_lengths));
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_aligned_space_size = math::integer_least_multiple(
+            a_block_desc_k0_m0_m1_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_aligned_space_size = math::integer_least_multiple(
+            b_block_desc_k0_n0_n1_k1.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block_double = p_shared_block;
+        FloatAB* p_b_block_double = p_shared_block + 2 * a_block_aligned_space_size;
+
+        // register allocation for output
+        auto c_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAcc>(
+            c_thread_desc_m10_m11_n10_n11.GetElementSpaceSize());
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0, 0);
+
+        auto a_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block_double, a_block_desc_k0_m0_m1_k1.GetElementSpaceSize());
+        auto b_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block_double, b_block_desc_k0_n0_n1_k1.GetElementSpaceSize());
+
+        auto a_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block_double + a_block_aligned_space_size,
+            a_block_desc_k0_m0_m1_k1.GetElementSpaceSize());
+        auto b_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block_double + b_block_aligned_space_size,
+            b_block_desc_k0_n0_n1_k1.GetElementSpaceSize());
+
+        // LDS double buffer: preload data into LDS
+        {
+            a_blockwise_copy.RunRead(a_grid_desc_b_k0_m0_m1_k1, a_global_buf);
+            b_blockwise_copy.RunRead(b_grid_desc_b_k0_n0_n1_k1, b_global_buf);
+
+            a_blockwise_copy.RunWrite(a_block_desc_b_k0_m0_m1_k1, a_block_even_buf);
+            b_blockwise_copy.RunWrite(b_block_desc_b_k0_n0_n1_k1, b_block_even_buf);
+        }
+
+        if constexpr(HasMainKBlockLoop)
+        {
+            const auto K0 = a_grid_desc_b_k0_m0_m1_k1.GetLength(I1);
+
+            index_t k_block_data_begin = 0;
+
+            // LDS double buffer: main body
+            // use Do-While loop instead of For loop to simplify control flow
+            do
+            {
+                // even iteration
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_b_k0_m0_m1_k1,
+                                                    a_block_slice_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_b_k0_n0_n1_k1,
+                                                    b_block_slice_copy_step);
+
+                // LDS double buffer: load next data from device mem
+                a_blockwise_copy.RunRead(a_grid_desc_b_k0_m0_m1_k1, a_global_buf);
+                b_blockwise_copy.RunRead(b_grid_desc_b_k0_n0_n1_k1, b_global_buf);
+
+                block_sync_lds();
+
+                // LDS double buffer: GEMM on current data
+                blockwise_gemm.Run(c_thread_desc_m10_m11_n10_n11,
+                                   a_block_even_buf,
+                                   b_block_even_buf,
+                                   c_thread_buf);
+
+                // LDS double buffer: store next data to LDS
+                a_blockwise_copy.RunWrite(a_block_desc_b_k0_m0_m1_k1, a_block_odd_buf);
+                b_blockwise_copy.RunWrite(b_block_desc_b_k0_n0_n1_k1, b_block_odd_buf);
+
+                // odd iteration
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_b_k0_m0_m1_k1,
+                                                    a_block_slice_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_b_k0_n0_n1_k1,
+                                                    b_block_slice_copy_step);
+
+                // LDS doubel buffer: load next data from device mem
+                a_blockwise_copy.RunRead(a_grid_desc_b_k0_m0_m1_k1, a_global_buf);
+                b_blockwise_copy.RunRead(b_grid_desc_b_k0_n0_n1_k1, b_global_buf);
+
+                block_sync_lds();
+
+                // LDS double buffer: GEMM on current data
+                blockwise_gemm.Run(
+                    c_thread_desc_m10_m11_n10_n11, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
+
+                // LDS double buffer: store next data to LDS
+                a_blockwise_copy.RunWrite(a_block_desc_b_k0_m0_m1_k1, a_block_even_buf);
+                b_blockwise_copy.RunWrite(b_block_desc_b_k0_n0_n1_k1, b_block_even_buf);
+
+                k_block_data_begin += 2 * K0PerBlock;
+            } while(k_block_data_begin < K0 - 2 * K0PerBlock);
+        }
+
+        // LDS double buffer: tail
+        if constexpr(HasDoubleTailKBlockLoop) // if has 2 iteration left
+        {
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_b_k0_m0_m1_k1, a_block_slice_copy_step);
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_b_k0_n0_n1_k1, b_block_slice_copy_step);
+
+            block_sync_lds();
+
+            // LDS double buffer: load last data from device mem
+            a_blockwise_copy.RunRead(a_grid_desc_b_k0_m0_m1_k1, a_global_buf);
+            b_blockwise_copy.RunRead(b_grid_desc_b_k0_n0_n1_k1, b_global_buf);
+
+            // LDS double buffer: GEMM on 2nd-last data
+            blockwise_gemm.Run(
+                c_thread_desc_m10_m11_n10_n11, a_block_even_buf, b_block_even_buf, c_thread_buf);
+
+            // LDS double buffer: store last data to LDS
+            a_blockwise_copy.RunWrite(a_block_desc_b_k0_m0_m1_k1, a_block_odd_buf);
+            b_blockwise_copy.RunWrite(b_block_desc_b_k0_n0_n1_k1, b_block_odd_buf);
+
+            block_sync_lds();
+
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(
+                c_thread_desc_m10_m11_n10_n11, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
+        }
+        else // if has 1 iteration left
+        {
+            __syncthreads();
+
+            // LDS double buffer: GEMM on last data
+            blockwise_gemm.Run(
+                c_thread_desc_m10_m11_n10_n11, a_block_even_buf, b_block_even_buf, c_thread_buf);
+        }
+
+        // output: register to global memory
+        {
+            constexpr auto c_thread_desc_m0_m10_m11_n0_n10_n11 =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1,
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I0]>{},
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I1]>{},
+                               I1,
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I2]>{},
+                               Number<c_m10_m11_n10_n11_thread_tensor_lengths[I3]>{}));
+
+            const auto c_m10_m11_n10_n11_thread_origin_idx_on_block =
+                blockwise_gemm.CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
+                    get_thread_local_1d_id());
+
+            ThreadwiseTensorSliceTransfer_v1r3<
+                FloatAcc,
+                FloatC,
+                decltype(c_thread_desc_m0_m10_m11_n0_n10_n11),
+                decltype(c_grid_desc_m0_m10_m11_n0_n10_n11),
+                ck::tensor_operation::element_wise::PassThrough,
+                Sequence<1,
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I0],
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I1],
+                         1,
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I2],
+                         c_m10_m11_n10_n11_thread_tensor_lengths[I3]>,
+                CThreadTransferSrcDstAccessOrder,
+                CThreadTransferSrcDstVectorDim,
+                CThreadTransferDstScalarPerVector,
+                CGlobalMemoryDataOperation,
+                1,
+                true>{c_grid_desc_m0_m10_m11_n0_n10_n11,
+                      make_multi_index(m_block_data_idx_on_grid,
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I0],
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I1],
+                                       n_block_data_idx_on_grid,
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I2],
+                                       c_m10_m11_n10_n11_thread_origin_idx_on_block[I3]),
+                      ck::tensor_operation::element_wise::PassThrough{}}
+                .Run(c_thread_desc_m0_m10_m11_n0_n10_n11,
+                     make_tuple(I0, I0, I0, I0, I0, I0),
+                     c_thread_buf,
+                     c_grid_desc_m0_m10_m11_n0_n10_n11,
+                     c_grid_buf);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..d710fc1894
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
@@ -0,0 +1,945 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename DsDataType,
+          typename FloatE,
+          typename FloatReduceAcc,
+          typename RsDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename QsElementwiseOperation,
+          typename RsElementwiseOperation,
+          typename ThreadReduceOperations,
+          InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+          typename RsGlobalMemoryDataOperation,
+          typename AGridDesc_M_K,
+          typename BGridDesc_N_K,
+          typename EGridDesc_M_N,
+          typename RGridDesc_M,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDRThreadTransferClusterLengths_MPerBlock_NPerBlock,
+          index_t CDEReduceThreadTransferScalarPerVector_NPerBlock,
+          index_t RThreadTransferDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmMultipleDMultipleR_k0mk1_k0nk1_mn_xdl_cshuffle_v1
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+    static constexpr index_t NumRTensor = RsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // ck::Tuple<const T0DataType*, const T1DataType*, ...>
+    template <typename Ts, bool isConst = true>
+    static constexpr auto MakeTsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using T = remove_cvref_t<tuple_element_t<i.value, Ts>>;
+                if constexpr(isConst)
+                    return static_cast<const T*>(nullptr);
+                else
+                    return static_cast<T*>(nullptr);
+            },
+            Number<Ts::Size()>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // A desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultBGridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2ETileMap>
+    __host__ __device__ static constexpr bool CheckValidity(const AGridDesc_M_K& a_grid_desc_m_k,
+                                                            const BGridDesc_N_K& b_grid_desc_n_k,
+                                                            const EGridDesc_M_N& e_grid_desc_m_n,
+                                                            const RGridDesc_M& r_grid_desc_m,
+                                                            const Block2ETileMap& block_2_etile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        static_assert(AGridDesc_M_K::GetNumOfDimension() == 2);
+        static_assert(BGridDesc_N_K::GetNumOfDimension() == 2);
+        static_assert(EGridDesc_M_N::GetNumOfDimension() == 2);
+
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+            return false;
+
+        if(M != r_grid_desc_m.GetLength(I0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_etile_map.CheckValidity(e_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeRGridDescriptor_MBlock_MPerBlock(const RGridDesc_M& r_grid_desc_m)
+    {
+        const auto M      = r_grid_desc_m.GetLength(I0);
+        const auto MBlock = M / MPerBlock;
+
+        const auto r_grid_desc_mblock_mperblock = transform_tensor_descriptor(
+            r_grid_desc_m,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{}))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1>{}));
+
+        return r_grid_desc_mblock_mperblock;
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    using DefaultAGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using DefaultBGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+
+    // Support 2 dimension in the future. Not only M
+    using RGridDescriptor_MBlock_MPerBlock =
+        remove_cvref_t<decltype(MakeRGridDescriptor_MBlock_MPerBlock(RGridDesc_M{}))>;
+
+    using DefaultBlock2ETileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    using DsGridPointer = decltype(MakeTsGridPointer<DsDataType, true>());
+    using RsGridPointer = decltype(MakeTsGridPointer<RsDataType, false>());
+
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AK0_M_AK1,
+              typename BGridDesc_BK0_N_BK1,
+              typename Block2ETileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        DsGridPointer p_ds_grid,
+        FloatE* __restrict__ p_e_grid,
+        RsGridPointer p_rs_grid,
+        void* __restrict__ p_shared,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op,
+        const QsElementwiseOperation& qs_element_op,
+        const RsElementwiseOperation& rs_element_op,
+        const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+        const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+        const StaticallyIndexedArray<EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                                     NumDTensor>&
+            ds_grid_desc_mblock_mperblock_nblock_nperblock, // FIXME: Ds desc may be of different
+        const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            e_grid_desc_mblock_mperblock_nblock_nperblock,
+        const StaticallyIndexedArray<RGridDescriptor_MBlock_MPerBlock,
+                                     NumRTensor>&
+            rs_grid_desc_mblock_mperblock, // FIXME: Rs desc may be of different
+        const Block2ETileMap& block_2_etile_map)
+    {
+        // FIXME - Share code with other gemm kernel
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        auto rs_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_rs_grid(i), rs_grid_desc_mblock_mperblock[i].GetElementSpaceSize());
+            },
+            Number<NumRTensor>{});
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_etile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C + Ds + reduction + write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_der_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            // TODO: this should be implemented as a blockwise reduction
+            // LDS c_reduce_block_desc_mperblock_nperblock
+            constexpr auto c_reduce_block_desc_mperblock_nperblock = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1)),
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I3))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<>{}, Sequence<1>{}));
+
+            static_assert(CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I0) *
+                                  CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                              BlockSize,
+                          "wrong!");
+
+            static_assert((CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) %
+                                      CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I0) ==
+                                  0 &&
+                              (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) %
+                                      CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                                  0,
+                          "wrong!");
+
+            constexpr index_t mreduce_per_thread =
+                (CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) /
+                CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I0);
+
+            constexpr index_t nreduce_per_thread =
+                (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) /
+                CDRThreadTransferClusterLengths_MPerBlock_NPerBlock::At(I1);
+
+            constexpr auto c_reduce_thread_lengths_mperblock_nperblock =
+                Sequence<mreduce_per_thread, nreduce_per_thread>{};
+
+            // VGPR cde_reduce_thread_desc_mperblock_nperblock
+            constexpr auto cde_reduce_thread_desc_mperblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(Number<mreduce_per_thread>{}, Number<nreduce_per_thread>{}));
+
+            constexpr auto r_thread_desc_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(Number<mreduce_per_thread>{}));
+
+            constexpr auto r_thread_desc_mblock_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(I1, Number<mreduce_per_thread>{}));
+
+            auto e_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                cde_reduce_thread_desc_mperblock_nperblock.GetElementSpaceSize());
+
+            // reduce: threadwise copy from LDS to VGPR
+            constexpr auto c_reduce_thread_cluster_desc = make_cluster_descriptor(
+                CDRThreadTransferClusterLengths_MPerBlock_NPerBlock{}, Sequence<1, 0>{});
+
+            const auto c_reduce_thread_cluster_idx =
+                c_reduce_thread_cluster_desc.CalculateBottomIndex(
+                    make_multi_index(get_thread_local_1d_id()));
+
+            const auto c_reduce_thread_data_idx_begin =
+                c_reduce_thread_cluster_idx * c_reduce_thread_lengths_mperblock_nperblock;
+
+            // To apply D0, D1, ... and reduction.
+            // Copy c shuffle from LDS back to VGPR
+            auto c_reduce_thread_copy_lds_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatCShuffle,
+                FloatReduceAcc,
+                decltype(c_reduce_block_desc_mperblock_nperblock),
+                decltype(cde_reduce_thread_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_lengths_mperblock_nperblock),
+                Sequence<0, 1>,
+                1,
+                CDEReduceThreadTransferScalarPerVector_NPerBlock,
+                1,
+                true>{c_reduce_block_desc_mperblock_nperblock, c_reduce_thread_data_idx_begin};
+
+            // Copy result of reduction back from VGPR to global
+            auto reduce_tuple_thread_copy_vgpr_to_global = generate_tuple(
+                [&](auto I) {
+                    auto p_r_grid                     = p_rs_grid[I];
+                    auto r_element_op                 = rs_element_op[I];
+                    auto r_grid_desc_mblock_mperblock = rs_grid_desc_mblock_mperblock[I];
+
+                    return ThreadwiseTensorSliceTransfer_v1r3<
+                        FloatReduceAcc,
+                        remove_pointer_t<decltype(p_r_grid)>,
+                        decltype(r_thread_desc_mblock_mperblock),
+                        decltype(r_grid_desc_mblock_mperblock),
+                        decltype(r_element_op),
+                        Sequence<1, mreduce_per_thread>,
+                        Sequence<0, 1>,
+                        1,
+                        RThreadTransferDstScalarPerVector_MPerBlock,
+                        RsGlobalMemoryDataOperation::At(I),
+                        1,
+                        false>{r_grid_desc_mblock_mperblock,
+                               make_multi_index(block_work_idx[I0],                  // mblock
+                                                c_reduce_thread_data_idx_begin[I0]), // mperblock
+                               r_element_op};
+                },
+                Number<NumRTensor>{});
+
+            // D0, D1, ..., Dn
+            constexpr auto cde_reduce_thread_desc_I1_mperblock_I1_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mreduce_per_thread>{}, I1, Number<nreduce_per_thread>{}));
+
+            // FIXME: Decrease usage of VGPR
+            // Apply pointwise lambda function from multi-source (Global and LDS) into VGPR
+            auto ds_thread_buf = generate_tuple(
+                [&](auto) {
+                    return make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                        cde_reduce_thread_desc_I1_mperblock_I1_nperblock.GetElementSpaceSize());
+                },
+                Number<NumDTensor>{});
+
+            // Copy D0, D1, ..., Dn from global to VGPR
+            auto ds_thread_copy_global_to_vgpr = generate_tuple(
+                [&](auto I) {
+                    using DDataType = remove_cvref_t<tuple_element_t<I.value, DsDataType>>;
+                    return ThreadwiseTensorSliceTransfer_v2<
+                        DDataType,
+                        FloatReduceAcc,
+                        decltype(ds_grid_desc_mblock_mperblock_nblock_nperblock[I]),
+                        decltype(cde_reduce_thread_desc_I1_mperblock_I1_nperblock),
+                        Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>,
+                        Sequence<0, 1, 2, 3>,
+                        3,
+                        CDEReduceThreadTransferScalarPerVector_NPerBlock,
+                        1,
+                        true>(ds_grid_desc_mblock_mperblock_nblock_nperblock[I],
+                              make_multi_index(
+                                  I0,
+                                  m_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I0],
+                                  I0,
+                                  n_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I1]));
+                },
+                Number<NumDTensor>{});
+
+            auto e_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
+                FloatReduceAcc,
+                FloatE,
+                decltype(cde_reduce_thread_desc_I1_mperblock_I1_nperblock),
+                decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                tensor_operation::element_wise::PassThrough,
+                Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>, // SliceLengths
+                Sequence<0, 1, 2, 3>,                                     // DimAccessOrder
+                3,                                                        // DstVectorDim
+                CDEReduceThreadTransferScalarPerVector_NPerBlock,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                make_multi_index(I0,
+                                 m_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I0],
+                                 I0,
+                                 n_block_data_idx_on_grid + c_reduce_thread_data_idx_begin[I1]),
+                tensor_operation::element_wise::PassThrough{}};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_der_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each thread shuffle data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // Get shuffle data from LDS to VGPR
+                c_reduce_thread_copy_lds_to_vgpr.Run(c_reduce_block_desc_mperblock_nperblock,
+                                                     c_shuffle_block_buf,
+                                                     cde_reduce_thread_desc_mperblock_nperblock,
+                                                     make_tuple(I0, I0),
+                                                     e_thread_buf);
+
+                // Global read D0, D1, ...
+                static_for<0, NumDTensor, 1>{}([&](auto Id) {
+                    auto& d_thread_copy_global_to_vgpr = ds_thread_copy_global_to_vgpr(Id);
+                    d_thread_copy_global_to_vgpr.Run(
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock[Id],
+                        ds_grid_buf[Id],
+                        cde_reduce_thread_desc_I1_mperblock_I1_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        ds_thread_buf(Id));
+
+                    if constexpr(access_id < num_access - 1)
+                    {
+                        // move on D0, D1, ...
+                        constexpr auto de_global_step = sfc_der_global.GetForwardStep(access_id);
+                        d_thread_copy_global_to_vgpr.MoveSrcSliceWindow(
+                            ds_grid_desc_mblock_mperblock_nblock_nperblock[Id], de_global_step);
+                    }
+                });
+
+                // cde_element_op(e, c, d0, d1, ...);
+                static_for<0, cde_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                    [&](auto i) {
+                        const auto c_ds_src_data_refs = concat_tuple_of_reference(
+                            tie(e_thread_buf[i]),
+                            generate_tie(
+                                [&](auto Id) -> const auto& { return ds_thread_buf[Id][i]; },
+                                Number<NumDTensor>{}));
+                        auto e_dst_data_refs = tie(e_thread_buf(i));
+                        unpack2(cde_element_op, e_dst_data_refs, c_ds_src_data_refs);
+                    });
+
+                // Global write E
+                e_thread_copy_vgpr_to_global.Run(cde_reduce_thread_desc_I1_mperblock_I1_nperblock,
+                                                 make_tuple(I0, I0, I0, I0),
+                                                 e_thread_buf,
+                                                 e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                 e_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    // move on E
+                    constexpr auto de_global_step = sfc_der_global.GetForwardStep(access_id);
+                    e_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                        e_grid_desc_mblock_mperblock_nblock_nperblock, de_global_step);
+                }
+
+                // reduction
+                static_for<0, NumRTensor, 1>{}([&](auto Ir) {
+                    auto r_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                        r_thread_desc_mperblock.GetElementSpaceSize());
+
+                    auto& reduce_thread_copy_vgpr_to_global =
+                        reduce_tuple_thread_copy_vgpr_to_global(Ir);
+
+                    using ThreadReduceOperation =
+                        remove_cvref_t<decltype(ThreadReduceOperations{}[Ir])>;
+
+                    using ThreadwiseReduce =
+                        ThreadwiseReduction<FloatReduceAcc,
+                                            decltype(cde_reduce_thread_desc_mperblock_nperblock),
+                                            decltype(r_thread_desc_mperblock),
+                                            ThreadReduceOperation,
+                                            false>;
+
+                    // threadwise reduction
+                    const auto reduce_identityVal =
+                        ThreadReduceOperation::template GetIdentityValue<FloatReduceAcc>();
+                    static_for<0, mreduce_per_thread, 1>{}(
+                        [&](auto I) { r_thread_buf(I) = reduce_identityVal; });
+                    static_for<0, mreduce_per_thread, 1>{}([&](auto im) {
+                        static_for<0, nreduce_per_thread, 1>{}([&](auto in) {
+                            constexpr auto offset =
+                                Number<cde_reduce_thread_desc_mperblock_nperblock.CalculateOffset(
+                                    make_tuple(im, in))>{};
+
+                            qs_element_op[Ir](e_thread_buf(offset), e_thread_buf(offset));
+                        });
+                    });
+                    ThreadwiseReduce::Reduce(e_thread_buf, r_thread_buf);
+
+                    // gridwise reduction
+                    reduce_thread_copy_vgpr_to_global.Run(r_thread_desc_mblock_mperblock,
+                                                          make_tuple(I0, I0),
+                                                          r_thread_buf,
+                                                          rs_grid_desc_mblock_mperblock[Ir],
+                                                          rs_grid_buf(Ir));
+
+                    if constexpr(access_id < num_access - 1)
+                    {
+                        // move on R0, R1, ...
+                        constexpr auto de_global_step = sfc_der_global.GetForwardStep(access_id);
+                        reduce_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                            rs_grid_desc_mblock_mperblock[Ir],
+                            make_tuple(de_global_step[I0], de_global_step[I1]));
+                    }
+                });
+            }); // copy c, d, e + reduction
+
+        } // shuffle C + Ds + reduction + write out
+    }     // Run
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp
new file mode 100644
index 0000000000..98ade85a30
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp
@@ -0,0 +1,958 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseOp,
+          typename ADataType,
+          typename BDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2CTileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_conv_fwd_multiple_d_wmma_cshuffle(
+            const ADataType* __restrict__ p_a_grid,
+            const BDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const index_t batch_count,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock_,
+            const Block2CTileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__) || defined(__gfx1101__) || \
+    defined(__gfx1102__))
+    // offset base pointer for each work-group
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    __shared__ char p_shared[GridwiseOp::GetSharedMemoryNumberOfByte()];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    GridwiseOp::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                p_b_grid + b_batch_offset,
+                                                p_ds_grid_grp,
+                                                p_e_grid + e_batch_offset,
+                                                p_shared,
+                                                a_grid_desc_k0_m_k1,
+                                                b_grid_desc_k0_n_k1,
+                                                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                                a_element_op,
+                                                b_element_op,
+                                                cde_element_op,
+                                                block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+#endif
+}
+
+template <typename GridwiseOp,
+          typename ADataType,
+          typename BDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename ComputePtrOffsetOfBatch,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_contraction_multiple_d_wmma_cshuffle(
+            const ADataType* __restrict__ p_a_grid,
+            const BDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const index_t batch_count,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
+            const Block2CTileMap block_2_etile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__) || defined(__gfx1101__) || \
+    defined(__gfx1102__))
+    // printf("entry kernel launch");
+    __shared__ char p_shared[GridwiseOp::GetSharedMemoryNumberOfByte()];
+
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    // printf("before compute_ptr_offset call");
+    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+
+    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    DsPointer p_ds_grid_grp;
+
+    // printf("before allocate pointer d");
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    // printf("before entry");
+
+    GridwiseOp::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
+                                                p_b_grid + b_batch_offset,
+                                                p_ds_grid_grp,
+                                                p_e_grid + e_batch_offset,
+                                                p_shared,
+                                                a_grid_desc_k0_m_k1,
+                                                b_grid_desc_k0_n_k1,
+                                                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                a_element_op,
+                                                b_element_op,
+                                                cde_element_op,
+                                                block_2_etile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = block_2_etile_map;
+    ignore = compute_ptr_offset_of_batch;
+#endif
+}
+
+template <typename GridwiseOp,
+          typename ADataType,
+          typename BDataType,
+          typename DsPointer,
+          typename EDataType,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_mupltipe_d_wmma_cshuffle(
+            const ADataType* __restrict__ p_a_grid,
+            const BDataType* __restrict__ p_b_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__) || defined(__gfx1101__) || \
+    defined(__gfx1102__))
+    __shared__ char p_shared[GridwiseOp::GetSharedMemoryNumberOfByte()];
+
+    GridwiseOp::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                p_b_grid,
+                                                p_ds_grid,
+                                                p_e_grid,
+                                                p_shared,
+                                                a_grid_desc_k0_m_k1,
+                                                b_grid_desc_k0_n_k1,
+                                                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                a_element_op,
+                                                b_element_op,
+                                                cde_element_op,
+                                                block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx1100__ ))
+}
+
+template < // DataType Family
+    typename ADataType,
+    typename BDataType,
+    typename AccDataType,
+    typename CShuffleDataType,
+    typename DsDataType,
+    typename EDataType,
+    // InMemory Data Descriptor
+    typename AGridDesc_K0_M_K1,
+    typename BGridDesc_K0_N_K1,
+    typename DsGridDesc_M_N,
+    typename EGridDesc_M_N,
+    // ElementwiseOp Family
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CDEElementwiseOperation,
+    InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+    // Tiling Family
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t K0PerBlock,
+    index_t MPerWmma,
+    index_t NPerWmma,
+    index_t K1Value,
+    index_t MRepeat,
+    index_t NRepeat,
+    // ThreadCluster Family
+    index_t BlockSize,
+    typename ABlockTransferThreadClusterLengths_K0_M_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    index_t ABlockTransferSrcVectorDim,
+    index_t ABlockTransferSrcScalarPerVector,
+    index_t ABlockTransferDstScalarPerVector_K1,
+    bool AThreadTransferSrcResetCoordinateAfterRun,
+    bool ABlockLdsExtraM,
+    typename BBlockTransferThreadClusterLengths_K0_N_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    index_t BBlockTransferSrcVectorDim,
+    index_t BBlockTransferSrcScalarPerVector,
+    index_t BBlockTransferDstScalarPerVector_K1,
+    bool BThreadTransferSrcResetCoordinateAfterRun,
+    bool BBlockLdsExtraN,
+    index_t CShuffleMRepeatPerShuffle,
+    index_t CShuffleNRepeatPerShuffle,
+    typename CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+    index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+    index_t NumGemmKPrefetchStage = 1,
+    LoopScheduler LoopSched       = make_default_loop_scheduler(),
+    PipelineVersion PipelineVer   = PipelineVersion::v1>
+struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_wmma_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0perblock_mperblock_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return a_block_desc_k0perblock_mperblock_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0perblock_nperblock_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return b_block_desc_k0perblock_nperblock_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    // *Caution Here repeat is shuffle repeat
+    GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat()
+    {
+        constexpr index_t MWave = MPerBlock / (MRepeat * MPerWmma);
+        constexpr index_t NWave = NPerBlock / (NRepeat * NPerWmma);
+
+        constexpr auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMRepeatPerShuffle * MWave * MPerWmma>{},
+                           I1,
+                           Number<CShuffleNRepeatPerShuffle * NWave * NPerWmma>{}));
+
+        return c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0perblock_mperblock_k1 =
+            GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        constexpr auto b_block_desc_k0perblock_nperblock_k1 =
+            GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        constexpr auto cshuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
+            GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat();
+
+        constexpr auto max_lds_align = K1;
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_k0perblock_nperblock_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto c_block_space_size_aligned = math::integer_least_multiple(
+            cshuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat.GetElementSpaceSize(),
+            max_lds_align);
+
+        return math::max((a_block_space_size_aligned * sizeof(ADataType) +
+                          b_block_space_size_aligned * sizeof(BDataType)),
+                         c_block_space_size_aligned * sizeof(CShuffleDataType));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const DsGridDesc_M_N& ds_grid_desc_m_n,
+                  const EGridDesc_M_N& e_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerWmma * MRepeat) == 0) &&
+                          (NPerBlock % (NRepeat * NPerWmma)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
+                              N == ds_grid_desc_m_n[i].GetLength(I1));
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1) &&
+             K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+             K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K0 / K0PerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(e_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_k0_m_k1.GetElementSpaceSize() * sizeof(ADataType) <= TwoGB &&
+             b_grid_desc_k0_n_k1.GetElementSpaceSize() * sizeof(BDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / (K0PerBlock * K1);
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDesc_M_N_>
+    __host__ __device__ static constexpr auto
+    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N_& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDesc_M_N_>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const DsGridDesc_M_N_& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap(
+        const EGridDesc_M_N& e_grid_desc_m_n, index_t /* M01 */, index_t /* N01 */)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    using DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
+    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(EGridDesc_M_N{}, 1, 1))>;
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
+    __device__ static void Run(const ADataType* __restrict__ p_a_grid,
+                               const BDataType* __restrict__ p_b_grid,
+                               DsGridPointer p_ds_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                               const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                               const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const Block2CTileMap& block_2_ctile_map)
+    {
+        // printf("safe entry");
+        // clang-format off
+/*******************************************************************************/
+// Memory buffer zone.
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor>{});
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+/*******************************************************************************/
+// BlockIdx.x -> [BlockId.m, BlockId.n]
+        const auto block_work_idx = block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        { return; }
+
+        // Store BlockId into SGPR
+        const index_t m_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+        const index_t n_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+/*******************************************************************************/
+// BlockLevel, A/B Matrix ThreadMapping in LDS, As Destinaion of BlockWise_Copy
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+        constexpr auto max_lds_align = K1;
+        constexpr auto a_block_desc_k0perblock_mperblock_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+        constexpr auto b_block_desc_k0perblock_nperblock_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<        ThisThreadBlock,
+/* typename SrcElementwiseOperation,              */    AElementwiseOperation,
+/* typename DstElementwiseOperation,              */    ck::tensor_operation::element_wise::PassThrough,
+/* InMemoryDataOperationEnum DstInMemOp,          */    InMemoryDataOperationEnum::Set,
+/* typename BlockSliceLengths,                    */    Sequence<K0PerBlock, MPerBlock, K1>,
+/* typename ThreadClusterLengths,                 */    ABlockTransferThreadClusterLengths_K0_M_K1,
+/* typename ThreadClusterArrangeOrder,            */    ABlockTransferThreadClusterArrangeOrder,
+/* typename SrcData,                              */    ADataType,
+/* typename DstData,                              */    ADataType,
+/* typename SrcDesc,                              */    decltype(a_grid_desc_k0_m_k1),
+/* typename DstDesc,                              */    decltype(a_block_desc_k0perblock_mperblock_k1),
+/* typename SrcDimAccessOrder,                    */    ABlockTransferSrcAccessOrder,
+/* typename DstDimAccessOrder,                    */    Sequence<0, 1, 2>,
+/* index_t SrcVectorDim,                          */    ABlockTransferSrcVectorDim,
+/* index_t DstVectorDim,                          */    2,
+/* index_t SrcScalarPerVector,                    */    ABlockTransferSrcScalarPerVector,
+/* index_t DstScalarPerVector,                    */    ABlockTransferDstScalarPerVector_K1,
+/* index_t SrcScalarStrideInVector,               */    1,
+/* index_t DstScalarStrideInVector,               */    1,
+/* bool ThreadTransferSrcResetCoordinateAfterRun, */    AThreadTransferSrcResetCoordinateAfterRun,
+/* bool ThreadTransferDstResetCoordinateAfterRun, */    true>(
+                a_grid_desc_k0_m_k1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_k0perblock_mperblock_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<K0PerBlock, NPerBlock, K1>,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                BDataType,
+                                                BDataType,
+                                                decltype(b_grid_desc_k0_n_k1),
+                                                decltype(b_block_desc_k0perblock_nperblock_k1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<0, 1, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                b_grid_desc_k0_n_k1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_k0perblock_nperblock_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+/*******************************************************************************/
+        // GEMM
+        constexpr auto WmmaK = 16;
+        constexpr auto KPack = math::integer_least_multiple(K1, WmmaK);
+
+        auto blockwise_gemm =
+            BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle<BlockSize,
+                                                         ADataType,
+                                                         BDataType,
+                                                         AccDataType,
+                                                         decltype(a_block_desc_k0perblock_mperblock_k1),
+                                                         decltype(b_block_desc_k0perblock_nperblock_k1),
+                                                         MPerWmma,
+                                                         NPerWmma,
+                                                         MRepeat,
+                                                         NRepeat,
+                                                         KPack>{};
+
+        // Prepare Register for C matrix
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+/*******************************************************************************/
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize(), max_lds_align);
+        // LDS allocation for A and B: be careful of alignment
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(static_cast<ADataType*>(p_shared), a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize());
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(static_cast<BDataType*>(p_shared) + a_block_space_size_aligned, b_block_desc_k0perblock_nperblock_k1.GetElementSpaceSize());
+        
+        // Shift Per SUB_K
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+
+        // gridwise GEMM pipeline
+        const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
+        GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc_k0_m_k1,
+                                                          a_block_desc_k0perblock_mperblock_k1,
+                                                          a_blockwise_copy,
+                                                          a_grid_buf,
+                                                          a_block_buf,
+                                                          a_block_slice_copy_step,
+                                                          b_grid_desc_k0_n_k1,
+                                                          b_block_desc_k0perblock_nperblock_k1,
+                                                          b_blockwise_copy,
+                                                          b_grid_buf,
+                                                          b_block_buf,
+                                                          b_block_slice_copy_step,
+                                                          blockwise_gemm,
+                                                          c_thread_buf,
+                                                          K0BlockMainLoop);
+/*******************************************************************************/
+        // write out to C, implement shuffle
+        {
+            constexpr auto c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs =  
+            blockwise_gemm.GetCThreadDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
+
+            // This API Provide All dimension (size) you need
+            constexpr auto c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp =
+                blockwise_gemm.GetCBlockDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
+
+            constexpr auto MWave              = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I1);
+            constexpr auto MSubGroup          = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I2);
+            constexpr auto NWave              = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I4);
+            constexpr auto NThreadPerSubGroup = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I5);
+            constexpr auto MAccVgprs          = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I6);
+
+            // LDS descriptor, shuffle and write out in MRepeat x NRepeat times
+            constexpr auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
+                GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs = transform_tensor_descriptor(
+                c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMRepeatPerShuffle>{}, // MRepeat per shuffle repeat
+                        MWave,                               // MWave
+                        MSubGroup,                           // MSubGroup * MAccVgprs = MPerWmma
+                        MAccVgprs)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNRepeatPerShuffle>{}, // NRepeat per shuffle repeat
+                        NWave,                               // NWave
+                        NThreadPerSubGroup))),               // NThreadPerSubGroup = NPerWmma
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0, 1, 2, 6>{}, Sequence<>{}, Sequence<3, 4, 5>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block = blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_mrepeat_mwave_msubgroup_maccvgprs_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(MRepeat, MWave, MSubGroup, MAccVgprs))),
+                make_tuple(Sequence<0, 1, 2, 3>{}),
+                make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(NRepeat, NWave, NThreadPerSubGroup))),
+                make_tuple(Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0>{}));
+            
+            const auto m_thread_data_on_block_idx = m_thread_data_on_block_to_mrepeat_mwave_msubgroup_maccvgprs_adaptor.CalculateBottomIndex(
+                make_multi_index(m_thread_data_on_block));
+            
+            const auto n_thread_data_on_block_idx = n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_adaptor.CalculateBottomIndex(
+                make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
+                                                   decltype(c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMRepeatPerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            CShuffleNRepeatPerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            MAccVgprs>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6>,
+                                                   6,
+                                                   1, // vector write pixel
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
+                    make_multi_index(0,
+                                     m_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     0,
+                                     n_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+            
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor>{}));
+            
+            // tuple of reference to C/Ds tensor buffers
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
+                    },
+                    Number<NumDTensor>{}));
+
+            // shuffle: blockwise copy C from LDS to global
+            auto cde_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v7<
+                ThisThreadBlock,            // ThreadGroup
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CDEElementwiseOperation,    // ElementwiseOperation,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMRepeatPerShuffle * MWave * MPerWmma,
+                         1,
+                         CShuffleNRepeatPerShuffle * NWave * NPerWmma>, // BlockSliceLengths,
+                CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor,
+                                           false>>,   // bool ThreadTransferSrcResetCoordinateAfterRun,
+                Sequence<false>> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0)),
+                 cde_element_op};
+
+            // space filling curve for local reg & global memory
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MRepeat, 1, 1, NRepeat, 1, 1, MAccVgprs>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6>,
+                                  Sequence<CShuffleMRepeatPerShuffle,
+                                           1,
+                                           1,
+                                           CShuffleNRepeatPerShuffle,
+                                           1,
+                                           1,
+                                           MAccVgprs>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_cde_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMRepeatPerShuffle * MWave * MPerWmma,
+                                           1,
+                                           CShuffleNRepeatPerShuffle * NWave * NPerWmma>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_cde_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde_shuffle_block_copy_lds_to_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(e_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_global_step = sfc_cde_global.GetForwardStep(access_id);
+                    // move on Ds
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        cde_shuffle_block_copy_lds_to_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_global_step);
+                    });
+
+                    // move on E
+                    cde_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_global_step);
+                }
+            });
+        }
+        // clang-format on
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..1d920fb44d
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,937 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+namespace ck {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ADataType,
+          typename BDataType,
+          typename ComputeDataType_,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmMultipleD_xdl_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    using GemmSpecialization = ck::tensor_operation::device::GemmSpecialization;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+#if CK_WORKAROUND_DENORM_FIX
+    using ComputeDataType =
+        conditional_t<is_same_v<ComputeDataType_, ck::half_t>, ck::bhalf_t, ComputeDataType_>;
+#else
+    using ComputeDataType = ComputeDataType_;
+#endif
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(ComputeDataType),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    // A desc for source in blockwise copy
+    template <typename AGridDesc_M_K>
+    __host__ __device__ static constexpr auto
+    MakeDefaultAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B desc for source in blockwise copy
+    template <typename BGridDesc_N_K>
+    __host__ __device__ static constexpr auto
+    MakeDefaultBGridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const DsGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename AGridDesc_M_K,
+              typename BGridDesc_N_K,
+              typename DsGridDesc_M_N,
+              typename EGridDesc_M_N,
+              typename Block2ETileMap>
+    __host__ __device__ static constexpr bool CheckValidity(const AGridDesc_M_K& a_grid_desc_m_k,
+                                                            const BGridDesc_N_K& b_grid_desc_n_k,
+                                                            const DsGridDesc_M_N& ds_grid_desc_m_n,
+                                                            const EGridDesc_M_N& e_grid_desc_m_n,
+                                                            const Block2ETileMap& block_2_etile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+        static_assert(KPerBlock % AK1Value == 0 && KPerBlock % BK1Value == 0,
+                      "KPerBlock must be divisible by AK1Value and BK1Value!");
+
+        const auto M  = a_grid_desc_m_k.GetLength(I0);
+        const auto N  = b_grid_desc_n_k.GetLength(I0);
+        const auto AK = a_grid_desc_m_k.GetLength(I1);
+        const auto BK = b_grid_desc_n_k.GetLength(I1);
+
+        // check consistency of desc
+        if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1) && AK == BK))
+        {
+            return false;
+        }
+
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
+                              N == ds_grid_desc_m_n[i].GetLength(I1));
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check tile size
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && AK % KPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = AK / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // check block-to-E-tile
+        if(!block_2_etile_map.CheckValidity(e_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        // check tensor size: cannot be larger than 2GB each
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_m_k.GetElementSpaceSize() * sizeof(ADataType) <= TwoGB &&
+             b_grid_desc_n_k.GetElementSpaceSize() * sizeof(BDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <typename ALayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    template <typename BLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    template <typename ELayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    template <typename DsLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                             const std::array<index_t, NumDTensor>& NRaws,
+                             const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return MakeEGridDescriptor_M_N<DLayout, GemmSpec>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __device__ __host__ static constexpr auto GetMPerBlock() { return MPerBlock; }
+
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AK0_M_AK1,
+              typename BGridDesc_BK0_N_BK1,
+              typename DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename Block2ETileMap>
+    __device__ static void Run(const ADataType* __restrict__ p_a_grid,
+                               const BDataType* __restrict__ p_b_grid,
+                               DsGridPointer p_ds_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_etile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ADataType,
+                                                ComputeDataType,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                BDataType,
+                                                ComputeDataType,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<ComputeDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ComputeDataType,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeDataType*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
+                    },
+                    Number<NumDTensor>{}));
+
+            // blockwise copy C/D/E between LDS and global
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CDEElementwiseOperation,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
+                                                                            // support arbitray type
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
+                3,                    // index_t VectorDim,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0)),
+                 cde_element_op};
+
+            // space filling curve for threadwise C in VGPR before shuffle
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C/D/E
+            constexpr auto sfc_cde_block =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(e_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_lds_and_global_step =
+                        sfc_cde_block.GetForwardStep(access_id);
+
+                    // move on Ds
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                    });
+
+                    // move on E
+                    cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_lds_and_global_step);
+                }
+            });
+        }
+    }
+
+    template <bool HasMainKBlockLoop,
+              GemmSpecialization GemmSpec,
+              typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              typename Block2ETileMap>
+    __device__ static void Run(const void* __restrict__ p_a_grid_,
+                               const void* __restrict__ p_b_grid_,
+                               DsGridPointer p_ds_grid,
+                               void* __restrict__ p_e_grid_,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const index_t M,
+                               const index_t N,
+                               const index_t K,
+                               const index_t StrideA,
+                               const index_t StrideB,
+                               const std::array<index_t, NumDTensor> StrideDs,
+                               const index_t StrideE,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto p_a_grid = reinterpret_cast<const ADataType*>(p_a_grid_);
+        const auto p_b_grid = reinterpret_cast<const BDataType*>(p_b_grid_);
+        const auto p_e_grid = reinterpret_cast<EDataType*>(p_e_grid_);
+
+        // tensor descriptors for problem definiton
+        const auto a_grid_desc_m_k = MakeAGridDescriptor_M_K<ALayout, GemmSpec>(M, K, StrideA);
+        const auto b_grid_desc_n_k = MakeBGridDescriptor_N_K<BLayout, GemmSpec>(K, N, StrideB);
+
+        using DsGridDesc_M_N =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_M_N<DsLayout, GemmSpec>({}, {}, {}))>;
+
+        DsGridDesc_M_N ds_grid_desc_m_n;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+            ds_grid_desc_m_n(j) = MakeEGridDescriptor_M_N<DLayout, GemmSpec>(M, N, StrideDs[j]);
+        });
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+        // tensor descriptors for block/thread-wise copy
+        const auto a_grid_desc_ak0_m_ak1 = MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k);
+
+        const auto b_grid_desc_bk0_n_bk1 = MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k);
+
+        using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                DsGridDesc_M_N{}))>;
+
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock ds_grid_desc_mblock_mperblock_nblock_nperblock;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            ds_grid_desc_mblock_mperblock_nblock_nperblock(j) =
+                MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[j]);
+        });
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n);
+
+        Run<HasMainKBlockLoop>(p_a_grid,
+                               p_b_grid,
+                               p_ds_grid,
+                               p_e_grid,
+                               p_shared,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op,
+                               a_grid_desc_ak0_m_ak1,
+                               b_grid_desc_bk0_n_bk1,
+                               ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               block_2_etile_map);
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp
new file mode 100644
index 0000000000..e22391293e
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp
@@ -0,0 +1,1086 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+
+namespace ck {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ABDataType, // FIXME: don't assume A/B have same datatype
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_KBatch_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_KBatch_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    using GemmSpecialization = ck::tensor_operation::device::GemmSpecialization;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    // denorm test fix, required to work around fp16 mfma issue
+    // we convert fp16->fp32->bf16 and execute bf16 mfma instruction
+    // when mfma if fixed, remove this section and update
+    // ABDataTypeAdjusted -> ABDataType throughout this file
+#if CK_WORKAROUND_DENORM_FIX
+    using ABDataTypeAdjusted =
+        conditional_t<is_same_v<ABDataType, ck::half_t>, ck::bhalf_t, ABDataType>;
+#else
+    using ABDataTypeAdjusted = ABDataType;
+#endif
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_KBatch_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(AK0PerBlock * Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       AK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_KBatch_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(BK0PerBlock * Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       BK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(ABDataType),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    __host__ __device__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_least_multiple(M, MPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_least_multiple(N, NPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch)
+    {
+        return math::integer_least_multiple(K, KPerBlock * K_Batch);
+    }
+
+    template <typename ALayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeAGridDescriptor_KBatch_AK0_M_AK1(index_t M, index_t K, index_t StrideA, index_t KBatch)
+    {
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto MPad = CalculateMPadded(M);
+        const auto KPad = CalculateKPadded(K, KBatch);
+
+        const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto AK0 = KPad / (KBatch * AK1);
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
+        {
+            // const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, AK0, AK1)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, AK0, AK1)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    template <typename BLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeBGridDescriptor_KBatch_BK0_N_BK1(index_t K, index_t N, index_t StrideB, index_t KBatch)
+    {
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        const auto NPad = CalculateNPadded(N);
+        const auto KPad = CalculateKPadded(K, KBatch);
+
+        const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
+            b_grid_desc_k_n,
+            make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto BK0 = KPad / (KBatch * BK1);
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
+        {
+            // const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, BK0, BK1)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, BK0, BK1)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const DsGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    template <typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              GemmSpecialization GemmSpec>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const index_t M,
+                  const index_t N,
+                  const index_t K,
+                  const index_t StrideA,
+                  const index_t StrideB,
+                  const std::array<index_t, NumDTensor> StrideDs,
+                  const index_t StrideE,
+                  const index_t KBatch)
+    {
+        const auto a_grid_desc_kbatch_ak0_m_ak1 =
+            MakeAGridDescriptor_KBatch_AK0_M_AK1<ALayout, GemmSpec>(M, K, StrideA, KBatch);
+        const auto b_grid_desc_kbatch_bk0_n_bk1 =
+            MakeBGridDescriptor_KBatch_BK0_N_BK1<BLayout, GemmSpec>(K, N, StrideB, KBatch);
+
+        ignore = StrideDs;
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+#if 0
+        // check tile size
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+        {
+            return false;
+        }
+#endif
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        // check tensor size: cannot be larger than 2GB each
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <typename ELayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    template <typename DsLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                             const std::array<index_t, NumDTensor>& NRaws,
+                             const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return MakeEGridDescriptor_M_N<DLayout, GemmSpec>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __device__ __host__ static constexpr auto GetMPerBlock() { return MPerBlock; }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+              index_t NumDTensor_,
+              typename DsDataType_,
+              typename AGridDesc_KBatch_AK0_M_AK1,
+              typename BGridDesc_KBatch_BK0_N_BK1,
+              typename DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename CDEElementwiseOperation_,
+              typename Block2ETileMap>
+    __device__ static void Run(const ABDataType* __restrict__ p_a_grid,
+                               const ABDataType* __restrict__ p_b_grid,
+                               DsGridPointer p_ds_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               uint32_t* barrier_count_finished,
+                               const index_t KBatch,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation_& cde_element_op,
+                               const AGridDesc_KBatch_AK0_M_AK1& a_grid_desc_kbatch_ak0_m_ak1,
+                               const BGridDesc_KBatch_BK0_N_BK1& b_grid_desc_kbatch_bk0_n_bk1,
+                               const DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor_>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t kbatch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_kbatch_ak0_m_ak1 =
+            GetABlockDescriptor_KBatch_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_kbatch_bk0_n_bk1 =
+            GetBBlockDescriptor_KBatch_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_KBatch_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataTypeAdjusted,
+                                                decltype(a_grid_desc_kbatch_ak0_m_ak1),
+                                                decltype(a_block_desc_kbatch_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<2, 0, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_kbatch_ak0_m_ak1,
+                make_multi_index(kbatch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_kbatch_ak0_m_ak1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_KBatch_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataTypeAdjusted,
+                                                decltype(b_grid_desc_kbatch_bk0_n_bk1),
+                                                decltype(b_block_desc_kbatch_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<2, 0, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_kbatch_bk0_n_bk1,
+                make_multi_index(kbatch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_kbatch_bk0_n_bk1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<ABDataTypeAdjusted, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ABDataTypeAdjusted,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+#if 1
+        if(block_work_idx[I0] == 0)
+        {
+            const index_t nThreadSize = CDEShuffleBlockTransferScalarPerVector_NPerBlock;
+            const index_t numNThreads = NPerBlock / nThreadSize;
+            const index_t numMThreads = BlockSize / numNThreads;
+            const index_t mThreadSize = MPerBlock / numMThreads;
+
+            const index_t m_tid = get_thread_local_1d_id() / numNThreads;
+            const index_t n_tid = get_thread_local_1d_id() % numNThreads;
+
+            auto c_thread_desc_mblock_mperblock_nblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mThreadSize>{}, I1, Number<nThreadSize>{}));
+
+            StaticBuffer<AddressSpaceEnum::Vgpr,
+                         EDataType,
+                         c_thread_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize(),
+                         true>
+                e_thread_zero_buf;
+
+            auto c_thread_copy = ThreadwiseTensorSliceTransfer_v1r3<
+                EDataType,
+                EDataType,
+                decltype(c_thread_desc_mblock_mperblock_nblock_nperblock),
+                decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                ck::tensor_operation::element_wise::PassThrough,
+                Sequence<1, mThreadSize, 1, nThreadSize>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{e_grid_desc_mblock_mperblock_nblock_nperblock,
+                      make_multi_index(block_work_idx[I1],
+                                       m_tid * mThreadSize,
+                                       block_work_idx[I2],
+                                       n_tid * nThreadSize),
+                      ck::tensor_operation::element_wise::PassThrough{}};
+
+            c_thread_copy.Run(c_thread_desc_mblock_mperblock_nblock_nperblock,
+                              make_tuple(I0, I0, I0, I0),
+                              e_thread_zero_buf,
+                              e_grid_desc_mblock_mperblock_nblock_nperblock,
+                              e_grid_buf);
+
+            __syncthreads();
+
+            if(threadIdx.x == 0)
+            {
+                atomicAdd(barrier_count_finished, 1);
+            }
+        }
+#endif
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataTypeAdjusted*>(p_shared),
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataTypeAdjusted*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop =
+            __builtin_amdgcn_readfirstlane((a_grid_desc_kbatch_ak0_m_ak1.GetLength(I1) *
+                                            a_grid_desc_kbatch_ak0_m_ak1.GetLength(I3)) /
+                                           KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_kbatch_ak0_m_ak1,
+                                                               a_block_desc_kbatch_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_kbatch_bk0_n_bk1,
+                                                               b_block_desc_kbatch_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            if(threadIdx.x == 0)
+            {
+                while(__atomic_load_n(barrier_count_finished, __ATOMIC_RELAXED) == 0) {}
+            }
+
+            __syncthreads();
+
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor_>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor_>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0);
+                    },
+                    Number<NumDTensor_>{}));
+
+            // space filling curve for threadwise C in VGPR before shuffle
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C/D/E
+            constexpr auto sfc_cde_block =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+
+            // blockwise copy C/D/E between LDS and global
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType_{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CDEElementwiseOperation_,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make
+                                                                            // Sequence support
+                                                                            // arbitray type
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
+                3,                    // index_t VectorDim,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor_,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0)),
+                 cde_element_op};
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(e_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_lds_and_global_step =
+                        sfc_cde_block.GetForwardStep(access_id);
+
+                    // move on Ds
+                    static_for<0, NumDTensor_, 1>{}([&](auto i) {
+                        cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                    });
+
+                    // move on E
+                    cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_lds_and_global_step);
+                }
+            });
+
+            if(threadIdx.x == 0)
+            {
+                index_t k_id_finished_t = atomicAdd(barrier_count_finished, 1);
+
+                if(k_id_finished_t == KBatch)
+                {
+                    *barrier_count_finished = 0;
+                }
+            }
+        }
+    }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+              GemmSpecialization GemmSpec,
+              typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              typename Block2ETileMap>
+    __device__ static void Run(const void* __restrict__ p_a_grid_,
+                               const void* __restrict__ p_b_grid_,
+                               DsGridPointer p_ds_grid,
+                               void* __restrict__ p_e_grid_,
+                               void* __restrict__ p_shared,
+                               uint32_t* barrier_count_finished,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const index_t M,
+                               const index_t N,
+                               const index_t K,
+                               const index_t StrideA,
+                               const index_t StrideB,
+                               const std::array<index_t, NumDTensor> StrideDs,
+                               const index_t StrideE,
+                               const index_t KBatch,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto p_a_grid = reinterpret_cast<const ABDataType*>(p_a_grid_);
+        const auto p_b_grid = reinterpret_cast<const ABDataType*>(p_b_grid_);
+        const auto p_e_grid = reinterpret_cast<EDataType*>(p_e_grid_);
+
+        using DsGridDesc_M_N =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_M_N<DsLayout, GemmSpec>({}, {}, {}))>;
+
+        DsGridDesc_M_N ds_grid_desc_m_n;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+            ds_grid_desc_m_n(j) = MakeEGridDescriptor_M_N<DLayout, GemmSpec>(M, N, StrideDs[j]);
+        });
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+        // tensor descriptors for block/thread-wise copy
+        const auto a_grid_desc_kbatch_ak0_m_ak1 =
+            MakeAGridDescriptor_KBatch_AK0_M_AK1<ALayout, GemmSpec>(M, K, StrideA, KBatch);
+
+        const auto b_grid_desc_kbatch_bk0_n_bk1 =
+            MakeBGridDescriptor_KBatch_BK0_N_BK1<BLayout, GemmSpec>(K, N, StrideB, KBatch);
+
+        using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                DsGridDesc_M_N{}))>;
+
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock ds_grid_desc_mblock_mperblock_nblock_nperblock;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            ds_grid_desc_mblock_mperblock_nblock_nperblock(j) =
+                MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[j]);
+        });
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n);
+
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t kbatch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+
+        if(kbatch_id == KBatch - 1)
+        {
+            Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, NumDTensor, DsDataType>(
+                p_a_grid,
+                p_b_grid,
+                p_ds_grid,
+                p_e_grid,
+                p_shared,
+                barrier_count_finished,
+                KBatch,
+                a_element_op,
+                b_element_op,
+                cde_element_op,
+                a_grid_desc_kbatch_ak0_m_ak1,
+                b_grid_desc_kbatch_bk0_n_bk1,
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                block_2_etile_map);
+        }
+        else
+        {
+            Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, 0, Tuple<>>(
+                p_a_grid,
+                p_b_grid,
+                p_ds_grid,
+                p_e_grid,
+                p_shared,
+                barrier_count_finished,
+                KBatch,
+                a_element_op,
+                b_element_op,
+                ck::tensor_operation::element_wise::PassThrough{},
+                a_grid_desc_kbatch_ak0_m_ak1,
+                b_grid_desc_kbatch_bk0_n_bk1,
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                block_2_etile_map);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp
new file mode 100644
index 0000000000..f760feb2ed
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp"
+
+namespace ck {
+
+enum struct PipelineVersion
+{
+    v1,
+    v2,
+};
+
+template <PipelineVersion PipelineVer,
+          index_t NumPrefetch     = 1,
+          LoopScheduler LoopSched = LoopScheduler::Default>
+constexpr auto GridwiseGemmPipeline_Selector()
+{
+    if constexpr(PipelineVer == PipelineVersion::v1)
+    {
+        if constexpr(LoopSched == LoopScheduler::Default)
+        {
+            return GridwiseGemmPipeline_v1<NumPrefetch>{};
+        }
+        else if constexpr(LoopSched == LoopScheduler::Interwave)
+        {
+            return GridwiseGemmPipelineInterwave_v1<NumPrefetch>{};
+        }
+    }
+    else if constexpr(PipelineVer == PipelineVersion::v2)
+    {
+        return GridwiseGemmPipeline_v2{};
+    }
+    else
+    {
+        std::cerr << "GridwiseGemmPipeline configuration is not available" << std::endl;
+    }
+}
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp
new file mode 100644
index 0000000000..d1209636de
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp
@@ -0,0 +1,369 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+
+namespace ck {
+
+template <index_t NumPrefetch>
+struct GridwiseGemmPipeline_v1;
+
+// 1-stage prefetch
+template <>
+struct GridwiseGemmPipeline_v1<1>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    __host__ __device__ static constexpr bool IsSupported(index_t /* num_loop */) { return true; }
+
+    __host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
+    {
+        return num_loop > 1;
+    }
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename BlockwiseGemm,
+              typename CThreadBuffer>
+    __device__ static void Run(const AGridDesc& a_grid_desc,
+                               const ABlockDesc& a_block_desc,
+                               ABlockTransfer& a_blockwise_copy,
+                               const AGridBuffer& a_grid_buf,
+                               ABlockBuffer& a_block_buf,
+                               const ABlockTransferStep& a_block_copy_step,
+                               const BGridDesc& b_grid_desc,
+                               const BBlockDesc& b_block_desc,
+                               BBlockTransfer& b_blockwise_copy,
+                               const BGridBuffer& b_grid_buf,
+                               BBlockBuffer& b_block_buf,
+                               const BBlockTransferStep& b_block_copy_step,
+                               const BlockwiseGemm& blockwise_gemm,
+                               CThreadBuffer& c_thread_buf,
+                               index_t num_loop)
+    {
+        // preload data into LDS
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+
+            do
+            {
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+
+                block_sync_lds();
+
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                block_sync_lds();
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+                ++i;
+            } while(i < (num_loop - 1));
+        }
+
+        // tail
+        {
+            block_sync_lds();
+
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+    }
+};
+
+// 2-stage prefetch
+template <>
+struct GridwiseGemmPipeline_v1<2>
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    __host__ __device__ static constexpr bool IsSupported(index_t num_loop)
+    {
+        // TODO: improve applicability
+        return num_loop % 2 == 0;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
+    {
+        return (num_loop / 2) > 1;
+    }
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename BlockwiseGemm,
+              typename CThreadBuffer>
+    static __device__ void Run(const AGridDesc& a_grid_desc,
+                               const ABlockDesc& a_block_desc,
+                               ABlockTransfer& a_blockwise_copy,
+                               const AGridBuffer& a_grid_buf,
+                               ABlockBuffer& a_block_buf,
+                               const ABlockTransferStep& a_block_copy_step,
+                               const BGridDesc& b_grid_desc,
+                               const BBlockDesc& b_block_desc,
+                               BBlockTransfer& b_blockwise_copy,
+                               const BGridBuffer& b_grid_buf,
+                               BBlockBuffer& b_block_buf,
+                               const BBlockTransferStep& b_block_copy_step,
+                               const BlockwiseGemm& blockwise_gemm,
+                               CThreadBuffer& c_thread_buf,
+                               index_t num_loop)
+    {
+        // preload data into LDS
+        {
+            // Read 0
+            a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+            b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
+
+            // Move
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+            // Read 1
+            a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1);
+            b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I1);
+        }
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+
+            do
+            {
+                // Move
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                // Write i
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I0);
+
+                // Read i+2
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
+
+                // Sync
+                block_sync_lds();
+
+                // Gemm i
+                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                // Sync
+                block_sync_lds();
+
+                // Move
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                // Write i+1
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I1);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I1);
+
+                // Read i+3
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I1);
+
+                // Sync
+                block_sync_lds();
+
+                // Gemm i+1
+                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                // Sync
+                block_sync_lds();
+
+                i += 2;
+            } while(i < (num_loop - 2));
+        }
+
+        // tail
+        {
+            // Write num_loop - 2
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
+            b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I0);
+
+            // Sync
+            block_sync_lds();
+
+            // Gemm num_loop - 2
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+            // Sync
+            block_sync_lds();
+
+            // Write num_loop - 1
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I1);
+            b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I1);
+
+            // Sync
+            block_sync_lds();
+
+            // Gemm num_loop - 1
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+    }
+};
+
+template <index_t NumPrefetch>
+struct GridwiseGemmPipelineInterwave_v1;
+
+template <>
+struct GridwiseGemmPipelineInterwave_v1<1>
+{
+    __host__ __device__ static constexpr bool IsSupported(index_t /* num_loop */) { return true; }
+
+    __host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
+    {
+        return num_loop > 1;
+    }
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename BlockwiseGemm,
+              typename CThreadBuffer>
+    static __device__ void Run(const AGridDesc& a_grid_desc,
+                               const ABlockDesc& a_block_desc,
+                               ABlockTransfer& a_blockwise_copy,
+                               const AGridBuffer& a_grid_buf,
+                               ABlockBuffer& a_block_buf,
+                               const ABlockTransferStep& a_block_copy_step,
+                               const BGridDesc& b_grid_desc,
+                               const BBlockDesc& b_block_desc,
+                               BBlockTransfer& b_blockwise_copy,
+                               const BGridBuffer& b_grid_buf,
+                               BBlockBuffer& b_block_buf,
+                               const BBlockTransferStep& b_block_copy_step,
+                               const BlockwiseGemm& blockwise_gemm,
+                               CThreadBuffer& c_thread_buf,
+                               index_t num_loop)
+    {
+        // preload data into LDS
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+
+            do
+            {
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+
+                block_sync_lds();
+
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                // block_sync_lds(); // moved into blockwise_gemm
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+                ++i;
+            } while(i < (num_loop - 1));
+        }
+
+        // tail
+        {
+            block_sync_lds();
+
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+    }
+};
+
+// Note: 2 stage prefetch not optimized for inter-wave loop scheduler
+template <>
+struct GridwiseGemmPipelineInterwave_v1<2> : public GridwiseGemmPipeline_v1<2>
+{
+};
+
+// TODO: deprecate as GridwiseGemmPipeline_Selector covers the functionality
+template <index_t NumPrefetch, LoopScheduler LoopSched>
+constexpr auto GridwiseGemmPipeline_v1_Selector()
+{
+    if constexpr(LoopSched == LoopScheduler::Default)
+    {
+        return GridwiseGemmPipeline_v1<NumPrefetch>{};
+    }
+    else if constexpr(LoopSched == LoopScheduler::Interwave)
+    {
+        return GridwiseGemmPipelineInterwave_v1<NumPrefetch>{};
+    }
+}
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp
new file mode 100644
index 0000000000..d3d7d5af85
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp
@@ -0,0 +1,132 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+
+namespace ck {
+
+struct GridwiseGemmPipeline_v2
+{
+    __host__ __device__ static constexpr bool IsSupported(index_t num_loop)
+    {
+        // TODO: improve applicability
+        return num_loop % 2 == 0;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
+    {
+        return (num_loop / 2) > 1;
+    }
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename BlockwiseGemm,
+              typename CThreadBuffer>
+    __device__ static void Run(const AGridDesc& a_grid_desc,
+                               const ABlockDesc& a_block_desc,
+                               ABlockTransfer& a_blockwise_copy,
+                               const AGridBuffer& a_grid_buf,
+                               ABlockBuffer& a_block_buf,
+                               const ABlockTransferStep& a_block_copy_step,
+                               const BGridDesc& b_grid_desc,
+                               const BBlockDesc& b_block_desc,
+                               BBlockTransfer& b_blockwise_copy,
+                               const BGridBuffer& b_grid_buf,
+                               BBlockBuffer& b_block_buf,
+                               const BBlockTransferStep& b_block_copy_step,
+                               const BlockwiseGemm& blockwise_gemm,
+                               CThreadBuffer& c_thread_buf,
+                               index_t num_loop)
+    {
+        // global read 0
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        // move to 1
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // LDS write 0
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        // global Read 1
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+
+        // LDS write 0
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+        // global Read 1
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+
+            do
+            {
+#if CK_EXPERIMENTAL_PIPELINE_V2_IGLP_OPT
+                __builtin_amdgcn_iglp_opt(CK_EXPERIMENTAL_PIPELINE_V2_IGLP_OPT);
+#endif
+
+                block_sync_lds();
+
+                // GEMM i
+                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                block_sync_lds();
+
+                // move to i + 2
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                // LDS write i + 1
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                // global read i + 2
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+
+                // LDS write i + 1
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+                // global read i + 2
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                ++i;
+            } while(i < (num_loop - 2));
+        }
+
+        // tail
+        {
+            block_sync_lds();
+
+            // GEMM num_loop - 2
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+            block_sync_lds();
+
+            // LDS write num_loop - 1
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+            b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+            block_sync_lds();
+
+            // GEMM num_loop - 1
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v3.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v3.hpp
new file mode 100644
index 0000000000..ced62241cd
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v3.hpp
@@ -0,0 +1,89 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+
+namespace ck {
+
+struct GridwiseGemmPipeline_v3
+{
+    __host__ __device__ static constexpr bool IsSupported(index_t)
+    {
+        // TODO: improve applicability
+        return true;
+    }
+
+    template <typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename BlockwiseGemm,
+              typename CThreadBuffer>
+    __device__ static void Run(const AGridDesc& a_grid_desc,
+                               const ABlockDesc& a_block_desc,
+                               ABlockTransfer& a_blockwise_copy,
+                               const AGridBuffer& a_grid_buf,
+                               ABlockBuffer& a_block_buf,
+                               const ABlockTransferStep& a_block_copy_step,
+                               const BGridDesc& b_grid_desc,
+                               const BBlockDesc& b_block_desc,
+                               BBlockTransfer& b_blockwise_copy,
+                               const BGridBuffer& b_grid_buf,
+                               BBlockBuffer& b_block_buf,
+                               const BBlockTransferStep& b_block_copy_step,
+                               const BlockwiseGemm& blockwise_gemm,
+                               CThreadBuffer& c_thread_buf,
+                               index_t num_loop)
+    {
+        // global read 0
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // LDS write 0
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+        num_loop--;
+
+        while(num_loop > 0)
+        {
+            a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+            block_sync_lds();
+            b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+            block_sync_lds();
+
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+            b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+            b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+            num_loop--;
+        }
+        // tail
+        {
+            block_sync_lds();
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
new file mode 100644
index 0000000000..99e410f688
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
@@ -0,0 +1,881 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename ReduceGridDescriptor_MBlock_MPerBlock,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_reduce_xdl_cshuffle_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            ReducePtrsGlobal p_reduces_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const ReduceInElementwiseOperations reduce_in_element_ops,
+            const ReduceAccElementwiseOperations reduce_out_element_ops,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const ReduceGridDescriptor_MBlock_MPerBlock reduce_grid_desc_mblock_mperblock,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_reduces_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  reduce_in_element_ops,
+                                                  reduce_out_element_ops,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  reduce_grid_desc_mblock_mperblock,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_reduces_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = reduce_in_element_ops;
+    ignore = reduce_out_element_ops;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = reduce_grid_desc_mblock_mperblock;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename FloatReduceAcc,
+          typename ReducePtrsGlobal,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename ReduceOperations,
+          typename ReduceInElementwiseOperations,
+          typename ReduceAccElementwiseOperations,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename ReduceGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDesc_M_N,
+          typename ReduceGridDesc_M,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+          index_t CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+          index_t CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        // static_assert(is_known_at_compile_time<remove_cv_t<decltype(AK1)>>::value &&
+        //               is_known_at_compile_time<remove_cv_t<decltype(BK1)>>::value,
+        //               "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeReduceGridDescriptor_MBlock_MPerBlock(const ReduceGridDesc_M& d_grid_desc_m)
+    {
+        const auto M      = d_grid_desc_m.GetLength(I0);
+        const auto MBlock = M / MPerBlock;
+
+        const auto reduce_grid_desc_mblock_mperblock = transform_tensor_descriptor(
+            d_grid_desc_m,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{}))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1>{}));
+
+        return reduce_grid_desc_mblock_mperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            CGridDesc_M_N{}))>;
+
+    using ReduceGridDescriptor_MBlock_MPerBlock =
+        remove_cvref_t<decltype(MakeReduceGridDescriptor_MBlock_MPerBlock(ReduceGridDesc_M{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        ReducePtrsGlobal p_reduces_grid,
+        void* __restrict__ p_shared,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const ReduceInElementwiseOperations& reduce_in_element_ops,
+        const ReduceAccElementwiseOperations& reduce_out_element_ops,
+        const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+        const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+        const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            c_grid_desc_mblock_mperblock_nblock_nperblock,
+        const ReduceGridDescriptor_MBlock_MPerBlock& reduce_grid_desc_mblock_mperblock,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C + reduction + write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            // TODO: this should be implemented as a blockwise reduction
+            // LDS c_reduce_block_desc_mperblock_nperblock
+            constexpr auto c_reduce_block_desc_mperblock_nperblock = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1)),
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I3))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<>{}, Sequence<1>{}));
+
+            static_assert(CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) *
+                                  CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                              BlockSize,
+                          "wrong!");
+
+            static_assert((CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) %
+                                      CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) ==
+                                  0 &&
+                              (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) %
+                                      CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                                  0,
+                          "wrong!");
+
+            constexpr index_t mreduce_per_thread =
+                (CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) /
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0);
+
+            constexpr index_t nreduce_per_thread =
+                (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) /
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1);
+
+            constexpr auto c_reduce_thread_lengths_mperblock_nperblock =
+                Sequence<mreduce_per_thread, nreduce_per_thread>{};
+
+            // VGPR c_reduce_thread_desc_mperblock_nperblock
+            constexpr auto c_reduce_thread_desc_mperblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(Number<mreduce_per_thread>{}, Number<nreduce_per_thread>{}));
+
+            // VGPR reduce_thread_desc_mperblock
+            constexpr auto reduce_thread_desc_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(Number<mreduce_per_thread>{}));
+
+            // VGPR reduce_thread_desc_mblock_mperblock
+            constexpr auto reduce_thread_desc_mblock_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(I1, Number<mreduce_per_thread>{}));
+
+            auto c_reduce_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                c_reduce_thread_desc_mperblock_nperblock.GetElementSpaceSize());
+
+            // reduce: threadwise copy from LDS to VGPR
+            constexpr auto c_reduce_thread_cluster_desc = make_cluster_descriptor(
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock{}, Sequence<1, 0>{});
+
+            const auto c_reduce_thread_cluster_idx =
+                c_reduce_thread_cluster_desc.CalculateBottomIndex(
+                    make_multi_index(get_thread_local_1d_id()));
+
+            const auto c_reduce_thread_data_idx_begin =
+                c_reduce_thread_cluster_idx * c_reduce_thread_lengths_mperblock_nperblock;
+
+            auto c_reduce_thread_copy_lds_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatCShuffle,
+                FloatReduceAcc,
+                decltype(c_reduce_block_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_lengths_mperblock_nperblock),
+                Sequence<0, 1>,
+                1,
+                CReduceThreadLds2VGprCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>{c_reduce_block_desc_mperblock_nperblock, c_reduce_thread_data_idx_begin};
+
+            auto reduce_tuple_thread_copy_vgpr_to_global = generate_tuple(
+                [&](auto I) {
+                    auto p_reduce_grid         = p_reduces_grid[I];
+                    auto reduce_acc_element_op = reduce_out_element_ops[I];
+
+                    return ThreadwiseTensorSliceTransfer_v1r3<
+                        FloatReduceAcc,
+                        remove_pointer_t<decltype(p_reduce_grid)>,
+                        decltype(reduce_thread_desc_mblock_mperblock),
+                        decltype(reduce_grid_desc_mblock_mperblock),
+                        decltype(reduce_acc_element_op),
+                        Sequence<1, mreduce_per_thread>,
+                        Sequence<0, 1>,
+                        1,
+                        CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock,
+                        ReduceGlobalMemoryDataOperation::At(I),
+                        1,
+                        false>{reduce_grid_desc_mblock_mperblock,
+                               make_multi_index(block_work_idx[I0],                  // mblock
+                                                c_reduce_thread_data_idx_begin[I0]), // mperblock
+                               reduce_acc_element_op};
+                },
+                Number<p_reduces_grid.Size()>{});
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                // TODO - extract following into reduction_blockwise
+                {
+                    c_reduce_thread_copy_lds_to_vgpr.Run(c_reduce_block_desc_mperblock_nperblock,
+                                                         c_shuffle_block_buf,
+                                                         c_reduce_thread_desc_mperblock_nperblock,
+                                                         make_tuple(I0, I0),
+                                                         c_reduce_thread_buf);
+
+                    static_for<0, p_reduces_grid.Size(), 1>{}([&](auto In) {
+                        auto& p_reduce_grid = p_reduces_grid[In];
+
+                        auto reduce_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                            p_reduce_grid, reduce_grid_desc_mblock_mperblock.GetElementSpaceSize());
+
+                        auto reduce_thread_buf =
+                            make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                                reduce_thread_desc_mperblock.GetElementSpaceSize());
+
+                        auto& reduce_in_element_op = reduce_in_element_ops[In];
+
+                        auto& reduce_thread_copy_vgpr_to_global =
+                            reduce_tuple_thread_copy_vgpr_to_global(In);
+
+                        using ReduceOperation = remove_cvref_t<decltype(ReduceOperations{}[In])>;
+                        using ThreadwiseReduce =
+                            ThreadwiseReduction<FloatReduceAcc,
+                                                decltype(c_reduce_thread_desc_mperblock_nperblock),
+                                                decltype(reduce_thread_desc_mperblock),
+                                                ReduceOperation,
+                                                false>;
+
+                        // Global write Gemm shuffle + reduction
+                        const auto reduce_identityVal =
+                            ReduceOperation::template GetIdentityValue<FloatReduceAcc>();
+
+                        static_for<0, mreduce_per_thread, 1>{}(
+                            [&](auto I) { reduce_thread_buf(I) = reduce_identityVal; });
+
+                        // reduce in VGPR
+                        static_for<0, mreduce_per_thread, 1>{}([&](auto im) {
+                            static_for<0, nreduce_per_thread, 1>{}([&](auto in) {
+                                constexpr auto offset =
+                                    Number<c_reduce_thread_desc_mperblock_nperblock.CalculateOffset(
+                                        make_tuple(im, in))>{};
+
+                                reduce_in_element_op(c_reduce_thread_buf(offset),
+                                                     c_reduce_thread_buf(offset));
+                            });
+                        });
+
+                        ThreadwiseReduce::Reduce(c_reduce_thread_buf, reduce_thread_buf);
+
+                        // copy from VGPR to Global
+                        reduce_thread_copy_vgpr_to_global.Run(reduce_thread_desc_mblock_mperblock,
+                                                              make_tuple(I0, I0),
+                                                              reduce_thread_buf,
+                                                              reduce_grid_desc_mblock_mperblock,
+                                                              reduce_grid_buf);
+
+                        if constexpr(access_id < num_access - 1)
+                        {
+                            constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+                            reduce_thread_copy_vgpr_to_global.MoveDstSliceWindow(
+                                reduce_grid_desc_mblock_mperblock,
+                                make_tuple(c_global_step[I0], c_global_step[I1]));
+                        }
+                    });
+                }
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+
+            // Reduction
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp
new file mode 100644
index 0000000000..18cf80041b
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp
@@ -0,0 +1,1265 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ABDataType, // FIXME: don't assume A/B have same datatype
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+          typename AGridDesc_M_K,
+          typename BGridDesc_N_K,
+          typename DsGridDesc_M_N,
+          typename EGridDesc_M_N,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched>
+struct GridwiseGemmSplitKMultipleD_xdl_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = GridwiseGemmPipeline_v1<NumGemmKPrefetchStage>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, src of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AKB_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(AK0PerBlock * Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       AK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, src of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BKB_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(BK0PerBlock * Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       BK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(ABDataType),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    // A desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultAGridDescriptor_AKB_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k,
+                                             const int split_k)
+    {
+        const auto MRaw = a_grid_desc_m_k.GetLength(I0);
+        const auto KRaw = a_grid_desc_m_k.GetLength(I1);
+
+        const index_t AK0 =
+            (math::integer_divide_ceil(KRaw, KPerBlock * split_k) * KPerBlock) / AK1;
+        const index_t K = split_k * AK0 * AK1;
+        const auto KPad = K - KRaw;
+
+        const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+        return transform_tensor_descriptor(
+            a_grid_desc_m_kpad,
+            make_tuple(make_unmerge_transform(make_tuple(split_k, AK0, AK1)),
+                       make_pass_through_transform(MRaw)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+    }
+
+    // B desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultBGridDescriptor_BKB_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k,
+                                             const int split_k)
+    {
+        const auto NRaw = b_grid_desc_n_k.GetLength(I0);
+        const auto KRaw = b_grid_desc_n_k.GetLength(I1);
+
+        const index_t BK0 =
+            (math::integer_divide_ceil(KRaw, KPerBlock * split_k) * KPerBlock) / BK1;
+        const index_t K = split_k * BK0 * BK1;
+        const auto KPad = K - KRaw;
+
+        const auto b_grid_desc_n_kpad = transform_tensor_descriptor(
+            b_grid_desc_n_k,
+            make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return transform_tensor_descriptor(
+            b_grid_desc_n_kpad,
+            make_tuple(make_unmerge_transform(make_tuple(split_k, BK0, BK1)),
+                       make_pass_through_transform(NRaw)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDescriptor_M_N>
+    __host__ __device__ static constexpr auto MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const EGridDescriptor_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDescriptor_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const DsGridDescriptor_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n, const int split_k)
+    {
+        return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n, 8, split_k);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename AGridDesc_AKB_AK0_M_AK1,
+              typename BGridDesc_BKB_BK0_N_BK1,
+              typename Block2ETileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AKB_AK0_M_AK1& a_grid_desc_akb_ak0_m_ak1,
+                  const BGridDesc_BKB_BK0_N_BK1& b_grid_desc_bkb_bk0_n_bk1,
+                  const DsGridDesc_M_N& ds_grid_desc_m_n,
+                  const EGridDesc_M_N& e_grid_desc_m_n,
+                  const Block2ETileMap& block_2_etile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_akb_ak0_m_ak1.GetLength(I2);
+        const auto N = b_grid_desc_bkb_bk0_n_bk1.GetLength(I2);
+        const auto K =
+            a_grid_desc_akb_ak0_m_ak1.GetLength(I1) * a_grid_desc_akb_ak0_m_ak1.GetLength(I3);
+
+        if(K != b_grid_desc_bkb_bk0_n_bk1.GetLength(I1) * b_grid_desc_bkb_bk0_n_bk1.GetLength(I3))
+        {
+            return false;
+        }
+        if(a_grid_desc_akb_ak0_m_ak1.GetLength(I0) != b_grid_desc_bkb_bk0_n_bk1.GetLength(I0))
+        {
+            return false;
+        }
+
+        // check consistency of desc
+        if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1)))
+        {
+            return false;
+        }
+
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
+                              N == ds_grid_desc_m_n[i].GetLength(I1));
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check tile size
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // check block-to-E-tile
+        if(!block_2_etile_map.CheckValidity(e_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        // check tensor size: cannot be larger than 2GB each
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_akb_ak0_m_ak1.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             b_grid_desc_bkb_bk0_n_bk1.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    using DefaultAGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(MakeDefaultAGridDescriptor_AKB_AK0_M_AK1(AGridDesc_M_K{}, 1))>;
+    using DefaultBGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(MakeDefaultBGridDescriptor_BKB_BK0_N_BK1(BGridDesc_N_K{}, 1))>;
+    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+    using DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
+
+    using DefaultBlock2ETileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}, 1))>;
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AKB_AK0_M_AK1,
+              typename BGridDesc_BKB_BK0_N_BK1,
+              typename Block2ETileMap>
+    __device__ static void Run(const ABDataType* __restrict__ p_a_grid,
+                               const ABDataType* __restrict__ p_b_grid,
+                               DsGridPointer p_ds_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const AGridDesc_AKB_AK0_M_AK1& a_grid_desc_akb_ak0_m_ak1,
+                               const BGridDesc_BKB_BK0_N_BK1& b_grid_desc_bkb_bk0_n_bk1,
+                               const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(block_work_idx[Number<0>{}] == 0)
+        {
+            Run0<HasMainKBlockLoop>(p_a_grid,
+                                    p_b_grid,
+                                    p_ds_grid,
+                                    p_e_grid,
+                                    p_shared,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op,
+                                    a_grid_desc_akb_ak0_m_ak1,
+                                    b_grid_desc_bkb_bk0_n_bk1,
+                                    ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                    e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                    block_2_etile_map);
+        }
+        else
+        {
+            Run1<HasMainKBlockLoop>(p_a_grid,
+                                    p_b_grid,
+                                    p_e_grid,
+                                    p_shared,
+                                    a_element_op,
+                                    b_element_op,
+                                    a_grid_desc_akb_ak0_m_ak1,
+                                    b_grid_desc_bkb_bk0_n_bk1,
+                                    ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                    e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                    block_2_etile_map);
+        }
+    }
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AKB_AK0_M_AK1,
+              typename BGridDesc_BKB_BK0_N_BK1,
+              typename Block2ETileMap>
+    __device__ static void Run0(const ABDataType* __restrict__ p_a_grid,
+                                const ABDataType* __restrict__ p_b_grid,
+                                DsGridPointer p_ds_grid,
+                                EDataType* __restrict__ p_e_grid,
+                                void* __restrict__ p_shared,
+                                const AElementwiseOperation& a_element_op,
+                                const BElementwiseOperation& b_element_op,
+                                const CDEElementwiseOperation& cde_element_op,
+                                const AGridDesc_AKB_AK0_M_AK1& a_grid_desc_akb_ak0_m_ak1,
+                                const BGridDesc_BKB_BK0_N_BK1& b_grid_desc_bkb_bk0_n_bk1,
+                                const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                    ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                    e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                const Block2ETileMap& block_2_etile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_akb_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bkb_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_etile_map.ValidCTileIndex(
+               make_tuple(block_work_idx[I1], block_work_idx[I2]),
+               make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t k_batch_id = block_work_idx[I0];
+
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto a_block_desc_akb_ak0_m_ak1 =
+            GetABlockDescriptor_AKB_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        constexpr auto b_block_desc_bkb_bk0_n_bk1 =
+            GetBBlockDescriptor_BKB_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<I1, AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(a_grid_desc_akb_ak0_m_ak1),
+                                                decltype(a_block_desc_akb_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_akb_ak0_m_ak1,
+                make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_akb_ak0_m_ak1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<I1, BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(b_grid_desc_bkb_bk0_n_bk1),
+                                                decltype(b_block_desc_bkb_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bkb_bk0_n_bk1,
+                make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bkb_bk0_n_bk1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<ABDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ABDataType,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_v1_Selector<NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_akb_ak0_m_ak1.GetLength(I1) * a_grid_desc_akb_ak0_m_ak1.GetLength(I3)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_akb_ak0_m_ak1,
+                                                               a_block_desc_akb_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bkb_bk0_n_bk1,
+                                                               b_block_desc_bkb_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+            {
+                // tuple of reference to C/Ds tensor descriptors
+                const auto c_ds_desc_refs = concat_tuple_of_reference(
+                    tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                    generate_tie(
+                        [&](auto i) -> const auto& // return type should be reference
+                        { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                        Number<NumDTensor>{}));
+
+                // tuple of reference to C/Ds tensor descriptors
+                const auto c_ds_buf_refs = concat_tuple_of_reference(
+                    tie(c_shuffle_block_buf),
+                    generate_tie(
+                        [&](auto i) -> const auto& // return type should be reference
+                        { return ds_grid_buf[i]; },
+                        Number<NumDTensor>{}));
+
+                // tuple of starting index of C/Ds blockwise copy
+                const auto idx_c_ds_block_begin = container_concat(
+                    make_tuple(make_multi_index(0, 0, 0, 0)),
+                    generate_tuple(
+                        [&](auto) {
+                            return make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0);
+                        },
+                        Number<NumDTensor>{}));
+
+                // blockwise copy C/D/E between LDS and global
+                auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
+                    ThisThreadBlock,
+                    decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
+                    Tuple<EDataType>,
+                    decltype(c_ds_desc_refs),
+                    decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                    CDEElementwiseOperation,
+                    Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make
+                                                                                // Sequence support
+                                                                                // arbitray type
+                    Sequence<1,
+                             CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                             1,
+                             CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                    CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                    Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
+                    3,                    // index_t VectorDim,
+                    CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                    sequence_merge_t<Sequence<true>,
+                                     uniform_sequence_gen_t<
+                                         NumDTensor,
+                                         false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                    Sequence<false>>              // ThreadTransferDstResetCoordinateAfterRunFlags
+                    {c_ds_desc_refs,
+                     idx_c_ds_block_begin,
+                     tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                     make_tuple(make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0)),
+                     cde_element_op};
+
+                // space filling curve for threadwise C in VGPR before shuffle
+                constexpr auto sfc_c_vgpr =
+                    SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                      Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                      Sequence<CShuffleMXdlPerWavePerShuffle,
+                                               CShuffleNXdlPerWavePerShuffle,
+                                               1,
+                                               1,
+                                               M2,
+                                               1,
+                                               M4,
+                                               1>>{};
+
+                // space filling curve for shuffled blockwise C/D/E
+                constexpr auto sfc_cde_block =
+                    SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                      Sequence<0, 2, 1, 3>,
+                                      Sequence<1,
+                                               CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                               1,
+                                               CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+                constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+                static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+
+                static_for<0, num_access, 1>{}([&](auto access_id) {
+                    // make sure it's safe to write to LDS
+                    block_sync_lds();
+
+                    // each thread write its data from VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                                  c_thread_buf,
+                                                  c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  c_shuffle_block_buf);
+
+                    // make sure it's safe to read from LDS
+                    block_sync_lds();
+
+                    // each block copy its data from LDS to global
+                    cde_block_copy_lds_and_global.Run(
+                        c_ds_desc_refs,
+                        c_ds_buf_refs,
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        tie(e_grid_buf));
+
+                    if constexpr(access_id < num_access - 1)
+                    {
+                        constexpr auto cde_lds_and_global_step =
+                            sfc_cde_block.GetForwardStep(access_id);
+
+                        // move on Ds
+                        static_for<0, NumDTensor, 1>{}([&](auto i) {
+                            cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                                c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                        });
+
+                        // move on E
+                        cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                            tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                            I0,
+                            cde_lds_and_global_step);
+                    }
+                });
+            }
+        }
+    }
+
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AKB_AK0_M_AK1,
+              typename BGridDesc_BKB_BK0_N_BK1,
+              typename Block2ETileMap>
+    __device__ static void Run1(const ABDataType* __restrict__ p_a_grid,
+                                const ABDataType* __restrict__ p_b_grid,
+                                EDataType* __restrict__ p_e_grid,
+                                void* __restrict__ p_shared,
+                                const AElementwiseOperation& a_element_op,
+                                const BElementwiseOperation& b_element_op,
+                                const AGridDesc_AKB_AK0_M_AK1& a_grid_desc_akb_ak0_m_ak1,
+                                const BGridDesc_BKB_BK0_N_BK1& b_grid_desc_bkb_bk0_n_bk1,
+                                const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&,
+                                const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                    e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                const Block2ETileMap& block_2_etile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_akb_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bkb_bk0_n_bk1.GetElementSpaceSize());
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_etile_map.ValidCTileIndex(
+               make_tuple(block_work_idx[I1], block_work_idx[I2]),
+               make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t k_batch_id = block_work_idx[I0];
+
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto a_block_desc_akb_ak0_m_ak1 =
+            GetABlockDescriptor_AKB_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+        constexpr auto b_block_desc_bkb_bk0_n_bk1 =
+            GetBBlockDescriptor_BKB_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<I1, AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(a_grid_desc_akb_ak0_m_ak1),
+                                                decltype(a_block_desc_akb_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_akb_ak0_m_ak1,
+                make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_akb_ak0_m_ak1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<I1, BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(b_grid_desc_bkb_bk0_n_bk1),
+                                                decltype(b_block_desc_bkb_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bkb_bk0_n_bk1,
+                make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bkb_bk0_n_bk1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<ABDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ABDataType,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_v1_Selector<NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_akb_ak0_m_ak1.GetLength(I1) * a_grid_desc_akb_ak0_m_ak1.GetLength(I3)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_akb_ak0_m_ak1,
+                                                               a_block_desc_akb_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bkb_bk0_n_bk1,
+                                                               b_block_desc_bkb_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+            {
+                // shuffle: blockwise copy C from LDS to global
+                auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                    ThisThreadBlock,                                 // ThreadGroup
+                    ck::tensor_operation::element_wise::PassThrough, // ElementwiseOperation,
+                    EGlobalMemoryDataOperation,                      // DstInMemOp,
+                    Sequence<1,
+                             CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                             1,
+                             CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                    CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                    CShuffleDataType,     // typename SrcData,
+                    EDataType,            // typename DstData,
+                    decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                    decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    Sequence<0, 1, 2, 3>,                             // typename DimAccessOrder,
+                    3,                                                // index_t VectorDim,
+                    CDEShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                    true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                    false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                    {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                     make_multi_index(0, 0, 0, 0),
+                     e_grid_desc_mblock_mperblock_nblock_nperblock,
+                     make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0),
+                     ck::tensor_operation::element_wise::PassThrough{}};
+
+                // space filling curve for threadwise C in VGPR
+                constexpr auto sfc_c_vgpr =
+                    SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                      Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                      Sequence<CShuffleMXdlPerWavePerShuffle,
+                                               CShuffleNXdlPerWavePerShuffle,
+                                               1,
+                                               1,
+                                               M2,
+                                               1,
+                                               M4,
+                                               1>>{};
+
+                // space filling curve for shuffled blockwise C in global mem
+                constexpr auto sfc_c_global =
+                    SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                      Sequence<0, 2, 1, 3>,
+                                      Sequence<1,
+                                               CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                               1,
+                                               CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+                constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+                static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+                static_for<0, num_access, 1>{}([&](auto access_id) {
+                    // make sure it's safe to write to LDS
+                    block_sync_lds();
+
+                    // each thread write its data from VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                                  c_thread_buf,
+                                                  c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  c_shuffle_block_buf);
+
+                    // make sure it's safe to read from LDS
+                    block_sync_lds();
+
+                    // each block copy its data from LDS to global
+                    c_shuffle_block_copy_lds_to_global.Run(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                        c_shuffle_block_buf,
+                        e_grid_desc_mblock_mperblock_nblock_nperblock,
+                        e_grid_buf);
+
+                    if constexpr(access_id < num_access - 1)
+                    {
+                        constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                        // move on C
+                        c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                            e_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                    }
+                });
+            }
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle_v2.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle_v2.hpp
new file mode 100644
index 0000000000..caf8f040f4
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle_v2.hpp
@@ -0,0 +1,1076 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+
+namespace ck {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ADataType, // FIXME: don't assume A/B have same datatype
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ComputeType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_KBatch_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_KBatch_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    using GemmSpecialization = ck::tensor_operation::device::GemmSpecialization;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_KBatch_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(AK0PerBlock * Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       AK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_KBatch_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(BK0PerBlock * Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       BK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max(a_block_space_size_aligned * sizeof(ADataType) +
+                             b_block_space_size_aligned * sizeof(BDataType),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    __host__ __device__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_least_multiple(M, MPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_least_multiple(N, NPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch)
+    {
+        return math::integer_least_multiple(K, KPerBlock * K_Batch);
+    }
+
+    template <typename ALayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeAGridDescriptor_KBatch_AK0_M_AK1(index_t M, index_t K, index_t StrideA, index_t KBatch)
+    {
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto MPad = CalculateMPadded(M);
+        const auto KPad = CalculateKPadded(K, KBatch);
+
+        const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto AK0 = KPad / (KBatch * AK1);
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
+        {
+            // const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, AK0, AK1)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, AK0, AK1)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    template <typename BLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeBGridDescriptor_KBatch_BK0_N_BK1(index_t K, index_t N, index_t StrideB, index_t KBatch)
+    {
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        const auto NPad = CalculateNPadded(N);
+        const auto KPad = CalculateKPadded(K, KBatch);
+
+        const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
+            b_grid_desc_k_n,
+            make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto BK0 = KPad / (KBatch * BK1);
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
+        {
+            // const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, BK0, BK1)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, BK0, BK1)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const DsGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    template <typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              GemmSpecialization GemmSpec>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const index_t M,
+                  const index_t N,
+                  const index_t K,
+                  const index_t StrideA,
+                  const index_t StrideB,
+                  const std::array<index_t, NumDTensor> StrideDs,
+                  const index_t StrideE,
+                  const index_t KBatch)
+    {
+        const auto a_grid_desc_kbatch_ak0_m_ak1 =
+            MakeAGridDescriptor_KBatch_AK0_M_AK1<ALayout, GemmSpec>(M, K, StrideA, KBatch);
+        const auto b_grid_desc_kbatch_bk0_n_bk1 =
+            MakeBGridDescriptor_KBatch_BK0_N_BK1<BLayout, GemmSpec>(K, N, StrideB, KBatch);
+
+        ignore = StrideDs;
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+#if 0
+        // check tile size
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+        {
+            return false;
+        }
+#endif
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        // check tensor size: cannot be larger than 2GB each
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType) <= TwoGB &&
+             b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <typename ELayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    template <typename DsLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                             const std::array<index_t, NumDTensor>& NRaws,
+                             const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return MakeEGridDescriptor_M_N<DLayout, GemmSpec>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __device__ __host__ static constexpr auto GetMPerBlock() { return MPerBlock; }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+              index_t NumDTensor_,
+              typename DsDataType_,
+              typename AGridDesc_KBatch_AK0_M_AK1,
+              typename BGridDesc_KBatch_BK0_N_BK1,
+              typename DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename CDEElementwiseOperation_,
+              typename Block2ETileMap>
+    __device__ static void Run(const ADataType* __restrict__ p_a_grid,
+                               const BDataType* __restrict__ p_b_grid,
+                               DsGridPointer p_ds_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               uint32_t* barrier_count_finished,
+                               const index_t KBatch,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation_& cde_element_op,
+                               const AGridDesc_KBatch_AK0_M_AK1& a_grid_desc_kbatch_ak0_m_ak1,
+                               const BGridDesc_KBatch_BK0_N_BK1& b_grid_desc_kbatch_bk0_n_bk1,
+                               const DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor_>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t kbatch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_kbatch_ak0_m_ak1 =
+            GetABlockDescriptor_KBatch_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_kbatch_bk0_n_bk1 =
+            GetBBlockDescriptor_KBatch_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_KBatch_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ADataType,
+                                                ComputeType,
+                                                decltype(a_grid_desc_kbatch_ak0_m_ak1),
+                                                decltype(a_block_desc_kbatch_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<2, 0, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_kbatch_ak0_m_ak1,
+                make_multi_index(kbatch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_kbatch_ak0_m_ak1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_KBatch_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                BDataType,
+                                                ComputeType,
+                                                decltype(b_grid_desc_kbatch_bk0_n_bk1),
+                                                decltype(b_block_desc_kbatch_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<2, 0, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_kbatch_bk0_n_bk1,
+                make_multi_index(kbatch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_kbatch_bk0_n_bk1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<ComputeType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ComputeType,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+#if 1
+        if(block_work_idx[I0] == 0)
+        {
+            const index_t nThreadSize = CDEShuffleBlockTransferScalarPerVector_NPerBlock;
+            const index_t numNThreads = NPerBlock / nThreadSize;
+            const index_t numMThreads = BlockSize / numNThreads;
+            const index_t mThreadSize = MPerBlock / numMThreads;
+
+            const index_t m_tid = get_thread_local_1d_id() / numNThreads;
+            const index_t n_tid = get_thread_local_1d_id() % numNThreads;
+
+            auto c_thread_desc_mblock_mperblock_nblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mThreadSize>{}, I1, Number<nThreadSize>{}));
+
+            StaticBuffer<AddressSpaceEnum::Vgpr,
+                         EDataType,
+                         c_thread_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize(),
+                         true>
+                e_thread_zero_buf;
+
+            auto c_thread_copy = ThreadwiseTensorSliceTransfer_v1r3<
+                EDataType,
+                EDataType,
+                decltype(c_thread_desc_mblock_mperblock_nblock_nperblock),
+                decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                ck::tensor_operation::element_wise::PassThrough,
+                Sequence<1, mThreadSize, 1, nThreadSize>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{e_grid_desc_mblock_mperblock_nblock_nperblock,
+                      make_multi_index(block_work_idx[I1],
+                                       m_tid * mThreadSize,
+                                       block_work_idx[I2],
+                                       n_tid * nThreadSize),
+                      ck::tensor_operation::element_wise::PassThrough{}};
+
+            c_thread_copy.Run(c_thread_desc_mblock_mperblock_nblock_nperblock,
+                              make_tuple(I0, I0, I0, I0),
+                              e_thread_zero_buf,
+                              e_grid_desc_mblock_mperblock_nblock_nperblock,
+                              e_grid_buf);
+
+            __syncthreads();
+
+            if(threadIdx.x == 0)
+            {
+                atomicAdd(barrier_count_finished, 1);
+            }
+        }
+#endif
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeType*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop =
+            __builtin_amdgcn_readfirstlane((a_grid_desc_kbatch_ak0_m_ak1.GetLength(I1) *
+                                            a_grid_desc_kbatch_ak0_m_ak1.GetLength(I3)) /
+                                           KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_kbatch_ak0_m_ak1,
+                                                               a_block_desc_kbatch_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_kbatch_bk0_n_bk1,
+                                                               b_block_desc_kbatch_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            if(threadIdx.x == 0)
+            {
+                while(__atomic_load_n(barrier_count_finished, __ATOMIC_RELAXED) == 0) {}
+            }
+
+            __syncthreads();
+
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor_>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor_>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0);
+                    },
+                    Number<NumDTensor_>{}));
+
+            // space filling curve for threadwise C in VGPR before shuffle
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C/D/E
+            constexpr auto sfc_cde_block =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+
+            // blockwise copy C/D/E between LDS and global
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType_{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CDEElementwiseOperation_,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make
+                                                                            // Sequence support
+                                                                            // arbitray type
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
+                3,                    // index_t VectorDim,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor_,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0)),
+                 cde_element_op};
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(e_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_lds_and_global_step =
+                        sfc_cde_block.GetForwardStep(access_id);
+
+                    // move on Ds
+                    static_for<0, NumDTensor_, 1>{}([&](auto i) {
+                        cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                    });
+
+                    // move on E
+                    cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_lds_and_global_step);
+                }
+            });
+
+            if(threadIdx.x == 0)
+            {
+                index_t k_id_finished_t = atomicAdd(barrier_count_finished, 1);
+
+                if(k_id_finished_t == KBatch)
+                {
+                    *barrier_count_finished = 0;
+                }
+            }
+        }
+    }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+              GemmSpecialization GemmSpec,
+              typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              typename Block2ETileMap>
+    __device__ static void Run(const void* __restrict__ p_a_grid_,
+                               const void* __restrict__ p_b_grid_,
+                               DsGridPointer p_ds_grid,
+                               void* __restrict__ p_e_grid_,
+                               void* __restrict__ p_shared,
+                               uint32_t* barrier_count_finished,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const index_t M,
+                               const index_t N,
+                               const index_t K,
+                               const index_t StrideA,
+                               const index_t StrideB,
+                               const std::array<index_t, NumDTensor> StrideDs,
+                               const index_t StrideE,
+                               const index_t KBatch,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto p_a_grid = reinterpret_cast<const ADataType*>(p_a_grid_);
+        const auto p_b_grid = reinterpret_cast<const BDataType*>(p_b_grid_);
+        const auto p_e_grid = reinterpret_cast<EDataType*>(p_e_grid_);
+
+        using DsGridDesc_M_N =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_M_N<DsLayout, GemmSpec>({}, {}, {}))>;
+
+        DsGridDesc_M_N ds_grid_desc_m_n;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+            ds_grid_desc_m_n(j) = MakeEGridDescriptor_M_N<DLayout, GemmSpec>(M, N, StrideDs[j]);
+        });
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+        // tensor descriptors for block/thread-wise copy
+        const auto a_grid_desc_kbatch_ak0_m_ak1 =
+            MakeAGridDescriptor_KBatch_AK0_M_AK1<ALayout, GemmSpec>(M, K, StrideA, KBatch);
+
+        const auto b_grid_desc_kbatch_bk0_n_bk1 =
+            MakeBGridDescriptor_KBatch_BK0_N_BK1<BLayout, GemmSpec>(K, N, StrideB, KBatch);
+
+        using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                DsGridDesc_M_N{}))>;
+
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock ds_grid_desc_mblock_mperblock_nblock_nperblock;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            ds_grid_desc_mblock_mperblock_nblock_nperblock(j) =
+                MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[j]);
+        });
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n);
+
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t kbatch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+
+        if(kbatch_id == KBatch - 1)
+        {
+            Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, NumDTensor, DsDataType>(
+                p_a_grid,
+                p_b_grid,
+                p_ds_grid,
+                p_e_grid,
+                p_shared,
+                barrier_count_finished,
+                KBatch,
+                a_element_op,
+                b_element_op,
+                cde_element_op,
+                a_grid_desc_kbatch_ak0_m_ak1,
+                b_grid_desc_kbatch_bk0_n_bk1,
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                block_2_etile_map);
+        }
+        else
+        {
+            Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, 0, Tuple<>>(
+                p_a_grid,
+                p_b_grid,
+                p_ds_grid,
+                p_e_grid,
+                p_shared,
+                barrier_count_finished,
+                KBatch,
+                a_element_op,
+                b_element_op,
+                ck::tensor_operation::element_wise::PassThrough{},
+                a_grid_desc_kbatch_ak0_m_ak1,
+                b_grid_desc_kbatch_bk0_n_bk1,
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                block_2_etile_map);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_waveletmodel.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_waveletmodel.hpp
new file mode 100644
index 0000000000..de5a424198
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_waveletmodel.hpp
@@ -0,0 +1,157 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+
+namespace ck {
+
+template <typename TileLoadThreadGroup, index_t NumGemmKPrefetchStage>
+struct GridwiseGemmLoadWave;
+
+// 1-stage prefetch
+template <typename TileLoadThreadGroup>
+struct GridwiseGemmLoadWave<TileLoadThreadGroup, 1>
+{
+    __host__ __device__ static constexpr bool IsSupported(index_t /* num_loop */)
+    {
+        // TODO: improve applicability
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
+    {
+        return num_loop > 1;
+    }
+
+    template <bool HasMainLoop,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep>
+    static __device__ void RunLoadWavePipeline(const AGridDesc& a_grid_desc,
+                                               const ABlockDesc& a_block_desc,
+                                               ABlockTransfer& a_blockwise_copy,
+                                               const AGridBuffer& a_grid_buf,
+                                               ABlockBuffer& a_block_buf,
+                                               const ABlockTransferStep& a_block_copy_step,
+                                               const BGridDesc& b_grid_desc,
+                                               const BBlockDesc& b_block_desc,
+                                               BBlockTransfer& b_blockwise_copy,
+                                               const BGridBuffer& b_grid_buf,
+                                               BBlockBuffer& b_block_buf,
+                                               const BBlockTransferStep& b_block_copy_step,
+                                               index_t num_loop)
+    {
+        // global read 0
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        // move to 1
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // LDS write 0
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+
+            do
+            {
+                // sync for Load threads()
+                block_sync_lds();
+                // global read i + 1
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                // move to i + 2
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                // sync with math threads()
+                block_sync_lds();
+
+                // LDS write i+1
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+                ++i;
+            } while(i < (num_loop - 1));
+        }
+
+        // tail
+        {
+            block_sync_lds();
+            // GEMM num_loop - 1
+        }
+    }
+};
+
+template <typename TileMathThreadGroup, index_t NumGemmKPrefetchStage>
+struct GridwiseGemmMathWave;
+// 1- stage prefetch
+template <typename TileMathThreadGroup>
+struct GridwiseGemmMathWave<TileMathThreadGroup, 1>
+{
+
+    __host__ __device__ static constexpr bool IsSupported(index_t /* num_loop */) { return true; }
+
+    __host__ __device__ static constexpr bool CalculateHasMainLoop(index_t num_loop)
+    {
+        return num_loop > 1;
+    }
+
+    template <bool HasMainLoop,
+              typename ABlockBuffer,
+              typename BBlockBuffer,
+              typename BlockwiseGemm,
+              typename CThreadBuffer>
+    static __device__ void RunMathWavePipeline(ABlockBuffer& a_block_buf,
+                                               BBlockBuffer& b_block_buf,
+                                               const BlockwiseGemm& block_gemm,
+                                               CThreadBuffer& c_thread_buf,
+                                               index_t num_loop)
+    {
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+
+            do
+            {
+                block_sync_lds();
+
+                // GEMM i
+                block_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                block_sync_lds();
+                ++i;
+            } while(i < (num_loop - 1));
+        }
+
+        // tail
+        {
+            block_sync_lds();
+
+            // GEMM num_loop - 1
+            block_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp
new file mode 100644
index 0000000000..d8b31311b1
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp
@@ -0,0 +1,650 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_wmma(
+            const FloatA* __restrict__ p_a_grid,
+            const FloatB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            // const
+            // CGridDescriptor_MBlockxRepeat_MWave_MSubGroup_MAccVgprs_NBlockxRepeat_NWave_NThreadPerSubGroup
+            //     c_grid_desc_mblockxrepeat_mwave_msubgroup_maccvgprs_nblockxrepeat_nwave_nthreadpersubgroup,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__) || defined(__gfx1101__) || \
+    defined(__gfx1102__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_shared,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx1100__))
+}
+
+template <index_t BlockSize,
+          typename FloatA,
+          typename FloatB,
+          typename FloatAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M_N,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerWmma,
+          index_t NPerWmma,
+          index_t K1Value,
+          index_t MRepeat,
+          index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          index_t NumGemmKPrefetchStage = 1,
+          LoopScheduler LoopSched       = make_default_loop_scheduler(),
+          PipelineVersion PipelineVer   = PipelineVersion::v1>
+struct GridwiseGemm_k0mk1_k0nk1_mn_wmma
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0perblock_mperblock_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return a_block_desc_k0perblock_mperblock_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0perblock_nperblock_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return b_block_desc_k0perblock_nperblock_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    // *Caution Here repeat is shuffle repeat
+    GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat()
+    {
+        constexpr index_t MWave = MPerBlock / (MRepeat * MPerWmma);
+        constexpr index_t NWave = NPerBlock / (NRepeat * NPerWmma);
+
+        constexpr auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMRepeatPerShuffle * MWave * MPerWmma>{},
+                           I1,
+                           Number<CShuffleNRepeatPerShuffle * NWave * NPerWmma>{}));
+
+        return c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0perblock_mperblock_k1 =
+            GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        constexpr auto b_block_desc_k0perblock_nperblock_k1 =
+            GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        constexpr auto max_lds_align = K1;
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_k0perblock_nperblock_k1.GetElementSpaceSize(), max_lds_align);
+
+        return (a_block_space_size_aligned * sizeof(FloatA) +
+                b_block_space_size_aligned * sizeof(FloatB));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerWmma * MRepeat) == 0) &&
+                          (NPerBlock % (NRepeat * NPerWmma)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+             K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+             K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K0 / K0PerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_k0_m_k1.GetElementSpaceSize() * sizeof(FloatA) <= TwoGB &&
+             b_grid_desc_k0_n_k1.GetElementSpaceSize() * sizeof(FloatB) <= TwoGB))
+        {
+            return false;
+        }
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / (K0PerBlock * K1);
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap(
+        const CGridDesc_M_N& c_grid_desc_m_n, index_t /* M01 */, index_t /* N01 */)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            CGridDesc_M_N{}))>;
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>;
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
+    __device__ static void Run(const FloatA* __restrict__ p_a_grid,
+                               const FloatB* __restrict__ p_b_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               void* __restrict__ p_shared,
+                               const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                               const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                               const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CElementwiseOperation& c_element_op,
+                               const Block2CTileMap& block_2_ctile_map)
+    {
+        // clang-format off
+/*******************************************************************************/
+// Memory buffer zone.
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+/*******************************************************************************/
+// BlockIdx.x -> [BlockId.m, BlockId.n]
+        const auto block_work_idx = block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        { return; }
+
+        // Store BlockId into SGPR
+        const index_t m_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+        const index_t n_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+/*******************************************************************************/
+// BlockLevel, A/B Matrix ThreadMapping in LDS, As Destinaion of BlockWise_Copy
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+        constexpr auto max_lds_align = K1;
+        constexpr auto a_block_desc_k0perblock_mperblock_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+        constexpr auto b_block_desc_k0perblock_nperblock_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<        ThisThreadBlock,
+/* typename SrcElementwiseOperation,              */    AElementwiseOperation,
+/* typename DstElementwiseOperation,              */    ck::tensor_operation::element_wise::PassThrough,
+/* InMemoryDataOperationEnum DstInMemOp,          */    InMemoryDataOperationEnum::Set,
+/* typename BlockSliceLengths,                    */    Sequence<K0PerBlock, MPerBlock, K1>,
+/* typename ThreadClusterLengths,                 */    ABlockTransferThreadClusterLengths_K0_M_K1,
+/* typename ThreadClusterArrangeOrder,            */    ABlockTransferThreadClusterArrangeOrder,
+/* typename SrcData,                              */    FloatA,
+/* typename DstData,                              */    FloatA,
+/* typename SrcDesc,                              */    decltype(a_grid_desc_k0_m_k1),
+/* typename DstDesc,                              */    decltype(a_block_desc_k0perblock_mperblock_k1),
+/* typename SrcDimAccessOrder,                    */    ABlockTransferSrcAccessOrder,
+/* typename DstDimAccessOrder,                    */    Sequence<0, 1, 2>,
+/* index_t SrcVectorDim,                          */    ABlockTransferSrcVectorDim,
+/* index_t DstVectorDim,                          */    2,
+/* index_t SrcScalarPerVector,                    */    ABlockTransferSrcScalarPerVector,
+/* index_t DstScalarPerVector,                    */    ABlockTransferDstScalarPerVector_K1,
+/* index_t SrcScalarStrideInVector,               */    1,
+/* index_t DstScalarStrideInVector,               */    1,
+/* bool ThreadTransferSrcResetCoordinateAfterRun, */    AThreadTransferSrcResetCoordinateAfterRun,
+/* bool ThreadTransferDstResetCoordinateAfterRun, */    true>(
+                a_grid_desc_k0_m_k1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_k0perblock_mperblock_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<K0PerBlock, NPerBlock, K1>,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatB,
+                                                FloatB,
+                                                decltype(b_grid_desc_k0_n_k1),
+                                                decltype(b_block_desc_k0perblock_nperblock_k1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<0, 1, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                b_grid_desc_k0_n_k1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_k0perblock_nperblock_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+/*******************************************************************************/
+        // GEMM
+        constexpr auto WmmaK = 16;
+        constexpr auto KPack = math::integer_least_multiple(K1, WmmaK);
+
+        auto blockwise_gemm =
+            BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle<BlockSize,
+                                                         FloatA,
+                                                         FloatB,
+                                                         FloatAcc,
+                                                         decltype(a_block_desc_k0perblock_mperblock_k1),
+                                                         decltype(b_block_desc_k0perblock_nperblock_k1),
+                                                         MPerWmma,
+                                                         NPerWmma,
+                                                         MRepeat,
+                                                         NRepeat,
+                                                         KPack>{};
+
+        // Prepare Register for C matrix
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+/*******************************************************************************/
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize(), max_lds_align);
+        // LDS allocation for A and B: be careful of alignment
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(static_cast<FloatA*>(p_shared), a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize());
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(static_cast<FloatB*>(p_shared) + a_block_space_size_aligned, b_block_desc_k0perblock_nperblock_k1.GetElementSpaceSize());
+        
+        // Shift Per SUB_K
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+
+        // gridwise GEMM pipeline
+        const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
+        GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc_k0_m_k1,
+                                                          a_block_desc_k0perblock_mperblock_k1,
+                                                          a_blockwise_copy,
+                                                          a_grid_buf,
+                                                          a_block_buf,
+                                                          a_block_slice_copy_step,
+                                                          b_grid_desc_k0_n_k1,
+                                                          b_block_desc_k0perblock_nperblock_k1,
+                                                          b_blockwise_copy,
+                                                          b_grid_buf,
+                                                          b_block_buf,
+                                                          b_block_slice_copy_step,
+                                                          blockwise_gemm,
+                                                          c_thread_buf,
+                                                          K0BlockMainLoop);
+/*******************************************************************************/
+        // write out to C, implement shuffle
+        {
+            constexpr auto c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs =  
+            blockwise_gemm.GetCThreadDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
+
+            // This API Provide All dimension (size) you need
+            constexpr auto c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp =
+                blockwise_gemm.GetCBlockDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
+
+            constexpr auto MWave              = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I1);
+            constexpr auto MSubGroup          = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I2);
+            constexpr auto NWave              = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I4);
+            constexpr auto NThreadPerSubGroup = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I5);
+            constexpr auto MAccVgprs          = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I6);
+
+            // LDS descriptor, shuffle and write out in MRepeat x NRepeat times
+            constexpr auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
+                GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs = transform_tensor_descriptor(
+                c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMRepeatPerShuffle>{}, // MRepeat per shuffle repeat
+                        MWave,                               // MWave
+                        MSubGroup,                           // MSubGroup * MAccVgprs = MPerWmma
+                        MAccVgprs)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNRepeatPerShuffle>{}, // NRepeat per shuffle repeat
+                        NWave,                               // NWave
+                        NThreadPerSubGroup))),               // NThreadPerSubGroup = NPerWmma
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0, 1, 2, 6>{}, Sequence<>{}, Sequence<3, 4, 5>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block = blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_mrepeat_mwave_msubgroup_maccvgprs_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(MRepeat, MWave, MSubGroup, MAccVgprs))),
+                make_tuple(Sequence<0, 1, 2, 3>{}),
+                make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(NRepeat, NWave, NThreadPerSubGroup))),
+                make_tuple(Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0>{}));
+            
+            const auto m_thread_data_on_block_idx = m_thread_data_on_block_to_mrepeat_mwave_msubgroup_maccvgprs_adaptor.CalculateBottomIndex(
+                make_multi_index(m_thread_data_on_block));
+            
+            const auto n_thread_data_on_block_idx = n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_adaptor.CalculateBottomIndex(
+                make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
+                                                   decltype(c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMRepeatPerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            CShuffleNRepeatPerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            MAccVgprs>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6>,
+                                                   6,
+                                                   1, // vector write pixel
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
+                    make_multi_index(0,
+                                     m_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     0,
+                                     n_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMRepeatPerShuffle * MWave * MPerWmma,
+                         1,
+                         CShuffleNRepeatPerShuffle * NWave * NPerWmma>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            // space filling curve for local reg & global memory
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MRepeat, 1, 1, NRepeat, 1, 1, MAccVgprs>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6>,
+                                  Sequence<CShuffleMRepeatPerShuffle,
+                                           1,
+                                           1,
+                                           CShuffleNRepeatPerShuffle,
+                                           1,
+                                           1,
+                                           MAccVgprs>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMRepeatPerShuffle * MWave * MPerWmma,
+                                           1,
+                                           CShuffleNRepeatPerShuffle * NWave * NPerWmma>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+        // clang-format on
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
new file mode 100644
index 0000000000..9c09f3a539
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
@@ -0,0 +1,1068 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm, bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdl_cshuffle_v1(typename GridwiseGemm::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        karg.p_a_grid, karg.p_b_grid, karg.p_c_grid, p_shared, karg);
+#else
+    ignore = karg;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename GridwiseGemm,
+          typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdl_cshuffle_v1(const FloatA* __restrict__ p_a_grid,
+                                    const FloatB* __restrict__ p_b_grid,
+                                    FloatC* __restrict__ p_c_grid,
+                                    typename GridwiseGemm::Problem problem)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid, p_b_grid, p_c_grid, p_shared, problem);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = problem;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename FloatA,
+          typename FloatB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          tensor_operation::device::GemmSpecialization GemmSpec,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1,
+          typename ComputeType        = FloatC>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK0Number = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0Number = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1Number = Number<AK1Value>{};
+    static constexpr auto BK1Number = Number<BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __host__ static auto CalculateGridSize(index_t M, index_t N)
+    {
+        return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, 1);
+    }
+
+    __host__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_divide_ceil(M, MPerBlock) * MPerBlock;
+    }
+
+    __host__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_divide_ceil(N, NPerBlock) * NPerBlock;
+    }
+
+    __host__ static auto CalculateKPadded(index_t K)
+    {
+        return math::integer_divide_ceil(K, KPerBlock) * KPerBlock;
+    }
+
+    __host__ static auto CalculateAK0(index_t K)
+    {
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding ||
+                     GemmSpec == GemmSpecialization::KPadding ||
+                     GemmSpec == GemmSpecialization::NKPadding)
+        {
+            return CalculateKPadded(K) / AK1Value;
+        }
+        else
+        {
+            return K / AK1Value;
+        }
+    }
+
+    __host__ static auto CalculateBK0(index_t K)
+    {
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding ||
+                     GemmSpec == GemmSpecialization::KPadding ||
+                     GemmSpec == GemmSpecialization::MKPadding)
+        {
+            return CalculateKPadded(K) / BK1Value;
+        }
+        else
+        {
+            return K / BK1Value;
+        }
+    }
+
+    __host__ static auto CalculateMBlock(index_t M)
+    {
+        return math::integer_divide_floor(M, MPerBlock);
+    }
+
+    __host__ static auto CalculateNBlock(index_t N)
+    {
+        return math::integer_divide_floor(N, NPerBlock);
+    }
+
+    __device__ static auto MakeAGridDescriptor_AK0_M_AK1(
+        index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            const auto a_grid_desc_m_k =
+                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
+                                            make_tuple(make_right_pad_transform(M, MPad - M),
+                                                       make_right_pad_transform(K, KPad - K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(MPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            const auto a_grid_desc_m_k = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+        else
+        {
+            // not pad M or K
+            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
+                a_grid_desc_mraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return a_grid_desc_ak0_m_ak1;
+        }
+    }
+
+    __device__ static auto MakeBGridDescriptor_BK0_N_BK1(
+        index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(StrideB, I1));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            const auto b_grid_desc_n_k =
+                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
+                                            make_tuple(make_right_pad_transform(N, NPad - N),
+                                                       make_right_pad_transform(K, KPad - K)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_n_k,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(NPad)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            const auto b_grid_desc_n_k = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(N), make_right_pad_transform(K, KPad - K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_n_k,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+        else
+        {
+            // not pad N or K
+            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
+                b_grid_desc_nraw_kraw,
+                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return b_grid_desc_bk0_n_bk1;
+        }
+    }
+
+    __host__ __device__ static auto
+    MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(M, MPad - M),
+                                                          make_right_pad_transform(N, NPad - N)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(M, MPad - M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_grid_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(M), make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_grid_desc_mraw_nraw;
+        }
+    }
+
+    struct Problem
+    {
+        __host__ Problem(index_t M_,
+                         index_t N_,
+                         index_t K_,
+                         index_t StrideA_,
+                         index_t StrideB_,
+                         index_t StrideC_)
+            : M{M_},
+              N{N_},
+              K{K_},
+              StrideA{StrideA_},
+              StrideB{StrideB_},
+              StrideC{StrideC_},
+              MPadded{CalculateMPadded(M_)},
+              NPadded{CalculateNPadded(N_)},
+              KPadded{CalculateKPadded(K_)},
+              AK0{CalculateAK0(K_)},
+              BK0{CalculateBK0(K_)},
+              MBlock{CalculateMBlock(M_)},
+              NBlock{CalculateNBlock(N_)}
+        {
+        }
+
+        __host__ void Print() const
+        {
+            std::cout << "problem {"
+                      << "M:" << M << ", "
+                      << "N:" << N << ", "
+                      << "K:" << K << ", "
+                      << "SA:" << StrideA << ", "
+                      << "SB:" << StrideB << ", "
+                      << "SC:" << StrideC << ", "
+                      << "MP:" << MPadded << ", "
+                      << "NP:" << NPadded << ", "
+                      << "KP:" << KPadded << ", "
+                      << "AK0:" << AK0 << ", "
+                      << "BK0:" << BK0 << ", "
+                      << "MBlock: " << MBlock << ", "
+                      << "NBlock: " << NBlock << "}" << std::endl;
+        }
+
+        index_t M;
+        index_t N;
+        index_t K;
+        index_t StrideA;
+        index_t StrideB;
+        index_t StrideC;
+        index_t MPadded;
+        index_t NPadded;
+        index_t KPadded;
+        index_t AK0;
+        index_t BK0;
+        index_t MBlock;
+        index_t NBlock;
+    };
+
+    // Argument
+    struct Argument : public tensor_operation::device::BaseArgument, public Problem
+    {
+        __host__ Argument(const FloatA* p_a_grid_,
+                          const FloatB* p_b_grid_,
+                          FloatC* p_c_grid_,
+                          index_t M_,
+                          index_t N_,
+                          index_t K_,
+                          index_t StrideA_,
+                          index_t StrideB_,
+                          index_t StrideC_)
+            : Problem{M_, N_, K_, StrideA_, StrideB_, StrideC_},
+              p_a_grid{p_a_grid_},
+              p_b_grid{p_b_grid_},
+              p_c_grid{p_c_grid_}
+        {
+        }
+
+        const FloatA* p_a_grid;
+        const FloatB* p_b_grid;
+        FloatC* p_c_grid;
+    };
+
+    // FIXME: pass GridwiseGemmPipe as a template arguement into GridwiseGemm
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0Number, Number<MPerBlock>{}, AK1Number),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1Number, AK1Number, I1));
+    }
+
+    __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0Number, Number<NPerBlock>{}, BK1Number),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1Number, BK1Number, I1));
+    }
+
+    __device__ static constexpr auto GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned * sizeof(ComputeType) +
+                          b_block_space_size_aligned * sizeof(ComputeType)),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    __host__ static constexpr bool CheckValidity(const Problem& problem)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+            if(!(problem.M % MPerBlock == 0))
+            {
+                return false;
+            }
+        }
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+            if(!(problem.N % NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::KPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding)
+        {
+            if(!(CalculateKPadded(problem.K) % AK1Value == 0) ||
+               !(CalculateKPadded(problem.K) % BK1Value == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(!(problem.K % AK1Value == 0) || !(problem.K % BK1Value == 0))
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            if(problem.K % ABlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.M % ABlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        {
+            if(problem.N % BBlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.K % BBlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+        {
+            if(problem.N % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.M % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+            {
+                return false;
+            }
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = (CalculateAK0(problem.K) * AK1Value) / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    template <typename CGridDesc>
+    __device__ static constexpr auto MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const CGridDesc& c_grid_desc_m_n, index_t MBlock, index_t NBlock)
+    {
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    using Block2CTileMap = BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock>;
+
+    template <bool HasMainKBlockLoop>
+    __device__ static void Run(const FloatA* __restrict__ p_a_grid,
+                               const FloatB* __restrict__ p_b_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               void* __restrict__ p_shared,
+                               const Problem& problem)
+    {
+        const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
+            problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
+        const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
+            problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0);
+        const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(
+            problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                c_grid_desc_m_n, problem.MBlock, problem.NBlock);
+
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        const AElementwiseOperation a_element_op{};
+        const BElementwiseOperation b_element_op{};
+        const CElementwiseOperation c_element_op{};
+
+        // divide block work by [M, N]
+        const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N};
+
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0Number, MPerBlock, AK1Number>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatA,
+                                                ComputeType,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0Number, NPerBlock, BK1Number>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatB,
+                                                ComputeType,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1Number, BK1Number),
+                      MfmaSelector<ComputeType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ComputeType,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ComputeType*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1Number, 0, 0);
+
+        // gridwise GEMM pipeline
+        static_assert(std::is_default_constructible_v<GridwiseGemmPipe>);
+        const auto gridwise_gemm_pipeline = GridwiseGemmPipe{};
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
new file mode 100644
index 0000000000..0404d88ab8
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
@@ -0,0 +1,1070 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+
+namespace ck {
+
+// D = Layernorm(acc_element_op(A * B + broadcast(bias)) + add) * broadcast(gamma) + broadcast(beta)
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename FloatC0,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename CElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename C0GridDescriptor_NBlock_NPerBlock,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_layernorm_xdl_cshuffle_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,               // MxN
+            const FloatC0* __restrict__ p_c0_bias_grid,  // 1xN
+            const FloatC0* __restrict__ p_c0_add_grid,   // MxN
+            const FloatC0* __restrict__ p_c0_gamma_grid, // 1xN
+            const FloatC0* __restrict__ p_c0_beta_grid,  // 1xN
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const AccElementwiseOperation acc_element_op,
+            const CElementwiseOperation c_element_op,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                c_grid_desc_mblock_mperblock_nblock_nperblock,
+            const C0GridDescriptor_NBlock_NPerBlock c0_grid_desc_nblock_nperblock,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    // TODO ANT: separate into MMA + Epilogue
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_c0_bias_grid,
+                                                  p_c0_add_grid,
+                                                  p_c0_gamma_grid,
+                                                  p_c0_beta_grid,
+                                                  p_shared,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  acc_element_op,
+                                                  c_element_op,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  c0_grid_desc_nblock_nperblock,
+                                                  block_2_ctile_map);
+
+    // TODO ANT: Run layernorm epilogue here
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_c0_bias_grid;
+    ignore = p_c0_add_grid;
+    ignore = p_c0_gamma_grid;
+    ignore = p_c0_beta_grid;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = acc_element_op;
+    ignore = c_element_op;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = c0_grid_desc_nblock_nperblock;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+// The GEMM + Layernorm implementation is a specialized kernel which allows fusing both layers
+// together given the condition GEMM extents N of MNK is spanned by a single workgroup. For example,
+// a kernel configured with NPerBlock = 128 allows to operate on all GEMM sizes if N <= 128
+template <typename FloatAB,
+          typename FloatGemmAcc,
+          typename FloatCShuffle,
+          typename FloatC,
+          typename FloatC0,
+          typename FloatReduceAcc, // Data type after shuffle
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename CElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDesc_M_N,
+          typename C0GridDesc_N,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          typename CReduceThreadClusterLengths_MPerBlock_NPerBlock,
+          index_t CReduceThreadCopySrcDstScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmLayernorm_k0mk1_k0nk1_mn_xdl_cshuffle_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        // Align 16 bytes (maximum LDS read/write width)
+        constexpr auto c_block_size_aligned =
+            math::integer_least_multiple(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize() *
+                    sizeof(FloatCShuffle),
+                16) /
+            sizeof(FloatCShuffle);
+
+        // LDS allocation for reduction workspace
+        constexpr index_t c_lds_workspace_size = BlockSize;
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size_aligned * sizeof(FloatCShuffle) +
+                             c_lds_workspace_size * sizeof(FloatReduceAcc));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+            return false;
+
+        // in order to reduce N dim without elaborate sync across CUs in single kernel, one
+        // workgroup must span the entire N extent
+        if(math::integer_divide_ceil(N, NPerBlock) > 1)
+        {
+            return false;
+        }
+
+        // static check: all waves in the workgroups combined must cover whole N extent in order
+        // to have efficient N-dim reduction
+        static_assert(CShuffleNXdlPerWavePerShuffle == NXdlPerWave,
+                      "condition not met for efficient layernorm");
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            c_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return c_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // for bias, beta, gamma
+    __host__ __device__ static constexpr auto
+    MakeC0GridDescriptor_NBlock_NPerBlock(const C0GridDesc_N& c0_grid_desc_n)
+    {
+        const auto N      = c0_grid_desc_n.GetLength(I0);
+        const auto NBlock = N / NPerBlock;
+
+        const auto c0_grid_desc_nblock_nperblock = transform_tensor_descriptor(
+            c0_grid_desc_n,
+            make_tuple(make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0, 1>{}));
+
+        return c0_grid_desc_nblock_nperblock;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            CGridDesc_M_N{}))>;
+
+    using C0GridDescriptor_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeC0GridDescriptor_NBlock_NPerBlock(C0GridDesc_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        const FloatC0* __restrict__ p_c0_bias_grid,  // 1xN
+        const FloatC0* __restrict__ p_c0_add_grid,   // MxN
+        const FloatC0* __restrict__ p_c0_gamma_grid, // 1xN
+        const FloatC0* __restrict__ p_c0_beta_grid,  // 1xN
+        void* __restrict__ p_shared,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const AccElementwiseOperation& acc_element_op,
+        const CElementwiseOperation& c_element_op,
+        const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+        const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+        const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+            c_grid_desc_mblock_mperblock_nblock_nperblock,
+        const C0GridDescriptor_NBlock_NPerBlock& c0_grid_desc_nblock_nperblock,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+        auto c0_bias_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c0_bias_grid, c0_grid_desc_nblock_nperblock.GetElementSpaceSize());
+        // Note: c0_add is of same layout as c so we don't declare new c0_add_desc here
+        auto c0_add_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c0_add_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+        auto c0_gamma_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c0_gamma_grid, c0_grid_desc_nblock_nperblock.GetElementSpaceSize());
+        auto c0_beta_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c0_beta_grid, c0_grid_desc_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            FloatAB,
+            FloatGemmAcc,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                               a_block_desc_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_bk0_n_bk1,
+                                                               b_block_desc_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    tensor_operation::element_wise::PassThrough{}};
+
+            // shuffle: blockwise copy C from LDS to global
+            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,        // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                3,                                              // index_t VectorDim,
+                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                 c_element_op};
+
+            const auto NBlock = c0_grid_desc_nblock_nperblock.GetLength(I0);
+
+            // for broadcasting bias, beta, gamma
+            const auto c0_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+                c0_grid_desc_nblock_nperblock,
+                make_tuple(make_insert_transform(I1),
+                           make_insert_transform(I1),
+                           make_pass_through_transform(NBlock),
+                           make_pass_through_transform(NPerBlock)),
+                make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            // LDS c_reduce_block_desc_mperblock_nperblock
+            constexpr auto c_reduce_block_desc_mperblock_nperblock = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I1)),
+                    make_freeze_transform(I0),
+                    make_pass_through_transform(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetLength(I3))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<>{}, Sequence<0>{}, Sequence<>{}, Sequence<1>{}));
+
+            static_assert(CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) *
+                                  CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                              BlockSize,
+                          "wrong!");
+
+            static_assert((CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) %
+                                      CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0) ==
+                                  0 &&
+                              (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) %
+                                      CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1) ==
+                                  0,
+                          "wrong!");
+
+            constexpr index_t mreduce_per_thread =
+                (CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl) /
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I0);
+
+            constexpr index_t nreduce_per_thread =
+                (CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl) /
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock::At(I1);
+
+            constexpr auto c_reduce_thread_lengths_mperblock_nperblock =
+                Sequence<mreduce_per_thread, nreduce_per_thread>{};
+
+            // pytorch default
+            // https://pytorch.org/docs/stable/generated/torch.nn.LayerNorm.html
+            static constexpr FloatReduceAcc epsilon = 1e-5;
+
+            // VGPR c_reduce_thread_desc_mperblock_nperblock
+            constexpr auto c_reduce_thread_desc_mperblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(Number<mreduce_per_thread>{}, Number<nreduce_per_thread>{}));
+
+            constexpr auto c_reduce_thread_desc_mblock_mperblock_nblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mreduce_per_thread>{}, I1, Number<nreduce_per_thread>{}));
+
+            // VGPR d_reduce_thread_desc_mperblock
+            constexpr auto d_reduce_thread_desc_mperblock =
+                make_naive_tensor_descriptor_packed(make_tuple(Number<mreduce_per_thread>{}));
+
+            // TODO: this should be implemented as a blockwise reduction
+            auto c_reduce_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                c_reduce_thread_desc_mperblock_nperblock.GetElementSpaceSize());
+
+            auto c0_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatC0>(
+                c_reduce_thread_desc_mperblock_nperblock.GetElementSpaceSize());
+
+            // Align 16 bytes (maximum LDS read/write width)
+            constexpr auto c_block_size_aligned =
+                math::integer_least_multiple(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize() *
+                        sizeof(FloatCShuffle),
+                    16) /
+                sizeof(FloatCShuffle);
+
+            auto d_reduce_work_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                reinterpret_cast<FloatReduceAcc*>(static_cast<FloatCShuffle*>(p_shared) +
+                                                  c_block_size_aligned),
+                BlockSize);
+
+            // Sum thread workspace
+            auto d0_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                d_reduce_thread_desc_mperblock.GetElementSpaceSize());
+
+            // Squared sum thread workspace
+            auto d1_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatReduceAcc>(
+                d_reduce_thread_desc_mperblock.GetElementSpaceSize());
+
+            // reduce: threadwise copy from LDS to VGPR
+            constexpr auto c_reduce_thread_cluster_desc = make_cluster_descriptor(
+                CReduceThreadClusterLengths_MPerBlock_NPerBlock{}, Sequence<1, 0>{});
+
+            const auto c_reduce_thread_cluster_idx =
+                c_reduce_thread_cluster_desc.CalculateBottomIndex(
+                    make_multi_index(get_thread_local_1d_id()));
+
+            const auto c_reduce_thread_data_idx_begin =
+                c_reduce_thread_cluster_idx * c_reduce_thread_lengths_mperblock_nperblock;
+
+            auto c_reduce_thread_copy_lds_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatCShuffle,
+                FloatReduceAcc,
+                decltype(c_reduce_block_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_desc_mperblock_nperblock),
+                decltype(c_reduce_thread_lengths_mperblock_nperblock),
+                Sequence<0, 1>,
+                1,
+                CReduceThreadCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>{c_reduce_block_desc_mperblock_nperblock, c_reduce_thread_data_idx_begin};
+
+            auto c_reduce_thread_copy_vgpr_to_lds = ThreadwiseTensorSliceTransfer_v1r3<
+                FloatReduceAcc,
+                FloatCShuffle,
+                decltype(c_reduce_thread_desc_mperblock_nperblock),
+                decltype(c_reduce_block_desc_mperblock_nperblock),
+                tensor_operation::element_wise::PassThrough,
+                decltype(c_reduce_thread_lengths_mperblock_nperblock),
+                Sequence<0, 1>,
+                1,
+                CReduceThreadCopySrcDstScalarPerVector_NPerBlock,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{c_reduce_block_desc_mperblock_nperblock,
+                      c_reduce_thread_data_idx_begin,
+                      tensor_operation::element_wise::PassThrough{}};
+
+            auto c0_thread_copy_global_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatC0,
+                FloatC0,
+                decltype(c0_grid_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_reduce_thread_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CReduceThreadCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>(c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                      make_multi_index(block_work_idx[I0],
+                                       c_reduce_thread_data_idx_begin[I0],
+                                       block_work_idx[I1],
+                                       c_reduce_thread_data_idx_begin[I1]));
+
+            // Note: c0_add is of same layout as c so we don't declare new c0_add_desc here
+            auto c0_add_thread_copy_global_to_vgpr = ThreadwiseTensorSliceTransfer_v2<
+                FloatC0,
+                FloatC0,
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_reduce_thread_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<I1, mreduce_per_thread, I1, nreduce_per_thread>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CReduceThreadCopySrcDstScalarPerVector_NPerBlock,
+                1,
+                true>(c_grid_desc_mblock_mperblock_nblock_nperblock,
+                      make_multi_index(block_work_idx[I0],
+                                       c_reduce_thread_data_idx_begin[I0],
+                                       block_work_idx[I1],
+                                       c_reduce_thread_data_idx_begin[I1]));
+
+            // space filling curve for threadwise C in VGPR
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C in global mem
+            constexpr auto sfc_c_global =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                block_sync_lds();
+
+                // load from LDS and global, add bias
+                c_reduce_thread_copy_lds_to_vgpr.Run(c_reduce_block_desc_mperblock_nperblock,
+                                                     c_shuffle_block_buf,
+                                                     c_reduce_thread_desc_mperblock_nperblock,
+                                                     make_tuple(I0, I0),
+                                                     c_reduce_thread_buf);
+
+                c0_thread_copy_global_to_vgpr.Run(
+                    c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c0_bias_grid_buf,
+                    c_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                    make_tuple(I0, I0, I0, I0),
+                    c0_thread_buf);
+
+                static_for<0, c_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                    [&](auto i) {
+                        FloatReduceAcc out;
+                        acc_element_op(out,
+                                       c_reduce_thread_buf(i) +
+                                           static_cast<FloatReduceAcc>(c0_thread_buf(i)));
+                        c_reduce_thread_buf(i) = out; // acc_element_op(acc + bias)
+                    });
+
+                c0_add_thread_copy_global_to_vgpr.Run(
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c0_add_grid_buf,
+                    c_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                    make_tuple(I0, I0, I0, I0),
+                    c0_thread_buf);
+
+                static_for<0, c_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                    [&](auto i) {
+                        c_reduce_thread_buf(i) +=
+                            static_cast<FloatReduceAcc>(c0_thread_buf(i)); // add
+                    });
+
+                // layernorm
+                {
+                    using ThreadwiseReduceD0 =
+                        ThreadwiseReduction<FloatReduceAcc,
+                                            decltype(c_reduce_thread_desc_mperblock_nperblock),
+                                            decltype(d_reduce_thread_desc_mperblock),
+                                            reduce::Add,
+                                            false>;
+                    using ThreadwiseReduceD1 =
+                        ThreadwiseReduction<FloatReduceAcc,
+                                            decltype(c_reduce_thread_desc_mperblock_nperblock),
+                                            decltype(d_reduce_thread_desc_mperblock),
+                                            reduce::SquaredAdd,
+                                            false>;
+
+                    const auto d0_zeroVal =
+                        ThreadwiseReduceD0::Op::template GetIdentityValue<FloatReduceAcc>();
+                    const auto d1_zeroVal =
+                        ThreadwiseReduceD1::Op::template GetIdentityValue<FloatReduceAcc>();
+                    static_for<0, mreduce_per_thread, 1>{}(
+                        [&](auto i) { d0_thread_buf(i) = d0_zeroVal; });
+                    static_for<0, mreduce_per_thread, 1>{}(
+                        [&](auto i) { d1_thread_buf(i) = d1_zeroVal; });
+
+                    // reduce sum in VGPR
+                    ThreadwiseReduceD0::Reduce(c_reduce_thread_buf, d0_thread_buf);
+
+                    // reduce squared sum in VGPR
+                    ThreadwiseReduceD1::Reduce(c_reduce_thread_buf, d1_thread_buf);
+
+                    // reduce within workgroup
+                    using BlockwiseReduce = PartitionedBlockwiseReduction<
+                        FloatReduceAcc,
+                        BlockSize,
+                        CReduceThreadClusterLengths_MPerBlock_NPerBlock, // ThreadClusterLengths_M_K
+                        Sequence<1, 0>, // ThreadClusterArrangeOrder
+                        reduce::Add,
+                        false>;
+
+                    static_for<0, mreduce_per_thread, 1>{}([&](auto i) {
+                        block_sync_lds();
+                        BlockwiseReduce::Reduce(d_reduce_work_buf,
+                                                d0_thread_buf(i)); // blockwise reduced sum
+                        block_sync_lds();
+                        BlockwiseReduce::Reduce(d_reduce_work_buf,
+                                                d1_thread_buf(i)); // blockwise reduced squared sum
+                    });
+
+                    // normalize
+                    const index_t NRaw =
+                        c_grid_desc_mblock_mperblock_nblock_nperblock.GetTransforms()[I0]
+                            .GetUpperLengths()[I1]; // TODO: proper handle
+
+                    static_for<0, mreduce_per_thread, 1>{}([&](auto im) {
+                        static_for<0, nreduce_per_thread, 1>{}([&](auto in) {
+                            constexpr auto dst_offset =
+                                Number<c_reduce_thread_desc_mperblock_nperblock.CalculateOffset(
+                                    make_tuple(im, in))>{};
+
+                            constexpr auto src_offset =
+                                Number<d_reduce_thread_desc_mperblock.CalculateOffset(
+                                    make_tuple(im))>{};
+
+                            FloatReduceAcc avg_sum         = d0_thread_buf(src_offset) / NRaw;
+                            FloatReduceAcc avg_squared_sum = d1_thread_buf(src_offset) / NRaw;
+
+                            FloatReduceAcc numerator = c_reduce_thread_buf(dst_offset) - avg_sum;
+                            FloatReduceAcc divisor = epsilon + avg_squared_sum - avg_sum * avg_sum;
+                            FloatReduceAcc divisor_sqrt;
+                            tensor_operation::element_wise::UnarySqrt{}(divisor_sqrt, divisor);
+
+                            c_reduce_thread_buf(dst_offset) = numerator / divisor_sqrt;
+                        });
+                    });
+
+                    // scaling
+                    c0_thread_copy_global_to_vgpr.Run(
+                        c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                        c0_gamma_grid_buf,
+                        c_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        c0_thread_buf);
+
+                    static_for<0, c_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                        [&](auto i) {
+                            c_reduce_thread_buf(i) *=
+                                static_cast<FloatReduceAcc>(c0_thread_buf(i)); // * gamma
+                        });
+
+                    c0_thread_copy_global_to_vgpr.Run(
+                        c0_grid_desc_mblock_mperblock_nblock_nperblock,
+                        c0_beta_grid_buf,
+                        c_reduce_thread_desc_mblock_mperblock_nblock_nperblock,
+                        make_tuple(I0, I0, I0, I0),
+                        c0_thread_buf);
+
+                    static_for<0, c_reduce_thread_desc_mperblock_nperblock.GetElementSize(), 1>{}(
+                        [&](auto i) {
+                            c_reduce_thread_buf(i) +=
+                                static_cast<FloatReduceAcc>(c0_thread_buf(i)); // + beta
+                        });
+
+                    block_sync_lds();
+
+                    c_reduce_thread_copy_vgpr_to_lds.Run(c_reduce_thread_desc_mperblock_nperblock,
+                                                         make_tuple(I0, I0),
+                                                         c_reduce_thread_buf,
+                                                         c_reduce_block_desc_mperblock_nperblock,
+                                                         c_shuffle_block_buf);
+
+                } // end layernorm
+
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                c_shuffle_block_copy_lds_to_global.Run(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    c_shuffle_block_buf,
+                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                    c_grid_buf);
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                    // move on C
+                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+
+                    // move on C0
+                    c0_thread_copy_global_to_vgpr.MoveSrcSliceWindow(
+                        c0_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+
+                    // move on C0_add
+                    c0_add_thread_copy_global_to_vgpr.MoveSrcSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp
new file mode 100644
index 0000000000..bbd01a238e
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp
@@ -0,0 +1,745 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_waveletmodel.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename ABDataType,
+          typename FloatGemmAcc,
+          typename EDataTypeShuffle,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename EElementwiseOperation,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_M_K,
+          typename BGridDesc_N_K,
+          typename EGridDesc_M_N,
+          index_t NumGemmKPrefetchStage,
+          index_t TileLoadThreadGroupSize,
+          index_t TileMathThreadGroupSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle
+{
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    struct TileLoadThreadGroup
+    {
+        __device__ static constexpr index_t GetNumOfThread() { return TileLoadThreadGroupSize; }
+
+        __device__ static constexpr bool IsBelong()
+        {
+            return (get_thread_local_1d_id() >= TileLoadThreadGroupSize);
+        }
+
+        __device__ static index_t GetThreadId()
+        {
+            return get_thread_local_1d_id() - TileMathThreadGroupSize;
+        }
+    };
+
+    struct TileMathThreadGroup
+    {
+        __device__ static constexpr index_t GetNumOfThread() { return TileMathThreadGroupSize; }
+
+        __device__ static constexpr bool IsBelong()
+        {
+            return get_thread_local_1d_id() < TileMathThreadGroupSize;
+        }
+
+        __device__ static index_t GetThreadId() { return get_thread_local_1d_id(); }
+    };
+
+    using CShuffleBlockTransferThreadGroup = ThisThreadBlock<TileMathThreadGroupSize>;
+
+    // load and math+store Wave pipelines.
+    // TODO: build pipelines blocks scheduling parallel tasks
+    using GridwiseGemmLoad = GridwiseGemmLoadWave<TileLoadThreadGroup, NumGemmKPrefetchStage>;
+    using GridwiseGemmMath = GridwiseGemmMathWave<TileMathThreadGroup, NumGemmKPrefetchStage>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(ABDataType),
+                         c_block_size * sizeof(EDataTypeShuffle));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2ETileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_M_K& a_grid_desc_m_k,
+                  const BGridDesc_N_K& b_grid_desc_n_k,
+                  const EGridDesc_M_N& e_grid_desc_m_n,
+                  const Block2ETileMap& /*block_2_etile_map*/)
+    {
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        // check consistency of desc
+        if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1) &&
+             K == b_grid_desc_n_k.GetLength(I1)))
+        {
+            return false;
+        }
+
+        // check tile size
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+        {
+            return false;
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmMath::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+
+        // check tensor size: cannot be larger than 2GB each
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_m_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             b_grid_desc_n_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmMath::CalculateHasMainLoop(num_loop);
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        constexpr auto M01 = I1;
+        constexpr auto N01 = I1;
+
+        const auto m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_unmerge_transform(make_tuple(M0, M01)),
+                           make_unmerge_transform(make_tuple(N0, N01))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}));
+
+        const auto cblockid_to_m00_m01_n00_n01_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(M0, N0, M01, N01))),
+                make_tuple(Sequence<0, 1, 2, 3>{}),
+                make_tuple(Sequence<0>{}));
+
+        const auto cblockid_to_m0_n0_block_cluster_adaptor =
+            chain_tensor_adaptors(m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor,
+                                  cblockid_to_m00_m01_n00_n01_block_cluster_adaptor);
+
+        return cblockid_to_m0_n0_block_cluster_adaptor;
+    }
+
+    __host__ __device__ static constexpr index_t
+    CalculateGridSize(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const index_t grid_size = (M / MPerBlock) * (N / NPerBlock);
+
+        return grid_size;
+    }
+
+    // A desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultBGridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDescriptor_M_N>
+    __host__ __device__ static constexpr auto MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const EGridDescriptor_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+
+    using DefaultBlock2ETileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AK0_M_AK1,
+              typename BGridDesc_BK0_N_BK1,
+              typename Block2ETileMap>
+    __device__ static void Run(const ABDataType* __restrict__ p_a_grid,
+                               const ABDataType* __restrict__ p_b_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const EElementwiseOperation& e_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        // build loadWave and MathWave pipelines
+        // loadWave and MathWave synchronized through LDS
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        if(TileLoadThreadGroup::IsBelong())
+        {
+
+            // LoadWave
+            const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+            const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+
+            // A matrix blockwise copy
+            auto a_blockwise_copy =
+                ThreadGroupTensorSliceTransfer_v4r1<TileLoadThreadGroup,
+                                                    AElementwiseOperation,
+                                                    ck::tensor_operation::element_wise::PassThrough,
+                                                    InMemoryDataOperationEnum::Set,
+                                                    Sequence<AK0PerBlock, MPerBlock, AK1>,
+                                                    ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                    ABlockTransferThreadClusterArrangeOrder,
+                                                    ABDataType,
+                                                    ABDataType,
+                                                    decltype(a_grid_desc_ak0_m_ak1),
+                                                    decltype(a_block_desc_ak0_m_ak1),
+                                                    ABlockTransferSrcAccessOrder,
+                                                    Sequence<1, 0, 2>,
+                                                    ABlockTransferSrcVectorDim,
+                                                    2,
+                                                    ABlockTransferSrcScalarPerVector,
+                                                    ABlockTransferDstScalarPerVector_AK1,
+                                                    1,
+                                                    1,
+                                                    AThreadTransferSrcResetCoordinateAfterRun,
+                                                    true,
+                                                    NumGemmKPrefetchStage>(
+                    a_grid_desc_ak0_m_ak1,
+                    make_multi_index(0, m_block_data_idx_on_grid, 0),
+                    a_element_op,
+                    a_block_desc_ak0_m_ak1,
+                    make_multi_index(0, 0, 0),
+                    ck::tensor_operation::element_wise::PassThrough{});
+
+            // B matrix blockwise copy
+            auto b_blockwise_copy =
+                ThreadGroupTensorSliceTransfer_v4r1<TileLoadThreadGroup,
+                                                    BElementwiseOperation,
+                                                    ck::tensor_operation::element_wise::PassThrough,
+                                                    InMemoryDataOperationEnum::Set,
+                                                    Sequence<BK0PerBlock, NPerBlock, BK1>,
+                                                    BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                    BBlockTransferThreadClusterArrangeOrder,
+                                                    ABDataType,
+                                                    ABDataType,
+                                                    decltype(b_grid_desc_bk0_n_bk1),
+                                                    decltype(b_block_desc_bk0_n_bk1),
+                                                    BBlockTransferSrcAccessOrder,
+                                                    Sequence<1, 0, 2>,
+                                                    BBlockTransferSrcVectorDim,
+                                                    2,
+                                                    BBlockTransferSrcScalarPerVector,
+                                                    BBlockTransferDstScalarPerVector_BK1,
+                                                    1,
+                                                    1,
+                                                    BThreadTransferSrcResetCoordinateAfterRun,
+                                                    true,
+                                                    NumGemmKPrefetchStage>(
+                    b_grid_desc_bk0_n_bk1,
+                    make_multi_index(0, n_block_data_idx_on_grid, 0),
+                    b_element_op,
+                    b_block_desc_bk0_n_bk1,
+                    make_multi_index(0, 0, 0),
+                    ck::tensor_operation::element_wise::PassThrough{});
+
+            GridwiseGemmLoad::template RunLoadWavePipeline<HasMainKBlockLoop>(
+                a_grid_desc_ak0_m_ak1,
+                a_block_desc_ak0_m_ak1,
+                a_blockwise_copy,
+                a_grid_buf,
+                a_block_buf,
+                a_block_slice_copy_step,
+                b_grid_desc_bk0_n_bk1,
+                b_block_desc_bk0_n_bk1,
+                b_blockwise_copy,
+                b_grid_buf,
+                b_block_buf,
+                b_block_slice_copy_step,
+                num_k_block_main_loop);
+
+            block_sync_lds();
+            block_sync_lds();
+        }
+        else if(TileMathThreadGroup::IsBelong())
+        {
+            // branch early for math wave
+            constexpr index_t KPack =
+                math::max(math::lcm(AK1, BK1),
+                          MfmaSelector<ABDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+            auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<
+                TileMathThreadGroupSize,
+                ABDataType,
+                FloatGemmAcc,
+                decltype(a_block_desc_ak0_m_ak1),
+                decltype(b_block_desc_bk0_n_bk1),
+                MPerXdl,
+                NPerXdl,
+                MXdlPerWave,
+                NXdlPerWave,
+                KPack>{};
+
+            auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+            auto c_grid_buf   = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            // TODO re-architect LDS+math stages
+            // Writing data to GMEM: only math wave is doing the work in cshuffle
+            GridwiseGemmMath::template RunMathWavePipeline<HasMainKBlockLoop>(
+                a_block_buf, b_block_buf, blockwise_gemm, c_thread_buf, num_k_block_main_loop);
+
+            // GEMM definition
+            //   c_mtx += transpose(a_mtx) * b_mtx
+            //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+            //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+            //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+            //       register
+            // sanity check
+
+            // shuffle C and write out
+            {
+                static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                                  NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                              "wrong!");
+
+                constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+                constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+                // TODO: hacky, fix it!
+                constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                    blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+                // TODO: hacky, fix it!
+                // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+                constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                    blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+                constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+                constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+                constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+                constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+                constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+                constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+                constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+                constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+                constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+                auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                    static_cast<EDataTypeShuffle*>(p_shared),
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+                constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                    make_tuple(
+                        make_freeze_transform(I0),
+                        make_unmerge_transform(make_tuple(
+                            Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                            M1,                                      // M1 = MWave
+                            M2,                                      // M2 * M3 * M4 = MPerXdl
+                            M3,
+                            M4)),
+                        make_freeze_transform(I0),
+                        make_unmerge_transform(make_tuple(
+                            Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                            N1,                                      // N1 = NWave
+                            N2))),                                   // N2 = NPerXdl
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<>{},
+                               Sequence<0, 2, 4, 5, 6>{},
+                               Sequence<>{},
+                               Sequence<1, 3, 7>{}));
+
+                // calculate origin of thread output tensor on global memory
+                // blockwise GEMM c matrix starting index
+                const auto c_thread_mtx_on_block =
+                    blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+                const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+                const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+                const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                    make_single_stage_tensor_adaptor(
+                        make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                        make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                        make_tuple(Sequence<0>{}));
+
+                const auto m_thread_data_on_block_idx =
+                    m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                        make_multi_index(m_thread_data_on_block));
+
+                const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                    make_single_stage_tensor_adaptor(
+                        make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                        make_tuple(Sequence<0, 1, 2>{}),
+                        make_tuple(Sequence<0>{}));
+
+                const auto n_thread_data_on_block_idx =
+                    n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                        make_multi_index(n_thread_data_on_block));
+
+                // shuffle: threadwise copy C from VGPR to LDS
+                auto c_thread_copy_vgpr_to_lds = ThreadwiseTensorSliceTransfer_v1r3<
+                    FloatGemmAcc,
+                    EDataTypeShuffle,
+                    decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                    decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                    ck::tensor_operation::element_wise::PassThrough,
+                    Sequence<CShuffleMXdlPerWavePerShuffle,
+                             CShuffleNXdlPerWavePerShuffle,
+                             I1,
+                             I1,
+                             M2,
+                             I1,
+                             M4,
+                             I1>,
+                    Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                    7,
+                    1,
+                    InMemoryDataOperationEnum::Set,
+                    1,
+                    true>{c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                          make_multi_index(0,
+                                           0,
+                                           m_thread_data_on_block_idx[I1],
+                                           n_thread_data_on_block_idx[I1],
+                                           m_thread_data_on_block_idx[I2],
+                                           m_thread_data_on_block_idx[I3],
+                                           m_thread_data_on_block_idx[I4],
+                                           n_thread_data_on_block_idx[I2]),
+                          ck::tensor_operation::element_wise::PassThrough{}};
+
+                // shuffle: blockwise copy C from LDS to global
+                auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                    CShuffleBlockTransferThreadGroup, // ThreadGroup
+                    EElementwiseOperation,            // ElementwiseOperation,
+                    CGlobalMemoryDataOperation,       // DstInMemOp,
+                    Sequence<1,
+                             CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                             1,
+                             CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                    CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                    EDataTypeShuffle,     // typename SrcData,
+                    EDataType,            // typename DstData,
+                    decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                    decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
+                    3,                                              // index_t VectorDim,
+                    CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
+                    true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                    false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                    {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                     make_multi_index(0, 0, 0, 0),
+                     e_grid_desc_mblock_mperblock_nblock_nperblock,
+                     make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
+                     e_element_op};
+
+                // space filling curve for threadwise C in VGPR
+                constexpr auto sfc_c_vgpr =
+                    SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                      Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                      Sequence<CShuffleMXdlPerWavePerShuffle,
+                                               CShuffleNXdlPerWavePerShuffle,
+                                               1,
+                                               1,
+                                               M2,
+                                               1,
+                                               M4,
+                                               1>>{};
+
+                // space filling curve for shuffled blockwise C in global mem
+                constexpr auto sfc_c_global =
+                    SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                      Sequence<0, 2, 1, 3>,
+                                      Sequence<1,
+                                               CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                               1,
+                                               CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+                constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+                static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
+
+                // Different way of getting coalesced writes:
+                // We can get rid of doing cshuffle. Instead of reading A rows in contiguous manner
+                // do it interleaved, then mfma can have nice c-mat layout as below:
+                //
+                // TODO
+                //      We do not need to do LDS swizzle to align global writes writing cache lines:
+                //         v_mfma  cmat, amat, bmat, cmat   - c-mat register layout   are 1xN
+                //                                            elments  (N is vertical or strided
+                //                                            dimension)
+                //         v_mfma  cmat, bmat, amat, cmat   - c-mat register layout   are Mx1
+                //         elments  (M is coalescing
+                //                                            dimension) by enumerating M index in
+                //                                            amat, bmat you can align cmat
+                //                                            register(s) to contiguous M elements
+                //                                            for example
+                //              1st mfma instruction  output space : 0 4 8  12 16 ....
+                //              2nd mfma instruction  output space : 1 5 9  13 17 ....
+                //              3rd mfma instruction  output space : 2 6 10 14 18 ....
+                //              4th mfma instruction  output space : 3 7 11 15 19 ....
+                //              you can pack 4 registers output space into 2WORD and do global write
+                //              (no LDS swizzling required)
+
+                static_for<0, num_access, 1>{}([&](auto access_id) {
+                    // make sure it's safe to write to LDS
+                    block_sync_lds();
+
+                    // each thread write its data from VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                                  c_thread_buf,
+                                                  c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                  c_shuffle_block_buf);
+                    // make sure it's safe to read from LDS
+                    block_sync_lds();
+
+                    // each block copy its data from LDS to global
+                    c_shuffle_block_copy_lds_to_global.Run(
+                        c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                        c_shuffle_block_buf,
+                        e_grid_desc_mblock_mperblock_nblock_nperblock,
+                        c_grid_buf);
+
+                    if constexpr(access_id < num_access - 1)
+                    {
+                        constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
+
+                        // move on C
+                        c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
+                            e_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
+                    }
+                });
+            }
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
new file mode 100644
index 0000000000..0920a17fc7
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
@@ -0,0 +1,988 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+// Implementation of "Merge" transformation primitive that uses division and mod. It is supposed to
+// be used for low_lengths that are known at compile time and are power of 2, otherwise performance
+// will be very bad
+template <typename LowLengths>
+struct Merge_v4_no_carry
+{
+    static constexpr index_t NDimLow = LowLengths::Size();
+
+    using LowerIndex = MultiIndex<NDimLow>;
+    using UpperIndex = MultiIndex<1>;
+
+    using LowLengthsScan =
+        decltype(container_reverse_exclusive_scan(LowLengths{}, math::multiplies{}, Number<1>{}));
+
+    using UpLengths =
+        decltype(make_tuple(container_reduce(LowLengths{}, math::multiplies{}, Number<1>{})));
+
+    LowLengths low_lengths_;
+    LowLengthsScan low_lengths_scan_;
+    UpLengths up_lengths_;
+
+    __host__ __device__ constexpr Merge_v4_no_carry() = default;
+
+    __host__ __device__ constexpr Merge_v4_no_carry(const LowLengths& low_lengths)
+        : low_lengths_{low_lengths},
+          low_lengths_scan_{
+              container_reverse_exclusive_scan(low_lengths, math::multiplies{}, Number<1>{})},
+          up_lengths_{make_tuple(container_reduce(low_lengths, math::multiplies{}, Number<1>{}))}
+    {
+        static_assert(LowerIndex::Size() == NDimLow, "wrong!");
+    }
+
+    __host__ __device__ static constexpr index_t GetNumOfLowerDimension() { return NDimLow; }
+
+    __host__ __device__ static constexpr index_t GetNumOfUpperDimension() { return 1; }
+
+    __host__ __device__ constexpr const auto& GetUpperLengths() const { return up_lengths_; }
+
+    template <typename LowIdx, typename UpIdx>
+    __host__ __device__ constexpr void CalculateLowerIndex(LowIdx& idx_low,
+                                                           const UpIdx& idx_up) const
+    {
+        static_assert(LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        index_t tmp = idx_up[Number<0>{}];
+
+        // division and mod
+        static_for<0, NDimLow - 1, 1>{}([&](auto i) {
+            idx_low(i) = tmp / this->low_lengths_scan_[i];
+            tmp %= this->low_lengths_scan_[i];
+        });
+
+        idx_low(Number<NDimLow - 1>{}) = tmp;
+    }
+
+    template <typename LowIdxDiff,
+              typename UpIdxDiff,
+              typename LowIdx,
+              typename UpIdx,
+              index_t Hack>
+    __host__ __device__ void UpdateLowerIndex(LowIdxDiff& idx_diff_low,
+                                              const UpIdxDiff& idx_up_diff,
+                                              LowIdx& idx_low,
+                                              const UpIdx& idx_up_new,
+                                              Number<Hack>) const
+    {
+        static_assert(LowIdxDiff::Size() == NDimLow && UpIdxDiff::Size() == 1 &&
+                          LowIdx::Size() == NDimLow && UpIdx::Size() == 1,
+                      "wrong! inconsistent # of dimension");
+
+        constexpr auto I0   = Number<0>{};
+        constexpr auto INm1 = Number<NDimLow - 1>{};
+
+        index_t tmp = idx_up_new[I0];
+
+        idx_low(INm1)      = tmp;
+        idx_diff_low(INm1) = idx_up_diff[I0];
+    }
+
+    __host__ __device__ static constexpr bool IsLinearTransform() { return false; }
+
+    __host__ __device__ static constexpr bool IsValidUpperIndexAlwaysMappedToValidLowerIndex()
+    {
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool IsKnownAtCompileTime()
+    {
+        return is_known_at_compile_time<LowLengths>::value &&
+               is_known_at_compile_time<LowLengthsScan>::value &&
+               is_known_at_compile_time<UpLengths>::value;
+    }
+
+    template <typename UpIdx>
+    __host__ __device__ static constexpr bool
+    IsValidUpperIndexMappedToValidLowerIndex(const UpIdx& /* idx_up */)
+    {
+        return true;
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("{");
+        printf("Merge_v3_direct_division_mod_wrw, ");
+        printf("low_lengths_ ");
+        print_multi_index(low_lengths_);
+        printf("low_lengths_scan_ ");
+        print_multi_index(low_lengths_scan_);
+        printf("up_lengths_ ");
+        print_multi_index(up_lengths_);
+        printf("}");
+    }
+};
+
+template <typename LowLengths>
+__host__ __device__ constexpr auto make_merge_transform_v4_no_carry(const LowLengths& low_lengths)
+{
+    return Merge_v4_no_carry<LowLengths>{low_lengths};
+}
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_B_K0_M_K1,
+          typename BGridDesc_B_K0_N_K1,
+          typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename CBlockClusterAdaptor,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_bwd_weight(const FloatAB* __restrict__ p_a_grid,
+                                      const FloatAB* __restrict__ p_b_grid,
+                                      FloatC* __restrict__ p_c_grid,
+                                      const AGridDesc_B_K0_M_K1 a_b_k0_m_k1_grid_desc,
+                                      const BGridDesc_B_K0_N_K1 b_b_k0_n_k1_grid_desc,
+                                      const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                                          c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                      const AElementwiseOperation a_element_op,
+                                      const BElementwiseOperation b_element_op,
+                                      const CElementwiseOperation c_element_op,
+                                      const CBlockClusterAdaptor c_block_cluster_adaptor)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_shared,
+                                                  a_b_k0_m_k1_grid_desc,
+                                                  b_b_k0_n_k1_grid_desc,
+                                                  c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  c_block_cluster_adaptor);
+#else
+    ignore                = p_a_grid;
+    ignore                = p_b_grid;
+    ignore                = p_c_grid;
+    ignore                = a_b_k0_m_k1_grid_desc;
+    ignore                = b_b_k0_n_k1_grid_desc;
+    ignore                = c_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore                = a_element_op;
+    ignore                = b_element_op;
+    ignore                = c_element_op;
+    ignore                = c_block_cluster_adaptor;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_B_K0_M_K1,
+          typename BGridDesc_B_K0_N_K1,
+          typename CMNGridDesc,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MRepeat,
+          index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          bool ABlockLdsExtraM,
+          index_t ABlockLdsM1PerBlock,
+          index_t ABlockLdsM0PerBlock,
+          index_t ABlockLdsM1Padding,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          bool BBlockLdsExtraN,
+          index_t BBlockLdsN1PerBlock,
+          index_t BBlockLdsN0PerBlock,
+          index_t BBlockLdsN1Padding,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXDL,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          bool ABlockLdsExtraM1Wrw      = false,
+          bool BBlockLdsExtraN1Wrw      = false,
+          index_t NumGemmKPrefetchStage = 1,
+          PipelineVersion PipelineVer   = PipelineVersion::v1>
+struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    // denorm test fix, required to work around fp16 mfma issue
+    // we convert fp16->fp32->bf16 and execute bf16 mfma instruction
+    // when mfma if fixed, remove this section and update
+    // FloatABAdjusted -> FloatAB throughout this file
+#if CK_WORKAROUND_DENORM_FIX
+    using FloatABAdjusted = conditional_t<is_same_v<FloatAB, ck::half_t>, ck::bhalf_t, FloatAB>;
+#else
+    using FloatABAdjusted = FloatAB;
+#endif
+
+    // M0/M1/M1Padding
+    static constexpr auto M1PerBlock = Number<ABlockLdsM1PerBlock>{};
+    static constexpr auto M0PerBlock = Number<ABlockLdsM0PerBlock>{};
+    static constexpr auto M1Padding  = Number<ABlockLdsM1Padding>{};
+
+    // N0/N1/N1Padding
+    static constexpr auto N1PerBlock = Number<BBlockLdsN1PerBlock>{};
+    static constexpr auto N0PerBlock = Number<BBlockLdsN0PerBlock>{};
+    static constexpr auto N1Padding  = Number<BBlockLdsN1Padding>{};
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                if constexpr(ABlockLdsExtraM1Wrw)
+                {
+                    constexpr auto a_block_desc_k0_m0_m1_k1 = make_naive_tensor_descriptor(
+                        make_tuple(
+                            Number<K0PerBlock>{}, Number<M0PerBlock>{}, Number<M1PerBlock>{}, K1),
+                        make_tuple(Number<M0PerBlock>{} * (Number<M1PerBlock>{} * K1 + M1Padding),
+                                   Number<M1PerBlock>{} * K1 + M1Padding,
+                                   K1,
+                                   I1));
+
+                    constexpr auto a_block_desc_k0_m_k1_tmp = transform_tensor_descriptor(
+                        a_block_desc_k0_m0_m1_k1,
+                        make_tuple(make_pass_through_transform(Number<K0PerBlock>{}),
+                                   make_merge_transform_v3_division_mod(
+                                       make_tuple(Number<M0PerBlock>{}, Number<M1PerBlock>{})),
+                                   make_pass_through_transform(K1)),
+                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
+                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+                    return a_block_desc_k0_m_k1_tmp;
+                }
+                else
+                {
+                    return make_naive_tensor_descriptor(
+                        make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                        make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+                }
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return a_block_desc_k0_m_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_Batch_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_b_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                if constexpr(ABlockLdsExtraM1Wrw)
+                {
+                    constexpr auto a_block_desc_b_k0_m0_m1_k1 = make_naive_tensor_descriptor(
+                        make_tuple(Number<1>{},
+                                   Number<K0PerBlock>{},
+                                   Number<M0PerBlock>{},
+                                   Number<M1PerBlock>{},
+                                   K1),
+                        make_tuple(Number<K0PerBlock>{} * Number<M0PerBlock>{} *
+                                       (Number<M1PerBlock>{} * K1 + M1Padding),
+                                   Number<M0PerBlock>{} * (Number<M1PerBlock>{} * K1 + M1Padding),
+                                   Number<M1PerBlock>{} * K1 + M1Padding,
+                                   K1,
+                                   I1));
+
+                    constexpr auto a_block_desc_b_k0_m_k1_tmp = transform_tensor_descriptor(
+                        a_block_desc_b_k0_m0_m1_k1,
+                        make_tuple(make_pass_through_transform(Number<1>{}),
+                                   make_pass_through_transform(Number<K0PerBlock>{}),
+                                   make_merge_transform_v4_no_carry(
+                                       make_tuple(Number<M0PerBlock>{}, Number<M1PerBlock>{})),
+                                   make_pass_through_transform(K1)),
+                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}),
+                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+                    return a_block_desc_b_k0_m_k1_tmp;
+                }
+                else
+                {
+                    return make_naive_tensor_descriptor(
+                        make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                        make_tuple(Number<K0PerBlock>{} * Number<MPerBlock + 1>{} * K1,
+                                   Number<MPerBlock + 1>{} * K1,
+                                   K1,
+                                   I1));
+                }
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+
+        return a_block_desc_b_k0_m_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                if constexpr(BBlockLdsExtraN1Wrw)
+                {
+                    constexpr auto b_block_desc_k0_n0_n1_k1 = make_naive_tensor_descriptor(
+                        make_tuple(
+                            Number<K0PerBlock>{}, Number<N0PerBlock>{}, Number<N1PerBlock>{}, K1),
+                        make_tuple(Number<N0PerBlock>{} * (Number<N1PerBlock>{} * K1 + N1Padding),
+                                   Number<N1PerBlock>{} * K1 + N1Padding,
+                                   K1,
+                                   I1));
+
+                    constexpr auto b_block_desc_k0_n_k1_tmp = transform_tensor_descriptor(
+                        b_block_desc_k0_n0_n1_k1,
+                        make_tuple(make_pass_through_transform(Number<K0PerBlock>{}),
+                                   make_merge_transform_v3_division_mod(
+                                       make_tuple(Number<N0PerBlock>{}, Number<N1PerBlock>{})),
+                                   make_pass_through_transform(K1)),
+                        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
+                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+                    return b_block_desc_k0_n_k1_tmp;
+                }
+                else
+                {
+                    return make_naive_tensor_descriptor(
+                        make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                        make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+                }
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return b_block_desc_k0_n_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_Batch_K0PerBlock_NPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_b_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                if constexpr(BBlockLdsExtraN1Wrw)
+                {
+                    constexpr auto b_block_desc_b_k0_n0_n1_k1 = make_naive_tensor_descriptor(
+                        make_tuple(Number<1>{},
+                                   Number<K0PerBlock>{},
+                                   Number<N0PerBlock>{},
+                                   Number<N1PerBlock>{},
+                                   K1),
+                        make_tuple(Number<K0PerBlock>{} * Number<N0PerBlock>{} *
+                                       (Number<N1PerBlock>{} * K1 + N1Padding),
+                                   Number<N0PerBlock>{} * (Number<N1PerBlock>{} * K1 + N1Padding),
+                                   Number<N1PerBlock>{} * K1 + N1Padding,
+                                   K1,
+                                   I1));
+
+                    constexpr auto b_block_desc_b_k0_n_k1_tmp = transform_tensor_descriptor(
+                        b_block_desc_b_k0_n0_n1_k1,
+                        make_tuple(make_pass_through_transform(Number<1>{}),
+                                   make_pass_through_transform(Number<K0PerBlock>{}),
+                                   make_merge_transform_v4_no_carry(
+                                       make_tuple(Number<N0PerBlock>{}, Number<N1PerBlock>{})),
+                                   make_pass_through_transform(K1)),
+                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}),
+                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+                    return b_block_desc_b_k0_n_k1_tmp;
+                }
+                else
+                {
+                    return make_naive_tensor_descriptor(
+                        make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                        make_tuple(Number<K0PerBlock>{} * Number<NPerBlock + 1>{} * K1,
+                                   Number<NPerBlock + 1>{} * K1,
+                                   K1,
+                                   I1));
+                }
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+
+        return b_block_desc_b_k0_n_k1;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_b_k0_m_k1_block_desc = GetABlockDescriptor_Batch_K0PerBlock_MPerBlock_K1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_b_k0_n_k1_block_desc = GetBBlockDescriptor_Batch_K0PerBlock_NPerBlock_K1();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size = math::integer_least_multiple(
+            a_b_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size = math::integer_least_multiple(
+            b_b_k0_n_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto c_block_size =
+            GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock().GetElementSpaceSize();
+
+        return math::max((a_block_space_size + b_block_space_size) * sizeof(FloatAB),
+                         c_block_size * sizeof(FloatC));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
+                  const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
+                  const CMNGridDesc& c_m_n_grid_desc,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXDL * MRepeat) == 0) &&
+                          (NPerBlock % (NRepeat * NPerXDL)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M      = a_b_k0_m_k1_grid_desc.GetLength(I2);
+        const auto N      = b_b_k0_n_k1_grid_desc.GetLength(I2);
+        const auto K0     = a_b_k0_m_k1_grid_desc.GetLength(I1);
+        const auto KBatch = a_b_k0_m_k1_grid_desc.GetLength(I0);
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K0 / K0PerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!(M == c_m_n_grid_desc.GetLength(I0) && N == c_m_n_grid_desc.GetLength(I1) &&
+             K0 == b_b_k0_n_k1_grid_desc.GetLength(I1) &&
+             K1 == a_b_k0_m_k1_grid_desc.GetLength(I3) &&
+             K1 == b_b_k0_n_k1_grid_desc.GetLength(I3) &&
+             KBatch == b_b_k0_n_k1_grid_desc.GetLength(I0)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        if(!block_2_ctile_map.CheckValidity(c_m_n_grid_desc))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
+    {
+        // const bool has_main_k0_block_loop = K0 > K0PerBlock;
+        const index_t num_loop = K0 / K0PerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+
+        // return has_main_k0_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(const CMNGridDesc& c_m_n_grid_desc)
+    {
+        const auto M = c_m_n_grid_desc.GetLength(I0);
+        const auto N = c_m_n_grid_desc.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        return transform_tensor_descriptor(
+            c_m_n_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
+        const CMNGridDesc& c_m_n_grid_desc, index_t M01, index_t N01, index_t KBatch)
+    {
+        return BlockToCTileMap_KSplit_M00_N00_M01_N01<MPerBlock, NPerBlock, CMNGridDesc>(
+            c_m_n_grid_desc, M01, N01, KBatch);
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+        constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1,
+                       Number<CShuffleMRepeatPerShuffle * MWave * MPerXDL>{},
+                       I1,
+                       Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
+    }
+
+    using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        decltype(MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CMNGridDesc{}));
+    using CBlockClusterAdaptor = decltype(MakeCBlockClusterAdaptor(CMNGridDesc{}, 1, 1, 1));
+
+    template <bool HasMainKBlockLoop>
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               void* __restrict__ p_shared,
+                               const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
+                               const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
+                               const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CElementwiseOperation& c_element_op,
+                               const CBlockClusterAdaptor& c_block_cluster_adaptor)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_b_k0_n_k1_grid_desc.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            c_block_cluster_adaptor.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t k_batch_id = block_work_idx[I0];
+
+        if(!c_block_cluster_adaptor.ValidCTileIndex(
+               make_tuple(block_work_idx[I1], block_work_idx[I2]),
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_k0_m_k1_block_desc = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        constexpr auto a_b_k0_m_k1_block_desc = GetABlockDescriptor_Batch_K0PerBlock_MPerBlock_K1();
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_k0_n_k1_block_desc = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        constexpr auto b_b_k0_n_k1_block_desc = GetBBlockDescriptor_Batch_K0PerBlock_NPerBlock_K1();
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, K0PerBlock, MPerBlock, K1>,
+                                                ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatABAdjusted,
+                                                decltype(a_b_k0_m_k1_grid_desc),
+                                                decltype(a_b_k0_m_k1_block_desc),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                a_b_k0_m_k1_grid_desc,
+                make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_b_k0_m_k1_block_desc,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, K0PerBlock, NPerBlock, K1>,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatABAdjusted,
+                                                decltype(b_b_k0_n_k1_grid_desc),
+                                                decltype(b_b_k0_n_k1_block_desc),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                b_b_k0_n_k1_grid_desc,
+                make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_b_k0_n_k1_block_desc,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+
+        constexpr index_t KPack =
+            math::max(K1, MfmaSelector<FloatABAdjusted, MPerXDL, NPerXDL>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatABAdjusted,
+                                                                FloatAcc,
+                                                                decltype(a_k0_m_k1_block_desc),
+                                                                decltype(b_k0_n_k1_block_desc),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MRepeat,
+                                                                NRepeat,
+                                                                KPack>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatABAdjusted*>(p_shared), a_k0_m_k1_block_desc.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatABAdjusted*>(p_shared) + a_block_space_size,
+            b_k0_n_k1_block_desc.GetElementSpaceSize());
+
+        // gridwise GEMM pipeline
+        const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
+
+        GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_b_k0_m_k1_grid_desc,
+                                                          a_b_k0_m_k1_block_desc,
+                                                          a_blockwise_copy,
+                                                          a_grid_buf,
+                                                          a_block_buf,
+                                                          a_block_slice_copy_step,
+                                                          b_b_k0_n_k1_grid_desc,
+                                                          b_b_k0_n_k1_block_desc,
+                                                          b_blockwise_copy,
+                                                          b_grid_buf,
+                                                          b_block_buf,
+                                                          b_block_slice_copy_step,
+                                                          blockwise_gemm,
+                                                          c_thread_buf,
+                                                          K0BlockMainLoop);
+
+        // output: register to global memory
+        {
+            constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+            constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I0);
+            constexpr auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I1);
+            constexpr auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I2);
+            constexpr auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I3);
+            constexpr auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I4);
+            constexpr auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I5);
+            constexpr auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I6);
+            constexpr auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I7);
+
+            constexpr auto c_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatC*>(p_shared),
+                c_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            static_assert(M1 == MWave, "");
+            static_assert(N1 == NWave, "");
+            static_assert(M2 * M3 * M4 == MPerXDL, "");
+            static_assert(N2 == NPerXDL, "");
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0), // freeze mblock
+                    make_unmerge_transform(make_tuple(CShuffleMRepeatPerShuffle,
+                                                      M1,
+                                                      M2,
+                                                      M3,
+                                                      M4)), // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                    make_freeze_transform(I0),              // freeze nblock
+                    make_unmerge_transform(make_tuple(CShuffleNRepeatPerShuffle,
+                                                      N1,
+                                                      N2))), // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMRepeatPerShuffle,
+                                                            CShuffleNRepeatPerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // LDS to global
+            auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // index_t BlockSize,
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMRepeatPerShuffle * MWave * MPerXDL,
+                         1,
+                         CShuffleNRepeatPerShuffle * NWave * NPerXDL>, // BlockSliceLengths,
+                CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatC,               // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                       // typename DimAccessOrder,
+                3,                                          // index_t VectorDim,
+                CBlockTransferScalarPerVector_NWaveNPerXDL, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun
+                {c_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0),
+                 c_element_op};
+
+            constexpr auto mxdlperwave_forward_step =
+                make_multi_index(0, CShuffleMRepeatPerShuffle * MWave * MPerXDL, 0, 0);
+            constexpr auto nxdlperwave_forward_step =
+                make_multi_index(0, 0, 0, CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+            constexpr auto nxdlperwave_backward_step =
+                make_multi_index(0, 0, 0, -CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+
+            static_for<0, MRepeat, CShuffleMRepeatPerShuffle>{}([&](auto mxdlperwave_iter) {
+                constexpr auto mxdlperwave = mxdlperwave_iter;
+
+                static_for<0, NRepeat, CShuffleNRepeatPerShuffle>{}([&](auto nxdlperwave_iter) {
+                    constexpr bool nxdlperwave_forward_sweep =
+                        (mxdlperwave % (2 * CShuffleMRepeatPerShuffle) == 0);
+
+                    constexpr index_t nxdlperwave_value =
+                        nxdlperwave_forward_sweep
+                            ? nxdlperwave_iter
+                            : (NRepeat - nxdlperwave_iter - CShuffleNRepeatPerShuffle);
+
+                    constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
+
+                    // make sure it's safe to do ds_write
+                    block_sync_lds();
+
+                    // VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(
+                        c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
+                        make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
+                        c_thread_buf,
+                        c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        c_block_buf);
+
+                    // make sure it's safe to do ds_read
+                    block_sync_lds();
+
+                    // LDS to global
+                    c_block_copy_lds_to_global.Run(c_block_desc_mblock_mperblock_nblock_nperblock,
+                                                   c_block_buf,
+                                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                   c_grid_buf);
+
+                    // move on nxdlperwave dimension
+                    if constexpr(nxdlperwave_forward_sweep &&
+                                 (nxdlperwave < NRepeat - CShuffleNRepeatPerShuffle))
+                    {
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mperblock_nblock_nperblock,
+                            nxdlperwave_forward_step);
+                    }
+                    else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
+                    {
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mperblock_nblock_nperblock,
+                            nxdlperwave_backward_step);
+                    }
+                });
+
+                // move on mxdlperwave dimension
+                if constexpr(mxdlperwave < MRepeat - CShuffleMRepeatPerShuffle)
+                {
+                    c_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, mxdlperwave_forward_step);
+                }
+            });
+        }
+    }
+}; // namespace ck
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
new file mode 100644
index 0000000000..b12bcee0f4
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
@@ -0,0 +1,679 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3,
+          typename CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainK0BlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_skip_b_lds_v1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3 b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+            const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainK0BlockLoop>(p_a_grid,
+                                                   p_b_grid,
+                                                   p_c_grid,
+                                                   p_shared,
+                                                   a_grid_desc_k0_m_k1,
+                                                   b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                                   c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                                   a_element_op,
+                                                   b_element_op,
+                                                   c_element_op,
+                                                   block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3;
+    ignore = c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M_N,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          bool ABlockLdsExtraM,
+          index_t BBlockTransferSrcScalarPerVector,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockBufferSize,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_skip_b_lds_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    static constexpr index_t WaveSize = 64;
+    static constexpr index_t MWaves   = MPerBlock / (MXdlPerWave * MPerXDL);
+    static constexpr index_t NWaves   = NPerBlock / (NXdlPerWave * NPerXDL);
+
+    static constexpr auto xdlops_gemm    = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, K1>{};
+    static constexpr index_t K0PerThread = K0PerBlock / xdlops_gemm.K0PerXdlops;
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock * BBlockBufferSize>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock * BBlockBufferSize>{}, Number<MPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+
+        return a_block_desc_k0_m_k1;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        constexpr auto max_lds_align = K1;
+
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        return (a_block_space_size_aligned) * sizeof(FloatAB);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  index_t M01,
+                  index_t N01)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXDL * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXDL)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+             K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+             K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        // 2-stage prefetch currently only support even number of K0 loop
+        // TODO: add support for odd number of K0 loop
+        if(!((K0 / K0PerBlock) % BBlockBufferSize == 0))
+        {
+            return false;
+        }
+
+        // check M01, N01
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        if(!(M0 % M01 == 0 && N0 % N01 == 0))
+            return false;
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr index_t
+    CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const index_t grid_size = (M / MPerBlock) * (N / NPerBlock);
+
+        return grid_size;
+    }
+
+    // TODO move this function into GEMM-pipeline class
+    __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
+    {
+        const bool has_main_k0_block_loop = (K0 / (BBlockBufferSize * K0PerBlock)) > 1;
+
+        return has_main_k0_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeBGridDescriptor_K0_K1_K2_N0_N1_N2_N3_K3(const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1)
+    {
+        const auto K0 = b_grid_desc_k0_n_k1.GetLength(I0);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+
+        const auto b_griddesc_k0_nblockid_nrepeat_waves_nperxdlops_k1 = transform_tensor_descriptor(
+            b_grid_desc_k0_n_k1,
+            make_tuple(make_unmerge_transform(
+                           make_tuple(K0 / K0PerBlock, xdlops_gemm.K0PerXdlops, K0PerThread)),
+                       make_unmerge_transform(make_tuple(
+                           N / (NXdlPerWave * NWaves * NPerXDL), NXdlPerWave, NWaves, NPerXDL)),
+                       make_pass_through_transform(K1)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5, 6>{}, Sequence<7>{}));
+        return b_griddesc_k0_nblockid_nrepeat_waves_nperxdlops_k1;
+    }
+
+    __device__ static auto GetWaveIdx()
+    {
+        const index_t thread_id = get_thread_local_1d_id();
+
+        constexpr auto threadid_to_wave_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(MWaves, NWaves, WaveSize))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        return threadid_to_wave_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __device__ static auto GetWaveKNIdx(const index_t thread_id)
+    {
+        constexpr auto wave_threadid_to_nk_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(make_tuple(xdlops_gemm.K0PerXdlops, NPerXDL))),
+            make_tuple(Sequence<0, 1>{}),
+            make_tuple(Sequence<0>{}));
+
+        return wave_threadid_to_nk_idx_adaptor.CalculateBottomIndex(make_multi_index(thread_id));
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix threadwise copy
+        constexpr auto b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           I1,
+                                                           Number<K0PerThread>{}, // K0PerThread
+                                                           I1,                    // NBlockId
+                                                           Number<NXdlPerWave>{}, // repeat
+                                                           I1,                    // waves
+                                                           I1,                    // NPerXdlops
+                                                           Number<K1>{}));
+
+        using BlockwiseGemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1r1<
+            BlockSize,
+            FloatAB,
+            FloatAcc,
+            decltype(a_block_desc_k0_m_k1),
+            decltype(b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3),
+            MPerBlock,
+            NPerBlock,
+            K0PerBlock,
+            MPerXDL,
+            NPerXDL,
+            MXdlPerWave,
+            NXdlPerWave,
+            K1>;
+
+        return BlockwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n);
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n, index_t M01, index_t N01)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        constexpr auto M1 = Number<MPerBlock>{};
+        constexpr auto N1 = Number<NPerBlock>{};
+
+        const auto M0 = M / M1;
+        const auto N0 = N / N1;
+
+        const auto M00 = M0 / M01;
+        const auto N00 = N0 / N01;
+
+        const auto m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_unmerge_transform(make_tuple(M00, M01)),
+                           make_unmerge_transform(make_tuple(N00, N01))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}));
+
+        const auto cblockid_to_m00_m01_n00_n01_block_cluster_adaptor =
+            make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(M00, N00, M01, N01))),
+                make_tuple(Sequence<0, 1, 2, 3>{}),
+                make_tuple(Sequence<0>{}));
+
+        const auto cblockid_to_m0_n0_block_cluster_adaptor =
+            chain_tensor_adaptors(m00_m01_n00_n01_to_m0_n0_block_cluster_adaptor,
+                                  cblockid_to_m00_m01_n00_n01_block_cluster_adaptor);
+
+        return cblockid_to_m0_n0_block_cluster_adaptor;
+    }
+
+    using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
+        decltype(MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
+    using DefaultBlock2CTileMap = decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
+    using BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3 =
+        decltype(MakeBGridDescriptor_K0_K1_K2_N0_N1_N2_N3_K3(BGridDesc_K0_N_K1{}));
+
+    template <bool HasMainK0BlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        void* __restrict__ p_shared,
+        const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+        const BGridDesc_K0_K1_K2_N0_N1_N2_N3_K3 b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+        const CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2& c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetElementSpaceSize());
+
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy = ThreadGroupTensorSliceTransfer_v4r1<
+            ThisThreadBlock,
+            AElementwiseOperation,
+            ck::tensor_operation::element_wise::PassThrough,
+            InMemoryDataOperationEnum::Set,
+            Sequence<K0PerBlock * BBlockBufferSize, MPerBlock, K1>,
+            ABlockTransferThreadClusterLengths_K0_M_K1,
+            ABlockTransferThreadClusterArrangeOrder,
+            FloatAB,
+            FloatAB,
+            decltype(a_grid_desc_k0_m_k1),
+            decltype(a_block_desc_k0_m_k1),
+            ABlockTransferSrcAccessOrder,
+            Sequence<1, 0, 2>,
+            ABlockTransferSrcVectorDim,
+            2,
+            ABlockTransferSrcScalarPerVector,
+            ABlockTransferDstScalarPerVector_K1,
+            1,
+            1,
+            AThreadTransferSrcResetCoordinateAfterRun,
+            true,
+            1>(a_grid_desc_k0_m_k1,
+               make_multi_index(0, m_block_data_idx_on_grid, 0),
+               a_element_op,
+               a_block_desc_k0_m_k1,
+               make_multi_index(0, 0, 0),
+               ck::tensor_operation::element_wise::PassThrough{});
+
+        ignore = b_element_op;
+        // B matrix threadwise copy
+        constexpr auto b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3 =
+            make_naive_tensor_descriptor_packed(make_tuple(I1,
+                                                           I1,
+                                                           Number<K0PerThread>{}, // K0PerThread
+                                                           I1,                    // NBlockId
+                                                           Number<NXdlPerWave>{}, // repeat
+                                                           I1,                    // waves
+                                                           I1,                    // NPerXdlops
+                                                           Number<K1>{}));
+
+        auto b_thread_buf = generate_tuple(
+            [&](auto i) {
+                ignore = i;
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    FloatAB,
+                                    b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3.GetElementSpaceSize(),
+                                    true>{};
+            },
+            Number<BBlockBufferSize>{});
+
+        const auto wave_id     = GetWaveIdx();
+        const auto wave_k_n_id = GetWaveKNIdx(wave_id[I2]);
+
+#if 0
+        const index_t block_id  = get_block_1d_id();
+        const index_t thread_id = get_thread_local_1d_id();
+        printf("block id: %d  m blockid: %d n block id: %d ,thread id: %d, wave id :{%d %d %d} "
+               "kn id: {%d %d}\n",
+               block_id,
+               block_work_idx[I0],
+               block_work_idx[I1],
+               thread_id,
+               wave_id[I0],
+               wave_id[I1],
+               wave_id[I2],
+               wave_k_n_id[I0],
+               wave_k_n_id[I1]);
+        printf("mfma thread k per xdlops: %d K0PerThread: %d HasMainK0BlockLoop: %d K0: %d  \t", 
+                xdlops_gemm.K0PerXdlops, K0PerThread, HasMainK0BlockLoop, b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3.GetLength(I0));
+#endif
+
+        auto b_threadwise_copy =
+            ThreadwiseTensorSliceTransfer_v2<FloatAB,
+                                             FloatAB,
+                                             decltype(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3),
+                                             decltype(b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3),
+                                             Sequence<I1,
+                                                      I1,
+                                                      Number<K0PerThread>{},
+                                                      I1,
+                                                      Number<NXdlPerWave>{},
+                                                      I1,
+                                                      I1,
+                                                      Number<K1>{}>,
+                                             Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                             7,
+                                             BBlockTransferSrcScalarPerVector,
+                                             BThreadTransferSrcResetCoordinateAfterRun,
+                                             true>(
+                b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                make_multi_index(
+                    0, wave_k_n_id[I0], 0, block_work_idx[I1], 0, wave_id[I1], wave_k_n_id[I1], 0));
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1r1<
+            BlockSize,
+            FloatAB,
+            FloatAcc,
+            decltype(a_block_desc_k0_m_k1),
+            decltype(b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3),
+            MPerBlock,
+            NPerBlock,
+            K0PerBlock,
+            MPerXDL,
+            NPerXDL,
+            MXdlPerWave,
+            NXdlPerWave,
+            K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
+
+        // gridwise GEMM pipeline
+        constexpr auto a_block_slice_copy_step =
+            make_multi_index(K0PerBlock * BBlockBufferSize, 0, 0);
+        constexpr auto b_thread_slice_copy_step = make_multi_index(1, 0, 0, 0, 0, 0, 0, 0);
+        // preload data to regiester and LDS
+        {
+            // Read
+            a_blockwise_copy.RunRead(a_grid_desc_k0_m_k1, a_grid_buf);
+            a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m_k1, a_block_slice_copy_step);
+
+            static_for<0, BBlockBufferSize, 1>{}([&](auto ii) {
+                b_threadwise_copy.Run(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                      b_grid_buf,
+                                      b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                      make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                                      b_thread_buf(Number<ii>{}));
+                b_threadwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                                     b_thread_slice_copy_step);
+            });
+
+            // Initialize C
+            c_thread_buf.Clear();
+            // a data write to lds
+            a_blockwise_copy.RunWrite(a_block_desc_k0_m_k1, a_block_buf);
+            // main body
+            if constexpr(HasMainK0BlockLoop)
+            {
+                index_t K0BlockMainLoop =
+                    __builtin_amdgcn_readfirstlane(K0 / (BBlockBufferSize * K0PerBlock));
+                index_t i = 0;
+                do
+                {
+                    a_blockwise_copy.RunRead(a_grid_desc_k0_m_k1, a_grid_buf);
+                    blockwise_gemm.ResetABlockStartWindow();
+                    block_sync_lds();
+
+                    static_for<0, BBlockBufferSize, 1>{}([&](auto ii) {
+                        blockwise_gemm.Run(a_block_buf, b_thread_buf(Number<ii>{}), c_thread_buf);
+                        blockwise_gemm.MoveABlockSliceWindow();
+                        s_nop();
+
+                        b_threadwise_copy.Run(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                              b_grid_buf,
+                                              b_thread_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                                              b_thread_buf(Number<ii>{}));
+                        b_threadwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_k1_k2_n0_n1_n2_n3_k3,
+                                                             b_thread_slice_copy_step);
+                    });
+
+                    block_sync_lds();
+                    a_blockwise_copy.RunWrite(a_block_desc_k0_m_k1, a_block_buf);
+                    // move a and b window
+                    a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m_k1,
+                                                        a_block_slice_copy_step);
+
+                    i += 1;
+                } while(i < (K0BlockMainLoop - 1));
+            }
+
+            // tail
+            {
+                block_sync_lds();
+
+                blockwise_gemm.ResetABlockStartWindow();
+
+                static_for<0, BBlockBufferSize, 1>{}([&](auto ii) {
+                    blockwise_gemm.Run(a_block_buf, b_thread_buf(Number<ii>{}), c_thread_buf);
+                    blockwise_gemm.MoveABlockSliceWindow();
+                });
+            }
+        }
+
+        // output: register to global memory
+        {
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I7);
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_grid =
+                m_block_data_idx_on_grid + c_thread_mtx_on_block[I0];
+
+            const index_t n_thread_data_on_grid =
+                n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_grid_idx =
+                m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_grid));
+
+            const auto n_thread_data_on_grid_to_n0_n1_n2_adaptor = make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                make_tuple(Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_grid_idx =
+                n_thread_data_on_grid_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_grid));
+
+            auto c_thread_copy =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   CElementwiseOperation,
+                                                   Sequence<M0, N0, I1, I1, M2, I1, M4, I1>,
+                                                   CThreadTransferSrcDstAccessOrder,
+                                                   CThreadTransferSrcDstVectorDim,
+                                                   CThreadTransferDstScalarPerVector,
+                                                   CGlobalMemoryDataOperation,
+                                                   1,
+                                                   true>{
+                    c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(m_thread_data_on_grid_idx[I0],
+                                     n_thread_data_on_grid_idx[I0],
+                                     m_thread_data_on_grid_idx[I1],
+                                     n_thread_data_on_grid_idx[I1],
+                                     m_thread_data_on_grid_idx[I2],
+                                     m_thread_data_on_grid_idx[I3],
+                                     m_thread_data_on_grid_idx[I4],
+                                     n_thread_data_on_grid_idx[I2]),
+                    c_element_op};
+
+            c_thread_copy.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                              c_thread_buf,
+                              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                              c_grid_buf);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp
new file mode 100644
index 0000000000..70abcac0b1
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp
@@ -0,0 +1,1184 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1r2.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v3.hpp"
+#include "ck/utility/workgroup_barrier.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_streamk(const typename GridwiseGemm::FloatAB* p_a_grid,
+                                   const typename GridwiseGemm::FloatAB* p_b_grid,
+                                   typename GridwiseGemm::FloatC* p_c_grid,
+                                   void* p_workspace,
+                                   index_t M,
+                                   index_t N,
+                                   index_t K,
+                                   index_t StrideA,
+                                   index_t StrideB,
+                                   index_t StrideC,
+                                   typename GridwiseGemm::Block2CTileMap block_mapping)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
+
+    __shared__ uint8_t p_shared[shared_size];
+
+    GridwiseGemm::Run(p_a_grid,
+                      p_b_grid,
+                      p_c_grid,
+                      p_workspace,
+                      M,
+                      N,
+                      K,
+                      StrideA,
+                      StrideB,
+                      StrideC,
+                      block_mapping,
+                      static_cast<void*>(p_shared));
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_workspace;
+    ignore = M;
+    ignore = N;
+    ignore = K;
+    ignore = StrideA;
+    ignore = StrideB;
+    ignore = StrideC;
+    ignore = block_mapping;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <index_t BlockSize,
+          typename Block2CTileMap_,
+          typename FloatAB_,
+          typename FloatAcc_,
+          typename FloatC_,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MRepeat,
+          index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXDL,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>
+struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_streamk
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1        = Number<K1Value>{};
+    static constexpr auto M01       = 1;
+    static constexpr auto N01       = 1;
+    static constexpr auto KPerBlock = K0PerBlock * K1;
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+    using FloatAcc        = FloatAcc_;
+    using FloatCShuffle   = FloatAcc;
+
+    using Block2CTileMap = Block2CTileMap_;
+    using FloatAB        = FloatAB_;
+    using FloatC         = FloatC_;
+
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        const FloatAB* p_a_grid;
+        const FloatAB* p_b_grid;
+        FloatC* p_c_grid;
+        index_t M;
+        index_t N;
+        index_t K;
+        index_t StrideA;
+        index_t StrideB;
+        index_t StrideC;
+        Block2CTileMap block_mapping;
+
+        Argument(const FloatAB* p_a_grid_,
+                 const FloatAB* p_b_grid_,
+                 FloatC* p_c_grid_,
+                 index_t M_,
+                 index_t N_,
+                 index_t K_,
+                 index_t StrideA_,
+                 index_t StrideB_,
+                 index_t StrideC_,
+                 uint32_t num_cu,
+                 uint32_t occupancy,
+                 uint32_t num_sk_blocks_)
+            : p_a_grid(p_a_grid_),
+              p_b_grid(p_b_grid_),
+              p_c_grid(p_c_grid_),
+              M(M_),
+              N(N_),
+              K(K_),
+              StrideA(StrideA_),
+              StrideB(StrideB_),
+              StrideC(StrideC_),
+              block_mapping(M, N, K, num_cu, occupancy, num_sk_blocks_)
+        {
+        }
+
+        void Print() const
+        {
+            std::cout << "arg {"
+                      << "M:" << M << ", "
+                      << "N:" << N << ", "
+                      << "K:" << K << ", "
+                      << "SA:" << StrideA << ", "
+                      << "SB:" << StrideB << ", "
+                      << "SC:" << StrideC << std::endl;
+        }
+    };
+
+    __host__ __device__ static auto CalculateGridSize(const Argument& karg)
+    {
+        return std::make_tuple(math::integer_divide_ceil(karg.N, NPerBlock),
+                               math::integer_divide_ceil(karg.M, MPerBlock),
+                               karg.k_batch);
+    }
+
+    __host__ __device__ static auto CalculateK0(index_t KPad) { return KPad / K1; }
+
+    __host__ __device__ static auto
+    MakeAGridDescriptor_K0_M_K1(index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA)
+    {
+        const index_t K0 = CalculateK0(KPad);
+
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return transform_tensor_descriptor(a_grid_desc_m_kpad,
+                                           make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                      make_right_pad_transform(M, MPad - M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    __host__ __device__ static auto
+    MakeBGridDescriptor_K0_N_K1(index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB)
+    {
+        const index_t K0 = CalculateK0(KPad);
+
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
+            b_grid_desc_k_n,
+            make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return transform_tensor_descriptor(b_grid_desc_kpad_n,
+                                           make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
+                                                      make_right_pad_transform(N, NPad - N)),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    __host__ __device__ static auto
+    MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        return transform_tensor_descriptor(c_grid_desc_m_n,
+                                           make_tuple(make_right_pad_transform(M, MPad - M),
+                                                      make_right_pad_transform(N, NPad - N)),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                           make_tuple(Sequence<0>{}, Sequence<1>{}));
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * K1, K1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * K1, K1, I1));
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned =
+            math::integer_least_multiple(b_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto c_block_size =
+            GetCBlockDescriptor_MBlock_MPerShuffle_NBlock_NPerShuffle().GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    __host__ __device__ static constexpr bool CheckValidity(const Argument& karg)
+    {
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            if(karg.K % ABlockTransferSrcScalarPerVector != 0)
+                return false;
+        }
+        else
+        {
+            if(karg.M % ABlockTransferSrcScalarPerVector != 0)
+                return false;
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        {
+            if(karg.N % BBlockTransferSrcScalarPerVector != 0)
+                return false;
+        }
+        else
+        {
+            if(karg.K % BBlockTransferSrcScalarPerVector != 0)
+                return false;
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+        {
+            if(karg.N % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
+                return false;
+        }
+        else
+        {
+            if(karg.M % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
+                return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
+    {
+        const bool has_main_k0_block_loop = K0 > K0PerBlock;
+
+        return has_main_k0_block_loop;
+    }
+
+    template <typename CGridDesc>
+    __host__ __device__ static constexpr auto
+    MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc& c_m_n_grid_desc)
+    {
+        const auto M = c_m_n_grid_desc.GetLength(I0);
+        const auto N = c_m_n_grid_desc.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        return transform_tensor_descriptor(
+            c_m_n_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    template <typename CGridDesc>
+    __host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
+        const CGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch)
+    {
+        return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc>(
+            c_m_n_grid_desc, 8, KBatch);
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MBlock_MPerShuffle_NBlock_NPerShuffle()
+    {
+        constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+        constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1,
+                       Number<CShuffleMRepeatPerShuffle * MWave * MPerXDL>{},
+                       I1,
+                       Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MShuffleRepeat_MPerShuffle_NShuffleRepeat_NPerShuffle()
+    {
+        constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+        constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MRepeat / CShuffleMRepeatPerShuffle>{},
+                       Number<CShuffleMRepeatPerShuffle * MWave * MPerXDL>{},
+                       Number<NRepeat / CShuffleNRepeatPerShuffle>{},
+                       Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
+    }
+
+    __host__ __device__ static constexpr auto GetClusterLengthReduction()
+    {
+        // TODO: assume C is row major
+        // TODO: we always first loop over N, then M
+        constexpr auto NPerBlockPow2 = math::next_power_of_two<NPerBlock>();
+        constexpr auto NPerBlockReduction =
+            NPerBlockPow2 / CBlockTransferScalarPerVector_NWaveNPerXDL;
+        constexpr auto MPerBlockReduction =
+            (BlockSize + NPerBlockReduction - 1) / NPerBlockReduction;
+        return Sequence<MPerBlockReduction, NPerBlockReduction>{};
+    }
+
+    __host__ __device__ static constexpr auto GetPartialAccBlockDescriptor()
+    {
+        const auto c_partial_acc_block_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MPerBlock, NPerBlock),
+                                                    make_tuple(NPerBlock, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MPerBlock, NPerBlock),
+                                                    make_tuple(I1, MPerBlock));
+            }
+        }();
+        return c_partial_acc_block_m_n;
+    }
+
+    using CGridDesc_M_N = remove_cvref_t<decltype(MakeCGridDescriptor_M_N(1, 1, 1, 1, 1))>;
+
+    __device__ static void Run(const FloatAB* p_a_grid,
+                               const FloatAB* p_b_grid,
+                               FloatC* p_c_grid,
+                               void* p_workspace,
+                               index_t M,
+                               index_t N,
+                               index_t K,
+                               index_t StrideA,
+                               index_t StrideB,
+                               index_t StrideC,
+                               Block2CTileMap block_mapping,
+                               void* __restrict__ p_shared_block)
+    {
+        uint32_t m        = M;
+        uint32_t n        = N;
+        uint32_t k        = K;
+        uint32_t pad_m    = (m + MPerBlock - 1) / MPerBlock * MPerBlock;
+        uint32_t pad_n    = (n + NPerBlock - 1) / NPerBlock * NPerBlock;
+        uint32_t pad_k    = (k + KPerBlock - 1) / KPerBlock * KPerBlock;
+        uint32_t stride_a = StrideA;
+        uint32_t stride_b = StrideB;
+        uint32_t stride_c = StrideC;
+
+        const auto a_k0_m_k1_grid_desc = MakeAGridDescriptor_K0_M_K1(m, pad_m, k, pad_k, stride_a);
+        const auto b_k0_n_k1_grid_desc = MakeBGridDescriptor_K0_N_K1(k, pad_k, n, pad_n, stride_b);
+        const auto c_grid_desc_m_n     = MakeCGridDescriptor_M_N(m, pad_m, n, pad_n, stride_c);
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n);
+        const AElementwiseOperation a_element_op = AElementwiseOperation{};
+        const BElementwiseOperation b_element_op = BElementwiseOperation{};
+        const CElementwiseOperation c_element_op = CElementwiseOperation{};
+
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_k0_m_k1_grid_desc.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_k0_n_k1_grid_desc.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_block_desc_k0_m_k1),
+                                                                decltype(b_block_desc_k0_n_k1),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MRepeat,
+                                                                NRepeat,
+                                                                K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block = static_cast<FloatAB*>(p_shared_block);
+        FloatAB* p_b_block = static_cast<FloatAB*>(p_shared_block) + a_block_space_size;
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block, a_block_desc_k0_m_k1.GetElementSpaceSize());
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block, b_block_desc_k0_n_k1.GetElementSpaceSize());
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline = GridwiseGemmPipeline_v3();
+
+        uint32_t block_idx = block_mapping.get_block_idx();
+        bool is_sk_block   = block_idx < block_mapping.sk_num_blocks;
+        bool is_dp_block   = block_idx >= block_mapping.dp_start_block_idx &&
+                           block_idx < block_mapping.reduction_start_block_idx;
+        bool is_reduction_block = block_idx >= block_mapping.reduction_start_block_idx;
+        bool is_padding_block   = block_idx >= block_mapping.sk_num_blocks &&
+                                block_idx < block_mapping.dp_start_block_idx;
+        uint32_t iter_start, iter_end;
+        block_mapping.get_block_itr(block_idx, iter_start, iter_end);
+        uint32_t total_iter_length = iter_end - iter_start;
+
+        if(is_padding_block)
+            return;
+
+        uint32_t* p_semaphore =
+            reinterpret_cast<uint32_t*>(reinterpret_cast<char*>(p_workspace) +
+                                        block_mapping.get_workspace_size_for_acc(sizeof(FloatAcc)));
+
+        if constexpr(Block2CTileMap::ReductionStrategy == StreamKReductionStrategy::Reduction)
+        {
+            if(is_reduction_block)
+            {
+                // descriptors
+                constexpr auto cluster_length_reduce = GetClusterLengthReduction();
+                constexpr auto reduce_desc = make_cluster_descriptor(cluster_length_reduce);
+                const auto reduce_thread_cluster_idx =
+                    reduce_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
+                const auto thread_m_cluster_id = reduce_thread_cluster_idx[I0];
+                const auto thread_n_cluster_id = reduce_thread_cluster_idx[I1];
+
+                constexpr auto MReduceIters =
+                    math::integer_divide_ceil(Number<MPerBlock>{}, cluster_length_reduce.At(I0));
+                constexpr auto NReduceIters = math::integer_divide_ceil(
+                    Number<NPerBlock>{},
+                    cluster_length_reduce.At(I1) *
+                        Number<CBlockTransferScalarPerVector_NWaveNPerXDL>{});
+
+                constexpr auto acc_thread_buf_load_desc = make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<CBlockTransferScalarPerVector_NWaveNPerXDL>{}));
+                constexpr auto acc_thread_buf_store_desc = make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, I1, I1, Number<CBlockTransferScalarPerVector_NWaveNPerXDL>{}));
+
+                constexpr auto c_partial_acc_block_m_n = GetPartialAccBlockDescriptor();
+
+                constexpr auto partial_acc_load_step_n = make_multi_index(
+                    0, cluster_length_reduce.At(I1) * CBlockTransferScalarPerVector_NWaveNPerXDL);
+                constexpr auto partial_acc_load_step_n_reverse =
+                    make_multi_index(0,
+                                     -1 * cluster_length_reduce.At(I1).value * (NReduceIters - 1) *
+                                         CBlockTransferScalarPerVector_NWaveNPerXDL);
+                constexpr auto partial_acc_load_step_m =
+                    make_multi_index(cluster_length_reduce.At(I0), 0);
+
+                constexpr auto partial_acc_store_step_n = make_multi_index(
+                    0,
+                    0,
+                    0,
+                    cluster_length_reduce.At(I1) * CBlockTransferScalarPerVector_NWaveNPerXDL);
+                constexpr auto partial_acc_store_step_n_reverse =
+                    make_multi_index(0,
+                                     0,
+                                     0,
+                                     -1 * cluster_length_reduce.At(I1).value * (NReduceIters - 1) *
+                                         CBlockTransferScalarPerVector_NWaveNPerXDL);
+                constexpr auto partial_acc_store_step_m =
+                    make_multi_index(0, cluster_length_reduce.At(I0), 0, 0);
+
+                StaticBuffer<AddressSpaceEnum::Vgpr,
+                             FloatAcc,
+                             CBlockTransferScalarPerVector_NWaveNPerXDL,
+                             true>
+                    parcial_acc_buf;
+                StaticBuffer<AddressSpaceEnum::Vgpr,
+                             FloatAcc,
+                             CBlockTransferScalarPerVector_NWaveNPerXDL,
+                             true>
+                    acc_buf;
+
+                // start to compute
+                auto reduction_idx = blockIdx.x - block_mapping.reduction_start_block_idx;
+                auto spatial_idx   = block_mapping.tile_to_spatial(reduction_idx, m, n);
+
+                workgroup_barrier wg_barrier(p_semaphore);
+
+                uint32_t tile_acc_offset_start =
+                    block_mapping.get_acc_buffer_offset_from_tile(reduction_idx);
+                uint32_t tile_acc_offset_end =
+                    block_mapping.get_acc_buffer_offset_from_tile(reduction_idx + 1);
+
+                auto acc_load = ThreadwiseTensorSliceTransfer_v2<
+                    FloatAcc,                                                // SrcData,
+                    FloatAcc,                                                // DstData,
+                    decltype(c_partial_acc_block_m_n),                       // SrcDesc,
+                    decltype(acc_thread_buf_load_desc),                      // DstDesc,
+                    Sequence<1, CBlockTransferScalarPerVector_NWaveNPerXDL>, // SliceLengths,
+                    Sequence<0, 1>,                                          // DimAccessOrder,
+                    1,                                                       // SrcVectorDim,
+                    CBlockTransferScalarPerVector_NWaveNPerXDL,              // SrcScalarPerVector,
+                    1,    // SrcScalarStrideInVector,
+                    false // SrcResetCoordinateAfterRun,
+                    >{c_partial_acc_block_m_n,
+                      make_multi_index(thread_m_cluster_id,
+                                       thread_n_cluster_id *
+                                           CBlockTransferScalarPerVector_NWaveNPerXDL)};
+
+                auto acc_store = ThreadwiseTensorSliceTransfer_v1r3<
+                    FloatAcc,                                                // SrcData,
+                    FloatC,                                                  // DstData,
+                    decltype(acc_thread_buf_store_desc),                     // SrcDesc,
+                    decltype(c_grid_desc_mblock_mperblock_nblock_nperblock), // DstDesc,
+                    CElementwiseOperation, // ElementwiseOperation,
+                    Sequence<1, 1, 1, CBlockTransferScalarPerVector_NWaveNPerXDL>, // SliceLengths,
+                    Sequence<0, 1, 2, 3>,                       // DimAccessOrder,
+                    3,                                          // DstVectorDim,
+                    CBlockTransferScalarPerVector_NWaveNPerXDL, // DstScalarPerVector,
+                    InMemoryDataOperationEnum::Set, // InMemoryDataOperationEnum DstInMemOp,
+                    1,                              // DstScalarStrideInVector,
+                    false                           // DstResetCoordinateAfterRun,
+                    >{c_grid_desc_mblock_mperblock_nblock_nperblock,
+                      make_multi_index(__builtin_amdgcn_readfirstlane(spatial_idx[I0]),
+                                       thread_m_cluster_id,
+                                       __builtin_amdgcn_readfirstlane(spatial_idx[I1]),
+                                       thread_n_cluster_id *
+                                           CBlockTransferScalarPerVector_NWaveNPerXDL),
+                      CElementwiseOperation{}};
+
+                // block synchronization
+                wg_barrier.wait_eq(reduction_idx, tile_acc_offset_end - tile_acc_offset_start);
+
+#if 0
+                if(threadIdx.x == 0) {
+                    printf("bid:%d, rid:%d, os:%d,%d, spatial:%d,%d\n", static_cast<int>(blockIdx.x),
+                        reduction_idx, __builtin_amdgcn_readfirstlane(tile_acc_offset_start), __builtin_amdgcn_readfirstlane(tile_acc_offset_end),
+                        __builtin_amdgcn_readfirstlane(spatial_idx[I0]),
+                        __builtin_amdgcn_readfirstlane(spatial_idx[I1]));
+                }
+#endif
+
+                using Accumulation = ck::detail::
+                    AccumulateWithNanCheck<false /*PropagateNan*/, reduce::Add, FloatAcc>;
+
+                for(int i_m = 0; i_m < MReduceIters; i_m++)
+                {
+                    static_for<0, NReduceIters, 1>{}([&](auto i_n_reduce) {
+                        acc_buf.Clear();
+                        for(auto i = tile_acc_offset_start; i < tile_acc_offset_end; i++)
+                        {
+                            auto c_partial_acc_buf =
+                                make_dynamic_buffer<AddressSpaceEnum::Global,
+                                                    AmdBufferCoherenceEnum::GLC>(
+                                    reinterpret_cast<FloatAcc*>(p_workspace) +
+                                        i * c_partial_acc_block_m_n.GetElementSpaceSize(),
+                                    c_partial_acc_block_m_n.GetElementSpaceSize());
+
+                            acc_load.Run(c_partial_acc_block_m_n,
+                                         c_partial_acc_buf,
+                                         acc_thread_buf_load_desc,
+                                         make_tuple(I0, I0),
+                                         parcial_acc_buf);
+
+                            static_for<0, CBlockTransferScalarPerVector_NWaveNPerXDL, 1>{}(
+                                [&](auto i_vec) {
+                                    constexpr auto offset =
+                                        acc_thread_buf_load_desc.CalculateOffset(
+                                            make_tuple(0, i_vec));
+                                    Accumulation::Calculate(acc_buf(Number<offset>{}),
+                                                            parcial_acc_buf[Number<offset>{}]);
+                                });
+                        }
+
+                        if(thread_n_cluster_id * CBlockTransferScalarPerVector_NWaveNPerXDL <
+                           NPerBlock)
+                        {
+                            acc_store.Run(acc_thread_buf_store_desc,
+                                          make_tuple(I0, I0, I0, I0),
+                                          acc_buf,
+                                          c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                          c_grid_buf);
+                        }
+                        if constexpr(NReduceIters != 1)
+                        {
+                            if constexpr(i_n_reduce != (NReduceIters - 1))
+                            {
+                                acc_load.MoveSrcSliceWindow(c_partial_acc_block_m_n,
+                                                            partial_acc_load_step_n);
+                                acc_store.MoveDstSliceWindow(
+                                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                    partial_acc_store_step_n);
+                            }
+                            else
+                            {
+                                acc_load.MoveSrcSliceWindow(c_partial_acc_block_m_n,
+                                                            partial_acc_load_step_n_reverse);
+                                acc_store.MoveDstSliceWindow(
+                                    c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                    partial_acc_store_step_n_reverse);
+                            }
+                        }
+                    });
+                    {
+                        acc_load.MoveSrcSliceWindow(c_partial_acc_block_m_n,
+                                                    partial_acc_load_step_m);
+                        acc_store.MoveDstSliceWindow(c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                     partial_acc_store_step_m);
+                    }
+                }
+                return;
+            }
+        }
+
+        // offset for last acc buffer of this block
+        uint32_t block_acc_offset =
+            (block_mapping.get_acc_buffer_offset_from_block(block_idx + 1) - 1) * MPerBlock *
+            NPerBlock;
+
+        while(true)
+        {
+            uint32_t current_iter_length = __builtin_amdgcn_readfirstlane(
+                block_mapping.get_current_iter_length(iter_start, iter_end, total_iter_length));
+            uint32_t tile_idx, iter_offset;
+            block_mapping.get_tile_idx_with_offset(iter_end - 1, tile_idx, iter_offset);
+            iter_offset = __builtin_amdgcn_readfirstlane(iter_offset - current_iter_length + 1);
+            auto spatial_idx = block_mapping.tile_to_spatial(tile_idx, m, n);
+
+            const index_t m_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(spatial_idx[I0] * MPerBlock);
+
+            const index_t n_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(spatial_idx[I1] * NPerBlock);
+
+            const index_t k0_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(iter_offset * K0PerBlock);
+
+            // A matrix blockwise copy
+            auto a_blockwise_copy =
+                ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                    AElementwiseOperation,
+                                                    ck::tensor_operation::element_wise::PassThrough,
+                                                    InMemoryDataOperationEnum::Set,
+                                                    Sequence<K0PerBlock, MPerBlock, K1>,
+                                                    ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                    ABlockTransferThreadClusterArrangeOrder,
+                                                    FloatAB,
+                                                    FloatAB,
+                                                    decltype(a_k0_m_k1_grid_desc),
+                                                    decltype(a_block_desc_k0_m_k1),
+                                                    ABlockTransferSrcAccessOrder,
+                                                    Sequence<1, 0, 2>,
+                                                    ABlockTransferSrcVectorDim,
+                                                    2,
+                                                    ABlockTransferSrcScalarPerVector,
+                                                    ABlockTransferDstScalarPerVector_K1,
+                                                    1,
+                                                    1,
+                                                    AThreadTransferSrcResetCoordinateAfterRun,
+                                                    true>(
+                    a_k0_m_k1_grid_desc,
+                    make_multi_index(k0_block_data_idx_on_grid, m_block_data_idx_on_grid, 0),
+                    a_element_op,
+                    a_block_desc_k0_m_k1,
+                    make_multi_index(0, 0, 0),
+                    ck::tensor_operation::element_wise::PassThrough{});
+
+            // B matrix blockwise copy
+            auto b_blockwise_copy =
+                ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                    BElementwiseOperation,
+                                                    ck::tensor_operation::element_wise::PassThrough,
+                                                    InMemoryDataOperationEnum::Set,
+                                                    Sequence<K0PerBlock, NPerBlock, K1>,
+                                                    BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                    BBlockTransferThreadClusterArrangeOrder,
+                                                    FloatAB,
+                                                    FloatAB,
+                                                    decltype(b_k0_n_k1_grid_desc),
+                                                    decltype(b_block_desc_k0_n_k1),
+                                                    BBlockTransferSrcAccessOrder,
+                                                    Sequence<1, 0, 2>,
+                                                    BBlockTransferSrcVectorDim,
+                                                    2,
+                                                    BBlockTransferSrcScalarPerVector,
+                                                    BBlockTransferDstScalarPerVector_K1,
+                                                    1,
+                                                    1,
+                                                    BThreadTransferSrcResetCoordinateAfterRun,
+                                                    true>(
+                    b_k0_n_k1_grid_desc,
+                    make_multi_index(k0_block_data_idx_on_grid, n_block_data_idx_on_grid, 0),
+                    b_element_op,
+                    b_block_desc_k0_n_k1,
+                    make_multi_index(0, 0, 0),
+                    ck::tensor_operation::element_wise::PassThrough{});
+
+            const index_t num_k_block_main_loop = current_iter_length;
+
+            gridwise_gemm_pipeline.Run(a_k0_m_k1_grid_desc,
+                                       a_block_desc_k0_m_k1,
+                                       a_blockwise_copy,
+                                       a_grid_buf,
+                                       a_block_buf,
+                                       a_block_slice_copy_step,
+                                       b_k0_n_k1_grid_desc,
+                                       b_block_desc_k0_n_k1,
+                                       b_blockwise_copy,
+                                       b_grid_buf,
+                                       b_block_buf,
+                                       b_block_slice_copy_step,
+                                       blockwise_gemm,
+                                       c_thread_buf,
+                                       num_k_block_main_loop);
+
+            // output: register to global memory
+            {
+                constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+                constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+                constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
+                    blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+                constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
+                    blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+                constexpr auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I0);
+                constexpr auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I1);
+                constexpr auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I2);
+                constexpr auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I3);
+                constexpr auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I4);
+                constexpr auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I5);
+                constexpr auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I6);
+                constexpr auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I7);
+
+                constexpr auto c_block_desc_mblock_mpershuffle_nblock_npershuffle =
+                    GetCBlockDescriptor_MBlock_MPerShuffle_NBlock_NPerShuffle();
+
+                constexpr auto c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle =
+                    GetCBlockDescriptor_MShuffleRepeat_MPerShuffle_NShuffleRepeat_NPerShuffle();
+
+                auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                    reinterpret_cast<FloatCShuffle*>(p_shared_block),
+                    c_block_desc_mblock_mpershuffle_nblock_npershuffle.GetElementSpaceSize());
+
+                auto c_partial_acc_buf =
+                    make_dynamic_buffer<AddressSpaceEnum::Global, AmdBufferCoherenceEnum::GLC>(
+                        reinterpret_cast<FloatAcc*>(p_workspace) + block_acc_offset,
+                        c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle
+                            .GetElementSpaceSize());
+
+                constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                    c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                    make_tuple(make_freeze_transform(I0), // freeze mblock
+                               make_unmerge_transform(
+                                   make_tuple(CShuffleMRepeatPerShuffle,
+                                              M1,
+                                              M2,
+                                              M3,
+                                              M4)),       // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                               make_freeze_transform(I0), // freeze nblock
+                               make_unmerge_transform(
+                                   make_tuple(CShuffleNRepeatPerShuffle,
+                                              N1,
+                                              N2))), // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<>{},
+                               Sequence<0, 2, 4, 5, 6>{},
+                               Sequence<>{},
+                               Sequence<1, 3, 7>{}));
+
+                // calculate origin of thread output tensor on global memory
+                //     blockwise GEMM c matrix starting index
+                const auto c_thread_mtx_on_block =
+                    blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+                const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+                const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+                const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                    make_single_stage_tensor_adaptor(
+                        make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                        make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                        make_tuple(Sequence<0>{}));
+
+                const auto m_thread_data_on_block_idx =
+                    m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                        make_multi_index(m_thread_data_on_block));
+
+                const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                    make_single_stage_tensor_adaptor(
+                        make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                        make_tuple(Sequence<0, 1, 2>{}),
+                        make_tuple(Sequence<0>{}));
+
+                const auto n_thread_data_on_block_idx =
+                    n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                        make_multi_index(n_thread_data_on_block));
+
+                // VGPR to LDS
+                auto c_thread_copy_vgpr_to_lds = ThreadwiseTensorSliceTransfer_v1r3<
+                    FloatAcc,
+                    FloatCShuffle,
+                    decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
+                    decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                    ck::tensor_operation::element_wise::PassThrough,
+                    Sequence<CShuffleMRepeatPerShuffle,
+                             CShuffleNRepeatPerShuffle,
+                             I1,
+                             I1,
+                             M2,
+                             I1,
+                             M4,
+                             I1>,
+                    Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                    7,
+                    1,
+                    InMemoryDataOperationEnum::Set,
+                    1,
+                    true>{c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                          make_multi_index(0,
+                                           0,
+                                           m_thread_data_on_block_idx[I1],
+                                           n_thread_data_on_block_idx[I1],
+                                           m_thread_data_on_block_idx[I2],
+                                           m_thread_data_on_block_idx[I3],
+                                           m_thread_data_on_block_idx[I4],
+                                           n_thread_data_on_block_idx[I2]),
+                          ck::tensor_operation::element_wise::PassThrough{}};
+
+                // LDS to global
+                auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1r2<
+                    ThisThreadBlock,       // index_t BlockSize,
+                    CElementwiseOperation, // ElementwiseOperation,
+                                           // InMemoryDataOperationEnum::Set, // DstInMemOp,
+                    Sequence<1,
+                             CShuffleMRepeatPerShuffle * MWave * MPerXDL,
+                             1,
+                             CShuffleNRepeatPerShuffle * NWave * NPerXDL>, // BlockSliceLengths,
+                    CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                    FloatCShuffle,        // typename SrcData,
+                    FloatC,               // typename DstData,
+                    decltype(c_block_desc_mblock_mpershuffle_nblock_npershuffle),
+                    decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                    Sequence<0, 1, 2, 3>,                       // typename DimAccessOrder,
+                    3,                                          // index_t VectorDim,
+                    CBlockTransferScalarPerVector_NWaveNPerXDL, // index_t ScalarPerVector,
+                    false, // bool ThreadTransferSrcResetCoordinateAfterRun,
+                    false> // bool ThreadTransferDstResetCoordinateAfterRun
+                    {c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                     make_multi_index(0, 0, 0, 0),
+                     c_grid_desc_mblock_mperblock_nblock_nperblock,
+                     make_multi_index(__builtin_amdgcn_readfirstlane(spatial_idx[I0]),
+                                      0,
+                                      __builtin_amdgcn_readfirstlane(spatial_idx[I1]),
+                                      0),
+                     c_element_op};
+
+                // LDS to global partial acc
+                auto c_block_copy_lds_to_partial_acc = ThreadGroupTensorSliceTransfer_v6r1r2<
+                    ThisThreadBlock,       // index_t BlockSize,
+                    CElementwiseOperation, // ElementwiseOperation,
+                                           // InMemoryDataOperationEnum::Set, // DstInMemOp,
+                    Sequence<1,
+                             CShuffleMRepeatPerShuffle * MWave * MPerXDL,
+                             1,
+                             CShuffleNRepeatPerShuffle * NWave * NPerXDL>, // BlockSliceLengths,
+                    CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                    FloatCShuffle,        // typename SrcData,
+                    FloatCShuffle,        // typename DstData,
+                    decltype(c_block_desc_mblock_mpershuffle_nblock_npershuffle),
+                    decltype(c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle),
+                    Sequence<0, 1, 2, 3>,                       // typename DimAccessOrder,
+                    3,                                          // index_t VectorDim,
+                    CBlockTransferScalarPerVector_NWaveNPerXDL, // index_t ScalarPerVector,
+                    false, // bool ThreadTransferSrcResetCoordinateAfterRun, => need to be false,
+                           // othre wise has scratch
+                    false> // bool ThreadTransferDstResetCoordinateAfterRun, => need to be false,
+                           // othre wise has scratch
+                    {c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                     make_multi_index(0, 0, 0, 0),
+                     c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle,
+                     make_multi_index(0, 0, 0, 0),
+                     c_element_op};
+
+                constexpr auto mxdlperwave_forward_step =
+                    make_multi_index(0, CShuffleMRepeatPerShuffle * MWave * MPerXDL, 0, 0);
+                constexpr auto nxdlperwave_forward_step =
+                    make_multi_index(0, 0, 0, CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+                constexpr auto nxdlperwave_backward_step =
+                    make_multi_index(0, 0, 0, -CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+
+                static_for<0, MRepeat, CShuffleMRepeatPerShuffle>{}([&](auto mxdlperwave_iter) {
+                    constexpr auto mxdlperwave = mxdlperwave_iter;
+
+                    static_for<0, NRepeat, CShuffleNRepeatPerShuffle>{}([&](auto nxdlperwave_iter) {
+                        constexpr bool nxdlperwave_forward_sweep =
+                            (mxdlperwave % (2 * CShuffleMRepeatPerShuffle) == 0);
+
+                        constexpr index_t nxdlperwave_value =
+                            nxdlperwave_forward_sweep
+                                ? nxdlperwave_iter
+                                : (NRepeat - nxdlperwave_iter - CShuffleNRepeatPerShuffle);
+
+                        constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
+
+                        // make sure it's safe to do ds_write
+                        block_sync_lds();
+
+                        // VGPR to LDS
+                        c_thread_copy_vgpr_to_lds.Run(
+                            c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
+                            make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
+                            c_thread_buf,
+                            c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                            c_block_buf);
+
+                        // make sure it's safe to do ds_read
+                        block_sync_lds();
+
+                        c_block_copy_lds_to_global.SetSrcSliceOrigin(
+                            c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                            make_tuple(0, 0, 0, 0));
+
+                        // LDS to global
+                        if(is_dp_block)
+                            c_block_copy_lds_to_global.template Run<decltype(c_block_buf),
+                                                                    decltype(c_grid_buf),
+                                                                    InMemoryDataOperationEnum::Set>(
+                                c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                                c_block_buf,
+                                c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                c_grid_buf);
+                        else if(is_sk_block)
+                        {
+                            if constexpr(Block2CTileMap::ReductionStrategy ==
+                                         StreamKReductionStrategy::Reduction)
+                            {
+                                // constexpr offset
+                                c_block_copy_lds_to_partial_acc.SetSrcSliceOrigin(
+                                    c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                                    make_tuple(0, 0, 0, 0));
+
+                                c_block_copy_lds_to_partial_acc.SetDstSliceOrigin(
+                                    c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle,
+                                    make_tuple(mxdlperwave.value, 0, nxdlperwave.value, 0));
+
+                                c_block_copy_lds_to_partial_acc
+                                    .template Run<decltype(c_block_buf),
+                                                  decltype(c_partial_acc_buf),
+                                                  InMemoryDataOperationEnum::Set>(
+                                        c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                                        c_block_buf,
+                                        c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle,
+                                        c_partial_acc_buf);
+                            }
+                            else if constexpr(Block2CTileMap::ReductionStrategy ==
+                                              StreamKReductionStrategy::Atomic)
+                            {
+                                c_block_copy_lds_to_global
+                                    .template Run<decltype(c_block_buf),
+                                                  decltype(c_grid_buf),
+                                                  InMemoryDataOperationEnum::AtomicAdd>(
+                                        c_block_desc_mblock_mpershuffle_nblock_npershuffle,
+                                        c_block_buf,
+                                        c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                        c_grid_buf);
+                            }
+                        }
+
+                        // move on nxdlperwave dimension
+                        if constexpr(nxdlperwave_forward_sweep &&
+                                     (nxdlperwave < NRepeat - CShuffleNRepeatPerShuffle))
+                        {
+                            c_block_copy_lds_to_global.MoveDstSliceWindow(
+                                c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                nxdlperwave_forward_step);
+                        }
+                        else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
+                        {
+                            c_block_copy_lds_to_global.MoveDstSliceWindow(
+                                c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                nxdlperwave_backward_step);
+                        }
+                    });
+
+                    // move on mxdlperwave dimension
+                    if constexpr(mxdlperwave < MRepeat - CShuffleMRepeatPerShuffle)
+                    {
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mperblock_nblock_nperblock,
+                            mxdlperwave_forward_step);
+                    }
+                });
+
+                if constexpr(Block2CTileMap::ReductionStrategy ==
+                             StreamKReductionStrategy::Reduction)
+                {
+                    if(is_sk_block)
+                    {
+                        // increase the counter for this tile
+                        workgroup_barrier wg_barrier(p_semaphore);
+                        wg_barrier.inc(tile_idx);
+                    }
+                }
+            }
+
+            // exit condition
+            iter_end -= current_iter_length;
+            if(iter_end <= iter_start)
+                break;
+
+            if constexpr(Block2CTileMap::ReductionStrategy == StreamKReductionStrategy::Reduction)
+            {
+                block_acc_offset -= MPerBlock * NPerBlock;
+            }
+            // make sure next loop LDS is ready for use
+            block_sync_lds();
+        }
+    }
+
+    template <typename Layout>
+    struct LStr
+    {
+        static std::string Get() { return ""; }
+    };
+
+    template <>
+    struct LStr<ck::tensor_layout::gemm::RowMajor>
+    {
+        static std::string Get() { return "R"; }
+    };
+
+    template <>
+    struct LStr<ck::tensor_layout::gemm::ColumnMajor>
+    {
+        static std::string Get() { return "C"; }
+    };
+
+    static std::string GetTypeString()
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "GemmXdlStreamK_"
+            << std::string(ALayout::name)[0]
+            << std::string(BLayout::name)[0]
+            << std::string(CLayout::name)[0]
+            << "_"
+            << "B" << BlockSize << "_"
+            << "Vec" << ABlockTransferSrcScalarPerVector << "x"
+            << BBlockTransferSrcScalarPerVector << "x"
+            << CBlockTransferScalarPerVector_NWaveNPerXDL << "_"
+            << MPerBlock << "x"
+            << NPerBlock << "x"
+            << K0PerBlock << "x"
+            << K1 ;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
new file mode 100644
index 0000000000..31aee70db2
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
@@ -0,0 +1,1005 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDesc_M_N,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+#if CK_USE_WAVES_PER_EU
+        __attribute__((amdgpu_waves_per_eu(CK_MIN_WAVES_PER_EU, CK_MAX_WAVES_PER_EU)))
+#endif
+        kernel_gemm_xdlops_v2r3(const FloatAB* __restrict__ p_a_grid,
+                                const FloatAB* __restrict__ p_b_grid,
+                                FloatC* __restrict__ p_c_grid,
+                                const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+                                const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+                                const CGridDesc_M_N c_grid_desc_m_n)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_shared,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  c_grid_desc_m_n);
+#else
+    ignore                = p_a_grid;
+    ignore                = p_b_grid;
+    ignore                = p_c_grid;
+    ignore                = a_grid_desc_k0_m_k1;
+    ignore                = b_grid_desc_k0_n_k1;
+    ignore                = c_grid_desc_m_n;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <typename GridwiseGemm, bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+#if CK_USE_WAVES_PER_EU
+        __attribute__((amdgpu_waves_per_eu(CK_MIN_WAVES_PER_EU, CK_MAX_WAVES_PER_EU)))
+#endif
+        kernel_gemm_xdlops_v2r3(const typename GridwiseGemm::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const auto a_grid_desc_k0_m_k1 =
+        amd_wave_read_first_lane(GridwiseGemm::MakeAGridDescriptor_K0_M_K1(
+            karg.M, karg.MPadded, karg.K, karg.K0, karg.StrideA));
+    const auto b_grid_desc_k0_n_k1 =
+        amd_wave_read_first_lane(GridwiseGemm::MakeBGridDescriptor_K0_N_K1(
+            karg.K, karg.N, karg.NPadded, karg.K0, karg.StrideB));
+    const auto c_grid_desc_m_n = amd_wave_read_first_lane(GridwiseGemm::MakeCGridDescriptor_M_N(
+        karg.M, karg.MPadded, karg.N, karg.NPadded, karg.StrideC));
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(karg.p_a_grid,
+                                                  karg.p_b_grid,
+                                                  karg.p_c_grid,
+                                                  p_shared,
+                                                  a_grid_desc_k0_m_k1,
+                                                  b_grid_desc_k0_n_k1,
+                                                  c_grid_desc_m_n);
+#else
+    ignore                = karg;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          bool BBlockLdsExtraN,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          index_t NumGemmKPrefetchStage = 1,
+          LoopScheduler LoopSched       = make_default_loop_scheduler(),
+          PipelineVersion PipelineVer   = PipelineVersion::v1>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __host__ static auto CalculateGridSize(index_t M, index_t N)
+    {
+        return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, 1);
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ static auto CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        return std::make_tuple(Block2CTileMap::CalculateGridSize(c_grid_desc_m_n), 1, 1);
+    }
+
+    template <typename>
+    __host__ static auto CalculateGridSize(index_t M, index_t N)
+    {
+        return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, 1);
+    }
+
+    __host__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_divide_ceil(M, MPerBlock) * MPerBlock;
+    }
+
+    __host__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_divide_ceil(N, NPerBlock) * NPerBlock;
+    }
+
+    __host__ static auto CalculateK0(index_t K) { return math::integer_divide_floor(K, K1Value); }
+
+    // Argument
+    struct Problem
+    {
+        __host__ Problem(index_t M_,
+                         index_t N_,
+                         index_t K_,
+                         index_t StrideA_,
+                         index_t StrideB_,
+                         index_t StrideC_)
+            : M{M_},
+              N{N_},
+              K{K_},
+              StrideA{StrideA_},
+              StrideB{StrideB_},
+              StrideC{StrideC_},
+              MPadded{CalculateMPadded(M_)},
+              NPadded{CalculateNPadded(N_)},
+              K0{CalculateK0(K)}
+        {
+        }
+
+        __host__ void Print() const
+        {
+            std::cout << "problem {"
+                      << "M:" << M << ", "
+                      << "N:" << N << ", "
+                      << "K:" << K << ", "
+                      << "SA:" << StrideA << ", "
+                      << "SB:" << StrideB << ", "
+                      << "SC:" << StrideC << ", "
+                      << "MP:" << MPadded << ", "
+                      << "NP:" << NPadded << ", "
+                      << "K0:" << K0 << "}" << std::endl;
+        }
+
+        index_t M;
+        index_t N;
+        index_t K;
+        index_t StrideA;
+        index_t StrideB;
+        index_t StrideC;
+        index_t MPadded;
+        index_t NPadded;
+        index_t K0;
+    };
+
+    // Argument
+    struct Argument : public Problem, public tensor_operation::device::BaseArgument
+    {
+        __host__ Argument(const FloatAB* p_a_grid_,
+                          const FloatAB* p_b_grid_,
+                          FloatC* p_c_grid_,
+                          index_t M_,
+                          index_t N_,
+                          index_t K_,
+                          index_t StrideA_,
+                          index_t StrideB_,
+                          index_t StrideC_)
+            : Problem{M_, N_, K_, StrideA_, StrideB_, StrideC_},
+              p_a_grid{p_a_grid_},
+              p_b_grid{p_b_grid_},
+              p_c_grid{p_c_grid_}
+        {
+        }
+
+        const FloatAB* p_a_grid;
+        const FloatAB* p_b_grid;
+        FloatC* p_c_grid;
+    };
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    // denorm test fix, required to work around fp16 mfma issue
+    // we convert fp16->fp32->bf16 and execute bf16 mfma instruction
+    // when mfma if fixed, remove this section and update
+    // FloatABAdjusted -> FloatAB throughout this file
+#if CK_WORKAROUND_DENORM_FIX
+    using FloatABAdjusted = conditional_t<is_same_v<FloatAB, ck::half_t>, ck::bhalf_t, FloatAB>;
+#else
+    using FloatABAdjusted = FloatAB;
+#endif
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return a_block_desc_k0_m_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return b_block_desc_k0_n_k1;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        constexpr auto max_lds_align = K1;
+
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned =
+            math::integer_least_multiple(b_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
+
+        return (a_block_space_size_aligned + b_block_space_size_aligned) * sizeof(FloatAB);
+    }
+
+    template <typename AGridDesc_K0_M_K1, typename BGridDesc_K0_N_K1, typename CGridDesc_M_N>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXDL * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXDL)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+             K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+             K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K0 / K0PerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ static constexpr bool CheckValidity(const Problem& problem)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXDL * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXDL)) == 0,
+                      "Invalid tuning param!");
+
+        // check gridwise gemm pipeline
+        const index_t K0      = problem.K / K1Value;
+        const auto num_k_loop = K0 / K0PerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / (K0PerBlock * K1);
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    template <typename CGridDesc>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CGridDesc& c_grid_desc_m_n)
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        using BlockwiseGemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatABAdjusted,
+                                                                FloatAcc,
+                                                                decltype(a_block_desc_k0_m_k1),
+                                                                decltype(b_block_desc_k0_n_k1),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MXdlPerWave,
+                                                                NXdlPerWave,
+                                                                K1>;
+
+        return BlockwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n);
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    using Block2CTileMap = BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock>;
+
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_K0_M_K1,
+              typename BGridDesc_K0_N_K1,
+              typename CGridDesc_M_N>
+    __device__ static void Run(const FloatAB* p_a_grid,
+                               const FloatAB* p_b_grid,
+                               FloatC* p_c_grid,
+                               void* __restrict__ p_shared,
+                               const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                               const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                               const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+            MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n);
+
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetElementSpaceSize());
+
+        const AElementwiseOperation a_element_op{};
+        const BElementwiseOperation b_element_op{};
+        const CElementwiseOperation c_element_op{};
+
+        const auto block_2_ctile_map =
+            Block2CTileMap{c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1)};
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I0),
+                          c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I1))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<K0PerBlock, MPerBlock, K1>,
+                                                ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatABAdjusted,
+                                                decltype(a_grid_desc_k0_m_k1),
+                                                decltype(a_block_desc_k0_m_k1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_k0_m_k1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_k0_m_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<K0PerBlock, NPerBlock, K1>,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatABAdjusted,
+                                                decltype(b_grid_desc_k0_n_k1),
+                                                decltype(b_block_desc_k0_n_k1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_k0_n_k1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_k0_n_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            FloatABAdjusted,
+            FloatAcc,
+            decltype(a_block_desc_k0_m_k1),
+            decltype(b_block_desc_k0_n_k1),
+            MPerXDL,
+            NPerXDL,
+            MXdlPerWave,
+            NXdlPerWave,
+            K1,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatABAdjusted*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatABAdjusted*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_k0_n_k1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto K0                       = a_grid_desc_k0_m_k1.GetLength(I0);
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
+
+        GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc_k0_m_k1,
+                                                          a_block_desc_k0_m_k1,
+                                                          a_blockwise_copy,
+                                                          a_grid_buf,
+                                                          a_block_buf,
+                                                          a_block_slice_copy_step,
+                                                          b_grid_desc_k0_n_k1,
+                                                          b_block_desc_k0_n_k1,
+                                                          b_blockwise_copy,
+                                                          b_grid_buf,
+                                                          b_block_buf,
+                                                          b_block_slice_copy_step,
+                                                          blockwise_gemm,
+                                                          c_thread_buf,
+                                                          num_k_block_main_loop);
+
+        // output: register to global memory
+        {
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetLength(I7);
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_grid =
+                m_block_data_idx_on_grid + c_thread_mtx_on_block[I0];
+
+            const index_t n_thread_data_on_grid =
+                n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_grid_idx =
+                m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_grid));
+
+            const auto n_thread_data_on_grid_to_n0_n1_n2_adaptor = make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                make_tuple(Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_grid_idx =
+                n_thread_data_on_grid_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_grid));
+
+            auto c_thread_copy =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   CElementwiseOperation,
+                                                   Sequence<M0, N0, I1, I1, M2, I1, M4, I1>,
+                                                   CThreadTransferSrcDstAccessOrder,
+                                                   CThreadTransferSrcDstVectorDim,
+                                                   CThreadTransferDstScalarPerVector,
+                                                   CGlobalMemoryDataOperation,
+                                                   1,
+                                                   true>{
+                    c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(m_thread_data_on_grid_idx[I0],
+                                     n_thread_data_on_grid_idx[I0],
+                                     m_thread_data_on_grid_idx[I1],
+                                     n_thread_data_on_grid_idx[I1],
+                                     m_thread_data_on_grid_idx[I2],
+                                     m_thread_data_on_grid_idx[I3],
+                                     m_thread_data_on_grid_idx[I4],
+                                     n_thread_data_on_grid_idx[I2]),
+                    c_element_op};
+
+            c_thread_copy.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                              c_thread_buf,
+                              c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                              c_grid_buf);
+        }
+    }
+};
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          tensor_operation::device::GemmSpecialization GemmSpec,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          bool BBlockLdsExtraN,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          index_t NumGemmKPrefetchStage = 1,
+          LoopScheduler LoopSched       = make_default_loop_scheduler(),
+          PipelineVersion PipelineVer   = PipelineVersion::v1>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext
+    : GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<BlockSize,
+                                              FloatAB,
+                                              FloatAcc,
+                                              FloatC,
+                                              CGlobalMemoryDataOperation,
+                                              AElementwiseOperation,
+                                              BElementwiseOperation,
+                                              CElementwiseOperation,
+                                              MPerBlock,
+                                              NPerBlock,
+                                              K0PerBlock,
+                                              MPerXDL,
+                                              NPerXDL,
+                                              K1Value,
+                                              MXdlPerWave,
+                                              NXdlPerWave,
+                                              ABlockTransferThreadClusterLengths_K0_M_K1,
+                                              ABlockTransferThreadClusterArrangeOrder,
+                                              ABlockTransferSrcAccessOrder,
+                                              ABlockTransferSrcVectorDim,
+                                              ABlockTransferSrcScalarPerVector,
+                                              ABlockTransferDstScalarPerVector_K1,
+                                              AThreadTransferSrcResetCoordinateAfterRun,
+                                              ABlockLdsExtraM,
+                                              BBlockTransferThreadClusterLengths_K0_N_K1,
+                                              BBlockTransferThreadClusterArrangeOrder,
+                                              BBlockTransferSrcAccessOrder,
+                                              BBlockTransferSrcVectorDim,
+                                              BBlockTransferSrcScalarPerVector,
+                                              BBlockTransferDstScalarPerVector_K1,
+                                              BThreadTransferSrcResetCoordinateAfterRun,
+                                              BBlockLdsExtraN,
+                                              CThreadTransferSrcDstAccessOrder,
+                                              CThreadTransferSrcDstVectorDim,
+                                              CThreadTransferDstScalarPerVector,
+                                              NumGemmKPrefetchStage,
+                                              LoopSched,
+                                              PipelineVer>
+{
+    using Parent =
+        GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<BlockSize,
+                                                FloatAB,
+                                                FloatAcc,
+                                                FloatC,
+                                                CGlobalMemoryDataOperation,
+                                                AElementwiseOperation,
+                                                BElementwiseOperation,
+                                                CElementwiseOperation,
+                                                MPerBlock,
+                                                NPerBlock,
+                                                K0PerBlock,
+                                                MPerXDL,
+                                                NPerXDL,
+                                                K1Value,
+                                                MXdlPerWave,
+                                                NXdlPerWave,
+                                                ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ABlockTransferSrcAccessOrder,
+                                                ABlockTransferSrcVectorDim,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                ABlockLdsExtraM,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                BBlockTransferSrcAccessOrder,
+                                                BBlockTransferSrcVectorDim,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                BBlockLdsExtraN,
+                                                CThreadTransferSrcDstAccessOrder,
+                                                CThreadTransferSrcDstVectorDim,
+                                                CThreadTransferDstScalarPerVector,
+                                                NumGemmKPrefetchStage,
+                                                LoopSched,
+                                                PipelineVer>;
+
+    using typename Parent::GridwiseGemmPipe;
+    using typename Parent::Problem;
+
+    using Parent::I1;
+
+    using Parent::K1;
+
+    __device__ static auto
+    MakeAGridDescriptor_K0_M_K1(index_t M, index_t MPad, index_t K, index_t K0, index_t StrideA)
+    {
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Value)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Value)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    __device__ static auto
+    MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t NPad, index_t K0, index_t StrideB)
+    {
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Value)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0, K1Value)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+    }
+
+    __device__ static auto
+    MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
+        {
+            return transform_tensor_descriptor(c_grid_desc_m_n,
+                                               make_tuple(make_right_pad_transform(M, MPad - M),
+                                                          make_right_pad_transform(N, NPad - N)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+
+            return transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+    }
+
+    __host__ static constexpr bool CheckValidity(const Problem& problem)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXDL * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXDL)) == 0,
+                      "Invalid tuning param!");
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+            if(!(problem.M % MPerBlock == 0))
+            {
+                return false;
+            }
+        }
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+            if(!(problem.N % NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            if(problem.K % ABlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.M % ABlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        {
+            if(problem.N % BBlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.K % BBlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+
+        // check gridwise gemm pipeline
+        const index_t K0      = problem.K / K1;
+        const auto num_k_loop = K0 / K0PerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
new file mode 100644
index 0000000000..19fbee727f
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
@@ -0,0 +1,617 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename ABK0MK1GridDesc,
+          typename BBK0NK1GridDesc,
+          typename CM0N0M1N1M2M3M4N2GridDesc,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename CBlockClusterAdaptor,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v2r4(const FloatAB* __restrict__ p_a_grid,
+                                const FloatAB* __restrict__ p_b_grid,
+                                FloatC* __restrict__ p_c_grid,
+                                const ABK0MK1GridDesc a_b_k0_m_k1_grid_desc,
+                                const BBK0NK1GridDesc b_b_k0_n_k1_grid_desc,
+                                const CM0N0M1N1M2M3M4N2GridDesc c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                                const AElementwiseOperation a_element_op,
+                                const BElementwiseOperation b_element_op,
+                                const CElementwiseOperation c_element_op,
+                                const CBlockClusterAdaptor c_block_cluster_adaptor)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    constexpr index_t shared_block_size =
+        GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
+
+    __shared__ FloatAB p_shared_block[shared_block_size];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_c_grid,
+                                                  p_shared_block,
+                                                  a_b_k0_m_k1_grid_desc,
+                                                  b_b_k0_n_k1_grid_desc,
+                                                  c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                                                  a_element_op,
+                                                  b_element_op,
+                                                  c_element_op,
+                                                  c_block_cluster_adaptor);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = a_b_k0_m_k1_grid_desc;
+    ignore = b_b_k0_n_k1_grid_desc;
+    ignore = c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = c_block_cluster_adaptor;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <index_t BlockSize,
+          typename FloatAB,
+          typename FloatAcc,
+          typename FloatC,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename ABK0MK1GridDesc,
+          typename BBK0NK1GridDesc,
+          typename CMNGridDesc,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MRepeat,
+          index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          bool BBlockLdsExtraN,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector>
+struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size =
+            math::integer_least_multiple(b_k0_n_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        return (a_block_space_size + b_block_space_size) * sizeof(FloatAB);
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const ABK0MK1GridDesc& a_b_k0_m_k1_grid_desc,
+                  const BBK0NK1GridDesc& b_b_k0_n_k1_grid_desc,
+                  const CMNGridDesc& c_m_n_grid_desc,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXDL * MRepeat) == 0) &&
+                          (NPerBlock % (NRepeat * NPerXDL)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M      = a_b_k0_m_k1_grid_desc.GetLength(I2);
+        const auto N      = b_b_k0_n_k1_grid_desc.GetLength(I2);
+        const auto K0     = a_b_k0_m_k1_grid_desc.GetLength(I1);
+        const auto KBatch = a_b_k0_m_k1_grid_desc.GetLength(I0);
+
+        if(!(M == c_m_n_grid_desc.GetLength(I0) && N == c_m_n_grid_desc.GetLength(I1) &&
+             K0 == b_b_k0_n_k1_grid_desc.GetLength(I1) &&
+             K1 == a_b_k0_m_k1_grid_desc.GetLength(I3) &&
+             K1 == b_b_k0_n_k1_grid_desc.GetLength(I3) &&
+             KBatch == b_b_k0_n_k1_grid_desc.GetLength(I0)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        if(!block_2_ctile_map.CheckValidity(c_m_n_grid_desc))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
+    {
+        const bool has_main_k0_block_loop = K0 > K0PerBlock;
+
+        return has_main_k0_block_loop;
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCM0N0M1N1M2M3M4N2GridDescriptor(const CMNGridDesc& c_m_n_grid_desc)
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        using BlockwiseGemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_k0_m_k1_block_desc),
+                                                                decltype(b_k0_n_k1_block_desc),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MRepeat,
+                                                                NRepeat,
+                                                                K1>;
+
+        return BlockwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_m_n_grid_desc);
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
+        const CMNGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch)
+    {
+        return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CMNGridDesc>(
+            c_m_n_grid_desc, 8, KBatch);
+    }
+
+    using CM0N0M1N1M2M3M4N2GridDesc = decltype(MakeCM0N0M1N1M2M3M4N2GridDescriptor(CMNGridDesc{}));
+    using CBlockClusterAdaptor      = decltype(MakeCBlockClusterAdaptor(CMNGridDesc{}, 1, 1, 1));
+
+    template <bool HasMainKBlockLoop>
+    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
+                               const FloatAB* __restrict__ p_b_grid,
+                               FloatC* __restrict__ p_c_grid,
+                               FloatAB* __restrict__ p_shared_block,
+                               const ABK0MK1GridDesc& a_b_k0_m_k1_grid_desc,
+                               const BBK0NK1GridDesc& b_b_k0_n_k1_grid_desc,
+                               const CM0N0M1N1M2M3M4N2GridDesc& c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CElementwiseOperation& c_element_op,
+                               const CBlockClusterAdaptor& c_block_cluster_adaptor)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_b_k0_n_k1_grid_desc.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetElementSpaceSize());
+
+        const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            c_block_cluster_adaptor.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!c_block_cluster_adaptor.ValidCTileIndex(
+               make_tuple(block_work_idx[I1], block_work_idx[I2]),
+               make_tuple(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetLength(I0),
+                          c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetLength(I1))))
+        {
+            return;
+        }
+
+        const index_t k_batch_id = block_work_idx[I0];
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        constexpr auto a_b_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<K0PerBlock>{} * Number<MPerBlock + 1>{} * K1,
+                               Number<MPerBlock + 1>{} * K1,
+                               K1,
+                               I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        constexpr auto b_b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<K0PerBlock>{} * Number<NPerBlock + 1>{} * K1,
+                               Number<NPerBlock + 1>{} * K1,
+                               K1,
+                               I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, K0PerBlock, MPerBlock, K1>,
+                                                ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_b_k0_m_k1_grid_desc),
+                                                decltype(a_b_k0_m_k1_block_desc),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                a_b_k0_m_k1_grid_desc,
+                make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_b_k0_m_k1_block_desc,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, K0PerBlock, NPerBlock, K1>,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_b_k0_n_k1_grid_desc),
+                                                decltype(b_b_k0_n_k1_block_desc),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                b_b_k0_n_k1_grid_desc,
+                make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_b_k0_n_k1_block_desc,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_k0_m_k1_block_desc),
+                                                                decltype(b_k0_n_k1_block_desc),
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MRepeat,
+                                                                NRepeat,
+                                                                K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        FloatAB* p_a_block = p_shared_block;
+        FloatAB* p_b_block = p_shared_block + a_block_space_size;
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
+
+        // preload data into LDS
+        {
+            a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
+            b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
+
+            a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
+            b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
+        }
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // main body
+        if constexpr(HasMainKBlockLoop)
+        {
+            index_t k0_block_data_begin = 0;
+
+            do
+            {
+                a_blockwise_copy.MoveSrcSliceWindow(a_b_k0_m_k1_grid_desc, a_block_slice_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_b_k0_n_k1_grid_desc, b_block_slice_copy_step);
+
+                a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
+
+                block_sync_lds();
+
+                b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
+
+                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+
+                block_sync_lds();
+
+                a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
+                b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
+
+                k0_block_data_begin += K0PerBlock;
+            } while(k0_block_data_begin < (K0 - K0PerBlock));
+        }
+
+        // tail
+        {
+            block_sync_lds();
+
+            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
+        }
+
+        // output: register to global memory
+        {
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I0);
+            constexpr auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I1);
+            constexpr auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I2);
+            constexpr auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I3);
+            constexpr auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I4);
+            constexpr auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I5);
+            constexpr auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I6);
+            constexpr auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I7);
+
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(
+                    Number<M0>{}, Number<N0>{}, I1, I1, Number<M2>{}, I1, Number<M4>{}, I1));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_grid =
+                m_block_data_idx_on_grid + c_thread_mtx_on_block[I0];
+
+            const index_t n_thread_data_on_grid =
+                n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_grid_idx =
+                m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_grid));
+
+            const auto n_thread_data_on_grid_to_n0_n1_n2_adaptor = make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                make_tuple(Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_grid_idx =
+                n_thread_data_on_grid_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_grid));
+
+            auto c_thread_copy =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
+                                                   decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc),
+                                                   CElementwiseOperation,
+                                                   Sequence<M0, N0, I1, I1, M2, I1, M4, I1>,
+                                                   CThreadTransferSrcDstAccessOrder,
+                                                   CThreadTransferSrcDstVectorDim,
+                                                   CThreadTransferDstScalarPerVector,
+                                                   CGlobalMemoryDataOperation,
+                                                   1,
+                                                   true>{
+
+                    c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                    make_multi_index(m_thread_data_on_grid_idx[I0],
+                                     n_thread_data_on_grid_idx[I0],
+                                     m_thread_data_on_grid_idx[I1],
+                                     n_thread_data_on_grid_idx[I1],
+                                     m_thread_data_on_grid_idx[I2],
+                                     m_thread_data_on_grid_idx[I3],
+                                     m_thread_data_on_grid_idx[I4],
+                                     n_thread_data_on_grid_idx[I2]),
+                    c_element_op};
+
+            c_thread_copy.Run(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
+                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
+                              c_thread_buf,
+                              c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
+                              c_grid_buf);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
new file mode 100644
index 0000000000..371281dfe2
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
@@ -0,0 +1,1037 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          bool HasMainKBlockLoop,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename Block2CTileMap>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v2r4r2_simplified(typename GridwiseGemm::Argument karg,
+                                             const Block2CTileMap& b2c_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
+
+    __shared__ uint8_t p_shared[shared_size];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation>(
+        karg, static_cast<void*>(p_shared), b2c_map);
+#else
+    ignore = karg;
+    ignore = b2c_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <index_t BlockSize,
+          typename FloatA,
+          typename FloatB,
+          typename FloatAcc,
+          typename FloatC,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          tensor_operation::device::GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t K0PerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t K1Value,
+          index_t MRepeat,
+          index_t NRepeat,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMRepeatPerShuffle,
+          index_t CShuffleNRepeatPerShuffle,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXDL,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          LoopScheduler LoopSched     = make_default_loop_scheduler(),
+          PipelineVersion PipelineVer = PipelineVersion::v1,
+          typename ComputeType        = FloatC>
+struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1  = Number<K1Value>{};
+    static constexpr auto M01 = 1;
+    static constexpr auto N01 = 1;
+
+    static constexpr auto gemm_padder =
+        tensor_operation::device::GemmPadder<GemmSpec, index_t, index_t, index_t>{
+            MPerBlock, NPerBlock, K1* K0PerBlock};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        const FloatA* p_a_grid;
+        const FloatB* p_b_grid;
+        FloatC* p_c_grid;
+        index_t M;
+        index_t N;
+        index_t K;
+        index_t StrideA;
+        index_t StrideB;
+        index_t StrideC;
+        index_t MPadded;
+        index_t NPadded;
+        index_t KPadded;
+        index_t K0;
+        index_t k_batch;
+
+        Argument(const FloatA* p_a_grid_,
+                 const FloatB* p_b_grid_,
+                 FloatC* p_c_grid_,
+                 index_t M_,
+                 index_t N_,
+                 index_t K_,
+                 index_t StrideA_,
+                 index_t StrideB_,
+                 index_t StrideC_,
+                 index_t MPadded_,
+                 index_t NPadded_,
+                 index_t KPadded_,
+                 index_t K0_,
+                 index_t k_batch_)
+            : p_a_grid(p_a_grid_),
+              p_b_grid(p_b_grid_),
+              p_c_grid(p_c_grid_),
+              M(M_),
+              N(N_),
+              K(K_),
+              StrideA(StrideA_),
+              StrideB(StrideB_),
+              StrideC(StrideC_),
+              MPadded(MPadded_),
+              NPadded(NPadded_),
+              KPadded(KPadded_),
+              K0(K0_),
+              k_batch(k_batch_)
+        {
+        }
+
+        void Print() const
+        {
+            std::cout << "arg {"
+                      << "M:" << M << ", "
+                      << "N:" << N << ", "
+                      << "K:" << K << ", "
+                      << "SA:" << StrideA << ", "
+                      << "SB:" << StrideB << ", "
+                      << "SC:" << StrideC << ", "
+                      << "MP:" << MPadded << ", "
+                      << "NP:" << NPadded << ", "
+                      << "KP:" << KPadded << ", "
+                      << "K0:" << K0 << ", "
+                      << "KB:" << k_batch << "}" << std::endl;
+        }
+    };
+
+    __host__ __device__ static auto CalculateGridSize(const Argument& karg)
+    {
+        return std::make_tuple(math::integer_divide_ceil(karg.N, NPerBlock),
+                               math::integer_divide_ceil(karg.M, MPerBlock),
+                               karg.k_batch);
+    }
+
+    // prefer this to be called on host
+    __host__ __device__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_least_multiple(M, MPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_least_multiple(N, NPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateK0(index_t K, index_t K_Batch = 1)
+    {
+        // k_batch * k0 * k0_per_block * k1
+        auto K_t = K_Batch * K0PerBlock * K1;
+        return (K + K_t - 1) / K_t * K0PerBlock;
+    }
+
+    __host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch = 1)
+    {
+        auto K0 = CalculateK0(K, K_Batch);
+        return K_Batch * K0 * K1;
+    }
+
+    __host__ __device__ static auto MakeAGridDescriptor_KBatch_K0_M_K1(index_t M,
+                                                                       index_t MPad,
+                                                                       index_t K,
+                                                                       index_t StrideA,
+                                                                       index_t KBatch,
+                                                                       index_t K0,
+                                                                       index_t KPad)
+    {
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
+        {
+            // const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    __host__ __device__ static auto MakeBGridDescriptor_KBatch_K0_N_K1(index_t K,
+                                                                       index_t NPad,
+                                                                       index_t N,
+                                                                       index_t StrideB,
+                                                                       index_t KBatch,
+                                                                       index_t K0,
+                                                                       index_t KPad)
+    {
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
+            b_grid_desc_k_n,
+            make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
+        {
+            // const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    __host__ __device__ static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
+    {
+        const auto c_grid_desc_m_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        return gemm_padder.PadCDescriptor_M_N(c_grid_desc_m_n);
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size =
+            math::integer_least_multiple(b_k0_n_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto c_block_size =
+            GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock().GetElementSpaceSize();
+
+        return math::max((a_block_space_size + b_block_space_size) * sizeof(ComputeType),
+                         c_block_size * sizeof(FloatC));
+    }
+
+    __host__ __device__ static constexpr bool CheckValidity(const Argument& karg)
+    {
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+            if(!(karg.M % MPerBlock == 0))
+            {
+#if DEBUG_LOG
+                std::cout << "Arg M value is not a multiple of MPerBlock! M: " << karg.M << " "
+                          << __FILE__ << ":" << __LINE__ << ", in function: " << __func__
+                          << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+            if(!(karg.N % NPerBlock == 0))
+            {
+#if DEBUG_LOG
+                std::cout << "Arg N value is not a multiple of NPerBlock! N: " << karg.N << " "
+                          << __FILE__ << ":" << __LINE__ << ", in function: " << __func__
+                          << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            if(karg.K % ABlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg K (" << karg.K
+                          << ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
+                          << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.M % ABlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg M (" << karg.M
+                          << ") value is not a multiple of ABlockTransferSrcScalarPerVector ("
+                          << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        {
+            if(karg.N % BBlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg N (" << karg.N
+                          << ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
+                          << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.K % BBlockTransferSrcScalarPerVector != 0)
+            {
+#if DEBUG_LOG
+                std::cout << "Arg K (" << karg.K
+                          << ") value is not a multiple of BBlockTransferSrcScalarPerVector ("
+                          << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":"
+                          << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+        {
+            if(karg.N % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
+            {
+#if DEBUG_LOG
+                std::cout
+                    << "Arg N (" << karg.N
+                    << ") value is not a multiple of CBlockTransferScalarPerVector_NWaveNPerXDL ("
+                    << CBlockTransferScalarPerVector_NWaveNPerXDL << " )! " << __FILE__ << ":"
+                    << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+        else
+        {
+            if(karg.M % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
+            {
+#if DEBUG_LOG
+                std::cout
+                    << "Arg M (" << karg.M
+                    << ") value is not a multiple of CBlockTransferScalarPerVector_NWaveNPerXDL ("
+                    << CBlockTransferScalarPerVector_NWaveNPerXDL << " )! " << __FILE__ << ":"
+                    << __LINE__ << ", in function: " << __func__ << std::endl;
+
+#endif // DEBUG_LOG
+                return false;
+            }
+        }
+
+        const auto num_k_loop = karg.K0 / K0PerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+#if DEBUG_LOG
+            std::cout << "The number of k loops (" << num_k_loop
+                      << ") value is not supported by GridwiseGemm Pipeline."
+                      << " K0: " << karg.K0 << ", K0PerBlock: " << K0PerBlock << " " << __FILE__
+                      << ":" << __LINE__ << ", in function: " << __func__ << std::endl;
+#endif // DEBUG_LOG
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static auto GetKPad(index_t K, index_t KBatch)
+    {
+        const index_t K0   = math::integer_divide_ceil(K, K1 * K0PerBlock * KBatch) * K0PerBlock;
+        const index_t KPad = KBatch * K0 * K1;
+        return KPad;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
+    {
+        const index_t num_loop = K0 / K0PerBlock;
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    template <typename CGridDesc>
+    __host__ __device__ static constexpr auto
+    MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc& c_m_n_grid_desc)
+    {
+        const auto M = c_m_n_grid_desc.GetLength(I0);
+        const auto N = c_m_n_grid_desc.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        return transform_tensor_descriptor(
+            c_m_n_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    template <typename CGridDesc>
+    __host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
+        const CGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch)
+    {
+        return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc>(
+            c_m_n_grid_desc, 8, KBatch);
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+        constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+        return make_naive_tensor_descriptor_packed(
+            make_tuple(I1,
+                       Number<CShuffleMRepeatPerShuffle * MWave * MPerXDL>{},
+                       I1,
+                       Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
+    }
+
+    // return block_id to C matrix tile idx (m0, n0, k_split) mapping
+    __host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap()
+    {
+        return BlockToCTileMap_3DGrid_KSplit<MPerBlock, NPerBlock>();
+    }
+
+    using CGridDesc_M_N         = remove_cvref_t<decltype(MakeCGridDescriptor_M_N(1, 1, 1))>;
+    using DefaultBlock2CTileMap = remove_cvref_t<decltype(MakeDefaultBlock2CTileMap())>;
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+              typename Block2CTileMap>
+    __device__ static void Run(const Argument& karg,
+                               void* __restrict__ p_shared_block,
+                               const Block2CTileMap& block_2_ctile_map)
+    {
+        const FloatA* p_a_grid           = karg.p_a_grid;
+        const FloatB* p_b_grid           = karg.p_b_grid;
+        FloatC* p_c_grid                 = karg.p_c_grid;
+        const auto a_b_k0_m_k1_grid_desc = MakeAGridDescriptor_KBatch_K0_M_K1(
+            karg.M, karg.MPadded, karg.K, karg.StrideA, karg.k_batch, karg.K0, karg.KPadded);
+        const auto b_b_k0_n_k1_grid_desc = MakeBGridDescriptor_KBatch_K0_N_K1(
+            karg.K, karg.NPadded, karg.N, karg.StrideB, karg.k_batch, karg.K0, karg.KPadded);
+        const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(karg.M, karg.N, karg.StrideC);
+
+        const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n);
+        const AElementwiseOperation a_element_op = AElementwiseOperation{};
+        const BElementwiseOperation b_element_op = BElementwiseOperation{};
+        const CElementwiseOperation c_element_op = CElementwiseOperation{};
+
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_b_k0_n_k1_grid_desc.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [KBatch, M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        {
+            return;
+        }
+
+        const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
+        const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I2]);
+        const index_t k_batch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_n_id * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        constexpr auto a_b_k0_m_k1_block_desc = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<K0PerBlock>{} * Number<MPerBlock + 1>{} * K1,
+                               Number<MPerBlock + 1>{} * K1,
+                               K1,
+                               I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        constexpr auto b_b_k0_n_k1_block_desc = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<K0PerBlock>{} * Number<NPerBlock + 1>{} * K1,
+                               Number<NPerBlock + 1>{} * K1,
+                               K1,
+                               I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    max_lds_align);
+            }
+        }();
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, K0PerBlock, MPerBlock, K1>,
+                                                ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatA,
+                                                ComputeType,
+                                                decltype(a_b_k0_m_k1_grid_desc),
+                                                decltype(a_b_k0_m_k1_block_desc),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                a_b_k0_m_k1_grid_desc,
+                make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_b_k0_m_k1_block_desc,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, K0PerBlock, NPerBlock, K1>,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatB,
+                                                ComputeType,
+                                                decltype(b_b_k0_n_k1_grid_desc),
+                                                decltype(b_b_k0_n_k1_block_desc),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<0, 2, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                b_b_k0_n_k1_grid_desc,
+                make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_b_k0_n_k1_block_desc,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ComputeType,
+            FloatAcc,
+            decltype(a_k0_m_k1_block_desc),
+            decltype(b_k0_n_k1_block_desc),
+            MPerXDL,
+            NPerXDL,
+            MRepeat,
+            NRepeat,
+            K1,
+            LoopSched>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size =
+            math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+        ComputeType* p_a_block = static_cast<ComputeType*>(p_shared_block);
+        ComputeType* p_b_block = static_cast<ComputeType*>(p_shared_block) + a_block_space_size;
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
+
+        // gridwise GEMM pipeline
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_b_k0_m_k1_grid_desc.GetLength(I1) * a_b_k0_m_k1_grid_desc.GetLength(I3)) /
+            (K0PerBlock * K1));
+
+        const auto gridwise_gemm_pipeline = GridwiseGemmPipe{};
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_b_k0_m_k1_grid_desc,
+                                                               a_b_k0_m_k1_block_desc,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_b_k0_n_k1_grid_desc,
+                                                               b_b_k0_n_k1_block_desc,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // output: register to global memory
+        {
+            constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+            constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I0);
+            constexpr auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I1);
+            constexpr auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I2);
+            constexpr auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I3);
+            constexpr auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I4);
+            constexpr auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I5);
+            constexpr auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I6);
+            constexpr auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I7);
+
+            constexpr auto c_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatC*>(p_shared_block),
+                c_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0), // freeze mblock
+                    make_unmerge_transform(make_tuple(CShuffleMRepeatPerShuffle,
+                                                      M1,
+                                                      M2,
+                                                      M3,
+                                                      M4)), // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                    make_freeze_transform(I0),              // freeze nblock
+                    make_unmerge_transform(make_tuple(CShuffleNRepeatPerShuffle,
+                                                      N1,
+                                                      N2))), // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMRepeatPerShuffle,
+                                                            CShuffleNRepeatPerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // LDS to global
+            auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // index_t BlockSize,
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMRepeatPerShuffle * MWave * MPerXDL,
+                         1,
+                         CShuffleNRepeatPerShuffle * NWave * NPerXDL>, // BlockSliceLengths,
+                CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                FloatC,               // typename SrcData,
+                FloatC,               // typename DstData,
+                decltype(c_block_desc_mblock_mperblock_nblock_nperblock),
+                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                Sequence<0, 1, 2, 3>,                       // typename DimAccessOrder,
+                3,                                          // index_t VectorDim,
+                CBlockTransferScalarPerVector_NWaveNPerXDL, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun
+                {c_block_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(0, 0, 0, 0),
+                 c_grid_desc_mblock_mperblock_nblock_nperblock,
+                 make_multi_index(block_m_id, 0, block_n_id, 0),
+                 c_element_op};
+
+            constexpr auto mxdlperwave_forward_step =
+                make_multi_index(0, CShuffleMRepeatPerShuffle * MWave * MPerXDL, 0, 0);
+            constexpr auto nxdlperwave_forward_step =
+                make_multi_index(0, 0, 0, CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+            constexpr auto nxdlperwave_backward_step =
+                make_multi_index(0, 0, 0, -CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+
+            static_for<0, MRepeat, CShuffleMRepeatPerShuffle>{}([&](auto mxdlperwave_iter) {
+                constexpr auto mxdlperwave = mxdlperwave_iter;
+
+                static_for<0, NRepeat, CShuffleNRepeatPerShuffle>{}([&](auto nxdlperwave_iter) {
+                    constexpr bool nxdlperwave_forward_sweep =
+                        (mxdlperwave % (2 * CShuffleMRepeatPerShuffle) == 0);
+
+                    constexpr index_t nxdlperwave_value =
+                        nxdlperwave_forward_sweep
+                            ? nxdlperwave_iter
+                            : (NRepeat - nxdlperwave_iter - CShuffleNRepeatPerShuffle);
+
+                    constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
+
+                    // make sure it's safe to do ds_write
+                    block_sync_lds();
+
+                    // VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(
+                        c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
+                        make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
+                        c_thread_buf,
+                        c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        c_block_buf);
+
+                    // make sure it's safe to do ds_read
+                    block_sync_lds();
+
+                    // LDS to global
+                    c_block_copy_lds_to_global.Run(c_block_desc_mblock_mperblock_nblock_nperblock,
+                                                   c_block_buf,
+                                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                                   c_grid_buf);
+
+                    // move on nxdlperwave dimension
+                    if constexpr(nxdlperwave_forward_sweep &&
+                                 (nxdlperwave < NRepeat - CShuffleNRepeatPerShuffle))
+                    {
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mperblock_nblock_nperblock,
+                            nxdlperwave_forward_step);
+                    }
+                    else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
+                    {
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mperblock_nblock_nperblock,
+                            nxdlperwave_backward_step);
+                    }
+                });
+
+                // move on mxdlperwave dimension
+                if constexpr(mxdlperwave < MRepeat - CShuffleMRepeatPerShuffle)
+                {
+                    c_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mperblock_nblock_nperblock, mxdlperwave_forward_step);
+                }
+            });
+        }
+    }
+
+    static std::string GetTypeString()
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "GemmXdlSplitKCShuffle_"
+            << getGemmSpecializationString(GemmSpec) << "_"
+            << std::string(ALayout::name)[0]
+            << std::string(BLayout::name)[0]
+            << std::string(CLayout::name)[0]
+            << "_"
+            << "B" << BlockSize << "_"
+            << "Vec" << ABlockTransferSrcScalarPerVector << "x"
+            << BBlockTransferSrcScalarPerVector << "x"
+            << CBlockTransferScalarPerVector_NWaveNPerXDL << "_"
+            << MPerBlock << "x"
+            << NPerBlock << "x"
+            << K0PerBlock << "x"
+            << K1 ;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
new file mode 100644
index 0000000000..8d7bd9a8d1
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
@@ -0,0 +1,722 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainK0BlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v3r1(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
+            const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainK0BlockLoop>(
+        p_a_grid,
+        p_b_grid,
+        p_c_grid,
+        p_shared,
+        a_grid_desc_ak0_m_ak1,
+        b_grid_desc_bk0_n_bk1,
+        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        a_element_op,
+        b_element_op,
+        c_element_op,
+        block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = a_grid_desc_ak0_m_ak1;
+    ignore = b_grid_desc_bk0_n_bk1;
+    ignore = c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <
+    index_t BlockSize,
+    typename FloatAB,
+    typename FloatAcc,
+    typename FloatCShuffle,
+    typename FloatC,
+    InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+    typename AGridDesc_AK0_M_AK1,
+    typename BGridDesc_BK0_N_BK1,
+    typename CGridDesc_M_N,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t KPerBlock,
+    index_t AK1Value,
+    index_t BK1Value,
+    index_t MPerXdl,
+    index_t NPerXdl,
+    index_t MXdlPerWave,
+    index_t NXdlPerWave,
+    typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    index_t ABlockTransferSrcVectorDim,
+    index_t ABlockTransferSrcScalarPerVector,
+    index_t ABlockTransferDstScalarPerVector_K1,
+    bool AThreadTransferSrcResetCoordinateAfterRun,
+    bool ABlockLdsExtraM,
+    typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    index_t BBlockTransferSrcVectorDim,
+    index_t BBlockTransferSrcScalarPerVector,
+    index_t BBlockTransferDstScalarPerVector_K1,
+    bool BThreadTransferSrcResetCoordinateAfterRun,
+    bool BBlockLdsExtraN,
+    index_t CShuffleMXdlPerWavePerShuffle,
+    index_t CShuffleNXdlPerWavePerShuffle,
+    typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+    index_t CBlockTransferScalarPerVector_NWaveNPerXdl,
+    index_t NumGemmKPrefetchStage = 1,
+    PipelineVersion PipelineVer   = PipelineVersion::v1>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
+    static constexpr auto AK1 = Number<AK1Value>{};
+    static constexpr auto BK1 = Number<BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        constexpr auto max_lds_align = AK1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(AK0, Number<MPerBlock>{}, AK1),
+                    make_tuple(Number<MPerBlock + 1>{} * AK1, AK1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(AK0, Number<MPerBlock>{}, AK1), max_lds_align);
+            }
+        }();
+
+        return a_block_desc_ak0_m_ak1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        constexpr auto max_lds_align = BK1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(BK0, Number<NPerBlock>{}, BK1),
+                    make_tuple(Number<NPerBlock + 1>{} * BK1, BK1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(BK0, Number<NPerBlock>{}, BK1), max_lds_align);
+            }
+        }();
+
+        return b_block_desc_bk0_n_bk1;
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto
+            c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1,
+                               Number<CShuffleMXdlPerWavePerShuffle>{},
+                               Number<MWave * MPerXdl>{},
+                               I1,
+                               Number<CShuffleNXdlPerWavePerShuffle>{},
+                               Number<NWave * NPerXdl>{}));
+
+        return c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_ak0_m_ak1.GetElementSpaceSize(), AK1);
+
+        constexpr auto b_block_space_size_aligned =
+            math::integer_least_multiple(b_block_desc_bk0_n_bk1.GetElementSpaceSize(), BK1);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+            GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl();
+
+        constexpr auto c_block_size =
+            c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatCShuffle));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        // static_assert(is_known_at_compile_time<remove_cv_t<decltype(AK1)>>::value &&
+        //               is_known_at_compile_time<remove_cv_t<decltype(BK1)>>::value,
+        //               "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
+        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
+        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        const auto c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+            transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_unmerge_transform(make_tuple(
+                               MBlock, Number<MXdlPerWave>{}, Number<MWave * MPerXdl>{})),
+                           make_unmerge_transform(make_tuple(
+                               NBlock, Number<NXdlPerWave>{}, Number<NWave * NPerXdl>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
+
+        return c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap(
+        const CGridDesc_M_N& c_grid_desc_m_n, index_t /* M01 */, index_t /* N01 */)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+    using CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                CGridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>;
+
+    template <bool HasMainK0BlockLoop, typename Block2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        void* __restrict__ p_shared,
+        const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+        const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+        const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
+            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid,
+            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(
+                   c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                       .GetLength(I0),
+                   c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                       .GetLength(I3))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t k_pack = math::max(
+            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_block_desc_ak0_m_ak1),
+                                                                decltype(b_block_desc_bk0_n_bk1),
+                                                                MPerXdl,
+                                                                NPerXdl,
+                                                                MXdlPerWave,
+                                                                NXdlPerWave,
+                                                                k_pack>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
+            KPerBlock);
+
+        GridwiseGemmPipe::template Run<HasMainK0BlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                           a_block_desc_ak0_m_ak1,
+                                                           a_blockwise_copy,
+                                                           a_grid_buf,
+                                                           a_block_buf,
+                                                           a_block_slice_copy_step,
+                                                           b_grid_desc_bk0_n_bk1,
+                                                           b_block_desc_bk0_n_bk1,
+                                                           b_blockwise_copy,
+                                                           b_grid_buf,
+                                                           b_block_buf,
+                                                           b_block_slice_copy_step,
+                                                           blockwise_gemm,
+                                                           c_thread_buf,
+                                                           num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+                GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatCShuffle*>(p_shared),
+                c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                    .GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                make_tuple(
+                    make_freeze_transform(I0), // freeze mblock
+                    make_pass_through_transform(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}), // M0 (MXdlPerWave) per shuffle
+                    make_unmerge_transform(
+                        make_tuple(M1, M2, M3, M4)), // M1 = MWave, M2 * M3 * M4 = MPerXdl
+                    make_freeze_transform(I0),       // freeze nblock
+                    make_pass_through_transform(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}), // N0 (NXdlPerWave) per shuffle
+                    make_unmerge_transform(
+                        make_tuple(N1, N2))), // M1 = MWave, M2 * M3 * M4 = MPerXdl
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<>{},
+                           Sequence<0>{},
+                           Sequence<2, 4, 5, 6>{},
+                           Sequence<>{},
+                           Sequence<1>{},
+                           Sequence<3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatCShuffle,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // LDS to global
+            auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle,
+                         MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle,
+                         NWave * NPerXdl>, // BlockSliceLengths,
+                CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+                Sequence<0, 1, 2, 3, 4, 5>, // typename ThreadClusterArrangeOrder,
+                FloatCShuffle,              // typename SrcData,
+                FloatC,                     // typename DstData,
+                decltype(c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                Sequence<0, 1, 2, 3, 4, 5>,                 // typename DimAccessOrder,
+                5,                                          // index_t VectorDim,
+                CBlockTransferScalarPerVector_NWaveNPerXdl, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(0, 0, 0, 0, 0, 0),
+                 c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(block_work_idx[I0], 0, 0, block_work_idx[I1], 0, 0),
+                 c_element_op};
+
+            constexpr auto mxdlperwave_forward_step =
+                make_multi_index(0, CShuffleMXdlPerWavePerShuffle, 0, 0, 0, 0);
+            constexpr auto nxdlperwave_forward_step =
+                make_multi_index(0, 0, 0, 0, CShuffleNXdlPerWavePerShuffle, 0);
+            constexpr auto nxdlperwave_backward_step =
+                make_multi_index(0, 0, 0, 0, -CShuffleNXdlPerWavePerShuffle, 0);
+
+            static_for<0, MXdlPerWave, CShuffleMXdlPerWavePerShuffle>{}([&](auto mxdlperwave_iter) {
+                constexpr auto mxdlperwave = mxdlperwave_iter;
+
+                static_for<0,
+                           NXdlPerWave,
+                           CShuffleNXdlPerWavePerShuffle>{}([&](auto nxdlperwave_iter) {
+                    constexpr bool nxdlperwave_forward_sweep =
+                        (mxdlperwave % (2 * CShuffleMXdlPerWavePerShuffle) == 0);
+
+                    constexpr index_t nxdlperwave_value =
+                        nxdlperwave_forward_sweep
+                            ? nxdlperwave_iter
+                            : (NXdlPerWave - nxdlperwave_iter - CShuffleNXdlPerWavePerShuffle);
+
+                    constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
+
+                    // make sure it's safe to do ds_write
+                    block_sync_lds();
+
+                    // VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(
+                        c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
+                        c_thread_buf,
+                        c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        c_shuffle_block_buf);
+
+                    // make sure it's safe to do ds_read
+                    block_sync_lds();
+
+                    // LDS to global
+                    c_block_copy_lds_to_global.Run(
+                        c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c_shuffle_block_buf,
+                        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c_grid_buf);
+
+                    // move on nxdlperwave dimension
+                    if constexpr(nxdlperwave_forward_sweep &&
+                                 (nxdlperwave < NXdlPerWave - CShuffleNXdlPerWavePerShuffle))
+                    {
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_forward_step);
+                    }
+                    else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
+                    {
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_backward_step);
+                    }
+                });
+
+                // move on mxdlperwave dimension
+                if constexpr(mxdlperwave < MXdlPerWave - CShuffleMXdlPerWavePerShuffle)
+                {
+                    c_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        mxdlperwave_forward_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
new file mode 100644
index 0000000000..79202cb5cf
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
@@ -0,0 +1,760 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+          typename C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v3r2(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const FloatC* __restrict__ p_c0_grid,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+            const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+            const C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        p_a_grid,
+        p_b_grid,
+        p_c_grid,
+        p_c0_grid,
+        p_shared,
+        a_grid_desc_k0_m_k1,
+        b_grid_desc_k0_n_k1,
+        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        a_element_op,
+        b_element_op,
+        c_element_op,
+        block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_c0_grid;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    ignore = c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <
+    index_t BlockSize,
+    typename FloatAB,
+    typename FloatAcc,
+    typename FloatC,
+    InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+    typename AGridDesc_K0_M_K1,
+    typename BGridDesc_K0_N_K1,
+    typename CGridDesc_M_N,
+    typename C0GridDesc_M_N,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t K0PerBlock,
+    index_t MPerXdl,
+    index_t NPerXdl,
+    index_t K1Value,
+    index_t MXdlPerWave,
+    index_t NXdlPerWave,
+    typename ABlockTransferThreadClusterLengths_K0_M_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    index_t ABlockTransferSrcVectorDim,
+    index_t ABlockTransferSrcScalarPerVector,
+    index_t ABlockTransferDstScalarPerVector_K1,
+    bool AThreadTransferSrcResetCoordinateAfterRun,
+    bool ABlockLdsExtraM,
+    typename BBlockTransferThreadClusterLengths_K0_N_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    index_t BBlockTransferSrcVectorDim,
+    index_t BBlockTransferSrcScalarPerVector,
+    index_t BBlockTransferDstScalarPerVector_K1,
+    bool BThreadTransferSrcResetCoordinateAfterRun,
+    bool BBlockLdsExtraN,
+    index_t CShuffleMXdlPerWavePerShuffle,
+    index_t CShuffleNXdlPerWavePerShuffle,
+    typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+    index_t CBlockTransferScalarPerVector_NWaveNPerXdl,
+    index_t NumGemmKPrefetchStage = 1,
+    PipelineVersion PipelineVer   = PipelineVersion::v1>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return a_block_desc_k0_m_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return b_block_desc_k0_n_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto
+            c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1,
+                               Number<CShuffleMXdlPerWavePerShuffle>{},
+                               Number<MWave * MPerXdl>{},
+                               I1,
+                               Number<CShuffleNXdlPerWavePerShuffle>{},
+                               Number<NWave * NPerXdl>{}));
+
+        return c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        constexpr auto max_lds_align = K1;
+
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned =
+            math::integer_least_multiple(b_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+            GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl();
+
+        constexpr auto c_block_size =
+            c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatC));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+             K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+             K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K0 / K0PerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / (K0PerBlock * K1);
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    template <typename CGridDesc_M_N_>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        const CGridDesc_M_N_& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        const auto c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+            transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_unmerge_transform(make_tuple(
+                               MBlock, Number<MXdlPerWave>{}, Number<MWave * MPerXdl>{})),
+                           make_unmerge_transform(make_tuple(
+                               NBlock, Number<NXdlPerWave>{}, Number<NWave * NPerXdl>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
+
+        return c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap(
+        const CGridDesc_M_N& c_grid_desc_m_n, index_t /* M01 */, index_t /* N01 */)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+
+    using CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                CGridDesc_M_N{}))>;
+
+    using C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                C0GridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>;
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        const FloatC* __restrict__ p_c0_grid,
+        void* __restrict__ p_shared,
+        const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+        const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+        const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
+            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        const C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
+            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid,
+            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize());
+        auto c0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c0_grid,
+            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize());
+
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(
+                   c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                       .GetLength(I0),
+                   c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                       .GetLength(I3))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<K0PerBlock, MPerBlock, K1>,
+                                                ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_k0_m_k1),
+                                                decltype(a_block_desc_k0_m_k1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_k0_m_k1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_k0_m_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<K0PerBlock, NPerBlock, K1>,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_k0_n_k1),
+                                                decltype(b_block_desc_k0_n_k1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_k0_n_k1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_k0_n_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_block_desc_k0_m_k1),
+                                                                decltype(b_block_desc_k0_n_k1),
+                                                                MPerXdl,
+                                                                NPerXdl,
+                                                                MXdlPerWave,
+                                                                NXdlPerWave,
+                                                                K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_k0_n_k1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+
+        // gridwise GEMM pipeline
+        const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
+
+        GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc_k0_m_k1,
+                                                          a_block_desc_k0_m_k1,
+                                                          a_blockwise_copy,
+                                                          a_grid_buf,
+                                                          a_block_buf,
+                                                          a_block_slice_copy_step,
+                                                          b_grid_desc_k0_n_k1,
+                                                          b_block_desc_k0_n_k1,
+                                                          b_blockwise_copy,
+                                                          b_grid_buf,
+                                                          b_block_buf,
+                                                          b_block_slice_copy_step,
+                                                          blockwise_gemm,
+                                                          c_thread_buf,
+                                                          K0BlockMainLoop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+                GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl();
+
+            auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatC*>(p_shared),
+                c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                    .GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                make_tuple(
+                    make_freeze_transform(I0), // freeze mblock
+                    make_pass_through_transform(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}), // M0 (MXdlPerWave) per shuffle
+                    make_unmerge_transform(
+                        make_tuple(M1, M2, M3, M4)), // M1 = MWave, M2 * M3 * M4 = MPerXdl
+                    make_freeze_transform(I0),       // freeze nblock
+                    make_pass_through_transform(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}), // N0 (NXdlPerWave) per shuffle
+                    make_unmerge_transform(
+                        make_tuple(N1, N2))), // M1 = MWave, M2 * M3 * M4 = MPerXdl
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<>{},
+                           Sequence<0>{},
+                           Sequence<2, 4, 5, 6>{},
+                           Sequence<>{},
+                           Sequence<1>{},
+                           Sequence<3, 7>{})
+
+            );
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r2<
+                ThisThreadBlock,            // index_t BlockSize,
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle,
+                         MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle,
+                         NWave * NPerXdl>, // BlockSliceLengths,
+                CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+                Sequence<0, 1, 2, 3, 4, 5>, // typename ThreadClusterArrangeOrder,
+                FloatC,                     // typename Src0Data,
+                FloatC,                     // typename Src1Data,
+                FloatC,                     // typename DstData,
+                decltype(c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                Sequence<0, 1, 2, 3, 4, 5>,                 // typename DimAccessOrder,
+                5,                                          // index_t VectorDim,
+                CBlockTransferScalarPerVector_NWaveNPerXdl, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrc0ResetCoordinateAfterRun,
+                false, // bool ThreadTransferSrc1ResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(0, 0, 0, 0, 0, 0),
+                 c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(block_work_idx[I0], 0, 0, block_work_idx[I1], 0, 0),
+                 c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(block_work_idx[I0], 0, 0, block_work_idx[I1], 0, 0),
+                 c_element_op};
+
+            constexpr auto mxdlperwave_forward_step =
+                make_multi_index(0, CShuffleMXdlPerWavePerShuffle, 0, 0, 0, 0);
+            constexpr auto nxdlperwave_forward_step =
+                make_multi_index(0, 0, 0, 0, CShuffleNXdlPerWavePerShuffle, 0);
+            constexpr auto nxdlperwave_backward_step =
+                make_multi_index(0, 0, 0, 0, -CShuffleNXdlPerWavePerShuffle, 0);
+
+            static_for<0, MXdlPerWave, CShuffleMXdlPerWavePerShuffle>{}([&](auto mxdlperwave_iter) {
+                constexpr auto mxdlperwave = mxdlperwave_iter;
+
+                static_for<0,
+                           NXdlPerWave,
+                           CShuffleNXdlPerWavePerShuffle>{}([&](auto nxdlperwave_iter) {
+                    constexpr bool nxdlperwave_forward_sweep =
+                        (mxdlperwave % (2 * CShuffleMXdlPerWavePerShuffle) == 0);
+
+                    constexpr index_t nxdlperwave_value =
+                        nxdlperwave_forward_sweep
+                            ? nxdlperwave_iter
+                            : (NXdlPerWave - nxdlperwave_iter - CShuffleNXdlPerWavePerShuffle);
+
+                    constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
+
+                    // make sure it's safe to do ds_write
+                    block_sync_lds();
+
+                    // VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(
+                        c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
+                        c_thread_buf,
+                        c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        c_block_buf);
+
+                    // make sure it's safe to do ds_read
+                    block_sync_lds();
+
+                    // LDS to global
+                    c_block_copy_lds_to_global.Run(
+                        c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c_block_buf,
+                        c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c0_grid_buf,
+                        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c_grid_buf);
+
+                    // move on nxdlperwave dimension
+                    if constexpr(nxdlperwave_forward_sweep &&
+                                 (nxdlperwave < NXdlPerWave - CShuffleNXdlPerWavePerShuffle))
+                    {
+                        c_block_copy_lds_to_global.MoveSrc1SliceWindow(
+                            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_forward_step);
+
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_forward_step);
+                    }
+                    else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
+                    {
+                        c_block_copy_lds_to_global.MoveSrc1SliceWindow(
+                            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_backward_step);
+
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_backward_step);
+                    }
+                });
+
+                // move on mxdlperwave dimension
+                if constexpr(mxdlperwave < MXdlPerWave - CShuffleMXdlPerWavePerShuffle)
+                {
+                    c_block_copy_lds_to_global.MoveSrc1SliceWindow(
+                        c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        mxdlperwave_forward_step);
+
+                    c_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        mxdlperwave_forward_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
new file mode 100644
index 0000000000..0d461b4fbf
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
@@ -0,0 +1,798 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm,
+          typename FloatAB,
+          typename FloatC,
+          typename AGridDesc_K0_M_K1,
+          typename BGridDesc_K0_N_K1,
+          typename CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+          typename C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+          typename C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename Block2CTileMap,
+          bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_gemm_xdlops_v3r3(
+            const FloatAB* __restrict__ p_a_grid,
+            const FloatAB* __restrict__ p_b_grid,
+            FloatC* __restrict__ p_c_grid,
+            const FloatC* __restrict__ p_c0_grid,
+            const FloatC* __restrict__ p_c1_grid,
+            const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
+            const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
+            const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+            const C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+            const C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+                c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CElementwiseOperation c_element_op,
+            const Block2CTileMap block_2_ctile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(
+        p_a_grid,
+        p_b_grid,
+        p_c_grid,
+        p_c0_grid,
+        p_c1_grid,
+        p_shared,
+        a_grid_desc_k0_m_k1,
+        b_grid_desc_k0_n_k1,
+        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        a_element_op,
+        b_element_op,
+        c_element_op,
+        block_2_ctile_map);
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = p_c0_grid;
+    ignore = p_c1_grid;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    ignore = c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    ignore = c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = c_element_op;
+    ignore = block_2_ctile_map;
+#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+}
+
+template <
+    index_t BlockSize,
+    typename FloatAB,
+    typename FloatAcc,
+    typename FloatC,
+    InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+    typename AGridDesc_K0_M_K1,
+    typename BGridDesc_K0_N_K1,
+    typename CGridDesc_M_N,
+    typename C0GridDesc_M_N,
+    typename C1GridDesc_M_N,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    index_t MPerBlock,
+    index_t NPerBlock,
+    index_t K0PerBlock,
+    index_t MPerXdl,
+    index_t NPerXdl,
+    index_t K1Value,
+    index_t MXdlPerWave,
+    index_t NXdlPerWave,
+    typename ABlockTransferThreadClusterLengths_K0_M_K1,
+    typename ABlockTransferThreadClusterArrangeOrder,
+    typename ABlockTransferSrcAccessOrder,
+    index_t ABlockTransferSrcVectorDim,
+    index_t ABlockTransferSrcScalarPerVector,
+    index_t ABlockTransferDstScalarPerVector_K1,
+    bool AThreadTransferSrcResetCoordinateAfterRun,
+    bool ABlockLdsExtraM,
+    typename BBlockTransferThreadClusterLengths_K0_N_K1,
+    typename BBlockTransferThreadClusterArrangeOrder,
+    typename BBlockTransferSrcAccessOrder,
+    index_t BBlockTransferSrcVectorDim,
+    index_t BBlockTransferSrcScalarPerVector,
+    index_t BBlockTransferDstScalarPerVector_K1,
+    bool BThreadTransferSrcResetCoordinateAfterRun,
+    bool BBlockLdsExtraN,
+    index_t CShuffleMXdlPerWavePerShuffle,
+    index_t CShuffleNXdlPerWavePerShuffle,
+    typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+    index_t CBlockTransferScalarPerVector_NWaveNPerXdl,
+    index_t NumGemmKPrefetchStage = 1,
+    PipelineVersion PipelineVer   = PipelineVersion::v1>
+struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+    static constexpr auto I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto K1 = Number<K1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return a_block_desc_k0_m_k1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
+    {
+        constexpr auto max_lds_align = K1;
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+            }
+        }();
+
+        return b_block_desc_k0_n_k1;
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto
+            c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1,
+                               Number<CShuffleMXdlPerWavePerShuffle>{},
+                               Number<MWave * MPerXdl>{},
+                               I1,
+                               Number<CShuffleNXdlPerWavePerShuffle>{},
+                               Number<NWave * NPerXdl>{}));
+
+        return c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        constexpr auto max_lds_align = K1;
+
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned =
+            math::integer_least_multiple(b_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+            GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl();
+
+        constexpr auto c_block_size =
+            c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(FloatAB),
+                         c_block_size * sizeof(FloatC));
+    }
+
+    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
+    template <typename Block2CTileMap>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+                  const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+                  const CGridDesc_M_N& c_grid_desc_m_n,
+                  const Block2CTileMap& block_2_ctile_map)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
+                      "wrong! K1 need to be known at compile-time");
+
+        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
+                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
+                      "Invalid tuning param!");
+
+        const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
+        const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
+             K0 == b_grid_desc_k0_n_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
+             K1 == b_grid_desc_k0_n_k1.GetLength(I2)))
+            return false;
+
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
+            return false;
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K0 / K0PerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / (K0PerBlock * K1);
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    template <typename CGridDesc_M_N_>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        const CGridDesc_M_N_& c_grid_desc_m_n)
+    {
+        const auto M = c_grid_desc_m_n.GetLength(I0);
+        const auto N = c_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        const auto c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+            transform_tensor_descriptor(
+                c_grid_desc_m_n,
+                make_tuple(make_unmerge_transform(make_tuple(
+                               MBlock, Number<MXdlPerWave>{}, Number<MWave * MPerXdl>{})),
+                           make_unmerge_transform(make_tuple(
+                               NBlock, Number<NXdlPerWave>{}, Number<NWave * NPerXdl>{}))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
+
+        return c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
+    }
+
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap(
+        const CGridDesc_M_N& c_grid_desc_m_n, index_t /* M01 */, index_t /* N01 */)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
+            c_grid_desc_m_n);
+    }
+    using CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                CGridDesc_M_N{}))>;
+
+    using C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                C0GridDesc_M_N{}))>;
+
+    using C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+                C1GridDesc_M_N{}))>;
+
+    using DefaultBlock2CTileMap =
+        remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>;
+
+    template <bool HasMainKBlockLoop, typename Block2CTileMap>
+    __device__ static void
+    Run(const FloatAB* __restrict__ p_a_grid,
+        const FloatAB* __restrict__ p_b_grid,
+        FloatC* __restrict__ p_c_grid,
+        const FloatC* __restrict__ p_c0_grid,
+        const FloatC* __restrict__ p_c1_grid,
+        void* __restrict__ p_shared,
+        const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
+        const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
+        const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
+            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        const C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
+            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        const C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
+            c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+        const AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CElementwiseOperation& c_element_op,
+        const Block2CTileMap& block_2_ctile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid,
+            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize());
+        auto c0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c0_grid,
+            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize());
+        auto c1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c1_grid,
+            c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                .GetElementSpaceSize());
+
+        const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(
+                   c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                       .GetLength(I0),
+                   c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                       .GetLength(I3))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = K1;
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<K0PerBlock, MPerBlock, K1>,
+                                                ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(a_grid_desc_k0_m_k1),
+                                                decltype(a_block_desc_k0_m_k1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                a_grid_desc_k0_m_k1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_k0_m_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<K0PerBlock, NPerBlock, K1>,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                FloatAB,
+                                                FloatAB,
+                                                decltype(b_grid_desc_k0_n_k1),
+                                                decltype(b_block_desc_k0_n_k1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true>(
+                b_grid_desc_k0_n_k1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_k0_n_k1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+
+        auto blockwise_gemm =
+            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
+                                                                FloatAB,
+                                                                FloatAcc,
+                                                                decltype(a_block_desc_k0_m_k1),
+                                                                decltype(b_block_desc_k0_n_k1),
+                                                                MPerXdl,
+                                                                NPerXdl,
+                                                                MXdlPerWave,
+                                                                NXdlPerWave,
+                                                                K1>{};
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned =
+            math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_k0_n_k1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
+
+        // gridwise GEMM pipeline
+        const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
+
+        GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc_k0_m_k1,
+                                                          a_block_desc_k0_m_k1,
+                                                          a_blockwise_copy,
+                                                          a_grid_buf,
+                                                          a_block_buf,
+                                                          a_block_slice_copy_step,
+                                                          b_grid_desc_k0_n_k1,
+                                                          b_block_desc_k0_n_k1,
+                                                          b_blockwise_copy,
+                                                          b_grid_buf,
+                                                          b_block_buf,
+                                                          b_block_slice_copy_step,
+                                                          blockwise_gemm,
+                                                          c_thread_buf,
+                                                          K0BlockMainLoop);
+
+        // shuffle C and write out
+        {
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
+                GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl();
+
+            auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<FloatC*>(p_shared),
+                c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
+                    .GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                make_tuple(make_freeze_transform(I0), // freeze mblock
+                           make_pass_through_transform(
+                               Number<CShuffleMXdlPerWavePerShuffle>{}), // M0 (MXdlPerWave) per
+                                                                         // shuffle
+                           make_unmerge_transform(
+                               make_tuple(M1, M2, M3, M4)), // M1 = MWave, M2 * M3 * M4 = MPerXdl
+                           make_freeze_transform(I0),       // freeze nblock
+                           make_pass_through_transform(
+                               Number<CShuffleNXdlPerWavePerShuffle>{}), // N0 (NXdlPerWave) per
+                                                                         // shuffle
+                           make_unmerge_transform(
+                               make_tuple(N1, N2))), // M1 = MWave, M2 * M3 * M4 = MPerXdl
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{}),
+                make_tuple(Sequence<>{},
+                           Sequence<0>{},
+                           Sequence<2, 4, 5, 6>{},
+                           Sequence<>{},
+                           Sequence<1>{},
+                           Sequence<3, 7>{})
+
+            );
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
+                                                   FloatC,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r3<
+                ThisThreadBlock,            // ThreadGroup
+                CElementwiseOperation,      // ElementwiseOperation,
+                CGlobalMemoryDataOperation, // DstInMemOp,
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle,
+                         MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle,
+                         NWave * NPerXdl>, // BlockSliceLengths,
+                CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
+                Sequence<0, 1, 2, 3, 4, 5>, // typename ThreadClusterArrangeOrder,
+                FloatC,                     // typename Src0Data,
+                FloatC,                     // typename Src1Data,
+                FloatC,                     // typename Src2Data,
+                FloatC,                     // typename DstData,
+                decltype(c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                Sequence<0, 1, 2, 3, 4, 5>,                 // typename DimAccessOrder,
+                5,                                          // index_t VectorDim,
+                CBlockTransferScalarPerVector_NWaveNPerXdl, // index_t ScalarPerVector,
+                true,  // bool ThreadTransferSrc0ResetCoordinateAfterRun,
+                false, // bool ThreadTransferSrc1ResetCoordinateAfterRun,
+                false, // bool ThreadTransferSrc2ResetCoordinateAfterRun,
+                false> // bool ThreadTransferDstResetCoordinateAfterRun>
+                {c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(0, 0, 0, 0, 0, 0),
+                 c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(block_work_idx[I0], 0, 0, block_work_idx[I1], 0, 0),
+                 c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(block_work_idx[I0], 0, 0, block_work_idx[I1], 0, 0),
+                 c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                 make_multi_index(block_work_idx[I0], 0, 0, block_work_idx[I1], 0, 0),
+                 c_element_op};
+
+            constexpr auto mxdlperwave_forward_step =
+                make_multi_index(0, CShuffleMXdlPerWavePerShuffle, 0, 0, 0, 0);
+            constexpr auto nxdlperwave_forward_step =
+                make_multi_index(0, 0, 0, 0, CShuffleNXdlPerWavePerShuffle, 0);
+            constexpr auto nxdlperwave_backward_step =
+                make_multi_index(0, 0, 0, 0, -CShuffleNXdlPerWavePerShuffle, 0);
+
+            static_for<0, MXdlPerWave, CShuffleMXdlPerWavePerShuffle>{}([&](auto mxdlperwave_iter) {
+                constexpr auto mxdlperwave = mxdlperwave_iter;
+
+                static_for<0,
+                           NXdlPerWave,
+                           CShuffleNXdlPerWavePerShuffle>{}([&](auto nxdlperwave_iter) {
+                    constexpr bool nxdlperwave_forward_sweep =
+                        (mxdlperwave % (2 * CShuffleMXdlPerWavePerShuffle) == 0);
+
+                    constexpr index_t nxdlperwave_value =
+                        nxdlperwave_forward_sweep
+                            ? nxdlperwave_iter
+                            : (NXdlPerWave - nxdlperwave_iter - CShuffleNXdlPerWavePerShuffle);
+
+                    constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
+
+                    // make sure it's safe to do ds_write
+                    block_sync_lds();
+
+                    // VGPR to LDS
+                    c_thread_copy_vgpr_to_lds.Run(
+                        c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
+                        c_thread_buf,
+                        c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                        c_block_buf);
+
+                    // make sure it's safe to do ds_read
+                    block_sync_lds();
+
+                    // LDS to global
+                    c_block_copy_lds_to_global.Run(
+                        c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c_block_buf,
+                        c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c0_grid_buf,
+                        c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c1_grid_buf,
+                        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        c_grid_buf);
+
+                    // move on nxdlperwave dimension
+                    if constexpr(nxdlperwave_forward_sweep &&
+                                 (nxdlperwave < NXdlPerWave - CShuffleNXdlPerWavePerShuffle))
+                    {
+                        c_block_copy_lds_to_global.MoveSrc1SliceWindow(
+                            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_forward_step);
+
+                        c_block_copy_lds_to_global.MoveSrc2SliceWindow(
+                            c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_forward_step);
+
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_forward_step);
+                    }
+                    else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
+                    {
+                        c_block_copy_lds_to_global.MoveSrc1SliceWindow(
+                            c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_backward_step);
+
+                        c_block_copy_lds_to_global.MoveSrc2SliceWindow(
+                            c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_backward_step);
+
+                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                            c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                            nxdlperwave_backward_step);
+                    }
+                });
+
+                // move on mxdlperwave dimension
+                if constexpr(mxdlperwave < MXdlPerWave - CShuffleMXdlPerWavePerShuffle)
+                {
+                    c_block_copy_lds_to_global.MoveSrc1SliceWindow(
+                        c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        mxdlperwave_forward_step);
+
+                    c_block_copy_lds_to_global.MoveSrc2SliceWindow(
+                        c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        mxdlperwave_forward_step);
+
+                    c_block_copy_lds_to_global.MoveDstSliceWindow(
+                        c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
+                        mxdlperwave_forward_step);
+                }
+            });
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_image_to_column.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_image_to_column.hpp
new file mode 100644
index 0000000000..93625a324e
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_image_to_column.hpp
@@ -0,0 +1,97 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename InputGridDesc,
+          typename InputDataType,
+          typename OutputGridDesc,
+          typename OutputDataType,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t KPerBlock,
+          typename ThreadClusterLengths,
+          index_t ScalarPerVector,
+          typename Block2ETileMap>
+struct GridwiseImageToColumn
+{
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __device__ static void Run(const InputGridDesc& in_grid_desc,
+                               const InputDataType* __restrict__ p_in_global,
+                               const OutputGridDesc& out_grid_desc,
+                               OutputDataType* __restrict__ p_out_global,
+                               const Block2ETileMap& block_2_tile_map)
+    {
+        const auto block_work_idx =
+            block_2_tile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t k_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * KPerBlock);
+
+        // Global Memory
+        const auto in_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_global, in_grid_desc.GetElementSpaceSize());
+        auto out_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_global, out_grid_desc.GetElementSpaceSize());
+
+        auto copy_global_to_global = ThreadGroupTensorSliceTransfer_v7<
+            ThisThreadBlock,
+            Tuple<InputDataType>,
+            Tuple<OutputDataType>,
+            decltype(tie(in_grid_desc)),
+            decltype(tie(out_grid_desc)),
+            tensor_operation::element_wise::PassThrough,
+            Sequence<static_cast<index_t>(InMemoryDataOperationEnum::Set)>,
+            Sequence<MPerBlock, KPerBlock>,
+            ThreadClusterLengths,
+            Sequence<0, 1>,
+            Sequence<0, 1>,
+            I1,
+            ScalarPerVector,
+            Sequence<true>,
+            Sequence<true>>{
+            in_grid_desc,
+            make_tuple(make_multi_index(m_block_data_idx_on_grid, k_block_data_idx_on_grid)),
+            out_grid_desc,
+            make_tuple(make_multi_index(m_block_data_idx_on_grid, k_block_data_idx_on_grid)),
+            tensor_operation::element_wise::PassThrough{}};
+
+        copy_global_to_global.Run(
+            tie(in_grid_desc), tie(in_global_buf), tie(out_grid_desc), tie(out_global_buf));
+    }
+
+    __host__ static constexpr bool CheckValidity(const InputGridDesc& in_grid_desc,
+                                                 const OutputGridDesc& out_grid_desc)
+    {
+        if(in_grid_desc.GetLength(I0) % MPerBlock != 0 ||
+           in_grid_desc.GetLength(I1) % KPerBlock != 0)
+            return false;
+        if(out_grid_desc.GetLength(I0) % MPerBlock != 0 ||
+           out_grid_desc.GetLength(I1) % KPerBlock != 0)
+            return false;
+        return true;
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp
new file mode 100644
index 0000000000..61d0f9e0d5
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp
@@ -0,0 +1,339 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <numeric>
+#include <iterator>
+
+#include "ck/tensor_description/cluster_descriptor.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwisePermute,
+          typename InGridDesc,
+          typename OutGridDesc,
+          typename InDataType,
+          typename OutDataType,
+          typename ElementwiseOperation,
+          typename Block2TileMap>
+__global__ void kernel_nd_permute(const InGridDesc in_grid_desc,
+                                  const OutGridDesc out_grid_desc,
+                                  const InDataType* p_in_global,
+                                  OutDataType* p_out_global,
+                                  const ElementwiseOperation elementwise_op,
+                                  const Block2TileMap block_2_tile_map)
+{
+    __shared__ char p_shared[GridwisePermute::GetSharedMemoryNumberOfByte()];
+
+    GridwisePermute::Run(in_grid_desc,
+                         out_grid_desc,
+                         p_in_global,
+                         p_out_global,
+                         p_shared,
+                         elementwise_op,
+                         block_2_tile_map);
+}
+
+template <typename InGridDesc,
+          typename OutGridDesc,
+          typename InDataType,
+          typename OutDataType,
+          typename ElementwiseOperation,
+          index_t BlockSize,
+          index_t NPerBlock,
+          index_t HPerBlock,
+          index_t WPerBlock,
+          index_t InBlockLdsExtraW,
+          typename InBlockTransferThreadClusterLengths,
+          typename InBlockTransferThreadClusterArrangeOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector>
+struct GridwisePermute
+{
+    static_assert(InGridDesc::GetNumOfDimension() == OutGridDesc::GetNumOfDimension());
+    static_assert(3 <= InGridDesc::GetNumOfDimension());
+    static_assert((InGridDesc::GetNumOfDimension() - 2) <= SrcVectorDim &&
+                  SrcVectorDim < InGridDesc::GetNumOfDimension());
+    static_assert((OutGridDesc::GetNumOfDimension() - 2) <= DstVectorDim &&
+                  DstVectorDim < OutGridDesc::GetNumOfDimension());
+    static_assert(SrcVectorDim != DstVectorDim);
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    struct Block2TileMap
+    {
+        static constexpr index_t NumDim = InGridDesc::GetNumOfDimension();
+        static_assert(3 <= NumDim);
+
+        static constexpr auto I0 = Number<0>{};
+
+        Block2TileMap()                     = delete;
+        Block2TileMap(const Block2TileMap&) = default;
+        Block2TileMap(Block2TileMap&&)      = delete;
+
+        ~Block2TileMap() = default;
+
+        Block2TileMap& operator=(const Block2TileMap&) = delete;
+        Block2TileMap& operator=(Block2TileMap&&) = delete;
+
+        explicit Block2TileMap(const InGridDesc& desc) : desc_(desc) {}
+
+        __host__ constexpr index_t CalculateGridSize(const InGridDesc& desc) const
+        {
+            const auto N0 =
+                math::integer_divide_ceil(desc.GetLength(Number<NumDim - 3>{}), NPerBlock);
+            const auto H0 =
+                math::integer_divide_ceil(desc.GetLength(Number<NumDim - 2>{}), HPerBlock);
+            const auto W0 =
+                math::integer_divide_ceil(desc.GetLength(Number<NumDim - 1>{}), WPerBlock);
+
+            const index_t grid_size = N0 * H0 * W0;
+
+            return grid_size;
+        }
+
+        template <typename TopIdx>
+        __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
+        {
+            static_assert(TopIdx::Size() == 1);
+
+            auto block_1d_id = idx_top[I0];
+
+            const auto N0 =
+                math::integer_divide_ceil(desc_.GetLength(Number<NumDim - 3>{}), NPerBlock);
+            const auto H0 =
+                math::integer_divide_ceil(desc_.GetLength(Number<NumDim - 2>{}), HPerBlock);
+            const auto W0 =
+                math::integer_divide_ceil(desc_.GetLength(Number<NumDim - 1>{}), WPerBlock);
+
+            block_1d_id = block_1d_id % (N0 * H0 * W0);
+
+            index_t idx_N0 = block_1d_id / (H0 * W0);
+            index_t idx_H0 = (block_1d_id % (H0 * W0)) / W0;
+            index_t idx_W0 = block_1d_id % W0;
+
+            return make_tuple(idx_N0, idx_H0, idx_W0);
+        }
+
+        private:
+        const InGridDesc desc_;
+    };
+
+    using DefaultBlock2TileMap = Block2TileMap;
+
+    // use an [NPerBlock, HPerBlock, WPerBlock] tensor as element-copy relay
+    __host__ __device__ static constexpr auto GetInBlockDesc_NPerBlock_HPerBlock_WPerBlock()
+    {
+        return make_naive_tensor_descriptor(
+            make_tuple(Number<NPerBlock>{}, Number<HPerBlock>{}, Number<WPerBlock>{}),
+            make_tuple(Number<HPerBlock*(WPerBlock + InBlockLdsExtraW)>{},
+                       Number<WPerBlock + InBlockLdsExtraW>{},
+                       I1));
+    }
+
+    // for N-dimension descriptor, reserve its last 2 dimensions, then merge its leading dimensions
+    // into single one. finally, form a 3D descriptor: [d(0), d(1), ..., d(N - 2), d(N - 1)] ->
+    // [(d(0) x d(1) x ...), d(N - 2), d(N - 1)]
+    template <typename GridDesc>
+    __host__ __device__ static constexpr auto GetMergedDesc(const GridDesc& desc)
+    {
+        constexpr index_t NumDim = GridDesc::GetNumOfDimension();
+        static_assert(3 <= NumDim);
+
+        const auto merged_desc = transform_tensor_descriptor(
+            desc,
+            make_tuple(make_merge_transform(generate_tuple(
+                           [&](auto I) { return desc.GetLength(I); }, Number<NumDim - 2>{})),
+                       make_pass_through_transform(desc.GetLength(Number<NumDim - 2>{})),
+                       make_pass_through_transform(desc.GetLength(Number<NumDim - 1>{}))),
+            make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<NumDim - 2>{}),
+                       Sequence<NumDim - 2>{},
+                       Sequence<NumDim - 1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+        return merged_desc;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        constexpr auto in_block_desc_nperblock_hperblock_wperblock =
+            GetInBlockDesc_NPerBlock_HPerBlock_WPerBlock();
+
+        return in_block_desc_nperblock_hperblock_wperblock.GetElementSpaceSize() *
+               sizeof(InDataType);
+    }
+
+    __host__ __device__ static constexpr auto MakeDefaultBlock2TileMap(const InGridDesc& desc)
+    {
+        return DefaultBlock2TileMap{desc};
+    }
+
+    __host__ __device__ static constexpr bool CheckValidity(const InGridDesc& in_grid_desc,
+                                                            const OutGridDesc& out_grid_desc)
+    {
+        constexpr index_t NumDim = InGridDesc::GetNumOfDimension();
+
+        // check if we only swap last 2 dimensions
+        bool valid = true;
+        static_for<0, NumDim - 2, 1>{}([&](auto I) {
+            if(valid && in_grid_desc.GetLength(I) != out_grid_desc.GetLength(I))
+            {
+                valid = false;
+            }
+        });
+
+        return valid &&
+               (in_grid_desc.GetLength(Number<NumDim - 1>{}) ==
+                out_grid_desc.GetLength(Number<NumDim - 2>{})) &&
+               (in_grid_desc.GetLength(Number<NumDim - 2>{}) ==
+                out_grid_desc.GetLength(Number<NumDim - 1>{}));
+    }
+
+    template <typename Block2TileMap>
+    __device__ static void Run(const InGridDesc in_grid_desc,
+                               const OutGridDesc out_grid_desc,
+                               const InDataType* p_in_global,
+                               OutDataType* p_out_global,
+                               void* __restrict__ p_shared,
+                               const ElementwiseOperation elementwise_op,
+                               const Block2TileMap& block_2_tile_map)
+    {
+        auto in_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_global, in_grid_desc.GetElementSpaceSize());
+
+        auto out_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_global, out_grid_desc.GetElementSpaceSize());
+
+        // each workgroup handles an [NPerBlock, HPerBlock, WPerBLock] slice-transpose problem
+        const auto block_work_idx =
+            block_2_tile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * NPerBlock);
+
+        const index_t h_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * HPerBlock);
+
+        const index_t w_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * WPerBlock);
+
+        // create [NPerBlock, HPerBlock, WPerBLock] shaped LDS buffer
+        constexpr auto in_block_desc_nperblock_hperblock_wperblock =
+            GetInBlockDesc_NPerBlock_HPerBlock_WPerBlock();
+
+        auto in_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<InDataType*>(p_shared),
+            in_block_desc_nperblock_hperblock_wperblock.GetElementSpaceSize());
+
+        using BlockSliceLengths          = Sequence<NPerBlock, HPerBlock, WPerBlock>;
+        using InBlockTransferAccessOrder = Sequence<0, 1, 2>;
+
+        constexpr index_t SrcVectorDimAfterMerge =
+            SrcVectorDim - (InGridDesc::GetNumOfDimension() - 3);
+        constexpr index_t DstVectorDimAfterMerge = SrcVectorDimAfterMerge;
+
+        using ck::tensor_operation::element_wise::PassThrough;
+
+        // merge input descriptor into [(in_grid_desc.GetLength(0) x in_grid_desc.GetLength(1) x
+        // ...), in_grid_desc.GetLength(NumDim - 2), in_grid_desc.GetLength(NumDim - 1)]
+        const auto in_grid_desc_n_h_w = GetMergedDesc(in_grid_desc);
+
+        // a workgroup copies an [NPerBlock, HPerBlock, WPerBlock] slice from global memory to LDS
+        auto in_global_load = ThreadGroupTensorSliceTransfer_v4r1<
+            ThisThreadBlock,
+            ElementwiseOperation,
+            PassThrough,
+            InMemoryDataOperationEnum::Set,
+            BlockSliceLengths,
+            InBlockTransferThreadClusterLengths,
+            InBlockTransferThreadClusterArrangeOrder,
+            InDataType,
+            InDataType,
+            decltype(in_grid_desc_n_h_w),
+            decltype(in_block_desc_nperblock_hperblock_wperblock),
+            InBlockTransferAccessOrder,
+            InBlockTransferAccessOrder,
+            SrcVectorDimAfterMerge,
+            2,
+            SrcScalarPerVector,
+            1,
+            1,
+            1,
+            true,
+            true>(in_grid_desc_n_h_w,
+                  make_multi_index(
+                      n_block_data_idx_on_grid, h_block_data_idx_on_grid, w_block_data_idx_on_grid),
+                  PassThrough{},
+                  in_block_desc_nperblock_hperblock_wperblock,
+                  make_multi_index(0, 0, 0),
+                  PassThrough{});
+
+        // merge output descriptor into [(out_grid_desc.GetLength(0) x out_grid_desc.GetLength(1) x
+        // ...), out_grid_desc.GetLength(NumDim - 2), out_grid_desc.GetLength(NumDim - 1)]
+        const auto out_grid_desc_n_w_h = GetMergedDesc(out_grid_desc);
+
+        // create transposed view of output tensor
+        const auto out_grid_desc_n_h_w = transform_tensor_descriptor(
+            out_grid_desc_n_w_h,
+            make_tuple(make_pass_through_transform(out_grid_desc_n_w_h.GetLength(I0)),
+                       make_pass_through_transform(out_grid_desc_n_w_h.GetLength(I1)),
+                       make_pass_through_transform(out_grid_desc_n_w_h.GetLength(I2))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<2>{}, Sequence<1>{}));
+
+        // a workgroup copies an [NPerBlock, HPerBlock, WPerBlock] slice from LDS to global memory
+        auto out_global_store = ThreadGroupTensorSliceTransfer_v4r1<
+            ThisThreadBlock,
+            ElementwiseOperation,
+            PassThrough,
+            InMemoryDataOperationEnum::Set,
+            BlockSliceLengths,
+            InBlockTransferThreadClusterLengths,
+            InBlockTransferThreadClusterArrangeOrder,
+            InDataType,
+            OutDataType,
+            decltype(in_block_desc_nperblock_hperblock_wperblock),
+            decltype(out_grid_desc_n_h_w),
+            InBlockTransferAccessOrder,
+            InBlockTransferAccessOrder,
+            2,
+            DstVectorDimAfterMerge,
+            1,
+            DstScalarPerVector,
+            1,
+            1,
+            true,
+            true>(in_block_desc_nperblock_hperblock_wperblock,
+                  make_multi_index(0, 0, 0),
+                  PassThrough{},
+                  out_grid_desc_n_h_w,
+                  make_multi_index(
+                      n_block_data_idx_on_grid, h_block_data_idx_on_grid, w_block_data_idx_on_grid),
+                  elementwise_op);
+
+        in_global_load.Run(in_grid_desc_n_h_w,
+                           in_global_buf,
+                           in_block_desc_nperblock_hperblock_wperblock,
+                           in_block_buf,
+                           I0);
+
+        out_global_store.Run(in_block_desc_nperblock_hperblock_wperblock,
+                             in_block_buf,
+                             out_grid_desc_n_h_w,
+                             out_global_buf,
+                             I0);
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp
new file mode 100644
index 0000000000..c41eef8c45
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp
@@ -0,0 +1,155 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+
+namespace ck {
+
+template <typename GridwisePutElementwise1dFunctor,
+          typename InGrid1dDesc,
+          typename InDataType,
+          typename IndexDataType,
+          typename OutDataType,
+          typename ElementwiseOperation>
+__global__ void kernel_put_element_1d(const InGrid1dDesc in_grid_1d_desc,
+                                      const InDataType* __restrict__ p_in_global,
+                                      const IndexDataType* __restrict__ p_indices_global,
+                                      OutDataType* __restrict__ p_out_global,
+                                      const ElementwiseOperation elementwise_op)
+{
+    GridwisePutElementwise1dFunctor::Run(
+        in_grid_1d_desc, p_in_global, p_indices_global, p_out_global, elementwise_op);
+}
+
+// output[indices] = input
+template <typename InGrid1dDesc,
+          typename InDataType,
+          typename IndexDataType,
+          typename OutDataType,
+          typename ElementwiseOperation,
+          InMemoryDataOperationEnum MemOp,
+          index_t InVectorSize>
+struct GridwisePutElement_1D
+{
+    static constexpr auto I0 = Number<0>{};
+
+    static constexpr auto thread_buffer_desc_m =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<InVectorSize>{}));
+
+    __device__ static void Run(const InGrid1dDesc& in_grid_1d_desc,
+                               const InDataType* __restrict__ p_in_global,
+                               const IndexDataType* __restrict__ p_indices_global,
+                               OutDataType* __restrict__ p_out_global,
+                               const ElementwiseOperation& elementwise_op)
+    {
+        // Global Memory
+        const auto in_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_global, in_grid_1d_desc.GetElementSpaceSize());
+
+        const auto indices_global_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Global>(p_indices_global,
+                                                          in_grid_1d_desc.GetElementSpaceSize(),
+                                                          NumericLimits<IndexDataType>::Lowest());
+
+        // VGPR
+        StaticBuffer<AddressSpaceEnum::Vgpr, InDataType, InVectorSize, true> in_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, IndexDataType, InVectorSize, true> indices_thread_buf;
+
+        // Thread id, Block id and index
+        const index_t thread_global_id  = get_thread_global_1d_id();
+        const auto thread_global_offset = make_multi_index(thread_global_id * InVectorSize);
+        const index_t blockSize         = get_block_size();
+        const index_t blockPerGrid      = get_grid_size();
+        const auto M                    = in_grid_1d_desc.GetLength(I0);
+        const index_t loop_step         = blockPerGrid * blockSize * InVectorSize;
+        const auto loop_step_index      = make_multi_index(loop_step);
+
+        auto in_global_load =
+            ThreadwiseTensorSliceTransfer_v2<InDataType,
+                                             InDataType,
+                                             decltype(in_grid_1d_desc),
+                                             decltype(thread_buffer_desc_m),
+                                             Sequence<InVectorSize>, // SliceLengths
+                                             Sequence<0>,            // DimAccessOrder
+                                             0,                      // SrcVectorDim
+                                             InVectorSize,           // ScalarPerVector
+                                             1,                      // SrcScalarStrideInVector
+                                             false>{in_grid_1d_desc, thread_global_offset};
+
+        auto indices_global_load =
+            ThreadwiseTensorSliceTransfer_v2<IndexDataType,
+                                             IndexDataType,
+                                             decltype(in_grid_1d_desc),
+                                             decltype(thread_buffer_desc_m),
+                                             Sequence<InVectorSize>, // SliceLengths
+                                             Sequence<0>,            // DimAccessOrder
+                                             0,                      // SrcVectorDim
+                                             InVectorSize,           // ScalarPerVector
+                                             1,                      // SrcScalarStrideInVector
+                                             false>{in_grid_1d_desc, thread_global_offset};
+
+        index_t num_iter = M / loop_step;
+        do
+        {
+            in_global_load.Run(in_grid_1d_desc,
+                               in_global_buf,
+                               thread_buffer_desc_m,
+                               make_tuple(I0),
+                               in_thread_buf);
+
+            in_global_load.MoveSrcSliceWindow(in_grid_1d_desc, loop_step_index);
+
+            static_for<0, InVectorSize, 1>{}(
+                [&](auto iM) { elementwise_op(in_thread_buf(iM), in_thread_buf[iM]); });
+
+            indices_global_load.Run(in_grid_1d_desc,
+                                    indices_global_buf,
+                                    thread_buffer_desc_m,
+                                    make_tuple(I0),
+                                    indices_thread_buf);
+
+            indices_global_load.MoveSrcSliceWindow(in_grid_1d_desc, loop_step_index);
+
+            static_for<0, InVectorSize, 1>{}([&](auto iM) {
+                if(indices_thread_buf[iM] >= 0)
+                {
+                    if constexpr(MemOp == InMemoryDataOperationEnum::Set)
+                    {
+                        // User should guarantee each index in p_indices_global is different
+                        *(p_out_global + indices_thread_buf[iM]) =
+                            ck::type_convert<OutDataType>(in_thread_buf[iM]);
+                    }
+                    else if constexpr(MemOp == InMemoryDataOperationEnum::AtomicAdd)
+                    {
+                        atomic_add<OutDataType>(p_out_global + indices_thread_buf[iM],
+                                                ck::type_convert<OutDataType>(in_thread_buf[iM]));
+                    }
+                    else if constexpr(MemOp == InMemoryDataOperationEnum::AtomicMax)
+                    {
+                        atomic_max<OutDataType>(p_out_global + indices_thread_buf[iM],
+                                                ck::type_convert<OutDataType>(in_thread_buf[iM]));
+                    }
+                    else if constexpr(MemOp == InMemoryDataOperationEnum::Add)
+                    {
+                        // User should guarantee each index in p_indices_global is different
+                        *(p_out_global + indices_thread_buf[iM]) +=
+                            ck::type_convert<OutDataType>(in_thread_buf[iM]);
+                    }
+                    else
+                    {
+                        static_assert(MemOp == InMemoryDataOperationEnum::Set ||
+                                      MemOp == InMemoryDataOperationEnum::AtomicAdd ||
+                                      MemOp == InMemoryDataOperationEnum::AtomicMax ||
+                                      MemOp == InMemoryDataOperationEnum::Add);
+                    }
+                }
+            });
+
+        } while(--num_iter);
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp
new file mode 100644
index 0000000000..41352fabeb
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp
@@ -0,0 +1,57 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+
+namespace ck {
+
+template <index_t BlockSize, typename DataType, typename Grid1dBufferDescType>
+__global__ void kernel_buffer_set_value(const Grid1dBufferDescType grid_1d_buffer_desc,
+                                        DataType* const __restrict__ p_global,
+                                        DataType value)
+
+{
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    constexpr auto I0 = Number<0>{};
+
+    const index_t thread_local_id = get_thread_local_1d_id();
+    const index_t block_global_id = get_block_1d_id();
+
+    const index_t thread_global_id = block_global_id * BlockSize + thread_local_id;
+
+    StaticBuffer<AddressSpaceEnum::Vgpr, DataType, 1, true> value_buf;
+
+    value_buf(I0) = value;
+
+    constexpr auto val_buff_desc = make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}));
+
+    auto global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+        p_global, grid_1d_buffer_desc.GetElementSpaceSize());
+
+    if(thread_global_id < grid_1d_buffer_desc.GetElementSize())
+    {
+        auto threadwise_store = ThreadwiseTensorSliceTransfer_v1r3<DataType,
+                                                                   DataType,
+                                                                   decltype(val_buff_desc),
+                                                                   Grid1dBufferDescType,
+                                                                   PassThroughOp,
+                                                                   Sequence<1>,
+                                                                   Sequence<0>,
+                                                                   0,
+                                                                   1,
+                                                                   InMemoryDataOperationEnum::Set,
+                                                                   1,
+                                                                   true>(
+            grid_1d_buffer_desc, make_multi_index(thread_global_id), PassThroughOp{});
+
+        threadwise_store.Run(
+            val_buff_desc, make_tuple(I0), value_buf, grid_1d_buffer_desc, global_buf);
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp
new file mode 100644
index 0000000000..0ad36b418a
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+
+namespace ck {
+
+template <typename Grid1dBufferDescTuple,
+          index_t NumBuffer,
+          index_t BlockSize,
+          typename DataTypePointerTuple,
+          typename DataTypeTuple>
+__global__ void
+kernel_multiple_buffer_set_value(const Grid1dBufferDescTuple grid_1d_buffer_desc_tuple,
+                                 DataTypePointerTuple p_global_tuple,
+                                 DataTypeTuple value_tuple)
+
+{
+    static_assert(NumBuffer == DataTypePointerTuple::Size() && NumBuffer == DataTypeTuple::Size(),
+                  "The tuple size should be same as NumBuffer!");
+
+    static_for<0, NumBuffer, 1>{}([&](auto iB) {
+        using DataTypePointer     = remove_cvref_t<decltype(DataTypePointerTuple{}[iB])>;
+        using DataTypeFromPointer = remove_pointer_t<DataTypePointer>;
+        using DataType            = remove_cvref_t<decltype(DataTypeTuple{}[iB])>;
+
+        static_assert(is_same<DataType, DataTypeFromPointer>::value,
+                      "Types in tuples does not match!");
+    });
+
+    constexpr auto I0 = Number<0>{};
+
+    const index_t thread_global_id = get_thread_global_1d_id();
+
+    auto value_buf_tuple = generate_tuple(
+        [&](auto iB) {
+            using DataType = remove_cvref_t<decltype(DataTypeTuple{}[iB])>;
+
+            return StaticBuffer<AddressSpaceEnum::Vgpr, DataType, 1, true>{};
+        },
+        Number<NumBuffer>{});
+
+    static_for<0, NumBuffer, 1>{}([&](auto iB) {
+        static_for<0, 1, 1>{}([&](auto J) { value_buf_tuple(iB)(J) = value_tuple[iB]; });
+    });
+
+    auto global_buf_tuple = generate_tuple(
+        [&](auto iB) {
+            return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_global_tuple(iB), grid_1d_buffer_desc_tuple[iB].GetElementSpaceSize());
+        },
+        Number<NumBuffer>{});
+
+    constexpr auto val_buff_desc = make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}));
+
+    static_for<0, NumBuffer, 1>{}([&](auto iB) {
+        using DataType      = remove_cvref_t<decltype(DataTypeTuple{}[iB])>;
+        using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+        auto threadwise_store =
+            ThreadwiseTensorSliceTransfer_v1r3<DataType,
+                                               DataType,
+                                               decltype(val_buff_desc),
+                                               decltype(Grid1dBufferDescTuple{}[iB]),
+                                               PassThroughOp,
+                                               Sequence<1>,
+                                               Sequence<0>,
+                                               0,
+                                               1,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                grid_1d_buffer_desc_tuple[iB], make_multi_index(thread_global_id), PassThroughOp{});
+
+        threadwise_store.Run(val_buff_desc,
+                             make_tuple(I0),
+                             value_buf_tuple(iB),
+                             grid_1d_buffer_desc_tuple[iB],
+                             global_buf_tuple(iB));
+    });
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_softmax.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_softmax.hpp
new file mode 100644
index 0000000000..5f56ac6fc4
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_softmax.hpp
@@ -0,0 +1,407 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseReduction,
+          typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename GridDesc_M_K>
+__global__ void kernel_softmax(const GridDesc_M_K in_grid_desc_m_k,
+                               const GridDesc_M_K out_grid_desc_m_k,
+                               index_t block_group_size,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType alpha,
+                               const InDataType* const __restrict__ p_in_value_global,
+                               AccDataType beta,
+                               OutDataType* const __restrict__ p_out_value_global)
+{
+    GridwiseReduction::Run(in_grid_desc_m_k,
+                           out_grid_desc_m_k,
+                           block_group_size,
+                           num_k_block_tile_iteration,
+                           alpha,
+                           p_in_value_global,
+                           beta,
+                           p_out_value_global);
+};
+
+template <typename InDataType,
+          typename OutDataType,
+          typename AccDataType,
+          typename GridDesc_M_K,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          index_t OutDstVectorSize,
+          bool SweepOnce>
+struct GridwiseSoftmax_mk_to_mk
+{
+    static_assert(((InSrcVectorDim == 0 && MThreadSliceSize % InSrcVectorSize == 0) ||
+                   (InSrcVectorDim == 1 && KThreadSliceSize % InSrcVectorSize == 0)) &&
+                      (KThreadSliceSize % OutDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (InSrcVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    __device__ static void Run(const GridDesc_M_K& in_grid_desc_m_k,
+                               const GridDesc_M_K& out_grid_desc_m_k,
+                               index_t block_group_size,
+                               index_t num_k_block_tile_iteration,
+                               AccDataType alpha,
+                               const InDataType* const __restrict__ p_in_value_global,
+                               AccDataType beta,
+                               OutDataType* const __restrict__ p_out_value_global)
+    {
+        if constexpr(SweepOnce)
+        {
+            num_k_block_tile_iteration = 1;
+        }
+
+        // LDS
+        __shared__ AccDataType p_reduce_work_buffer[BlockSize];
+
+        auto out_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_value_global, out_grid_desc_m_k.GetElementSpaceSize());
+
+        auto reduce_work_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            in_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
+            out_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> max_value_buf;
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            max_value_buf(I) = reduce::Max::template GetIdentityValue<AccDataType>();
+        });
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> accu_value_buf;
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            accu_value_buf(I) = reduce::Add::template GetIdentityValue<AccDataType>();
+        });
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / block_group_size;
+        const index_t block_local_id  = block_global_id % block_group_size;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        const index_t reduceSizePerBlock = K_BlockTileSize * num_k_block_tile_iteration;
+
+        using ThreadBufferLengths         = Sequence<MThreadSliceSize, KThreadSliceSize>;
+        constexpr auto thread_buffer_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
+
+        // Normally, 0 as invalid element value is adequate since 0 makes no contribution to
+        // accumulated result. However, in stable softmax, all values 0s or not are subtracted by
+        // another value_max. As numbers become non-zero, effectively it allows invalid values to
+        // slip through and contribute to the accumulated result.
+        //
+        // The trick here is leveraging the fact that many math functions (add, sub, exp, ...)
+        // propagate NaNs when operands have NaNs involved. By initialiing invalid element value
+        // with NaN, an invalid value doing math manipulations is still NaN, which in turn can still
+        // be identified as an invalid value. We can then discard the invalid values which
+        // originally failed the bound check during accumulation. This allows to ignore values that
+        // failed bound check even after multiple math manipulations.
+        //
+        // NOTE: reset coordinate after every step because the same threadwise copy will sweep
+        // through global memory 3 times back and forth
+        auto threadwise_src_load = ThreadwiseTensorSliceTransfer_v2<InDataType,
+                                                                    AccDataType,
+                                                                    GridDesc_M_K,
+                                                                    decltype(thread_buffer_desc),
+                                                                    ThreadBufferLengths,
+                                                                    ThreadBufferDimAccessOrder,
+                                                                    InSrcVectorDim,
+                                                                    InSrcVectorSize,
+                                                                    1,
+                                                                    true /* ResetCoordAfterRun */,
+                                                                    true /* InvalidElementAsNaN */>(
+            in_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             block_local_id * reduceSizePerBlock +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_dst_load = ThreadwiseTensorSliceTransfer_v2<OutDataType,
+                                                                    AccDataType,
+                                                                    GridDesc_M_K,
+                                                                    decltype(thread_buffer_desc),
+                                                                    ThreadBufferLengths,
+                                                                    ThreadBufferDimAccessOrder,
+                                                                    InSrcVectorDim,
+                                                                    InSrcVectorSize,
+                                                                    1,
+                                                                    false>(
+            out_grid_desc_m_k,
+            make_multi_index(blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                             block_local_id * reduceSizePerBlock +
+                                 thread_k_cluster_id * KThreadSliceSize));
+
+        auto threadwise_dst_store =
+            ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                               OutDataType,
+                                               decltype(thread_buffer_desc),
+                                               GridDesc_M_K,
+                                               PassThroughOp,
+                                               ThreadBufferLengths,
+                                               ThreadBufferDimAccessOrder,
+                                               InSrcVectorDim,
+                                               OutDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                out_grid_desc_m_k,
+                make_multi_index(
+                    blkgroup_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                    block_local_id * reduceSizePerBlock + thread_k_cluster_id * KThreadSliceSize),
+                PassThroughOp{});
+
+        constexpr auto in_thread_copy_fwd_step =
+            make_multi_index(0, SweepOnce ? 0 : K_BlockTileSize);
+        constexpr auto in_thread_copy_bwd_step =
+            make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize);
+
+        ///
+        /// max(x)
+        ///
+        using BlockwiseMaxReduce = PartitionedBlockwiseReduction<
+            AccDataType,
+            BlockSize,
+            ThreadClusterLengths_M_K,
+            ThreadClusterArrangeOrder,
+            reduce::Max,
+            false, // param ignored
+            detail::AccumulateWithNanIgnore<reduce::Max, AccDataType>>;
+
+        using ThreadwiseMaxReduce =
+            ThreadwiseReduction<AccDataType,
+                                ThreadReduceSrcDesc_M_K,
+                                ThreadReduceDstDesc_M,
+                                reduce::Max,
+                                false, // param ignored
+                                detail::AccumulateWithNanIgnore<reduce::Max, AccDataType>>;
+
+        const auto in_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_value_global, in_grid_desc_m_k.GetElementSpaceSize());
+
+        index_t reducedTiles = 0;
+        do
+        {
+            threadwise_src_load.Run(in_grid_desc_m_k,
+                                    in_global_val_buf,
+                                    thread_buffer_desc,
+                                    make_tuple(I0, I0),
+                                    in_thread_buf);
+
+            ThreadwiseMaxReduce::Reduce(in_thread_buf, max_value_buf);
+
+            threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_fwd_step);
+
+            reducedTiles++;
+        } while(reducedTiles < num_k_block_tile_iteration);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            BlockwiseMaxReduce::Reduce(reduce_work_buf, max_value_buf(I));
+            block_sync_lds();
+        });
+
+        threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_bwd_step);
+
+        ///
+        /// sum(exp(x - max(x)))
+        ///
+        using BlockwiseSumReduce = PartitionedBlockwiseReduction<
+            AccDataType,
+            BlockSize,
+            ThreadClusterLengths_M_K,
+            ThreadClusterArrangeOrder,
+            reduce::Add,
+            false, // ignored
+            detail::AccumulateWithNanIgnore<reduce::Add, AccDataType>>;
+
+        using ThreadwiseSumReduce =
+            ThreadwiseReduction<AccDataType,
+                                ThreadReduceSrcDesc_M_K,
+                                ThreadReduceDstDesc_M,
+                                reduce::Add,
+                                false, // ignored
+                                detail::AccumulateWithNanIgnore<reduce::Add, AccDataType>>;
+
+        reducedTiles = 0;
+        do
+        {
+            if constexpr(!SweepOnce)
+            {
+                threadwise_src_load.Run(in_grid_desc_m_k,
+                                        in_global_val_buf,
+                                        thread_buffer_desc,
+                                        make_tuple(I0, I0),
+                                        in_thread_buf);
+            }
+
+            // do element-wise pre-reduction operation
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                    constexpr auto offset = thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+                    out_thread_buf(Number<offset>{}) =
+                        math::exp(in_thread_buf(Number<offset>{}) - max_value_buf(iM));
+                });
+            });
+
+            ThreadwiseSumReduce::Reduce(out_thread_buf, accu_value_buf);
+
+            threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_bwd_step);
+
+            reducedTiles++;
+        } while(reducedTiles < num_k_block_tile_iteration);
+
+        block_sync_lds(); // wait for reading being complete before writing to LDS
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            BlockwiseSumReduce::Reduce(reduce_work_buf, accu_value_buf(I));
+            block_sync_lds();
+        });
+
+        threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_fwd_step);
+
+        ///
+        /// softmax
+        ///
+        reducedTiles = 0;
+        if(float_equal_zero{}(beta))
+        {
+            do
+            {
+                if constexpr(!SweepOnce)
+                {
+                    threadwise_src_load.Run(in_grid_desc_m_k,
+                                            in_global_val_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_buf);
+                }
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    // out = alpha * exp(x - max(x)) / sum(exp(x - max(x)))
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+                        out_thread_buf(Number<offset>{}) =
+                            alpha * math::exp(in_thread_buf(Number<offset>{}) - max_value_buf(iM)) /
+                            accu_value_buf(iM);
+                    });
+                });
+
+                threadwise_dst_store.Run(thread_buffer_desc,
+                                         make_tuple(I0, I0),
+                                         out_thread_buf,
+                                         out_grid_desc_m_k,
+                                         out_global_val_buf);
+
+                threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_fwd_step);
+                threadwise_dst_store.MoveDstSliceWindow(out_grid_desc_m_k, in_thread_copy_fwd_step);
+
+                reducedTiles++;
+            } while(reducedTiles < num_k_block_tile_iteration);
+        }
+        else
+        {
+            StaticBuffer<AddressSpaceEnum::Vgpr,
+                         AccDataType,
+                         MThreadSliceSize * KThreadSliceSize,
+                         true>
+                in_prior_dst_buf;
+            do
+            {
+                if constexpr(!SweepOnce)
+                {
+                    threadwise_src_load.Run(in_grid_desc_m_k,
+                                            in_global_val_buf,
+                                            thread_buffer_desc,
+                                            make_tuple(I0, I0),
+                                            in_thread_buf);
+                }
+                threadwise_dst_load.Run(out_grid_desc_m_k,
+                                        out_global_val_buf,
+                                        thread_buffer_desc,
+                                        make_tuple(I0, I0),
+                                        in_prior_dst_buf);
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    // out = alpha * exp(x - max(x)) / sum(exp(x - max(x))) + beta * prior_out
+                    static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
+                        constexpr auto offset =
+                            thread_buffer_desc.CalculateOffset(make_tuple(iM, iK));
+                        out_thread_buf(Number<offset>{}) =
+                            alpha * math::exp(in_thread_buf(Number<offset>{}) - max_value_buf(iM)) /
+                                accu_value_buf(iM) +
+                            beta * in_prior_dst_buf(Number<offset>{});
+                    });
+                });
+
+                threadwise_dst_store.Run(thread_buffer_desc,
+                                         make_tuple(I0, I0),
+                                         out_thread_buf,
+                                         out_grid_desc_m_k,
+                                         out_global_val_buf);
+
+                threadwise_src_load.MoveSrcSliceWindow(in_grid_desc_m_k, in_thread_copy_fwd_step);
+                threadwise_dst_store.MoveDstSliceWindow(out_grid_desc_m_k, in_thread_copy_fwd_step);
+                threadwise_dst_load.MoveSrcSliceWindow(out_grid_desc_m_k, in_thread_copy_fwd_step);
+
+                reducedTiles++;
+            } while(reducedTiles < num_k_block_tile_iteration);
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_sparse_embeddings_forward_layernorm.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_sparse_embeddings_forward_layernorm.hpp
new file mode 100644
index 0000000000..287b4e5421
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_sparse_embeddings_forward_layernorm.hpp
@@ -0,0 +1,319 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+
+namespace ck {
+
+template <typename GridwiseSparseEmbedding,
+          typename EmbType,
+          typename IndexType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename OutType,
+          typename OutGridDesc,
+          typename EmbElementwiseOperation,
+          ck::index_t NumEmbeddings>
+#if CK_USE_LAUNCH_BOUNDS
+__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+    __global__ void kernel_sparse_embeddings_forward_layernorm(
+        OutType* p_out,
+        const ck::Array<EmbType*, NumEmbeddings> p_embs,
+        const ck::Array<IndexType*, NumEmbeddings> p_indexes,
+        const GammaDataType* p_gamma,
+        const BetaDataType* p_beta,
+        const OutGridDesc out_grid_desc,
+        const AccDataType epsilon,
+        const EmbElementwiseOperation emb_elementwise_op)
+{
+    GridwiseSparseEmbedding::Run(
+        p_out, p_embs, p_indexes, p_gamma, p_beta, out_grid_desc, epsilon, emb_elementwise_op);
+}
+
+template <typename EmbType,
+          typename IndexType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename OutType,
+          typename OutGridDesc,
+          typename EmbElementwiseOperation,
+          ck::index_t BlockSize,
+          ck::index_t DimClusterSize,
+          ck::index_t RowClusterSize,
+          ck::index_t DimPerBlock,   // Row x Dim, along Dim
+          ck::index_t RowPerBlock,   // Row x Dim, along Row
+          ck::index_t DimThreadSize, // this is actually not vector, but number of registers
+          ck::index_t RowVectorSize,
+          ck::index_t NumEmbeddings>
+struct GridwiseSparseEmbeddingsForwardLayernorm
+{
+    static constexpr auto I0          = Number<0>{};
+    static constexpr auto I1          = Number<1>{};
+    static constexpr auto I2          = Number<2>{};
+    static constexpr auto I3          = Number<3>{};
+    static constexpr index_t WaveSize = 64;
+
+    static_assert(BlockSize == RowClusterSize * DimClusterSize,
+                  "Invalid cluster distribution within block");
+    static_assert(RowClusterSize % WaveSize == 0, "need to be wavewise");
+
+    static_assert(DimPerBlock % (DimClusterSize * DimThreadSize) == 0, "");
+    static_assert(RowPerBlock % (RowClusterSize * RowVectorSize) == 0, "");
+
+    static constexpr auto DimSubBlocks = DimPerBlock / (DimClusterSize * DimThreadSize);
+    static constexpr auto RowSubBlocks = RowPerBlock / (RowClusterSize * RowVectorSize);
+
+    static_assert((DimPerBlock % DimSubBlocks == 0) && (RowPerBlock % RowSubBlocks == 0), "");
+    static constexpr auto DimPerSubBlock = DimPerBlock / DimSubBlocks;
+    static constexpr auto RowPerSubBlock = RowPerBlock / RowSubBlocks;
+
+    using ThreadwiseWolfordDesc2D = decltype(make_naive_tensor_descriptor_packed(make_tuple(
+        Number<DimSubBlocks * DimThreadSize>{}, Number<RowSubBlocks * RowVectorSize>{})));
+
+    using ThreadwiseWolfordDescReduce = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<DimSubBlocks * DimThreadSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelford<AccDataType, ThreadwiseWolfordDesc2D, ThreadwiseWolfordDescReduce>;
+
+    using ThreadClusterLength = Sequence<DimClusterSize, RowClusterSize>;
+
+    using BlockwiseWelford =
+        BlockwiseWelford<AccDataType, BlockSize, ThreadClusterLength, Sequence<0, 1>>;
+
+    __device__ static void Run(OutType* p_out,
+                               const ck::Array<EmbType*, NumEmbeddings> p_embs,
+                               const ck::Array<IndexType*, NumEmbeddings> p_indexes,
+                               const GammaDataType* p_gamma,
+                               const BetaDataType* p_beta,
+                               const OutGridDesc,
+                               const AccDataType epsilon,
+                               const EmbElementwiseOperation emb_elementwise_op)
+    {
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+
+        constexpr auto thread_cluster_desc =
+            make_cluster_descriptor(Sequence<DimClusterSize, RowClusterSize>{}, Sequence<0, 1>{});
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_dim_cluster_id = thread_cluster_idx[I0];
+        const auto thread_row_cluster_id = thread_cluster_idx[I1];
+
+        const auto wave_dim_id = __builtin_amdgcn_readfirstlane(thread_dim_cluster_id / WaveSize);
+
+        const auto index_start = block_global_id * DimPerBlock + wave_dim_id * DimThreadSize;
+
+        auto threadwise_welford       = ThreadwiseWelford();
+        threadwise_welford.max_count_ = RowSubBlocks * RowVectorSize;
+
+        constexpr auto thread_buf_size =
+            DimSubBlocks * DimThreadSize * RowSubBlocks * RowVectorSize;
+        constexpr auto thread_buf_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(DimSubBlocks, DimThreadSize, RowSubBlocks, RowVectorSize));
+        constexpr auto mean_var_buf_size = DimSubBlocks * DimThreadSize;
+        constexpr auto mean_var_buf_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(DimSubBlocks, DimThreadSize));
+        constexpr auto gamma_beta_buf_size = RowSubBlocks * RowVectorSize;
+        constexpr auto gamma_beta_buf_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(RowSubBlocks, RowVectorSize));
+
+        ck::Array<StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, thread_buf_size, true>,
+                  NumEmbeddings>
+            in_thread_bufs;
+        ck::Array<StaticBuffer<AddressSpaceEnum::Vgpr, IndexType, DimPerBlock, true>, NumEmbeddings>
+            index_bufs;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, thread_buf_size, true> acc_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, gamma_beta_buf_size, true>
+            gamma_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, gamma_beta_buf_size, true>
+            beta_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, mean_var_buf_size, true> mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, mean_var_buf_size, true> var_thread_buf;
+
+        auto load_current_sub_row = [&](auto i_dim_sub_, auto i_row_sub_) {
+            ck::Array<vector_type_maker_t<EmbType, RowVectorSize>, NumEmbeddings> emb_vectors;
+            auto emb_a         = emb_vectors[0];
+            using src_vector_t = typename decltype(emb_a)::type;
+            static_for<0, DimThreadSize, 1>{}([&](auto i_dim_vec_) {
+                constexpr auto current_dim = i_dim_sub_ * DimPerSubBlock + i_dim_vec_;
+
+                auto thread_offset = (thread_row_cluster_id + i_row_sub_ * RowClusterSize) *
+                                     sizeof(EmbType) * RowVectorSize;
+                static_for<0, NumEmbeddings, 1>{}([&](auto i_embedding_) {
+                    IndexType index = index_bufs[i_embedding_][Number<current_dim>{}];
+
+                    int32x4_t emb_res = make_wave_buffer_resource_with_default_range(
+                        p_embs[i_embedding_] + index * RowPerBlock);
+                    emb_vectors(i_embedding_).template AsType<src_vector_t>()(I0) =
+                        amd_buffer_load_impl<EmbType, RowVectorSize>(emb_res, thread_offset, 0);
+                });
+
+                static_for<0, RowVectorSize, 1>{}([&](auto i_row_vec_) {
+                    constexpr auto register_offset = thread_buf_desc.CalculateOffset(
+                        make_tuple(i_dim_sub_, i_dim_vec_, i_row_sub_, i_row_vec_));
+                    static_for<0, NumEmbeddings, 1>{}([&](auto i_embedding_) {
+                        in_thread_bufs(i_embedding_)(Number<register_offset>{}) =
+                            ck::type_convert<AccDataType>(
+                                emb_vectors[i_embedding_].template AsType<EmbType>()[i_row_vec_]);
+                    });
+                });
+            });
+        };
+
+        auto accumulate_current_sub_row = [&](auto i_dim_sub_, auto i_row_sub_) {
+            static_for<0, DimThreadSize, 1>{}([&](auto i_dim_vec_) {
+                static_for<0, RowVectorSize, 1>{}([&](auto i_row_vec_) {
+                    constexpr auto register_offset = thread_buf_desc.CalculateOffset(
+                        make_tuple(i_dim_sub_, i_dim_vec_, i_row_sub_, i_row_vec_));
+                    auto in_data_refs = generate_tie(
+                        [&](auto i_embedding_) -> const auto& {
+                            return in_thread_bufs(i_embedding_)(Number<register_offset>{});
+                        },
+                        Number<NumEmbeddings>{});
+                    auto out_data_refs = generate_tie(
+                        [&](auto) -> auto& { return acc_thread_buf(Number<register_offset>{}); },
+                        Number<1>{});
+                    unpack2(emb_elementwise_op, out_data_refs, in_data_refs);
+                });
+            });
+        };
+
+        auto threadwise_welford_sub_row = [&](auto i_dim_sub_, auto i_row_sub_) {
+            static_for<0, DimThreadSize, 1>{}([&](auto i_dim_vec_) {
+                static_for<0, RowVectorSize, 1>{}([&](auto i_row_vec_) {
+                    constexpr auto register_offset = thread_buf_desc.CalculateOffset(
+                        make_tuple(i_dim_sub_, i_dim_vec_, i_row_sub_, i_row_vec_));
+                    constexpr auto mean_var_offset =
+                        mean_var_buf_desc.CalculateOffset(make_tuple(i_dim_sub_, i_dim_vec_));
+
+                    threadwise_welford.cur_count_++;
+                    threadwise_welford.Update(mean_thread_buf(Number<mean_var_offset>{}),
+                                              var_thread_buf(Number<mean_var_offset>{}),
+                                              acc_thread_buf(Number<register_offset>{}));
+                });
+            });
+        };
+
+        auto threadwise_normalize_store_out = [&](auto i_dim_sub_, auto i_row_sub_) {
+            int32x4_t out_res =
+                make_wave_buffer_resource_with_default_range(p_out + index_start * RowPerBlock);
+            static_for<0, DimThreadSize, 1>{}([&](auto i_dim_vec_) {
+                vector_type_maker_t<OutType, RowVectorSize> out_vector;
+                using dst_vector_t = typename decltype(out_vector)::type;
+
+                constexpr auto mean_var_offset =
+                    mean_var_buf_desc.CalculateOffset(make_tuple(i_dim_sub_, i_dim_vec_));
+                auto divisor =
+                    1 / __builtin_amdgcn_sqrtf(var_thread_buf(Number<mean_var_offset>{}) + epsilon);
+                static_for<0, RowVectorSize, 1>{}([&](auto i_row_vec_) {
+                    constexpr auto register_offset = thread_buf_desc.CalculateOffset(
+                        make_tuple(i_dim_sub_, i_dim_vec_, i_row_sub_, i_row_vec_));
+                    constexpr auto gamma_beta_offset =
+                        gamma_beta_buf_desc.CalculateOffset(make_tuple(i_row_sub_, i_row_vec_));
+
+                    auto acc_val = acc_thread_buf[Number<register_offset>{}];
+                    acc_val      = (acc_val - mean_thread_buf(Number<mean_var_offset>{})) * divisor;
+                    acc_val      = acc_val * gamma_thread_buf[Number<gamma_beta_offset>{}] +
+                              beta_thread_buf[Number<gamma_beta_offset>{}];
+
+                    out_vector.template AsType<OutType>()(Number<i_row_vec_>{}) =
+                        type_convert<OutType>(acc_val);
+                });
+
+                index_t thread_offset = (thread_row_cluster_id + i_row_sub_ * RowClusterSize) *
+                                        sizeof(OutType) * RowVectorSize;
+
+                amd_buffer_store_impl<OutType, RowVectorSize>(
+                    out_vector.template AsType<dst_vector_t>()[Number<0>{}],
+                    out_res,
+                    thread_offset,
+                    0);
+            });
+        };
+
+        // first load index
+        ck::static_for<0, DimPerBlock, 1>{}([&](auto i_idx_) {
+            // prefer use s_load
+            ck::static_for<0, NumEmbeddings, 1>{}([&](auto i_embedding_) {
+                index_bufs(i_embedding_)(i_idx_) =
+                    p_indexes[i_embedding_][index_start + i_idx_.value];
+            });
+        });
+
+        // load gamma/beta
+        static_for<0, RowSubBlocks, 1>{}([&](auto i_row_sub_) {
+            vector_type_maker_t<GammaDataType, RowVectorSize> gamma_vector;
+            vector_type_maker_t<BetaDataType, RowVectorSize> beta_vector;
+
+            index_t thread_offset_gamma = (thread_row_cluster_id + i_row_sub_ * RowClusterSize) *
+                                          sizeof(GammaDataType) * RowVectorSize;
+            index_t thread_offset_beta = (thread_row_cluster_id + i_row_sub_ * RowClusterSize) *
+                                         sizeof(BetaDataType) * RowVectorSize;
+
+            int32x4_t gamma_res = make_wave_buffer_resource_with_default_range(p_gamma);
+            int32x4_t beta_res  = make_wave_buffer_resource_with_default_range(p_beta);
+
+            gamma_vector.template AsType<typename decltype(gamma_vector)::type>()(I0) =
+                amd_buffer_load_impl<GammaDataType, RowVectorSize>(
+                    gamma_res, thread_offset_gamma, 0);
+            beta_vector.template AsType<typename decltype(beta_vector)::type>()(I0) =
+                amd_buffer_load_impl<BetaDataType, RowVectorSize>(beta_res, thread_offset_beta, 0);
+
+            static_for<0, RowVectorSize, 1>{}([&](auto i_row_vec_) {
+                constexpr auto offset =
+                    gamma_beta_buf_desc.CalculateOffset(make_tuple(i_row_sub_, i_row_vec_));
+                gamma_thread_buf(Number<offset>{}) = type_convert<AccDataType>(
+                    gamma_vector.template AsType<GammaDataType>()[Number<i_row_vec_>{}]);
+                beta_thread_buf(Number<offset>{}) = type_convert<AccDataType>(
+                    beta_vector.template AsType<BetaDataType>()[Number<i_row_vec_>{}]);
+            });
+        });
+
+        static_for<0, thread_buf_size, 1>{}(
+            [&](auto I) { acc_thread_buf(I) = type_convert<AccDataType>(0.0f); });
+
+        static_for<0, mean_var_buf_size, 1>{}([&](auto I) {
+            mean_thread_buf(I) = type_convert<AccDataType>(0.0f);
+            var_thread_buf(I)  = type_convert<AccDataType>(0.0f);
+        });
+
+        static_for<0, DimSubBlocks, 1>{}([&](auto i_dim_sub) {
+            load_current_sub_row(i_dim_sub, Number<0>{});
+            static_for<0, RowSubBlocks - 1, 1>{}([&](auto i_row) {
+                load_current_sub_row(i_dim_sub, Number<1>{} + i_row);
+                accumulate_current_sub_row(i_dim_sub, i_row);
+                threadwise_welford_sub_row(i_dim_sub, i_row);
+            });
+            accumulate_current_sub_row(i_dim_sub, Number<RowSubBlocks - 1>{});
+            threadwise_welford_sub_row(i_dim_sub, Number<RowSubBlocks - 1>{});
+
+            // blockwise welford
+            static_for<0, mean_var_buf_size, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+                BlockwiseWelford::Run(
+                    mean_thread_buf(I), var_thread_buf(I), threadwise_welford.cur_count_);
+            });
+
+            // store
+            static_for<0, RowSubBlocks, 1>{}(
+                [&](auto i_row) { threadwise_normalize_store_out(i_dim_sub, i_row); });
+        });
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_naive_variance.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_naive_variance.hpp
new file mode 100644
index 0000000000..c3f122106d
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_naive_variance.hpp
@@ -0,0 +1,505 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp"
+#include "ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+// Y = Normalization(X, Beta, Gamma)
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename ComputeDataType,
+          typename YElementwiseOperation,
+          typename GridDesc_M_K,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcVectorDim,
+          index_t XSrcVectorSize,
+          index_t GammaSrcVectorDim,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorDim,
+          index_t BetaSrcVectorSize,
+          index_t YDstVectorDim,
+          index_t YDstVectorSize,
+          bool SweepOnce>
+struct GridwiseNormalizationNaiveVariance_mk_to_mk
+{
+    static_assert((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert((YDstVectorDim == 0 && MThreadSliceSize % YDstVectorSize == 0) ||
+                      (YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert(XSrcVectorSize == YDstVectorSize);
+    static_assert(XSrcVectorSize == GammaSrcVectorSize);
+    static_assert(XSrcVectorSize == BetaSrcVectorSize);
+
+    static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadBufferLengths_M_K                = Sequence<MThreadSliceSize, XSrcVectorSize>;
+    static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using BlockwiseSumReduce = PartitionedBlockwiseReduction<ComputeDataType,
+                                                             BlockSize,
+                                                             ThreadClusterLengths_M_K,
+                                                             ThreadClusterArrangeOrder,
+                                                             reduce::Add,
+                                                             true>;
+
+    using ThreadwiseSumReduce = ThreadwiseReduction<ComputeDataType,
+                                                    ThreadReduceSrcDesc_M_K,
+                                                    ThreadReduceDstDesc_M,
+                                                    reduce::Add,
+                                                    true>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr index_t M_BlockTileSize     = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize     = KThreadClusterSize * KThreadSliceSize;
+    static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
+
+    static constexpr auto ThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
+
+    __device__ static void Run(const GridDesc_M_K& x_grid_desc_m_k,
+                               const GridDesc_M_K& gamma_grid_desc_m_k,
+                               const GridDesc_M_K& beta_grid_desc_m_k,
+                               const GridDesc_M_K& y_grid_desc_m_k,
+                               index_t num_k_block_tile_iteration,
+                               ComputeDataType epsilon,
+                               const XDataType* const __restrict__ p_x_global,
+                               const GammaDataType* const __restrict__ p_gamma_global,
+                               const BetaDataType* const __restrict__ p_beta_global,
+                               YDataType* const __restrict__ p_y_global,
+                               const YElementwiseOperation y_elementwise_op)
+    {
+        // LDS
+        __shared__ ComputeDataType p_reduce_work_buffer[BlockSize];
+
+        auto reduce_work_buf =
+            make_dynamic_buffer<AddressSpaceEnum::Lds>(p_reduce_work_buffer, BlockSize);
+
+        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
+
+        auto x_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    ComputeDataType,
+                                    MThreadSliceSize * XSrcVectorSize,
+                                    true>{};
+            },
+            Number<ThreadBufferNumber>{});
+
+        auto gamma_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    ComputeDataType,
+                                    MThreadSliceSize * GammaSrcVectorSize,
+                                    true>{};
+            },
+            Number<ThreadBufferNumber>{});
+
+        auto& beta_thread_buf = gamma_thread_buf;
+
+        auto y_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    ComputeDataType,
+                                    MThreadSliceSize * YDstVectorSize,
+                                    true>{};
+            },
+            Number<ThreadBufferNumber>{});
+
+        auto& x_square_thread_buf = y_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            mean_square_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>&
+            var_thread_buf = mean_square_thread_buf;
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  ComputeDataType,
+                                                                  GridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize,
+                             thread_k_cluster_id * XSrcVectorSize));
+
+        auto threadwise_gamma_load =
+            ThreadwiseTensorSliceTransfer_v2<GammaDataType,
+                                             ComputeDataType,
+                                             GridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_k),
+                                             ThreadBufferLengths_M_K,
+                                             ThreadBufferDimAccessOrder,
+                                             GammaSrcVectorDim,
+                                             GammaSrcVectorSize,
+                                             1,
+                                             true>(
+                gamma_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * GammaSrcVectorSize));
+
+        auto threadwise_beta_load =
+            ThreadwiseTensorSliceTransfer_v2<BetaDataType,
+                                             ComputeDataType,
+                                             GridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_k),
+                                             ThreadBufferLengths_M_K,
+                                             ThreadBufferDimAccessOrder,
+                                             BetaSrcVectorDim,
+                                             BetaSrcVectorSize,
+                                             1,
+                                             true>(
+                beta_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * BetaSrcVectorSize));
+
+        auto threadwise_y_store =
+            ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
+                                               YDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               GridDesc_M_K,
+                                               YElementwiseOperation,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               YDstVectorDim,
+                                               YDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                y_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * YDstVectorSize),
+                y_elementwise_op);
+
+        constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
+        constexpr auto thread_copy_bwd_step_m_k =
+            make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize);
+
+        const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x_global, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto gamma_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_gamma_global, gamma_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto beta_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_beta_global, beta_grid_desc_m_k.GetElementSpaceSize());
+
+        // E(x), E[x^2], var(x)
+        // FIXME: Should not hack the transform from deviceOP
+        int reduce_length = x_grid_desc_m_k.GetTransforms()[I2].GetUpperLengths()[I0];
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            mean_thread_buf(I)        = reduce::Add::template GetIdentityValue<ComputeDataType>();
+            mean_square_thread_buf(I) = reduce::Add::template GetIdentityValue<ComputeDataType>();
+        });
+
+        // Separate sweep once and sweep twice pipeline
+        if constexpr(SweepOnce)
+        {
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_x_load.Run(x_grid_desc_m_k,
+                                      x_global_val_buf,
+                                      thread_buffer_desc_m_k,
+                                      make_tuple(I0, I0),
+                                      x_thread_buf(i));
+
+                threadwise_gamma_load.Run(gamma_grid_desc_m_k,
+                                          gamma_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          gamma_thread_buf(i));
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+                        x_square_thread_buf(i)(Number<offset_m_k>{}) =
+                            x_thread_buf(i)(Number<offset_m_k>{}) *
+                            x_thread_buf(i)(Number<offset_m_k>{});
+                    });
+                });
+
+                ThreadwiseSumReduce::Reduce(x_thread_buf[i], mean_thread_buf);
+                ThreadwiseSumReduce::Reduce(x_square_thread_buf[i], mean_square_thread_buf);
+
+                if constexpr(i != ThreadBufferNumber - 1)
+                {
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                    threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                             thread_copy_fwd_step_m_k);
+                }
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+
+                BlockwiseSumReduce::Reduce(reduce_work_buf, mean_thread_buf(I));
+                mean_thread_buf(I) = mean_thread_buf(I) / reduce_length;
+
+                block_sync_lds();
+
+                BlockwiseSumReduce::Reduce(reduce_work_buf, mean_square_thread_buf(I));
+                mean_square_thread_buf(I) = mean_square_thread_buf(I) / reduce_length;
+
+                // var(x) = E[x^2] - E[x]^2
+                var_thread_buf(I) =
+                    mean_square_thread_buf(I) - (mean_thread_buf(I) * mean_thread_buf(I));
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // normalize
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
+                            divisor;
+
+                        // gamma & beta
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) *
+                            gamma_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
+
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_beta_load.Run(beta_grid_desc_m_k,
+                                         beta_global_val_buf,
+                                         thread_buffer_desc_m_k,
+                                         make_tuple(I0, I0),
+                                         beta_thread_buf(i));
+
+                if constexpr(i != ThreadBufferNumber - 1)
+                    threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                            thread_copy_fwd_step_m_k);
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // beta
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) +
+                            beta_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
+
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                       make_tuple(I0, I0),
+                                       y_thread_buf(i),
+                                       y_grid_desc_m_k,
+                                       y_global_val_buf);
+
+                if constexpr(i != ThreadBufferNumber - 1)
+                    threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                          thread_copy_fwd_step_m_k);
+            });
+        } // end of sweep once
+        else
+        {
+            for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+            {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_x_load.Run(x_grid_desc_m_k,
+                                          x_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          x_thread_buf(i));
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+
+                    static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                        static_for<0, XSrcVectorSize, 1>{}([&](auto iK) {
+                            constexpr auto offset_m_k =
+                                thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+                            x_square_thread_buf(i)(Number<offset_m_k>{}) =
+                                x_thread_buf(i)(Number<offset_m_k>{}) *
+                                x_thread_buf(i)(Number<offset_m_k>{});
+                        });
+                    });
+
+                    ThreadwiseSumReduce::Reduce(x_thread_buf[i], mean_thread_buf);
+                    ThreadwiseSumReduce::Reduce(x_square_thread_buf[i], mean_square_thread_buf);
+                });
+            }
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+
+                BlockwiseSumReduce::Reduce(reduce_work_buf, mean_thread_buf(I));
+                mean_thread_buf(I) = mean_thread_buf(I) / reduce_length;
+
+                block_sync_lds();
+
+                BlockwiseSumReduce::Reduce(reduce_work_buf, mean_square_thread_buf(I));
+                mean_square_thread_buf(I) = mean_square_thread_buf(I) / reduce_length;
+
+                // var(x) = E[x^2] - E[x]^2
+                var_thread_buf(I) =
+                    mean_square_thread_buf(I) - (mean_thread_buf(I) * mean_thread_buf(I));
+            });
+
+            auto thread_copy_tail_m_k =
+                (num_k_block_tile_iteration - 1) * ThreadBufferNumber * thread_copy_fwd_step_m_k;
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
+            threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
+
+            for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+            {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_x_load.Run(x_grid_desc_m_k,
+                                          x_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          x_thread_buf(i));
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_gamma_load.Run(gamma_grid_desc_m_k,
+                                              gamma_global_val_buf,
+                                              thread_buffer_desc_m_k,
+                                              make_tuple(I0, I0),
+                                              gamma_thread_buf(i));
+
+                    threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                             thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
+                    static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                        static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                            constexpr auto offset_m_k =
+                                thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                            // normalize
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
+                                divisor;
+
+                            // gamma
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                y_thread_buf(iK0)(Number<offset_m_k>{}) *
+                                gamma_thread_buf(iK0)(Number<offset_m_k>{});
+                        });
+                    });
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_beta_load.Run(beta_grid_desc_m_k,
+                                             beta_global_val_buf,
+                                             thread_buffer_desc_m_k,
+                                             make_tuple(I0, I0),
+                                             beta_thread_buf(i));
+                    threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                            thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                        static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                            constexpr auto offset_m_k =
+                                thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                            // beta
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                y_thread_buf(iK0)(Number<offset_m_k>{}) +
+                                beta_thread_buf(iK0)(Number<offset_m_k>{});
+                        });
+                    });
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                           make_tuple(I0, I0),
+                                           y_thread_buf(i),
+                                           y_grid_desc_m_k,
+                                           y_global_val_buf);
+                    threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                          thread_copy_fwd_step_m_k);
+                });
+
+                threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
+                threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                         2 * thread_copy_bwd_step_m_k);
+                threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                        2 * thread_copy_bwd_step_m_k);
+                threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                      2 * thread_copy_bwd_step_m_k);
+            }
+        } // end of sweep twice
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp
new file mode 100644
index 0000000000..e50fb98133
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp
@@ -0,0 +1,195 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_naive_variance.hpp"
+#include "ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_welford_variance.hpp"
+
+namespace ck {
+template <typename GridwiseReduction,
+          typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename ComputeDataType,
+          typename YElementwiseOperation,
+          typename GridDesc_M_K>
+__global__ void kernel_normalization(const GridDesc_M_K x_grid_desc_m_k,
+                                     const GridDesc_M_K gamma_grid_desc_m_k,
+                                     const GridDesc_M_K beta_grid_desc_m_k,
+                                     const GridDesc_M_K y_grid_desc_m_k,
+                                     index_t num_k_block_tile_iteration,
+                                     ComputeDataType epsilon,
+                                     const XDataType* const __restrict__ p_x_global,
+                                     const GammaDataType* const __restrict__ p_gamma_global,
+                                     const BetaDataType* const __restrict__ p_beta_global,
+                                     YDataType* const __restrict__ p_y_global,
+                                     const YElementwiseOperation y_elementwise_op)
+{
+    GridwiseReduction::Run(x_grid_desc_m_k,
+                           gamma_grid_desc_m_k,
+                           beta_grid_desc_m_k,
+                           y_grid_desc_m_k,
+                           num_k_block_tile_iteration,
+                           epsilon,
+                           p_x_global,
+                           p_gamma_global,
+                           p_beta_global,
+                           p_y_global,
+                           y_elementwise_op);
+};
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename ComputeDataType,
+          typename YElementwiseOperation,
+          typename GridDesc_M_K,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcVectorDim,
+          index_t XSrcVectorSize,
+          index_t GammaSrcVectorDim,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorDim,
+          index_t BetaSrcVectorSize,
+          index_t YDstVectorDim,
+          index_t YDstVectorSize,
+          bool UseWelford>
+auto NormalizationKernelSelector(bool isSweepOnce)
+{
+    using GridwiseNormalizationGenericNaive =
+        GridwiseNormalizationNaiveVariance_mk_to_mk<XDataType,
+                                                    GammaDataType,
+                                                    BetaDataType,
+                                                    YDataType,
+                                                    ComputeDataType,
+                                                    YElementwiseOperation,
+                                                    GridDesc_M_K,
+                                                    BlockSize,
+                                                    MThreadClusterSize,
+                                                    KThreadClusterSize,
+                                                    MThreadSliceSize,
+                                                    KThreadSliceSize,
+                                                    XSrcVectorDim,
+                                                    XSrcVectorSize,
+                                                    GammaSrcVectorDim,
+                                                    GammaSrcVectorSize,
+                                                    BetaSrcVectorDim,
+                                                    BetaSrcVectorSize,
+                                                    YDstVectorDim,
+                                                    YDstVectorSize,
+                                                    false>;
+    using GridwiseNormalizationSweepOnceNaive =
+        GridwiseNormalizationNaiveVariance_mk_to_mk<XDataType,
+                                                    GammaDataType,
+                                                    BetaDataType,
+                                                    YDataType,
+                                                    ComputeDataType,
+                                                    YElementwiseOperation,
+                                                    GridDesc_M_K,
+                                                    BlockSize,
+                                                    MThreadClusterSize,
+                                                    KThreadClusterSize,
+                                                    MThreadSliceSize,
+                                                    KThreadSliceSize,
+                                                    XSrcVectorDim,
+                                                    XSrcVectorSize,
+                                                    GammaSrcVectorDim,
+                                                    GammaSrcVectorSize,
+                                                    BetaSrcVectorDim,
+                                                    BetaSrcVectorSize,
+                                                    YDstVectorDim,
+                                                    YDstVectorSize,
+                                                    true>;
+    using GridwiseNormalizationGenericWelford =
+        GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
+                                                      GammaDataType,
+                                                      BetaDataType,
+                                                      YDataType,
+                                                      ComputeDataType,
+                                                      YElementwiseOperation,
+                                                      GridDesc_M_K,
+                                                      BlockSize,
+                                                      MThreadClusterSize,
+                                                      KThreadClusterSize,
+                                                      MThreadSliceSize,
+                                                      KThreadSliceSize,
+                                                      XSrcVectorDim,
+                                                      XSrcVectorSize,
+                                                      GammaSrcVectorDim,
+                                                      GammaSrcVectorSize,
+                                                      BetaSrcVectorDim,
+                                                      BetaSrcVectorSize,
+                                                      YDstVectorDim,
+                                                      YDstVectorSize,
+                                                      false>;
+    using GridwiseNormalizationSweepOnceWelford =
+        GridwiseNormalizationWelfordVariance_mk_to_mk<XDataType,
+                                                      GammaDataType,
+                                                      BetaDataType,
+                                                      YDataType,
+                                                      ComputeDataType,
+                                                      YElementwiseOperation,
+                                                      GridDesc_M_K,
+                                                      BlockSize,
+                                                      MThreadClusterSize,
+                                                      KThreadClusterSize,
+                                                      MThreadSliceSize,
+                                                      KThreadSliceSize,
+                                                      XSrcVectorDim,
+                                                      XSrcVectorSize,
+                                                      GammaSrcVectorDim,
+                                                      GammaSrcVectorSize,
+                                                      BetaSrcVectorDim,
+                                                      BetaSrcVectorSize,
+                                                      YDstVectorDim,
+                                                      YDstVectorSize,
+                                                      true>;
+
+    if constexpr(UseWelford)
+    {
+        return isSweepOnce ? kernel_normalization<GridwiseNormalizationSweepOnceWelford,
+                                                  XDataType,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  YDataType,
+                                                  ComputeDataType,
+                                                  YElementwiseOperation,
+                                                  GridDesc_M_K>
+                           : kernel_normalization<GridwiseNormalizationGenericWelford,
+                                                  XDataType,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  YDataType,
+                                                  ComputeDataType,
+                                                  YElementwiseOperation,
+                                                  GridDesc_M_K>;
+    }
+    else
+    {
+        return isSweepOnce ? kernel_normalization<GridwiseNormalizationSweepOnceNaive,
+                                                  XDataType,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  YDataType,
+                                                  ComputeDataType,
+                                                  YElementwiseOperation,
+                                                  GridDesc_M_K>
+                           : kernel_normalization<GridwiseNormalizationGenericNaive,
+                                                  XDataType,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  YDataType,
+                                                  ComputeDataType,
+                                                  YElementwiseOperation,
+                                                  GridDesc_M_K>;
+    }
+}
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp
new file mode 100644
index 0000000000..80e9a84f96
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp
@@ -0,0 +1,256 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename XDataType,
+          typename ComputeDataType,
+          typename MeanVarDataType,
+          typename XGridDesc_M_K,
+          typename MeanVarGridDesc_M_KBlock,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcVectorDim,
+          index_t XSrcVectorSize>
+struct GridwiseNormalizationSplitK1st
+{
+    static_assert((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadBufferLengths_M_K                = Sequence<MThreadSliceSize, XSrcVectorSize>;
+    static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
+
+    using ThreadBufferLengths_M_1 = Sequence<MThreadSliceSize, 1>;
+    static constexpr auto thread_buffer_desc_m_1 =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}, I1));
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelford<ComputeDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<ComputeDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder,
+                                              false>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr index_t M_BlockTileSize     = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize     = KThreadClusterSize * KThreadSliceSize;
+    static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
+
+    static constexpr auto ThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
+
+    __device__ static int
+    GetKPerThread(int k, int kRaw, int kGridSize, int block_k_cluster_id, int thread_k_cluster_id)
+    {
+        bool is_rightmost_block = block_k_cluster_id == kGridSize - 1;
+
+        if(is_rightmost_block)
+        {
+            int left_kPerBlock  = math::integer_divide_ceil(k, kGridSize);
+            int kRightmostBlock = kRaw - left_kPerBlock * (kGridSize - 1);
+            int kPerThread      = kRightmostBlock < K_BlockTileSize
+                                      ? 0
+                                      : KThreadSliceSize * (kRightmostBlock / K_BlockTileSize);
+            int kPerBlockTail   = kRightmostBlock - kPerThread * KThreadClusterSize;
+
+            if(kPerBlockTail > 0)
+            {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    int thread_max_len =
+                        (thread_k_cluster_id + 1) * XSrcVectorSize + K_BlockTileStepSize * i;
+                    int delta = thread_max_len - kPerBlockTail;
+                    delta     = math::clamp(thread_max_len - kPerBlockTail, 0, XSrcVectorSize);
+                    kPerThread += XSrcVectorSize - delta;
+                });
+            }
+
+            return kPerThread;
+        }
+        else
+        {
+            int kPerBlock = math::integer_divide_ceil(k, kGridSize);
+            return KThreadSliceSize * (kPerBlock / K_BlockTileSize);
+        }
+    }
+
+    // Calculate mean and variance by welford along k dimension
+    __device__ static void Run(const XGridDesc_M_K& x_grid_desc_m_k,
+                               const MeanVarGridDesc_M_KBlock& mean_var_grid_desc_m_kblock,
+                               index_t num_k_block_tile_iteration,
+                               const XDataType* const __restrict__ p_x_global,
+                               MeanVarDataType* const p_mean_global,
+                               MeanVarDataType* const p_variance_global,
+                               int32_t* const p_welford_count_global)
+    {
+        auto x_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    ComputeDataType,
+                                    MThreadSliceSize * XSrcVectorSize,
+                                    true>{};
+            },
+            Number<ThreadBufferNumber>{});
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            var_thread_buf;
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+
+        const index_t k_grid_size        = mean_var_grid_desc_m_kblock.GetLength(I1);
+        const index_t block_m_cluster_id = block_global_id / k_grid_size;
+        const index_t block_k_cluster_id = block_global_id % k_grid_size;
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        const index_t reduceSizePerBlock = K_BlockTileSize * num_k_block_tile_iteration;
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  ComputeDataType,
+                                                                  XGridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(
+                block_m_cluster_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+                block_k_cluster_id * reduceSizePerBlock + thread_k_cluster_id * XSrcVectorSize));
+
+        auto mean_var_count_store_index = make_multi_index(
+            block_m_cluster_id * M_BlockTileSize + thread_m_cluster_id * MThreadSliceSize,
+            block_k_cluster_id);
+
+        auto threadwise_welford_mean_var_store =
+            ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
+                                               MeanVarDataType,
+                                               decltype(thread_buffer_desc_m_1),
+                                               MeanVarGridDesc_M_KBlock,
+                                               PassThroughOp,
+                                               ThreadBufferLengths_M_1,
+                                               Sequence<0, 1>,
+                                               1,
+                                               1,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                mean_var_grid_desc_m_kblock, mean_var_count_store_index, PassThroughOp{});
+
+        constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
+
+        const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x_global, x_grid_desc_m_k.GetElementSpaceSize());
+
+        auto mean_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_mean_global, mean_var_grid_desc_m_kblock.GetElementSpaceSize());
+
+        auto var_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_variance_global, mean_var_grid_desc_m_kblock.GetElementSpaceSize());
+
+        auto threadwise_welford       = ThreadwiseWelford();
+        int kRaw                      = x_grid_desc_m_k.GetTransforms()[I2].GetUpperLengths()[I0];
+        threadwise_welford.max_count_ = GetKPerThread(x_grid_desc_m_k.GetLength(I1),
+                                                      kRaw,
+                                                      k_grid_size,
+                                                      block_k_cluster_id,
+                                                      thread_k_cluster_id);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            mean_thread_buf(I) = type_convert<ComputeDataType>(0.0f);
+            var_thread_buf(I)  = type_convert<ComputeDataType>(0.0f);
+        });
+
+        for(index_t k = 0; k < num_k_block_tile_iteration; ++k)
+        {
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_x_load.Run(x_grid_desc_m_k,
+                                      x_global_val_buf,
+                                      thread_buffer_desc_m_k,
+                                      make_tuple(I0, I0),
+                                      x_thread_buf(i));
+                threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
+            });
+        }
+
+        int welford_count = 0;
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            int count = threadwise_welford.cur_count_;
+            BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
+
+            // The value of count is same for all I
+            if constexpr(I == MThreadSliceSize - 1)
+                welford_count = count;
+        });
+
+        if(thread_k_cluster_id == 0)
+        {
+            threadwise_welford_mean_var_store.Run(thread_buffer_desc_m_1,
+                                                  make_tuple(I0, I0),
+                                                  mean_thread_buf,
+                                                  mean_var_grid_desc_m_kblock,
+                                                  mean_global_val_buf);
+
+            threadwise_welford_mean_var_store.Run(thread_buffer_desc_m_1,
+                                                  make_tuple(I0, I0),
+                                                  var_thread_buf,
+                                                  mean_var_grid_desc_m_kblock,
+                                                  var_global_val_buf);
+
+            if(block_m_cluster_id == 0 && thread_m_cluster_id == 0)
+                p_welford_count_global[block_k_cluster_id] = welford_count;
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_2nd.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_2nd.hpp
new file mode 100644
index 0000000000..136ac94e7f
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_2nd.hpp
@@ -0,0 +1,418 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename MeanVarDataType,
+          typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename ComputeDataType,
+          typename YElementwiseOperation,
+          typename MeanVarGridDesc_M_KBlock,
+          typename CountGridDesc_M_KBlock,
+          typename XYGammaBetaGridDesc_M_K,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcVectorDim,
+          index_t XSrcVectorSize,
+          index_t GammaSrcVectorDim,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorDim,
+          index_t BetaSrcVectorSize,
+          index_t YDstVectorDim,
+          index_t YDstVectorSize>
+struct GridwiseNormalizationSplitK2nd
+{
+    static_assert((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert((YDstVectorDim == 0 && MThreadSliceSize % YDstVectorSize == 0) ||
+                      (YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert(XSrcVectorSize == YDstVectorSize);
+    static_assert(XSrcVectorSize == GammaSrcVectorSize);
+    static_assert(XSrcVectorSize == BetaSrcVectorSize);
+
+    static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadBufferLengths_M_K                = Sequence<MThreadSliceSize, XSrcVectorSize>;
+    static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
+
+    using ThreadBufferLengths_M_1 = Sequence<MThreadSliceSize, 1>;
+    static constexpr auto thread_buffer_desc_m_1 =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{}, I1));
+
+    using ThreadWelfordSrcDesc_M_1 = decltype(thread_buffer_desc_m_1);
+    using ThreadWelfordDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelfordMerge<ComputeDataType, ThreadWelfordSrcDesc_M_1, ThreadWelfordDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<ComputeDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder>;
+
+    using PassThroughOp = tensor_operation::element_wise::PassThrough;
+
+    static constexpr index_t M_BlockTileSize     = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize     = KThreadClusterSize * KThreadSliceSize;
+    static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
+
+    static constexpr auto ThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
+
+    __device__ static void Run(const MeanVarGridDesc_M_KBlock& mean_var_grid_desc_m_kblock,
+                               const CountGridDesc_M_KBlock& count_grid_desc_m_kblock,
+                               const XYGammaBetaGridDesc_M_K& x_grid_desc_m_k,
+                               const XYGammaBetaGridDesc_M_K& gamma_grid_desc_m_k,
+                               const XYGammaBetaGridDesc_M_K& beta_grid_desc_m_k,
+                               const XYGammaBetaGridDesc_M_K& y_grid_desc_m_k,
+                               index_t num_k_mean_var_count_iteration,
+                               index_t num_k_block_tile_iteration,
+                               index_t k_grid_size,
+                               ComputeDataType epsilon,
+                               const MeanVarDataType* const p_mean_global,
+                               const MeanVarDataType* const p_variance_global,
+                               const int32_t* const p_welford_count_global,
+                               const XDataType* const __restrict__ p_x_global,
+                               const GammaDataType* const __restrict__ p_gamma_global,
+                               const BetaDataType* const __restrict__ p_beta_global,
+                               YDataType* const __restrict__ p_y_global,
+                               const YElementwiseOperation y_elementwise_op)
+    {
+        // Thread/Block id
+        const index_t thread_local_id    = get_thread_local_1d_id();
+        const index_t block_global_id    = get_block_1d_id();
+        const index_t block_m_cluster_id = block_global_id / k_grid_size;
+        const index_t block_k_cluster_id = block_global_id % k_grid_size;
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        // Global Memory
+        const auto mean_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_mean_global, mean_var_grid_desc_m_kblock.GetElementSpaceSize());
+
+        const auto var_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_variance_global, mean_var_grid_desc_m_kblock.GetElementSpaceSize());
+
+        const auto welford_count_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_welford_count_global, count_grid_desc_m_kblock.GetElementSpaceSize());
+
+        const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x_global, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto gamma_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_gamma_global, gamma_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto beta_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_beta_global, beta_grid_desc_m_k.GetElementSpaceSize());
+
+        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
+
+        // VGPR
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            in_mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            in_var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true>
+            in_welford_count_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            var_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, int32_t, MThreadSliceSize, true>
+            welford_count_thread_buf;
+
+        auto x_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    ComputeDataType,
+                                    MThreadSliceSize * XSrcVectorSize,
+                                    true>{};
+            },
+            Number<ThreadBufferNumber>{});
+
+        auto gamma_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    ComputeDataType,
+                                    MThreadSliceSize * GammaSrcVectorSize,
+                                    true>{};
+            },
+            Number<ThreadBufferNumber>{});
+
+        auto& beta_thread_buf = gamma_thread_buf;
+        auto& y_thread_buf    = x_thread_buf;
+
+        // IO
+        auto threadwise_mean_var_load_m_kblock =
+            ThreadwiseTensorSliceTransfer_v2<MeanVarDataType,
+                                             ComputeDataType,
+                                             MeanVarGridDesc_M_KBlock,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             Sequence<0, 1>,
+                                             1,
+                                             1,
+                                             1,
+                                             true>(
+                mean_var_grid_desc_m_kblock,
+                make_multi_index(block_m_cluster_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id));
+
+        auto threadwise_count_load_m_kblock =
+            ThreadwiseTensorSliceTransfer_v2<int32_t,
+                                             int32_t,
+                                             CountGridDesc_M_KBlock,
+                                             decltype(thread_buffer_desc_m_1),
+                                             ThreadBufferLengths_M_1,
+                                             Sequence<0, 1>,
+                                             1,
+                                             1,
+                                             1,
+                                             true>(
+                count_grid_desc_m_kblock,
+                make_multi_index(block_m_cluster_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id));
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  ComputeDataType,
+                                                                  XYGammaBetaGridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(block_m_cluster_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize,
+                             block_k_cluster_id * K_BlockTileSize * num_k_block_tile_iteration +
+                                 thread_k_cluster_id * XSrcVectorSize));
+
+        auto threadwise_gamma_load =
+            ThreadwiseTensorSliceTransfer_v2<GammaDataType,
+                                             ComputeDataType,
+                                             XYGammaBetaGridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_k),
+                                             ThreadBufferLengths_M_K,
+                                             ThreadBufferDimAccessOrder,
+                                             GammaSrcVectorDim,
+                                             GammaSrcVectorSize,
+                                             1,
+                                             true>(
+                gamma_grid_desc_m_k,
+                make_multi_index(block_m_cluster_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 block_k_cluster_id * K_BlockTileSize * num_k_block_tile_iteration +
+                                     thread_k_cluster_id * GammaSrcVectorSize));
+
+        auto threadwise_beta_load =
+            ThreadwiseTensorSliceTransfer_v2<BetaDataType,
+                                             ComputeDataType,
+                                             XYGammaBetaGridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_k),
+                                             ThreadBufferLengths_M_K,
+                                             ThreadBufferDimAccessOrder,
+                                             BetaSrcVectorDim,
+                                             BetaSrcVectorSize,
+                                             1,
+                                             true>(
+                beta_grid_desc_m_k,
+                make_multi_index(block_m_cluster_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 block_k_cluster_id * K_BlockTileSize * num_k_block_tile_iteration +
+                                     thread_k_cluster_id * BetaSrcVectorSize));
+
+        auto threadwise_y_store =
+            ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
+                                               YDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               XYGammaBetaGridDesc_M_K,
+                                               YElementwiseOperation,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               YDstVectorDim,
+                                               YDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                y_grid_desc_m_k,
+                make_multi_index(block_m_cluster_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 block_k_cluster_id * K_BlockTileSize * num_k_block_tile_iteration +
+                                     thread_k_cluster_id * YDstVectorSize),
+                y_elementwise_op);
+
+        // step1: Merge mean and variance
+        constexpr auto mean_var_count_thread_copy_step_I0_k =
+            make_multi_index(I0, KThreadClusterSize);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            mean_thread_buf(I)          = type_convert<ComputeDataType>(0.0f);
+            var_thread_buf(I)           = type_convert<ComputeDataType>(0.0f);
+            welford_count_thread_buf(I) = 0;
+        });
+
+        for(index_t k = 0; k < num_k_mean_var_count_iteration; ++k)
+        {
+            threadwise_mean_var_load_m_kblock.Run(mean_var_grid_desc_m_kblock,
+                                                  mean_global_val_buf,
+                                                  thread_buffer_desc_m_1,
+                                                  make_tuple(I0, I0),
+                                                  in_mean_thread_buf);
+
+            threadwise_mean_var_load_m_kblock.Run(mean_var_grid_desc_m_kblock,
+                                                  var_global_val_buf,
+                                                  thread_buffer_desc_m_1,
+                                                  make_tuple(I0, I0),
+                                                  in_var_thread_buf);
+
+            threadwise_count_load_m_kblock.Run(count_grid_desc_m_kblock,
+                                               welford_count_global_val_buf,
+                                               thread_buffer_desc_m_1,
+                                               make_tuple(I0, I0),
+                                               in_welford_count_thread_buf);
+
+            ThreadwiseWelford::Run(in_mean_thread_buf,
+                                   in_var_thread_buf,
+                                   in_welford_count_thread_buf,
+                                   mean_thread_buf,
+                                   var_thread_buf,
+                                   welford_count_thread_buf);
+
+            threadwise_mean_var_load_m_kblock.MoveSrcSliceWindow(
+                mean_var_grid_desc_m_kblock, mean_var_count_thread_copy_step_I0_k);
+            threadwise_count_load_m_kblock.MoveSrcSliceWindow(count_grid_desc_m_kblock,
+                                                              mean_var_count_thread_copy_step_I0_k);
+        }
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            if constexpr(I > 0)
+                block_sync_lds();
+
+            BlockwiseWelford::Run(
+                mean_thread_buf(I), var_thread_buf(I), welford_count_thread_buf(I));
+        });
+
+        // step2: normalization
+        constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
+
+        for(index_t k = 0; k < num_k_block_tile_iteration; ++k)
+        {
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_x_load.Run(x_grid_desc_m_k,
+                                      x_global_val_buf,
+                                      thread_buffer_desc_m_k,
+                                      make_tuple(I0, I0),
+                                      x_thread_buf(i));
+                threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+            });
+
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_gamma_load.Run(gamma_grid_desc_m_k,
+                                          gamma_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          gamma_thread_buf(i));
+
+                threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                         thread_copy_fwd_step_m_k);
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // normalize
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
+                            divisor;
+
+                        // gamma
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) *
+                            gamma_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
+
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_beta_load.Run(beta_grid_desc_m_k,
+                                         beta_global_val_buf,
+                                         thread_buffer_desc_m_k,
+                                         make_tuple(I0, I0),
+                                         beta_thread_buf(i));
+                threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                        thread_copy_fwd_step_m_k);
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // beta
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) +
+                            beta_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
+
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                       make_tuple(I0, I0),
+                                       y_thread_buf(i),
+                                       y_grid_desc_m_k,
+                                       y_global_val_buf);
+                threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_fwd_step_m_k);
+            });
+        } // end for (normalization)
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_welford_variance.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_welford_variance.hpp
new file mode 100644
index 0000000000..ff9712276c
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_welford_variance.hpp
@@ -0,0 +1,461 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_welford.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+// Y = Normalization(X, Beta, Gamma)
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename ComputeDataType,
+          typename YElementwiseOperation,
+          typename GridDesc_M_K,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t XSrcVectorDim,
+          index_t XSrcVectorSize,
+          index_t GammaSrcVectorDim,
+          index_t GammaSrcVectorSize,
+          index_t BetaSrcVectorDim,
+          index_t BetaSrcVectorSize,
+          index_t YDstVectorDim,
+          index_t YDstVectorSize,
+          bool SweepOnce>
+struct GridwiseNormalizationWelfordVariance_mk_to_mk
+{
+    static_assert((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                      (XSrcVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert((YDstVectorDim == 0 && MThreadSliceSize % YDstVectorSize == 0) ||
+                      (YDstVectorDim == 1 && KThreadSliceSize % YDstVectorSize == 0),
+                  "Invalid thread slice sizes and/or vector sizes configuration, please check!");
+
+    static_assert(XSrcVectorSize == YDstVectorSize);
+    static_assert(XSrcVectorSize == GammaSrcVectorSize);
+    static_assert(XSrcVectorSize == BetaSrcVectorSize);
+
+    static constexpr bool reorder_thread_cluster = (XSrcVectorDim == 0);
+
+    using ThreadClusterLengths_M_K = Sequence<MThreadClusterSize, KThreadClusterSize>;
+
+    using ThreadBufferDimAccessOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    using ThreadClusterArrangeOrder =
+        typename conditional<reorder_thread_cluster, Sequence<1, 0>, Sequence<0, 1>>::type;
+
+    static constexpr auto thread_cluster_desc =
+        make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
+
+    using ThreadBufferLengths_M_K                = Sequence<MThreadSliceSize, XSrcVectorSize>;
+    static constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
+
+    using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
+    using ThreadReduceDstDesc_M =
+        decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
+
+    using ThreadwiseWelford =
+        ThreadwiseWelford<ComputeDataType, ThreadReduceSrcDesc_M_K, ThreadReduceDstDesc_M>;
+
+    using BlockwiseWelford = BlockwiseWelford<ComputeDataType,
+                                              BlockSize,
+                                              ThreadClusterLengths_M_K,
+                                              ThreadClusterArrangeOrder>;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr index_t M_BlockTileSize     = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize     = KThreadClusterSize * KThreadSliceSize;
+    static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
+
+    static constexpr auto ThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
+
+    __device__ static int GetKPerThread(const GridDesc_M_K& x_grid_desc_m_k,
+                                        int thread_k_cluster_id)
+    {
+        // FIXME: Should not hack the transform from deviceOP
+        int kPerBlock = x_grid_desc_m_k.GetTransforms()[I2].GetUpperLengths()[I0];
+        int kPerThread =
+            kPerBlock < K_BlockTileSize ? 0 : KThreadSliceSize * (kPerBlock / K_BlockTileSize);
+        int kPerBlockTail = kPerBlock - kPerThread * KThreadClusterSize;
+
+        if(kPerBlockTail > 0)
+        {
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                int thread_max_len =
+                    (thread_k_cluster_id + 1) * XSrcVectorSize + K_BlockTileStepSize * i;
+                int delta = thread_max_len - kPerBlockTail;
+                delta     = math::clamp(thread_max_len - kPerBlockTail, 0, XSrcVectorSize);
+                kPerThread += XSrcVectorSize - delta;
+            });
+        }
+
+        return kPerThread;
+    }
+
+    __device__ static void Run(const GridDesc_M_K& x_grid_desc_m_k,
+                               const GridDesc_M_K& gamma_grid_desc_m_k,
+                               const GridDesc_M_K& beta_grid_desc_m_k,
+                               const GridDesc_M_K& y_grid_desc_m_k,
+                               index_t num_k_block_tile_iteration,
+                               ComputeDataType epsilon,
+                               const XDataType* const __restrict__ p_x_global,
+                               const GammaDataType* const __restrict__ p_gamma_global,
+                               const BetaDataType* const __restrict__ p_beta_global,
+                               YDataType* const __restrict__ p_y_global,
+                               const YElementwiseOperation y_elementwise_op)
+    {
+        auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
+
+        auto x_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    ComputeDataType,
+                                    MThreadSliceSize * XSrcVectorSize,
+                                    true>{};
+            },
+            Number<ThreadBufferNumber>{});
+
+        auto gamma_thread_buf = generate_tuple(
+            [&](auto) {
+                return StaticBuffer<AddressSpaceEnum::Vgpr,
+                                    ComputeDataType,
+                                    MThreadSliceSize * GammaSrcVectorSize,
+                                    true>{};
+            },
+            Number<ThreadBufferNumber>{});
+
+        auto& beta_thread_buf = gamma_thread_buf;
+        auto& y_thread_buf    = x_thread_buf;
+
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            mean_thread_buf;
+        StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MThreadSliceSize, true>
+            var_thread_buf;
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+
+        const auto thread_cluster_idx =
+            thread_cluster_desc.CalculateBottomIndex(make_multi_index(thread_local_id));
+
+        const auto thread_m_cluster_id = thread_cluster_idx[I0];
+        const auto thread_k_cluster_id = thread_cluster_idx[I1];
+
+        auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
+                                                                  ComputeDataType,
+                                                                  GridDesc_M_K,
+                                                                  decltype(thread_buffer_desc_m_k),
+                                                                  ThreadBufferLengths_M_K,
+                                                                  ThreadBufferDimAccessOrder,
+                                                                  XSrcVectorDim,
+                                                                  XSrcVectorSize,
+                                                                  1,
+                                                                  true>(
+            x_grid_desc_m_k,
+            make_multi_index(block_global_id * M_BlockTileSize +
+                                 thread_m_cluster_id * MThreadSliceSize,
+                             thread_k_cluster_id * XSrcVectorSize));
+
+        auto threadwise_gamma_load =
+            ThreadwiseTensorSliceTransfer_v2<GammaDataType,
+                                             ComputeDataType,
+                                             GridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_k),
+                                             ThreadBufferLengths_M_K,
+                                             ThreadBufferDimAccessOrder,
+                                             GammaSrcVectorDim,
+                                             GammaSrcVectorSize,
+                                             1,
+                                             true>(
+                gamma_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * GammaSrcVectorSize));
+
+        auto threadwise_beta_load =
+            ThreadwiseTensorSliceTransfer_v2<BetaDataType,
+                                             ComputeDataType,
+                                             GridDesc_M_K,
+                                             decltype(thread_buffer_desc_m_k),
+                                             ThreadBufferLengths_M_K,
+                                             ThreadBufferDimAccessOrder,
+                                             BetaSrcVectorDim,
+                                             BetaSrcVectorSize,
+                                             1,
+                                             true>(
+                beta_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * BetaSrcVectorSize));
+
+        auto threadwise_y_store =
+            ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
+                                               YDataType,
+                                               decltype(thread_buffer_desc_m_k),
+                                               GridDesc_M_K,
+                                               YElementwiseOperation,
+                                               ThreadBufferLengths_M_K,
+                                               ThreadBufferDimAccessOrder,
+                                               YDstVectorDim,
+                                               YDstVectorSize,
+                                               InMemoryDataOperationEnum::Set,
+                                               1,
+                                               true>(
+                y_grid_desc_m_k,
+                make_multi_index(block_global_id * M_BlockTileSize +
+                                     thread_m_cluster_id * MThreadSliceSize,
+                                 thread_k_cluster_id * YDstVectorSize),
+                y_elementwise_op);
+
+        constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
+        constexpr auto thread_copy_bwd_step_m_k =
+            make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize);
+
+        const auto x_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_x_global, x_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto gamma_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_gamma_global, gamma_grid_desc_m_k.GetElementSpaceSize());
+
+        const auto beta_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_beta_global, beta_grid_desc_m_k.GetElementSpaceSize());
+
+        auto threadwise_welford       = ThreadwiseWelford();
+        threadwise_welford.max_count_ = GetKPerThread(x_grid_desc_m_k, thread_k_cluster_id);
+
+        static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+            mean_thread_buf(I) = type_convert<ComputeDataType>(0.0f);
+            var_thread_buf(I)  = type_convert<ComputeDataType>(0.0f);
+        });
+
+        // Separate sweep once and sweep twice pipeline
+        if constexpr(SweepOnce)
+        {
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_x_load.Run(x_grid_desc_m_k,
+                                      x_global_val_buf,
+                                      thread_buffer_desc_m_k,
+                                      make_tuple(I0, I0),
+                                      x_thread_buf(i));
+
+                threadwise_gamma_load.Run(gamma_grid_desc_m_k,
+                                          gamma_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          gamma_thread_buf(i));
+
+                threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
+
+                if constexpr(i != ThreadBufferNumber - 1)
+                {
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                    threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                             thread_copy_fwd_step_m_k);
+                }
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+
+                int count = threadwise_welford.cur_count_;
+                BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // normalize
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
+                            divisor;
+
+                        // gamma & beta
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) *
+                            gamma_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
+
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_beta_load.Run(beta_grid_desc_m_k,
+                                         beta_global_val_buf,
+                                         thread_buffer_desc_m_k,
+                                         make_tuple(I0, I0),
+                                         beta_thread_buf(i));
+
+                if constexpr(i != ThreadBufferNumber - 1)
+                    threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                            thread_copy_fwd_step_m_k);
+            });
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                    static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                        constexpr auto offset_m_k =
+                            thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                        // beta
+                        y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) +
+                            beta_thread_buf(iK0)(Number<offset_m_k>{});
+                    });
+                });
+            });
+
+            static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                       make_tuple(I0, I0),
+                                       y_thread_buf(i),
+                                       y_grid_desc_m_k,
+                                       y_global_val_buf);
+
+                if constexpr(i != ThreadBufferNumber - 1)
+                    threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                          thread_copy_fwd_step_m_k);
+            });
+        } // end of sweep once
+        else
+        {
+            for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+            {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_x_load.Run(x_grid_desc_m_k,
+                                          x_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          x_thread_buf(i));
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                    threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
+                });
+            }
+
+            static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
+                if constexpr(I > 0)
+                    block_sync_lds();
+
+                int count = threadwise_welford.cur_count_;
+                BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
+            });
+
+            auto thread_copy_tail_m_k =
+                (num_k_block_tile_iteration - 1) * ThreadBufferNumber * thread_copy_fwd_step_m_k;
+
+            threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
+            threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
+            threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_tail_m_k);
+            threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_tail_m_k);
+
+            for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
+            {
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_x_load.Run(x_grid_desc_m_k,
+                                          x_global_val_buf,
+                                          thread_buffer_desc_m_k,
+                                          make_tuple(I0, I0),
+                                          x_thread_buf(i));
+                    threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_gamma_load.Run(gamma_grid_desc_m_k,
+                                              gamma_global_val_buf,
+                                              thread_buffer_desc_m_k,
+                                              make_tuple(I0, I0),
+                                              gamma_thread_buf(i));
+
+                    threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                             thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    auto divisor = 1 / ck::math::sqrt(var_thread_buf(iM) + epsilon);
+                    static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                        static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                            constexpr auto offset_m_k =
+                                thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                            // normalize
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                (x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
+                                divisor;
+
+                            // gamma
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                y_thread_buf(iK0)(Number<offset_m_k>{}) *
+                                gamma_thread_buf(iK0)(Number<offset_m_k>{});
+                        });
+                    });
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_beta_load.Run(beta_grid_desc_m_k,
+                                             beta_global_val_buf,
+                                             thread_buffer_desc_m_k,
+                                             make_tuple(I0, I0),
+                                             beta_thread_buf(i));
+                    threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                            thread_copy_fwd_step_m_k);
+                });
+
+                static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
+                    static_for<0, ThreadBufferNumber, 1>{}([&](auto iK0) {
+                        static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
+                            constexpr auto offset_m_k =
+                                thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
+
+                            // beta
+                            y_thread_buf(iK0)(Number<offset_m_k>{}) =
+                                y_thread_buf(iK0)(Number<offset_m_k>{}) +
+                                beta_thread_buf(iK0)(Number<offset_m_k>{});
+                        });
+                    });
+                });
+
+                static_for<0, ThreadBufferNumber, 1>{}([&](auto i) {
+                    threadwise_y_store.Run(thread_buffer_desc_m_k,
+                                           make_tuple(I0, I0),
+                                           y_thread_buf(i),
+                                           y_grid_desc_m_k,
+                                           y_global_val_buf);
+                    threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                          thread_copy_fwd_step_m_k);
+                });
+
+                threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
+                threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
+                                                         2 * thread_copy_bwd_step_m_k);
+                threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
+                                                        2 * thread_copy_bwd_step_m_k);
+                threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k,
+                                                      2 * thread_copy_bwd_step_m_k);
+            }
+        } // end of sweep twice
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp b/include/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp
new file mode 100644
index 0000000000..c6eecc067d
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp
@@ -0,0 +1,99 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_functions_accumulate.hpp"
+
+namespace ck {
+
+// Assume
+//  1) SrcDesc is known at compile-time
+//  2) DstDesc is known at compile-time
+//  3) SrcBuffer is static buffer
+//  4) DstBuffer is static buffer
+template <typename AccDataType,
+          typename SrcThreadDesc_M_K,
+          typename DstThreadDesc_M,
+          typename OpReduce,
+          bool PropagateNan,
+          typename Accumulation =
+              detail::AccumulateWithNanCheck<PropagateNan, OpReduce, AccDataType>>
+struct ThreadwiseReduction
+{
+    static constexpr auto src_thread_desc_m_k = SrcThreadDesc_M_K{};
+    static constexpr auto dst_thread_desc_m   = DstThreadDesc_M{};
+
+    static constexpr auto src_length_m = src_thread_desc_m_k.GetLength(Number<0>{});
+    static constexpr auto src_length_k = src_thread_desc_m_k.GetLength(Number<1>{});
+    static constexpr auto dst_length_m = dst_thread_desc_m.GetLength(Number<0>{});
+
+    static_assert(src_length_m == dst_length_m, "lengths of source and dst buffer must match!");
+
+    using Op = OpReduce;
+
+    template <typename SrcBufferType, typename DstBufferType>
+    __device__ static void Reduce(const SrcBufferType& src_buf, DstBufferType& dst_buf)
+    {
+        static_for<0, src_length_m, 1>{}([&](auto iM) {
+            constexpr index_t out_offset = dst_thread_desc_m.CalculateOffset(make_tuple(iM));
+
+            static_for<0, src_length_k, 1>{}([&](auto iK) {
+                constexpr auto offset = src_thread_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                Accumulation::Calculate(dst_buf(Number<out_offset>{}), src_buf[Number<offset>{}]);
+            });
+        });
+    };
+};
+
+// Assume
+//  1) SrcDesc is known at compile-time
+//  2) DstDesc is known at compile-time
+//  3) SrcBuffer is static buffer
+//  4) DstBuffer is static buffer
+template <
+    typename AccDataType,
+    typename IndexDataType,
+    typename SrcThreadDesc_M_K,
+    typename DstThreadDesc_M,
+    typename OpReduce,
+    bool PropagateNan,
+    typename Accumulation =
+        detail::AccumulateWithIndexAndNanCheck<PropagateNan, OpReduce, AccDataType, IndexDataType>>
+struct ThreadwiseReductionWithIndex
+{
+    static constexpr auto src_thread_desc_m_k = SrcThreadDesc_M_K{};
+    static constexpr auto dst_thread_desc_m   = DstThreadDesc_M{};
+
+    static constexpr auto src_length_m = src_thread_desc_m_k.GetLength(Number<0>{});
+    static constexpr auto src_length_k = src_thread_desc_m_k.GetLength(Number<1>{});
+    static constexpr auto dst_length_m = dst_thread_desc_m.GetLength(Number<0>{});
+
+    static_assert(src_length_m == dst_length_m, "lengths of source and dst buffer must match!");
+
+    template <typename SrcValueBufferType,
+              typename SrcIndexBufferType,
+              typename DstValueBufferType,
+              typename DstIndexBufferType>
+    __device__ static void Reduce(const SrcValueBufferType& src_val_buf,
+                                  const SrcIndexBufferType& src_idx_buf,
+                                  DstValueBufferType& dst_val_buf,
+                                  DstIndexBufferType& dst_idx_buf)
+    {
+        static_for<0, src_length_m, 1>{}([&](auto iM) {
+            constexpr index_t out_offset = dst_thread_desc_m.CalculateOffset(make_tuple(iM));
+
+            static_for<0, src_length_k, 1>{}([&](auto iK) {
+                constexpr auto offset = src_thread_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                Accumulation::Calculate(dst_val_buf(Number<out_offset>{}),
+                                        src_val_buf[Number<offset>{}],
+                                        dst_idx_buf(Number<out_offset>{}),
+                                        src_idx_buf[Number<offset>{}]);
+            });
+        });
+    };
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp
new file mode 100644
index 0000000000..44730d551c
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp
@@ -0,0 +1,224 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/math.hpp"
+
+namespace ck {
+
+// C[TM0, TM1, TN0, TN1] += A[TK, TM0, TM1] * B[TK, TN0, TN1]
+//   Tensor element can be vectorized data
+// Assume:
+//   1. AThreadDesc_TK0_TM0_TM1_TK1, BThreadDesc_TK0_TN0_TN1_TK1, CThreadDesc_TM0_TM1_TN0_TN1 are
+//   known at compile-time
+//   2. AOriginIdx, BOriginIdx, COriginIdx are known at compile-time
+template <typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename AThreadDesc_TK0_TM0_TM1_TK1,
+          typename BThreadDesc_TK0_TN0_TN1_TK1,
+          typename CThreadDesc_TM0_TM1_TN0_TN1,
+          typename TKLengths,
+          typename TMLengths,
+          typename TNLengths,
+          typename enable_if<AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
+                                 BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
+                                 CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseGemmDl_km0m1_kn0n1_m0m1n0n1
+{
+    __device__ constexpr ThreadwiseGemmDl_km0m1_kn0n1_m0m1n0n1()
+    {
+        static_assert(AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
+                          BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
+                          CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        // TODO: sanity-check: compare AThreadDesc_TK0_TM0_TM1_TK1, BThreadDesc_TK0_TN0_TN1_TK1,
+        // CThreadDesc_TM0_TM1_TN0_TN1 Size with KLenghts, TMLengths and TNLengths
+
+        // TODO remove this restriction
+        static_assert(TKLengths::Size() == 1 && TMLengths::Size() == 2 && TNLengths::Size() == 2,
+                      "wrong!");
+    }
+
+    template <typename ABuffer,
+              typename AOriginIdx,
+              typename BBuffer,
+              typename BOriginIdx,
+              typename CBuffer,
+              typename COriginIdx>
+    __device__ static void Run(const ABuffer& a_buf,
+                               AOriginIdx,
+                               const BBuffer& b_buf,
+                               BOriginIdx,
+                               CBuffer& c_buf,
+                               COriginIdx)
+    {
+        static_assert(is_known_at_compile_time<remove_cvref_t<AOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<BOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<COriginIdx>>::value,
+                      "wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename ABuffer::type>, remove_cvref_t<FloatA>>::value &&
+            is_same<remove_cvref_t<typename BBuffer::type>, remove_cvref_t<FloatB>>::value &&
+            is_same<remove_cvref_t<typename CBuffer::type>, remove_cvref_t<FloatC>>::value &&
+            "wrong! inconsistent type");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        constexpr auto TK  = TKLengths{}[I0];
+        constexpr auto TM0 = TMLengths{}[I0];
+        constexpr auto TM1 = TMLengths{}[I1];
+        constexpr auto TN0 = TNLengths{}[I0];
+        constexpr auto TN1 = TNLengths{}[I1];
+
+        constexpr auto a_origin_idx = to_multi_index(AOriginIdx{});
+        constexpr auto b_origin_idx = to_multi_index(BOriginIdx{});
+        constexpr auto c_origin_idx = to_multi_index(COriginIdx{});
+
+        static_for<0, TK, 1>{}([&](auto tk) {
+            static_for<0, TM0, 1>{}([&](auto tm0) {
+                static_for<0, TM1, 1>{}([&](auto tm1) {
+                    static_for<0, TN0, 1>{}([&](auto tn0) {
+                        static_for<0, TN1, 1>{}([&](auto tn1) {
+                            constexpr index_t a_offset =
+                                AThreadDesc_TK0_TM0_TM1_TK1{}.CalculateOffset(
+                                    a_origin_idx + make_multi_index(tk, tm0, tm1));
+                            constexpr index_t b_offset =
+                                BThreadDesc_TK0_TN0_TN1_TK1{}.CalculateOffset(
+                                    b_origin_idx + make_multi_index(tk, tn0, tn1));
+                            constexpr index_t c_offset =
+                                CThreadDesc_TM0_TM1_TN0_TN1{}.CalculateOffset(
+                                    c_origin_idx + make_multi_index(tm0, tm1, tn0, tn1));
+
+                            inner_product<FloatA, FloatB, FloatC>(a_buf[Number<a_offset>{}],
+                                                                  b_buf[Number<b_offset>{}],
+                                                                  c_buf(Number<c_offset>{}));
+                        });
+                    });
+                });
+            });
+        });
+    }
+};
+
+// C[TM0, TM1, TN0, TN1] += A[TK0, TM0, TM1, TK1] * B[TK0, TN0, TN1, TK1]
+//   Tensor element can be vectorized data
+// Assume:
+//   1. AThreadDesc_TK0_TM0_TM1_TK1, BThreadDesc_TK0_TN0_TN1_TK1, CThreadDesc_TM0_TM1_TN0_TN1 are
+//      known at compile-time
+//   2. AOriginIdx, BOriginIdx, COriginIdx are known at compile-time
+template <typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename AThreadDesc_TK0_TM0_TM1_TK1,
+          typename BThreadDesc_TK0_TN0_TN1_TK1,
+          typename CThreadDesc_TM0_TM1_TN0_TN1,
+          typename TKLengths,
+          typename TMLengths,
+          typename TNLengths,
+          typename enable_if<AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
+                                 BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
+                                 CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseContractionDl_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1
+{
+    __device__ constexpr ThreadwiseContractionDl_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1()
+    {
+        static_assert(AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
+                          BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
+                          CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        // TODO: sanity-check: compare AThreadDesc_TK0_TM0_TM1_TK1, BThreadDesc_TK0_TN0_TN1_TK1,
+        // CThreadDesc_TM0_TM1_TN0_TN1 Size with KLenghts, TMLengths and TNLengths
+
+        // TODO remove this restriction
+        static_assert(TKLengths::Size() == 2 && TMLengths::Size() == 2 && TNLengths::Size() == 2,
+                      "wrong!");
+    }
+
+    template <typename ABuffer,
+              typename AOriginIdx,
+              typename BBuffer,
+              typename BOriginIdx,
+              typename CBuffer,
+              typename COriginIdx>
+    __device__ static void Run(const ABuffer& a_buf,
+                               AOriginIdx,
+                               const BBuffer& b_buf,
+                               BOriginIdx,
+                               CBuffer& c_buf,
+                               COriginIdx)
+    {
+        static_assert(is_known_at_compile_time<remove_cvref_t<AOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<BOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<COriginIdx>>::value,
+                      "wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename ABuffer::type>, remove_cvref_t<FloatA>>::value &&
+            is_same<remove_cvref_t<typename BBuffer::type>, remove_cvref_t<FloatB>>::value &&
+            is_same<remove_cvref_t<typename CBuffer::type>, remove_cvref_t<FloatC>>::value &&
+            "wrong! inconsistent type");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        constexpr index_t TK0 = TKLengths{}[I0];
+        constexpr index_t TK1 = TKLengths{}[I1];
+        constexpr index_t TM0 = TMLengths{}[I0];
+        constexpr index_t TM1 = TMLengths{}[I1];
+        constexpr index_t TN0 = TNLengths{}[I0];
+        constexpr index_t TN1 = TNLengths{}[I1];
+
+        constexpr auto a_origin_idx = to_multi_index(AOriginIdx{});
+        constexpr auto b_origin_idx = to_multi_index(BOriginIdx{});
+        constexpr auto c_origin_idx = to_multi_index(COriginIdx{});
+
+        static_for<0, TK0, 1>{}([&](auto tk0) {
+            static_for<0, TM0, 1>{}([&](auto tm0) {
+                static_for<0, TM1, 1>{}([&](auto tm1) {
+                    static_for<0, TN0, 1>{}([&](auto tn0) {
+                        static_for<0, TN1, 1>{}([&](auto tn1) {
+                            vector_type<FloatA, TK1> a_vec;
+                            vector_type<FloatB, TK1> b_vec;
+
+                            static_for<0, TK1, 1>{}([&](auto tk1) {
+                                constexpr index_t a_offset =
+                                    AThreadDesc_TK0_TM0_TM1_TK1{}.CalculateOffset(
+                                        a_origin_idx + make_multi_index(tk0, tm0, tm1, tk1));
+
+                                constexpr index_t b_offset =
+                                    BThreadDesc_TK0_TN0_TN1_TK1{}.CalculateOffset(
+                                        b_origin_idx + make_multi_index(tk0, tn0, tn1, tk1));
+
+                                a_vec.template AsType<FloatA>()(tk1) = a_buf[Number<a_offset>{}];
+                                b_vec.template AsType<FloatB>()(tk1) = b_buf[Number<b_offset>{}];
+                            });
+
+                            using a_vector_t = typename vector_type<FloatA, TK1>::type;
+                            using b_vector_t = typename vector_type<FloatB, TK1>::type;
+
+                            constexpr index_t c_offset =
+                                CThreadDesc_TM0_TM1_TN0_TN1{}.CalculateOffset(
+                                    c_origin_idx + make_multi_index(tm0, tm1, tn0, tn1));
+
+                            inner_product<a_vector_t, b_vector_t, FloatC>(
+                                a_vec.template AsType<a_vector_t>()[I0],
+                                b_vec.template AsType<b_vector_t>()[I0],
+                                c_buf(Number<c_offset>{}));
+                        });
+                    });
+                });
+            });
+        });
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl_dpp8.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl_dpp8.hpp
new file mode 100644
index 0000000000..d0d214381d
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl_dpp8.hpp
@@ -0,0 +1,136 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/amd_gemm_dpp.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/inner_product_dpp8.hpp"
+#include "ck/utility/math.hpp"
+
+namespace ck {
+
+/**
+ * Threadwise contraction using dot instructions with DPP8 modifier.
+ *
+ * Assumptions:
+ *   1. `AThreadDesc_TK0_TM0_TM1_TK1`, `BThreadDesc_TK0_TN0_TN1_TK1`, `CThreadDesc_TM0_TM1_TN0_TN1`
+ *      are known at compile-time;
+ *   2. `AOriginIdx`, `BOriginIdx`, `COriginIdx` are known at compile-time;
+ *   3. `TM0` is equal to 1 and `TN0` is equal to 1;
+ *   4. When `ShareA` is set (unset, respectively), `TM1` (`TN1`, respectively) is divisible by
+ *      the size of the lane group (`dpp8::lane_group_size`).
+ */
+template <typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename AThreadDesc_TK0_TM0_TM1_TK1,
+          typename BThreadDesc_TK0_TN0_TN1_TK1,
+          typename CThreadDesc_TM0_TM1_TN0_TN1,
+          typename TKLengths,
+          typename TMLengths,
+          typename TNLengths,
+          bool ShareA,
+          typename enable_if<AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
+                                 BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
+                                 CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseContractionDlDpp8_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1
+{
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t TK0 = TKLengths{}[I0];
+    static constexpr index_t TK1 = TKLengths{}[I1];
+    static constexpr index_t TM0 = TMLengths{}[I0];
+    static constexpr index_t TM1 = TMLengths{}[I1];
+    static constexpr index_t TN0 = TNLengths{}[I0];
+    static constexpr index_t TN1 = TNLengths{}[I1];
+
+    static_assert(TM0 == 1 && TN0 == 1);
+
+    static_assert((ShareA && TM1 % dpp8::lane_group_size == 0) ||
+                  (!ShareA && TN1 % dpp8::lane_group_size == 0));
+    static constexpr index_t shared_elems_per_lane =
+        ShareA ? TM1 / dpp8::lane_group_size : TN1 / dpp8::lane_group_size;
+
+    __device__ constexpr ThreadwiseContractionDlDpp8_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1()
+    {
+        static_assert(AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
+                          BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
+                          CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(TKLengths::Size() == 2 && TMLengths::Size() == 2 && TNLengths::Size() == 2,
+                      "wrong!");
+    }
+
+    template <typename ABuffer,
+              typename AOriginIdx,
+              typename BBuffer,
+              typename BOriginIdx,
+              typename CBuffer,
+              typename COriginIdx>
+    __device__ static void Run(const ABuffer& a_buf,
+                               AOriginIdx,
+                               const BBuffer& b_buf,
+                               BOriginIdx,
+                               CBuffer& c_buf,
+                               COriginIdx)
+    {
+        static_assert(is_known_at_compile_time<remove_cvref_t<AOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<BOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<COriginIdx>>::value,
+                      "wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename ABuffer::type>, remove_cvref_t<FloatA>>::value &&
+            is_same<remove_cvref_t<typename BBuffer::type>, remove_cvref_t<FloatB>>::value &&
+            is_same<remove_cvref_t<typename CBuffer::type>, remove_cvref_t<FloatC>>::value &&
+            "wrong! inconsistent type");
+
+        constexpr auto a_origin_idx = to_multi_index(AOriginIdx{});
+        constexpr auto b_origin_idx = to_multi_index(BOriginIdx{});
+        constexpr auto c_origin_idx = to_multi_index(COriginIdx{});
+
+        static_for<0, TK0, 1>{}([&](auto tk0) {
+            static_for<0, TM1, 1>{}([&](auto tm1) {
+                static_for<0, TN1, 1>{}([&](auto tn1) {
+                    vector_type<FloatA, TK1> a_vec;
+                    vector_type<FloatB, TK1> b_vec;
+
+                    static_for<0, TK1, 1>{}([&](auto tk1) {
+                        constexpr index_t local_tm1 = ShareA ? tm1 % shared_elems_per_lane : tm1;
+                        constexpr index_t a_offset  = AThreadDesc_TK0_TM0_TM1_TK1{}.CalculateOffset(
+                            a_origin_idx + make_multi_index(tk0, 0, local_tm1, tk1));
+
+                        constexpr index_t local_tn1 = ShareA ? tn1 : tn1 % shared_elems_per_lane;
+                        constexpr index_t b_offset  = BThreadDesc_TK0_TN0_TN1_TK1{}.CalculateOffset(
+                            b_origin_idx + make_multi_index(tk0, 0, local_tn1, tk1));
+
+                        a_vec.template AsType<FloatA>()(tk1) = a_buf[Number<a_offset>{}];
+                        b_vec.template AsType<FloatB>()(tk1) = b_buf[Number<b_offset>{}];
+                    });
+
+                    using a_vector_t = typename vector_type<FloatA, TK1>::type;
+                    using b_vector_t = typename vector_type<FloatB, TK1>::type;
+
+                    constexpr index_t c_offset = CThreadDesc_TM0_TM1_TN0_TN1{}.CalculateOffset(
+                        c_origin_idx + make_multi_index(0, tm1, 0, tn1));
+
+                    constexpr int src_lane =
+                        ShareA ? (tm1 / shared_elems_per_lane) % dpp8::lane_group_size
+                               : (tn1 / shared_elems_per_lane) % dpp8::lane_group_size;
+
+                    dpp8::inner_product_dpp<a_vector_t, b_vector_t, FloatC, src_lane, ShareA>(
+                        a_vec.template AsType<a_vector_t>()[I0],
+                        b_vec.template AsType<b_vector_t>()[I0],
+                        c_buf(Number<c_offset>{}));
+                });
+            });
+        });
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp
new file mode 100644
index 0000000000..e97aa433a6
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp
@@ -0,0 +1,168 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_THREADWISE_GEMM_DLOPS_V3_HPP
+#define CK_THREADWISE_GEMM_DLOPS_V3_HPP
+
+#include "common_header.hpp"
+#include "math.hpp"
+
+namespace ck {
+
+// C[M, N] += transpose(A[K, M]) * B[K, N]
+//   Element of matrix can be vectorized data
+// Assume:
+//   1. AThreadDesc_E1_K_E2, BThreadDesc_E1_N_Ho_Wo_E2, CThreadDesc_K_N_Ho_Wo are known at
+//   compile-time
+//   2. AOriginIdx, BOriginIdx, COriginIdx are known at compile-time
+template <typename FloatA,
+          typename FloatB,
+          typename FloatC,
+          typename AThreadDesc_E1_K_E2,
+          typename BThreadDesc_E1_N_Ho_Wo_E2,
+          typename CThreadDesc_K_N_Ho_Wo,
+          typename enable_if<AThreadDesc_E1_K_E2::IsKnownAtCompileTime() &&
+                                 BThreadDesc_E1_N_Ho_Wo_E2::IsKnownAtCompileTime() &&
+                                 CThreadDesc_K_N_Ho_Wo::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseGemmDlops_km_kn_mn_v3
+{
+
+    template <typename ABuffer,
+              typename AOriginIdx,
+              typename BBuffer,
+              typename BOriginIdx,
+              typename CBuffer,
+              typename COriginIdx>
+    __device__ static void Run(const ABuffer& a_buf,
+                               AOriginIdx,
+                               const BBuffer& b_buf,
+                               BOriginIdx,
+                               CBuffer& c_buf,
+                               COriginIdx)
+    {
+
+        static_assert(AThreadDesc_E1_K_E2::IsKnownAtCompileTime() &&
+                          BThreadDesc_E1_N_Ho_Wo_E2::IsKnownAtCompileTime() &&
+                          CThreadDesc_K_N_Ho_Wo::IsKnownAtCompileTime(),
+                      "wrong! Desc should be known at compile-time");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<AOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<BOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<COriginIdx>>::value,
+                      "wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename ABuffer::type>, remove_cvref_t<FloatA>>::value &&
+            is_same<remove_cvref_t<typename BBuffer::type>, remove_cvref_t<FloatB>>::value &&
+            is_same<remove_cvref_t<typename CBuffer::type>, remove_cvref_t<FloatC>>::value &&
+            "wrong! inconsistent type");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+        constexpr auto I2 = Number<2>{};
+        constexpr auto I3 = Number<3>{};
+
+        constexpr auto E1 = AThreadDesc_E1_K_E2{}.GetLength(I0);
+        constexpr auto K  = AThreadDesc_E1_K_E2{}.GetLength(I1);
+        constexpr auto E2 = AThreadDesc_E1_K_E2{}.GetLength(I2);
+
+        constexpr auto Ho = BThreadDesc_E1_N_Ho_Wo_E2{}.GetLength(I2);
+        constexpr auto Wo = BThreadDesc_E1_N_Ho_Wo_E2{}.GetLength(I3);
+
+        constexpr auto a_origin_idx = to_multi_index(AOriginIdx{});
+        constexpr auto b_origin_idx = to_multi_index(BOriginIdx{});
+        constexpr auto c_origin_idx = to_multi_index(COriginIdx{});
+
+        if constexpr((Ho % 2 == 0) && (Wo % 2 == 0))
+        {
+            constexpr auto SubHW = 2;
+
+            static_for<0, K, 1>{}([&](auto k) {
+                static_for<0, Ho, SubHW>{}([&](auto h) {
+                    static_for<0, Wo, SubHW>{}([&](auto w) {
+                        static_for<0, E1, 1>{}([&](auto e1) {
+                            static_for<0, E2, 1>{}([&](auto e2) {
+                                constexpr index_t a_offset = AThreadDesc_E1_K_E2{}.CalculateOffset(
+                                    a_origin_idx + make_tuple(e1, k, e2));
+
+                                constexpr index_t b0_offset =
+                                    BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
+                                        b_origin_idx + make_tuple(e1, 0, h, w, e2));
+
+                                constexpr index_t b1_offset =
+                                    BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
+                                        b_origin_idx + make_tuple(e1, 0, h, w + 1, e2));
+
+                                constexpr index_t b2_offset =
+                                    BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
+                                        b_origin_idx + make_tuple(e1, 0, h + 1, w, e2));
+
+                                constexpr index_t b3_offset =
+                                    BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
+                                        b_origin_idx + make_tuple(e1, 0, h + 1, w + 1, e2));
+
+                                constexpr index_t c0_offset =
+                                    CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(c_origin_idx +
+                                                                            make_tuple(k, 0, h, w));
+
+                                constexpr index_t c1_offset =
+                                    CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(
+                                        c_origin_idx + make_tuple(k, 0, h, w + 1));
+
+                                constexpr index_t c2_offset =
+                                    CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(
+                                        c_origin_idx + make_tuple(k, 0, h + 1, w));
+
+                                constexpr index_t c3_offset =
+                                    CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(
+                                        c_origin_idx + make_tuple(k, 0, h + 1, w + 1));
+
+                                amd_assembly_outer_product_1x4(a_buf[Number<a_offset>{}],
+                                                               b_buf[Number<b0_offset>{}],
+                                                               b_buf[Number<b1_offset>{}],
+                                                               b_buf[Number<b2_offset>{}],
+                                                               b_buf[Number<b3_offset>{}],
+                                                               c_buf(Number<c0_offset>{}),
+                                                               c_buf(Number<c1_offset>{}),
+                                                               c_buf(Number<c2_offset>{}),
+                                                               c_buf(Number<c3_offset>{}));
+                            });
+                        });
+                    });
+                });
+            });
+        }
+        else
+        {
+
+            static_for<0, K, 1>{}([&](auto k) {
+                static_for<0, Ho, 1>{}([&](auto h) {
+                    static_for<0, Wo, 1>{}([&](auto w) {
+                        static_for<0, E1, 1>{}([&](auto e1) {
+                            static_for<0, E2, 1>{}([&](auto e2) {
+                                constexpr index_t a_offset = AThreadDesc_E1_K_E2{}.CalculateOffset(
+                                    a_origin_idx + make_tuple(e1, k, e2));
+
+                                constexpr index_t b_offset =
+                                    BThreadDesc_E1_N_Ho_Wo_E2{}.CalculateOffset(
+                                        b_origin_idx + make_tuple(e1, 0, h, w, e2));
+
+                                constexpr index_t c_offset =
+                                    CThreadDesc_K_N_Ho_Wo{}.CalculateOffset(c_origin_idx +
+                                                                            make_tuple(k, 0, h, w));
+
+                                inner_product<FloatA, FloatB, FloatC>(a_buf[Number<a_offset>{}],
+                                                                      b_buf[Number<b_offset>{}],
+                                                                      c_buf(Number<c_offset>{}));
+                            });
+                        });
+                    });
+                });
+            });
+        }
+    }
+};
+
+} // namespace ck
+#endif
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp
new file mode 100644
index 0000000000..a8bf383d90
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// Assume:
+//   1. Desc is known at compile-time
+//   2. Buffer is StaticBuffer
+//   3. OriginIdx is known at compile-time
+//   4. use #-step
+template <typename Data,
+          typename Desc,
+          typename SliceLengths,
+          typename enable_if<Desc::IsKnownAtCompileTime(), bool>::type = false>
+struct ThreadwiseTensorSliceSet_v1
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    template <typename OriginIdx, typename Buffer>
+    __device__ void Run(const Desc&, const OriginIdx&, Buffer& buf, const Data& initial_value) const
+    {
+        static_assert(Desc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc and DstDesc need to known at compile-time");
+
+        static_assert(Buffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<OriginIdx>>::value,
+                      "wrong! OriginIdx need to be known at compile-time");
+
+        // Desc is known at compile-time
+        constexpr auto desc = remove_cvref_t<Desc>{};
+
+        // OriginIdx is known at compile-time
+        constexpr auto origin_idx = to_multi_index(OriginIdx{});
+
+        static_ford<SliceLengths>{}([&](auto access_idx) {
+            constexpr auto coord = make_tensor_coordinate(desc, origin_idx + access_idx);
+
+            constexpr bool is_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(desc, coord);
+
+            constexpr index_t offset = coord.GetOffset();
+
+            if constexpr(is_valid)
+            {
+                buf(Number<offset>{}) = initial_value;
+            }
+        });
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
new file mode 100644
index 0000000000..8a6d2f4ab5
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
@@ -0,0 +1,1264 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+namespace detail {
+// TODO: How to fix this? It uses an struct instead of lambda because lambda
+// doesn't have constructor
+template <index_t VectorDim, index_t ScalarPerVector>
+struct lambda_scalar_per_access
+{
+    __host__ __device__ constexpr auto operator()(index_t i) const
+    {
+        return (i == VectorDim) ? ScalarPerVector : 1;
+    }
+};
+
+template <index_t VectorDim>
+struct lambda_scalar_step_in_vector
+{
+    __host__ __device__ constexpr auto operator()(index_t i) const
+    {
+        return (i == VectorDim) ? 1 : 0;
+    }
+};
+} // namespace detail
+
+// Assume:
+//   1. src:
+//     1. SrcDesc is known at compile-time
+//     2. SrcBuffer is StaticBuffer
+//     3. SrcSliceOrginIdx is known at compile-time
+//   2. dst:
+//     1. DstDesc is not known at compile-time
+//     2. DstBuffer is DynamicBuffer
+//     3. DstSliceOrginIdx is not known at compile time
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t DstVectorDim,
+          index_t DstScalarPerVector,
+          InMemoryDataOperationEnum DstInMemOp,
+          index_t DstScalarStrideInVector,
+          bool DstResetCoordinateAfterRun,
+          typename enable_if<SrcDesc::IsKnownAtCompileTime(), bool>::type = false>
+struct ThreadwiseTensorSliceTransfer_v1r3
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v1r3(const DstDesc& dst_desc,
+                                                            const Index& dst_slice_origin_idx,
+                                                            const ElementwiseOperation& element_op)
+        : dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin_idx)),
+          element_op_{element_op}
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc need to known at compile-time");
+        static_assert(SliceLengths::At(Number<DstVectorDim>{}) % DstScalarPerVector == 0,
+                      "wrong! Not divisible");
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcSliceOriginIdx, typename SrcBuffer, typename DstBuffer>
+    __device__ void Run(const SrcDesc&,
+                        const SrcSliceOriginIdx&,
+                        const SrcBuffer& src_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc need to known at compile-time");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<SrcSliceOriginIdx>>::value,
+                      "wrong! SrcSliceOrigin need to known at compile-time");
+
+        static_assert(SrcBuffer::IsStaticBuffer(), "wrong! SrcBuffer need to be StaticBuffer");
+
+        // SrcDesc and src_slice_origin_idx are known at compile-time
+        constexpr auto src_desc             = remove_cvref_t<SrcDesc>{};
+        constexpr auto src_slice_origin_idx = to_multi_index(SrcSliceOriginIdx{});
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_scalar_step_in_vector =
+            generate_sequence(detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(dst_scalar_per_access)>>;
+
+        // TODO: Use SpaceFillingCurve::ScalarsPerAccess instread of DstScalarPerVector?
+        static_assert(DstScalarPerVector == SpaceFillingCurve::ScalarPerVector,
+                      "wrong!DstScalarPerVector != SpaceFillingCurve::ScalarPerVector");
+        typename vector_type_maker<DstData, DstScalarPerVector>::type dst_vector;
+        using dst_vector_t = typename vector_type_maker<DstData, DstScalarPerVector>::type::type;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            constexpr auto idx_md = SpaceFillingCurve::GetIndex(idx_1d);
+
+            // copy data from src_buf into dst_vector
+            // TODO: It's a hack here to use \p dst_scalar_step_in_vector. Use SpaceFillingCurve?
+            static_for<0, DstScalarPerVector, 1>{}([&](auto i) {
+                constexpr index_t src_offset = src_desc.CalculateOffset(
+                    src_slice_origin_idx + idx_md + i * dst_scalar_step_in_vector);
+
+                SrcData v;
+
+                // apply element-wise operation
+                element_op_(v, src_buf[Number<src_offset>{}]);
+
+                // apply type convert
+                dst_vector.template AsType<DstData>()(i) = type_convert<DstData>(v);
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(dst_desc, dst_coord_);
+
+            // copy data from dst_vector into dst_buf
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector.template AsType<dst_vector_t>()[Number<0>{}]);
+
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+
+                move_tensor_coordinate(dst_desc, dst_coord_, forward_step);
+            }
+        });
+
+        // move dst coordinate back to slice origin (or not)
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            move_tensor_coordinate(dst_desc, dst_coord_, GetDstCoordinateResetStep());
+        }
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(dst_scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
+                                       : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    DstCoord dst_coord_;
+    const ElementwiseOperation element_op_;
+}; // namespace ThreadwiseTensorSliceTransfer_v1r3
+
+// Assume:
+//   1. src:
+//     1. SrcDesc is not known at compile-time
+//     2. SrcBuffer is DynamicBuffer
+//     3. src_slice_origin_idx is not known at compile-time
+//   2. dst:
+//     1. DstDesc is known at compile-time
+//     2. DstBuffer is StaticBuffer
+//     3. dst_slice_origin_idx is known at compile-time
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t SrcVectorDim,
+          index_t SrcScalarPerVector,
+          index_t SrcScalarStrideInVector,
+          bool SrcResetCoordinateAfterRun,
+          bool InvalidElementAsNaN                                        = false,
+          typename enable_if<DstDesc::IsKnownAtCompileTime(), bool>::type = false>
+struct ThreadwiseTensorSliceTransfer_v2
+{
+    static_assert((InvalidElementAsNaN && !std::is_integral<DstData>::value) ||
+                      (!InvalidElementAsNaN),
+                  "Filling invalid element as NaN is only for floating point types");
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v2(const SrcDesc& src_desc,
+                                                          const Index& src_slice_origin_idx)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin_idx))
+    {
+        static_assert(DstDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc need to known at compile-time");
+        static_assert(SliceLengths::At(Number<SrcVectorDim>{}) % SrcScalarPerVector == 0,
+                      "wrong! Not divisible");
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, typename DstBuffer, typename DstSliceOriginIdx>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc&,
+                        const DstSliceOriginIdx&,
+                        DstBuffer& dst_buf)
+    {
+        static_assert(DstDesc::IsKnownAtCompileTime(),
+                      "wrong! DstDesc need to known at compile-time");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<DstSliceOriginIdx>>::value,
+                      "wrong! DstSliceOrigin need to known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value &&
+            "wrong! inconsistent type");
+
+        // DstDesc and dst_slice_origin_idx are known at compile-time
+        constexpr auto dst_desc             = remove_cvref_t<DstDesc>{};
+        constexpr auto dst_slice_origin_idx = DstSliceOriginIdx{};
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_scalar_step_in_vector =
+            generate_sequence(detail::lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(src_scalar_per_access)>>;
+
+        // loop over tensor and copy
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            typename vector_type_maker<SrcData, SrcScalarPerVector>::type src_vector;
+
+            using src_vector_t =
+                typename vector_type_maker<SrcData, SrcScalarPerVector>::type::type;
+            constexpr auto src_data_idx = SpaceFillingCurve::GetIndex(idx_1d);
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(src_desc, src_coord_);
+
+            // copy data from src_buf into src_vector
+            src_vector.template AsType<src_vector_t>()(Number<0>{}) =
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid);
+
+            // copy data from src_vector into dst_buf
+            static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
+                constexpr index_t dst_offset =
+                    dst_desc.CalculateOffset(to_multi_index(dst_slice_origin_idx) + src_data_idx +
+                                             i * src_scalar_step_in_vector);
+
+                if constexpr(InvalidElementAsNaN)
+                {
+                    dst_buf(Number<dst_offset>{}) =
+                        is_src_valid
+                            ? type_convert<DstData>(src_vector.template AsType<SrcData>()[i])
+                            : NumericLimits<DstData>::QuietNaN();
+                }
+                else
+                {
+                    dst_buf(Number<dst_offset>{}) =
+                        type_convert<DstData>(src_vector.template AsType<SrcData>()[i]);
+                }
+            });
+
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+
+                move_tensor_coordinate(src_desc, src_coord_, forward_step);
+            }
+        });
+
+        // move src coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            move_tensor_coordinate(src_desc, src_coord_, GetSrcCoordinateResetStep());
+        }
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(src_scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step_idx);
+    }
+
+    private:
+    SrcCoord src_coord_;
+}; // namespace ck
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+//   4. Use thread buffer
+template <typename SliceLengths,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector,
+          index_t SrcScalarStrideInVector,
+          index_t DstScalarStrideInVector,
+          bool SrcResetCoordinateAfterRun, // control whether to move back src coordinate after each
+                                           // RunRead(),  will be fused with MoveSrcSliceWindow to
+                                           // save addr computation
+          bool DstResetCoordinateAfterRun> // control whether to move back dst coordinate after each
+                                           // RunWrite(),  will be fused with MoveDstSliceWindow to
+                                           // save addr computation
+struct ThreadwiseTensorSliceTransfer_v3
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+    using Index                   = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v3(const SrcDesc& src_desc,
+                                                          const Index& src_slice_origin,
+                                                          const DstDesc& dst_desc,
+                                                          const Index& dst_slice_origin)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin))
+    {
+        static_assert(SliceLengths::At(Number<SrcVectorDim>{}) % SrcScalarPerVector == 0,
+                      "wrong! Not divisible");
+        static_assert(SliceLengths::At(Number<DstVectorDim>{}) % DstScalarPerVector == 0,
+                      "wrong! Not divisible");
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, typename SrcStepHacks>
+    __device__ void
+    RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf, const SrcStepHacks& src_step_hacks)
+    {
+        static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value,
+            "wrong! SrcBuffer and SrcData data type are inconsistent");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_scalar_step_in_vector =
+            generate_sequence(detail::lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // make forward steps
+        const auto src_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? src_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    src_desc, forward_step_idx, src_step_hacks[I0][i]);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto src_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -src_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    src_desc, backward_step_idx, src_step_hacks[I1][i]);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_src_access_lengths)>{}([&](auto ordered_src_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_src_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate src data index
+            constexpr auto src_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_src_access_idx[i]
+                                                      : ordered_src_access_lengths[i] - 1 -
+                                                            ordered_src_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                       src_scalar_per_access;
+            }();
+
+            vector_type_maker_t<SrcData, SrcScalarPerVector> src_tmp_vector;
+
+            using src_vector_t = typename decltype(src_tmp_vector)::type;
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(src_desc, src_coord_);
+
+            // copy data from src_buf to src_tmp_vector
+            src_tmp_vector.template AsType<src_vector_t>()(Number<0>{}) =
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid);
+
+            // copy data from src_tmp_vector to buffer_
+            static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
+                constexpr index_t buffer_offset =
+                    buffer_desc_.CalculateOffset(src_data_idx + i * src_scalar_step_in_vector);
+
+                buffer_(Number<buffer_offset>{}) = src_tmp_vector.template AsType<SrcData>()[i];
+            });
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_src_access_idx[i] < ordered_src_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_src_access_idx[j] == ordered_src_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_forward_steps[src_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_backward_steps[src_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move src coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetSrcCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+    }
+
+    template <typename DstBuffer, typename DstStepHacks>
+    __device__ void
+    RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf, const DstStepHacks& dst_step_hacks)
+    {
+        static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        constexpr auto I0 = Number<0>{};
+        constexpr auto I1 = Number<1>{};
+
+        // src scalar per access on each dim
+        // TODO: don't use this
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_scalar_step_in_vector =
+            generate_sequence(detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // make forward steps
+        const auto dst_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    dst_desc, forward_step_idx, dst_step_hacks[I0][i]);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto dst_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    dst_desc, backward_step_idx, dst_step_hacks[I1][i]);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_dst_access_lengths)>{}([&](auto ordered_dst_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_dst_access_idx[I0];
+
+                    static_for<1, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate dst data index
+            constexpr auto dst_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_idx[i]
+                                                      : ordered_dst_access_lengths[i] - 1 -
+                                                            ordered_dst_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                       dst_scalar_per_access;
+            }();
+
+            vector_type_maker_t<DstData, DstScalarPerVector> dst_tmp_vector;
+
+            // copy data from buffer_ to dst_tmp_vector
+            static_for<0, DstScalarPerVector, 1>{}([&](auto i) {
+                constexpr index_t buffer_offset =
+                    buffer_desc_.CalculateOffset(dst_data_idx + i * dst_scalar_step_in_vector);
+
+                dst_tmp_vector.template AsType<DstData>()(i) =
+                    type_convert<DstData>(buffer_[Number<buffer_offset>{}]);
+            });
+
+            using dst_vector_t = typename decltype(dst_tmp_vector)::type;
+
+            // copy data from dst_tmp_vector to dst_buf
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(dst_desc, dst_coord_);
+
+            dst_buf.template Set<dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_tmp_vector.template AsType<dst_vector_t>()[Number<0>{}]);
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_dst_access_idx[i] < ordered_dst_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_dst_access_idx[j] == ordered_dst_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_forward_steps[dst_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_backward_steps[dst_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move dst coordinate back to slice origin (or not)
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetDstCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    template <typename SrcBuffer>
+    __device__ void RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf)
+    {
+        constexpr index_t ntransform_src = SrcDesc::GetNumOfTransform();
+
+        constexpr auto zeros = typename uniform_sequence_gen<ntransform_src, 0>::type{};
+
+        constexpr auto src_step_hacks =
+            make_tuple(generate_tuple([&](auto) { return zeros; }, Number<nDim>{}),
+                       generate_tuple([&](auto) { return zeros; }, Number<nDim>{}));
+
+        RunRead(src_desc, src_buf, src_step_hacks);
+    }
+
+    template <typename DstBuffer>
+    __device__ void RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf)
+    {
+        constexpr index_t ntransform_dst = DstDesc::GetNumOfTransform();
+
+        constexpr auto zeros = typename uniform_sequence_gen<ntransform_dst, 0>::type{};
+
+        constexpr auto dst_step_hacks =
+            make_tuple(generate_tuple([&](auto) { return zeros; }, Number<nDim>{}),
+                       generate_tuple([&](auto) { return zeros; }, Number<nDim>{}));
+
+        RunWrite(dst_desc, dst_buf, dst_step_hacks);
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        constexpr auto I0 = Number<0>{};
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_src_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate src data index after last iteration in RunRead(), if it has not being reset by
+        // RunRead()
+        constexpr auto src_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_src_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                   src_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_src_data_step = [&]() {
+            Index reset_src_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_src_data_step_(i) = -src_data_idx[i]; });
+
+            return reset_src_data_step_;
+        }();
+
+        return reset_src_data_step;
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        constexpr auto I0 = Number<0>{};
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_dst_access_lengths[I0] - 1;
+
+                static_for<1, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate dst data index after last iteration in RunWrite(), if it has not being reset by
+        // RunWrite()
+        constexpr auto dst_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                   dst_scalar_per_access;
+        }();
+
+        //
+        constexpr auto reset_dst_data_step = [&]() {
+            Index reset_dst_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_dst_data_step_(i) = -dst_data_idx[i]; });
+
+            return reset_dst_data_step_;
+        }();
+
+        return reset_dst_data_step;
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <typename SrcMoveSliceWindowStepHack>
+    __device__ void
+    MoveSrcSliceWindow(const SrcDesc& src_desc,
+                       const Index& src_slice_origin_step_idx,
+                       const SrcMoveSliceWindowStepHack& src_move_slice_window_step_hack)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(
+            src_desc, adjusted_step_idx, src_move_slice_window_step_hack);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by RunWrite(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
+                                       : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    static constexpr auto buffer_desc_ =
+        make_naive_tensor_descriptor_packed(sequence_to_tuple_of_number(SliceLengths{}));
+
+    static constexpr auto buffer_size_ = buffer_desc_.GetElementSpaceSize();
+
+    StaticBuffer<AddressSpaceEnum::Vgpr, SrcData, buffer_size_, true> buffer_;
+
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+};
+
+// Assume:
+//   1. src:
+//     1. SrcDesc is known at compile-time
+//     2. SrcBuffer is DynamicBuffer
+//     3. src_ref_idx is known at run-time
+//     4. SrcRefToOriginDisplacement is known at compile-time
+//     5. use #-step
+//   2. dst:
+//     1. DstDesc is known at compile-time
+//     2. DstBuffer is StaticBuffer
+//     3. DstOriginIdx is known at compile-time
+//     4. use direct address calculation
+//   3. vector access on src
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t SrcVectorDim,
+          index_t SrcScalarPerVector,
+          index_t SrcScalarStrideInVector,
+          typename enable_if<SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseTensorSliceTransfer_v4
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v4(const Index& src_ref_idx)
+        : src_ref_coord_(make_tensor_coordinate(SrcDesc{}, src_ref_idx))
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc and DstDesc need to known at compile-time");
+
+        static_assert(SliceLengths::At(Number<SrcVectorDim>{}) % SrcScalarPerVector == 0,
+                      "wrong! Not divisible");
+    }
+
+    template <typename SrcRefToOriginDisplacement,
+              typename DstOriginIdx,
+              typename SrcBuffer,
+              typename DstBuffer>
+    __device__ void Run(const SrcDesc&,
+                        const SrcRefToOriginDisplacement&,
+                        const SrcBuffer& src_buf,
+                        const DstDesc&,
+                        const DstOriginIdx&,
+                        DstBuffer& dst_buf) const
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc and DstDesc need to known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value &&
+                is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        static_assert(DstBuffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<SrcRefToOriginDisplacement>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<DstOriginIdx>>::value,
+                      "wrong! SrcOriginToRefDistance and DstOriginToRefDistance need to be known "
+                      "at compile-time");
+
+        // SrcDesc and DstDesc are known at compile-time
+        constexpr auto src_desc = remove_cvref_t<SrcDesc>{};
+        constexpr auto dst_desc = remove_cvref_t<DstDesc>{};
+
+        // SrcOriginToRefDisttance and DstOriginToRefDistance are known at compile-time
+        constexpr auto src_ref_to_origin_disp_idx = to_multi_index(SrcRefToOriginDisplacement{});
+        constexpr auto dst_origin_idx             = to_multi_index(DstOriginIdx{});
+
+        // scalar per access of each dim
+        constexpr auto src_scalar_per_access = generate_sequence_v2(
+            [&](auto i) constexpr {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return Number<SrcScalarPerVector>{};
+                }
+                else
+                {
+                    return Number<1>{};
+                }
+            },
+            Number<nDim>{});
+
+        // scalar step (if steping on SrcVectorDim) of each dim
+        constexpr auto src_scalar_step_in_vector = generate_sequence_v2(
+            [&](auto i) constexpr {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return Number<1>{};
+                }
+                else
+                {
+                    return Number<0>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto dim_access_order = DimAccessOrder{};
+
+        constexpr auto ordered_access_lengths =
+            container_reorder_given_new2old(access_lengths, dim_access_order);
+
+        static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) {
+#if 0
+            // TODO: unable to compile
+            // position in slice window
+            constexpr auto data_to_origin_disp_idx =
+                container_reorder_given_old2new(ordered_access_idx, dim_access_order) *
+                src_scalar_per_access;
+#else
+            // position in slice window
+            constexpr auto data_to_origin_disp_idx =
+                ordered_access_idx.ReorderGivenOld2New(dim_access_order) * src_scalar_per_access;
+#endif
+            // src coordinate
+            constexpr auto src_ref_to_data_disp_idx =
+                src_ref_to_origin_disp_idx + data_to_origin_disp_idx;
+
+            auto src_data_coord = src_ref_coord_;
+
+            move_tensor_coordinate(src_desc, src_data_coord, src_ref_to_data_disp_idx);
+
+            vector_type_maker_t<SrcData, SrcScalarPerVector> src_tmp_vector;
+
+            using src_vector_t = typename decltype(src_tmp_vector)::type;
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(src_desc, src_data_coord);
+
+            // copy data from src_buf into src_tmp_vector
+            if constexpr(SrcBuffer::IsDynamicBuffer())
+            {
+                src_tmp_vector.template AsType<src_vector_t>()(Number<0>{}) =
+                    src_buf.template Get<src_vector_t>(src_data_coord.GetOffset(), is_src_valid);
+            }
+            else if constexpr(SrcBuffer::IsStaticBuffer())
+            {
+                static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
+                    constexpr index_t src_offset = src_desc.CalculateOffset(
+                        src_ref_to_origin_disp_idx + data_to_origin_disp_idx +
+                        i * src_scalar_step_in_vector);
+
+                    // apply type convert
+                    src_tmp_vector.template AsType<SrcData>()(i) = src_buf[Number<src_offset>{}];
+                });
+            }
+            // copy data from src_tmp_vector to dst_tmp_vector (data cast data from SrcData to
+            // DstData)
+            vector_type_maker_t<DstData, SrcScalarPerVector> dst_tmp_vector;
+
+            // TODO: if SrcData and DstData are vetor type, then static_cast may not compile
+            static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
+                dst_tmp_vector.template AsType<DstData>()(i) =
+                    type_convert<DstData>(src_tmp_vector.template AsType<SrcData>()[i]);
+            });
+
+            // copy data from dst_tmp_vector into dst_buf
+            static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
+                constexpr index_t dst_offset = dst_desc.CalculateOffset(
+                    dst_origin_idx + data_to_origin_disp_idx + i * src_scalar_step_in_vector);
+
+                dst_buf(Number<dst_offset>{}) = dst_tmp_vector.template AsType<DstData>()[i];
+            });
+        });
+    }
+
+    template <typename SrcSliceMoveStepIdx>
+    __device__ void MoveSrcSliceWindow(const SrcDesc&,
+                                       const SrcSliceMoveStepIdx& src_slice_move_step_idx)
+    {
+        constexpr auto src_desc = SrcDesc{};
+
+        const auto src_slice_move_step_iter =
+            make_tensor_coordinate_step(src_desc, to_multi_index(src_slice_move_step_idx));
+
+        move_tensor_coordinate(SrcDesc{}, src_ref_coord_, src_slice_move_step_iter);
+    }
+    __device__ void SetSrcCoord(const Index& src_ref_idx)
+    {
+        src_ref_coord_ = make_tensor_coordinate(SrcDesc{}, src_ref_idx);
+    }
+
+    private:
+    SrcCoord src_ref_coord_;
+};
+
+/**
+ * @brief Threadwise data transfer
+ *
+ * Do NOT involve any tensor coordinates with StaticBuffer
+ *
+ */
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t DstVectorDim,
+          index_t DstScalarPerVector,
+          typename enable_if<SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseTensorSliceTransfer_StaticToStatic
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_StaticToStatic(
+        const ElementwiseOperation& element_op)
+        : element_op_{element_op}
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc need to known at compile-time");
+
+        static_assert(SliceLengths::At(Number<DstVectorDim>{}) % DstScalarPerVector == 0,
+                      "wrong! Not divisible");
+    }
+
+    template <typename SrcSliceOriginIdx,
+              typename DstSliceOriginIdx,
+              typename SrcBuffer,
+              typename DstBuffer>
+    __device__ void Run(const SrcDesc&,
+                        const SrcSliceOriginIdx&,
+                        const SrcBuffer& src_buf,
+                        const DstDesc&,
+                        const DstSliceOriginIdx&,
+                        DstBuffer& dst_buf)
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! Desc need to known at compile-time");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<SrcSliceOriginIdx>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<DstSliceOriginIdx>>::value,
+                      "wrong! SliceOrigin need to known at compile-time");
+
+        static_assert(SrcBuffer::IsStaticBuffer() && DstBuffer::IsStaticBuffer(),
+                      "wrong! Buffer need to be StaticBuffer");
+
+        // SrcDesc and src_slice_origin_idx are known at compile-time
+        constexpr auto src_desc             = remove_cvref_t<SrcDesc>{};
+        constexpr auto dst_desc             = remove_cvref_t<DstDesc>{};
+        constexpr auto src_slice_origin_idx = to_multi_index(SrcSliceOriginIdx{});
+        constexpr auto dst_slice_origin_idx = to_multi_index(DstSliceOriginIdx{});
+
+        // scalar per access on each dim
+        constexpr auto dst_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto dst_scalar_step_in_vector =
+            generate_sequence(detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(dst_scalar_per_access)>>;
+
+        static_assert(DstScalarPerVector == SpaceFillingCurve::ScalarPerVector,
+                      "wrong!DstScalarPerVector != SpaceFillingCurve::ScalarPerVector");
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            constexpr auto idx_md = SpaceFillingCurve::GetIndex(idx_1d);
+
+            // copy data from src_buf into dst_vector
+            static_for<0, DstScalarPerVector, 1>{}([&](auto i) {
+                constexpr index_t src_offset = src_desc.CalculateOffset(
+                    src_slice_origin_idx + idx_md + i * dst_scalar_step_in_vector);
+
+                constexpr index_t dst_offset = dst_desc.CalculateOffset(
+                    dst_slice_origin_idx + idx_md + i * dst_scalar_step_in_vector);
+
+                SrcData v;
+
+                // apply element-wise operation
+                element_op_(v, src_buf[Number<src_offset>{}]);
+
+                // apply type convert
+                dst_buf(Number<dst_offset>{}) = type_convert<DstData>(v);
+            });
+        });
+    }
+
+    ElementwiseOperation element_op_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
new file mode 100644
index 0000000000..e56658eb80
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
@@ -0,0 +1,520 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_coordinate.hpp"
+#include "ck/tensor/thread_private_tensor.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+// FIXME: remove
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor/static_tensor.hpp"
+
+namespace ck {
+
+namespace detail {
+// TODO: How to fix this? It uses an struct instead of lambda because lambda
+// doesn't have constructor
+template <index_t SrcVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstVectorDim,
+          index_t DstScalarPerVector>
+struct lambda_scalar_per_access_for_src_and_dst
+{
+    __host__ __device__ constexpr auto operator()(index_t i) const
+    {
+        if(i == SrcVectorDim && i == DstVectorDim)
+        {
+            return math::lcm(SrcScalarPerVector, DstScalarPerVector);
+        }
+        else if(i == SrcVectorDim)
+        {
+            return SrcScalarPerVector;
+        }
+        else if(i == DstVectorDim)
+        {
+            return DstScalarPerVector;
+        }
+        else
+        {
+            return 1;
+        }
+    }
+};
+
+} // namespace detail
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+//   4. Use thread buffer
+template <typename SliceLengths,
+          typename SrcElementwiseOperation,
+          typename DstElementwiseOperation,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SrcDataTmp,
+          typename DstDataTmp,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          index_t SrcVectorDim,
+          index_t DstVectorDim,
+          index_t SrcScalarPerVector,
+          index_t DstScalarPerVector,
+          index_t SrcScalarStrideInVector,
+          index_t DstScalarStrideInVector,
+          bool SrcResetCoordinateAfterRun, // control whether to move back src coordinate after each
+                                           // RunRead(),  will be fused with MoveSrcSliceWindow to
+                                           // save addr computation
+          bool DstResetCoordinateAfterRun, // control whether to move back dst coordinate after each
+                                           // RunWrite(),  will be fused with MoveDstSliceWindow to
+                                           // save addr computation
+          index_t NumThreadScratch = 1>
+struct ThreadwiseTensorSliceTransfer_v3r1
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+    using Index                   = MultiIndex<nDim>;
+
+    using SrcData = remove_cvref_t<SrcDataTmp>;
+    using DstData = remove_cvref_t<DstDataTmp>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    // scalar per access on each dim
+    // TODO: don't use lambda_scalar_per_access
+    static constexpr auto src_scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+    static constexpr auto dst_scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
+
+    using SrcSpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                   SrcDimAccessOrder,
+                                                   remove_cv_t<decltype(src_scalar_per_access)>>;
+
+    using DstSpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                   DstDimAccessOrder,
+                                                   remove_cv_t<decltype(dst_scalar_per_access)>>;
+
+    __host__ __device__ constexpr ThreadwiseTensorSliceTransfer_v3r1()
+        : src_coord_{}, dst_coord_{}, src_element_op_{}, dst_element_op_{}
+    {
+    }
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v3r1(
+        const SrcDesc& src_desc,
+        const Index& src_slice_origin,
+        const SrcElementwiseOperation& src_element_op,
+        const DstDesc& dst_desc,
+        const Index& dst_slice_origin,
+        const DstElementwiseOperation& dst_element_op)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          src_element_op_(src_element_op),
+          dst_element_op_(dst_element_op)
+    {
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunRead(const SrcDesc& src_desc,
+                            const SrcBuffer& src_buf,
+                            Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value,
+            "wrong! SrcBuffer and SrcData data type are inconsistent");
+
+        constexpr auto num_access = SrcSpaceFillingCurve::GetNumOfAccess();
+
+        // loop over space-filling curve
+        static_for<0, num_access, 1>{}([&](auto iAccess) {
+            // calculate src data index
+            constexpr auto src_data_idx = SrcSpaceFillingCurve::GetIndexTupleOfNumber(iAccess);
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(src_desc, src_coord_);
+
+            using src_vector_type = vector_type_maker_t<SrcData, SrcScalarPerVector>;
+            using src_vector_t    = typename src_vector_type::type;
+
+            // copy data from src_buf into src_vector_container
+            auto src_vector_container = src_vector_type{
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
+
+            // copy data from src_vector_container into src_thread_scratch_
+#if 1 // debug
+            src_thread_scratch_tuple_(thread_scratch_id)
+                .template SetAsType<src_vector_t>(
+                    src_data_idx, src_vector_container.template AsType<src_vector_t>()[I0]);
+#else
+            src_thread_scratch_tuple_(thread_scratch_id)
+                .template Set<src_vector_t>(
+                    src_data_idx, src_vector_container.template AsType<src_vector_t>()[I0]);
+#endif
+
+            // move src coordinate
+            if constexpr(iAccess.value != num_access - 1)
+            {
+                constexpr auto step = SrcSpaceFillingCurve::GetForwardStep(iAccess);
+
+                move_tensor_coordinate(src_desc, src_coord_, step);
+            }
+        });
+
+        // move src coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step = GetSrcCoordinateResetStep();
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+    }
+
+    template <index_t ThreadScratchId>
+    __device__ void
+    TransferDataFromSrcThreadScratchToDstThreadScratch(Number<ThreadScratchId> thread_scratch_id)
+    {
+        // sub-dword transpose between src_thread_scratch_ and dst_thread_scratch_
+        // TODO make this logic more generic for more sub-dword datatype
+        if constexpr(SrcVectorDim != DstVectorDim &&
+                     ((is_same<half_t, remove_cvref_t<SrcData>>::value &&
+                       is_same<half_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) ||
+                      (is_same<int8_t, remove_cvref_t<SrcData>>::value &&
+                       is_same<int8_t, remove_cvref_t<DstData>>::value &&
+                       SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0)))
+        {
+            // each transpose does
+            // DstScalarPerVector # of src vectors in src_thread_scratch_
+            // SrcScalarPerVector # of dst vectors in dst_thread_scratch_
+            constexpr index_t num_src_vector = Number<DstScalarPerVector>{};
+            constexpr index_t num_dst_vector = Number<SrcScalarPerVector>{};
+
+            // Assume SrcVectorDim is not the same as DstVectorDim, so we do transpose
+            // TODO: make this logic generic for all scenario
+            static_assert(SrcVectorDim != DstVectorDim, "wrong");
+
+            constexpr auto src_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{});
+
+            constexpr auto dst_scalar_step_in_vector = generate_sequence(
+                detail::lambda_scalar_step_in_vector<DstVectorDim>{}, Number<nDim>{});
+
+            constexpr auto scalar_per_access = generate_sequence(
+                detail::lambda_scalar_per_access_for_src_and_dst<SrcVectorDim,
+                                                                 SrcScalarPerVector,
+                                                                 DstVectorDim,
+                                                                 DstScalarPerVector>{},
+                Number<nDim>{});
+
+            constexpr auto access_lengths = SliceLengths{} / scalar_per_access;
+
+            static_ford<decltype(access_lengths)>{}([&](auto access_idx) {
+                constexpr auto data_idx = access_idx * scalar_per_access;
+
+                constexpr auto data_idx_seq = generate_sequence_v2(
+                    [&](auto i) { return Number<data_idx[i]>{}; }, Number<nDim>{});
+
+                using src_vector_t = vector_type_maker_t<SrcData, SrcScalarPerVector>;
+                using dst_vector_t = vector_type_maker_t<DstData, DstScalarPerVector>;
+
+                // get DstScalarPerVector # of read-only references to src vectors from
+                // src_thread_scratch_
+                const auto src_vector_refs = generate_tie(
+                    [&](auto i) -> const src_vector_t& {
+                        // i increment corresponds to movement in DstVectorDim
+                        return src_thread_scratch_tuple_[thread_scratch_id].GetVectorTypeReference(
+                            data_idx_seq + i * dst_scalar_step_in_vector);
+                    },
+                    Number<num_src_vector>{});
+
+                // get SrcScalarPerVector # of references to dst vectors from dst_thread_scratch_
+                auto dst_vector_refs = generate_tie(
+                    [&](auto i) -> dst_vector_t& {
+                        // i increment corresponds to movement in SrcVectorDim
+                        return dst_thread_scratch_.GetVectorTypeReference(
+                            data_idx_seq + i * src_scalar_step_in_vector);
+                    },
+                    Number<num_dst_vector>{});
+
+                // do data transpose
+                transpose_vectors<SrcData, DstScalarPerVector, SrcScalarPerVector>{}(
+                    src_vector_refs, dst_vector_refs);
+            });
+        }
+
+        static_ford<SliceLengths>{}([&](auto idx) {
+            // apply the src elementwise op and convert to DstData under the hood if needed
+            DstData dst_v;
+            src_element_op_(dst_v, src_thread_scratch_tuple_[thread_scratch_id][idx]);
+            dst_thread_scratch_(idx) = dst_v;
+        });
+    }
+
+    template <typename DstBuffer, index_t ThreadScratchId = 0>
+    __device__ void RunWrite(const DstDesc& dst_desc,
+                             DstBuffer& dst_buf,
+                             Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
+    {
+        // if there is transpose, it's done here
+        // TODO move this elsewhere
+        TransferDataFromSrcThreadScratchToDstThreadScratch(thread_scratch_id);
+
+        static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        constexpr auto num_access = DstSpaceFillingCurve::GetNumOfAccess();
+
+        // loop over tensor and copy
+        static_for<0, num_access, 1>{}([&](auto iAccess) {
+            // calculate src data index
+            constexpr auto dst_data_idx = DstSpaceFillingCurve::GetIndexTupleOfNumber(iAccess);
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(dst_desc, dst_coord_);
+
+            using dst_vector_type = vector_type_maker_t<DstData, DstScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            // copy data from dst_thread_scratch_ into dst_vector_container
+            auto dst_vector_container =
+                dst_vector_type{dst_thread_scratch_.template GetAsType<dst_vector_t>(dst_data_idx)};
+
+            static_for<0, DstScalarPerVector, 1>{}([&](auto i) {
+                DstData dst_v;
+
+                // apply DstElementwiseOperation
+                dst_element_op_(dst_v, dst_vector_container.template AsType<DstData>()[i]);
+
+                dst_vector_container.template AsType<DstData>()(i) = dst_v;
+            });
+
+            // copy data from dst_vector_container to dst_buf
+            dst_buf.template Set<dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            // move dst coord
+            if constexpr(iAccess.value != num_access - 1)
+            {
+                constexpr auto step = DstSpaceFillingCurve::GetForwardStep(iAccess);
+
+                move_tensor_coordinate(dst_desc, dst_coord_, step);
+            }
+        });
+
+        // move dst coordinate back to slice origin (or not)
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            move_tensor_coordinate(dst_desc, dst_coord_, GetDstCoordinateResetStep());
+        }
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        constexpr auto num_access = SrcSpaceFillingCurve::GetNumOfAccess();
+
+        if constexpr(num_access == 0)
+        {
+            return typename SrcSpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SrcSpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        constexpr auto num_access = DstSpaceFillingCurve::GetNumOfAccess();
+
+        if constexpr(num_access == 0)
+        {
+            return typename DstSpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                DstSpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step_idx);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by RunWrite(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
+                                       : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step_idx);
+    }
+
+    __device__ static constexpr auto GetSrcThreadScratchDescriptor()
+    {
+        constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
+
+        constexpr auto src_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(src_access_lengths), Number<SrcScalarPerVector>{});
+
+        // 1st stage of transforms
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(src_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(src_access_lengths_and_vector_length[i],
+                                   src_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(src_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == SrcVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    __device__ static constexpr auto GetDstThreadScratchDescriptor()
+    {
+        // 1st stage of transforms
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access;
+
+        constexpr auto dst_access_lengths_and_vector_length = container_push_back(
+            sequence_to_tuple_of_number(dst_access_lengths), Number<DstScalarPerVector>{});
+
+        constexpr auto desc0 =
+            make_naive_tensor_descriptor_packed(dst_access_lengths_and_vector_length);
+
+        // 2nd stage of transforms
+        constexpr auto transforms = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return make_merge_transform_v3_division_mod(
+                        make_tuple(dst_access_lengths_and_vector_length[i],
+                                   dst_access_lengths_and_vector_length[Number<nDim>{}]));
+                }
+                else
+                {
+                    return make_pass_through_transform(dst_access_lengths_and_vector_length[i]);
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto low_dim_idss = generate_tuple(
+            [&](auto i) {
+                if constexpr(i == DstVectorDim)
+                {
+                    return Sequence<i.value, nDim>{};
+                }
+                else
+                {
+                    return Sequence<i.value>{};
+                }
+            },
+            Number<nDim>{});
+
+        constexpr auto up_dim_idss =
+            generate_tuple([&](auto i) { return Sequence<i.value>{}; }, Number<nDim>{});
+
+        return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss);
+    }
+
+    private:
+    static constexpr auto src_thread_scratch_desc_ = decltype(GetSrcThreadScratchDescriptor()){};
+    static constexpr auto dst_thread_scratch_desc_ = decltype(GetDstThreadScratchDescriptor()){};
+
+#if 1 // debug
+    using SrcThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                                             SrcData,
+                                                             SrcScalarPerVector,
+                                                             decltype(src_thread_scratch_desc_),
+                                                             true>;
+#else
+    using SrcThreadScratch = ThreadPrivateTensor<SrcData, decltype(src_thread_scratch_desc_)>;
+#endif
+
+    using DstThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
+                                                             DstData,
+                                                             DstScalarPerVector,
+                                                             decltype(dst_thread_scratch_desc_),
+                                                             true>;
+
+    StaticallyIndexedArray<SrcThreadScratch, NumThreadScratch> src_thread_scratch_tuple_;
+
+    DstThreadScratch dst_thread_scratch_;
+
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+    const SrcElementwiseOperation src_element_op_;
+    const DstElementwiseOperation dst_element_op_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp
new file mode 100644
index 0000000000..50ee8cf324
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp
@@ -0,0 +1,175 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+
+namespace ck {
+// Assume:
+//   1. src:
+//     1. SrcDesc is known at compile-time
+//     2. SrcBuffer is DynamicBuffer
+//     3. src_ref_idx is known at run-time
+//     4. SrcRefToOriginDisplacement is known at compile-time
+//     5. use #-step
+//   2. dst:
+//     1. DstDesc is known at compile-time
+//     2. DstBuffer is StaticBuffer
+//     3. DstOriginIdx is known at compile-time
+//     4. use direct address calculation
+//   3. vector access on src
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          typename SrcVectorTensorLengths,
+          typename SrcVectorTensorContiguousDimOrder,
+          typename enable_if<SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                             bool>::type = false>
+struct ThreadwiseTensorSliceTransfer_v4r1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+
+    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v4r1(const Index& src_ref_idx)
+        : src_ref_coord_(make_tensor_coordinate(SrcDesc{}, src_ref_idx))
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc and DstDesc need to known at compile-time");
+
+        static_for<0, nDim, 1>{}([](auto i) {
+            static_assert(SliceLengths::At(i) % SrcVectorTensorLengths::At(i) == 0, "wrong!");
+        });
+    }
+
+    template <typename SrcRefToOriginDisplacement,
+              typename DstOriginIdx,
+              typename SrcBuffer,
+              typename DstBuffer>
+    __device__ void Run(const SrcDesc&,
+                        const SrcRefToOriginDisplacement&,
+                        const SrcBuffer& src_buf,
+                        const DstDesc&,
+                        const DstOriginIdx&,
+                        DstBuffer& dst_buf) const
+    {
+        static_assert(SrcDesc::IsKnownAtCompileTime() && DstDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc and DstDesc need to known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value &&
+                is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        static_assert(DstBuffer::IsStaticBuffer(), "wrong! DstBuffer need to be StaticBuffer");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<SrcRefToOriginDisplacement>>::value &&
+                          is_known_at_compile_time<remove_cvref_t<DstOriginIdx>>::value,
+                      "wrong! SrcOriginToRefDistance and DstOriginToRefDistance need to be known "
+                      "at compile-time");
+
+        // SrcDesc and DstDesc are known at compile-time
+        constexpr auto src_desc = remove_cvref_t<SrcDesc>{};
+        constexpr auto dst_desc = remove_cvref_t<DstDesc>{};
+
+        // SrcOriginToRefDisttance and DstOriginToRefDistance are known at compile-time
+        constexpr auto src_ref_to_origin_disp_idx = to_multi_index(SrcRefToOriginDisplacement{});
+        constexpr auto dst_origin_idx             = to_multi_index(DstOriginIdx{});
+
+        // tensor descriptor for src_vector
+        constexpr auto src_vector_tensor_lengths = SrcVectorTensorLengths{};
+
+        constexpr auto src_vector_tensor_strides = container_reorder_given_old2new(
+            container_reverse_exclusive_scan(
+                container_reorder_given_new2old(src_vector_tensor_lengths,
+                                                SrcVectorTensorContiguousDimOrder{}),
+                math::multiplies{},
+                I1),
+            SrcVectorTensorContiguousDimOrder{});
+
+        constexpr auto src_vector_desc =
+            make_naive_tensor_descriptor(sequence_to_tuple_of_number(src_vector_tensor_lengths),
+                                         sequence_to_tuple_of_number(src_vector_tensor_strides));
+
+        // access order and lengths
+        constexpr auto access_lengths = SliceLengths{} / src_vector_tensor_lengths;
+
+        constexpr auto dim_access_order = DimAccessOrder{};
+
+        constexpr auto ordered_access_lengths =
+            container_reorder_given_new2old(access_lengths, dim_access_order);
+
+        static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) {
+            // position in slice window
+            constexpr auto data_to_origin_disp_idx =
+                ordered_access_idx.ReorderGivenOld2New(dim_access_order) *
+                src_vector_tensor_lengths;
+
+            // src coordinate at starting point of src_vector
+            constexpr auto src_ref_to_data_disp_idx =
+                src_ref_to_origin_disp_idx + data_to_origin_disp_idx;
+
+            constexpr auto src_ref_to_data_disp_coord_step =
+                make_tensor_coordinate_step(src_desc, src_ref_to_data_disp_idx);
+
+            auto src_data_coord = src_ref_coord_;
+
+            move_tensor_coordinate(src_desc, src_data_coord, src_ref_to_data_disp_coord_step);
+
+            vector_type_maker_t<SrcData, src_vector_desc.GetElementSpaceSize()> src_vector;
+
+            using src_vector_t = typename decltype(src_vector)::type;
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(src_desc, src_data_coord);
+
+            // copy data from src_buf into src_vector
+            src_vector.template AsType<src_vector_t>()(I0) =
+                src_buf.template Get<src_vector_t>(src_data_coord.GetOffset(), is_src_valid);
+
+            // copy data from src_vector into dst_buf (also cast from SrcData to DstData)
+            static_ford<SrcVectorTensorLengths>{}([&](auto src_vector_idx_) {
+                constexpr auto src_vector_idx = to_multi_index(src_vector_idx_);
+
+                constexpr index_t src_vector_offset =
+                    src_vector_desc.CalculateOffset(src_vector_idx);
+
+                constexpr index_t dst_offset = dst_desc.CalculateOffset(
+                    dst_origin_idx + data_to_origin_disp_idx + src_vector_idx);
+
+                dst_buf(Number<dst_offset>{}) = type_convert<DstData>(
+                    src_vector.template AsType<DstData>()[Number<src_vector_offset>{}]);
+            });
+        });
+    }
+
+    template <typename SrcSliceMoveStepIdx>
+    __device__ void MoveSrcSliceWindow(const SrcDesc&,
+                                       const SrcSliceMoveStepIdx& src_slice_move_step_idx)
+    {
+        constexpr auto src_desc = SrcDesc{};
+
+        const auto src_slice_move_step_iter =
+            make_tensor_coordinate_step(src_desc, to_multi_index(src_slice_move_step_idx));
+
+        move_tensor_coordinate(SrcDesc{}, src_ref_coord_, src_slice_move_step_iter);
+    }
+
+    private:
+    SrcCoord src_ref_coord_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp
new file mode 100644
index 0000000000..7cae17d511
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp
@@ -0,0 +1,614 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+//   4. Use thread buffer
+template <typename SliceLengths,
+          InMemoryDataOperationEnum DstInMemOp,
+          typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SrcDimAccessOrder,
+          typename DstDimAccessOrder,
+          typename SrcVectorTensorLengths,
+          typename DstVectorTensorLengths,
+          typename SrcVectorTensorContiguousDimOrder,
+          typename DstVectorTensorContiguousDimOrder,
+          bool SrcResetCoordinateAfterRun, // control whether to move back src coordinate after each
+                                           // RunRead(),  will be fused with MoveSrcSliceWindow to
+                                           // save addr computation
+          bool DstResetCoordinateAfterRun> // control whether to move back dst coordinate after each
+                                           // RunWrite(),  will be fused with MoveDstSliceWindow to
+                                           // save addr computation
+struct ThreadwiseTensorSliceTransfer_v5r1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr index_t nDim = SliceLengths::Size();
+    using Index                   = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+    using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v5r1(const SrcDesc& src_desc,
+                                                            const Index& src_slice_origin,
+                                                            const DstDesc& dst_desc,
+                                                            const Index& dst_slice_origin)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin))
+    {
+        // TODO: fix this
+        static_assert(is_same<SrcData, DstData>::value,
+                      "wrong! current implementation assume SrcData and DstData are same type");
+
+        static_for<0, nDim, 1>{}([](auto i) {
+            static_assert(SliceLengths::At(i) % SrcVectorTensorLengths::At(i) == 0 &&
+                              SliceLengths::At(i) % DstVectorTensorLengths::At(i) == 0,
+                          "wrong!");
+        });
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, typename SrcStepHacks>
+    __device__ void
+    RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf, const SrcStepHacks& src_step_hacks)
+    {
+        static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename SrcBuffer::type>, remove_cvref_t<SrcData>>::value,
+            "wrong! SrcBuffer and SrcData data type are inconsistent");
+
+        // tensor descriptor for src_vector
+        constexpr auto src_vector_tensor_lengths = SrcVectorTensorLengths{};
+
+        constexpr auto src_vector_tensor_strides = container_reorder_given_old2new(
+            container_reverse_exclusive_scan(
+                container_reorder_given_new2old(src_vector_tensor_lengths,
+                                                SrcVectorTensorContiguousDimOrder{}),
+                math::multiplies{},
+                I1),
+            SrcVectorTensorContiguousDimOrder{});
+
+        constexpr auto src_vector_desc =
+            make_naive_tensor_descriptor(sequence_to_tuple_of_number(src_vector_tensor_lengths),
+                                         sequence_to_tuple_of_number(src_vector_tensor_strides));
+
+        // access order and lengths
+        constexpr auto src_access_lengths = SliceLengths{} / src_vector_tensor_lengths;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // make forward steps
+        const auto src_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? src_vector_tensor_lengths[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    src_desc, forward_step_idx, src_step_hacks[I0][i]);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto src_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -src_vector_tensor_lengths[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    src_desc, backward_step_idx, src_step_hacks[I1][i]);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_src_access_lengths)>{}([&](auto ordered_src_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_src_access_idx[I0];
+
+                    static_for<0, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate src data index
+            constexpr auto src_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_src_access_idx[i]
+                                                      : ordered_src_access_lengths[i] - 1 -
+                                                            ordered_src_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                       src_vector_tensor_lengths;
+            }();
+
+            vector_type_maker_t<SrcData, src_vector_desc.GetElementSpaceSize()> src_vector;
+
+            using src_vector_t = typename decltype(src_vector)::type;
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(src_desc, src_coord_);
+
+            // copy data from src_buf to src_vector
+            src_vector.template AsType<src_vector_t>()(I0) =
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid);
+
+            // copy data from src_vector to buffer_
+            static_ford<SrcVectorTensorLengths>{}([&](auto src_vector_idx_) {
+                constexpr auto src_vector_idx = to_multi_index(src_vector_idx_);
+
+                constexpr index_t src_vector_offset =
+                    src_vector_desc.CalculateOffset(src_vector_idx);
+
+                constexpr index_t buffer_offset =
+                    buffer_desc_.CalculateOffset(src_data_idx + src_vector_idx);
+
+                buffer_(Number<buffer_offset>{}) =
+                    src_vector.template AsType<SrcData>()[Number<src_vector_offset>{}];
+            });
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_src_access_idx[i] < ordered_src_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_src_access_idx[j] == ordered_src_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_forward_steps[src_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            src_desc, src_coord_, src_backward_steps[src_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move src coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetSrcCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+    }
+
+    template <typename DstBuffer, typename DstStepHacks>
+    __device__ void
+    RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf, const DstStepHacks& dst_step_hacks)
+    {
+        static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
+                          DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
+                      "wrong!");
+
+        static_assert(
+            is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value,
+            "wrong! SrcBuffer or DstBuffer data type is wrong");
+
+        // tensor descriptor for dst_vector
+        constexpr auto dst_vector_tensor_lengths = DstVectorTensorLengths{};
+
+        constexpr auto dst_vector_tensor_strides = container_reorder_given_old2new(
+            container_reverse_exclusive_scan(
+                container_reorder_given_new2old(dst_vector_tensor_lengths,
+                                                DstVectorTensorContiguousDimOrder{}),
+                math::multiplies{},
+                I1),
+            DstVectorTensorContiguousDimOrder{});
+
+        constexpr auto dst_vector_desc =
+            make_naive_tensor_descriptor(sequence_to_tuple_of_number(dst_vector_tensor_lengths),
+                                         sequence_to_tuple_of_number(dst_vector_tensor_strides));
+
+        // dst access order and lengths
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_vector_tensor_lengths;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // make forward steps
+        const auto dst_forward_steps = generate_tuple(
+            [&](auto i) {
+                Index forward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    forward_step_idx(j) = (i.value == j.value) ? dst_vector_tensor_lengths[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    dst_desc, forward_step_idx, dst_step_hacks[I0][i]);
+            },
+            Number<nDim>{});
+
+        // make backward steps
+        const auto dst_backward_steps = generate_tuple(
+            [&](auto i) {
+                Index backward_step_idx;
+
+                static_for<0, nDim, 1>{}([&](auto j) {
+                    backward_step_idx(j) = (i.value == j.value) ? -dst_vector_tensor_lengths[i] : 0;
+                });
+
+                return make_tensor_coordinate_step(
+                    dst_desc, backward_step_idx, dst_step_hacks[I1][i]);
+            },
+            Number<nDim>{});
+
+        // loop over tensor and copy
+        static_ford<decltype(ordered_dst_access_lengths)>{}([&](auto ordered_dst_access_idx) {
+            // judge move forward or move backward
+            constexpr auto forward_sweep = [&]() {
+                StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+                forward_sweep_(I0) = true;
+
+                static_for<1, nDim, 1>{}([&](auto i) {
+                    index_t tmp = ordered_dst_access_idx[I0];
+
+                    static_for<0, i, 1>{}([&](auto j) {
+                        tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_idx[j];
+                    });
+
+                    forward_sweep_(i) = tmp % 2 == 0;
+                });
+
+                return forward_sweep_;
+            }();
+
+            // calculate dst data index
+            constexpr auto dst_data_idx = [&]() {
+                Index ordered_idx;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_idx[i]
+                                                      : ordered_dst_access_lengths[i] - 1 -
+                                                            ordered_dst_access_idx[i];
+                });
+
+                return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                       dst_vector_tensor_lengths;
+            }();
+
+            vector_type_maker_t<DstData, dst_vector_desc.GetElementSpaceSize()> dst_vector;
+
+            // copy data from buffer_ to dst_vector (also cast from SrcData to DstData)
+            static_ford<DstVectorTensorLengths>{}([&](auto dst_vector_idx_) {
+                constexpr auto dst_vector_idx = to_multi_index(dst_vector_idx_);
+
+                constexpr index_t buffer_offset =
+                    buffer_desc_.CalculateOffset(dst_data_idx + dst_vector_idx);
+
+                constexpr index_t dst_vector_offset =
+                    dst_vector_desc.CalculateOffset(dst_vector_idx);
+
+                dst_vector.template AsType<DstData>()(Number<dst_vector_offset>{}) =
+                    type_convert<DstData>(buffer_[Number<buffer_offset>{}]);
+            });
+
+            using dst_vector_t = typename decltype(dst_vector)::type;
+
+            // copy data from dst_vector to dst_buf
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(dst_desc, dst_coord_);
+
+            dst_buf.template Set<dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector.template AsType<dst_vector_t>()[Number<0>{}]);
+
+            constexpr auto move_on_dim = [&]() constexpr
+            {
+                StaticallyIndexedArray<bool, nDim> move_on_dim_;
+
+                static_for<0, nDim, 1>{}([&](auto i) {
+                    move_on_dim_(i) = ordered_dst_access_idx[i] < ordered_dst_access_lengths[i] - 1;
+
+                    static_for<i + 1, nDim, 1>{}([&](auto j) {
+                        move_on_dim_(i) &=
+                            ordered_dst_access_idx[j] == ordered_dst_access_lengths[j] - 1;
+                    });
+                });
+
+                return move_on_dim_;
+            }
+            ();
+
+            // move
+            static_for<0, nDim, 1>{}([&](auto i) {
+                if constexpr(move_on_dim[i])
+                {
+                    if constexpr(forward_sweep[i])
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_forward_steps[dst_dim_access_order[i]]);
+                    }
+                    else
+                    {
+                        move_tensor_coordinate(
+                            dst_desc, dst_coord_, dst_backward_steps[dst_dim_access_order[i]]);
+                    }
+                }
+            });
+        });
+
+        // move dst coordinate back to slice origin (or not)
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetDstCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    template <typename SrcBuffer>
+    __device__ void RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf)
+    {
+        constexpr index_t ntransform_src = SrcDesc::GetNumOfTransform();
+
+        constexpr auto zeros = typename uniform_sequence_gen<ntransform_src, 0>::type{};
+
+        constexpr auto src_step_hacks =
+            make_tuple(generate_tuple([&](auto) { return zeros; }, Number<nDim>{}),
+                       generate_tuple([&](auto) { return zeros; }, Number<nDim>{}));
+
+        RunRead(src_desc, src_buf, src_step_hacks);
+    }
+
+    template <typename DstBuffer>
+    __device__ void RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf)
+    {
+        constexpr index_t ntransform_dst = DstDesc::GetNumOfTransform();
+
+        constexpr auto zeros = typename uniform_sequence_gen<ntransform_dst, 0>::type{};
+
+        constexpr auto dst_step_hacks =
+            make_tuple(generate_tuple([&](auto) { return zeros; }, Number<nDim>{}),
+                       generate_tuple([&](auto) { return zeros; }, Number<nDim>{}));
+
+        RunWrite(dst_desc, dst_buf, dst_step_hacks);
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        constexpr auto src_vector_tensor_lengths = SrcVectorTensorLengths{};
+
+        constexpr auto src_access_lengths = SliceLengths{} / src_vector_tensor_lengths;
+
+        constexpr auto src_dim_access_order = SrcDimAccessOrder{};
+
+        constexpr auto ordered_src_access_lengths =
+            container_reorder_given_new2old(src_access_lengths, src_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_src_access_lengths[I0] - 1;
+
+                static_for<0, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_src_access_lengths[j] + ordered_src_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate src data index after last iteration in RunRead(), if it has not being reset by
+        // RunRead()
+        constexpr auto src_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_src_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, src_dim_access_order) *
+                   src_vector_tensor_lengths;
+        }();
+
+        //
+        constexpr auto reset_src_data_step = [&]() {
+            Index reset_src_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_src_data_step_(i) = -src_data_idx[i]; });
+
+            return reset_src_data_step_;
+        }();
+
+        return reset_src_data_step;
+    }
+
+    __device__ static constexpr auto GetDstCoordinateResetStep()
+    {
+        constexpr auto dst_vector_tensor_lengths = DstVectorTensorLengths{};
+
+        constexpr auto dst_access_lengths = SliceLengths{} / dst_vector_tensor_lengths;
+
+        constexpr auto dst_dim_access_order = DstDimAccessOrder{};
+
+        constexpr auto ordered_dst_access_lengths =
+            container_reorder_given_new2old(dst_access_lengths, dst_dim_access_order);
+
+        // judge move forward or move backward during the last iteration
+        constexpr auto forward_sweep = [&]() {
+            StaticallyIndexedArray<bool, nDim> forward_sweep_;
+
+            forward_sweep_(I0) = true;
+
+            static_for<1, nDim, 1>{}([&](auto i) {
+                index_t tmp = ordered_dst_access_lengths[I0] - 1;
+
+                static_for<0, i, 1>{}([&](auto j) {
+                    tmp = tmp * ordered_dst_access_lengths[j] + ordered_dst_access_lengths[j] - 1;
+                });
+
+                forward_sweep_(i) = tmp % 2 == 0;
+            });
+
+            return forward_sweep_;
+        }();
+
+        // calculate dst data index after last iteration in RunWrite(), if it has not being reset by
+        // RunWrite()
+        constexpr auto dst_data_idx = [&]() {
+            Index ordered_idx;
+
+            static_for<0, nDim, 1>{}([&](auto i) {
+                ordered_idx(i) = forward_sweep[i] ? ordered_dst_access_lengths[i] - 1 : 0;
+            });
+
+            return container_reorder_given_old2new(ordered_idx, dst_dim_access_order) *
+                   dst_vector_tensor_lengths;
+        }();
+
+        //
+        constexpr auto reset_dst_data_step = [&]() {
+            Index reset_dst_data_step_;
+
+            static_for<0, nDim, 1>{}([&](auto i) { reset_dst_data_step_(i) = -dst_data_idx[i]; });
+
+            return reset_dst_data_step_;
+        }();
+
+        return reset_dst_data_step;
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <typename SrcMoveSliceWindowStepHack>
+    __device__ void
+    MoveSrcSliceWindow(const SrcDesc& src_desc,
+                       const Index& src_slice_origin_step_idx,
+                       const SrcMoveSliceWindowStepHack& src_move_slice_window_step_hack)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(
+            src_desc, adjusted_step_idx, src_move_slice_window_step_hack);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by RunWrite(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
+                                       : dst_slice_origin_step_idx + GetDstCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    static constexpr auto buffer_desc_ =
+        make_naive_tensor_descriptor_packed(sequence_to_tuple_of_number(SliceLengths{}));
+
+    static constexpr auto buffer_size_ = buffer_desc_.GetElementSpaceSize();
+
+    StaticBuffer<AddressSpaceEnum::Vgpr, SrcData, buffer_size_, true> buffer_;
+
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp
new file mode 100644
index 0000000000..d4a37e0432
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp
@@ -0,0 +1,203 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          InMemoryDataOperationEnum DstInMemOp,
+          bool SrcResetCoordinateAfterRun,
+          bool DstResetCoordinateAfterRun>
+struct ThreadwiseTensorSliceTransfer_v6r1
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v6r1(const SrcDesc& src_desc,
+                                                            const Index& src_slice_origin,
+                                                            const DstDesc& dst_desc,
+                                                            const Index& dst_slice_origin,
+                                                            const ElementwiseOperation& element_op)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, typename DstBuffer>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        // loop over space-filling curve
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            using src_vector_type = vector_type_maker_t<SrcData, ScalarPerVector>;
+            using src_vector_t    = typename src_vector_type::type;
+
+            using dst_vector_type = vector_type_maker_t<DstData, ScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(src_desc, src_coord_);
+
+            // copy data from src_buf into src_vector_container
+            auto src_vector_container = src_vector_type{
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
+
+            auto dst_vector_container = dst_vector_type{};
+
+            // apply pointwise operation
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                DstData v;
+
+                // apply element-wise operation
+                element_op_(v, src_vector_container.template AsType<SrcData>()[i]);
+
+                // apply type convert
+                dst_vector_container.template AsType<DstData>()(i) = v;
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(dst_desc, dst_coord_);
+
+            // copy data from dst_vector into dst_buf
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            // move coordinate
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+
+                move_tensor_coordinate(src_desc, src_coord_, forward_step);
+                move_tensor_coordinate(dst_desc, dst_coord_, forward_step);
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            move_tensor_coordinate(src_desc, src_coord_, GetCoordinateResetStep());
+        }
+
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            move_tensor_coordinate(dst_desc, dst_coord_, GetCoordinateResetStep());
+        }
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = SrcResetCoordinateAfterRun
+                                           ? src_slice_origin_step_idx
+                                           : src_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRun
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step_idx);
+    }
+
+    private:
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1r2.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1r2.hpp
new file mode 100644
index 0000000000..88ed217547
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1r2.hpp
@@ -0,0 +1,213 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+// Assume:
+//   1. src_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          bool SrcResetCoordinateAfterRun,
+          bool DstResetCoordinateAfterRun>
+struct ThreadwiseTensorSliceTransfer_v6r1r2
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+    using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v6r1r2(
+        const SrcDesc& src_desc,
+        const Index& src_slice_origin,
+        const DstDesc& dst_desc,
+        const Index& dst_slice_origin,
+        const ElementwiseOperation& element_op)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename SrcBuffer, typename DstBuffer, InMemoryDataOperationEnum DstInMemOp>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        // loop over space-filling curve
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            using src_vector_type = vector_type_maker_t<SrcData, ScalarPerVector>;
+            using src_vector_t    = typename src_vector_type::type;
+
+            using dst_vector_type = vector_type_maker_t<DstData, ScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            const bool is_src_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc, src_coord_);
+
+            // copy data from src_buf into src_vector_container
+            auto src_vector_container = src_vector_type{
+                src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
+
+            auto dst_vector_container = dst_vector_type{};
+
+            // apply pointwise operation
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                SrcData v;
+
+                // apply element-wise operation
+                element_op_(v, src_vector_container.template AsType<SrcData>()[i]);
+
+                // apply type convert
+                dst_vector_container.template AsType<DstData>()(i) = type_convert<DstData>(v);
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
+
+            // copy data from dst_vector into dst_buf
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            // move coordinate
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+                move_tensor_coordinate(
+                    src_desc, src_coord_, make_tensor_coordinate_step(src_desc, forward_step));
+                move_tensor_coordinate(
+                    dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = SrcResetCoordinateAfterRun
+                                           ? src_slice_origin_step_idx
+                                           : src_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRun
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    SrcCoord src_coord_;
+    DstCoord dst_coord_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp
new file mode 100644
index 0000000000..d9da8fa6a1
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp
@@ -0,0 +1,260 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+// Assume:
+//   1. src0_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+template <typename Src0Data,
+          typename Src1Data,
+          typename DstData,
+          typename Src0Desc,
+          typename Src1Desc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          InMemoryDataOperationEnum DstInMemOp,
+          bool Src0ResetCoordinateAfterRun,
+          bool Src1ResetCoordinateAfterRun,
+          bool DstResetCoordinateAfterRun>
+struct ThreadwiseTensorSliceTransfer_v6r2
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using Src0Coord = decltype(make_tensor_coordinate(Src0Desc{}, Index{}));
+    using Src1Coord = decltype(make_tensor_coordinate(Src1Desc{}, Index{}));
+    using DstCoord  = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v6r2(const Src0Desc& src0_desc,
+                                                            const Index& src0_slice_origin,
+                                                            const Src1Desc& src1_desc,
+                                                            const Index& src1_slice_origin,
+                                                            const DstDesc& dst_desc,
+                                                            const Index& dst_slice_origin,
+                                                            const ElementwiseOperation& element_op)
+        : src0_coord_(make_tensor_coordinate(src0_desc, src0_slice_origin)),
+          src1_coord_(make_tensor_coordinate(src1_desc, src1_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    __device__ void SetSrc0SliceOrigin(const Src0Desc& src0_desc,
+                                       const Index& src0_slice_origin_idx)
+    {
+        src0_coord_ = make_tensor_coordinate(src0_desc, src0_slice_origin_idx);
+    }
+
+    __device__ void SetSrc1SliceOrigin(const Src1Desc& src1_desc,
+                                       const Index& src1_slice_origin_idx)
+    {
+        src1_coord_ = make_tensor_coordinate(src1_desc, src1_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename Src0Buffer, typename Src1Buffer, typename DstBuffer>
+    __device__ void Run(const Src0Desc& src0_desc,
+                        const Src0Buffer& src0_buf,
+                        const Src1Desc& src1_desc,
+                        const Src1Buffer& src1_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        // loop over space-filling curve
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            using src0_vector_type = vector_type_maker_t<Src0Data, ScalarPerVector>;
+            using src0_vector_t    = typename src0_vector_type::type;
+
+            using src1_vector_type = vector_type_maker_t<Src1Data, ScalarPerVector>;
+            using src1_vector_t    = typename src1_vector_type::type;
+
+            using dst_vector_type = vector_type_maker_t<DstData, ScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            const bool is_src0_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(src0_desc, src0_coord_);
+
+            const bool is_src1_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(src1_desc, src1_coord_);
+
+            // copy data from src0_buf into src0_vector_container
+            auto src0_vector_container = src0_vector_type{
+                src0_buf.template Get<src0_vector_t>(src0_coord_.GetOffset(), is_src0_valid)};
+
+            auto src1_vector_container = src1_vector_type{
+                src1_buf.template Get<src1_vector_t>(src1_coord_.GetOffset(), is_src1_valid)};
+
+            auto dst_vector_container = dst_vector_type{};
+
+            // apply pointwise operation
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                element_op_(dst_vector_container.template AsType<DstData>()(i),
+                            src0_vector_container.template AsType<Src0Data>()[i],
+                            src1_vector_container.template AsType<Src1Data>()[i]);
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(dst_desc, dst_coord_);
+
+            // copy data from dst_vector into dst_buf
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            // move coordinate
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+                move_tensor_coordinate(
+                    src0_desc, src0_coord_, make_tensor_coordinate_step(src0_desc, forward_step));
+                move_tensor_coordinate(
+                    src1_desc, src1_coord_, make_tensor_coordinate_step(src1_desc, forward_step));
+                move_tensor_coordinate(
+                    dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        if constexpr(Src0ResetCoordinateAfterRun)
+        {
+            const auto src0_reset_step =
+                make_tensor_coordinate_step(src0_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src0_desc, src0_coord_, src0_reset_step);
+        }
+
+        if constexpr(Src1ResetCoordinateAfterRun)
+        {
+            const auto src1_reset_step =
+                make_tensor_coordinate_step(src1_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src1_desc, src1_coord_, src1_reset_step);
+        }
+
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrc0SliceWindow(const Src0Desc& src0_desc,
+                                        const Index& src0_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = Src0ResetCoordinateAfterRun
+                                           ? src0_slice_origin_step_idx
+                                           : src0_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src0_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src0_desc, src0_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrc1SliceWindow(const Src1Desc& src1_desc,
+                                        const Index& src1_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = Src1ResetCoordinateAfterRun
+                                           ? src1_slice_origin_step_idx
+                                           : src1_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src1_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src1_desc, src1_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRun
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    Src0Coord src0_coord_;
+    Src1Coord src1_coord_;
+    DstCoord dst_coord_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp
new file mode 100644
index 0000000000..05f8c05c9f
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp
@@ -0,0 +1,310 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
+// and sometimes useless instructions:
+//   1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
+//   instead
+//   2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
+//   tensor coordinate instead
+//   3. Don't use a pointer to VGPR buffer, use vector instead
+
+// Assume:
+//   1. src0_desc and dst_desc are not known at compile-time
+//   2. SrcBuffer and DstBuffer are DynamicBuffer
+//   3. src_slice_origin and dst_slice_origin are not known at compile-time,
+template <typename Src0Data,
+          typename Src1Data,
+          typename Src2Data,
+          typename DstData,
+          typename Src0Desc,
+          typename Src1Desc,
+          typename Src2Desc,
+          typename DstDesc,
+          typename ElementwiseOperation,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          InMemoryDataOperationEnum DstInMemOp,
+          bool Src0ResetCoordinateAfterRun,
+          bool Src1ResetCoordinateAfterRun,
+          bool Src2ResetCoordinateAfterRun,
+          bool DstResetCoordinateAfterRun>
+struct ThreadwiseTensorSliceTransfer_v6r3
+{
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using Src0Coord = decltype(make_tensor_coordinate(Src0Desc{}, Index{}));
+    using Src1Coord = decltype(make_tensor_coordinate(Src1Desc{}, Index{}));
+    using Src2Coord = decltype(make_tensor_coordinate(Src2Desc{}, Index{}));
+    using DstCoord  = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
+
+    static constexpr auto I0 = Number<0>{};
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v6r3(const Src0Desc& src0_desc,
+                                                            const Index& src0_slice_origin,
+                                                            const Src1Desc& src1_desc,
+                                                            const Index& src1_slice_origin,
+                                                            const Src2Desc& src2_desc,
+                                                            const Index& src2_slice_origin,
+                                                            const DstDesc& dst_desc,
+                                                            const Index& dst_slice_origin,
+                                                            const ElementwiseOperation& element_op)
+        : src0_coord_(make_tensor_coordinate(src0_desc, src0_slice_origin)),
+          src1_coord_(make_tensor_coordinate(src1_desc, src1_slice_origin)),
+          src2_coord_(make_tensor_coordinate(src2_desc, src2_slice_origin)),
+          dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    __device__ void SetSrc0SliceOrigin(const Src0Desc& src0_desc,
+                                       const Index& src0_slice_origin_idx)
+    {
+        src0_coord_ = make_tensor_coordinate(src0_desc, src0_slice_origin_idx);
+    }
+
+    __device__ void SetSrc1SliceOrigin(const Src1Desc& src1_desc,
+                                       const Index& src1_slice_origin_idx)
+    {
+        src1_coord_ = make_tensor_coordinate(src1_desc, src1_slice_origin_idx);
+    }
+
+    __device__ void SetSrc2SliceOrigin(const Src2Desc& src2_desc,
+                                       const Index& src2_slice_origin_idx)
+    {
+        src2_coord_ = make_tensor_coordinate(src2_desc, src2_slice_origin_idx);
+    }
+
+    __device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
+    {
+        dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
+    }
+
+    template <typename Src0Buffer, typename Src1Buffer, typename Src2Buffer, typename DstBuffer>
+    __device__ void Run(const Src0Desc& src0_desc,
+                        const Src0Buffer& src0_buf,
+                        const Src1Desc& src1_desc,
+                        const Src1Buffer& src1_buf,
+                        const Src2Desc& src2_desc,
+                        const Src2Buffer& src2_buf,
+                        const DstDesc& dst_desc,
+                        DstBuffer& dst_buf)
+    {
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        // loop over space-filling curve
+        static_for<0, num_access, 1>{}([&](auto idx_1d) {
+            using src0_vector_type = vector_type_maker_t<Src0Data, ScalarPerVector>;
+            using src0_vector_t    = typename src0_vector_type::type;
+
+            using src1_vector_type = vector_type_maker_t<Src1Data, ScalarPerVector>;
+            using src1_vector_t    = typename src1_vector_type::type;
+
+            using src2_vector_type = vector_type_maker_t<Src2Data, ScalarPerVector>;
+            using src2_vector_t    = typename src2_vector_type::type;
+
+            using dst_vector_type = vector_type_maker_t<DstData, ScalarPerVector>;
+            using dst_vector_t    = typename dst_vector_type::type;
+
+            const bool is_src0_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(src0_desc, src0_coord_);
+
+            const bool is_src1_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(src1_desc, src1_coord_);
+
+            const bool is_src2_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(src2_desc, src2_coord_);
+
+            // copy data from src0_buf into src0_vector_container
+            auto src0_vector_container = src0_vector_type{
+                src0_buf.template Get<src0_vector_t>(src0_coord_.GetOffset(), is_src0_valid)};
+
+            auto src1_vector_container = src1_vector_type{
+                src1_buf.template Get<src1_vector_t>(src1_coord_.GetOffset(), is_src1_valid)};
+
+            auto src2_vector_container = src2_vector_type{
+                src2_buf.template Get<src2_vector_t>(src2_coord_.GetOffset(), is_src2_valid)};
+
+            auto dst_vector_container = dst_vector_type{};
+
+            // apply pointwise operation
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                element_op_(dst_vector_container.template AsType<DstData>()(i),
+                            src0_vector_container.template AsType<Src0Data>()[i],
+                            src1_vector_container.template AsType<Src1Data>()[i],
+                            src2_vector_container.template AsType<Src2Data>()[i]);
+            });
+
+            const bool is_dst_valid =
+                coordinate_has_valid_offset_assuming_top_index_is_valid(dst_desc, dst_coord_);
+
+            dst_buf.template Update<DstInMemOp, dst_vector_t>(
+                dst_coord_.GetOffset(),
+                is_dst_valid,
+                dst_vector_container.template AsType<dst_vector_t>()[I0]);
+
+            // move coordinate
+            if constexpr(idx_1d.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+                move_tensor_coordinate(
+                    src0_desc, src0_coord_, make_tensor_coordinate_step(src0_desc, forward_step));
+                move_tensor_coordinate(
+                    src1_desc, src1_coord_, make_tensor_coordinate_step(src1_desc, forward_step));
+                move_tensor_coordinate(
+                    src2_desc, src2_coord_, make_tensor_coordinate_step(src2_desc, forward_step));
+                move_tensor_coordinate(
+                    dst_desc, dst_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        if constexpr(Src0ResetCoordinateAfterRun)
+        {
+            const auto src0_reset_step =
+                make_tensor_coordinate_step(src0_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src0_desc, src0_coord_, src0_reset_step);
+        }
+
+        if constexpr(Src1ResetCoordinateAfterRun)
+        {
+            const auto src1_reset_step =
+                make_tensor_coordinate_step(src1_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src1_desc, src1_coord_, src1_reset_step);
+        }
+
+        if constexpr(Src2ResetCoordinateAfterRun)
+        {
+            const auto src2_reset_step =
+                make_tensor_coordinate_step(src2_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(src2_desc, src2_coord_, src2_reset_step);
+        }
+
+        if constexpr(DstResetCoordinateAfterRun)
+        {
+            const auto dst_reset_step =
+                make_tensor_coordinate_step(dst_desc, GetCoordinateResetStep());
+
+            move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrc0SliceWindow(const Src0Desc& src0_desc,
+                                        const Index& src0_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = Src0ResetCoordinateAfterRun
+                                           ? src0_slice_origin_step_idx
+                                           : src0_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src0_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src0_desc, src0_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrc1SliceWindow(const Src1Desc& src1_desc,
+                                        const Index& src1_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = Src1ResetCoordinateAfterRun
+                                           ? src1_slice_origin_step_idx
+                                           : src1_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src1_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src1_desc, src1_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrc2SliceWindow(const Src2Desc& src2_desc,
+                                        const Index& src2_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = Src2ResetCoordinateAfterRun
+                                           ? src2_slice_origin_step_idx
+                                           : src2_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src2_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src2_desc, src2_coord_, adjusted_step);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRun
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
+    }
+
+    private:
+    Src0Coord src0_coord_;
+    Src1Coord src1_coord_;
+    Src2Coord src2_coord_;
+    DstCoord dst_coord_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp
new file mode 100644
index 0000000000..f3f7e1f78e
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp
@@ -0,0 +1,280 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+namespace ck {
+
+// Thread-level multi-source, multi-destination tensor slice data movement
+// Assume:
+//   1. All sources and destinations are DynamicBuffer
+//   2. Same VectorDim and ScalerPerVector for all sources and destinations
+//   3. DstInMemOps are per destination tensor
+//   4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
+//   5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
+//   6. Does not need to know src_descs and dst_descs at compile-time
+//   7. Does not need to know src_slice_origins and dst_slice_origins at compile-time,
+//
+// Does following things to avoid scratch memory issue
+//   1. Use StaticallyIndexedArray or vector_type instead of C array for thread buffer
+//   2. Pass tensor descritpors by reference (or tuple of references)
+//   3. Does not keep reference to tensor descriptor
+//   4. Does not construct new tensor coordinate when call Run()
+template <typename SrcDatas,
+          typename DstDatas,
+          typename SrcDescs,
+          typename DstDescs,
+          typename ElementwiseOperation,
+          typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t VectorDim,
+          index_t ScalarPerVector,
+          typename SrcResetCoordinateAfterRunFlags, // Sequence<bool ...>
+          typename DstResetCoordinateAfterRunFlags> // Sequence<bool ...>
+struct ThreadwiseTensorSliceTransfer_v7
+{
+    static constexpr auto I0 = Number<0>{};
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    static constexpr index_t nSrc = SrcDescs::Size();
+    static constexpr index_t nDst = DstDescs::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    // return a tuple of coordiantes for a tuple of tensor
+    template <typename Descs,
+              typename Indices,
+              enable_if_t<Descs::Size() == Indices::Size(), bool> = false>
+    static constexpr auto MakeCoordinates(const Descs& descs, const Indices& indices)
+    {
+        return generate_tuple([&](auto i) { return make_tensor_coordinate(descs[i], indices[i]); },
+                              Number<Descs::Size()>{});
+    }
+
+    using SrcCoords = decltype(MakeCoordinates(SrcDescs{}, StaticallyIndexedArray<Index, nSrc>{}));
+    using DstCoords = decltype(MakeCoordinates(DstDescs{}, StaticallyIndexedArray<Index, nDst>{}));
+
+    // scalar per access on each dim
+    // FIXME: don't use lambda_scalar_per_access
+    static constexpr auto scalar_per_access = generate_sequence(
+        detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
+
+    using SpaceFillingCurve =
+        SpaceFillingCurve<SliceLengths, DimAccessOrder, remove_cv_t<decltype(scalar_per_access)>>;
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v7(
+        const SrcDescs& src_descs,
+        const StaticallyIndexedArray<Index, nSrc>& src_slice_origins,
+        const DstDescs& dst_descs,
+        const StaticallyIndexedArray<Index, nDst>& dst_slice_origins,
+        const ElementwiseOperation& element_op)
+        : src_coords_(MakeCoordinates(src_descs, src_slice_origins)),
+          dst_coords_(MakeCoordinates(dst_descs, dst_slice_origins)),
+          element_op_(element_op)
+    {
+        static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
+                      "wrong! cannot evenly divide");
+    }
+
+    template <typename Indices, enable_if_t<SrcDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetSrcSliceOrigins(const SrcDescs& src_descs,
+                                       const Indices& src_slice_origin_idxs)
+    {
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            src_coords_(i) = make_tensor_coordinate(src_descs[i], src_slice_origin_idxs[i]);
+        });
+    }
+
+    template <typename Indices, enable_if_t<DstDescs::Size() == Indices::Size(), bool> = false>
+    __device__ void SetDstSliceOrigins(const DstDescs& dst_descs,
+                                       const Indices& dst_slice_origin_idxs)
+    {
+        static_for<0, nDst, 1>{}([&](auto i) {
+            dst_coords_(i) = make_tensor_coordinate(dst_descs[i], dst_slice_origin_idxs[i]);
+        });
+    }
+
+    // SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
+    // SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
+    // DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
+    // DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
+    template <typename SrcBuffers,
+              typename DstBuffers,
+              enable_if_t<SrcDescs::Size() == SrcBuffers::Size() &&
+                              DstDescs::Size() == DstBuffers::Size(),
+                          bool> = false>
+    __device__ void Run(const SrcDescs& src_descs,
+                        const SrcBuffers& src_bufs,
+                        const DstDescs& dst_descs,
+                        DstBuffers dst_bufs)
+    {
+        auto generate_vectors = [&](auto data_types) {
+            constexpr index_t num = data_types.Size();
+
+            return generate_tuple(
+                [&](auto i) {
+                    using DataType = remove_cvref_t<decltype(data_types[i])>;
+
+                    return vector_type_maker_t<DataType, ScalarPerVector>{};
+                },
+                Number<num>{});
+        };
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        // loop over space-filling curve
+        static_for<0, num_access, 1>{}([&](auto iAccess) {
+            auto src_vectors = generate_vectors(SrcDatas{});
+            auto dst_vectors = generate_vectors(DstDatas{});
+
+            // copy data from src_bufs into src_vectors
+            static_for<0, nSrc, 1>{}([&](auto i) {
+                using src_vector_t = typename remove_cvref_t<decltype(src_vectors[i])>::type;
+
+                const bool is_src_valid = coordinate_has_valid_offset_assuming_top_index_is_valid(
+                    src_descs[i], src_coords_[i]);
+
+                src_vectors(i).template AsType<src_vector_t>()(I0) =
+                    src_bufs[i].template Get<src_vector_t>(src_coords_[i].GetOffset(),
+                                                           is_src_valid);
+            });
+
+            // apply pointwise function
+            static_for<0, ScalarPerVector, 1>{}([&](auto i) {
+                // get reference to src data
+                const auto src_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iSrc) -> const auto& {
+                        using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
+
+                        return src_vectors[iSrc].template AsType<SrcData>()[i];
+                    },
+                    Number<nSrc>{});
+
+                // get reference to dst data
+                auto dst_data_refs = generate_tie(
+                    // return type should be lvalue
+                    [&](auto iDst) -> auto& {
+                        using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
+
+                        return dst_vectors(iDst).template AsType<DstData>()(i);
+                    },
+                    Number<nDst>{});
+
+                // apply pointwise function
+                // pointwise function signature:
+                // element_op_(dst_data_refs[I0],
+                //             dst_data_refs[I1],
+                //             ...,
+                //             src_data_refs[I0],
+                //             src_data_refs[I1],
+                //             ...)
+                unpack2(element_op_, dst_data_refs, src_data_refs);
+            });
+
+            // copy data from buf_vectors into dst_bufs
+            static_for<0, nDst, 1>{}([&](auto i) {
+                using dst_vector_t = typename remove_cvref_t<decltype(dst_vectors[i])>::type;
+
+                const bool is_dst_valid = coordinate_has_valid_offset_assuming_top_index_is_valid(
+                    dst_descs[i], dst_coords_[i]);
+
+                constexpr InMemoryDataOperationEnum DstInMemOp =
+                    static_cast<InMemoryDataOperationEnum>(DstInMemOps::At(i.value));
+
+                dst_bufs(i).template Update<DstInMemOp, dst_vector_t>(
+                    dst_coords_[i].GetOffset(),
+                    is_dst_valid,
+                    dst_vectors[i].template AsType<dst_vector_t>()[I0]);
+            });
+
+            // move coordinate
+            if constexpr(iAccess.value != num_access - 1)
+            {
+                constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(iAccess);
+
+                static_for<0, nSrc, 1>{}([&](auto i) {
+                    move_tensor_coordinate(src_descs[i], src_coords_(i), forward_step);
+                });
+
+                static_for<0, nDst, 1>{}([&](auto i) {
+                    move_tensor_coordinate(dst_descs[i], dst_coords_(i), forward_step);
+                });
+            }
+        });
+
+        // move coordinate back to slice origin (or not)
+        static_for<0, nSrc, 1>{}([&](auto i) {
+            if constexpr(SrcResetCoordinateAfterRunFlags::At(i))
+            {
+                move_tensor_coordinate(src_descs[i], src_coords_(i), GetCoordinateResetStep());
+            }
+        });
+
+        static_for<0, nDst, 1>{}([&](auto i) {
+            if constexpr(DstResetCoordinateAfterRunFlags::At(i))
+            {
+                move_tensor_coordinate(dst_descs[i], dst_coords_(i), GetCoordinateResetStep());
+            }
+        });
+    }
+
+    __device__ static constexpr auto GetCoordinateResetStep()
+    {
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t ISrc>
+    __device__ void MoveSrcSliceWindow(const SrcDescs& src_descs,
+                                       Number<ISrc> iSrc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx = SrcResetCoordinateAfterRunFlags::At(iSrc)
+                                           ? src_slice_origin_step_idx
+                                           : src_slice_origin_step_idx + GetCoordinateResetStep();
+
+        move_tensor_coordinate(src_descs[iSrc], src_coords_(iSrc), adjusted_step_idx);
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <index_t IDst>
+    __device__ void MoveDstSliceWindow(const DstDescs& dst_descs,
+                                       Number<IDst> iDst,
+                                       const Index& dst_slice_origin_step_idx)
+    {
+        // if dst coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx = DstResetCoordinateAfterRunFlags::At(iDst)
+                                           ? dst_slice_origin_step_idx
+                                           : dst_slice_origin_step_idx + GetCoordinateResetStep();
+
+        move_tensor_coordinate(dst_descs[iDst], dst_coords_(iDst), adjusted_step_idx);
+    }
+
+    private:
+    SrcCoords src_coords_;
+    DstCoords dst_coords_;
+    const ElementwiseOperation element_op_;
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_welford.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_welford.hpp
new file mode 100644
index 0000000000..eb6715e8eb
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/thread/threadwise_welford.hpp
@@ -0,0 +1,137 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/math_v2.hpp"
+
+namespace ck {
+
+// Assume
+//  1) XDesc is known at compile-time
+//  2) MeanVarDesc is known at compile-time
+//  3) XBuffer is static buffer
+//  4) MeanBuffer is static buffer
+//  5) VarBuffer is static buffer
+template <typename T, typename XThreadDesc_M_K, typename MeanVarThreadDesc_M>
+struct ThreadwiseWelford
+{
+    static constexpr auto x_thread_desc_m_k      = XThreadDesc_M_K{};
+    static constexpr auto mean_var_thread_desc_m = MeanVarThreadDesc_M{};
+
+    static constexpr auto thread_x_length_m        = x_thread_desc_m_k.GetLength(Number<0>{});
+    static constexpr auto thread_x_length_k        = x_thread_desc_m_k.GetLength(Number<1>{});
+    static constexpr auto thread_mean_var_length_m = mean_var_thread_desc_m.GetLength(Number<0>{});
+
+    static_assert(thread_x_length_m == thread_mean_var_length_m,
+                  "lengths of source and mean/var buffer must match!");
+
+    __device__ constexpr ThreadwiseWelford() : cur_count_(0), max_count_(0) {}
+
+    __device__ inline void Update(T& mean, T& var, T x)
+    {
+        using ck::math::isnan;
+
+        if(isnan(x))
+        {
+            mean = x;
+            var  = x;
+        }
+        else
+        {
+            T delta = x - mean;
+            mean += delta / cur_count_;
+            T delta2 = x - mean;
+            var += delta * delta2;
+        }
+    }
+
+    template <typename XBufferType, typename MeanBufferType, typename VarBufferType>
+    __device__ void
+    Run(const XBufferType& x_buf_m_k, MeanBufferType& mean_buf_m, VarBufferType& var_buf_m)
+    {
+        // FIXME - Better naming for var_buf_m
+
+        static_for<0, thread_x_length_k, 1>{}([&](auto iK) {
+            if(cur_count_ < max_count_)
+            {
+                ++cur_count_;
+
+                static_for<0, thread_x_length_m, 1>{}([&](auto iM) {
+                    constexpr index_t out_offset =
+                        mean_var_thread_desc_m.CalculateOffset(make_tuple(iM));
+
+                    constexpr auto in_offset =
+                        x_thread_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+                    Update(mean_buf_m(Number<out_offset>{}),
+                           var_buf_m(Number<out_offset>{}),
+                           x_buf_m_k[Number<in_offset>{}]);
+                });
+            }
+        });
+    };
+
+    int cur_count_;
+    int max_count_;
+};
+
+template <typename T,
+          typename SrcMeanVarCountThreadDesc_M_K,
+          typename DstMeanVarThreadDesc_M,
+          bool GetActualVariance = false>
+struct ThreadwiseWelfordMerge
+{
+    static constexpr auto src_thread_desc_m_k = SrcMeanVarCountThreadDesc_M_K{};
+    static constexpr auto dst_thread_desc_m   = DstMeanVarThreadDesc_M{};
+
+    static constexpr auto src_length_m = src_thread_desc_m_k.GetLength(Number<0>{});
+    static constexpr auto src_length_k = src_thread_desc_m_k.GetLength(Number<1>{});
+    static constexpr auto dst_length_m = dst_thread_desc_m.GetLength(Number<0>{});
+
+    static_assert(src_length_m == dst_length_m, "lengths of source and dst buffer must match!");
+
+    __device__ static void
+    Merge(T& mean_a, T& var_a, int32_t& count_a, T mean_b, T var_b, int32_t count_b)
+    {
+        int count            = count_a + count_b;
+        T count_b_over_count = count == 0 ? type_convert<T>(0) : type_convert<T>(count_b) / count;
+        T delta              = mean_b - mean_a;
+        mean_a += delta * count_b_over_count;
+        var_a += var_b + delta * delta * count_a * count_b_over_count;
+        count_a = count;
+    }
+
+    template <typename SrcMeanBufferType,
+              typename SrcVarBufferType,
+              typename SrcCountBufferType,
+              typename DstMeanBufferType,
+              typename DstVarBufferType,
+              typename DstCountBufferType>
+    __device__ static void Run(const SrcMeanBufferType& src_mean_buf,
+                               const SrcVarBufferType& src_var_buf,
+                               const SrcCountBufferType& src_count_buf,
+                               DstMeanBufferType& dst_mean_buf,
+                               DstVarBufferType& dst_var_buf,
+                               DstCountBufferType& dst_count_buf)
+    {
+        static_for<0, src_length_m, 1>{}([&](auto iM) {
+            static_for<0, src_length_k, 1>{}([&](auto iK) {
+                constexpr auto src_offset = src_thread_desc_m_k.CalculateOffset(make_tuple(iM, iK));
+
+                Merge(dst_mean_buf(iM),
+                      dst_var_buf(iM),
+                      dst_count_buf(iM),
+                      src_mean_buf[Number<src_offset>{}],
+                      src_var_buf[Number<src_offset>{}],
+                      src_count_buf[Number<src_offset>{}]);
+            });
+
+            if constexpr(GetActualVariance)
+            {
+                dst_var_buf(iM) = dst_var_buf[iM] / dst_count_buf[iM];
+            };
+        });
+    };
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp b/include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp
new file mode 100644
index 0000000000..979f3567e9
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp
@@ -0,0 +1,507 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/math.hpp"
+#include "ck/utility/amd_wmma.hpp"
+
+namespace ck {
+
+enum struct WmmaInstr
+{
+    wmma_f32_16x16x16_f16 = 0,
+    wmma_f32_16x16x16_bf16,
+    wmma_f16_16x16x16_f16,
+    wmma_bf16_16x16x16_bf16,
+    wmma_i32_16x16x16_iu8,
+    wmma_i32_16x16x16_iu4
+};
+
+/*
+ *  WMMA Wave Tile Always MxNxK = 16x16x16
+ *  WAVE32
+        -----------------------------------
+        |RC0| | | | | | | | | | | | | | | |	   SubGroup 0
+        |RC1| | | | | | | | | | | | | | | |
+        |RC2| | | | | | | | | | | | | | | |
+        |RC3|T|T|T|T|T|T|T|T|T|T|T|T|T|T|T|
+        |RC4|0|0|0|0|0|0|0|0|0|1|1|1|1|1|1|
+        |RC5|1|2|3|4|5|6|7|8|9|0|1|2|3|4|5|
+        |RC6| | | | | | | | | | | | | | | |
+        |RC7| | | | | | | | | | | | | | | |
+        -----------------------------------
+        |   | | | | | | | | | | | | | | | |	   SubGroup 1
+        |   | | | | | | | | | | | | | | | |
+        | T |T|T|T|T|T|T|T|T|T|T|T|T|T|T|T|
+        | 1 |1|1|1|2|2|2|2|2|2|2|2|2|2|3|3|
+        | 6 |7|8|9|0|1|2|3|4|5|6|7|8|9|0|1|
+        |   | | | | | | | | | | | | | | | |
+        |   | | | | | | | | | | | | | | | |
+        |   | | | | | | | | | | | | | | | |
+        -----------------------------------
+
+
+ *  WAVE64
+        -----------------------------------
+        |RC0|T|T|T|T|T|T|T|T|T|T|T|T|T|T|T|	   SubGroup 0
+        |RC1|0|0|0|0|0|0|0|0|0|1|1|1|1|1|1|
+        |RC2|1|2|3|4|5|6|7|8|9|0|1|2|3|4|5|
+        |RC3|T|T|T|T|T|T|T|T|T|T|T|T|T|T|T|
+        -----------------------------------
+        | T |T|T|T|T|T|T|T|T|T|T|T|T|T|T|T|    SubGroup 1
+        | 1 |1|1|1|2|2|2|2|2|2|2|2|2|2|3|3|
+        | 6 |7|8|9|0|1|2|3|4|5|6|7|8|9|0|1|
+        |   | | | | | | | | | | | | | | | |
+        -----------------------------------
+        | T |T|T|T|T|T|T|T|T|T|T|T|T|T|T|T|	   SubGroup 2
+        | 3 |3|3|3|3|3|3|3|4|4|4|4|4|4|4|4|
+        | 2 |3|4|5|6|7|8|9|0|1|2|3|4|5|6|7|
+        |   | | | | | | | | | | | | | | | |
+        -----------------------------------
+        | T |T|T|T|T|T|T|T|T|T|T|T|T|T|T|T|    SubGroup 3
+        | 4 |4|5|5|5|5|5|5|5|5|5|5|6|6|6|6|
+        | 8 |9|0|1|2|3|4|5|6|7|8|9|0|1|2|3|
+        |   | | | | | | | | | | | | | | | |
+        -----------------------------------
+
+*   RC = Register for storing accumalted result
+*	T  = Thread ID
+*/
+
+template <WmmaInstr Instr, index_t WaveSize, typename = void>
+struct wmma_type
+{
+};
+
+// A-swizzled
+template <index_t WaveSize>
+struct wmma_type<WmmaInstr::wmma_f32_16x16x16_f16,
+                 WaveSize,
+                 typename std::enable_if_t<WaveSize == 32 || WaveSize == 64>>
+{
+    // Absolute fixing property
+    // * Data Pixel
+    static constexpr index_t m_per_wmma      = 16;
+    static constexpr index_t n_per_wmma      = 16;
+    static constexpr index_t k_per_wmma      = 16;
+    static constexpr index_t src_a_data_size = 2;
+    static constexpr index_t src_b_data_size = 2;
+    static constexpr index_t acc_data_size   = 4;
+    // * Thread mapping inside wave, num_thread_per_subgroups always alone N direction
+    static constexpr index_t num_thread_per_subgroups = n_per_wmma;
+
+    // Wave mode dependent propety
+    static constexpr index_t wave_size = Number<WaveSize>{};
+    // * Fixed in Navi3x, Will be wave mode dependent on Navi4x
+    static constexpr index_t num_src_a_vgprs_per_wave = m_per_wmma * src_a_data_size / 4;
+    static constexpr index_t num_src_b_vgprs_per_wave = n_per_wmma * src_b_data_size / 4;
+    // * num_acc_vgprs_per_wave alone M direction
+    // * num_subgroups alone M direction
+    static constexpr index_t num_acc_vgprs_per_wave =
+        m_per_wmma * n_per_wmma * acc_data_size / wave_size / 4;
+    static constexpr index_t num_subgroups = wave_size / num_thread_per_subgroups;
+
+    template <index_t MPerWmma, index_t NPerWmma, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        if constexpr(wave_size == 32)
+        {
+            intrin_wmma_f32_16x16x16_f16_w32<MPerWmma, NPerWmma>::Run(a, b, reg_c);
+        }
+        else if constexpr(wave_size == 64)
+        {
+            intrin_wmma_f32_16x16x16_f16_w64<MPerWmma, NPerWmma>::Run(a, b, reg_c);
+        }
+    }
+};
+
+template <index_t WaveSize>
+struct wmma_type<WmmaInstr::wmma_f32_16x16x16_bf16,
+                 WaveSize,
+                 typename std::enable_if_t<WaveSize == 32 || WaveSize == 64>>
+{
+    // Absolute fixing property
+    static constexpr index_t m_per_wmma               = 16;
+    static constexpr index_t n_per_wmma               = 16;
+    static constexpr index_t k_per_wmma               = 16;
+    static constexpr index_t src_a_data_size          = 2;
+    static constexpr index_t src_b_data_size          = 2;
+    static constexpr index_t acc_data_size            = 4;
+    static constexpr index_t num_thread_per_subgroups = n_per_wmma;
+
+    // Wave mode dependent propety
+    static constexpr index_t wave_size                = Number<WaveSize>{};
+    static constexpr index_t num_src_a_vgprs_per_wave = m_per_wmma * src_a_data_size / 4;
+    static constexpr index_t num_src_b_vgprs_per_wave = n_per_wmma * src_b_data_size / 4;
+    static constexpr index_t num_acc_vgprs_per_wave =
+        m_per_wmma * n_per_wmma * acc_data_size / wave_size / 4;
+    static constexpr index_t num_subgroups = wave_size / num_thread_per_subgroups;
+
+    template <index_t MPerWmma, index_t NPerWmma, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        if constexpr(wave_size == 32)
+        {
+            intrin_wmma_f32_16x16x16_bf16_w32<MPerWmma, NPerWmma>::Run(a, b, reg_c);
+        }
+        else if constexpr(wave_size == 64)
+        {
+            intrin_wmma_f32_16x16x16_bf16_w64<MPerWmma, NPerWmma>::Run(a, b, reg_c);
+        }
+    }
+};
+
+#ifdef CK_UNPACKED_ACC_DESC_LOGIC
+template <index_t WaveSize>
+struct wmma_type<WmmaInstr::wmma_f16_16x16x16_f16,
+                 WaveSize,
+                 typename std::enable_if_t<WaveSize == 32 || WaveSize == 64>>
+{
+    // Absolute fixing property
+    static constexpr index_t m_per_wmma               = 16;
+    static constexpr index_t n_per_wmma               = 16;
+    static constexpr index_t k_per_wmma               = 16;
+    static constexpr index_t src_a_data_size          = 2;
+    static constexpr index_t src_b_data_size          = 2;
+    static constexpr index_t acc_data_size            = 2;
+    static constexpr index_t num_thread_per_subgroups = n_per_wmma;
+
+    // Wave mode dependent propety
+    static constexpr index_t wave_size                = Number<WaveSize>{};
+    static constexpr index_t num_src_a_vgprs_per_wave = m_per_wmma * src_a_data_size / 4;
+    static constexpr index_t num_src_b_vgprs_per_wave = n_per_wmma * src_b_data_size / 4;
+    static constexpr index_t num_acc_vgprs_per_wave =
+        m_per_wmma * n_per_wmma * acc_data_size / wave_size / 4;
+    static constexpr index_t num_subgroups = wave_size / num_thread_per_subgroups;
+
+    template <index_t MPerWmma,
+              index_t NPerWmma,
+              index_t Opsel,
+              class FloatA,
+              class FloatB,
+              class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        if constexpr(wave_size == 32)
+        {
+            intrin_wmma_f16_16x16x16_f16_w32<MPerWmma, NPerWmma, Opsel>::Run(a, b, reg_c);
+        }
+        else if constexpr(wave_size == 64)
+        {
+            intrin_wmma_f16_16x16x16_f16_w64<MPerWmma, NPerWmma, Opsel>::Run(a, b, reg_c);
+        }
+    }
+};
+
+template <index_t WaveSize>
+struct wmma_type<WmmaInstr::wmma_bf16_16x16x16_bf16,
+                 WaveSize,
+                 typename std::enable_if_t<WaveSize == 32 || WaveSize == 64>>
+{
+    // Absolute fixing property
+    static constexpr index_t m_per_wmma               = 16;
+    static constexpr index_t n_per_wmma               = 16;
+    static constexpr index_t k_per_wmma               = 16;
+    static constexpr index_t src_a_data_size          = 2;
+    static constexpr index_t src_b_data_size          = 2;
+    static constexpr index_t acc_data_size            = 2;
+    static constexpr index_t num_thread_per_subgroups = n_per_wmma;
+
+    // Wave mode dependent propety
+    static constexpr index_t wave_size                = Number<WaveSize>{};
+    static constexpr index_t num_src_a_vgprs_per_wave = m_per_wmma * src_a_data_size / 4;
+    static constexpr index_t num_src_b_vgprs_per_wave = n_per_wmma * src_b_data_size / 4;
+    static constexpr index_t num_acc_vgprs_per_wave =
+        m_per_wmma * n_per_wmma * acc_data_size / wave_size / 4;
+    static constexpr index_t num_subgroups = wave_size / num_thread_per_subgroups;
+
+    template <index_t MPerWmma,
+              index_t NPerWmma,
+              index_t Opsel,
+              class FloatA,
+              class FloatB,
+              class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        if constexpr(wave_size == 32)
+        {
+            intrin_wmma_bf16_16x16x16_bf16_w32<MPerWmma, NPerWmma, Opsel>::Run(a, b, reg_c);
+        }
+        else if constexpr(wave_size == 64)
+        {
+            intrin_wmma_bf16_16x16x16_bf16_w64<MPerWmma, NPerWmma, Opsel>::Run(a, b, reg_c);
+        }
+    }
+};
+
+#endif
+
+template <index_t WaveSize>
+struct wmma_type<WmmaInstr::wmma_i32_16x16x16_iu8,
+                 WaveSize,
+                 typename std::enable_if_t<WaveSize == 32 || WaveSize == 64>>
+{
+    // Absolute fixing property
+    static constexpr index_t m_per_wmma               = 16;
+    static constexpr index_t n_per_wmma               = 16;
+    static constexpr index_t k_per_wmma               = 16;
+    static constexpr index_t src_a_data_size          = 2;
+    static constexpr index_t src_b_data_size          = 2;
+    static constexpr index_t acc_data_size            = 4;
+    static constexpr index_t num_thread_per_subgroups = n_per_wmma;
+
+    // Wave mode dependent propety
+    static constexpr index_t wave_size                = Number<WaveSize>{};
+    static constexpr index_t num_src_a_vgprs_per_wave = m_per_wmma * src_a_data_size / 4;
+    static constexpr index_t num_src_b_vgprs_per_wave = n_per_wmma * src_b_data_size / 4;
+    static constexpr index_t num_acc_vgprs_per_wave =
+        m_per_wmma * n_per_wmma * acc_data_size / wave_size / 4;
+    static constexpr index_t num_subgroups = wave_size / num_thread_per_subgroups;
+
+    template <index_t MPerWmma,
+              index_t NPerWmma,
+              class FloatA,
+              class FloatB,
+              class FloatC,
+              bool neg_a = false,
+              bool neg_b = false,
+              bool clamp = false>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        if constexpr(wave_size == 32)
+        {
+            intrin_wmma_i32_16x16x16_iu8_w32<MPerWmma, NPerWmma, neg_a, neg_b, clamp>::Run(
+                a, b, reg_c);
+        }
+        else if constexpr(wave_size == 64)
+        {
+            intrin_wmma_i32_16x16x16_iu8_w64<MPerWmma, NPerWmma, neg_a, neg_b, clamp>::Run(
+                a, b, reg_c);
+        }
+    }
+};
+
+template <typename src_type_a,
+          typename src_type_b,
+          typename dst_type,
+          index_t MPerWmma,
+          index_t NPerWmma>
+struct WmmaSelector
+{
+    template <typename src_type_a_,
+              typename src_type_b_,
+              typename dst_type_,
+              index_t MPerWmma_,
+              index_t NPerWmma_>
+    static constexpr auto GetWmma();
+
+    template <>
+    static constexpr auto GetWmma<half_t, half_t, float, 16, 16>()
+    {
+        return WmmaInstr::wmma_f32_16x16x16_f16;
+    }
+
+    template <>
+    static constexpr auto GetWmma<bhalf_t, bhalf_t, float, 16, 16>()
+    {
+        return WmmaInstr::wmma_f32_16x16x16_bf16;
+    }
+
+    template <>
+    static constexpr auto GetWmma<half_t, half_t, half_t, 16, 16>()
+    {
+        return WmmaInstr::wmma_f16_16x16x16_f16;
+    }
+
+    template <>
+    static constexpr auto GetWmma<bhalf_t, bhalf_t, bhalf_t, 16, 16>()
+    {
+        return WmmaInstr::wmma_bf16_16x16x16_bf16;
+    }
+
+    template <>
+    static constexpr auto GetWmma<int8_t, int8_t, int, 16, 16>()
+    {
+        return WmmaInstr::wmma_i32_16x16x16_iu8;
+    }
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+    template <>
+    static constexpr auto GetWmma<int4_t, int, 16, 16>()
+    {
+        return WmmaInstr::wmma_i32_16x16x16_iu4;
+    }
+#endif
+    // get_warp_size do not return the correct wavesize, hardcode to 32 as workaround
+    static constexpr auto selected_wmma =
+        wmma_type<GetWmma<src_type_a, src_type_b, dst_type, MPerWmma, NPerWmma>(), Number<32>{}>{};
+
+    __host__ __device__ constexpr WmmaSelector()
+    {
+        static_assert(selected_wmma.m_per_wmma == 16, "WRONG! WMMA_M must equal to 16");
+
+        static_assert(selected_wmma.m_per_wmma == 16, "WRONG! WMMA_M must equal to 16");
+
+        static_assert(selected_wmma.k_per_wmma == 16, "WRONG! WMMA_M must equal to 16");
+
+        static_assert(selected_wmma.wave_size * selected_wmma.num_acc_vgprs_per_wave *
+                              selected_wmma.acc_data_size ==
+                          selected_wmma.m_per_wmma * selected_wmma.n_per_wmma * 4,
+                      "WRONG! Invalid Number of Accumulator Register");
+    }
+};
+
+template <typename src_type_a,
+          typename src_type_b,
+          typename dst_type,
+          index_t MPerWmma,
+          index_t NPerWmma,
+          index_t KPack,
+          bool TransposeC = false>
+struct WmmaGemm
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    using CIndex   = MultiIndex<2>;
+    using CIndex4D = MultiIndex<4>;
+
+    __host__ __device__ constexpr WmmaGemm()
+    {
+        static_assert(NPerWmma == 16 && MPerWmma == 16,
+                      "Only support GemmNPerWmma == 16 and GemmMPerWmma == 16 for wmma");
+
+        static_assert(KPack == wmma_instr.k_per_wmma, "KPack should be k_per_wmma");
+    }
+
+    // WMMA output supporting C = A * B
+    // Vector Write
+    // MPerWMMA_NPerWMMA -> MSubGroup_..._NPerWMMA_MAccVgprPerWave
+    template <typename CDesc_MBlockxRepeat_MWave_MPerWMMA_NBlockxRepeat_NWave_NPerWMMA>
+    __host__ __device__ static constexpr auto
+    MakeCDesc_MBlockxRepeat_MWave_MSubGroup_NBlockxRepeat_NWave_NThreadPerSubGroup_MAccVgprs(
+        const CDesc_MBlockxRepeat_MWave_MPerWMMA_NBlockxRepeat_NWave_NPerWMMA&
+            c_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma)
+    {
+        const auto MBlockxRepeat =
+            c_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma.GetLength(I0);
+        const auto NBlockxRepeat =
+            c_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma.GetLength(I3);
+        const auto MWave =
+            c_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma.GetLength(I1);
+        const auto NWave =
+            c_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma.GetLength(I4);
+
+        return transform_tensor_descriptor(
+            c_desc_mblockxrepeat_mwave_mperwmma_nblockxrepeat_nwave_nperwmma,
+            make_tuple(
+                make_pass_through_transform(MBlockxRepeat),
+                make_pass_through_transform(MWave),
+                make_unmerge_transform(make_tuple(Number<wmma_instr.num_subgroups>{},
+                                                  Number<wmma_instr.num_acc_vgprs_per_wave>{})),
+                make_pass_through_transform(NBlockxRepeat),
+                make_pass_through_transform(NWave),
+                make_pass_through_transform(Number<wmma_instr.num_thread_per_subgroups>{})),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5>{}),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2, 6>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5>{}));
+    }
+
+    __device__ static constexpr index_t GetRegSizePerWmma()
+    {
+        return wmma_instr.num_acc_vgprs_per_wave;
+    }
+
+    __device__ static constexpr index_t GetWaveSize() { return wmma_instr.wave_size; }
+
+    template <class FloatA, class FloatB, class FloatC>
+    __device__ void Run(const FloatA& p_a_wave, const FloatB& p_b_wave, FloatC& p_c_thread) const
+    {
+        static_assert(
+            (is_same<src_type_a, half_t>::value && is_same<src_type_b, half_t>::value &&
+             is_same<dst_type, float>::value) ||
+                (is_same<src_type_a, bhalf_t>::value && is_same<src_type_b, bhalf_t>::value &&
+                 is_same<dst_type, float>::value) ||
+                (is_same<src_type_a, half_t>::value && is_same<src_type_b, half_t>::value &&
+                 is_same<dst_type, half_t>::value) ||
+                (is_same<src_type_a, bhalf_t>::value && is_same<src_type_b, bhalf_t>::value &&
+                 is_same<dst_type, bhalf_t>::value) ||
+                (is_same<src_type_a, int8_t>::value && is_same<src_type_b, int8_t>::value &&
+                 is_same<dst_type, int32_t>::value)
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+                || (is_same<src_type_a, int4_t>::value && is_same<src_type_b, int4_t>::value &&
+                    is_same<dst_type, int32_t>::value)
+#endif
+                ,
+            "base type couple must be (half, float), (bhalf, float), (half, half), (bhalf, bhalf), "
+            "(int8, int32) or (int4, int32)!");
+        if constexpr(!TransposeC)
+        {
+            wmma_instr.template run<MPerWmma, NPerWmma>(p_a_wave, p_b_wave, p_c_thread);
+        }
+        else
+        {
+            wmma_instr.template run<MPerWmma, NPerWmma>(p_b_wave, p_a_wave, p_c_thread);
+        }
+    }
+
+    __device__ static auto GetLaneId() { return get_thread_local_1d_id() % wmma_instr.wave_size; }
+
+    __device__ static auto GetSubGroupId()
+    {
+        return (GetLaneId() / wmma_instr.num_thread_per_subgroups) % wmma_instr.num_subgroups;
+    }
+
+    __device__ static auto GetLaneIdUnderSubGroup()
+    {
+        return GetLaneId() % wmma_instr.num_thread_per_subgroups;
+    }
+    __device__ static auto GetSwizzledLaneIdLow()
+    {
+        return ((GetLaneIdUnderSubGroup() & 1) << 3) | (GetLaneIdUnderSubGroup() >> 1);
+    }
+
+    __host__ __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        return GetSwizzledLaneIdLow();
+    }
+
+    __host__ __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        return GetLaneIdUnderSubGroup();
+    }
+
+    __device__ static CIndex GetBeginOfThreadBlk()
+    {
+        index_t n_offset = GetLaneIdUnderSubGroup();
+        index_t m_offset = GetSubGroupId() * wmma_instr.num_acc_vgprs_per_wave;
+
+        return TransposeC ? CIndex{n_offset, m_offset} : CIndex{m_offset, n_offset};
+    }
+
+    static constexpr auto wmma =
+        WmmaSelector<src_type_a, src_type_b, dst_type, MPerWmma, NPerWmma>{};
+    static constexpr auto wmma_instr = wmma.selected_wmma;
+
+    __host__ __device__ static constexpr auto
+    GetCMSubGroupNThreadPerSubGroupMAccVgprsThreadBlkLengths()
+    {
+        return make_tuple(I1, I1, Number<wmma_instr.num_acc_vgprs_per_wave>{});
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp b/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
new file mode 100644
index 0000000000..814969ef42
--- /dev/null
+++ b/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
@@ -0,0 +1,967 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/math.hpp"
+#include "ck/utility/amd_xdlops.hpp"
+
+namespace ck {
+
+enum struct MfmaInstr
+{
+    mfma_f32_32x32x1xf32 = 0,
+    mfma_f32_16x16x1xf32,
+    mfma_f32_4x4x1xf32,
+    mfma_f32_32x32x2xf32,
+    mfma_f32_16x16x4xf32,
+    mfma_f32_32x32x4f16,
+    mfma_f32_16x16x4f16,
+    mfma_f32_4x4x4f16,
+    mfma_f32_32x32x8f16,
+    mfma_f32_16x16x16f16,
+    mfma_f32_32x32x8bf16_1k,
+    mfma_f32_16x16x16bf16_1k,
+    mfma_f32_32x32x4bf16,
+    mfma_f32_16x16x8bf16,
+    mfma_i32_32x32x8i8,
+    mfma_i32_16x16x16i8,
+    mfma_i32_32x32x16i8,
+    mfma_i32_16x16x32i8,
+    mfma_f64_16x16x4f64,
+    mfma_f32_32x32x16f8f8,
+    mfma_f32_16x16x32f8f8
+};
+
+template <MfmaInstr instr>
+struct mfma_type;
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_32x32x1xf32>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 2;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 1;
+    static constexpr bool is_k_reduction         = false;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_32x32x1f32<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_32x32x2xf32>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 1;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_32x32x2f32<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_16x16x4xf32>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 1;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_16x16x4f32<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_16x16x1xf32>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 4;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 1;
+    static constexpr bool is_k_reduction         = false;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_16x16x1f32<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+// treat 4x4x1 as a single-blk 4x64 mfma
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_4x4x1xf32>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 64;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 1;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 4;
+    static constexpr index_t n_per_blk           = 64;
+    static constexpr index_t k_per_blk           = 1;
+    static constexpr bool is_k_reduction         = false;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_4x4x1f32<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_32x32x4f16>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 2;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = false;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_32x32x4f16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_32x32x8f16>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_32x32x8f16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_16x16x16f16>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_16x16x16f16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_16x16x4f16>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 4;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = false;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_16x16x4f16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_4x4x4f16>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 64;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 1;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 4;
+    static constexpr index_t n_per_blk           = 64;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = false;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_4x4x4f16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_32x32x8bf16_1k>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_32x32x8bf16_1k<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_16x16x16bf16_1k>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_16x16x16bf16_1k<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_32x32x4bf16>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 2;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_32x32x4bf16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_16x16x8bf16>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 2;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_16x16x8bf16<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_i32_32x32x8i8>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_i32_32x32x8i8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_i32_16x16x16i8>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 4;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_i32_16x16x16i8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_i32_32x32x16i8>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 8;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_i32_32x32x16i8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_i32_16x16x32i8>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 8;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_i32_16x16x32i8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f64_16x16x4f64>
+{
+    static constexpr index_t group_size          = 1;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 4; // group_size * num_groups_per_blk;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4; // wave_size / num_threads_per_blk;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 1;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f64_16x16x4f64<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_32x32x16f8f8>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 4;
+    static constexpr index_t num_regs_per_blk    = 16;
+    static constexpr index_t num_threads_per_blk = 32;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 2;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 32;
+    static constexpr index_t n_per_blk           = 32;
+    static constexpr index_t k_per_blk           = 8;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_32x32x16f8f8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct mfma_type<MfmaInstr::mfma_f32_16x16x32f8f8>
+{
+    static constexpr index_t group_size          = 4;
+    static constexpr index_t num_groups_per_blk  = 1;
+    static constexpr index_t num_regs_per_blk    = 4;
+    static constexpr index_t num_threads_per_blk = 16;
+    static constexpr index_t wave_size           = 64;
+    static constexpr index_t num_input_blks      = 4;
+    static constexpr index_t num_output_blks     = 1;
+    static constexpr index_t m_per_blk           = 16;
+    static constexpr index_t n_per_blk           = 16;
+    static constexpr index_t k_per_blk           = 8;
+    static constexpr bool is_k_reduction         = true;
+
+    template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
+    __device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
+    {
+        intrin_mfma_f32_16x16x32f8f8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
+    }
+};
+
+template <typename base_type, index_t MPerXdlops, index_t NPerXdlops>
+struct MfmaSelector
+{
+    template <typename base_type_, index_t MPerXdlops_, index_t NPerXdlops_>
+    static constexpr auto GetMfma();
+
+    template <>
+    static constexpr auto GetMfma<double, 16, 16>()
+    {
+        return MfmaInstr::mfma_f64_16x16x4f64;
+    }
+
+    template <>
+    static constexpr auto GetMfma<float, 64, 64>()
+    {
+        return MfmaInstr::mfma_f32_32x32x1xf32;
+    }
+
+    template <>
+    static constexpr auto GetMfma<float, 32, 64>()
+    {
+        return MfmaInstr::mfma_f32_32x32x1xf32;
+    }
+
+    template <>
+    static constexpr auto GetMfma<float, 16, 64>()
+    {
+        return MfmaInstr::mfma_f32_16x16x1xf32;
+    }
+
+    template <>
+    static constexpr auto GetMfma<float, 8, 64>()
+    {
+        return MfmaInstr::mfma_f32_4x4x1xf32;
+    }
+
+    template <>
+    static constexpr auto GetMfma<float, 4, 64>()
+    {
+        return MfmaInstr::mfma_f32_4x4x1xf32;
+    }
+
+    template <>
+    static constexpr auto GetMfma<float, 32, 32>()
+    {
+        return MfmaInstr::mfma_f32_32x32x2xf32;
+    }
+
+    template <>
+    static constexpr auto GetMfma<float, 16, 16>()
+    {
+        return MfmaInstr::mfma_f32_16x16x4xf32;
+    }
+
+    template <>
+    static constexpr auto GetMfma<half_t, 64, 64>()
+    {
+        return MfmaInstr::mfma_f32_32x32x4f16;
+    }
+
+    template <>
+    static constexpr auto GetMfma<half_t, 32, 64>()
+    {
+        return MfmaInstr::mfma_f32_32x32x4f16;
+    }
+
+    template <>
+    static constexpr auto GetMfma<half_t, 32, 32>()
+    {
+        return MfmaInstr::mfma_f32_32x32x8f16;
+    }
+
+    template <>
+    static constexpr auto GetMfma<half_t, 16, 16>()
+    {
+        return MfmaInstr::mfma_f32_16x16x16f16;
+    }
+
+    template <>
+    static constexpr auto GetMfma<half_t, 16, 64>()
+    {
+        return MfmaInstr::mfma_f32_16x16x4f16;
+    }
+
+    template <>
+    static constexpr auto GetMfma<half_t, 8, 64>()
+    {
+        return MfmaInstr::mfma_f32_4x4x4f16;
+    }
+
+    template <>
+    static constexpr auto GetMfma<half_t, 4, 64>()
+    {
+        return MfmaInstr::mfma_f32_4x4x4f16;
+    }
+
+    template <>
+    static constexpr auto GetMfma<bhalf_t, 32, 32>()
+    {
+#if defined(CK_USE_AMD_MFMA_BF16_1K_OP)
+        return MfmaInstr::mfma_f32_32x32x8bf16_1k;
+#else
+        return MfmaInstr::mfma_f32_32x32x4bf16;
+#endif
+    }
+
+    template <>
+    static constexpr auto GetMfma<bhalf_t, 16, 16>()
+    {
+#if defined(CK_USE_AMD_MFMA_BF16_1K_OP)
+        return MfmaInstr::mfma_f32_16x16x16bf16_1k;
+#else
+        return MfmaInstr::mfma_f32_16x16x8bf16;
+#endif
+    }
+
+#if defined(CK_USE_AMD_MFMA_GFX940)
+    template <>
+    static constexpr auto GetMfma<int8_t, 32, 32>()
+    {
+        return MfmaInstr::mfma_i32_32x32x16i8;
+    }
+    template <>
+    static constexpr auto GetMfma<int8_t, 16, 16>()
+    {
+        return MfmaInstr::mfma_i32_16x16x32i8;
+    }
+#else
+    template <>
+    static constexpr auto GetMfma<int8_t, 32, 32>()
+    {
+        return MfmaInstr::mfma_i32_32x32x8i8;
+    }
+    template <>
+    static constexpr auto GetMfma<int8_t, 16, 16>()
+    {
+        return MfmaInstr::mfma_i32_16x16x16i8;
+    }
+#endif
+
+    template <>
+    static constexpr auto GetMfma<f8_t, 32, 32>()
+    {
+        return MfmaInstr::mfma_f32_32x32x16f8f8;
+    }
+
+    template <>
+    static constexpr auto GetMfma<f8_t, 16, 16>()
+    {
+        return MfmaInstr::mfma_f32_16x16x32f8f8;
+    }
+
+    static constexpr auto selected_mfma = mfma_type<GetMfma<base_type, MPerXdlops, NPerXdlops>()>{};
+
+    __host__ __device__ constexpr MfmaSelector()
+    {
+        static_assert(selected_mfma.group_size * selected_mfma.num_groups_per_blk ==
+                          selected_mfma.num_regs_per_blk,
+                      "wrong! num_regs_per_blk");
+
+        static_assert(selected_mfma.num_threads_per_blk == selected_mfma.n_per_blk,
+                      "n_per_blk != num_threads_per_blk");
+
+        static_assert(selected_mfma.num_regs_per_blk * selected_mfma.num_input_blks ==
+                          selected_mfma.m_per_blk,
+                      "m_per_blk != num_input_blks * num_regs_per_blk");
+
+        static_assert(selected_mfma.num_output_blks == selected_mfma.num_input_blks ||
+                          selected_mfma.num_output_blks == 1,
+                      "incorrect num_output_blks");
+
+        static_assert(selected_mfma.num_regs_per_blk * selected_mfma.wave_size ==
+                          selected_mfma.m_per_blk * selected_mfma.n_per_blk,
+                      "num_regs_per_blk incorrect");
+
+        static_assert(selected_mfma.is_k_reduction ||
+                          (selected_mfma.num_input_blks == selected_mfma.num_output_blks),
+                      "is_k_reduction wrong!");
+    }
+
+    static constexpr bool IsABroadcast()
+    {
+        static_assert(NPerXdlops >= MPerXdlops, "only support ABroadcast");
+        return true;
+    }
+
+    static constexpr index_t GetKPerXdlops()
+    {
+        return (selected_mfma.is_k_reduction ? selected_mfma.num_input_blks : 1) *
+               selected_mfma.k_per_blk;
+    }
+
+    static constexpr index_t GetK1PerXdlops() { return selected_mfma.k_per_blk; }
+};
+
+template <typename base_type,
+          index_t MPerXdlops,
+          index_t NPerXdlops,
+          index_t KPack,
+          bool TransposeC = false>
+struct XdlopsGemm
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+    static constexpr auto I5 = Number<5>{};
+
+    using CIndex   = MultiIndex<2>;
+    using CIndex4D = MultiIndex<4>;
+
+    __device__ static constexpr index_t GetNumBlks() { return mfma_instr.num_output_blks; }
+
+    __device__ static constexpr index_t GetNumXdlops()
+    {
+        return MPerXdlops * NPerXdlops /
+               (mfma_instr.m_per_blk * mfma_instr.n_per_blk * mfma_instr.num_output_blks);
+    }
+
+    __host__ __device__ constexpr XdlopsGemm()
+    {
+        static_assert(NPerXdlops == 4 || NPerXdlops == 8 || NPerXdlops == 16 || NPerXdlops == 32 ||
+                          NPerXdlops == 64,
+                      "Only support GemmNPerXdlops == 4, 8, 16, 32 or 64 for xdlops");
+
+        static_assert(MPerXdlops == 4 || MPerXdlops == 8 || MPerXdlops == 16 || MPerXdlops == 32 ||
+                          MPerXdlops == 64,
+                      "Only support GemmMPerXdlops == 4, 8, 16, 32 or 64 for xdlops");
+
+        static_assert(KPack % mfma_instr.k_per_blk == 0, "KPack cannot be divided by k_per_blk");
+    }
+
+    // XDL output supporting C = A * B
+    // M2_N2 -> M2_M3_M4_N2
+    template <typename CDesc_M0_N0_M1_N1_M2_N2>
+    __host__ __device__ static constexpr auto
+    MakeCDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(const CDesc_M0_N0_M1_N1_M2_N2& c_desc_m0_n0_m1_n1_m2_n2)
+    {
+        const auto M0 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I0);
+        const auto N0 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I1);
+        const auto M1 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I2);
+        const auto N1 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I3);
+
+        return transform_tensor_descriptor(
+            c_desc_m0_n0_m1_n1_m2_n2,
+            make_tuple(make_pass_through_transform(M0),
+                       make_pass_through_transform(N0),
+                       make_pass_through_transform(M1),
+                       make_pass_through_transform(N1),
+                       make_unmerge_transform(make_tuple(Number<mfma_instr.num_groups_per_blk>{},
+                                                         Number<mfma_instr.num_input_blks>{},
+                                                         Number<mfma_instr.group_size>{})),
+                       make_pass_through_transform(Number<mfma_instr.num_threads_per_blk>{})),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5>{}),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4, 5, 6>{},
+                       Sequence<7>{}));
+    }
+
+    // transposed XDL output supporting C' = B' * A'
+    // M2_N2 -> M2_N2_N3_N4
+    template <typename CDesc_M0_N0_M1_N1_M2_N2>
+    __host__ __device__ static constexpr auto
+    MakeCDescriptor_M0_N0_M1_N1_M2_N2_N3_N4(const CDesc_M0_N0_M1_N1_M2_N2& c_desc_m0_n0_m1_n1_m2_n2)
+    {
+        const auto M0 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I0);
+        const auto N0 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I1);
+        const auto M1 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I2);
+        const auto N1 = c_desc_m0_n0_m1_n1_m2_n2.GetLength(I3);
+
+        return transform_tensor_descriptor(
+            c_desc_m0_n0_m1_n1_m2_n2,
+            make_tuple(make_pass_through_transform(M0),
+                       make_pass_through_transform(N0),
+                       make_pass_through_transform(M1),
+                       make_pass_through_transform(N1),
+                       make_pass_through_transform(Number<mfma_instr.num_threads_per_blk>{}),
+                       make_unmerge_transform(make_tuple(Number<mfma_instr.num_groups_per_blk>{},
+                                                         Number<mfma_instr.num_input_blks>{},
+                                                         Number<mfma_instr.group_size>{}))),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5>{}),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5, 6, 7>{}));
+    }
+
+    template <typename CDesc_G_M0_N0_M1_N1_M2_N2>
+    __host__ __device__ static constexpr auto MakeCDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2(
+        const CDesc_G_M0_N0_M1_N1_M2_N2& c_desc_g_m0_n0_m1_n1_m2_n2)
+    {
+        const auto G  = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I0);
+        const auto M0 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I1);
+        const auto N0 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I2);
+        const auto M1 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I3);
+        const auto N1 = c_desc_g_m0_n0_m1_n1_m2_n2.GetLength(I4);
+
+        return transform_tensor_descriptor(
+            c_desc_g_m0_n0_m1_n1_m2_n2,
+            make_tuple(make_pass_through_transform(G),
+                       make_pass_through_transform(M0),
+                       make_pass_through_transform(N0),
+                       make_pass_through_transform(M1),
+                       make_pass_through_transform(N1),
+                       make_unmerge_transform(make_tuple(mfma_instr.num_groups_per_blk,
+                                                         mfma_instr.num_input_blks,
+                                                         mfma_instr.group_size)),
+                       make_pass_through_transform(mfma_instr.num_threads_per_blk)),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5>{},
+                       Sequence<6>{}),
+            make_tuple(Sequence<0>{},
+                       Sequence<1>{},
+                       Sequence<2>{},
+                       Sequence<3>{},
+                       Sequence<4>{},
+                       Sequence<5, 6, 7>{},
+                       Sequence<8>{}));
+    }
+
+    __device__ static constexpr index_t GetRegSizePerXdlops()
+    {
+        return MPerXdlops * NPerXdlops / mfma_instr.wave_size;
+    }
+
+    __device__ static constexpr index_t GetWaveSize() { return mfma_instr.wave_size; }
+
+    template <class FloatA, class FloatB, class FloatC>
+    __device__ void Run(const FloatA& p_a_wave, const FloatB& p_b_wave, FloatC& p_c_thread) const
+    {
+        static_assert(is_same<base_type, double>::value || is_same<base_type, float>::value ||
+                          is_same<base_type, half_t>::value || is_same<base_type, bhalf_t>::value ||
+                          is_same<base_type, int8_t>::value || is_same<base_type, f8_t>::value,
+                      "base base_type must be double, float, half, bfloat16, and int8_t!");
+
+        static_for<0, KPack / mfma_instr.k_per_blk, 1>{}([&](auto k) {
+            if constexpr(!TransposeC)
+            {
+                mfma_instr.template run<MPerXdlops, NPerXdlops>(
+                    p_a_wave[k], p_b_wave[k], p_c_thread);
+            }
+            else
+            {
+                mfma_instr.template run<MPerXdlops, NPerXdlops>(
+                    p_b_wave[k], p_a_wave[k], p_c_thread);
+            }
+        });
+    }
+
+    __device__ static auto GetLaneId() { return get_thread_local_1d_id() % mfma_instr.wave_size; }
+
+    __device__ static auto GetBlkIdx()
+    {
+        const auto laneId = GetLaneId();
+
+        constexpr auto threadidx_to_blk_idx_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(make_merge_transform(
+                make_tuple(1, mfma_instr.num_input_blks, mfma_instr.num_threads_per_blk))),
+            make_tuple(Sequence<0, 1, 2>{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto blk_idx =
+            threadidx_to_blk_idx_adaptor.CalculateBottomIndex(make_multi_index(laneId));
+
+        const auto blk_id = blk_idx[I1];
+        const auto blk_td = blk_idx[I2];
+
+        return make_tuple(blk_id, blk_td);
+    }
+
+    __host__ __device__ static auto CalculateAThreadOriginDataIndex()
+    {
+        const auto laneId  = GetLaneId();
+        const auto blk_idx = GetBlkIdx();
+
+        const auto blk_id = blk_idx[I0];
+        const auto blk_td = blk_idx[I1];
+
+        if constexpr(mfma_instr.is_k_reduction)
+        {
+            return make_tuple(blk_id, blk_td);
+        }
+        else
+        {
+            return make_tuple(0, laneId);
+        }
+    }
+
+    __host__ __device__ static auto CalculateBThreadOriginDataIndex()
+    {
+        const auto laneId  = GetLaneId();
+        const auto blk_idx = GetBlkIdx();
+
+        const auto blk_id = blk_idx[I0];
+        const auto blk_td = blk_idx[I1];
+
+        if constexpr(mfma_instr.is_k_reduction)
+        {
+            return make_tuple(blk_id, blk_td);
+        }
+        else
+        {
+            return make_tuple(0, laneId);
+        }
+    }
+
+    __device__ static CIndex GetBeginOfThreadBlk(index_t xdlops_i, index_t blk_i)
+    {
+        const auto blk_idx = GetBlkIdx();
+
+        const auto blk_id = blk_idx[I0];
+        const auto blk_td = blk_idx[I1];
+
+        index_t n_offset = blk_i * mfma_instr.n_per_blk + blk_td;
+        index_t m_offset = xdlops_i * mfma_instr.m_per_blk + blk_id * mfma_instr.group_size;
+
+        return TransposeC ? CIndex{n_offset, m_offset} : CIndex{m_offset, n_offset};
+    }
+
+    __device__ static CIndex4D GetBeginOfThreadBlk4D(index_t /* xdlops_i */, index_t /* blk_i */)
+    {
+        const auto blk_idx = GetBlkIdx();
+
+        const auto blk_id = blk_idx[I0];
+        const auto blk_td = blk_idx[I1];
+
+        return TransposeC ? CIndex4D{blk_td, I0, blk_id, I0} : CIndex4D{I0, blk_id, I0, blk_td};
+    }
+
+    static constexpr auto mfma = MfmaSelector<base_type, MPerXdlops, NPerXdlops>{};
+
+    static constexpr auto mfma_instr = mfma.selected_mfma;
+
+    static constexpr auto KPerXdlops  = mfma.GetKPerXdlops();
+    static constexpr auto K1PerXdlops = mfma.GetK1PerXdlops();
+    static constexpr auto K0PerXdlops = KPerXdlops / K1PerXdlops;
+
+    __host__ __device__ static constexpr auto GetCM0M1M2NThreadBlkLengths()
+    {
+        return make_tuple(
+            Number<mfma_instr.num_groups_per_blk>{}, I1, Number<mfma_instr.group_size>{}, I1);
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp b/include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp
new file mode 100644
index 0000000000..ea27a40ce3
--- /dev/null
+++ b/include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp
@@ -0,0 +1,288 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
+
+namespace ck {
+namespace tensor_operation {
+
+// assume C[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          device::TensorSpecialization TensorSpec>
+static auto MakeGridDescriptorPair(const std::vector<index_t>& gs_ms_ns_lengths_vec,
+                                   const std::vector<index_t>& gs_ms_ns_strides_vec)
+{
+    if(!(gs_ms_ns_lengths_vec.size() == NumDimG + NumDimM + NumDimN &&
+         gs_ms_ns_strides_vec.size() == NumDimG + NumDimM + NumDimN))
+    {
+        throw std::runtime_error("wrong! dimension must match input lengths");
+    }
+
+    const auto to_tuple = [&](auto& vec, auto start, auto end) {
+        return generate_tuple([&](auto i) { return vec[start + i]; }, Number<end - start>{});
+    };
+
+    const auto gs_ms_ns_lengths =
+        to_tuple(gs_ms_ns_lengths_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+    const auto gs_ms_ns_strides =
+        to_tuple(gs_ms_ns_strides_vec, Number<0>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+    // dimension Ids for G0, G1, ...
+    constexpr auto gDimIds = typename arithmetic_sequence_gen<0, NumDimG, 1>::type{};
+
+    // dimension Ids for M0, M1, ...
+    constexpr auto mDimIds =
+        typename arithmetic_sequence_gen<NumDimG, NumDimG + NumDimM, 1>::type{};
+
+    // dimension Ids for N0, N1, ...
+    constexpr auto nDimIds =
+        typename arithmetic_sequence_gen<NumDimG + NumDimM, NumDimG + NumDimM + NumDimN, 1>::type{};
+
+    // lengths for G0, G1, ...
+    const auto gLengths = get_container_subset(gs_ms_ns_lengths, gDimIds);
+
+    // lengths for M0, M1, ...
+    const auto mLengths = get_container_subset(gs_ms_ns_lengths, mDimIds);
+
+    // lengths for N0, N1, ...
+    const auto nLengths = get_container_subset(gs_ms_ns_lengths, nDimIds);
+
+    if constexpr(TensorSpec == device::TensorSpecialization::Packed)
+    {
+        auto G = container_reduce(gLengths, math::multiplies{}, Number<1>{});
+        auto M = container_reduce(mLengths, math::multiplies{}, Number<1>{});
+        auto N = container_reduce(nLengths, math::multiplies{}, Number<1>{});
+        const auto grid_desc_g_mraw_nraw = make_naive_tensor_descriptor(
+            make_tuple(G, M, N),
+            make_tuple(gs_ms_ns_strides[Number<NumDimG - 1>{}],
+                       gs_ms_ns_strides[Number<NumDimG + NumDimM - 1>{}],
+                       gs_ms_ns_strides[Number<NumDimG + NumDimM + NumDimN - 1>{}]));
+
+        const auto grid_desc_mraw_nraw = make_naive_tensor_descriptor(
+            make_tuple(M, N),
+            make_tuple(gs_ms_ns_strides[Number<NumDimG + NumDimM - 1>{}],
+                       gs_ms_ns_strides[Number<NumDimG + NumDimM + NumDimN - 1>{}]));
+
+        return std::make_pair(grid_desc_g_mraw_nraw, grid_desc_mraw_nraw);
+    }
+    else
+    {
+        // naive tensor C[G0, G1, ..., M0, M1, M2, ..., N0, N1, N2...]
+        const auto grid_desc_gs_ms_ns =
+            make_naive_tensor_descriptor(gs_ms_ns_lengths, gs_ms_ns_strides);
+
+        // transformed tensor C[G = G0 * G1 * ..., MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
+        // N2 * ...]
+        // Note: This does not require padding as it only provides G offset calculation. Technically
+        // descriptor for only G is needed. Here we opt for backward compatibility purpose to return
+        // G_M_N
+        const auto grid_desc_g_mraw_nraw =
+            transform_tensor_descriptor(grid_desc_gs_ms_ns,
+                                        make_tuple(make_merge_transform(gLengths),
+                                                   make_merge_transform(mLengths),
+                                                   make_merge_transform(nLengths)),
+                                        make_tuple(gDimIds, mDimIds, nDimIds),
+                                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+        const auto c_ms_ns_lengths = to_tuple(
+            gs_ms_ns_lengths_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+        const auto c_ms_ns_strides = to_tuple(
+            gs_ms_ns_strides_vec, Number<NumDimG>{}, Number<NumDimG + NumDimM + NumDimN>{});
+
+        // transformed tensor C[MRaw = M0 * M1 * M2 * ... , NRaw = N0 * N1 *
+        // N2 * ...]
+        const auto grid_desc_ms_ns = make_naive_tensor_descriptor(c_ms_ns_lengths, c_ms_ns_strides);
+
+        const auto grid_desc_mraw_nraw = transform_tensor_descriptor(
+            grid_desc_ms_ns,
+            make_tuple(make_merge_transform(mLengths), make_merge_transform(nLengths)),
+            make_tuple(mDimIds - Number<NumDimG>{}, nDimIds - Number<NumDimG>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return std::make_pair(grid_desc_g_mraw_nraw, grid_desc_mraw_nraw);
+    }
+}
+
+template <typename NumDims_G_M_N_K_O, // Sequence<>
+          typename PerBlock_M_N_K_O,  // Sequence<>
+          device::GemmSpecialization GemmSpec,
+          device::TensorSpecialization ASpec,
+          device::TensorSpecialization B0Spec,
+          device::TensorSpecialization B1Spec,
+          device::TensorSpecialization CSpec>
+struct TransformBatchedContractionContractionToBatchedGemmGemm
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+    static constexpr auto I4 = Number<4>{};
+
+    static constexpr index_t NumDimG = NumDims_G_M_N_K_O::At(I0);
+    static constexpr index_t NumDimM = NumDims_G_M_N_K_O::At(I1);
+    static constexpr index_t NumDimN = NumDims_G_M_N_K_O::At(I2);
+    static constexpr index_t NumDimK = NumDims_G_M_N_K_O::At(I3);
+    static constexpr index_t NumDimO = NumDims_G_M_N_K_O::At(I4);
+
+    static constexpr index_t MPerBlock = PerBlock_M_N_K_O::At(I0);
+    static constexpr index_t NPerBlock = PerBlock_M_N_K_O::At(I1);
+    static constexpr index_t KPerBlock = PerBlock_M_N_K_O::At(I2);
+    static constexpr index_t OPerBlock = PerBlock_M_N_K_O::At(I3);
+
+    static constexpr auto matrix_padder =
+        device::GemmGemmPadder<GemmSpec, index_t, index_t, index_t, index_t>{
+            MPerBlock, NPerBlock, KPerBlock, OPerBlock};
+
+    //
+    // A
+    //
+    static auto MakeAGridDescriptorPair(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        return MakeGridDescriptorPair<NumDimG, NumDimM, NumDimK, ASpec>(a_gs_ms_ks_lengths_vec,
+                                                                        a_gs_ms_ks_strides_vec);
+    }
+
+    // TODO: rename to G_MRaw_KRaw
+    static auto MakeAGridDescriptor_G_M_K(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                          const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        return MakeAGridDescriptorPair(a_gs_ms_ks_lengths_vec, a_gs_ms_ks_strides_vec).first;
+    }
+    static auto MakeAGridDescriptor_M_K(const std::vector<index_t>& a_gs_ms_ks_lengths_vec,
+                                        const std::vector<index_t>& a_gs_ms_ks_strides_vec)
+    {
+        return matrix_padder.PadADescriptor_M_K(
+            MakeAGridDescriptorPair(a_gs_ms_ks_lengths_vec, a_gs_ms_ks_strides_vec).second);
+    }
+
+    template <typename AGridDesc_M_K, typename Number>
+    __host__ __device__ static constexpr auto
+    MakeAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k, const Number& AK1)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    //
+    // B (alias of B0)
+    //
+    static auto MakeB0GridDescriptorPair(const std::vector<index_t>& b0_gs_ns_ks_lengths_vec,
+                                         const std::vector<index_t>& b0_gs_ns_ks_strides_vec)
+    {
+        return MakeGridDescriptorPair<NumDimG, NumDimN, NumDimK, B0Spec>(b0_gs_ns_ks_lengths_vec,
+                                                                         b0_gs_ns_ks_strides_vec);
+    }
+
+    // TODO: rename to G_MRaw_NRaw
+    static auto MakeB0GridDescriptor_G_N_K(const std::vector<index_t>& b0_gs_ns_ks_lengths_vec,
+                                           const std::vector<index_t>& b0_gs_ns_ks_strides_vec)
+    {
+        return MakeB0GridDescriptorPair(b0_gs_ns_ks_lengths_vec, b0_gs_ns_ks_strides_vec).first;
+    }
+    static auto MakeB0GridDescriptor_N_K(const std::vector<index_t>& b0_gs_ns_ks_lengths_vec,
+                                         const std::vector<index_t>& b0_gs_ns_ks_strides_vec)
+    {
+        // alias of matrix_padder.PadB0Descriptor_N_K
+        return matrix_padder.PadBDescriptor_N_K(
+            MakeB0GridDescriptorPair(b0_gs_ns_ks_lengths_vec, b0_gs_ns_ks_strides_vec).second);
+    }
+
+    template <typename BGridDesc_N_K, typename Number>
+    __host__ __device__ static constexpr auto
+    MakeB0GridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k, const Number& BK1)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    //
+    // B1
+    //
+    static auto MakeB1GridDescriptorPair(const std::vector<index_t>& b1_gs_os_ns_lengths_vec,
+                                         const std::vector<index_t>& b1_gs_os_ns_strides_vec)
+    {
+        return MakeGridDescriptorPair<NumDimG, NumDimO, NumDimN, B1Spec>(b1_gs_os_ns_lengths_vec,
+                                                                         b1_gs_os_ns_strides_vec);
+    }
+
+    // TODO: rename to G_NRaw_KRaw
+    static auto MakeB1GridDescriptor_G_N_K(const std::vector<index_t>& b1_gs_os_ns_lengths_vec,
+                                           const std::vector<index_t>& b1_gs_os_ns_strides_vec)
+    {
+        return MakeB1GridDescriptorPair(b1_gs_os_ns_lengths_vec, b1_gs_os_ns_strides_vec).first;
+    }
+    static auto MakeB1GridDescriptor_N_K(const std::vector<index_t>& b1_gs_os_ns_lengths_vec,
+                                         const std::vector<index_t>& b1_gs_os_ns_strides_vec)
+    {
+        // alias of matrix_padder.PadB1Descriptor_O_N
+        return matrix_padder.PadB1Descriptor_N_K(
+            MakeB1GridDescriptorPair(b1_gs_os_ns_lengths_vec, b1_gs_os_ns_strides_vec).second);
+    }
+
+    template <typename B1GridDesc_N_K, typename Number>
+    __host__ __device__ static constexpr auto
+    MakeB1GridDescriptor_BK0_N_BK1(const B1GridDesc_N_K& b1_grid_desc_n_k, const Number& B1K1)
+    {
+        const auto N = b1_grid_desc_n_k.GetLength(I0);
+        const auto K = b1_grid_desc_n_k.GetLength(I1);
+
+        const auto B1K0 = K / B1K1;
+
+        return transform_tensor_descriptor(
+            b1_grid_desc_n_k,
+            make_tuple(make_unmerge_transform(make_tuple(B1K0, B1K1)),
+                       make_pass_through_transform(N)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    //
+    // C
+    //
+    static auto MakeCGridDescriptorPair(const std::vector<index_t>& c_gs_ms_os_lengths_vec,
+                                        const std::vector<index_t>& c_gs_ms_os_strides_vec)
+    {
+        return MakeGridDescriptorPair<NumDimG, NumDimM, NumDimO, CSpec>(c_gs_ms_os_lengths_vec,
+                                                                        c_gs_ms_os_strides_vec);
+    }
+
+    // TODO: rename to G_MRaw_NRaw
+    static auto MakeCGridDescriptor_G_M_N(const std::vector<index_t>& c_gs_ms_os_lengths_vec,
+                                          const std::vector<index_t>& c_gs_ms_os_strides_vec)
+    {
+        return MakeCGridDescriptorPair(c_gs_ms_os_lengths_vec, c_gs_ms_os_strides_vec).first;
+    }
+    static auto MakeCGridDescriptor_M_N(const std::vector<index_t>& c_gs_ms_os_lengths_vec,
+                                        const std::vector<index_t>& c_gs_ms_os_strides_vec)
+    {
+        return matrix_padder.PadCDescriptor_M_N(
+            MakeCGridDescriptorPair(c_gs_ms_os_lengths_vec, c_gs_ms_os_strides_vec).second);
+    }
+};
+
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp b/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp
new file mode 100644
index 0000000000..2be0b66812
--- /dev/null
+++ b/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp
@@ -0,0 +1,1137 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+
+namespace ck {
+namespace tensor_operation {
+
+namespace {
+template <
+    index_t NDimSpatial,
+    typename ALayout,
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization ConvBwdDataSpecialization>
+constexpr auto make_out_grid_desc(const index_t N,
+                                  const index_t Do,
+                                  const index_t Ho,
+                                  const index_t Wo,
+                                  const index_t K,
+                                  const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_strides)
+{
+    const auto KStride = Number<1>{};
+
+    if constexpr(is_same_v<ALayout, tensor_layout::convolution::NHWGK>)
+    {
+        const index_t NStride  = out_g_n_k_wos_strides[1];
+        const index_t HiStride = out_g_n_k_wos_strides[3];
+        const index_t WiStride = out_g_n_k_wos_strides[4];
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+
+            return make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, K),
+                                                make_tuple(WiStride, KStride));
+        }
+        else
+        {
+            return make_naive_tensor_descriptor(make_tuple(N, Ho, Wo, K),
+                                                make_tuple(NStride, HiStride, WiStride, KStride));
+        }
+    }
+    else if constexpr(is_same_v<ALayout, tensor_layout::convolution::NDHWGK>)
+    {
+        const index_t NStride  = out_g_n_k_wos_strides[1];
+        const index_t DoStride = out_g_n_k_wos_strides[3];
+        const index_t HoStride = out_g_n_k_wos_strides[4];
+        const index_t WoStride = out_g_n_k_wos_strides[5];
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+
+            return make_naive_tensor_descriptor(make_tuple(N * Do * Ho * Wo, K),
+                                                make_tuple(WoStride, KStride));
+        }
+        else
+        {
+            return make_naive_tensor_descriptor(
+                make_tuple(N, Do, Ho, Wo, K),
+                make_tuple(NStride, DoStride, HoStride, WoStride, KStride));
+        }
+    }
+    else if constexpr(is_same_v<ALayout, tensor_layout::convolution::GNHWK>)
+    {
+        // assume packed
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+        }
+        else
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(N, Ho, Wo, K));
+        }
+    }
+    else if constexpr(is_same_v<ALayout, tensor_layout::convolution::GNDHWK>)
+    {
+        // assume packed
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K));
+        }
+        else
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(N, Do, Ho, Wo, K));
+        }
+    }
+    else
+    {
+        throw std::runtime_error("wrong! unsupported layout: " + ALayout::name());
+    }
+}
+
+template <typename BLayout>
+constexpr auto make_wei_grid_desc(
+    const index_t K, const index_t Z, const index_t Y, const index_t X, const index_t C)
+{
+
+    if constexpr(is_same_v<BLayout, tensor_layout::convolution::GKYXC>)
+    {
+        return make_naive_tensor_descriptor_packed(make_tuple(K, Y, X, C));
+    }
+    else if constexpr(is_same_v<BLayout, tensor_layout::convolution::GKZYXC>)
+    {
+        return make_naive_tensor_descriptor_packed(make_tuple(K, Z, Y, X, C));
+    }
+    else
+    {
+        throw std::runtime_error("wrong! unsupported layout: " + BLayout::name());
+    }
+}
+
+template <index_t NDimSpatial, typename CLayout>
+constexpr auto make_in_grid_desc(const index_t N,
+                                 const index_t Di,
+                                 const index_t Hi,
+                                 const index_t Wi,
+                                 const index_t C,
+                                 const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_strides)
+{
+
+    if constexpr(is_same_v<CLayout, tensor_layout::convolution::GNHWC> ||
+                 is_same_v<CLayout, tensor_layout::convolution::NHWGC> ||
+                 is_same_v<CLayout, tensor_layout::convolution::G_NHW_C>)
+    {
+        return make_naive_tensor_descriptor(make_tuple(N, Hi, Wi, C),
+                                            make_tuple(in_g_n_c_wis_strides[1],
+                                                       in_g_n_c_wis_strides[3],
+                                                       in_g_n_c_wis_strides[4],
+                                                       in_g_n_c_wis_strides[2]));
+    }
+    else if constexpr(is_same_v<CLayout, tensor_layout::convolution::GNDHWC> ||
+                      is_same_v<CLayout, tensor_layout::convolution::NDHWGC>)
+    {
+        return make_naive_tensor_descriptor(make_tuple(N, Di, Hi, Wi, C),
+                                            make_tuple(in_g_n_c_wis_strides[1],
+                                                       in_g_n_c_wis_strides[3],
+                                                       in_g_n_c_wis_strides[4],
+                                                       in_g_n_c_wis_strides[5],
+                                                       in_g_n_c_wis_strides[2]));
+    }
+    else
+    {
+        throw std::runtime_error("wrong! unsupported layout: " + CLayout::name());
+    }
+}
+
+} // namespace
+
+template <
+    index_t NDimSpatial,
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization ConvBwdDataSpecialization,
+    index_t AK1,
+    index_t BK1,
+    index_t GemmMPerBlock,
+    index_t GemmNPerBlock,
+    index_t GemmKPerBlock,
+    bool DoPadGemmM,
+    bool DoPadGemmN>
+struct TransformConvBwdDataToGemm_v1
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr auto NonSpatialDimsNum = Number<3>{};
+
+    static constexpr auto DIdx = Number<NonSpatialDimsNum>{};
+    static constexpr auto HIdx =
+        NDimSpatial == 2 ? Number<NonSpatialDimsNum>{} : Number<NonSpatialDimsNum + 1>{};
+    static constexpr auto WIdx =
+        NDimSpatial == 2 ? Number<NonSpatialDimsNum + 1>{} : Number<NonSpatialDimsNum + 2>{};
+
+    static constexpr auto ZIdx = Number<NonSpatialDimsNum>{};
+    static constexpr auto YIdx =
+        NDimSpatial == 2 ? Number<NonSpatialDimsNum>{} : Number<NonSpatialDimsNum + 1>{};
+    static constexpr auto XIdx =
+        NDimSpatial == 2 ? Number<NonSpatialDimsNum + 1>{} : Number<NonSpatialDimsNum + 2>{};
+
+    template <typename ALayout,
+              typename std::enable_if<(NDimSpatial == 2 || NDimSpatial == 3) &&
+                                          (is_same_v<ALayout, tensor_layout::convolution::GNHWK> ||
+                                           is_same_v<ALayout, tensor_layout::convolution::GNDHWK> ||
+                                           is_same_v<ALayout, tensor_layout::convolution::NHWGK> ||
+                                           is_same_v<ALayout, tensor_layout::convolution::NDHWGK>),
+                                      bool>::type = false>
+    static auto MakeADescriptor_AK0_M_AK1(
+        const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& wei_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& /* wei_g_k_c_xs_strides */,
+        const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& /* in_g_n_c_wis_strides */,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& /* input_right_pads */,
+        const std::array<index_t, NDimSpatial>& tildes)
+    {
+        index_t i_ztilde = tildes[ZIdx - NonSpatialDimsNum];
+        index_t i_ytilde = tildes[YIdx - NonSpatialDimsNum];
+        index_t i_xtilde = tildes[XIdx - NonSpatialDimsNum];
+
+        const index_t N = in_g_n_c_wis_lengths[1];
+        const index_t K = wei_g_k_c_xs_lengths[1];
+
+        const index_t Di = NDimSpatial == 3 ? in_g_n_c_wis_lengths[DIdx] : 1;
+        const index_t Hi = in_g_n_c_wis_lengths[HIdx];
+        const index_t Wi = in_g_n_c_wis_lengths[WIdx];
+
+        const index_t Do = NDimSpatial == 3 ? out_g_n_k_wos_lengths[DIdx] : 1;
+        const index_t Ho = out_g_n_k_wos_lengths[HIdx];
+        const index_t Wo = out_g_n_k_wos_lengths[WIdx];
+
+        const index_t Z = NDimSpatial == 3 ? wei_g_k_c_xs_lengths[ZIdx] : 1;
+        const index_t Y = wei_g_k_c_xs_lengths[YIdx];
+        const index_t X = wei_g_k_c_xs_lengths[XIdx];
+
+        const index_t InLeftPadD = input_left_pads[DIdx - NonSpatialDimsNum];
+        const index_t InLeftPadH = input_left_pads[HIdx - NonSpatialDimsNum];
+        const index_t InLeftPadW = input_left_pads[WIdx - NonSpatialDimsNum];
+
+        const index_t ConvStrideD = conv_filter_strides[DIdx - NonSpatialDimsNum];
+        const index_t ConvStrideH = conv_filter_strides[HIdx - NonSpatialDimsNum];
+        const index_t ConvStrideW = conv_filter_strides[WIdx - NonSpatialDimsNum];
+
+        const index_t ConvDilationD = conv_filter_dilations[DIdx - NonSpatialDimsNum];
+        const index_t ConvDilationH = conv_filter_dilations[HIdx - NonSpatialDimsNum];
+        const index_t ConvDilationW = conv_filter_dilations[WIdx - NonSpatialDimsNum];
+
+        // n_do_ho_wo_k for 3d or n_ho_wo_k for 2d
+        const auto out_grid_desc =
+            make_out_grid_desc<NDimSpatial, ALayout, ConvBwdDataSpecialization>(
+                N, Do, Ho, Wo, K, out_g_n_k_wos_strides);
+
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+            const index_t AK0 = math::integer_divide_ceil(K, AK1);
+
+            // A: output tensor
+            const auto out_gemmak0_gemmmraw_gemmak1_grid_desc = transform_tensor_descriptor(
+                out_grid_desc,
+                make_tuple(make_pass_through_transform(N * Do * Ho * Wo),
+                           make_unmerge_transform(make_tuple(AK0, AK1))),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+            const auto out_gemmak0_gemmm_gemmak1_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    out_gemmak0_gemmmraw_gemmak1_grid_desc,
+                    make_tuple(AK0, GemmMPerBlock, AK1),
+                    Sequence<false, DoPadGemmM, false>{});
+
+            return out_gemmak0_gemmm_gemmak1_grid_desc;
+        }
+        else
+        {
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto ZDot = math::integer_divide_ceil(Z, ZTilde);
+            const auto YDot = math::integer_divide_ceil(Y, YTilde);
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            const auto DTilde =
+                Do + math::integer_divide_ceil(ConvDilationD * (Z - I1), ConvStrideD);
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IDTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadD - ConvDilationD * (ZTilde - I1)), ConvStrideD);
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IDTildeSliceEnd = math::min(
+                DTilde, math::integer_divide_ceil(InLeftPadD + Di - I1, ConvStrideD) + I1);
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto DTildeSlice = IDTildeSliceEnd - IDTildeSliceBegin;
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
+            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            if constexpr(NDimSpatial == 2)
+            {
+                // A: output tensor
+                const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                    out_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_pad_transform(Ho, I0, I0),
+                               make_pad_transform(Wo, I0, I0),
+                               make_pass_through_transform(K)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+                const auto out_n_ydot_htilde_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
+                    out_n_hop_wop_k_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(N),
+                        make_embed_transform(make_tuple(YDot, HTilde),
+                                             make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                        make_embed_transform(make_tuple(XDot, WTilde),
+                                             make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                        make_pass_through_transform(K)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+                const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k_grid_desc =
+                    transform_tensor_descriptor(
+                        out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_slice_transform(YDot, I0, YDotSlice),
+                                   make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                                   make_slice_transform(XDot, I0, XDotSlice),
+                                   make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                                   make_pass_through_transform(K)),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<5>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<5>{}));
+
+                const auto out_gemmk_gemmmraw_grid_desc = transform_tensor_descriptor(
+                    out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k_grid_desc,
+                    make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K)),
+                               make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice))),
+                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+                const auto out_gemmk_gemmm_padded_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        out_gemmk_gemmmraw_grid_desc,
+                        make_tuple(GemmKPerBlock, GemmMPerBlock),
+                        Sequence<true, DoPadGemmM>{});
+
+                const index_t AK0 = out_gemmk_gemmm_padded_grid_desc.GetLength(I0) / AK1;
+
+                const auto out_gemmak0_gemmm_gemmak1_grid_desc = transform_tensor_descriptor(
+                    out_gemmk_gemmm_padded_grid_desc,
+                    make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                               make_pass_through_transform(
+                                   out_gemmk_gemmm_padded_grid_desc.GetLength(I1))),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+                return out_gemmak0_gemmm_gemmak1_grid_desc;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+                // A: output tensor
+                const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                    out_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_pad_transform(Do, I0, I0),
+                               make_pad_transform(Ho, I0, I0),
+                               make_pad_transform(Wo, I0, I0),
+                               make_pass_through_transform(K)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+                const auto out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_k_grid_desc =
+                    transform_tensor_descriptor(
+                        out_n_hop_wop_k_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_embed_transform(
+                                       make_tuple(ZDot, DTilde),
+                                       make_tuple(-ConvDilationD / GcdStrideDilationD, I1)),
+                                   make_embed_transform(
+                                       make_tuple(YDot, HTilde),
+                                       make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                                   make_embed_transform(
+                                       make_tuple(XDot, WTilde),
+                                       make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                                   make_pass_through_transform(K)),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1, 2>{},
+                                   Sequence<3, 4>{},
+                                   Sequence<5, 6>{},
+                                   Sequence<7>{}));
+
+                const auto
+                    out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_k_grid_desc =
+                        transform_tensor_descriptor(
+                            out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_k_grid_desc,
+                            make_tuple(make_pass_through_transform(N),
+                                       make_slice_transform(ZDot, I0, ZDotSlice),
+                                       make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                                       make_slice_transform(YDot, I0, YDotSlice),
+                                       make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                                       make_slice_transform(XDot, I0, XDotSlice),
+                                       make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                                       make_pass_through_transform(K)),
+                            make_tuple(Sequence<0>{},
+                                       Sequence<1>{},
+                                       Sequence<2>{},
+                                       Sequence<3>{},
+                                       Sequence<4>{},
+                                       Sequence<5>{},
+                                       Sequence<6>{},
+                                       Sequence<7>{}),
+                            make_tuple(Sequence<0>{},
+                                       Sequence<1>{},
+                                       Sequence<2>{},
+                                       Sequence<3>{},
+                                       Sequence<4>{},
+                                       Sequence<5>{},
+                                       Sequence<6>{},
+                                       Sequence<7>{}));
+
+                const auto out_gemmk_gemmmraw_grid_desc = transform_tensor_descriptor(
+                    out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_k_grid_desc,
+                    make_tuple(
+                        make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, K)),
+                        make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice))),
+                    make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+                const auto out_gemmk_gemmm_padded_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        out_gemmk_gemmmraw_grid_desc,
+                        make_tuple(GemmKPerBlock, GemmMPerBlock),
+                        Sequence<true, DoPadGemmM>{});
+
+                const index_t AK0 = out_gemmk_gemmm_padded_grid_desc.GetLength(I0) / AK1;
+
+                const auto out_gemmak0_gemmm_gemmak1_grid_desc = transform_tensor_descriptor(
+                    out_gemmk_gemmm_padded_grid_desc,
+                    make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                               make_pass_through_transform(
+                                   out_gemmk_gemmm_padded_grid_desc.GetLength(I1))),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+                return out_gemmak0_gemmm_gemmak1_grid_desc;
+            }
+            else
+            {
+                throw std::runtime_error("wrong! only implemented for 2D and 3D now");
+            }
+        }
+    }
+
+    template <typename BLayout,
+              typename std::enable_if<(NDimSpatial == 2 || NDimSpatial == 3) &&
+                                          (is_same_v<BLayout, tensor_layout::convolution::GKYXC> ||
+                                           is_same_v<BLayout, tensor_layout::convolution::GKZYXC>),
+                                      bool>::type = false>
+    static auto MakeBDescriptor_BK0_N_BK1(
+        const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& /* out_g_n_k_wos_strides */,
+        const std::array<index_t, NDimSpatial + 3>& wei_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& /* wei_g_k_c_xs_strides */,
+        const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& /* in_g_n_c_wis_strides */,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& /* input_left_pads */,
+        const std::array<index_t, NDimSpatial>& /* input_right_pads */,
+        const std::array<index_t, NDimSpatial>& tildes)
+    {
+        index_t i_ztilde = tildes[ZIdx - NonSpatialDimsNum];
+        index_t i_ytilde = tildes[YIdx - NonSpatialDimsNum];
+        index_t i_xtilde = tildes[XIdx - NonSpatialDimsNum];
+
+        const index_t N = in_g_n_c_wis_lengths[1];
+        const index_t K = wei_g_k_c_xs_lengths[1];
+        const index_t C = wei_g_k_c_xs_lengths[2];
+
+        const index_t Do = NDimSpatial == 3 ? out_g_n_k_wos_lengths[DIdx] : 1;
+        const index_t Ho = out_g_n_k_wos_lengths[HIdx];
+        const index_t Wo = out_g_n_k_wos_lengths[WIdx];
+
+        const index_t Z = NDimSpatial == 3 ? wei_g_k_c_xs_lengths[ZIdx] : 1;
+        const index_t Y = wei_g_k_c_xs_lengths[YIdx];
+        const index_t X = wei_g_k_c_xs_lengths[XIdx];
+
+        const index_t ConvStrideD = conv_filter_strides[DIdx - NonSpatialDimsNum];
+        const index_t ConvStrideH = conv_filter_strides[HIdx - NonSpatialDimsNum];
+        const index_t ConvStrideW = conv_filter_strides[WIdx - NonSpatialDimsNum];
+
+        const index_t ConvDilationD = conv_filter_dilations[DIdx - NonSpatialDimsNum];
+        const index_t ConvDilationH = conv_filter_dilations[HIdx - NonSpatialDimsNum];
+        const index_t ConvDilationW = conv_filter_dilations[WIdx - NonSpatialDimsNum];
+
+        // assume packed
+        // k_y_x_c for 2d or k_z_y_x_c for 3d
+        const auto wei_grid_desc = make_wei_grid_desc<BLayout>(K, Z, Y, X, C);
+
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+            const index_t BK0 = math::integer_divide_ceil(K, BK1);
+
+            // B: weight tensor
+            const auto wei_gemmbk0_gemmnraw_gemmbk1_grid_desc =
+                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+            make_naive_tensor_descriptor(make_tuple(N * Do * Ho * Wo, C), make_tuple(I0, I1));
+
+            const auto wei_gemmbk0_gemmn_gemmbk1_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    wei_gemmbk0_gemmnraw_gemmbk1_grid_desc,
+                    make_tuple(BK0, GemmNPerBlock, BK1),
+                    Sequence<false, DoPadGemmN, false>{});
+
+            return wei_gemmbk0_gemmn_gemmbk1_grid_desc;
+        }
+        else
+        {
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto ZDot = math::integer_divide_ceil(Z, ZTilde);
+            const auto YDot = math::integer_divide_ceil(Y, YTilde);
+            const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+            // GemmK is different for each GEMM
+            const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
+            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+            // B weight tensor
+            if constexpr(NDimSpatial == 2)
+            {
+                const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
+                    wei_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(K),
+                        make_embed_transform(make_tuple(YDot, YTilde),
+                                             make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                        make_embed_transform(make_tuple(XDot, XTilde),
+                                             make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                        make_pass_through_transform(C)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+                const auto wei_k_ydotslice_xdotslice_c_grid_desc = transform_tensor_descriptor(
+                    wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                    make_tuple(make_pass_through_transform(K),
+                               make_slice_transform(YDot, I0, YDotSlice),
+                               make_slice_transform(XDot, I0, XDotSlice),
+                               make_freeze_transform(i_ytilde),
+                               make_freeze_transform(i_xtilde),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<3>{},
+                               Sequence<2>{},
+                               Sequence<4>{},
+                               Sequence<5>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<>{},
+                               Sequence<>{},
+                               Sequence<3>{}));
+
+                const auto wei_gemmk_gemmnraw_grid_desc = transform_tensor_descriptor(
+                    wei_k_ydotslice_xdotslice_c_grid_desc,
+                    make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K)),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<1, 2, 0>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+                const auto wei_gemmk_gemmn_padded_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        wei_gemmk_gemmnraw_grid_desc,
+                        make_tuple(GemmKPerBlock, GemmNPerBlock),
+                        Sequence<true, DoPadGemmN>{});
+
+                const index_t BK0 = wei_gemmk_gemmn_padded_grid_desc.GetLength(I0) / BK1;
+
+                const auto wei_gemmbk0_gemmn_gemmbk1_grid_desc = transform_tensor_descriptor(
+                    wei_gemmk_gemmn_padded_grid_desc,
+                    make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                               make_pass_through_transform(
+                                   wei_gemmk_gemmn_padded_grid_desc.GetLength(I1))),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+                return wei_gemmbk0_gemmn_gemmbk1_grid_desc;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+                const auto wei_k_zdot_ztilde_ydot_ytilde_xdot_xtilde_c_grid_desc =
+                    transform_tensor_descriptor(
+                        wei_grid_desc,
+                        make_tuple(
+                            make_pass_through_transform(K),
+                            make_embed_transform(make_tuple(ZDot, ZTilde),
+                                                 make_tuple(ConvStrideD / GcdStrideDilationD, I1)),
+                            make_embed_transform(make_tuple(YDot, YTilde),
+                                                 make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                            make_embed_transform(make_tuple(XDot, XTilde),
+                                                 make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                            make_pass_through_transform(C)),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1, 2>{},
+                                   Sequence<3, 4>{},
+                                   Sequence<5, 6>{},
+                                   Sequence<7>{}));
+
+                const auto wei_gemmk_zdotslice_ydotslice_xdotslice_c_grid_desc =
+                    transform_tensor_descriptor(
+                        wei_k_zdot_ztilde_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                        make_tuple(make_pass_through_transform(K),
+                                   make_slice_transform(ZDot, I0, ZDotSlice),
+                                   make_slice_transform(YDot, I0, YDotSlice),
+                                   make_slice_transform(XDot, I0, XDotSlice),
+                                   make_freeze_transform(i_ztilde),
+                                   make_freeze_transform(i_ytilde),
+                                   make_freeze_transform(i_xtilde),
+                                   make_pass_through_transform(C)),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<3>{},
+                                   Sequence<5>{},
+                                   Sequence<2>{},
+                                   Sequence<4>{},
+                                   Sequence<6>{},
+                                   Sequence<7>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<>{},
+                                   Sequence<>{},
+                                   Sequence<>{},
+                                   Sequence<4>{}));
+
+                const auto wei_gemmk_gemmnraw_grid_desc = transform_tensor_descriptor(
+                    wei_gemmk_zdotslice_ydotslice_xdotslice_c_grid_desc,
+                    make_tuple(make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, K)),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<1, 2, 3, 0>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+                const auto wei_gemmk_gemmn_padded_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        wei_gemmk_gemmnraw_grid_desc,
+                        make_tuple(GemmKPerBlock, GemmNPerBlock),
+                        Sequence<true, DoPadGemmN>{});
+
+                const index_t BK0 = wei_gemmk_gemmn_padded_grid_desc.GetLength(I0) / BK1;
+
+                const auto wei_gemmbk0_gemm_gemmbk1_grid_desc = transform_tensor_descriptor(
+                    wei_gemmk_gemmn_padded_grid_desc,
+                    make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                               make_pass_through_transform(
+                                   wei_gemmk_gemmn_padded_grid_desc.GetLength(I1))),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+                return wei_gemmbk0_gemm_gemmbk1_grid_desc;
+            }
+            else
+            {
+                throw std::runtime_error("wrong! only implemented for 2D and 3D now");
+            }
+        }
+    }
+
+    template <typename CLayout,
+              typename std::enable_if<(NDimSpatial == 2 || NDimSpatial == 3) &&
+                                          (is_same_v<CLayout, tensor_layout::convolution::GNHWC> ||
+                                           is_same_v<CLayout, tensor_layout::convolution::GNDHWC> ||
+                                           is_same_v<CLayout, tensor_layout::convolution::NHWGC> ||
+                                           is_same_v<CLayout, tensor_layout::convolution::NDHWGC> ||
+                                           is_same_v<CLayout, tensor_layout::convolution::G_NHW_C>),
+                                      bool>::type = false>
+    static auto
+    MakeCDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* out_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& wei_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* wei_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads,
+                        const std::array<index_t, NDimSpatial>& tildes)
+    {
+        index_t i_ztilde = tildes[ZIdx - NonSpatialDimsNum];
+        index_t i_ytilde = tildes[YIdx - NonSpatialDimsNum];
+        index_t i_xtilde = tildes[XIdx - NonSpatialDimsNum];
+
+        const index_t N = in_g_n_c_wis_lengths[1];
+        const index_t C = wei_g_k_c_xs_lengths[2];
+
+        const index_t Di = NDimSpatial == 3 ? in_g_n_c_wis_lengths[DIdx] : 1;
+        const index_t Hi = in_g_n_c_wis_lengths[HIdx];
+        const index_t Wi = in_g_n_c_wis_lengths[WIdx];
+
+        const index_t Do = NDimSpatial == 3 ? out_g_n_k_wos_lengths[DIdx] : 1;
+        const index_t Ho = out_g_n_k_wos_lengths[HIdx];
+        const index_t Wo = out_g_n_k_wos_lengths[WIdx];
+
+        const index_t Z = NDimSpatial == 3 ? wei_g_k_c_xs_lengths[ZIdx] : 1;
+        const index_t Y = wei_g_k_c_xs_lengths[YIdx];
+        const index_t X = wei_g_k_c_xs_lengths[XIdx];
+
+        const index_t InLeftPadD = input_left_pads[DIdx - NonSpatialDimsNum];
+        const index_t InLeftPadH = input_left_pads[HIdx - NonSpatialDimsNum];
+        const index_t InLeftPadW = input_left_pads[WIdx - NonSpatialDimsNum];
+
+        const index_t InRightPadD = input_right_pads[DIdx - NonSpatialDimsNum];
+        const index_t InRightPadH = input_right_pads[HIdx - NonSpatialDimsNum];
+        const index_t InRightPadW = input_right_pads[WIdx - NonSpatialDimsNum];
+
+        const index_t ConvStrideD = conv_filter_strides[DIdx - NonSpatialDimsNum];
+        const index_t ConvStrideH = conv_filter_strides[HIdx - NonSpatialDimsNum];
+        const index_t ConvStrideW = conv_filter_strides[WIdx - NonSpatialDimsNum];
+
+        const index_t ConvDilationD = conv_filter_dilations[DIdx - NonSpatialDimsNum];
+        const index_t ConvDilationH = conv_filter_dilations[HIdx - NonSpatialDimsNum];
+        const index_t ConvDilationW = conv_filter_dilations[WIdx - NonSpatialDimsNum];
+
+        // assume strided
+        // n_hi_wi_c for 2d n_di_hi_wi_c for 3d
+        const auto in_grid_desc =
+            make_in_grid_desc<NDimSpatial, CLayout>(N, Di, Hi, Wi, C, in_g_n_c_wis_strides);
+
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+            // C: input tensor
+            if constexpr(NDimSpatial == 2)
+            {
+                const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                    in_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(N),
+                        make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                        make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                        make_pass_through_transform(C)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+                const auto in_gemmmraw_gemmnraw_grid_desc = transform_tensor_descriptor(
+                    in_n_y_ho_x_wo_c_grid_desc,
+                    make_tuple(make_freeze_transform(I0),
+                               make_freeze_transform(I0),
+                               make_merge_transform(make_tuple(N, Ho, Wo)),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<1>{}, Sequence<3>{}, Sequence<0, 2, 4>{}, Sequence<5>{}),
+                    make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+                const auto in_gemmm_gemmn_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        in_gemmmraw_gemmnraw_grid_desc,
+                        make_tuple(GemmMPerBlock, GemmNPerBlock),
+                        Sequence<DoPadGemmM, DoPadGemmN>{});
+
+                return in_gemmm_gemmn_grid_desc;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+
+                // C: input tensor
+                const auto in_n_x_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                    in_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(N),
+                        make_embed_transform(make_tuple(I1, Do), make_tuple(I1, ConvStrideD)),
+                        make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                        make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
+                        make_pass_through_transform(C)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1, 2>{},
+                               Sequence<3, 4>{},
+                               Sequence<5, 6>{},
+                               Sequence<7>{}));
+
+                const auto in_gemmmraw_gemmnraw_grid_desc = transform_tensor_descriptor(
+                    in_n_x_do_y_ho_x_wo_c_grid_desc,
+                    make_tuple(make_freeze_transform(I0),
+                               make_freeze_transform(I0),
+                               make_freeze_transform(I0),
+                               make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<1>{},
+                               Sequence<3>{},
+                               Sequence<5>{},
+                               Sequence<0, 2, 4, 6>{},
+                               Sequence<7>{}),
+                    make_tuple(
+                        Sequence<>{}, Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+                const auto in_gemmm_gemmn_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        in_gemmmraw_gemmnraw_grid_desc,
+                        make_tuple(GemmMPerBlock, GemmNPerBlock),
+                        Sequence<DoPadGemmM, DoPadGemmN>{});
+
+                return in_gemmm_gemmn_grid_desc;
+            }
+            else
+            {
+                throw std::runtime_error("wrong! only implemented for 2D and 3D now");
+            }
+        }
+        else
+        {
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto DTilde =
+                Do + math::integer_divide_ceil(ConvDilationD * (Z - I1), ConvStrideD);
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on DTilde, HTilde and WTilde that contribute to
+            // non-padding area of input tensor
+            const auto IDTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadD - ConvDilationD * (ZTilde - I1)), ConvStrideD);
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IDTildeSliceEnd = math::min(
+                DTilde, math::integer_divide_ceil(InLeftPadD + Di - I1, ConvStrideD) + I1);
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto DTildeSlice = IDTildeSliceEnd - IDTildeSliceBegin;
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // C: input tensor
+            if constexpr(NDimSpatial == 2)
+            {
+                const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                    in_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                               make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+                const auto in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc =
+                    transform_tensor_descriptor(
+                        in_n_hip_wip_c_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_embed_transform(make_tuple(YTilde, HTilde),
+                                                        make_tuple(ConvDilationH, ConvStrideH)),
+                                   make_embed_transform(make_tuple(XTilde, WTilde),
+                                                        make_tuple(ConvDilationW, ConvStrideW)),
+                                   make_pass_through_transform(C)),
+                        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                        make_tuple(
+                            Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+                const auto in_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
+                    in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_freeze_transform(i_ytilde),
+                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                               make_freeze_transform(i_xtilde),
+                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<>{},
+                               Sequence<1>{},
+                               Sequence<>{},
+                               Sequence<2>{},
+                               Sequence<3>{}));
+
+                const auto in_gemmmraw_gemmnraw_grid_desc = transform_tensor_descriptor(
+                    in_n_htildeslice_wtildeslice_c_grid_desc,
+                    make_tuple(make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+                const auto in_gemmm_gemmn_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        in_gemmmraw_gemmnraw_grid_desc,
+                        make_tuple(GemmMPerBlock, GemmNPerBlock),
+                        Sequence<DoPadGemmM, DoPadGemmN>{});
+
+                return in_gemmm_gemmn_grid_desc;
+            }
+            else if(NDimSpatial == 3)
+            {
+                const auto in_n_dip_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                    in_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_pad_transform(Di, InLeftPadD, InRightPadD),
+                               make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                               make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                               make_pass_through_transform(C)),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                    make_tuple(
+                        Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+                const auto in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc =
+                    transform_tensor_descriptor(
+                        in_n_dip_hip_wip_c_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_embed_transform(make_tuple(ZTilde, DTilde),
+                                                        make_tuple(ConvDilationD, ConvStrideD)),
+                                   make_embed_transform(make_tuple(YTilde, HTilde),
+                                                        make_tuple(ConvDilationH, ConvStrideH)),
+                                   make_embed_transform(make_tuple(XTilde, WTilde),
+                                                        make_tuple(ConvDilationW, ConvStrideW)),
+                                   make_pass_through_transform(C)),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1, 2>{},
+                                   Sequence<3, 4>{},
+                                   Sequence<5, 6>{},
+                                   Sequence<7>{}));
+
+                const auto in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc =
+                    transform_tensor_descriptor(
+                        in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_freeze_transform(i_ztilde),
+                                   make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                                   make_freeze_transform(i_ytilde),
+                                   make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                                   make_freeze_transform(i_xtilde),
+                                   make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                                   make_pass_through_transform(C)),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<5>{},
+                                   Sequence<6>{},
+                                   Sequence<7>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<>{},
+                                   Sequence<1>{},
+                                   Sequence<>{},
+                                   Sequence<2>{},
+                                   Sequence<>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{}));
+
+                const auto in_gemmmraw_gemmnraw_grid_desc = transform_tensor_descriptor(
+                    in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc,
+                    make_tuple(
+                        make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice)),
+                        make_pass_through_transform(C)),
+                    make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+                const auto in_gemmm_gemmn_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        in_gemmmraw_gemmnraw_grid_desc,
+                        make_tuple(GemmMPerBlock, GemmNPerBlock),
+                        Sequence<DoPadGemmM, DoPadGemmN>{});
+                return in_gemmm_gemmn_grid_desc;
+            }
+            else
+            {
+                throw std::runtime_error("wrong! only implemented for 2D and 3D now");
+            }
+        }
+    }
+
+    // for input bias
+    template <typename CLayout,
+              typename std::enable_if<NDimSpatial == 2 &&
+                                          (is_same_v<CLayout, tensor_layout::convolution::GC> ||
+                                           is_same_v<CLayout, tensor_layout::convolution::G_C>),
+                                      bool>::type = false>
+    static auto
+    MakeCDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* out_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& wei_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* wei_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* in_g_n_c_wis_strides */,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& /* input_right_pads */,
+                        const std::array<index_t, NDimSpatial>& /* tildes */)
+    {
+        const index_t N = in_g_n_c_wis_lengths[1];
+        const index_t C = wei_g_k_c_xs_lengths[2];
+
+        const index_t Hi = in_g_n_c_wis_lengths[3];
+        const index_t Wi = in_g_n_c_wis_lengths[4];
+
+        const index_t Ho = out_g_n_k_wos_lengths[3];
+        const index_t Wo = out_g_n_k_wos_lengths[4];
+
+        const index_t Y = wei_g_k_c_xs_lengths[3];
+        const index_t X = wei_g_k_c_xs_lengths[4];
+
+        const index_t InLeftPadH = input_left_pads[0];
+        const index_t InLeftPadW = input_left_pads[1];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        const index_t ConvDilationH = conv_filter_dilations[0];
+        const index_t ConvDilationW = conv_filter_dilations[1];
+
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+            const auto in_gemmm_gemmn_grid_desc =
+                make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, C), make_tuple(I0, I1));
+
+            return in_gemmm_gemmn_grid_desc;
+        }
+        else
+        {
+            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+            const auto YTilde = ConvStrideH / GcdStrideDilationH;
+            const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+            const auto HTilde =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilde =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IHTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+            const auto IWTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+            const auto IHTildeSliceEnd = math::min(
+                HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildeSliceEnd = math::min(
+                WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+            // bias tensor
+            const auto in_gemmmraw_gemmnraw_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N * HTildeSlice * WTildeSlice, C), make_tuple(I0, I1));
+
+            const auto in_gemmm_gemmn_grid_desc = ck::tensor_operation::device::PadTensorDescriptor(
+                in_gemmmraw_gemmnraw_grid_desc,
+                make_tuple(GemmMPerBlock, GemmNPerBlock),
+                Sequence<DoPadGemmM, DoPadGemmN>{});
+
+            return in_gemmm_gemmn_grid_desc;
+        }
+    }
+};
+
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp b/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp
new file mode 100644
index 0000000000..76bdfad730
--- /dev/null
+++ b/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp
@@ -0,0 +1,882 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/library/utility/numeric.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+
+namespace ck {
+namespace tensor_operation {
+
+template <index_t NDimSpatial, device::ConvolutionForwardSpecialization ConvForwardSpecialization>
+struct TransformConvFwdToGemm
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    template <typename ALayout,
+              typename std::enable_if<NDimSpatial == 1 &&
+                                          is_same_v<ALayout, tensor_layout::convolution::GNWC>,
+                                      bool>::type = false>
+    __host__ __device__ static auto
+    MakeADescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* a_g_n_c_wis_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* b_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Wi = a_g_n_c_wis_lengths[3];
+
+        const index_t Wo = c_g_n_k_wos_lengths[3];
+
+        const index_t ConvStrideW = conv_filter_strides[0];
+
+        if constexpr(ConvForwardSpecialization ==
+                     device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NWo =
+                N * ck::accumulate_n<index_t>(
+                        c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+            const auto in_gemmm_gemmk_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(NWo, C));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          device::ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            const auto in_n_wi_c_desc = make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
+
+            const auto in_n_wo_c_desc = transform_tensor_descriptor(
+                in_n_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_gemmm_gemmk_desc = transform_tensor_descriptor(
+                in_n_wo_c_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Wo)), make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else
+        {
+            const index_t X             = b_g_k_c_xs_lengths[3];
+            const index_t ConvDilationW = conv_filter_dilations[0];
+            const index_t InLeftPadW    = input_left_pads[0];
+            const index_t InRightPadW   = input_right_pads[0];
+
+            const auto in_n_wi_c_desc = make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
+
+            const auto in_n_wip_c_desc = transform_tensor_descriptor(
+                in_n_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_n_x_wo_c_desc = transform_tensor_descriptor(
+                in_n_wip_c_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_gemmm_gemmk_desc =
+                transform_tensor_descriptor(in_n_x_wo_c_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Wo)),
+                                                       make_merge_transform(make_tuple(X, C))),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+    }
+
+    template <typename ALayout,
+              typename std::enable_if<NDimSpatial == 2 &&
+                                          is_same_v<ALayout, tensor_layout::convolution::GNHWC>,
+                                      bool>::type = false>
+    __host__ __device__ static auto
+    MakeADescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* a_g_n_c_wis_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* b_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Hi = a_g_n_c_wis_lengths[3];
+        const index_t Wi = a_g_n_c_wis_lengths[4];
+
+        const index_t Ho = c_g_n_k_wos_lengths[3];
+        const index_t Wo = c_g_n_k_wos_lengths[4];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        if constexpr(ConvForwardSpecialization ==
+                     device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NHoWo =
+                N * ck::accumulate_n<index_t>(
+                        c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+            const auto in_gemmm_gemmk_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(NHoWo, C));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          device::ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            const auto in_n_hi_wi_c_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_ho_wo_c_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmm_gemmk_desc =
+                transform_tensor_descriptor(in_n_ho_wo_c_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else
+        {
+            const index_t Y = b_g_k_c_xs_lengths[3];
+            const index_t X = b_g_k_c_xs_lengths[4];
+
+            const index_t ConvDilationH = conv_filter_dilations[0];
+            const index_t ConvDilationW = conv_filter_dilations[1];
+
+            const index_t InLeftPadH = input_left_pads[0];
+            const index_t InLeftPadW = input_left_pads[1];
+
+            const index_t InRightPadH = input_right_pads[0];
+            const index_t InRightPadW = input_right_pads[1];
+
+            const auto in_n_hi_wi_c_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmm_gemmk_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
+                                                       make_merge_transform(make_tuple(Y, X, C))),
+                                            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+    }
+
+    template <typename ALayout,
+              typename std::enable_if<NDimSpatial == 3 &&
+                                          is_same_v<ALayout, tensor_layout::convolution::GNDHWC>,
+                                      bool>::type = false>
+    __host__ __device__ static auto
+    MakeADescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* a_g_n_c_wis_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* b_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Di = a_g_n_c_wis_lengths[3];
+        const index_t Hi = a_g_n_c_wis_lengths[4];
+        const index_t Wi = a_g_n_c_wis_lengths[5];
+
+        const index_t Do = c_g_n_k_wos_lengths[3];
+        const index_t Ho = c_g_n_k_wos_lengths[4];
+        const index_t Wo = c_g_n_k_wos_lengths[5];
+
+        const index_t ConvStrideD = conv_filter_strides[0];
+        const index_t ConvStrideH = conv_filter_strides[1];
+        const index_t ConvStrideW = conv_filter_strides[2];
+
+        if constexpr(ConvForwardSpecialization ==
+                     device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NDoHoWo =
+                N * ck::accumulate_n<index_t>(
+                        c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+            const auto in_gemmm_gemmk_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(NDoHoWo, C));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          device::ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            const auto in_n_di_hi_wi_c_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
+
+            const auto in_n_do_ho_wo_c_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Do), make_tuple(ConvStrideD)),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_gemmm_gemmk_desc = transform_tensor_descriptor(
+                in_n_do_ho_wo_c_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else
+        {
+            const index_t Z = b_g_k_c_xs_lengths[3];
+            const index_t Y = b_g_k_c_xs_lengths[4];
+            const index_t X = b_g_k_c_xs_lengths[5];
+
+            const index_t ConvDilationD = conv_filter_dilations[0];
+            const index_t ConvDilationH = conv_filter_dilations[1];
+            const index_t ConvDilationW = conv_filter_dilations[2];
+
+            const index_t InLeftPadD = input_left_pads[0];
+            const index_t InLeftPadH = input_left_pads[1];
+            const index_t InLeftPadW = input_left_pads[2];
+
+            const index_t InRightPadD = input_right_pads[0];
+            const index_t InRightPadH = input_right_pads[1];
+            const index_t InRightPadW = input_right_pads[2];
+
+            const auto in_n_di_hi_wi_c_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Di, InLeftPadD, InRightPadD),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_n_z_do_y_ho_x_wo_c_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+            const auto in_gemmm_gemmk_desc = transform_tensor_descriptor(
+                in_n_z_do_y_ho_x_wo_c_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_merge_transform(make_tuple(Z, Y, X, C))),
+                make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+    }
+
+    // TODO: implement ck::tensor_layout::convolution that describe packed/strided dimemsion as
+    // properties
+    template <typename ALayout,
+              typename std::enable_if<NDimSpatial == 1 &&
+                                          (is_same_v<ALayout, tensor_layout::convolution::G_NW_C> ||
+                                           is_same_v<ALayout, tensor_layout::convolution::NWGC>),
+                                      bool>::type = false>
+    __host__ __device__ static auto
+    MakeADescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* b_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Wi = a_g_n_c_wis_lengths[3];
+
+        const index_t Wo = c_g_n_k_wos_lengths[3];
+
+        const index_t ConvStrideW = conv_filter_strides[0];
+
+        if constexpr(ConvForwardSpecialization ==
+                     device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NHoWo =
+                N * ck::accumulate_n<index_t>(
+                        c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+            // This is different
+            const index_t WiStride = a_g_n_c_wis_strides[2 + NDimSpatial];
+            const auto CStride     = I1;
+
+            const auto in_gemmm_gemmk_desc =
+                make_naive_tensor_descriptor(make_tuple(NHoWo, C), make_tuple(WiStride, CStride));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          device::ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t WiStride = a_g_n_c_wis_strides[3];
+            const auto CStride     = I1;
+
+            const auto in_n_wi_c_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Wi, C), make_tuple(NStride, WiStride, CStride));
+
+            const auto in_n_wo_c_desc = transform_tensor_descriptor(
+                in_n_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_gemmm_gemmk_desc = transform_tensor_descriptor(
+                in_n_wo_c_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Wo)), make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else
+        {
+            const index_t X             = b_g_k_c_xs_lengths[3];
+            const index_t ConvDilationW = conv_filter_dilations[0];
+            const index_t InLeftPadW    = input_left_pads[0];
+            const index_t InRightPadW   = input_right_pads[0];
+
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t WiStride = a_g_n_c_wis_strides[3];
+            const auto CStride     = I1;
+
+            const auto in_n_wi_c_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Wi, C), make_tuple(NStride, WiStride, CStride));
+
+            const auto in_n_wip_c_desc = transform_tensor_descriptor(
+                in_n_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_n_x_wo_c_desc = transform_tensor_descriptor(
+                in_n_wip_c_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_gemmm_gemmk_desc =
+                transform_tensor_descriptor(in_n_x_wo_c_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Wo)),
+                                                       make_merge_transform(make_tuple(X, C))),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+    }
+
+    template <typename ALayout,
+              typename std::enable_if<
+                  NDimSpatial == 2 && (is_same_v<ALayout, tensor_layout::convolution::G_NHW_C> ||
+                                       is_same_v<ALayout, tensor_layout::convolution::NHWGC>),
+                  bool>::type = false>
+    __host__ __device__ static auto
+    MakeADescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* b_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Hi = a_g_n_c_wis_lengths[3];
+        const index_t Wi = a_g_n_c_wis_lengths[4];
+
+        const index_t Ho = c_g_n_k_wos_lengths[3];
+        const index_t Wo = c_g_n_k_wos_lengths[4];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        if constexpr(ConvForwardSpecialization ==
+                     device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NHoWo =
+                N * ck::accumulate_n<index_t>(
+                        c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+            // This is different
+            const index_t WiStride = a_g_n_c_wis_strides[2 + NDimSpatial];
+            const auto CStride     = I1;
+
+            const auto in_gemmm_gemmk_desc =
+                make_naive_tensor_descriptor(make_tuple(NHoWo, C), make_tuple(WiStride, CStride));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          device::ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t HiStride = a_g_n_c_wis_strides[3];
+            const index_t WiStride = a_g_n_c_wis_strides[4];
+            const auto CStride     = I1;
+
+            const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Hi, Wi, C), make_tuple(NStride, HiStride, WiStride, CStride));
+
+            const auto in_n_ho_wo_c_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmm_gemmk_desc =
+                transform_tensor_descriptor(in_n_ho_wo_c_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else
+        {
+            const index_t Y = b_g_k_c_xs_lengths[3];
+            const index_t X = b_g_k_c_xs_lengths[4];
+
+            const index_t ConvDilationH = conv_filter_dilations[0];
+            const index_t ConvDilationW = conv_filter_dilations[1];
+
+            const index_t InLeftPadH = input_left_pads[0];
+            const index_t InLeftPadW = input_left_pads[1];
+
+            const index_t InRightPadH = input_right_pads[0];
+            const index_t InRightPadW = input_right_pads[1];
+
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t HiStride = a_g_n_c_wis_strides[3];
+            const index_t WiStride = a_g_n_c_wis_strides[4];
+            const auto CStride     = I1;
+
+            const auto in_n_hi_wi_c_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Hi, Wi, C), make_tuple(NStride, HiStride, WiStride, CStride));
+
+            const auto in_n_hip_wip_c_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmm_gemmk_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
+                                                       make_merge_transform(make_tuple(Y, X, C))),
+                                            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+    }
+
+    template <typename ALayout,
+              typename std::enable_if<
+                  NDimSpatial == 3 && (is_same_v<ALayout, tensor_layout::convolution::G_NDHW_C> ||
+                                       is_same_v<ALayout, tensor_layout::convolution::NDHWGC>),
+                  bool>::type = false>
+    __host__ __device__ static auto
+    MakeADescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* b_g_k_c_xs_strides */,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        const index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Di = a_g_n_c_wis_lengths[3];
+        const index_t Hi = a_g_n_c_wis_lengths[4];
+        const index_t Wi = a_g_n_c_wis_lengths[5];
+
+        const index_t Do = c_g_n_k_wos_lengths[3];
+        const index_t Ho = c_g_n_k_wos_lengths[4];
+        const index_t Wo = c_g_n_k_wos_lengths[5];
+
+        const index_t ConvStrideD = conv_filter_strides[0];
+        const index_t ConvStrideH = conv_filter_strides[1];
+        const index_t ConvStrideW = conv_filter_strides[2];
+
+        if constexpr(ConvForwardSpecialization ==
+                     device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NDoHoWo =
+                N * ck::accumulate_n<index_t>(
+                        c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+            // This is different
+            const index_t WiStride = a_g_n_c_wis_strides[2 + NDimSpatial];
+            const auto CStride     = I1;
+
+            const auto in_gemmm_gemmk_desc =
+                make_naive_tensor_descriptor(make_tuple(NDoHoWo, C), make_tuple(WiStride, CStride));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          device::ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t DiStride = a_g_n_c_wis_strides[3];
+            const index_t HiStride = a_g_n_c_wis_strides[4];
+            const index_t WiStride = a_g_n_c_wis_strides[5];
+            const auto CStride     = I1;
+
+            const auto in_n_di_hi_wi_c_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Di, Hi, Wi, C),
+                make_tuple(NStride, DiStride, HiStride, WiStride, CStride));
+
+            const auto in_n_do_ho_wo_c_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Do), make_tuple(ConvStrideD)),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_gemmm_gemmk_desc = transform_tensor_descriptor(
+                in_n_do_ho_wo_c_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+        else
+        {
+            const index_t Z = b_g_k_c_xs_lengths[3];
+            const index_t Y = b_g_k_c_xs_lengths[4];
+            const index_t X = b_g_k_c_xs_lengths[5];
+
+            const index_t ConvDilationD = conv_filter_dilations[0];
+            const index_t ConvDilationH = conv_filter_dilations[1];
+            const index_t ConvDilationW = conv_filter_dilations[2];
+
+            const index_t InLeftPadD = input_left_pads[0];
+            const index_t InLeftPadH = input_left_pads[1];
+            const index_t InLeftPadW = input_left_pads[2];
+
+            const index_t InRightPadD = input_right_pads[0];
+            const index_t InRightPadH = input_right_pads[1];
+            const index_t InRightPadW = input_right_pads[2];
+
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t DiStride = a_g_n_c_wis_strides[3];
+            const index_t HiStride = a_g_n_c_wis_strides[4];
+            const index_t WiStride = a_g_n_c_wis_strides[5];
+            const auto CStride     = I1;
+
+            const auto in_n_di_hi_wi_c_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Di, Hi, Wi, C),
+                make_tuple(NStride, DiStride, HiStride, WiStride, CStride));
+
+            const auto in_n_hip_wip_c_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Di, InLeftPadD, InRightPadD),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_n_z_do_y_ho_x_wo_c_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+            const auto in_gemmm_gemmk_desc = transform_tensor_descriptor(
+                in_n_z_do_y_ho_x_wo_c_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_merge_transform(make_tuple(Z, Y, X, C))),
+                make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return in_gemmm_gemmk_desc;
+        }
+    }
+
+    template <typename BLayout,
+              typename std::enable_if<is_same_v<BLayout, tensor_layout::convolution::GKXC> ||
+                                          is_same_v<BLayout, tensor_layout::convolution::GKYXC> ||
+                                          is_same_v<BLayout, tensor_layout::convolution::GKZYXC>,
+                                      bool>::type = false>
+    __host__ __device__ static auto
+    MakeBDescriptor_N_K(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* b_g_k_c_xs_strides */)
+    {
+        const index_t K = b_g_k_c_xs_lengths[1];
+        const index_t C = b_g_k_c_xs_lengths[2];
+
+        const index_t YX = ck::accumulate_n<index_t>(
+            b_g_k_c_xs_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+        const auto wei_gemmn_gemmk_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(K, YX * C));
+
+        return wei_gemmn_gemmk_desc;
+    }
+
+    template <
+        typename BLayout,
+        typename std::enable_if<is_same_v<BLayout, tensor_layout::convolution::G_K_X_C> ||
+                                    is_same_v<BLayout, tensor_layout::convolution::G_K_YX_C> ||
+                                    is_same_v<BLayout, tensor_layout::convolution::G_K_ZYX_C> ||
+                                    is_same_v<BLayout, tensor_layout::convolution::KXGC> ||
+                                    is_same_v<BLayout, tensor_layout::convolution::KYXGC> ||
+                                    is_same_v<BLayout, tensor_layout::convolution::KZYXGC>,
+                                bool>::type = false>
+    __host__ __device__ static auto
+    MakeBDescriptor_N_K(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const index_t K = b_g_k_c_xs_lengths[1];
+        const index_t C = b_g_k_c_xs_lengths[2];
+
+        const index_t YX = ck::accumulate_n<index_t>(
+            b_g_k_c_xs_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+        const index_t KStride = b_g_k_c_xs_strides[1];
+        const index_t XStride = b_g_k_c_xs_strides[2 + NDimSpatial];
+        const auto CStride    = I1;
+
+        const auto wei_k_yx_c_desc = make_naive_tensor_descriptor(
+            make_tuple(K, YX, C), make_tuple(KStride, XStride, CStride));
+
+        const auto wei_gemmn_gemmk_desc = transform_tensor_descriptor(
+            wei_k_yx_c_desc,
+            make_tuple(make_pass_through_transform(K), make_merge_transform(make_tuple(YX, C))),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return wei_gemmn_gemmk_desc;
+    }
+
+    template <typename CLayout,
+              typename std::enable_if<is_same_v<CLayout, tensor_layout::convolution::GNWK> ||
+                                          is_same_v<CLayout, tensor_layout::convolution::GNHWK> ||
+                                          is_same_v<CLayout, tensor_layout::convolution::GNDHWK>,
+                                      bool>::type = false>
+    __host__ __device__ static auto
+    MakeCDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */)
+    {
+        const index_t N = c_g_n_k_wos_lengths[1];
+        const index_t K = c_g_n_k_wos_lengths[2];
+
+        const index_t NHoWo =
+            N * ck::accumulate_n<index_t>(
+                    c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+        const auto out_gemmm_gemmn_desc = make_naive_tensor_descriptor_packed(make_tuple(NHoWo, K));
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    template <
+        typename CLayout,
+        typename std::enable_if<is_same_v<CLayout, tensor_layout::convolution::G_NW_K> ||
+                                    is_same_v<CLayout, tensor_layout::convolution::G_NHW_K> ||
+                                    is_same_v<CLayout, tensor_layout::convolution::G_NDHW_K> ||
+                                    is_same_v<CLayout, tensor_layout::convolution::NWGK> ||
+                                    is_same_v<CLayout, tensor_layout::convolution::NHWGK> ||
+                                    is_same_v<CLayout, tensor_layout::convolution::NDHWGK>,
+                                bool>::type = false>
+    __host__ __device__ static auto
+    MakeCDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_strides)
+    {
+        const index_t N = c_g_n_k_wos_lengths[1];
+        const index_t K = c_g_n_k_wos_lengths[2];
+
+        const auto KStride     = I1;
+        const index_t WoStride = c_g_n_k_wos_strides[NDimSpatial + 2];
+
+        const index_t NHoWo =
+            N * ck::accumulate_n<index_t>(
+                    c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+        const auto out_gemmm_gemmn_desc =
+            make_naive_tensor_descriptor(make_tuple(NHoWo, K), make_tuple(WoStride, KStride));
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    // for output bias
+    template <typename CLayout,
+              typename std::enable_if<is_same_v<CLayout, tensor_layout::convolution::GK> ||
+                                          is_same_v<CLayout, tensor_layout::convolution::G_K>,
+                                      bool>::type = false>
+    __host__ __device__ static auto
+    MakeCDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& c_g_n_k_wos_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& /* c_g_n_k_wos_strides */)
+    {
+        const index_t N = c_g_n_k_wos_lengths[1];
+        const index_t K = c_g_n_k_wos_lengths[2];
+
+        const index_t NHoWo =
+            N * ck::accumulate_n<index_t>(
+                    c_g_n_k_wos_lengths.begin() + 3, NDimSpatial, 1, std::multiplies<>());
+
+        const auto out_gemmm_gemmn_desc =
+            make_naive_tensor_descriptor(make_tuple(NHoWo, K), make_tuple(I0, I1));
+
+        return out_gemmm_gemmn_desc;
+    }
+};
+
+} // namespace tensor_operation
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile/block_gemm_areg_bgmem_creg_problem.hpp b/include/ck/tile_program/block_tile/block_gemm_areg_bgmem_creg_problem.hpp
new file mode 100644
index 0000000000..5e986243cc
--- /dev/null
+++ b/include/ck/tile_program/block_tile/block_gemm_areg_bgmem_creg_problem.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/type.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+// Problem Description for BlockGemmARegBGmemCReg
+template <typename ADataType_,
+          typename BDataType_,
+          typename CDataType_,
+          index_t kBlockSize_,
+          typename BlockGemmShape_>
+struct BlockGemmARegBGmemCRegProblem
+{
+    using ADataType      = remove_cvref_t<ADataType_>;
+    using BDataType      = remove_cvref_t<BDataType_>;
+    using CDataType      = remove_cvref_t<CDataType_>;
+    using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
+
+    static constexpr index_t kBlockSize = kBlockSize_;
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile/block_gemm_areg_bgmem_creg_v1.hpp b/include/ck/tile_program/block_tile/block_gemm_areg_bgmem_creg_v1.hpp
new file mode 100644
index 0000000000..39cab3bc7f
--- /dev/null
+++ b/include/ck/tile_program/block_tile/block_gemm_areg_bgmem_creg_v1.hpp
@@ -0,0 +1,146 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/static_tile_distribution_helper.hpp"
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile/block_gemm_areg_bgmem_creg_problem.hpp"
+#include "ck/tile_program/block_tile/block_gemm_areg_bgmem_creg_v1_default_policy.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+// A is block distributed tensor
+// B is block window on global memory
+// C is block distributed tensor
+// This will:
+//   1. Load B from global memory into shared memory and then
+//   2. Call BlockGemmARegSGmemCRegV1
+template <typename Problem, typename Policy = BlockGemmARegBGmemCRegV1DefaultPolicy>
+struct BlockGemmARegBGmemCRegV1
+{
+    using ADataType      = remove_cvref_t<typename Problem::ADataType>;
+    using BDataType      = remove_cvref_t<typename Problem::BDataType>;
+    using CDataType      = remove_cvref_t<typename Problem::CDataType>;
+    using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    // use BlockGemmARegBSmemCRegV1 as the underlying block-GEMM implementation
+    using BlockGemmARegBSmemCRegImpl = BlockGemmARegBSmemCRegV1<
+        BlockGemmARegBSmemCRegProblem<ADataType, BDataType, CDataType, kBlockSize, BlockGemmShape>,
+        BlockGemmARegBSmemCRegV1DefaultPolicy>;
+
+    __host__ __device__ static constexpr ck::index_t GetStaticLdsSize()
+    {
+        return sizeof(BDataType) *
+               Policy::template MakeBSmemBlockDescriptor<Problem>().GetElementSpaceSize();
+    }
+
+    // C += A * B
+    template <typename CBlockTensor, typename ABlockTensor, typename BBlockGmemWindowTmp>
+    __device__ void operator()(CBlockTensor& c_block_tensor,
+                               const ABlockTensor& a_block_tensor,
+                               const BBlockGmemWindowTmp& b_block_gmem_window_tmp,
+                               void* smem_ptr) const
+    {
+        static_assert(
+            is_same_v<ADataType, remove_cv_t<typename ABlockTensor::DataType>> &&
+                is_same_v<BDataType, remove_cv_t<typename BBlockGmemWindowTmp::DataType>> &&
+                is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>,
+            "wrong!");
+
+        constexpr index_t MPerBlock = ABlockTensor{}.GetLengths()[Number<0>{}];
+        constexpr index_t NPerBlock = BBlockGmemWindowTmp{}.GetWindowLengths()[Number<0>{}];
+        constexpr index_t KPerBlock = ABlockTensor{}.GetLengths()[Number<1>{}];
+
+        static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
+                          KPerBlock == BlockGemmShape::kK,
+                      "wrong!");
+
+        const auto b_block_gmem_window =
+            make_tile_window(b_block_gmem_window_tmp.GetBottomTensorView(),
+                             make_tuple(Number<NPerBlock>{}, Number<KPerBlock>{}),
+                             b_block_gmem_window_tmp.GetWindowOrigin(),
+                             Policy::template MakeBGmemTileDistribution<Problem>());
+
+        // B LDS and LDS window
+        auto b_block_smem = make_tensor_view<AddressSpaceEnum::Lds>(
+            reinterpret_cast<BDataType*>(smem_ptr),
+            Policy::template MakeBSmemBlockDescriptor<Problem>());
+
+        auto b_block_smem_window = make_tile_window(
+            b_block_smem, make_tuple(Number<MPerBlock>{}, Number<KPerBlock>{}), {0, 0});
+
+        // load B tile from global mem
+        const auto b_block_tile = load_tile(b_block_gmem_window);
+
+        // store B tile into shared mem
+        store_tile(b_block_smem_window, b_block_tile);
+
+        // wait for store_tile to finish
+        block_sync_lds();
+
+        // block GEMM
+        BlockGemmARegBSmemCRegImpl{}(c_block_tensor, a_block_tensor, b_block_smem_window);
+    }
+
+    // C = A * B
+    template <typename ABlockTensor, typename BBlockGmemWindowTmp>
+    __device__ auto operator()(const ABlockTensor& a_block_tensor,
+                               const BBlockGmemWindowTmp& b_block_gmem_window_tmp,
+                               void* smem_ptr) const
+    {
+        static_assert(is_same_v<ADataType, remove_cv_t<typename ABlockTensor::DataType>> &&
+                          is_same_v<BDataType, remove_cv_t<typename BBlockGmemWindowTmp::DataType>>,
+                      "wrong!");
+
+        constexpr index_t MPerBlock = ABlockTensor{}.GetLengths()[Number<0>{}];
+        constexpr index_t NPerBlock = BBlockGmemWindowTmp{}.GetWindowLengths()[Number<0>{}];
+        constexpr index_t KPerBlock = ABlockTensor{}.GetLengths()[Number<1>{}];
+
+        static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
+                          KPerBlock == BlockGemmShape::kK,
+                      "wrong!");
+
+        const auto b_block_gmem_window =
+            make_tile_window(b_block_gmem_window_tmp.GetBottomTensorView(),
+                             make_tuple(Number<NPerBlock>{}, Number<KPerBlock>{}),
+                             b_block_gmem_window_tmp.GetWindowOrigin(),
+                             Policy::template MakeBGmemTileDistribution<Problem>());
+
+        // B LDS and LDS window
+        auto b_block_smem = make_tensor_view<AddressSpaceEnum::Lds>(
+            reinterpret_cast<BDataType*>(smem_ptr),
+            Policy::template MakeBSmemBlockDescriptor<Problem>());
+
+        auto b_block_smem_window = make_tile_window(
+            b_block_smem, make_tuple(Number<MPerBlock>{}, Number<KPerBlock>{}), {0, 0});
+
+        // load B tile from global mem
+        const auto b_block_tile = load_tile(b_block_gmem_window);
+
+        // store B tile into shared mem
+        store_tile(b_block_smem_window, b_block_tile);
+
+        // wait for store_tile to finish
+        block_sync_lds();
+
+        // block GEMM
+        return BlockGemmARegBSmemCRegImpl{}(a_block_tensor, b_block_smem_window);
+    }
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile/block_gemm_areg_bgmem_creg_v1_default_policy.hpp b/include/ck/tile_program/block_tile/block_gemm_areg_bgmem_creg_v1_default_policy.hpp
new file mode 100644
index 0000000000..b90e006a8a
--- /dev/null
+++ b/include/ck/tile_program/block_tile/block_gemm_areg_bgmem_creg_v1_default_policy.hpp
@@ -0,0 +1,131 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+// Default policy for BlockGemmARegBGmemCRegV1
+// Default policy class should not be templated, put template on member functions instead
+struct BlockGemmARegBGmemCRegV1DefaultPolicy
+{
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeBGmemTileDistribution()
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+
+        using BDataType = remove_cvref_t<typename Problem::BDataType>;
+
+        constexpr index_t kBlockSize = Problem::kBlockSize;
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr index_t K1 = 16 / sizeof(BDataType);
+        constexpr index_t K0 = kKPerBlock / K1;
+        constexpr index_t N2 = get_warp_size() / K0;
+        constexpr index_t N1 = kBlockSize / get_warp_size();
+        constexpr index_t N0 = kNPerBlock / (N2 * N1);
+
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<1>,
+                                           Tuple<Sequence<N0, N1, N2>, Sequence<K0, K1>>,
+                                           Tuple<Sequence<1>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1>, Sequence<2, 0>>,
+                                           Sequence<1, 2>,
+                                           Sequence<0, 1>>{});
+    }
+
+#if 0
+    // 2d
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeBLdsBlockDescriptor()
+    {
+        using namespace ck;
+
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr auto b_lds_block_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(kNPerBlock, kKPerBlock), Number<32>{});
+
+        return b_lds_block_desc;
+    }
+#elif 0
+    // 3d + padding
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeBSmemBlockDescriptor()
+    {
+        using namespace ck;
+
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr auto b_lds_block_desc_0 = make_naive_tensor_descriptor(
+            make_tuple(Number<kKPerBlock / 8>{}, Number<kNPerBlock>{}, Number<8>{}),
+            make_tuple(Number<(kNPerBlock + 1) * 8>{}, Number<8>{}, Number<1>{}),
+            Number<8>{},
+            Number<1>{});
+
+        constexpr auto b_lds_block_desc = transform_tensor_descriptor(
+            b_lds_block_desc_0,
+            make_tuple(make_pass_through_transform(kNPerBlock),
+                       make_merge_transform(make_tuple(kKPerBlock / 8, 8))),
+            make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return b_lds_block_desc;
+    }
+#elif 1
+    // fake XOR
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeBSmemBlockDescriptor()
+    {
+        using namespace ck;
+
+        using BDataType = remove_cvref_t<typename Problem::BDataType>;
+
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr auto b_lds_block_desc_d1_d2_d3 = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<kNPerBlock / 2>{}, Number<2>{}, Number<kKPerBlock>{}),
+            Number<kKPerBlock>{});
+
+        constexpr index_t kK1 = 16 / sizeof(BDataType);
+
+        constexpr auto b_lds_block_desc_d4_d5_d6 = transform_tensor_descriptor(
+            b_lds_block_desc_d1_d2_d3,
+            make_tuple(
+                make_xor_transform(make_tuple(Number<kNPerBlock / 2>{}, Number<kKPerBlock>{}), kK1),
+                make_pass_through_transform(2)),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+        constexpr auto b_lds_block_desc_n_k = transform_tensor_descriptor(
+            b_lds_block_desc_d4_d5_d6,
+            make_tuple(make_merge_transform(make_tuple(Number<kNPerBlock / 2>{}, Number<2>{})),
+                       make_pass_through_transform(kKPerBlock)),
+            make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return b_lds_block_desc_n_k;
+    }
+#endif
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_problem.hpp b/include/ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_problem.hpp
new file mode 100644
index 0000000000..2b01568ed2
--- /dev/null
+++ b/include/ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_problem.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/type.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+// Problem Description for BlockGemmARegBSmemCReg
+template <typename ADataType_,
+          typename BDataType_,
+          typename CDataType_,
+          index_t kBlockSize_,
+          typename BlockGemmShape_>
+struct BlockGemmARegBSmemCRegProblem
+{
+    using ADataType      = remove_cvref_t<ADataType_>;
+    using BDataType      = remove_cvref_t<BDataType_>;
+    using CDataType      = remove_cvref_t<CDataType_>;
+    using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
+
+    static constexpr index_t kBlockSize = kBlockSize_;
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1.hpp b/include/ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1.hpp
new file mode 100644
index 0000000000..ffc37db87a
--- /dev/null
+++ b/include/ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1.hpp
@@ -0,0 +1,344 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/static_tile_distribution_helper.hpp"
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_problem.hpp"
+#include "ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1_default_policy.hpp"
+#include "ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1_iteratek_policy.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+// A is block distributed tensor
+// B is block window on shared memory
+// C is block distributed tensor
+template <typename Problem, typename Policy = BlockGemmARegBSmemCRegV1DefaultPolicy>
+struct BlockGemmARegBSmemCRegV1
+{
+    using ADataType       = remove_cvref_t<typename Problem::ADataType>;
+    using BDataType       = remove_cvref_t<typename Problem::BDataType>;
+    using CDataType       = remove_cvref_t<typename Problem::CDataType>;
+    using BlockGemmShape  = remove_cvref_t<typename Problem::BlockGemmShape>;
+    using BlockGemmPolicy = Policy;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    // C += A * B
+    template <typename CBlockTensor, typename ABlockTensorTmp, typename BBlockWindowTmp>
+    __device__ void operator()(CBlockTensor& c_block_tensor,
+                               const ABlockTensorTmp& a_block_tensor_tmp,
+                               const BBlockWindowTmp& b_block_window_tmp) const
+    {
+        static_assert(is_same_v<ADataType, remove_cv_t<typename ABlockTensorTmp::DataType>> &&
+                          is_same_v<BDataType, remove_cv_t<typename BBlockWindowTmp::DataType>> &&
+                          is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>,
+                      "wrong!");
+
+        constexpr index_t MPerBlock = ABlockTensorTmp{}.GetLengths()[Number<0>{}];
+        constexpr index_t NPerBlock = BBlockWindowTmp{}.GetWindowLengths()[Number<0>{}];
+        constexpr index_t KPerBlock = ABlockTensorTmp{}.GetLengths()[Number<1>{}];
+
+        static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
+                          KPerBlock == BlockGemmShape::kK,
+                      "wrong!");
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template At<0>())>;
+
+        constexpr index_t MWarp = config.template At<1>();
+        constexpr index_t NWarp = config.template At<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+        constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr index_t NPerBlockPerIter = NPerBlock / NIterPerWarp;
+        constexpr index_t KPerBlockPerIter = KPerBlock / KIterPerWarp;
+
+        const index_t iNWarp = get_warp_id() % NWarp;
+
+        constexpr auto a_block_outer_dstr_encoding = StaticTileDistributionEncoding<
+            Sequence<NWarp>,
+            Tuple<Sequence<MIterPerWarp, MWarp>, Sequence<KIterPerWarp>>,
+            Tuple<Sequence<1, 0>>,
+            Tuple<Sequence<1, 0>>,
+            Sequence<1, 2>,
+            Sequence<0, 0>>{};
+
+        constexpr auto c_block_outer_dstr_encoding = StaticTileDistributionEncoding<
+            Sequence<>,
+            Tuple<Sequence<MIterPerWarp, MWarp>, Sequence<NIterPerWarp, NWarp>>,
+            Tuple<Sequence<1, 2>>,
+            Tuple<Sequence<1, 1>>,
+            Sequence<1, 2>,
+            Sequence<0, 0>>{};
+
+        constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            a_block_outer_dstr_encoding, typename WG::AWarpDstrEncoding{});
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
+
+        constexpr auto a_block_dstr = make_static_tile_distribution(a_block_dstr_encode);
+
+        // constrcut from A-block-tensor from A-Block-tensor-tmp
+        // FIXME: need method to check a_block_tensor and a_block_tensor_tmp have equivalent
+        // distribution
+        auto a_block_tensor =
+            make_static_distributed_tensor<typename ABlockTensorTmp::DataType>(a_block_dstr);
+
+        a_block_tensor.GetThreadBuffer() = a_block_tensor_tmp.GetThreadBuffer();
+
+        // construct B-warp-window
+        auto b_warp_window_tmp = make_tile_window(
+            b_block_window_tmp.GetBottomTensorView(),
+            make_tuple(Number<WG::kN>{}, Number<WG::kK>{}),
+            b_block_window_tmp.GetWindowOrigin() + MultiIndex<2>{iNWarp * WG::kN, 0},
+            make_static_tile_distribution(typename WG::BWarpDstrEncoding{}));
+
+#if 0 // FIXME: using Array will cause register spill
+        Array<Array<decltype(b_warp_window_tmp), KIterPerWarp>, NIterPerWarp> b_warp_windows{
+            {b_warp_window_tmp}};
+
+        for(index_t nIter = 0; nIter < NIterPerWarp; nIter++)
+        {
+            for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
+            {
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            }
+        }
+#else
+        StaticallyIndexedArray<StaticallyIndexedArray<decltype(b_warp_window_tmp), KIterPerWarp>,
+                               NIterPerWarp>
+            b_warp_windows;
+
+        static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+            static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
+
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            });
+        });
+#endif
+
+        // check C-block-distribution
+        static_assert(is_same_v<remove_cvref_t<decltype(c_block_dstr_encode)>,
+                                remove_cvref_t<decltype(CBlockTensor::GetTileDistribution()
+                                                            .GetStaticTileDistributionEncoding())>>,
+                      "wrong!");
+
+        using AWarpDstr = typename WG::AWarpDstr;
+        using CWarpDstr = typename WG::CWarpDstr;
+
+        using AWarpTensor = typename WG::AWarpTensor;
+        using CWarpTensor = typename WG::CWarpTensor;
+
+        constexpr auto a_warp_y_lengths = to_sequence(AWarpDstr{}.GetYs2DDescriptor().GetLengths());
+        constexpr auto c_warp_y_lengths = to_sequence(CWarpDstr{}.GetYs2DDescriptor().GetLengths());
+
+        constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t<AWarpDstr::NDimY, 0>{};
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        // hot loop:
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                // read A warp tensor from A block tensor
+                AWarpTensor a_warp_tensor;
+
+                a_warp_tensor.GetThreadBuffer() = a_block_tensor.GetYSlicedThreadData(
+                    merge_sequences(Sequence<mIter, kIter>{}, a_warp_y_index_zeros),
+                    merge_sequences(Sequence<1, 1>{}, a_warp_y_lengths));
+
+                static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                    // read B warp tensor from B Block window
+                    const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
+                    // read C warp tensor from C block tensor
+                    CWarpTensor c_warp_tensor;
+
+                    c_warp_tensor.GetThreadBuffer() = c_block_tensor.GetYSlicedThreadData(
+                        merge_sequences(Sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(Sequence<1, 1>{}, c_warp_y_lengths));
+
+                    // warp GEMM
+                    WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
+
+                    // write C warp tensor into C block tensor
+                    c_block_tensor.SetYSlicedThreadData(
+                        merge_sequences(Sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(Sequence<1, 1>{}, c_warp_y_lengths),
+                        c_warp_tensor.GetThreadBuffer());
+                });
+            });
+        });
+    }
+
+    // C = A * B
+    template <typename ABlockTensorTmp, typename BBlockWindowTmp>
+    __device__ auto operator()(const ABlockTensorTmp& a_block_tensor_tmp,
+                               const BBlockWindowTmp& b_block_window_tmp) const
+    {
+        static_assert(is_same_v<ADataType, remove_cv_t<typename ABlockTensorTmp::DataType>> &&
+                          is_same_v<BDataType, remove_cv_t<typename BBlockWindowTmp::DataType>>,
+                      "wrong!");
+
+        constexpr index_t MPerBlock = ABlockTensorTmp{}.GetLengths()[Number<0>{}];
+        constexpr index_t NPerBlock = BBlockWindowTmp{}.GetWindowLengths()[Number<0>{}];
+        constexpr index_t KPerBlock = ABlockTensorTmp{}.GetLengths()[Number<1>{}];
+
+        static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
+                          KPerBlock == BlockGemmShape::kK,
+                      "wrong!");
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template At<0>())>;
+
+        constexpr index_t MWarp = config.template At<1>();
+        constexpr index_t NWarp = config.template At<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+        constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr index_t NPerBlockPerIter = NPerBlock / NIterPerWarp;
+        constexpr index_t KPerBlockPerIter = KPerBlock / KIterPerWarp;
+
+        const index_t iNWarp = get_warp_id() % NWarp;
+
+        constexpr auto a_block_outer_dstr_encoding = StaticTileDistributionEncoding<
+            Sequence<NWarp>,
+            Tuple<Sequence<MIterPerWarp, MWarp>, Sequence<KIterPerWarp>>,
+            Tuple<Sequence<1, 0>>,
+            Tuple<Sequence<1, 0>>,
+            Sequence<1, 2>,
+            Sequence<0, 0>>{};
+
+        constexpr auto c_block_outer_dstr_encoding = StaticTileDistributionEncoding<
+            Sequence<>,
+            Tuple<Sequence<MIterPerWarp, MWarp>, Sequence<NIterPerWarp, NWarp>>,
+            Tuple<Sequence<1, 2>>,
+            Tuple<Sequence<1, 1>>,
+            Sequence<1, 2>,
+            Sequence<0, 0>>{};
+
+        constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            a_block_outer_dstr_encoding, typename WG::AWarpDstrEncoding{});
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
+
+        constexpr auto a_block_dstr = make_static_tile_distribution(a_block_dstr_encode);
+        constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+
+        // constrcut from A-block-tensor from A-Block-tensor-tmp
+        // FIXME: need method to check a_block_tensor and a_block_tensor_tmp have equivalent
+        // distribution
+        auto a_block_tensor =
+            make_static_distributed_tensor<typename ABlockTensorTmp::DataType>(a_block_dstr);
+
+        a_block_tensor.GetThreadBuffer() = a_block_tensor_tmp.GetThreadBuffer();
+
+        // construct B-warp-window
+        auto b_warp_window_tmp = make_tile_window(
+            b_block_window_tmp.GetBottomTensorView(),
+            make_tuple(Number<WG::kN>{}, Number<WG::kK>{}),
+            b_block_window_tmp.GetWindowOrigin() + MultiIndex<2>{iNWarp * WG::kN, 0},
+            make_static_tile_distribution(typename WG::BWarpDstrEncoding{}));
+
+#if 0 // FIXME: using Array will cause register spill
+        Array<Array<decltype(b_warp_window_tmp), KIterPerWarp>, NIterPerWarp> b_warp_windows{
+            {b_warp_window_tmp}};
+
+        for(index_t nIter = 0; nIter < NIterPerWarp; nIter++)
+        {
+            for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
+            {
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            }
+        }
+#else
+        StaticallyIndexedArray<StaticallyIndexedArray<decltype(b_warp_window_tmp), KIterPerWarp>,
+                               NIterPerWarp>
+            b_warp_windows;
+
+        static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+            static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
+
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            });
+        });
+#endif
+
+        // Construct C-Block-Tensor
+        auto c_block_tensor = make_static_distributed_tensor<CDataType>(c_block_dstr);
+
+        using AWarpDstr = typename WG::AWarpDstr;
+        using CWarpDstr = typename WG::CWarpDstr;
+
+        using AWarpTensor = typename WG::AWarpTensor;
+        using CWarpTensor = typename WG::CWarpTensor;
+
+        constexpr auto a_warp_y_lengths = to_sequence(AWarpDstr{}.GetYs2DDescriptor().GetLengths());
+        constexpr auto c_warp_y_lengths = to_sequence(CWarpDstr{}.GetYs2DDescriptor().GetLengths());
+
+        constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t<AWarpDstr::NDimY, 0>{};
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        // hot loop:
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                // read A warp tensor from A block tensor
+                AWarpTensor a_warp_tensor;
+
+                a_warp_tensor.GetThreadBuffer() = a_block_tensor.GetYSlicedThreadData(
+                    merge_sequences(Sequence<mIter, kIter>{}, a_warp_y_index_zeros),
+                    merge_sequences(Sequence<1, 1>{}, a_warp_y_lengths));
+
+                static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                    // read B warp tensor from B Block window
+                    const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
+
+                    // read C warp tensor from C block tensor
+                    CWarpTensor c_warp_tensor;
+
+                    c_warp_tensor.GetThreadBuffer() = c_block_tensor.GetYSlicedThreadData(
+                        merge_sequences(Sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(Sequence<1, 1>{}, c_warp_y_lengths));
+
+                    // warp GEMM
+                    WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
+
+                    // write C warp tensor into C block tensor
+                    c_block_tensor.SetYSlicedThreadData(
+                        merge_sequences(Sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(Sequence<1, 1>{}, c_warp_y_lengths),
+                        c_warp_tensor.GetThreadBuffer());
+                });
+            });
+        });
+
+        return c_block_tensor;
+    }
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1_default_policy.hpp b/include/ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1_default_policy.hpp
new file mode 100644
index 0000000000..2a67ec2cb9
--- /dev/null
+++ b/include/ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1_default_policy.hpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+// Default policy for BlockGemmARegBSmemCRegV1
+// Default policy class should not be templated, put template on member functions instead
+struct BlockGemmARegBSmemCRegV1DefaultPolicy
+{
+    template <typename Problem>
+    __host__ __device__ static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+        using namespace ck::tile_program::warp;
+
+#if 0
+        constexpr index_t kBlockSize = Problem::kBlockSize;
+
+        constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        static_assert(kBlockSize % get_warp_size() == 0, "wrong!");
+
+        constexpr index_t NumWarp = kBlockSize / get_warp_size();
+
+        // FIXME
+        if constexpr(NumWarp == 4 && kMPerBlock % 128 == 0 &&
+                     kNPerBlock % 128 == 0 % kKPerBlock % 16 == 0)
+        {
+            return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
+        }
+        else
+        {
+            return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
+        }
+#else
+        return make_tuple(WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution{}, 4, 1);
+#endif
+    }
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1_iteratek_policy.hpp b/include/ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1_iteratek_policy.hpp
new file mode 100644
index 0000000000..17590a90e3
--- /dev/null
+++ b/include/ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1_iteratek_policy.hpp
@@ -0,0 +1,33 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+struct BlockGemmARegBSmemCRegV1K8Policy
+{
+    template <typename Problem>
+    __host__ __device__ static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+        using namespace ck::tile_program::warp;
+
+        return make_tuple(WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution{}, 4, 1);
+    }
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile/block_gemm_asmem_bsmem_creg_problem.hpp b/include/ck/tile_program/block_tile/block_gemm_asmem_bsmem_creg_problem.hpp
new file mode 100644
index 0000000000..b5a1a06af4
--- /dev/null
+++ b/include/ck/tile_program/block_tile/block_gemm_asmem_bsmem_creg_problem.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/type.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+// Problem Description for BlockGemmASmemBSmemCRegV1
+template <typename ADataType_,
+          typename BDataType_,
+          typename CDataType_,
+          index_t kBlockSize_,
+          typename BlockGemmShape_>
+struct BlockGemmASmemBSmemCRegProblem
+{
+    using ADataType      = remove_cvref_t<ADataType_>;
+    using BDataType      = remove_cvref_t<BDataType_>;
+    using CDataType      = remove_cvref_t<CDataType_>;
+    using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
+
+    static constexpr index_t kBlockSize = kBlockSize_;
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile/block_gemm_asmem_bsmem_creg_v1.hpp b/include/ck/tile_program/block_tile/block_gemm_asmem_bsmem_creg_v1.hpp
new file mode 100644
index 0000000000..166f887dba
--- /dev/null
+++ b/include/ck/tile_program/block_tile/block_gemm_asmem_bsmem_creg_v1.hpp
@@ -0,0 +1,349 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/static_tile_distribution_helper.hpp"
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/load_tile.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile/block_gemm_asmem_bsmem_creg_problem.hpp"
+#include "ck/tile_program/block_tile/block_gemm_asmem_bsmem_creg_v1_default_policy.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+// A is block window on shared memory
+// B is block window on shared memory
+// C is block distributed tensor
+template <typename Problem, typename Policy = BlockGemmASmemBSmemCRegV1DefaultPolicy>
+struct BlockGemmASmemBSmemCRegV1
+{
+    using ADataType      = remove_cvref_t<typename Problem::ADataType>;
+    using BDataType      = remove_cvref_t<typename Problem::BDataType>;
+    using CDataType      = remove_cvref_t<typename Problem::CDataType>;
+    using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    // C += A * B
+    template <typename CBlockTensor, typename ABlockWindowTmp, typename BBlockWindowTmp>
+    __device__ void operator()(CBlockTensor& c_block_tensor,
+                               const ABlockWindowTmp& a_block_window_tmp,
+                               const BBlockWindowTmp& b_block_window_tmp) const
+    {
+        static_assert(is_same_v<ADataType, typename ABlockWindowTmp::DataType> &&
+                          is_same_v<BDataType, typename BBlockWindowTmp::DataType> &&
+                          is_same_v<CDataType, typename CBlockTensor::DataType>,
+                      "wrong!");
+
+        constexpr index_t MPerBlock = ABlockWindowTmp{}.GetWindowLengths()[Number<0>{}];
+        constexpr index_t NPerBlock = BBlockWindowTmp{}.GetWindowLengths()[Number<0>{}];
+        constexpr index_t KPerBlock = ABlockWindowTmp{}.GetWindowLengths()[Number<1>{}];
+
+        static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
+                          KPerBlock == BlockGemmShape::kK,
+                      "wrong!");
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template At<0>())>;
+
+        constexpr index_t MWarp = config.template At<1>();
+        constexpr index_t NWarp = config.template At<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+        constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr index_t MPerBlockPerIter = MPerBlock / MIterPerWarp;
+        constexpr index_t NPerBlockPerIter = NPerBlock / NIterPerWarp;
+        constexpr index_t KPerBlockPerIter = KPerBlock / KIterPerWarp;
+
+        const index_t iMWarp = get_warp_id() / NWarp;
+        const index_t iNWarp = get_warp_id() % NWarp;
+
+        // construct A-warp-window
+        auto a_warp_window_tmp = make_tile_window(
+            a_block_window_tmp.GetBottomTensorView(),
+            make_tuple(Number<WG::kM>{}, Number<WG::kK>{}),
+            a_block_window_tmp.GetWindowOrigin() + MultiIndex<2>{iMWarp * WG::kM, 0},
+            make_static_tile_distribution(typename WG::AWarpDstrEncoding{}));
+
+#if 0 // FIXME: using Array will cause register spill
+        Array<Array<decltype(a_warp_window_tmp), KIterPerWarp>, MIterPerWarp> a_warp_windows{
+            {a_warp_window_tmp}};
+
+        for(index_t mIter = 0; mIter < MIterPerWarp; mIter++)
+        {
+            for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
+            {
+                move_tile_window(a_warp_windows(mIter)(kIter),
+                                 {mIter * MPerBlockPerIter, kIter * KPerBlockPerIter});
+            }
+        }
+#else
+        StaticallyIndexedArray<StaticallyIndexedArray<decltype(a_warp_window_tmp), KIterPerWarp>,
+                               MIterPerWarp>
+            a_warp_windows;
+
+        static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+            static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                a_warp_windows(mIter)(kIter) = a_warp_window_tmp;
+
+                move_tile_window(a_warp_windows(mIter)(kIter),
+                                 {mIter * MPerBlockPerIter, kIter * KPerBlockPerIter});
+            });
+        });
+#endif
+
+        // construct B-warp-window
+        auto b_warp_window_tmp = make_tile_window(
+            b_block_window_tmp.GetBottomTensorView(),
+            make_tuple(Number<WG::kN>{}, Number<WG::kK>{}),
+            b_block_window_tmp.GetWindowOrigin() + MultiIndex<2>{iNWarp * WG::kN, 0},
+            make_static_tile_distribution(typename WG::BWarpDstrEncoding{}));
+
+#if 0 // FIXME: using Array will cause register spill
+        Array<Array<decltype(b_warp_window_tmp), KIterPerWarp>, NIterPerWarp> b_warp_windows{
+            {b_warp_window_tmp}};
+
+        for(index_t nIter = 0; nIter < NIterPerWarp; nIter++)
+        {
+            for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
+            {
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            }
+        }
+#else
+        StaticallyIndexedArray<StaticallyIndexedArray<decltype(b_warp_window_tmp), KIterPerWarp>,
+                               NIterPerWarp>
+            b_warp_windows;
+
+        static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+            static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
+
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            });
+        });
+#endif
+
+        using CWarpDstr   = typename WG::CWarpDstr;
+        using CWarpTensor = typename WG::CWarpTensor;
+
+        constexpr auto c_warp_y_lengths = to_sequence(CWarpDstr{}.GetYs2DDescriptor().GetLengths());
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        // hot loop:
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                // read A warp tensor from A block window
+                const auto a_warp_tensor = load_tile(a_warp_windows(mIter)(kIter));
+
+                static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                    // read B warp tensor from B Block window
+                    const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
+
+                    // read C warp tensor from C block tensor
+                    CWarpTensor c_warp_tensor;
+
+                    c_warp_tensor.GetThreadBuffer() = c_block_tensor.GetYSlicedThreadData(
+                        merge_sequences(Sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(Sequence<1, 1>{}, c_warp_y_lengths));
+
+                    // warp GEMM
+                    WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
+
+                    // write C warp tensor into C block tensor
+                    c_block_tensor.SetYSlicedThreadData(
+                        merge_sequences(Sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(Sequence<1, 1>{}, c_warp_y_lengths),
+                        c_warp_tensor.GetThreadBuffer());
+                });
+            });
+        });
+    }
+
+    // C = A * B
+    template <typename ABlockWindowTmp, typename BBlockWindowTmp>
+    __device__ auto operator()(const ABlockWindowTmp& a_block_window_tmp,
+                               const BBlockWindowTmp& b_block_window_tmp) const
+    {
+        static_assert(is_same_v<ADataType, typename ABlockWindowTmp::DataType> &&
+                          is_same_v<BDataType, typename BBlockWindowTmp::DataType>,
+                      "wrong!");
+
+        constexpr index_t MPerBlock = ABlockWindowTmp{}.GetWindowLengths()[Number<0>{}];
+        constexpr index_t NPerBlock = BBlockWindowTmp{}.GetWindowLengths()[Number<0>{}];
+        constexpr index_t KPerBlock = ABlockWindowTmp{}.GetWindowLengths()[Number<1>{}];
+
+        static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
+                          KPerBlock == BlockGemmShape::kK,
+                      "wrong!");
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template At<0>())>;
+
+        constexpr index_t MWarp = config.template At<1>();
+        constexpr index_t NWarp = config.template At<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+        constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr index_t MPerBlockPerIter = MPerBlock / MIterPerWarp;
+        constexpr index_t NPerBlockPerIter = NPerBlock / NIterPerWarp;
+        constexpr index_t KPerBlockPerIter = KPerBlock / KIterPerWarp;
+
+        const index_t iMWarp = get_warp_id() / NWarp;
+        const index_t iNWarp = get_warp_id() % NWarp;
+
+        // construct A-warp-window
+        auto a_warp_window_tmp = make_tile_window(
+            a_block_window_tmp.GetBottomTensorView(),
+            make_tuple(Number<WG::kM>{}, Number<WG::kK>{}),
+            a_block_window_tmp.GetWindowOrigin() + MultiIndex<2>{iMWarp * WG::kM, 0},
+            make_static_tile_distribution(typename WG::AWarpDstrEncoding{}));
+
+#if 0 // FIXME: using Array will cause register spill
+        Array<Array<decltype(a_warp_window_tmp), KIterPerWarp>, MIterPerWarp> a_warp_windows{
+            {a_warp_window_tmp}};
+
+        for(index_t mIter = 0; mIter < MIterPerWarp; mIter++)
+        {
+            for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
+            {
+                move_tile_window(a_warp_windows(mIter)(kIter),
+                                 {mIter * MPerBlockPerIter, kIter * KPerBlockPerIter});
+            }
+        }
+#else
+        StaticallyIndexedArray<StaticallyIndexedArray<decltype(a_warp_window_tmp), KIterPerWarp>,
+                               MIterPerWarp>
+            a_warp_windows;
+
+        static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+            static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                a_warp_windows(mIter)(kIter) = a_warp_window_tmp;
+
+                move_tile_window(a_warp_windows(mIter)(kIter),
+                                 {mIter * MPerBlockPerIter, kIter * KPerBlockPerIter});
+            });
+        });
+#endif
+
+        // construct B-warp-window
+        auto b_warp_window_tmp = make_tile_window(
+            b_block_window_tmp.GetBottomTensorView(),
+            make_tuple(Number<WG::kN>{}, Number<WG::kK>{}),
+            b_block_window_tmp.GetWindowOrigin() + MultiIndex<2>{iNWarp * WG::kN, 0},
+            make_static_tile_distribution(typename WG::BWarpDstrEncoding{}));
+
+#if 0 // FIXME: using Array will cause register spill
+        Array<Array<decltype(b_warp_window_tmp), KIterPerWarp>, NIterPerWarp> b_warp_windows{
+            {b_warp_window_tmp}};
+
+        for(index_t nIter = 0; nIter < NIterPerWarp; nIter++)
+        {
+            for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
+            {
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            }
+        }
+#else
+        StaticallyIndexedArray<StaticallyIndexedArray<decltype(b_warp_window_tmp), KIterPerWarp>,
+                               NIterPerWarp>
+            b_warp_windows;
+
+        static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+            static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
+
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            });
+        });
+#endif
+
+        static_assert(is_same_v<CDataType, typename WG::CDataType>, "wrong!");
+
+        // Construct C-Block-Tensor
+        constexpr auto c_block_outer_dstr_encoding = StaticTileDistributionEncoding<
+            Sequence<>,
+            Tuple<Sequence<MIterPerWarp, MWarp>, Sequence<NIterPerWarp, NWarp>>,
+            Tuple<Sequence<1, 2>>,
+            Tuple<Sequence<1, 1>>,
+            Sequence<1, 2>,
+            Sequence<0, 0>>{};
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
+
+        constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+
+        auto c_block_tensor = make_static_distributed_tensor<CDataType>(c_block_dstr);
+
+        using CWarpDstr   = typename WG::CWarpDstr;
+        using CWarpTensor = typename WG::CWarpTensor;
+
+        constexpr auto c_warp_y_lengths = to_sequence(CWarpDstr{}.GetYs2DDescriptor().GetLengths());
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        // hot loop:
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                // read A warp tensor from A block window
+                const auto a_warp_tensor = load_tile(a_warp_windows(mIter)(kIter));
+
+                static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                    // read B warp tensor from B Block window
+                    const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
+
+                    // read C warp tensor from C block tensor
+                    CWarpTensor c_warp_tensor;
+
+                    // warp GEMM
+                    if constexpr(KIterPerWarp == 0)
+                    {
+                        // c = a * b
+                        c_warp_tensor = WG{}(a_warp_tensor, b_warp_tensor);
+                    }
+                    else
+                    {
+                        // c += a * b
+                        c_warp_tensor.GetThreadBuffer() = c_block_tensor.GetYSlicedThreadData(
+                            merge_sequences(Sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                            merge_sequences(Sequence<1, 1>{}, c_warp_y_lengths));
+
+                        WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
+                    }
+
+                    // write C warp tensor into C block tensor
+                    c_block_tensor.SetYSlicedThreadData(
+                        merge_sequences(Sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(Sequence<1, 1>{}, c_warp_y_lengths),
+                        c_warp_tensor.GetThreadBuffer());
+                });
+            });
+        });
+
+        return c_block_tensor;
+    }
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile/block_gemm_asmem_bsmem_creg_v1_default_policy.hpp b/include/ck/tile_program/block_tile/block_gemm_asmem_bsmem_creg_v1_default_policy.hpp
new file mode 100644
index 0000000000..2666c42bce
--- /dev/null
+++ b/include/ck/tile_program/block_tile/block_gemm_asmem_bsmem_creg_v1_default_policy.hpp
@@ -0,0 +1,57 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+// Default policy for BlockGemmASmemBSmemCRegV1
+// Default policy class should not be templated, put template on member functions instead
+struct BlockGemmASmemBSmemCRegV1DefaultPolicy
+{
+    template <typename Problem>
+    __host__ __device__ static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+        using namespace ck::tile_program::warp;
+
+#if 0
+        constexpr index_t kBlockSize = Problem::kBlockSize;
+
+        constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        static_assert(kBlockSize % get_warp_size() == 0, "wrong!");
+
+        constexpr index_t NumWarp = kBlockSize / get_warp_size();
+
+        if constexpr(NumWarp == 4 && kMPerBlock % 128 == 0 &&
+                     kNPerBlock % 128 == 0 % kKPerBlock % 16 == 0)
+        {
+            return make_tuple(WarpGemmMfmaF16F16F32M32N32K16{}, 2, 2);
+        }
+        else
+        {
+            return make_tuple(WarpGemmMfmaF16F16F32M32N32K16{}, 2, 2);
+        }
+#else
+        return make_tuple(WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution{}, 4, 1);
+#endif
+    }
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile/block_reduce.hpp b/include/ck/tile_program/block_tile/block_reduce.hpp
new file mode 100644
index 0000000000..1cba690c76
--- /dev/null
+++ b/include/ck/tile_program/block_tile/block_reduce.hpp
@@ -0,0 +1,215 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+
+#include "ck/tile_program/tile/static_distributed_tensor.hpp"
+#include "ck/tile_program/tile/static_tile_distribution_helper.hpp"
+#include "ck/tile_program/tile/distributed_tile_sweep.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+// synchronize reduce result (cross lane reduction and broadcast on replicated dimension)
+template <typename AccDistributedTensor_, typename ReduceFunc>
+__device__ void block_tile_reduce_sync(AccDistributedTensor_& acc_tensor,
+                                       const ReduceFunc& reduce_func)
+{
+    using Dstr             = typename AccDistributedTensor_::StaticTileDistribution;
+    using DstrEncode       = typename Dstr::DstrEncode;
+    using DstrEncodeDetail = typename DstrEncode::Detail;
+
+    constexpr index_t NDimP = Dstr::GetNumOfDimensionP();
+    constexpr index_t NDimR = Dstr::GetNumOfDimensionR();
+
+    constexpr index_t idim_p_lane = NDimP - 1;
+
+    const auto ps_idx = make_array<index_t>(get_block_id(), get_lane_id());
+    const auto rs_idx = acc_tensor.GetTileDistribution().CalculateRsIndexFromPsIndex(ps_idx);
+
+    constexpr index_t thread_buf_size = AccDistributedTensor_::GetThreadBufferSize();
+
+    // loop over thread data
+    static_for<0, thread_buf_size, 1>{}([&](auto i) {
+        auto v_local = acc_tensor.GetThreadBuffer()[i];
+
+        // cross-lane reduce for replication
+        // only reduce on R dimension correspond to lane
+        // (lane id maps to this R dimension)
+        static_for<0, NDimR, 1>{}([&](auto idim_r) {
+            // FIXME: nasty to use does_p_own_r_
+            if constexpr(DstrEncodeDetail::does_p_own_r_[idim_p_lane][idim_r])
+            {
+                constexpr index_t r_length = DstrEncode::rs_lengths_[idim_r];
+
+                constexpr index_t lid_over_rid_derivative =
+                    DstrEncodeDetail::ps_over_rs_derivative_[idim_p_lane][idim_r];
+
+                static_assert(math::is_power_of_two_integer(r_length),
+                              "wrong! only support power of 2 reduction");
+
+                constexpr index_t nstage = math::integer_log2_floor(r_length);
+
+                // reduction sweep forward
+                static_for<0, nstage, 1>{}([&](auto istage) {
+                    constexpr index_t lid_delta =
+                        lid_over_rid_derivative * (1 << (nstage - istage - 1));
+
+                    // pull data from remote lane
+                    const auto v_remote = warp_shuffle_down(v_local, lid_delta);
+
+                    // reduce
+                    v_local = reduce_func(v_local, v_remote);
+                });
+            }
+        });
+
+        // cross-lane broadcast for replication
+        // only broadcast on R dimension correspond to lane
+        // (lane id maps to this R dimension)
+        static_for<0, NDimR, 1>{}([&](auto idim_r) {
+            // FIXME: nasty to use does_p_own_r_
+            if constexpr(DstrEncodeDetail::does_p_own_r_[idim_p_lane][idim_r])
+            {
+                const index_t r_id = rs_idx[idim_r];
+
+                constexpr index_t r_length = DstrEncode::rs_lengths_[idim_r];
+
+                constexpr index_t lid_over_rid_derivative =
+                    DstrEncodeDetail::ps_over_rs_derivative_[NDimP - 1][idim_r];
+
+                static_assert(math::is_power_of_two_integer(r_length),
+                              "wrong! only support power of 2 reduction");
+
+                constexpr index_t nstage = math::integer_log2_floor(r_length);
+
+                // broadcast sweep backward
+                static_for<0, nstage, 1>{}([&](auto istage) {
+                    // do I hold reduced data?
+                    const bool do_i_hold_reduced_data = r_id < (1 << istage);
+
+                    constexpr index_t lid_delta = lid_over_rid_derivative * (1 << istage);
+
+                    // pull data from remote lane
+                    const auto v_remote = warp_shuffle_up(v_local, lid_delta);
+
+                    // decide whether to update local data with remote data
+                    v_local = do_i_hold_reduced_data ? v_local : v_remote;
+                });
+            }
+        });
+
+        acc_tensor.GetThreadBuffer()(i) = v_local;
+    });
+}
+
+// FIXME: this is for 2D to 1D reduce only, need to support n-D
+template <typename AccDistributedTensor_,
+          typename InDistributedTensor_,
+          index_t... InReduceDims,
+          typename ReduceFunc>
+__device__ void block_tile_reduce(AccDistributedTensor_& acc_tensor,
+                                  const InDistributedTensor_& in_tensor,
+                                  Sequence<InReduceDims...>,
+                                  const ReduceFunc& reduce_func)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+#if 0
+    constexpr auto in_reduce_dims = Sequence<InReduceDims...>{};
+
+    constexpr index_t ndim_in        = InDistributedTensor_::GetNumOfDimension();
+    constexpr index_t ndim_in_reduce = in_reduce_dims.Size();
+    constexpr index_t ndim_in_free   = ndim_in - ndim_in_reduce;
+
+    constexpr auto in_free_dims_arr = [&] {
+        Array<bool, ndim_free> is_free_dims{true};
+
+        for(index_t i = 0; i < ndim_reduce; i++)
+        {
+            is_free_dims(in_reduce_dims[i]) = false;
+        }
+
+        Array<index_t, ndim_free> in_free_dims{-1};
+
+        index_t cnt = 0;
+
+        for(index_t i = 0; i < ndim_in; i++)
+        {
+            if(is_free_dims[i])
+            {
+                in_free_dims(cnt) = i;
+
+                cnt++
+            }
+        }
+
+        return is_free_dims;
+    }();
+
+    constexpr auto in_free_dims = TO_SEQUENCE(is_free_dims_arr, ndim_in_free);
+#else
+
+    constexpr auto spans = InDistributedTensor_::GetDistributedSpans();
+
+    // in-thread reduction
+    // FIXME: hard coded to be 2D to 1D reduction
+    sweep_tile_span(spans[I0], [&](auto dstr_idx_i0) {
+        constexpr auto acc_dstr_idx = make_tuple(dstr_idx_i0);
+
+        auto acc = acc_tensor[acc_dstr_idx];
+
+        // FIXME
+        sweep_tile_span(spans[I1], [&](auto dstr_idx_i1) {
+            constexpr auto in_dstr_idx = make_tuple(dstr_idx_i0, dstr_idx_i1);
+
+            const auto in = in_tensor[in_dstr_idx];
+
+            acc = reduce_func(acc, in);
+        });
+
+        acc_tensor(acc_dstr_idx) = acc;
+    });
+#endif
+}
+
+template <typename AccDataType_,
+          typename InDistributedTensor_,
+          index_t... InReduceDims,
+          typename ReduceFunc,
+          typename InDataType_>
+__device__ auto block_tile_reduce(const InDistributedTensor_& in_tensor,
+                                  Sequence<InReduceDims...> in_reduce_dims,
+                                  const ReduceFunc& reduce_func,
+                                  const InDataType_& reduce_init)
+{
+    using InDataType  = typename InDistributedTensor_::DataType;
+    using AccDataType = remove_cvref_t<AccDataType_>;
+
+    static_assert(is_same_v<InDataType, remove_cvref_t<InDataType_>>, "wrong!");
+
+    // declare acc_tensor
+    constexpr auto acc_dstr = make_static_tile_distribution(
+        ck::tile_program::detail::make_reduce_tile_distribution_encoding(
+            InDistributedTensor_::GetTileDistribution().GetStaticTileDistributionEncoding(),
+            Sequence<InReduceDims...>{}));
+
+    auto acc_tensor = make_static_distributed_tensor<AccDataType>(acc_dstr);
+
+    // init acc_tensor
+    tile_elementwise_inout([&](auto& acc) { acc = type_convert<AccDataType>(reduce_init); },
+                           acc_tensor);
+
+    // warp reduce
+    block_tile_reduce(acc_tensor, in_tensor, in_reduce_dims, reduce_func);
+
+    return acc_tensor;
+}
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile_pipeline/block_fmha_pipeline_problem.hpp b/include/ck/tile_program/block_tile_pipeline/block_fmha_pipeline_problem.hpp
new file mode 100644
index 0000000000..7804ff2597
--- /dev/null
+++ b/include/ck/tile_program/block_tile_pipeline/block_fmha_pipeline_problem.hpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/type.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+template <typename QDataType_,
+          typename KDataType_,
+          typename VDataType_,
+          typename SaccDataType_,
+          typename SMPLComputeDataType_,
+          typename PDataType_,
+          typename OaccDataType_,
+          typename ODataType_,
+          index_t kBlockSize_,
+          typename BlockFmhaShape_>
+struct BlockFmhaPipelineProblem
+{
+    using QDataType           = remove_cvref_t<QDataType_>;
+    using KDataType           = remove_cvref_t<KDataType_>;
+    using VDataType           = remove_cvref_t<VDataType_>;
+    using SaccDataType        = remove_cvref_t<SaccDataType_>;
+    using SMPLComputeDataType = remove_cvref_t<SMPLComputeDataType_>;
+    using PDataType           = remove_cvref_t<PDataType_>;
+    using OaccDataType        = remove_cvref_t<OaccDataType_>;
+    using ODataType           = remove_cvref_t<ODataType_>;
+    using BlockFmhaShape      = remove_cvref_t<BlockFmhaShape_>;
+
+    static constexpr index_t kBlockSize = kBlockSize_;
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile_pipeline/block_fmha_pipeline_qkvs.hpp b/include/ck/tile_program/block_tile_pipeline/block_fmha_pipeline_qkvs.hpp
new file mode 100644
index 0000000000..33b0149fc0
--- /dev/null
+++ b/include/ck/tile_program/block_tile_pipeline/block_fmha_pipeline_qkvs.hpp
@@ -0,0 +1,341 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/load_tile.hpp"
+#include "ck/tile_program/tile/store_tile.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/tile/slice_tile.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_fmha_pipeline_qkvs_default_policy.hpp"
+#include "ck/tile_program/block_tile/block_reduce.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+// This pipeline is qkv all located in LDS
+template <typename Problem, typename Policy = BlockFmhaPipelineQKVSDefaultPolicy>
+struct BlockFmhaPipelineQKVS
+{
+    using QDataType           = remove_cvref_t<typename Problem::QDataType>;
+    using KDataType           = remove_cvref_t<typename Problem::KDataType>;
+    using VDataType           = remove_cvref_t<typename Problem::VDataType>;
+    using SaccDataType        = remove_cvref_t<typename Problem::SaccDataType>;
+    using SMPLComputeDataType = remove_cvref_t<typename Problem::SMPLComputeDataType>;
+    using PDataType           = remove_cvref_t<typename Problem::PDataType>;
+    using OaccDataType        = remove_cvref_t<typename Problem::OaccDataType>;
+    using ODataType           = remove_cvref_t<typename Problem::ODataType>;
+
+    using BlockFmhaShape = remove_cvref_t<typename Problem::BlockFmhaShape>;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    static constexpr index_t kM0 = BlockFmhaShape::kM0;
+    static constexpr index_t kN0 = BlockFmhaShape::kN0;
+    static constexpr index_t kK0 = BlockFmhaShape::kK0;
+    static constexpr index_t kN1 = BlockFmhaShape::kN1;
+    static constexpr index_t kK1 = BlockFmhaShape::kK1;
+
+    __host__ __device__ static constexpr ck::index_t GetSmemSize()
+    {
+        return Policy::template GetSmemSize<Problem>();
+    }
+
+    template <typename QDramBlockWindowTmp,
+              typename KDramBlockWindowTmp,
+              typename VDramBlockWindowTmp,
+              typename QElementFunction,
+              typename KElementFunction,
+              typename VElementFunction>
+    __host__ __device__ auto
+    operator()(const QDramBlockWindowTmp& q_dram_block_window_tmp, // M0*K0 tile
+               const QElementFunction& q_element_func,
+               const KDramBlockWindowTmp& k_dram_block_window_tmp, // N0*K0 tile
+               const KElementFunction& k_element_func,
+               const VDramBlockWindowTmp& v_dram_block_window_tmp, // N1*K1 tile
+               const VElementFunction& v_element_func,
+               index_t num_total_loop,
+               index_t num_sub_loop_qk,
+               void* smem_ptr) const
+    {
+        static_assert(
+            is_same_v<QDataType, remove_cvref_t<typename QDramBlockWindowTmp::DataType>> &&
+                is_same_v<KDataType, remove_cvref_t<typename KDramBlockWindowTmp::DataType>> &&
+                is_same_v<VDataType, remove_cvref_t<typename VDramBlockWindowTmp::DataType>>,
+            "wrong!");
+
+        static_assert(kM0 == QDramBlockWindowTmp{}.GetWindowLengths()[Number<0>{}] &&
+                          kK0 == QDramBlockWindowTmp{}.GetWindowLengths()[Number<1>{}] &&
+                          kN0 == KDramBlockWindowTmp{}.GetWindowLengths()[Number<0>{}] &&
+                          kK0 == KDramBlockWindowTmp{}.GetWindowLengths()[Number<1>{}] &&
+                          kN1 == VDramBlockWindowTmp{}.GetWindowLengths()[Number<0>{}] &&
+                          kK1 == VDramBlockWindowTmp{}.GetWindowLengths()[Number<1>{}],
+                      "wrong!");
+
+        // Q tile in LDS
+        auto q_lds = make_tensor_view<AddressSpaceEnum::Lds>(
+            reinterpret_cast<QDataType*>(smem_ptr),
+            Policy::template MakeQLdsBlockDescriptor<Problem>());
+        auto q_lds_window =
+            make_tile_window(q_lds, make_tuple(Number<kM0>{}, Number<kK0>{}), {0, 0});
+
+        // K tile in LDS
+        KDataType* k_lds_ptr = static_cast<KDataType*>(static_cast<void*>(
+            static_cast<char*>(smem_ptr) + Policy::template GetSmemSizeQ<Problem>()));
+        auto k_lds           = make_tensor_view<AddressSpaceEnum::Lds>(
+            k_lds_ptr, Policy::template MakeKLdsBlockDescriptor<Problem>());
+        auto k_lds_window =
+            make_tile_window(k_lds, make_tuple(Number<kN0>{}, Number<kK0>{}), {0, 0});
+
+        // V tile in LDS
+        auto v_lds = make_tensor_view<AddressSpaceEnum::Lds>(
+            reinterpret_cast<VDataType*>(smem_ptr),
+            Policy::template MakeVLdsBlockDescriptor<Problem>());
+        auto v_lds_window =
+            make_tile_window(v_lds, make_tuple(Number<kN1>{}, Number<kK1>{}), {0, 0});
+
+        // Block GEMM
+        constexpr auto gemm_0 = Policy::template GetQKBlockGemm<Problem>();
+        constexpr auto gemm_1 = Policy::template GetKVBlockGemm<Problem>();
+
+        auto s_acc = decltype(gemm_0(q_lds_window, k_lds_window)){};
+
+        // reduction function for softmax
+        const auto f_max = [](auto e0, auto e1) { return max(e0, e1); };
+        const auto f_sum = [](auto e0, auto e1) { return e0 + e1; };
+
+        // infer Sacc, S, P, M, L, Oacc type
+        using SBlockTileType =
+            decltype(tile_elementwise_in(type_convert<SMPLComputeDataType, SaccDataType>, s_acc));
+
+        using PBlockTileType =
+            decltype(tile_elementwise_in(type_convert<PDataType, SaccDataType>, s_acc));
+
+        using MLBlockTileType = decltype(block_tile_reduce<SMPLComputeDataType>(
+            SBlockTileType{}, Sequence<1>{}, f_max, SMPLComputeDataType{0}));
+
+        using OaccBlockTileType = decltype(gemm_1(
+            get_slice_tile(PBlockTileType{}, Sequence<0, 0>{}, Sequence<kM0, kK1>{}),
+            v_lds_window));
+
+        // init Oacc, M, L
+        auto o_acc = OaccBlockTileType{};
+        auto m     = MLBlockTileType{};
+        auto l     = MLBlockTileType{};
+
+        tile_elementwise_inout([](auto& e) { e = 0; }, o_acc);
+        tile_elementwise_inout([](auto& e) { e = NumericLimits<SMPLComputeDataType>::Lowest(); },
+                               m);
+        tile_elementwise_inout([](auto& e) { e = 0; }, l);
+
+        auto k_dram_block_window = k_dram_block_window_tmp;
+        auto v_dram_window =
+            make_tile_window(v_dram_block_window_tmp.GetBottomTensorView(),
+                             v_dram_block_window_tmp.GetWindowLengths(),
+                             v_dram_block_window_tmp.GetWindowOrigin(),
+                             Policy::template MakeVDramTileDistribution<Problem>());
+
+        index_t i_total_loops = 0;
+        do
+        {
+            // STAGE 1, QK gemm
+            auto q_dram_window = make_tile_window(
+                q_dram_block_window_tmp.GetBottomTensorView(),
+                q_dram_block_window_tmp.GetWindowLengths(),
+                q_dram_block_window_tmp.GetWindowOrigin(),
+                Policy::template MakeQDramTileDistribution<Problem>()); // Q DRAM tile window for
+                                                                        // load
+
+            auto k_dram_window = make_tile_window(
+                k_dram_block_window.GetBottomTensorView(),
+                k_dram_block_window.GetWindowLengths(),
+                k_dram_block_window.GetWindowOrigin(),
+                Policy::template MakeKDramTileDistribution<Problem>()); // K DRAM tile window for
+                                                                        // load
+
+            auto q_block_tile = load_tile(q_dram_window); // prefetch, global read 0
+            auto k_block_tile = load_tile(k_dram_window);
+            {
+                move_tile_window(q_dram_window, {0, kK0}); // move to 1
+                move_tile_window(k_dram_window, {0, kK0});
+
+                tile_elementwise_inout([](auto& c) { c = 0; }, s_acc); // Initialize C
+
+                store_tile(q_lds_window,
+                           tile_elementwise_in(q_element_func, q_block_tile)); // LDS write 0
+                q_block_tile = load_tile(q_dram_window);                       // global read 1
+                store_tile(k_lds_window,
+                           tile_elementwise_in(k_element_func, k_block_tile)); // LDS write 0
+                k_block_tile = load_tile(k_dram_window);                       // global read 1
+            }
+
+            index_t i_k0_loops = num_sub_loop_qk - 2;
+            do
+            {
+                block_sync_lds();
+                gemm_0(s_acc, q_lds_window, k_lds_window); // GEMM i
+                block_sync_lds();
+
+                move_tile_window(q_dram_window, {0, kK0}); // move to i + 2
+                move_tile_window(k_dram_window, {0, kK0});
+
+                store_tile(q_lds_window,
+                           tile_elementwise_in(q_element_func, q_block_tile)); // LDS write i + 1
+                q_block_tile = load_tile(q_dram_window);                       // global read i + 2
+                store_tile(k_lds_window,
+                           tile_elementwise_in(k_element_func, k_block_tile)); // LDS write i + 1
+                k_block_tile = load_tile(k_dram_window);                       // global read i + 2
+
+                i_k0_loops--;
+            } while(i_k0_loops > 0);
+
+            const auto v_prefetch = load_tile(v_dram_window); // prefetch load v tile
+
+            { // tail
+                block_sync_lds();
+                gemm_0(s_acc, q_lds_window, k_lds_window); // GEMM num_loop - 2
+                block_sync_lds();
+
+                store_tile(
+                    q_lds_window,
+                    tile_elementwise_in(q_element_func, q_block_tile)); // LDS write num_loop - 1
+                store_tile(k_lds_window, tile_elementwise_in(k_element_func, k_block_tile));
+                block_sync_lds();
+
+                gemm_0(s_acc, q_lds_window, k_lds_window); // GEMM num_loop - 1
+            }
+
+            // STAGE 2, scale softmax
+            const auto s =
+                tile_elementwise_in(type_convert<SMPLComputeDataType, SaccDataType>, s_acc); // S{j}
+            auto m_local = block_tile_reduce<SMPLComputeDataType>(
+                s,
+                Sequence<1>{},
+                f_max,
+                NumericLimits<SMPLComputeDataType>::Lowest()); // m_local = rowmax(S{j})
+            block_tile_reduce_sync(m_local, f_max);
+
+            const auto m_old = m; // m{j-1}
+            tile_elementwise_inout(
+                [](auto& e0, auto e1, auto e2) { e0 = max(e1, e2); }, m, m_old, m_local); // m{j}
+
+            auto p_compute = make_static_distributed_tensor<SMPLComputeDataType>(
+                s.GetTileDistribution()); // Pcompute{j}
+
+            constexpr auto p_spans = decltype(p_compute)::GetDistributedSpans();
+            sweep_tile_span(p_spans[Number<0>{}], [&](auto idx0) {
+                constexpr auto i_idx = make_tuple(idx0);
+                sweep_tile_span(p_spans[Number<1>{}], [&](auto idx1) {
+                    constexpr auto i_j_idx = make_tuple(idx0, idx1);
+                    p_compute(i_j_idx)     = math::exp(s[i_j_idx] - m[i_idx]);
+                });
+            });
+
+            auto rowsum_p = block_tile_reduce<SMPLComputeDataType>(
+                p_compute, Sequence<1>{}, f_sum, SMPLComputeDataType{0}); // rowsum(Pcompute{j})
+
+            block_tile_reduce_sync(rowsum_p, f_sum);
+            // l{j}, Oacc{j}
+            sweep_tile_span(p_spans[Number<0>{}], [&](auto idx0) {
+                constexpr auto i_idx = make_tuple(idx0);
+                const auto tmp       = math::exp(m_old[i_idx] - m[i_idx]);
+                l(i_idx)             = tmp * l[i_idx] + rowsum_p[i_idx];
+                sweep_tile_span(p_spans[Number<1>{}], [&](auto idx1) {
+                    constexpr auto i_j_idx = make_tuple(idx0, idx1);
+                    // FIXME: this use different equation from FA v2 paper,
+                    // but produce correc result.
+                    // Is the equation wrong?
+                    o_acc(i_j_idx) *= tmp;
+                });
+            });
+
+            block_sync_lds();
+            store_tile(v_lds_window,
+                       tile_elementwise_in(v_element_func, v_prefetch)); // store the prefetch
+            move_tile_window(v_dram_window, {0, kK1});
+
+            const auto p =
+                tile_elementwise_in(type_convert<PDataType, SMPLComputeDataType>, p_compute);
+
+            // STAGE 3, KV gemm
+            constexpr index_t k1_loops = kN0 / kK1;
+            if constexpr(k1_loops > 1)
+            {
+                static_for<0, k1_loops - 1, 1>{}([&](auto i_k1) {
+                    const auto v = load_tile(v_dram_window); // load next v
+                    block_sync_lds();
+                    gemm_1(o_acc,
+                           get_slice_tile(
+                               p, Sequence<0, i_k1 * kK1>{}, Sequence<kM0, (i_k1 + 1) * kK1>{}),
+                           v_lds_window);
+                    block_sync_lds();
+                    store_tile(v_lds_window,
+                               tile_elementwise_in(v_element_func, v)); // store next v
+                    move_tile_window(v_dram_window, {0, kK1});
+                });
+            }
+            // tail
+            {
+                block_sync_lds();
+                gemm_1(o_acc,
+                       get_slice_tile(p, Sequence<0, (k1_loops - 1) * kK1>{}, Sequence<kM0, kN0>{}),
+                       v_lds_window);
+                block_sync_lds();
+            }
+            // move K tile windows
+            move_tile_window(k_dram_block_window, {kN0, 0});
+
+            i_total_loops++;
+        } while(i_total_loops < num_total_loop);
+
+        // finally, O
+        constexpr auto o_spans = decltype(o_acc)::GetDistributedSpans();
+
+        sweep_tile_span(o_spans[Number<0>{}], [&](auto idx0) {
+            constexpr auto i_idx = make_tuple(idx0);
+            const auto tmp       = 1 / l[i_idx];
+            sweep_tile_span(o_spans[Number<1>{}], [&](auto idx1) {
+                constexpr auto i_j_idx = make_tuple(idx0, idx1);
+                o_acc(i_j_idx) *= tmp;
+            });
+        });
+
+        return o_acc;
+    }
+
+    template <typename QDramBlockWindowTmp,
+              typename KDramBlockWindowTmp,
+              typename VDramBlockWindowTmp>
+    __host__ __device__ auto
+    operator()(const QDramBlockWindowTmp& q_dram_block_window_tmp, // M0*K0 tile
+               const KDramBlockWindowTmp& k_dram_block_window_tmp, // N0*K0 tile
+               const VDramBlockWindowTmp& v_dram_block_window_tmp, // N1*K1 tile
+               index_t num_total_loop,
+               index_t num_sub_loop_qk,
+               void* smem_ptr) const
+    {
+        return operator()(
+            q_dram_block_window_tmp,
+            [](const QDataType& x) { return x; },
+            k_dram_block_window_tmp,
+            [](const KDataType& x) { return x; },
+            v_dram_block_window_tmp,
+            [](const VDataType& x) { return x; },
+            num_total_loop,
+            num_sub_loop_qk,
+            smem_ptr);
+    }
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile_pipeline/block_fmha_pipeline_qkvs_default_policy.hpp b/include/ck/tile_program/block_tile_pipeline/block_fmha_pipeline_qkvs_default_policy.hpp
new file mode 100644
index 0000000000..833e0787da
--- /dev/null
+++ b/include/ck/tile_program/block_tile_pipeline/block_fmha_pipeline_qkvs_default_policy.hpp
@@ -0,0 +1,264 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_problem.hpp"
+#include "ck/tile_program/block_tile/block_gemm_areg_bsmem_creg_v1.hpp"
+#include "ck/tile_program/block_tile/block_gemm_asmem_bsmem_creg_v1.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+// This pipeline is qkv all located in LDS
+struct BlockFmhaPipelineQKVSDefaultPolicy
+{
+    // 3d + padding
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeQLdsBlockDescriptor()
+    {
+        using namespace ck;
+
+        constexpr index_t kMPerBlock = Problem::BlockFmhaShape::kM0;
+        constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK0;
+
+        constexpr auto q_lds_block_desc_0 = make_naive_tensor_descriptor(
+            make_tuple(Number<kKPerBlock / 8>{}, Number<kMPerBlock>{}, Number<8>{}),
+            make_tuple(Number<(kMPerBlock + 1) * 8>{}, Number<8>{}, Number<1>{}),
+            Number<8>{},
+            Number<1>{});
+
+        constexpr auto q_lds_block_desc = transform_tensor_descriptor(
+            q_lds_block_desc_0,
+            make_tuple(make_pass_through_transform(kMPerBlock),
+                       make_merge_transform(make_tuple(kKPerBlock / 8, 8))),
+            make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return q_lds_block_desc;
+    }
+
+    // 3d + padding
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeKLdsBlockDescriptor()
+    {
+        using namespace ck;
+
+        constexpr index_t kNPerBlock = Problem::BlockFmhaShape::kN0;
+        constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK1;
+
+        constexpr auto k_lds_block_desc_0 = make_naive_tensor_descriptor(
+            make_tuple(Number<kKPerBlock / 8>{}, Number<kNPerBlock>{}, Number<8>{}),
+            make_tuple(Number<(kNPerBlock + 1) * 8>{}, Number<8>{}, Number<1>{}),
+            Number<8>{},
+            Number<1>{});
+
+        constexpr auto k_lds_block_desc = transform_tensor_descriptor(
+            k_lds_block_desc_0,
+            make_tuple(make_pass_through_transform(kNPerBlock),
+                       make_merge_transform(make_tuple(kKPerBlock / 8, 8))),
+            make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return k_lds_block_desc;
+    }
+
+    // 3d + padding
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeVLdsBlockDescriptor()
+    {
+        constexpr index_t kNPerBlock = Problem::BlockFmhaShape::kN1;
+        constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK1;
+        constexpr index_t kPad       = 1;
+        constexpr index_t kK1        = 8;
+
+        constexpr auto v_lds_block_desc_0 = make_naive_tensor_descriptor(
+            make_tuple(Number<kKPerBlock / kK1>{}, Number<kNPerBlock>{}, Number<kK1>{}),
+            make_tuple(Number<(kNPerBlock + kPad) * kK1>{}, Number<kK1>{}, Number<1>{}),
+            Number<kK1>{},
+            Number<1>{});
+
+        constexpr auto v_lds_block_desc = transform_tensor_descriptor(
+            v_lds_block_desc_0,
+            make_tuple(make_pass_through_transform(kNPerBlock),
+                       make_merge_transform(make_tuple(Number<kKPerBlock / kK1>{}, Number<kK1>{}))),
+            make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return v_lds_block_desc;
+    }
+
+    template <typename Problem>
+    __host__ __device__ static constexpr ck::index_t GetSmemSizeQ()
+    {
+        constexpr index_t lds_alignment = 16; // optional
+        constexpr index_t q_smem_size =
+            ck::math::integer_divide_ceil(
+                sizeof(typename Problem::QDataType) *
+                    MakeQLdsBlockDescriptor<Problem>().GetElementSpaceSize(),
+                lds_alignment) *
+            lds_alignment;
+        return q_smem_size;
+    }
+
+    template <typename Problem>
+    __host__ __device__ static constexpr ck::index_t GetSmemSize()
+    {
+        constexpr index_t smem_size_gemm_0 =
+            GetSmemSizeQ<Problem>() + sizeof(typename Problem::KDataType) *
+                                          MakeKLdsBlockDescriptor<Problem>().GetElementSpaceSize();
+        constexpr index_t smem_size_gemm_1 =
+            MakeVLdsBlockDescriptor<Problem>().GetElementSpaceSize() *
+            sizeof(typename Problem::VDataType);
+
+        // TODO: consider shuffle requirement
+        return math::max(smem_size_gemm_0, smem_size_gemm_1);
+    }
+
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeQDramTileDistribution()
+    {
+        using QDataType = remove_cvref_t<typename Problem::QDataType>;
+
+        constexpr index_t kBlockSize = Problem::kBlockSize;
+
+        constexpr index_t kMPerBlock = Problem::BlockFmhaShape::kM0;
+        constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK0;
+
+        constexpr index_t K1 = 16 / sizeof(QDataType); // use dwordx4. TODO: change this
+        constexpr index_t K0 = kKPerBlock / K1;
+        constexpr index_t M2 = get_warp_size() / K0;
+#if 1 // coalesce reading for each blocks
+        constexpr index_t M1 = kBlockSize / get_warp_size();
+        constexpr index_t M0 = kMPerBlock / (M2 * M1);
+
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<1>,
+                                           Tuple<Sequence<M0, M1, M2>, Sequence<K0, K1>>,
+                                           Tuple<Sequence<1>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1>, Sequence<2, 0>>,
+                                           Sequence<1, 2>,
+                                           Sequence<0, 1>>{});
+#else // coalesce reading for each warps
+        constexpr index_t M0 = kBlockSize / get_warp_size();
+        constexpr index_t M1 = kMPerBlock / (M2 * M0);
+
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<1>,
+                                           Tuple<Sequence<M0, M1, M2>, Sequence<K0, K1>>,
+                                           Tuple<Sequence<1>, Sequence<1, 2>>,
+                                           Tuple<Sequence<0>, Sequence<2, 0>>,
+                                           Sequence<1, 2>,
+                                           Sequence<1, 1>>{});
+#endif
+    }
+
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeKDramTileDistribution()
+    {
+        using KDataType = remove_cvref_t<typename Problem::KDataType>;
+
+        constexpr index_t kBlockSize = Problem::kBlockSize;
+
+        constexpr index_t kNPerBlock = Problem::BlockFmhaShape::kN0;
+        constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK0;
+
+        constexpr index_t K1 = 16 / sizeof(KDataType);
+        constexpr index_t K0 = kKPerBlock / K1;
+        constexpr index_t N2 = get_warp_size() / K0;
+#if 1 // coalesce reading for each blocks
+        constexpr index_t N1 = kBlockSize / get_warp_size();
+        constexpr index_t N0 = kNPerBlock / (N2 * N1);
+
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<1>,
+                                           Tuple<Sequence<N0, N1, N2>, Sequence<K0, K1>>,
+                                           Tuple<Sequence<1>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1>, Sequence<2, 0>>,
+                                           Sequence<1, 2>,
+                                           Sequence<0, 1>>{});
+#else // coalesce reading for each warps
+        constexpr index_t N0 = kBlockSize / get_warp_size();
+        constexpr index_t N1 = kNPerBlock / (N2 * N0);
+
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<1>,
+                                           Tuple<Sequence<N0, N1, N2>, Sequence<K0, K1>>,
+                                           Tuple<Sequence<1>, Sequence<1, 2>>,
+                                           Tuple<Sequence<0>, Sequence<2, 0>>,
+                                           Sequence<1, 2>,
+                                           Sequence<1, 1>>{});
+#endif
+    }
+
+    template <typename Problem>
+    __device__ static constexpr auto MakeVDramTileDistribution()
+    {
+        using VDataType = remove_cvref_t<typename Problem::VDataType>;
+        ;
+
+        constexpr index_t kBlockSize = Problem::kBlockSize;
+        constexpr index_t kNPerBlock = Problem::BlockFmhaShape::kN1;
+        constexpr index_t kKPerBlock = Problem::BlockFmhaShape::kK1;
+
+        constexpr index_t K1 = 16 / sizeof(VDataType);
+        constexpr index_t K0 = kKPerBlock / K1;
+        constexpr index_t N2 = get_warp_size() / K0;
+        constexpr index_t N1 = kBlockSize / get_warp_size();
+        constexpr index_t N0 = kNPerBlock / (N2 * N1);
+
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<1>,
+                                           Tuple<Sequence<N0, N1, N2>, Sequence<K0, K1>>,
+                                           Tuple<Sequence<1>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1>, Sequence<2, 0>>,
+                                           Sequence<1, 2>,
+                                           Sequence<0, 1>>{});
+    }
+
+    template <typename Problem>
+    __host__ __device__ static constexpr auto GetQKBlockGemm()
+    {
+        using BlockGemmProblem =
+            BlockGemmPipelineProblem<typename Problem::QDataType,
+                                     typename Problem::KDataType,
+                                     typename Problem::SaccDataType,
+                                     Problem::kBlockSize,
+                                     TileGemmShape<Problem::BlockFmhaShape::kM0,
+                                                   Problem::BlockFmhaShape::kN0,
+                                                   Problem::BlockFmhaShape::kK0>>;
+        using BlockGemmPolicy = BlockGemmASmemBSmemCRegV1DefaultPolicy;
+
+        return BlockGemmASmemBSmemCRegV1<BlockGemmProblem, BlockGemmPolicy>{};
+    }
+
+    template <typename Problem>
+    __host__ __device__ static constexpr auto GetKVBlockGemm()
+    {
+        using BlockGemmProblem =
+            BlockGemmPipelineProblem<typename Problem::PDataType,
+                                     typename Problem::VDataType,
+                                     typename Problem::OaccDataType,
+                                     Problem::kBlockSize,
+                                     TileGemmShape<Problem::BlockFmhaShape::kM0,
+                                                   Problem::BlockFmhaShape::kN1,
+                                                   Problem::BlockFmhaShape::kK1>>;
+        using BlockGemmPolicy = BlockGemmARegBSmemCRegV1DefaultPolicy;
+
+        return BlockGemmARegBSmemCRegV1<BlockGemmProblem, BlockGemmPolicy>{};
+    }
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v1.hpp b/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v1.hpp
new file mode 100644
index 0000000000..f1ef6ae367
--- /dev/null
+++ b/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v1.hpp
@@ -0,0 +1,214 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+//  A Tile Window: global memory
+//  B Tile Window: global memory
+//  C Distributed tensor: register
+template <typename Problem, typename Policy = BlockGemmPipelineAGmemBGmemCRegV1DefaultPolicy>
+struct BlockGemmPipelineAGmemBGmemCRegV1
+{
+    using ADataType      = remove_cvref_t<typename Problem::ADataType>;
+    using BDataType      = remove_cvref_t<typename Problem::BDataType>;
+    using CDataType      = remove_cvref_t<typename Problem::CDataType>;
+    using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    static constexpr index_t kMPerBlock = BlockGemmShape::kM;
+    static constexpr index_t kNPerBlock = BlockGemmShape::kN;
+    static constexpr index_t kKPerBlock = BlockGemmShape::kK;
+
+    __host__ __device__ static constexpr ck::index_t GetStaticLdsSize()
+    {
+        return ck::math::integer_divide_ceil(
+                   sizeof(ADataType) *
+                       Policy::template MakeALdsBlockDescriptor<Problem>().GetElementSpaceSize(),
+                   16) *
+                   16 +
+               sizeof(BDataType) *
+                   Policy::template MakeBLdsBlockDescriptor<Problem>().GetElementSpaceSize();
+    }
+
+    template <typename ADramBlockWindowTmp,
+              typename BDramBlockWindowTmp,
+              typename AElementFunction,
+              typename BElementFunction>
+    __host__ __device__ auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
+                                        const AElementFunction& a_element_func,
+                                        const BDramBlockWindowTmp& b_dram_block_window_tmp,
+                                        const BElementFunction& b_element_func,
+                                        index_t num_loop,
+                                        void* p_smem) const
+    {
+        static_assert(
+            is_same_v<ADataType, remove_cvref_t<typename ADramBlockWindowTmp::DataType>> &&
+                is_same_v<BDataType, remove_cvref_t<typename BDramBlockWindowTmp::DataType>>,
+            "wrong!");
+
+        static_assert(kMPerBlock == ADramBlockWindowTmp{}.GetWindowLengths()[Number<0>{}] &&
+                          kNPerBlock == BDramBlockWindowTmp{}.GetWindowLengths()[Number<0>{}] &&
+                          kKPerBlock == ADramBlockWindowTmp{}.GetWindowLengths()[Number<1>{}],
+                      "wrong!");
+
+        // A tile in LDS
+        ADataType* p_a_lds = static_cast<ADataType*>(p_smem);
+
+        constexpr auto a_lds_block_desc = Policy::template MakeALdsBlockDescriptor<Problem>();
+
+        auto a_lds_block = make_tensor_view<AddressSpaceEnum::Lds>(p_a_lds, a_lds_block_desc);
+
+        constexpr index_t a_lds_block_space_size_aligned =
+            math::integer_divide_ceil(sizeof(ADataType) * a_lds_block_desc.GetElementSpaceSize(),
+                                      16) *
+            16;
+
+        // B tile in LDS
+        BDataType* p_b_lds = static_cast<BDataType*>(
+            static_cast<void*>(static_cast<char*>(p_smem) + a_lds_block_space_size_aligned));
+
+        constexpr auto b_lds_block_desc = Policy::template MakeBLdsBlockDescriptor<Problem>();
+
+        auto b_lds_block = make_tensor_view<AddressSpaceEnum::Lds>(p_b_lds, b_lds_block_desc);
+
+        // A DRAM tile window for load
+        auto a_copy_dram_window =
+            make_tile_window(a_dram_block_window_tmp.GetBottomTensorView(),
+                             make_tuple(Number<kMPerBlock>{}, Number<kKPerBlock>{}),
+                             a_dram_block_window_tmp.GetWindowOrigin(),
+                             Policy::template MakeADramTileDistribution<Problem>());
+
+        // A LDS tile window for store
+        auto a_copy_lds_window =
+            make_tile_window(a_lds_block,
+                             make_tuple(Number<kMPerBlock>{}, Number<kKPerBlock>{}),
+                             {0, 0},
+                             a_copy_dram_window.GetTileDistribution());
+
+        // B DRAM tile window for load
+        auto b_copy_dram_window =
+            make_tile_window(b_dram_block_window_tmp.GetBottomTensorView(),
+                             make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}),
+                             b_dram_block_window_tmp.GetWindowOrigin(),
+                             Policy::template MakeBDramTileDistribution<Problem>());
+
+        // B LDS tile window for store
+        auto b_copy_lds_window =
+            make_tile_window(b_lds_block,
+                             make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}),
+                             {0, 0},
+                             b_copy_dram_window.GetTileDistribution());
+
+        // A LDS tile for block GEMM
+        auto a_lds_gemm_window = make_tile_window(
+            a_lds_block, make_tuple(Number<kMPerBlock>{}, Number<kKPerBlock>{}), {0, 0});
+
+        // B LDS tile for block GEMM
+        auto b_lds_gemm_window = make_tile_window(
+            b_lds_block, make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}), {0, 0});
+
+        // Block GEMM
+        constexpr auto block_gemm = Policy::template GetBlockGemm<Problem>();
+
+        // Acc register tile
+        auto c_block_tile = decltype(block_gemm(a_lds_gemm_window, b_lds_gemm_window)){};
+
+        // prefetch
+        // global read 0
+        auto a_block_tile = load_tile(a_copy_dram_window);
+        auto b_block_tile = load_tile(b_copy_dram_window);
+
+        {
+            // move to 1
+            move_tile_window(a_copy_dram_window, {0, kKPerBlock});
+            move_tile_window(b_copy_dram_window, {0, kKPerBlock});
+
+            // Initialize C
+            tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
+
+            // LDS write 0
+            const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile);
+            store_tile(a_copy_lds_window, a_block_tile_tmp);
+
+            // LDS write 0
+            const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
+            store_tile(b_copy_lds_window, b_block_tile_tmp);
+        }
+
+        index_t iCounter = num_loop - 1;
+
+        do
+        {
+            // global read i + 1
+            a_block_tile = load_tile(a_copy_dram_window);
+            b_block_tile = load_tile(b_copy_dram_window);
+
+            block_sync_lds();
+
+            // GEMM i
+            block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
+
+            block_sync_lds();
+
+            // move to i + 2
+            move_tile_window(a_copy_dram_window, {0, kKPerBlock});
+            move_tile_window(b_copy_dram_window, {0, kKPerBlock});
+
+            // LDS write i + 1
+            const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile);
+            store_tile(a_copy_lds_window, a_block_tile_tmp);
+
+            // LDS write i + 1
+            const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
+            store_tile(b_copy_lds_window, b_block_tile_tmp);
+
+            iCounter--;
+
+        } while(iCounter > 0);
+
+        // tail
+        {
+            block_sync_lds();
+
+            // GEMM num_loop - 1
+            block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
+        }
+
+        return c_block_tile;
+    }
+
+    template <typename ADramBlockWindowTmp, typename BDramBlockWindowTmp>
+    __device__ auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
+                               const BDramBlockWindowTmp& b_dram_block_window_tmp,
+                               index_t num_loop,
+                               void* p_smem) const
+    {
+        return operator()(
+            a_dram_block_window_tmp,
+            [](const ADataType& a) { return a; },
+            b_dram_block_window_tmp,
+            [](const BDataType& b) { return b; },
+            num_loop,
+            p_smem);
+    }
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp b/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp
new file mode 100644
index 0000000000..2b9f7de1fe
--- /dev/null
+++ b/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp
@@ -0,0 +1,264 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile/block_gemm_asmem_bsmem_creg_v1.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+// Default policy for BlockGemmPipelineAGmemBGmemCRegV1
+// Default policy class should not be templated, put template on member functions instead
+struct BlockGemmPipelineAGmemBGmemCRegV1DefaultPolicy
+{
+#if 0
+    // 2d
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeALdsBlockDescriptor()
+    {
+        using namespace ck;
+
+        constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr auto a_lds_block_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(kMPerBlock, kKPerBlock), Number<32>{});
+
+        return a_lds_block_desc;
+    }
+
+    // 2d
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeBLdsBlockDescriptor()
+    {
+        using namespace ck;
+
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr auto b_lds_block_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(kNPerBlock, kKPerBlock), Number<32>{});
+
+        return b_lds_block_desc;
+    }
+#elif 1
+    // 3d + padding
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeALdsBlockDescriptor()
+    {
+        using namespace ck;
+
+        constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr auto a_lds_block_desc_0 = make_naive_tensor_descriptor(
+            make_tuple(Number<kKPerBlock / 8>{}, Number<kMPerBlock>{}, Number<8>{}),
+            make_tuple(Number<(kMPerBlock + 1) * 8>{}, Number<8>{}, Number<1>{}),
+            Number<8>{},
+            Number<1>{});
+
+        constexpr auto a_lds_block_desc = transform_tensor_descriptor(
+            a_lds_block_desc_0,
+            make_tuple(make_pass_through_transform(kMPerBlock),
+                       make_merge_transform(make_tuple(kKPerBlock / 8, 8))),
+            make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return a_lds_block_desc;
+    }
+
+    // 3d + padding
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeBLdsBlockDescriptor()
+    {
+        using namespace ck;
+
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr auto b_lds_block_desc_0 = make_naive_tensor_descriptor(
+            make_tuple(Number<kKPerBlock / 8>{}, Number<kNPerBlock>{}, Number<8>{}),
+            make_tuple(Number<(kNPerBlock + 1) * 8>{}, Number<8>{}, Number<1>{}),
+            Number<8>{},
+            Number<1>{});
+
+        constexpr auto b_lds_block_desc = transform_tensor_descriptor(
+            b_lds_block_desc_0,
+            make_tuple(make_pass_through_transform(kNPerBlock),
+                       make_merge_transform(make_tuple(kKPerBlock / 8, 8))),
+            make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return b_lds_block_desc;
+    }
+#elif 1
+    // fake XOR
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeALdsBlockDescriptor()
+    {
+        using namespace ck;
+
+        using ADataType = remove_cvref_t<typename Problem::ADataType>;
+
+        constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr auto a_lds_block_desc_d1_d2_d3 = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<kMPerBlock / 2>{}, Number<2>{}, Number<kKPerBlock>{}),
+            Number<kKPerBlock>{});
+
+        constexpr index_t kK1 = 16 / sizeof(ADataType);
+
+        constexpr auto a_lds_block_desc_d4_d5_d6 = transform_tensor_descriptor(
+            a_lds_block_desc_d1_d2_d3,
+            make_tuple(
+                make_xor_transform(make_tuple(Number<kMPerBlock / 2>{}, Number<kKPerBlock>{}), kK1),
+                make_pass_through_transform(2)),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+        constexpr auto a_lds_block_desc_m_k = transform_tensor_descriptor(
+            a_lds_block_desc_d4_d5_d6,
+            make_tuple(make_merge_transform(make_tuple(Number<kMPerBlock / 2>{}, Number<2>{})),
+                       make_pass_through_transform(kKPerBlock)),
+            make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return a_lds_block_desc_m_k;
+    }
+
+    // fake XOR
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeBLdsBlockDescriptor()
+    {
+        using namespace ck;
+
+        using BDataType = remove_cvref_t<typename Problem::BDataType>;
+
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr auto b_lds_block_desc_d1_d2_d3 = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<kNPerBlock / 2>{}, Number<2>{}, Number<kKPerBlock>{}),
+            Number<kKPerBlock>{});
+
+        constexpr index_t kK1 = 16 / sizeof(BDataType);
+
+        constexpr auto b_lds_block_desc_d4_d5_d6 = transform_tensor_descriptor(
+            b_lds_block_desc_d1_d2_d3,
+            make_tuple(
+                make_xor_transform(make_tuple(Number<kNPerBlock / 2>{}, Number<kKPerBlock>{}), kK1),
+                make_pass_through_transform(2)),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+        constexpr auto b_lds_block_desc_n_k = transform_tensor_descriptor(
+            b_lds_block_desc_d4_d5_d6,
+            make_tuple(make_merge_transform(make_tuple(Number<kNPerBlock / 2>{}, Number<2>{})),
+                       make_pass_through_transform(kKPerBlock)),
+            make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return b_lds_block_desc_n_k;
+    }
+#endif
+
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeADramTileDistribution()
+    {
+        using ADataType = remove_cvref_t<typename Problem::ADataType>;
+
+        constexpr index_t kBlockSize = Problem::kBlockSize;
+
+        constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr index_t K1 = 16 / sizeof(ADataType);
+        constexpr index_t K0 = kKPerBlock / K1;
+        constexpr index_t M2 = get_warp_size() / K0;
+#if 1 // coalesce reading for each blocks
+        constexpr index_t M1 = kBlockSize / get_warp_size();
+        constexpr index_t M0 = kMPerBlock / (M2 * M1);
+
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<1>,
+                                           Tuple<Sequence<M0, M1, M2>, Sequence<K0, K1>>,
+                                           Tuple<Sequence<1>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1>, Sequence<2, 0>>,
+                                           Sequence<1, 2>,
+                                           Sequence<0, 1>>{});
+#else // coalesce reading for each warps
+        constexpr index_t M0 = kBlockSize / get_warp_size();
+        constexpr index_t M1 = kMPerBlock / (M2 * M0);
+
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<1>,
+                                           Tuple<Sequence<M0, M1, M2>, Sequence<K0, K1>>,
+                                           Tuple<Sequence<1>, Sequence<1, 2>>,
+                                           Tuple<Sequence<0>, Sequence<2, 0>>,
+                                           Sequence<1, 2>,
+                                           Sequence<1, 1>>{});
+#endif
+    }
+
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeBDramTileDistribution()
+    {
+        using BDataType = remove_cvref_t<typename Problem::BDataType>;
+
+        constexpr index_t kBlockSize = Problem::kBlockSize;
+
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr index_t K1 = 16 / sizeof(BDataType);
+        constexpr index_t K0 = kKPerBlock / K1;
+        constexpr index_t N2 = get_warp_size() / K0;
+#if 1 // coalesce reading for each blocks
+        constexpr index_t N1 = kBlockSize / get_warp_size();
+        constexpr index_t N0 = kNPerBlock / (N2 * N1);
+
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<1>,
+                                           Tuple<Sequence<N0, N1, N2>, Sequence<K0, K1>>,
+                                           Tuple<Sequence<1>, Sequence<1, 2>>,
+                                           Tuple<Sequence<1>, Sequence<2, 0>>,
+                                           Sequence<1, 2>,
+                                           Sequence<0, 1>>{});
+#else // coalesce reading for each warps
+        constexpr index_t N0 = kBlockSize / get_warp_size();
+        constexpr index_t N1 = kNPerBlock / (N2 * N0);
+
+        return make_static_tile_distribution(
+            StaticTileDistributionEncoding<Sequence<1>,
+                                           Tuple<Sequence<N0, N1, N2>, Sequence<K0, K1>>,
+                                           Tuple<Sequence<1>, Sequence<1, 2>>,
+                                           Tuple<Sequence<0>, Sequence<2, 0>>,
+                                           Sequence<1, 2>,
+                                           Sequence<1, 1>>{});
+#endif
+    }
+
+    template <typename Problem>
+    __host__ __device__ static constexpr auto GetBlockGemm()
+    {
+        using BlockGemmPolicy = BlockGemmASmemBSmemCRegV1DefaultPolicy;
+
+        return BlockGemmASmemBSmemCRegV1<Problem, BlockGemmPolicy>{};
+    }
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2.hpp b/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2.hpp
new file mode 100644
index 0000000000..91342a3895
--- /dev/null
+++ b/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2.hpp
@@ -0,0 +1,234 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/load_tile.hpp"
+#include "ck/tile_program/tile/store_tile.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2_default_policy.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+//  A Tile Window: global memory
+//  B Tile Window: global memory
+//  C Distributed tensor: register
+template <typename Problem, typename Policy = BlockGemmPipelineAGmemBGmemCRegV2DefaultPolicy>
+struct BlockGemmPipelineAGmemBGmemCRegV2
+{
+    using ADataType      = remove_cvref_t<typename Problem::ADataType>;
+    using BDataType      = remove_cvref_t<typename Problem::BDataType>;
+    using CDataType      = remove_cvref_t<typename Problem::CDataType>;
+    using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    static constexpr index_t kMPerBlock = BlockGemmShape::kM;
+    static constexpr index_t kNPerBlock = BlockGemmShape::kN;
+    static constexpr index_t kKPerBlock = BlockGemmShape::kK;
+
+    __host__ __device__ static constexpr ck::index_t GetStaticLdsSize()
+    {
+        return ck::math::integer_divide_ceil(
+                   sizeof(ADataType) *
+                       Policy::template MakeALdsBlockDescriptor<Problem>().GetElementSpaceSize(),
+                   16) *
+                   16 +
+               sizeof(BDataType) *
+                   Policy::template MakeBLdsBlockDescriptor<Problem>().GetElementSpaceSize();
+    }
+
+    template <typename ADramBlockWindowTmp,
+              typename BDramBlockWindowTmp,
+              typename AElementFunction,
+              typename BElementFunction>
+    __host__ __device__ auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
+                                        const AElementFunction& a_element_func,
+                                        const BDramBlockWindowTmp& b_dram_block_window_tmp,
+                                        const BElementFunction& b_element_func,
+                                        index_t num_loop,
+                                        void* p_smem) const
+    {
+        static_assert(
+            is_same_v<ADataType, remove_cvref_t<typename ADramBlockWindowTmp::DataType>> &&
+                is_same_v<BDataType, remove_cvref_t<typename BDramBlockWindowTmp::DataType>>,
+            "wrong!");
+
+        static_assert(kMPerBlock == ADramBlockWindowTmp{}.GetWindowLengths()[Number<0>{}] &&
+                          kNPerBlock == BDramBlockWindowTmp{}.GetWindowLengths()[Number<0>{}] &&
+                          kKPerBlock == ADramBlockWindowTmp{}.GetWindowLengths()[Number<1>{}],
+                      "wrong!");
+
+        // A tile in LDS
+        ADataType* p_a_lds = static_cast<ADataType*>(p_smem);
+
+        constexpr auto a_lds_block_desc = Policy::template MakeALdsBlockDescriptor<Problem>();
+
+        auto a_lds_block = make_tensor_view<AddressSpaceEnum::Lds>(p_a_lds, a_lds_block_desc);
+
+        constexpr index_t a_lds_block_space_size_aligned =
+            math::integer_divide_ceil(sizeof(ADataType) * a_lds_block_desc.GetElementSpaceSize(),
+                                      16) *
+            16;
+
+        // B tile in LDS
+        BDataType* p_b_lds = static_cast<BDataType*>(
+            static_cast<void*>(static_cast<char*>(p_smem) + a_lds_block_space_size_aligned));
+
+        constexpr auto b_lds_block_desc = Policy::template MakeBLdsBlockDescriptor<Problem>();
+
+        auto b_lds_block = make_tensor_view<AddressSpaceEnum::Lds>(p_b_lds, b_lds_block_desc);
+
+        // A DRAM tile window for load
+        auto a_copy_dram_window =
+            make_tile_window(a_dram_block_window_tmp.GetBottomTensorView(),
+                             make_tuple(Number<kMPerBlock>{}, Number<kKPerBlock>{}),
+                             a_dram_block_window_tmp.GetWindowOrigin(),
+                             Policy::template MakeADramTileDistribution<Problem>());
+
+        // A LDS tile window for store
+        auto a_copy_lds_window =
+            make_tile_window(a_lds_block,
+                             make_tuple(Number<kMPerBlock>{}, Number<kKPerBlock>{}),
+                             {0, 0},
+                             a_copy_dram_window.GetTileDistribution());
+
+        // B DRAM tile window for load
+        auto b_copy_dram_window =
+            make_tile_window(b_dram_block_window_tmp.GetBottomTensorView(),
+                             make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}),
+                             b_dram_block_window_tmp.GetWindowOrigin(),
+                             Policy::template MakeBDramTileDistribution<Problem>());
+
+        // B LDS tile window for store
+        auto b_copy_lds_window =
+            make_tile_window(b_lds_block,
+                             make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}),
+                             {0, 0},
+                             b_copy_dram_window.GetTileDistribution());
+
+        // A LDS tile for block GEMM
+        auto a_lds_gemm_window = make_tile_window(
+            a_lds_block, make_tuple(Number<kMPerBlock>{}, Number<kKPerBlock>{}), {0, 0});
+
+        // B LDS tile for block GEMM
+        auto b_lds_gemm_window = make_tile_window(
+            b_lds_block, make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}), {0, 0});
+
+        // Block GEMM
+        constexpr auto block_gemm = Policy::template GetBlockGemm<Problem>();
+
+        // Acc register tile
+        auto c_block_tile = decltype(block_gemm(a_lds_gemm_window, b_lds_gemm_window)){};
+
+        // prefetch
+        // global read 0
+        auto a_block_tile = load_tile(a_copy_dram_window);
+        auto b_block_tile = load_tile(b_copy_dram_window);
+
+        {
+            // move to 1
+            move_tile_window(a_copy_dram_window, {0, kKPerBlock});
+            move_tile_window(b_copy_dram_window, {0, kKPerBlock});
+
+            // Initialize C
+            tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
+
+            // LDS write 0
+            const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile);
+            store_tile(a_copy_lds_window, a_block_tile_tmp);
+            // global read 1
+            a_block_tile = load_tile(a_copy_dram_window);
+
+            // LDS write 0
+            const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
+            store_tile(b_copy_lds_window, b_block_tile_tmp);
+            // global read 1
+            b_block_tile = load_tile(b_copy_dram_window);
+        }
+
+        index_t iCounter = num_loop - 2;
+
+        do
+        {
+            block_sync_lds();
+
+            // GEMM i
+            block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
+
+            block_sync_lds();
+
+            // move to i + 2
+            move_tile_window(a_copy_dram_window, {0, kKPerBlock});
+            move_tile_window(b_copy_dram_window, {0, kKPerBlock});
+
+            // LDS write i + 1
+            const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile);
+            store_tile(a_copy_lds_window, a_block_tile_tmp);
+            // global read i + 2
+            a_block_tile = load_tile(a_copy_dram_window);
+
+            // LDS write i + 1
+            const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
+            store_tile(b_copy_lds_window, b_block_tile_tmp);
+            // global read i + 2
+            b_block_tile = load_tile(b_copy_dram_window);
+
+            iCounter--;
+
+        } while(iCounter > 0);
+
+        // tail
+        {
+            block_sync_lds();
+
+            // GEMM num_loop - 2
+            block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
+
+            block_sync_lds();
+
+            // LDS write num_loop - 1
+            const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile);
+            store_tile(a_copy_lds_window, a_block_tile_tmp);
+
+            const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
+            store_tile(b_copy_lds_window, b_block_tile_tmp);
+
+            block_sync_lds();
+
+            // GEMM num_loop - 1
+            block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
+        }
+
+        return c_block_tile;
+    }
+
+    template <typename ADramBlockWindowTmp, typename BDramBlockWindowTmp>
+    __device__ auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
+                               const BDramBlockWindowTmp& b_dram_block_window_tmp,
+                               index_t num_loop,
+                               void* p_smem) const
+    {
+        return operator()(
+            a_dram_block_window_tmp,
+            [](const ADataType& a) { return a; },
+            b_dram_block_window_tmp,
+            [](const BDataType& b) { return b; },
+            num_loop,
+            p_smem);
+    }
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2_askiplds.hpp b/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2_askiplds.hpp
new file mode 100644
index 0000000000..a096bfc2cb
--- /dev/null
+++ b/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2_askiplds.hpp
@@ -0,0 +1,548 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/load_tile.hpp"
+#include "ck/tile_program/tile/store_tile.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2_askiplds_policy.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+//  A Tile Window: global memory
+//  B Tile Window: global memory
+//  C Distributed tensor: register
+template <typename Problem>
+struct BlockGemmPipelineAGmemBGmemCRegV2<Problem, BlockGemmPipelineAGmemBGmemCRegV2SkipALdsPolicy>
+{
+    using ADataType      = remove_cvref_t<typename Problem::ADataType>;
+    using BDataType      = remove_cvref_t<typename Problem::BDataType>;
+    using CDataType      = remove_cvref_t<typename Problem::CDataType>;
+    using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
+    using Policy         = BlockGemmPipelineAGmemBGmemCRegV2SkipALdsPolicy;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    static constexpr index_t kMPerBlock = BlockGemmShape::kM;
+    static constexpr index_t kNPerBlock = BlockGemmShape::kN;
+    static constexpr index_t kKPerBlock = BlockGemmShape::kK;
+
+    // Move this part into Policy?
+    __host__ __device__ static constexpr ck::index_t GetStaticLdsSize()
+    {
+        return sizeof(BDataType) *
+               Policy::template MakeBLdsBlockDescriptor<Problem>().GetElementSpaceSize();
+    }
+
+    template <typename ADramBlockWindowTmp,
+              typename BDramBlockWindowTmp,
+              typename AElementFunction,
+              typename BElementFunction>
+    __host__ __device__ auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
+                                        const AElementFunction& a_element_func,
+                                        const BDramBlockWindowTmp& b_dram_block_window_tmp,
+                                        const BElementFunction& b_element_func,
+                                        index_t num_loop,
+                                        void* p_smem) const
+    {
+        static_assert(
+            is_same_v<ADataType, remove_cvref_t<typename ADramBlockWindowTmp::DataType>> &&
+                is_same_v<BDataType, remove_cvref_t<typename BDramBlockWindowTmp::DataType>>,
+            "wrong!");
+
+        static_assert(kMPerBlock == ADramBlockWindowTmp{}.GetWindowLengths()[Number<0>{}] &&
+                          kNPerBlock == BDramBlockWindowTmp{}.GetWindowLengths()[Number<0>{}] &&
+                          kKPerBlock == ADramBlockWindowTmp{}.GetWindowLengths()[Number<1>{}],
+                      "wrong!");
+
+        // A tile in Reg，blockTensor
+        // This tensor distribution used to construct both distributed tensor for local buffer store
+        // and read. without buffer address info
+        constexpr auto a_reg_block_dstr = Policy::template MakeARegBlockDescriptor<Problem>();
+
+        // B tile in LDS, blockWindow
+        BDataType* p_b_lds =
+            static_cast<BDataType*>(static_cast<void*>(static_cast<char*>(p_smem)));
+
+        constexpr auto b_lds_block_desc = Policy::template MakeBLdsBlockDescriptor<Problem>();
+
+        // This tensor view used to construct both tile window for lds store and read, with buffer
+        // address info
+        auto b_lds_block = make_tensor_view<AddressSpaceEnum::Lds>(p_b_lds, b_lds_block_desc);
+
+        // A DRAM tile window for load
+        auto a_copy_dram_window =
+            make_tile_window(a_dram_block_window_tmp.GetBottomTensorView(),
+                             make_tuple(Number<kMPerBlock>{}, Number<kKPerBlock>{}),
+                             a_dram_block_window_tmp.GetWindowOrigin(),
+                             Policy::template MakeADramTileDistribution<Problem>());
+
+        // A Reg tensor for store, also used for block GEMM
+        auto a_copy_reg_tensor = make_static_distributed_tensor<ADataType>(a_reg_block_dstr);
+
+        // B DRAM tile window for load
+        auto b_copy_dram_window =
+            make_tile_window(b_dram_block_window_tmp.GetBottomTensorView(),
+                             make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}),
+                             b_dram_block_window_tmp.GetWindowOrigin(),
+                             Policy::template MakeBDramTileDistribution<Problem>());
+
+        // B LDS tile window for store
+        auto b_copy_lds_window =
+            make_tile_window(b_lds_block,
+                             make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}),
+                             {0, 0},
+                             b_copy_dram_window.GetTileDistribution());
+
+        // B LDS tile for block GEMM
+        auto b_lds_gemm_window = make_tile_window(
+            b_lds_block, make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}), {0, 0});
+
+        // Block GEMM
+        constexpr auto block_gemm = Policy::template GetBlockGemm<Problem>();
+
+        // Acc register tile
+        auto c_block_tile = decltype(block_gemm(a_copy_reg_tensor, b_lds_gemm_window)){};
+
+        // prefetch
+        // global read 0
+        auto a_block_tile = load_tile(a_copy_dram_window);
+        auto b_block_tile = load_tile(b_copy_dram_window);
+
+        {
+            // move to 1
+            move_tile_window(a_copy_dram_window, {0, kKPerBlock});
+            move_tile_window(b_copy_dram_window, {0, kKPerBlock});
+
+            // Initialize C
+            tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
+
+            // block buffer write 0
+            const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile);
+            // store_tile -> shuffle store tile
+            store_tile(a_copy_reg_tensor, a_block_tile_tmp);
+            // global read 1
+            a_block_tile = load_tile(a_copy_dram_window);
+
+            // LDS write 0
+            const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
+            store_tile(b_copy_lds_window, b_block_tile_tmp);
+            // global read 1
+            b_block_tile = load_tile(b_copy_dram_window);
+        }
+
+        index_t iCounter = num_loop - 2;
+
+        do
+        {
+            block_sync_lds();
+
+            // GEMM i
+            block_gemm(c_block_tile, a_copy_reg_tensor, b_lds_gemm_window);
+
+            block_sync_lds();
+
+            // move to i + 2
+            move_tile_window(a_copy_dram_window, {0, kKPerBlock});
+            move_tile_window(b_copy_dram_window, {0, kKPerBlock});
+
+            // LDS write i + 1
+            const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile);
+            store_tile(a_copy_reg_tensor, a_block_tile_tmp);
+            // global read i + 2
+            a_block_tile = load_tile(a_copy_dram_window);
+
+            // LDS write i + 1
+            const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
+            store_tile(b_copy_lds_window, b_block_tile_tmp);
+            // global read i + 2
+            b_block_tile = load_tile(b_copy_dram_window);
+
+            iCounter--;
+
+        } while(iCounter > 0);
+
+        // tail
+        {
+            block_sync_lds();
+
+            // GEMM num_loop - 2
+            block_gemm(c_block_tile, a_copy_reg_tensor, b_lds_gemm_window);
+
+            block_sync_lds();
+
+            // LDS write num_loop - 1
+            const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile);
+            store_tile(a_copy_reg_tensor, a_block_tile_tmp);
+
+            const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
+            store_tile(b_copy_lds_window, b_block_tile_tmp);
+
+            block_sync_lds();
+
+            // GEMM num_loop - 1
+            block_gemm(c_block_tile, a_copy_reg_tensor, b_lds_gemm_window);
+        }
+
+        return c_block_tile;
+    }
+
+    template <typename ADramBlockWindowTmp, typename BDramBlockWindowTmp>
+    __device__ auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
+                               const BDramBlockWindowTmp& b_dram_block_window_tmp,
+                               index_t num_loop,
+                               void* p_smem) const
+    {
+        return operator()(
+            a_dram_block_window_tmp,
+            [](const ADataType& a) { return a; },
+            b_dram_block_window_tmp,
+            [](const BDataType& b) { return b; },
+            num_loop,
+            p_smem);
+    }
+};
+
+//  A Tile Window: global memory
+//  B Tile Window: global memory
+//  C Distributed tensor: register
+template <typename Problem, index_t kHeadDim>
+struct BlockGemmPipelineAGmemBGmemCRegV2<
+    Problem,
+    BlockGemmPipelineAGmemBGmemCRegV2SkipALdsPersistentQRegCachePolicy<kHeadDim>>
+{
+    using ADataType      = remove_cvref_t<typename Problem::ADataType>;
+    using BDataType      = remove_cvref_t<typename Problem::BDataType>;
+    using CDataType      = remove_cvref_t<typename Problem::CDataType>;
+    using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
+    using Policy = BlockGemmPipelineAGmemBGmemCRegV2SkipALdsPersistentQRegCachePolicy<kHeadDim>;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    static constexpr index_t kMPerBlock = BlockGemmShape::kM;
+    static constexpr index_t kNPerBlock = BlockGemmShape::kN;
+    static constexpr index_t kKPerBlock = BlockGemmShape::kK;
+
+    static constexpr index_t k_loops = Policy::AKDim / kKPerBlock;
+
+    // Move this part into Policy?
+    __host__ __device__ static constexpr ck::index_t GetStaticLdsSize()
+    {
+        return sizeof(BDataType) *
+               Policy::template MakeBLdsBlockDescriptor<Problem>().GetElementSpaceSize();
+    }
+
+    // Cold A Register Cache
+    template <typename ADramBlockWindowTmp,
+              typename BDramBlockWindowTmp,
+              typename AElementFunction,
+              typename BElementFunction,
+              typename ARegBlockTensorTmp>
+    __host__ __device__ auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
+                                        const AElementFunction& a_element_func,
+                                        const BDramBlockWindowTmp& b_dram_block_window_tmp,
+                                        const BElementFunction& b_element_func,
+                                        ARegBlockTensorTmp& a_reg_block_tensor_tmp,
+                                        void* p_smem) const
+    {
+        static_assert(
+            is_same_v<ADataType, remove_cvref_t<typename ADramBlockWindowTmp::DataType>> &&
+                is_same_v<BDataType, remove_cvref_t<typename BDramBlockWindowTmp::DataType>>,
+            "wrong!");
+
+        static_assert(kMPerBlock == ADramBlockWindowTmp{}.GetWindowLengths()[Number<0>{}] &&
+                          kNPerBlock == BDramBlockWindowTmp{}.GetWindowLengths()[Number<0>{}] &&
+                          kKPerBlock == ADramBlockWindowTmp{}.GetWindowLengths()[Number<1>{}],
+                      "wrong!");
+
+        ignore = a_element_func;
+        ignore = b_element_func;
+
+        // A tile in Reg，blockTensor
+        // This tensor distribution used to construct both distributed tensor for local buffer store
+        // and read. without buffer address info
+        constexpr auto a_reg_block_dstr = Policy::template MakeARegBlockDescriptor<Problem>();
+
+        // B tile in LDS, blockWindow
+        BDataType* p_b_lds =
+            static_cast<BDataType*>(static_cast<void*>(static_cast<char*>(p_smem)));
+
+        constexpr auto b_lds_block_desc = Policy::template MakeBLdsBlockDescriptor<Problem>();
+
+        // This tensor view used to construct both tile window for lds store and read, with buffer
+        // address info
+        auto b_lds_block = make_tensor_view<AddressSpaceEnum::Lds>(p_b_lds, b_lds_block_desc);
+
+        // A DRAM tile window for load
+        auto a_copy_dram_window =
+            make_tile_window(a_dram_block_window_tmp.GetBottomTensorView(),
+                             make_tuple(Number<kMPerBlock>{}, Number<kKPerBlock>{}),
+                             a_dram_block_window_tmp.GetWindowOrigin(),
+                             Policy::template MakeADramTileDistribution<Problem>());
+
+        // A Reg tensor for store, also used for block GEMM
+        auto a_copy_reg_tensor = make_static_distributed_tensor<ADataType>(a_reg_block_dstr);
+
+        // B DRAM tile window for load
+        auto b_copy_dram_window =
+            make_tile_window(b_dram_block_window_tmp.GetBottomTensorView(),
+                             make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}),
+                             b_dram_block_window_tmp.GetWindowOrigin(),
+                             Policy::template MakeBDramTileDistribution<Problem>());
+
+        // B LDS tile window for store
+        auto b_copy_lds_window =
+            make_tile_window(b_lds_block,
+                             make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}),
+                             {0, 0},
+                             b_copy_dram_window.GetTileDistribution());
+
+        // B LDS tile for block GEMM
+        auto b_lds_gemm_window = make_tile_window(
+            b_lds_block, make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}), {0, 0});
+
+        // Block GEMM
+        constexpr auto block_gemm = Policy::template GetBlockGemm<Problem>();
+
+        // Acc register tile
+        auto c_block_tile = decltype(block_gemm(
+            get_slice_tile(a_copy_reg_tensor, Sequence<0, 0>{}, Sequence<kMPerBlock, kKPerBlock>{}),
+            b_lds_gemm_window)){};
+
+        auto a_block_tile = load_tile(a_copy_dram_window);
+        auto b_block_tile = load_tile(b_copy_dram_window);
+        {
+            move_tile_window(a_copy_dram_window, {0, kKPerBlock});
+            move_tile_window(b_copy_dram_window, {0, kKPerBlock});
+
+            tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
+
+            set_slice_tile(a_copy_reg_tensor,
+                           a_block_tile,
+                           Sequence<0, 0>{},
+                           Sequence<kMPerBlock, kKPerBlock>{});
+            a_block_tile = load_tile(a_copy_dram_window);
+
+            store_tile(b_copy_lds_window, b_block_tile);
+            b_block_tile = load_tile(b_copy_dram_window);
+        }
+        if constexpr(k_loops > 2)
+        {
+            static_for<0, k_loops - 2, 1>{}([&](auto i_k0) {
+                block_sync_lds();
+
+                block_gemm(c_block_tile,
+                           get_slice_tile(a_copy_reg_tensor,
+                                          Sequence<0, (i_k0)*kKPerBlock>{},
+                                          Sequence<kMPerBlock, (i_k0 + 1) * kKPerBlock>{}),
+                           b_copy_lds_window);
+
+                block_sync_lds();
+
+                move_tile_window(a_copy_dram_window, {0, kKPerBlock});
+                move_tile_window(b_copy_dram_window, {0, kKPerBlock});
+
+                set_slice_tile(a_copy_reg_tensor,
+                               a_block_tile,
+                               Sequence<0, (i_k0 + 1) * kKPerBlock>{},
+                               Sequence<kMPerBlock, (i_k0 + 2) * kKPerBlock>{});
+                a_block_tile = load_tile(a_copy_dram_window);
+
+                store_tile(b_copy_lds_window, b_block_tile);
+                b_block_tile = load_tile(b_copy_dram_window);
+            });
+        }
+
+        // tail
+        {
+            block_sync_lds();
+
+            block_gemm(c_block_tile,
+                       get_slice_tile(a_copy_reg_tensor,
+                                      Sequence<0, (k_loops - 2) * kKPerBlock>{},
+                                      Sequence<kMPerBlock, (k_loops - 1) * kKPerBlock>{}),
+                       b_copy_lds_window);
+
+            block_sync_lds();
+
+            set_slice_tile(a_copy_reg_tensor,
+                           a_block_tile,
+                           Sequence<0, (k_loops - 1) * kKPerBlock>{},
+                           Sequence<kMPerBlock, k_loops * kKPerBlock>{});
+
+            store_tile(b_copy_lds_window, b_block_tile);
+
+            block_sync_lds();
+
+            block_gemm(c_block_tile,
+                       get_slice_tile(a_copy_reg_tensor,
+                                      Sequence<0, (k_loops - 1) * kKPerBlock>{},
+                                      Sequence<kMPerBlock, (k_loops)*kKPerBlock>{}),
+                       b_copy_lds_window);
+        }
+
+        store_tile(a_reg_block_tensor_tmp, a_copy_reg_tensor);
+
+        return c_block_tile;
+    }
+
+    // Hot A Register Cache
+    template <typename BDramBlockWindowTmp, typename BElementFunction, typename ARegBlockTensorTmp>
+    __host__ __device__ auto operator()(const BDramBlockWindowTmp& b_dram_block_window_tmp,
+                                        const BElementFunction& b_element_func,
+                                        const ARegBlockTensorTmp& a_reg_block_tensor_tmp,
+                                        void* p_smem) const
+    {
+        static_assert(is_same_v<BDataType, remove_cvref_t<typename BDramBlockWindowTmp::DataType>>,
+                      "wrong!");
+
+        static_assert(kNPerBlock == BDramBlockWindowTmp{}.GetWindowLengths()[Number<0>{}] &&
+                          kKPerBlock == BDramBlockWindowTmp{}.GetWindowLengths()[Number<1>{}],
+                      "wrong!");
+
+        ignore = b_element_func;
+
+        // A tile in Reg，blockTensor
+        // This tensor distribution used to construct both distributed tensor for local buffer store
+        // and read. without buffer address info
+        constexpr auto a_reg_block_dstr = Policy::template MakeARegBlockDescriptor<Problem>();
+
+        // A Reg tensor for store, also used for block GEMM
+        auto a_copy_reg_tensor = make_static_distributed_tensor<ADataType>(a_reg_block_dstr);
+        store_tile(a_copy_reg_tensor, a_reg_block_tensor_tmp);
+
+        // B tile in LDS, blockWindow
+        BDataType* p_b_lds =
+            static_cast<BDataType*>(static_cast<void*>(static_cast<char*>(p_smem)));
+
+        constexpr auto b_lds_block_desc = Policy::template MakeBLdsBlockDescriptor<Problem>();
+
+        // This tensor view used to construct both tile window for lds store and read, with buffer
+        // address info
+        auto b_lds_block = make_tensor_view<AddressSpaceEnum::Lds>(p_b_lds, b_lds_block_desc);
+
+        // B DRAM tile window for load
+        auto b_copy_dram_window =
+            make_tile_window(b_dram_block_window_tmp.GetBottomTensorView(),
+                             make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}),
+                             b_dram_block_window_tmp.GetWindowOrigin(),
+                             Policy::template MakeBDramTileDistribution<Problem>());
+
+        // B LDS tile window for store
+        auto b_copy_lds_window =
+            make_tile_window(b_lds_block,
+                             make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}),
+                             {0, 0},
+                             b_copy_dram_window.GetTileDistribution());
+
+        // B LDS tile for block GEMM
+        auto b_lds_gemm_window = make_tile_window(
+            b_lds_block, make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}), {0, 0});
+
+        // Block GEMM
+        constexpr auto block_gemm = Policy::template GetBlockGemm<Problem>();
+
+        // Acc register tile
+        auto c_block_tile = decltype(block_gemm(
+            get_slice_tile(a_copy_reg_tensor, Sequence<0, 0>{}, Sequence<kMPerBlock, kKPerBlock>{}),
+            b_lds_gemm_window)){};
+
+        auto b_block_tile = load_tile(b_copy_dram_window);
+        {
+            move_tile_window(b_copy_dram_window, {0, kKPerBlock});
+
+            tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
+
+            store_tile(b_copy_lds_window, b_block_tile);
+            b_block_tile = load_tile(b_copy_dram_window);
+        }
+        if constexpr(k_loops > 2)
+        {
+            static_for<0, k_loops - 2, 1>{}([&](auto i_k0) {
+                block_sync_lds();
+
+                block_gemm(c_block_tile,
+                           get_slice_tile(a_copy_reg_tensor,
+                                          Sequence<0, (i_k0)*kKPerBlock>{},
+                                          Sequence<kMPerBlock, (i_k0 + 1) * kKPerBlock>{}),
+                           b_copy_lds_window);
+
+                block_sync_lds();
+
+                move_tile_window(b_copy_dram_window, {0, kKPerBlock});
+
+                store_tile(b_copy_lds_window, b_block_tile);
+                b_block_tile = load_tile(b_copy_dram_window);
+            });
+        }
+
+        // tail
+        {
+            block_sync_lds();
+
+            block_gemm(c_block_tile,
+                       get_slice_tile(a_copy_reg_tensor,
+                                      Sequence<0, (k_loops - 2) * kKPerBlock>{},
+                                      Sequence<kMPerBlock, (k_loops - 1) * kKPerBlock>{}),
+                       b_copy_lds_window);
+
+            block_sync_lds();
+
+            store_tile(b_copy_lds_window, b_block_tile);
+
+            block_sync_lds();
+
+            block_gemm(c_block_tile,
+                       get_slice_tile(a_copy_reg_tensor,
+                                      Sequence<0, (k_loops - 1) * kKPerBlock>{},
+                                      Sequence<kMPerBlock, (k_loops)*kKPerBlock>{}),
+                       b_copy_lds_window);
+        }
+
+        return c_block_tile;
+    }
+
+    template <typename ADramBlockWindowTmp,
+              typename BDramBlockWindowTmp,
+              typename ARegBlockTensorTmp>
+    __device__ auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
+                               const BDramBlockWindowTmp& b_dram_block_window_tmp,
+                               ARegBlockTensorTmp& a_reg_block_tensor_tmp,
+                               void* p_smem) const
+    {
+        return operator()(
+            a_dram_block_window_tmp,
+            [](const ADataType& a) { return a; },
+            b_dram_block_window_tmp,
+            [](const BDataType& b) { return b; },
+            a_reg_block_tensor_tmp,
+            p_smem);
+    }
+
+    template <typename BDramBlockWindowTmp, typename ARegBlockTensorTmp>
+    __device__ auto operator()(const BDramBlockWindowTmp& b_dram_block_window_tmp,
+                               const ARegBlockTensorTmp& a_reg_block_tensor_tmp,
+                               void* p_smem) const
+    {
+        return operator()(
+            b_dram_block_window_tmp,
+            [](const BDataType& b) { return b; },
+            a_reg_block_tensor_tmp,
+            p_smem);
+    }
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2_askiplds_policy.hpp b/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2_askiplds_policy.hpp
new file mode 100644
index 0000000000..e54d3a4f55
--- /dev/null
+++ b/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2_askiplds_policy.hpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tile_program/block_tile_pipeline/blockgemm_pipeline_agmem_bgmem_creg_policy_impl.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v1.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+// NOTE: Assume A is K-Major
+struct BlockGemmPipelineAGmemBGmemCRegV2SkipALdsPolicy
+{
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeARegBlockDescriptor()
+    {
+        constexpr auto blockgemm = GetBlockGemm<Problem>();
+        using BlockGemm          = ck::remove_cvref_t<decltype(blockgemm)>;
+
+        return policy_impl::make_a_reg_block_descriptor<Problem, BlockGemm>();
+    }
+
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeBLdsBlockDescriptor()
+    {
+        return policy_impl::make_b_lds_block_descriptor_3d_pad<Problem>();
+    }
+
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeADramTileDistribution()
+    {
+        constexpr auto blockgemm = GetBlockGemm<Problem>();
+        using BlockGemm          = ck::remove_cvref_t<decltype(blockgemm)>;
+
+        return policy_impl::make_a_dram_tile_distribution_skip_lds<Problem, BlockGemm>();
+    }
+
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeBDramTileDistribution()
+    {
+        return policy_impl::make_b_dram_tile_distribution<Problem>();
+    }
+
+    template <typename Problem>
+    __host__ __device__ static constexpr auto GetBlockGemm()
+    {
+        using BlockGemmPolicy = BlockGemmARegBSmemCRegV1K8Policy;
+
+        return BlockGemmARegBSmemCRegV1<Problem, BlockGemmPolicy>{};
+    }
+};
+
+template <index_t AKDim_>
+struct BlockGemmPipelineAGmemBGmemCRegV2SkipALdsPersistentQRegCachePolicy
+    : BlockGemmPipelineAGmemBGmemCRegV2SkipALdsPolicy
+{
+    static constexpr index_t AKDim = AKDim_;
+
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeARegBlockDescriptor()
+    {
+        using namespace ck;
+
+        constexpr auto blockgemm = GetBlockGemm<Problem>();
+        using BlockGemm          = ck::remove_cvref_t<decltype(blockgemm)>;
+
+        constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+        constexpr index_t kKPerBlock = AKDim;
+
+        constexpr auto config =
+            BlockGemm::BlockGemmPolicy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template At<0>())>;
+
+        constexpr index_t MWarp = config.template At<1>();
+        constexpr index_t NWarp = config.template At<2>();
+
+        constexpr index_t MIterPerWarp = kMPerBlock / (MWarp * WG::kM);
+        constexpr index_t KIterPerWarp = kKPerBlock / WG::kK;
+
+        constexpr auto a_block_outer_dstr_encoding = StaticTileDistributionEncoding<
+            Sequence<NWarp>,
+            Tuple<Sequence<MIterPerWarp, MWarp>, Sequence<KIterPerWarp>>,
+            Tuple<Sequence<1, 0>>,
+            Tuple<Sequence<1, 0>>,
+            Sequence<1, 2>,
+            Sequence<0, 0>>{};
+
+        constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            a_block_outer_dstr_encoding, typename WG::AWarpDstrEncoding{});
+
+        constexpr auto a_block_dstr = make_static_tile_distribution(a_block_dstr_encode);
+
+        return a_block_dstr;
+    }
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2_default_policy.hpp b/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2_default_policy.hpp
new file mode 100644
index 0000000000..dc99a3bb94
--- /dev/null
+++ b/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v2_default_policy.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_agmem_bgmem_creg_v1.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+// Default policy for BlockGemmPipelineAGmemBGmemCRegV2
+// Default policy class should not be templated, put template on member functions instead
+// NOTE: policy should be binded to its corresponding operation. It's just a coincidence that
+//   BlockGemmPipelineAGmemBGmemCRegV2DefaultPolicy is the same as
+//   BlockGemmPipelineAGmemBGmemCRegV1DefaultPolicy
+using BlockGemmPipelineAGmemBGmemCRegV2DefaultPolicy =
+    BlockGemmPipelineAGmemBGmemCRegV1DefaultPolicy;
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_problem.hpp b/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_problem.hpp
new file mode 100644
index 0000000000..0465347543
--- /dev/null
+++ b/include/ck/tile_program/block_tile_pipeline/block_gemm_pipeline_problem.hpp
@@ -0,0 +1,30 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/type.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+
+template <typename ADataType_,
+          typename BDataType_,
+          typename CDataType_,
+          index_t kBlockSize_,
+          typename BlockGemmShape_>
+struct BlockGemmPipelineProblem
+{
+    using ADataType      = remove_cvref_t<ADataType_>;
+    using BDataType      = remove_cvref_t<BDataType_>;
+    using CDataType      = remove_cvref_t<CDataType_>;
+    using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
+
+    static constexpr index_t kBlockSize = kBlockSize_;
+};
+
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/block_tile_pipeline/blockgemm_pipeline_agmem_bgmem_creg_policy_impl.hpp b/include/ck/tile_program/block_tile_pipeline/blockgemm_pipeline_agmem_bgmem_creg_policy_impl.hpp
new file mode 100644
index 0000000000..d57ad3e3a6
--- /dev/null
+++ b/include/ck/tile_program/block_tile_pipeline/blockgemm_pipeline_agmem_bgmem_creg_policy_impl.hpp
@@ -0,0 +1,193 @@
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace block {
+namespace policy_impl {
+// 3d + padding
+template <typename Problem>
+__host__ __device__ static constexpr auto make_a_lds_block_descriptor_3d_pad()
+{
+    using namespace ck;
+
+    constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+    constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+    constexpr auto a_lds_block_desc_0 = make_naive_tensor_descriptor(
+        make_tuple(Number<kKPerBlock / 8>{}, Number<kMPerBlock>{}, Number<8>{}),
+        make_tuple(Number<(kMPerBlock + 1) * 8>{}, Number<8>{}, Number<1>{}),
+        Number<8>{},
+        Number<1>{});
+
+    constexpr auto a_lds_block_desc =
+        transform_tensor_descriptor(a_lds_block_desc_0,
+                                    make_tuple(make_pass_through_transform(kMPerBlock),
+                                               make_merge_transform(make_tuple(kKPerBlock / 8, 8))),
+                                    make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    return a_lds_block_desc;
+}
+
+// 3d + padding
+template <typename Problem>
+__host__ __device__ static constexpr auto make_b_lds_block_descriptor_3d_pad()
+{
+    using namespace ck;
+
+    constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+    constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+    constexpr auto b_lds_block_desc_0 = make_naive_tensor_descriptor(
+        make_tuple(Number<kKPerBlock / 8>{}, Number<kNPerBlock>{}, Number<8>{}),
+        make_tuple(Number<(kNPerBlock + 1) * 8>{}, Number<8>{}, Number<1>{}),
+        Number<8>{},
+        Number<1>{});
+
+    constexpr auto b_lds_block_desc =
+        transform_tensor_descriptor(b_lds_block_desc_0,
+                                    make_tuple(make_pass_through_transform(kNPerBlock),
+                                               make_merge_transform(make_tuple(kKPerBlock / 8, 8))),
+                                    make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    return b_lds_block_desc;
+}
+
+template <typename Problem, typename BlockGemm>
+__host__ __device__ static constexpr auto make_a_reg_block_descriptor()
+{
+    using namespace ck;
+
+    constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+    constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+    constexpr auto config = BlockGemm::BlockGemmPolicy::template GetWarpGemmMWarpNWarp<Problem>();
+
+    using WG = remove_cvref_t<decltype(config.template At<0>())>;
+
+    constexpr index_t MWarp = config.template At<1>();
+    constexpr index_t NWarp = config.template At<2>();
+
+    constexpr index_t MIterPerWarp = kMPerBlock / (MWarp * WG::kM);
+    constexpr index_t KIterPerWarp = kKPerBlock / WG::kK;
+
+    constexpr auto a_block_outer_dstr_encoding =
+        StaticTileDistributionEncoding<Sequence<NWarp>,
+                                       Tuple<Sequence<MIterPerWarp, MWarp>, Sequence<KIterPerWarp>>,
+                                       Tuple<Sequence<1, 0>>,
+                                       Tuple<Sequence<1, 0>>,
+                                       Sequence<1, 2>,
+                                       Sequence<0, 0>>{};
+
+    constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+        a_block_outer_dstr_encoding, typename WG::AWarpDstrEncoding{});
+
+    constexpr auto a_block_dstr = make_static_tile_distribution(a_block_dstr_encode);
+
+    return a_block_dstr;
+}
+
+template <typename Problem>
+__host__ __device__ static constexpr auto make_a_dram_tile_distribution()
+{
+    using ADataType = remove_cvref_t<typename Problem::ADataType>;
+
+    constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+    constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+    constexpr index_t K1 = 16 / sizeof(ADataType);
+    constexpr index_t K0 = kKPerBlock / K1;
+    constexpr index_t M2 = get_warp_size() / K0;
+
+    constexpr index_t M1 = kBlockSize / get_warp_size();
+    constexpr index_t M0 = kMPerBlock / (M2 * M1);
+
+    return make_static_tile_distribution(
+        StaticTileDistributionEncoding<Sequence<1>,
+                                       Tuple<Sequence<M0, M1, M2>, Sequence<K0, K1>>,
+                                       Tuple<Sequence<1>, Sequence<1, 2>>,
+                                       Tuple<Sequence<1>, Sequence<2, 0>>,
+                                       Sequence<1, 2>,
+                                       Sequence<0, 1>>{});
+}
+
+template <typename Problem, typename BlockGemm>
+__host__ __device__ static constexpr auto make_a_dram_tile_distribution_skip_lds()
+{
+    constexpr auto config = BlockGemm::BlockGemmPolicy::template GetWarpGemmMWarpNWarp<Problem>();
+
+    using WG = remove_cvref_t<decltype(config.template At<0>())>;
+
+    constexpr index_t MWarp = config.template At<1>();
+
+    constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+    constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+    constexpr index_t K2 =
+        WG::kK / WG::WarpGemmAttribute::Impl::kABKLane; // WG::WarpGemmAttribute::Impl::kABKPerLane;
+                                                        // // 16 / sizeof(ADataType);
+    constexpr index_t K1 = WG::WarpGemmAttribute::Impl::kABKLane;
+    constexpr index_t K0 = kKPerBlock / (K1 * K2);
+
+    constexpr index_t M2 = WG::WarpGemmAttribute::Impl::kAMLane;
+    constexpr index_t M1 = MWarp;
+    constexpr index_t M0 = kMPerBlock / (M2 * M1);
+
+    return make_static_tile_distribution(
+        StaticTileDistributionEncoding<Sequence<1>,
+                                       Tuple<Sequence<M0, M1, M2>, Sequence<K0, K1, K2>>,
+                                       Tuple<Sequence<1>, Sequence<2, 1>>,
+                                       Tuple<Sequence<1>, Sequence<1, 2>>,
+                                       Sequence<2, 1, 2>,
+                                       Sequence<0, 0, 2>>{});
+}
+
+template <typename Problem>
+__host__ __device__ static constexpr auto make_b_dram_tile_distribution()
+{
+    using BDataType = remove_cvref_t<typename Problem::BDataType>;
+
+    constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+    constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+    constexpr index_t K1 = 16 / sizeof(BDataType);
+    constexpr index_t K0 = kKPerBlock / K1;
+    constexpr index_t N2 = get_warp_size() / K0;
+
+    constexpr index_t N1 = kBlockSize / get_warp_size();
+    constexpr index_t N0 = kNPerBlock / (N2 * N1);
+
+    return make_static_tile_distribution(
+        StaticTileDistributionEncoding<Sequence<1>,
+                                       Tuple<Sequence<N0, N1, N2>, Sequence<K0, K1>>,
+                                       Tuple<Sequence<1>, Sequence<1, 2>>,
+                                       Tuple<Sequence<1>, Sequence<2, 0>>,
+                                       Sequence<1, 2>,
+                                       Sequence<0, 1>>{});
+}
+
+template <typename Problem>
+__host__ __device__ static constexpr auto get_block_gemm()
+{
+    using BlockGemmPolicy = BlockGemmASmemBSmemCRegV1DefaultPolicy;
+
+    return BlockGemmASmemBSmemCRegV1<Problem, BlockGemmPolicy>{};
+}
+
+} // namespace policy_impl
+} // namespace block
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/grid/block_to_c_tile.hpp b/include/ck/tile_program/grid/block_to_c_tile.hpp
new file mode 100644
index 0000000000..03cba6f437
--- /dev/null
+++ b/include/ck/tile_program/grid/block_to_c_tile.hpp
@@ -0,0 +1,127 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <type_traits>
+#include <utility>
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/utility/multi_index.hpp"
+#include "ck/utility/type.hpp"
+#include "ck/utility/tuple.hpp"
+
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+
+namespace ck {
+namespace detail {
+
+template <typename Descriptor>
+class DescToBlock2TileMapAdaptor
+{
+    static_assert(std::is_same_v<
+                  MultiIndex<2>,
+                  remove_cvref_t<decltype(std::declval<const Descriptor&>().CalculateBottomIndex(
+                      std::declval<MultiIndex<1>>()))>>);
+
+    Descriptor descriptor_;
+
+    public:
+    explicit constexpr DescToBlock2TileMapAdaptor(Descriptor descriptor)
+        : descriptor_(std::move(descriptor))
+    {
+    }
+
+    __host__ __device__ MultiIndex<2> operator()(index_t block_id) const
+    {
+        return descriptor_.CalculateBottomIndex(make_multi_index(block_id));
+    }
+};
+
+template <typename Descriptor>
+__host__ __device__ static auto make_desc_to_block2tile_map_adaptor(Descriptor&& descriptor)
+{
+    return DescToBlock2TileMapAdaptor<remove_cvref_t<Descriptor>>{
+        std::forward<Descriptor>(descriptor)};
+}
+} // namespace detail
+
+struct Block2TileMapNFast
+{
+    __host__ __device__ static constexpr auto MakeBlock2TileMap(index_t NumTilesM,
+                                                                index_t NumTilesN)
+    {
+        return ck::detail::make_desc_to_block2tile_map_adaptor(
+            make_cluster_descriptor(make_tuple(NumTilesM, NumTilesN)));
+    }
+};
+
+struct Block2TileMapMFast
+{
+    __host__ __device__ constexpr auto operator()(index_t NumTilesM, index_t NumTilesN) const
+    {
+        const auto unmerge = make_merge_transform(make_tuple(NumTilesN, NumTilesM));
+
+        return [unmerge](index_t block_id) {
+            MultiIndex<2> unmerged;
+            unmerge.CalculateLowerIndex(unmerged, make_multi_index(block_id));
+
+            return make_multi_index(unmerged.At<1>(), unmerged.At<0>());
+        };
+    }
+};
+
+/// NOTICE: This map will be compiled into considerable amount of instructions.
+///         Use with caution or replace it with more efficient implementation.
+template <index_t MaxCols = 8>
+struct Block2TileMapNAdapt
+{
+    __host__ __device__ static constexpr auto MakeBlock2TileMap(index_t NumTilesM,
+                                                                index_t NumTilesN)
+    {
+        return [=](index_t block_id) {
+            index_t idx_M0 = block_id % NumTilesM;
+            index_t idx_N0 = block_id / NumTilesM;
+
+            const auto LastCols =
+                (idx_N0 < NumTilesN - NumTilesN % MaxCols) ? MaxCols : NumTilesN % MaxCols;
+
+            index_t idx_N00          = idx_N0 / MaxCols;
+            index_t idx_N01          = idx_N0 % MaxCols;
+            index_t idx_M0_N01_local = idx_M0 + idx_N01 * NumTilesM;
+
+            return make_multi_index(idx_M0_N01_local / LastCols,
+                                    idx_M0_N01_local % LastCols + idx_N00 * MaxCols);
+        };
+    }
+};
+
+/// NOTICE: This map will be compiled into considerable amount of instructions.
+///         Use with caution or replace it with more efficient implementation.
+template <index_t MaxRows = 8>
+struct Block2TileMapMAdapt
+{
+    __host__ __device__ static constexpr auto MakeBlock2TileMap(index_t NumTilesM,
+                                                                index_t NumTilesN)
+    {
+        return [=](index_t block_id) {
+            index_t idx_N0 = block_id % NumTilesN;
+            index_t idx_M0 = block_id / NumTilesN;
+
+            const auto LastRows =
+                (idx_M0 < NumTilesM - NumTilesM % MaxRows) ? MaxRows : NumTilesM % MaxRows;
+
+            index_t idx_M00          = idx_M0 / MaxRows;
+            index_t idx_M01          = idx_M0 % MaxRows;
+            index_t idx_N0_M01_local = idx_N0 + idx_M01 * NumTilesN;
+
+            return make_multi_index(idx_N0_M01_local % LastRows + idx_M00 * MaxRows,
+                                    idx_N0_M01_local / LastRows);
+        };
+    }
+};
+
+using DefaultBlock2TileMap = Block2TileMapMFast;
+
+} // namespace ck
diff --git a/include/ck/tile_program/grid/grid_gemm_problem.hpp b/include/ck/tile_program/grid/grid_gemm_problem.hpp
new file mode 100644
index 0000000000..ca5483d0f1
--- /dev/null
+++ b/include/ck/tile_program/grid/grid_gemm_problem.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace tile_program {
+namespace grid {
+
+template <typename ADataType_,
+          typename BDataType_,
+          typename AccDataType_,
+          typename CDataType_,
+          typename AElementFunction_,
+          typename BElementFunction_,
+          typename CElementFunction_>
+struct GridGemmProblem
+{
+    using ADataType   = ADataType_;
+    using BDataType   = BDataType_;
+    using AccDataType = AccDataType_;
+    using CDataType   = CDataType_;
+
+    using AElementFunction = AElementFunction_;
+    using BElementFunction = BElementFunction_;
+    using CElementFunction = CElementFunction_;
+};
+
+} // namespace grid
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/grid/grid_gemm_v1.hpp b/include/ck/tile_program/grid/grid_gemm_v1.hpp
new file mode 100644
index 0000000000..2fb943d5d3
--- /dev/null
+++ b/include/ck/tile_program/grid/grid_gemm_v1.hpp
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+template <typename Problem, typename Policy>
+struct GridGemmV1
+{
+    using ADataType        = typename Problem::ADataType;
+    using BDataType        = typename Problem::BDataType;
+    using CDataType        = typename Problem::CDataType;
+    using AElementFunction = typename Problem::AElementFunction;
+    using BElementFunction = typename Problem::BElementFunction;
+    using CElementFunction = typename Problem::CElementFunction;
+
+    static constexpr auto kMPerBlock = Policy::kMPerBlock;
+    static constexpr auto kNPerBlock = Policy::kNPerBlock;
+    static constexpr auto kKPerBlock = Policy::kKPerBlock;
+
+    template <typename AGridTensorView, typename BGridTensorView, typename CGridTensorView>
+    __device__ void operator()(const AGridTensorView& a_grid,
+                               const BGridTensorView& b_grid,
+                               CGridTensorView& c_grid,
+                               const AElementFunction& a_element_func,
+                               const BElementFunction& b_element_func,
+                               const CElementFunction& c_element_func) const
+    {
+        using namespace ck;
+        using namespace ck::tile_program;
+        using namespace ck::tile_program::block;
+
+        const auto M = a_grid.GetTensorDescriptor().GetLength(Number<0>{});
+        const auto N = c_grid.GetTensorDescriptor().GetLength(Number<1>{});
+        const auto K = a_grid.GetTensorDescriptor().GetLength(Number<1>{});
+
+        // divide problem
+        const auto id_block = get_block_id();
+
+        const auto num_tile_m = M / kMPerBlock;
+        const auto num_tile_n = N / kNPerBlock;
+
+        const auto block2tile = Policy::template MakeBlock2TileMap<Problem>(num_tile_m, num_tile_n);
+
+        const auto id_tile = block2tile(id_block);
+
+        const auto iM = __builtin_amdgcn_readfirstlane(id_tile.template At<0>() * kMPerBlock);
+        const auto iN = __builtin_amdgcn_readfirstlane(id_tile.template At<1>() * kNPerBlock);
+
+        // A block window
+        auto a_block_window = make_tile_window(
+            a_grid, make_tuple(Number<kMPerBlock>{}, Number<kKPerBlock>{}), {iM, 0});
+
+        // B block window
+        auto b_block_window = make_tile_window(
+            b_grid, make_tuple(Number<kNPerBlock>{}, Number<kKPerBlock>{}), {iN, 0});
+
+        // Block GEMM pipeline
+        constexpr auto block_gemm_pipeline = Policy::template GetBlockGemmPipeline<Problem>();
+
+        __shared__ char p_smem_char[block_gemm_pipeline.GetStaticLdsSize()];
+
+        const auto acc_block_tile = block_gemm_pipeline(a_block_window,
+                                                        a_element_func,
+                                                        b_block_window,
+                                                        b_element_func,
+                                                        K / kKPerBlock,
+                                                        p_smem_char);
+
+        // cast to CDataType and apply CElementFunction
+        const auto c_block_tile = tile_elementwise_in(
+            [&](const auto& acc) { return c_element_func(type_convert<CDataType>(acc)); },
+            acc_block_tile);
+
+        // store C
+        auto c_window = make_tile_window(
+            c_grid, make_tuple(Number<kMPerBlock>{}, Number<kNPerBlock>{}), {iM, iN});
+
+        store_tile(c_window, c_block_tile);
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tile_program/grid/grid_gemm_v1_default_policy.hpp b/include/ck/tile_program/grid/grid_gemm_v1_default_policy.hpp
new file mode 100644
index 0000000000..b19cb4fc5f
--- /dev/null
+++ b/include/ck/tile_program/grid/grid_gemm_v1_default_policy.hpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tile_program/block_tile_pipeline/block_gemm_pipeline_problem.hpp"
+#include "ck/tile_program/tile/tile_gemm_shape.hpp"
+
+namespace ck {
+
+// Default policy for GridGemmV1
+// Default policy class should not be templated, put template on member functions instead
+struct GridGemmV1DefaultPolicy
+{
+    static constexpr index_t kBlockSize = 256;
+    static constexpr index_t kMPerBlock = 128;
+    static constexpr index_t kNPerBlock = 128;
+    static constexpr index_t kKPerBlock = 32;
+
+    template <typename Problem>
+    __host__ __device__ static constexpr auto MakeBlock2TileMap(index_t NumTilesM,
+                                                                index_t NumTilesN)
+    {
+        const auto unmerge = make_merge_transform(make_tuple(NumTilesN, NumTilesM));
+
+        return [unmerge](index_t block_id) {
+            MultiIndex<2> unmerged;
+            unmerge.CalculateLowerIndex(unmerged, make_multi_index(block_id));
+
+            return make_multi_index(unmerged.At<1>(), unmerged.At<0>());
+        };
+    }
+
+    template <typename Problem>
+    __host__ __device__ static constexpr auto GetBlockGemmPipeline()
+    {
+        using namespace ck::tile_program;
+        using namespace ck::tile_program::block;
+
+        using BlockGemmPipelineProblem_ =
+            BlockGemmPipelineProblem<typename Problem::ADataType,
+                                     typename Problem::BDataType,
+                                     typename Problem::AccDataType,
+                                     kBlockSize,
+                                     TileGemmShape<kMPerBlock, kNPerBlock, kKPerBlock>>;
+
+        return BlockGemmPipelineAGmemBGmemCRegV2<BlockGemmPipelineProblem_,
+                                                 BlockGemmPipelineAGmemBGmemCRegV2DefaultPolicy>{};
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/tile_program/tile/distributed_tile_sweep.hpp b/include/ck/tile_program/tile/distributed_tile_sweep.hpp
new file mode 100644
index 0000000000..6f438693fd
--- /dev/null
+++ b/include/ck/tile_program/tile/distributed_tile_sweep.hpp
@@ -0,0 +1,29 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+
+namespace ck {
+namespace tile_program {
+
+// sweep over a span of a distribted tile and apply lambda function F
+template <typename TileDistributedSpan_, // TileDistributedSpan<...>
+          typename F                     // signature: F(TileDistributedIndex<...>)
+          >
+__device__ void sweep_tile_span(TileDistributedSpan_, const F& f)
+{
+    using DstrSpan = remove_cvref_t<TileDistributedSpan_>;
+
+    static_ford<typename DstrSpan::Impl>{}([&](auto dstr_idx_impl) {
+        constexpr auto dstr_idx = detail::make_tile_distributed_index(dstr_idx_impl);
+
+        f(dstr_idx);
+    });
+}
+
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/load_tile.hpp b/include/ck/tile_program/tile/load_tile.hpp
new file mode 100644
index 0000000000..b2ea61afa9
--- /dev/null
+++ b/include/ck/tile_program/tile/load_tile.hpp
@@ -0,0 +1,32 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_window.hpp"
+#include "ck/tile_program/tile/static_distributed_tensor.hpp"
+
+namespace ck {
+namespace tile_program {
+
+template <typename BottomTensorView_,
+          typename WindowLengths_,
+          typename TileDistribution_,
+          index_t NumCoord>
+__device__ auto load_tile(const TileWindowWithStaticDistribution<BottomTensorView_,
+                                                                 WindowLengths_,
+                                                                 TileDistribution_,
+                                                                 NumCoord>& tile_window)
+{
+    return tile_window.Load();
+}
+
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/shuffle_distributed_tensor.hpp b/include/ck/tile_program/tile/shuffle_distributed_tensor.hpp
new file mode 100644
index 0000000000..6b13f0ad9e
--- /dev/null
+++ b/include/ck/tile_program/tile/shuffle_distributed_tensor.hpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tile_program/tile/tile_elementwise.hpp"
+#include "ck/tile_program/tile/shuffle_distributed_tensor_impl_in_thread.hpp"
+
+namespace ck {
+namespace tile_program {
+
+template <typename OutTensor, typename InTensor>
+__device__ void shuffle_distributed_tensor(OutTensor& out, const InTensor& in)
+{
+    using InDataType  = typename InTensor::DataType;
+    using OutDataType = typename OutTensor::DataType;
+
+    using InDstrEncode  = typename InTensor::StaticTileDistribution::DstrEncode;
+    using OutDstrEncode = typename OutTensor::StaticTileDistribution::DstrEncode;
+
+    // type convert
+    const auto in_tmp = tile_elementwise_in(type_convert<OutDataType, InDataType>, in);
+
+    // shuffle
+    if constexpr(InDstrEncode::rs_lengths_ == OutDstrEncode::rs_lengths_ &&
+                 InDstrEncode::hs_lengthss_ == OutDstrEncode::hs_lengthss_ &&
+                 InDstrEncode::ps_to_rhss_major_ == OutDstrEncode::ps_to_rhss_major_ &&
+                 InDstrEncode::ps_to_rhss_minor_ == OutDstrEncode::ps_to_rhss_minor_ &&
+                 InDstrEncode::NDimY == OutDstrEncode::NDimY)
+    {
+        detail::shuffle_distributed_tensor_impl_in_thread(out, in_tmp);
+    }
+    else
+    {
+        // NOT implemented
+    }
+}
+
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/shuffle_distributed_tensor_impl_in_thread.hpp b/include/ck/tile_program/tile/shuffle_distributed_tensor_impl_in_thread.hpp
new file mode 100644
index 0000000000..a8758460d4
--- /dev/null
+++ b/include/ck/tile_program/tile/shuffle_distributed_tensor_impl_in_thread.hpp
@@ -0,0 +1,152 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/static_distributed_tensor.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace detail {
+
+template <typename OutTensor, typename InTensor>
+__device__ void shuffle_distributed_tensor_impl_in_thread(OutTensor& out_tensor,
+                                                          const InTensor& in_tensor)
+{
+    constexpr auto I0 = Number<0>{};
+
+    using DataType = typename InTensor::DataType;
+
+    constexpr auto y_in_desc  = InTensor::GetTileDistribution().GetYs2DDescriptor();
+    constexpr auto y_out_desc = OutTensor::GetTileDistribution().GetYs2DDescriptor();
+
+    // y_dim_out_to_in
+    constexpr auto get_rh_major_minor_to_y = [](auto dstr_tensor) {
+        using DstrEncode = typename decltype(dstr_tensor.GetTileDistribution())::DstrEncode;
+
+        Map<Array<index_t, 2>, index_t> rh_major_minor_to_y_;
+
+        static_for<0, DstrEncode::NDimY, 1>{}([&](auto i) {
+            constexpr index_t rh_major = DstrEncode::ys_to_rhs_major_[i];
+            constexpr index_t rh_minor = DstrEncode::ys_to_rhs_minor_[i];
+
+            rh_major_minor_to_y_({rh_major, rh_minor}) = i;
+        });
+
+        return rh_major_minor_to_y_;
+    };
+
+    constexpr auto rh_major_minor_to_y_in  = get_rh_major_minor_to_y(InTensor{});
+    constexpr auto rh_major_minor_to_y_out = get_rh_major_minor_to_y(OutTensor{});
+
+    constexpr auto y_dim_out_to_in = [&] {
+        Map<index_t, index_t> y_dim_out_to_in_;
+
+        for(const auto& [rh_major_minor, y_out] : rh_major_minor_to_y_out)
+        {
+            y_dim_out_to_in_(y_out) = rh_major_minor_to_y_in[rh_major_minor];
+        }
+
+        return y_dim_out_to_in_;
+    }();
+
+    //
+    constexpr index_t NDimY = InTensor::GetTileDistribution().GetNumOfDimensionY();
+
+    constexpr auto y_lengths = to_sequence(y_in_desc.GetLengths());
+
+    // input and output vector dim in the order of input Y dims
+    constexpr index_t y_dim_vec_in  = NDimY - 1;
+    constexpr index_t y_dim_vec_out = y_dim_out_to_in[NDimY - 1];
+
+    // vector lengths
+    constexpr index_t vec_length_in  = y_lengths[y_dim_vec_in];
+    constexpr index_t vec_length_out = y_lengths[y_dim_vec_out];
+
+    // # of vectors
+    constexpr index_t num_vec_in  = vec_length_out;
+    constexpr index_t num_vec_out = vec_length_in;
+
+    using InVec  = vector_type<DataType, vec_length_in>;
+    using OutVec = vector_type<DataType, vec_length_out>;
+
+    using InVecType  = typename InVec::type;
+    using OutVecType = typename OutVec::type;
+
+    // SFC
+    constexpr auto scalars_per_access_arr = generate_array(
+        [&](auto i) { return (i == y_dim_vec_in or i == y_dim_vec_out) ? y_lengths[i] : 1; },
+        Number<NDimY>{});
+
+    constexpr auto scalars_per_access = TO_SEQUENCE(scalars_per_access_arr, NDimY);
+
+    using SFC_Y = SpaceFillingCurve<decltype(y_lengths),
+                                    typename arithmetic_sequence_gen<0, NDimY, 1>::type,
+                                    decltype(scalars_per_access)>;
+
+    constexpr index_t num_access = SFC_Y::GetNumOfAccess();
+
+    static_assert(num_access > 0, "wrong! num_access should be larger than 0");
+
+    // in/out vectors to be transposed
+    StaticallyIndexedArray<InVec, num_vec_in> in_vectors;
+    StaticallyIndexedArray<OutVec, num_vec_out> out_vectors;
+
+#if 0
+    print(y_dim_out_to_in);
+    printf("\n");
+    printf("y_dim_vec_in %d\n", y_dim_vec_in);
+    printf("y_dim_vec_out %d\n", y_dim_vec_out);
+    printf("num_vec_in %d\n", num_vec_in);
+    printf("num_vec_out %d\n", num_vec_out);
+#endif
+
+    // loop over SFC and do transpose
+    static_for<0, num_access, 1>{}([&](auto iAccess) {
+        // data index [y0, y1, ...] in the order of input tensor
+        constexpr auto idx_y_start = SFC_Y::GetIndex(iAccess);
+
+        // get input vectors
+        static_for<0, num_vec_in, 1>{}([&](auto i) {
+            constexpr auto idx_y_in = generate_array(
+                [&](auto ii) {
+                    return ii == y_dim_vec_out ? idx_y_start[ii] + i : idx_y_start[ii];
+                },
+                Number<NDimY>{});
+
+            constexpr index_t in_offset = y_in_desc.CalculateOffset(idx_y_in);
+
+            in_vectors(i).template AsType<InVecType>()(I0) =
+                in_tensor.GetThreadBuffer().template GetAsType<InVecType>(Number<in_offset>{});
+        });
+
+        // transpose
+        transpose_vectors<DataType, num_vec_in, num_vec_out>{}(in_vectors, out_vectors);
+
+        // set output vectors
+        static_for<0, num_vec_out, 1>{}([&](auto i) {
+            constexpr auto idx_y_out_tmp = generate_array(
+                [&](auto ii) { return ii == y_dim_vec_in ? idx_y_start[ii] + i : idx_y_start[ii]; },
+                Number<NDimY>{});
+
+            constexpr auto idx_y_out =
+                container_reorder_given_new2old(idx_y_out_tmp, y_dim_out_to_in);
+
+            constexpr index_t out_offset = y_out_desc.CalculateOffset(idx_y_out);
+
+            out_tensor.GetThreadBuffer().template SetAsType<OutVecType>(
+                Number<out_offset>{}, out_vectors[i].template AsType<OutVecType>()[I0]);
+        });
+    });
+}
+
+} // namespace detail
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/slice_tile.hpp b/include/ck/tile_program/tile/slice_tile.hpp
new file mode 100644
index 0000000000..78aa1208a7
--- /dev/null
+++ b/include/ck/tile_program/tile/slice_tile.hpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/static_tile_distribution_helper.hpp"
+#include "ck/tile_program/tile/tile_window.hpp"
+#include "ck/tile_program/tile/static_distributed_tensor.hpp"
+
+namespace ck {
+namespace tile_program {
+
+template <typename StaticDistributedTensor_, index_t... SliceBegins, index_t... SliceEnds>
+__host__ __device__ constexpr auto get_slice_tile(const StaticDistributedTensor_& tile,
+                                                  Sequence<SliceBegins...> slice_begins,
+                                                  Sequence<SliceEnds...> slice_ends)
+{
+    using Distribution = decltype(StaticDistributedTensor_::GetTileDistribution());
+    using DataType     = typename StaticDistributedTensor_::DataType;
+
+    constexpr auto sliced_dstr_yidx_ylen =
+        detail::slice_distribution_from_x(Distribution{}, slice_begins, slice_ends);
+
+    constexpr auto sliced_dstr      = sliced_dstr_yidx_ylen.template At<0>();
+    constexpr auto sliced_y_origins = sliced_dstr_yidx_ylen.template At<1>();
+    constexpr auto sliced_y_lengths = sliced_dstr_yidx_ylen.template At<2>();
+
+    auto sliced_tensor = make_static_distributed_tensor<DataType>(sliced_dstr);
+
+    sliced_tensor.GetThreadBuffer() = tile.GetYSlicedThreadData(sliced_y_origins, sliced_y_lengths);
+
+    return sliced_tensor;
+}
+
+template <typename DstStaticDistributedTensor_,
+          typename SrcStaticDistributedTensor_,
+          index_t... SliceBegins,
+          index_t... SliceEnds>
+__host__ __device__ constexpr auto set_slice_tile(DstStaticDistributedTensor_& dst_tile,
+                                                  const SrcStaticDistributedTensor_& src_tile,
+                                                  Sequence<SliceBegins...> slice_begins,
+                                                  Sequence<SliceEnds...> slice_ends)
+{
+    using DstDistribution = decltype(DstStaticDistributedTensor_::GetTileDistribution());
+    // using SrcDistribution = decltype(SrcStaticDistributedTensor_::GetTileDistribution());
+
+    constexpr auto sliced_dstr_yidx_ylen =
+        detail::slice_distribution_from_x(DstDistribution{}, slice_begins, slice_ends);
+
+    // constexpr auto sliced_dstr      = sliced_dstr_yidx_ylen.template At<0>();
+    constexpr auto sliced_y_origins = sliced_dstr_yidx_ylen.template At<1>();
+    constexpr auto sliced_y_lengths = sliced_dstr_yidx_ylen.template At<2>();
+
+    // static_assert(is_same_v<decltype(sliced_dstr), SrcDistribution>, "wrong!");
+
+    dst_tile.SetYSlicedThreadData(sliced_y_origins, sliced_y_lengths, src_tile.GetThreadBuffer());
+}
+
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/static_distributed_tensor.hpp b/include/ck/tile_program/tile/static_distributed_tensor.hpp
new file mode 100644
index 0000000000..19f420aa4d
--- /dev/null
+++ b/include/ck/tile_program/tile/static_distributed_tensor.hpp
@@ -0,0 +1,180 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+
+namespace ck {
+namespace tile_program {
+
+template <typename DataType_, typename StaticTileDistribution_>
+struct StaticDistributedTensor
+{
+    using DataType               = remove_cvref_t<DataType_>;
+    using StaticTileDistribution = remove_cvref_t<StaticTileDistribution_>;
+
+    static_assert(StaticTileDistribution::IsStatic(),
+                  "wrong! StaticTileDistribution should be known at compile tile");
+
+    using ThreadTensorDesc = remove_cvref_t<decltype(StaticTileDistribution{}.GetYs2DDescriptor())>;
+
+    static constexpr index_t kThreadElementSpaceSize = ThreadTensorDesc{}.GetElementSpaceSize();
+
+    __host__ __device__ static constexpr auto GetNumOfDimension()
+    {
+        return StaticTileDistribution::GetNumOfDimensionX();
+    }
+
+    __host__ __device__ static constexpr auto GetLengths()
+    {
+        return StaticTileDistribution::GetLengths();
+    }
+
+    __host__ __device__ static constexpr auto GetTileDistribution()
+    {
+        return StaticTileDistribution{};
+    }
+
+    __host__ __device__ static constexpr auto GetDistributedSpans()
+    {
+        return StaticTileDistribution::GetDistributedSpans();
+    }
+
+    __host__ __device__ void Initialize(const DataType& x) { thread_buf_.Initialize(x); }
+
+    __host__ __device__ constexpr const auto& GetThreadBuffer() const { return thread_buf_; }
+
+    __host__ __device__ constexpr auto& GetThreadBuffer() { return thread_buf_; }
+
+    __host__ __device__ static constexpr index_t GetThreadBufferSize()
+    {
+        return kThreadElementSpaceSize;
+    }
+
+    template <index_t... YSliceOrigins, index_t... YSliceLengths>
+    __host__ __device__ auto GetYSlicedThreadData(Sequence<YSliceOrigins...>,
+                                                  Sequence<YSliceLengths...>) const
+    {
+        static_assert(sizeof...(YSliceOrigins) == StaticTileDistribution::NDimY &&
+                          sizeof...(YSliceLengths) == StaticTileDistribution::NDimY,
+                      "wrong!");
+
+        constexpr auto sliced_thread_tensor_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(YSliceLengths...));
+
+        StaticBuffer<AddressSpaceEnum::Vgpr,
+                     DataType,
+                     sliced_thread_tensor_desc.GetElementSpaceSize(),
+                     true>
+            sliced_thread_data;
+
+        static_ford<Sequence<YSliceLengths...>>{}([&](auto idx) {
+            constexpr auto idx_ys = idx + Sequence<YSliceOrigins...>{};
+
+            sliced_thread_data(Number<sliced_thread_tensor_desc.CalculateOffset(idx)>{}) =
+                thread_buf_[Number<ThreadTensorDesc{}.CalculateOffset(idx_ys)>{}];
+        });
+
+        return sliced_thread_data;
+    }
+
+    template <index_t... YSliceOrigins, index_t... YSliceLengths, index_t NSlicedData>
+    __host__ __device__ void SetYSlicedThreadData(
+        Sequence<YSliceOrigins...>,
+        Sequence<YSliceLengths...>,
+        const StaticBuffer<AddressSpaceEnum::Vgpr, DataType, NSlicedData, true>& sliced_thread_data)
+    {
+        static_assert(sizeof...(YSliceOrigins) == StaticTileDistribution::NDimY &&
+                          sizeof...(YSliceLengths) == StaticTileDistribution::NDimY,
+                      "wrong!");
+
+        constexpr auto sliced_thread_tensor_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(YSliceLengths...));
+
+        static_ford<Sequence<YSliceLengths...>>{}([&](auto idx) {
+            constexpr auto idx_ys = idx + Sequence<YSliceOrigins...>{};
+
+            thread_buf_(Number<ThreadTensorDesc{}.CalculateOffset(idx_ys)>{}) =
+                sliced_thread_data[Number<sliced_thread_tensor_desc.CalculateOffset(idx)>{}];
+        });
+    }
+
+    template <typename TileDistributedIndices>
+    __host__ __device__ constexpr const DataType& operator[](TileDistributedIndices) const
+    {
+        static_assert(is_static_v<TileDistributedIndices>,
+                      "wrong! Tile Distributed Indices should be static");
+
+        constexpr auto y_idx =
+            GetTileDistribution().GetYIndicesFromDistributedIndices(TileDistributedIndices{});
+
+        return thread_buf_[Number<ThreadTensorDesc{}.CalculateOffset(y_idx)>{}];
+    }
+
+    template <typename TileDistributedIndices>
+    __host__ __device__ constexpr DataType& operator()(TileDistributedIndices)
+    {
+        static_assert(is_static_v<TileDistributedIndices>,
+                      "wrong! Tile Distributed Indices should be static");
+
+        constexpr auto y_idx =
+            GetTileDistribution().GetYIndicesFromDistributedIndices(TileDistributedIndices{});
+
+        return thread_buf_(Number<ThreadTensorDesc{}.CalculateOffset(y_idx)>{});
+    }
+
+#if 0
+    template <index_t... Ys>
+    __host__ __device__ auto GetElementFromYsIndex(Sequence<Ys...> idx_ys) const
+    {
+        return thread_buf_[Number<ThreadTensorDesc{}.CalculateOffset(idx_ys)>{}];
+    }
+
+    template <index_t... Ys>
+    __host__ __device__ void SetElementFromYsIndex(Sequence<Ys...> idx_ys, const DataType& v)
+    {
+        thread_buf_(Number<ThreadTensorDesc{}.CalculateOffset(idx_ys)>{}) = v;
+    }
+    template <typename TileDistributedIndices>
+    __host__ __device__ auto GetElementFromTileDistributedIndices(TileDistributedIndices) const
+    {
+        static_assert(is_static_v<TileDistributedIndices>, "wrong!");
+
+        constexpr auto y_idx =
+            GetTileDistribution().GetYIndicesFromDistributedIndices(TileDistributedIndices{});
+
+        return GetElementFromYsIndex(y_idx);
+    }
+
+    template <typename TileDistributedIndices>
+    __host__ __device__ void SetElementFromTileDistributedIndices(TileDistributedIndices,
+                                                                  const DataType& v)
+    {
+        static_assert(is_static_v<TileDistributedIndices>, "wrong!");
+
+        constexpr auto y_idx =
+            GetTileDistribution().GetYIndicesFromDistributedIndices(TileDistributedIndices{});
+
+        return SetElementFromYsIndex(y_idx, v);
+    }
+#endif
+
+    //
+    StaticBuffer<AddressSpaceEnum::Vgpr, DataType, kThreadElementSpaceSize, true> thread_buf_;
+};
+
+template <typename DataType, typename StaticTileDistribution>
+__host__ __device__ constexpr auto make_static_distributed_tensor(const StaticTileDistribution&)
+{
+    return StaticDistributedTensor<remove_cvref_t<DataType>,
+                                   remove_cvref_t<StaticTileDistribution>>{};
+}
+
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/static_tile_distribution_encoding.hpp b/include/ck/tile_program/tile/static_tile_distribution_encoding.hpp
new file mode 100644
index 0000000000..8511955864
--- /dev/null
+++ b/include/ck/tile_program/tile/static_tile_distribution_encoding.hpp
@@ -0,0 +1,411 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+
+namespace ck {
+namespace tile_program {
+
+template <typename RsLengths_,    // Sequence<...>
+          typename HsLengthss_,   // Tuple<Sequence<...>, ...>
+          typename Ps2RHssMajor_, // Tuple<Sequence<...>, ...>
+          typename Ps2RHssMinor_, // Tuple<Sequence<...>, ...>
+          typename Ys2RHsMajor_,  // Sequence<...>
+          typename Ys2RHsMinor_>  // Sequence<...>
+struct StaticTileDistributionEncoding
+{
+    using RsLengths    = remove_cvref_t<RsLengths_>;
+    using HsLengthss   = remove_cvref_t<HsLengthss_>;
+    using Ps2RHssMajor = remove_cvref_t<Ps2RHssMajor_>;
+    using Ps2RHssMinor = remove_cvref_t<Ps2RHssMinor_>;
+    using Ys2RHsMajor  = remove_cvref_t<Ys2RHsMajor_>;
+    using Ys2RHsMinor  = remove_cvref_t<Ys2RHsMinor_>;
+
+    static_assert(Ps2RHssMajor::Size() == Ps2RHssMinor::Size(), "wrong!");
+    static_assert(Ys2RHsMajor::Size() == Ys2RHsMinor::Size(), "wrong!");
+
+    static constexpr index_t NDimX = HsLengthss::Size();
+    static constexpr index_t NDimP = Ps2RHssMajor::Size();
+    static constexpr index_t NDimY = Ys2RHsMajor::Size();
+    static constexpr index_t NDimR = RsLengths::Size();
+
+    // FIXME: move into Detail
+    static constexpr auto rs_lengths_       = RsLengths{};
+    static constexpr auto hs_lengthss_      = HsLengthss{};
+    static constexpr auto ps_to_rhss_major_ = Ps2RHssMajor{};
+    static constexpr auto ps_to_rhss_minor_ = Ps2RHssMinor{};
+    static constexpr auto ys_to_rhs_major_  = Ys2RHsMajor{};
+    static constexpr auto ys_to_rhs_minor_  = Ys2RHsMinor{};
+
+    // redundant but useful info
+    // TODO: really bad code, should be over-hauled
+    struct Detail
+    {
+        // ndim_rh_major_, ndim_span_mainor_
+        static constexpr index_t ndim_rh_major_   = NDimX + 1;
+        static constexpr index_t ndim_span_major_ = NDimX;
+
+        // ndims_rhs_minor_[ndim_rh_major_]
+        static constexpr auto ndims_rhs_minor_ = generate_array(
+            [](auto i) {
+                if constexpr(i.value == 0)
+                {
+                    return rs_lengths_.Size();
+                }
+                else
+                {
+                    return hs_lengthss_[i - Number<1>{}].Size();
+                }
+            },
+            Number<ndim_rh_major_>{});
+
+        // max_ndim_rh_minor_
+        static constexpr index_t max_ndim_rh_minor_ =
+            container_reduce(ndims_rhs_minor_, math::maximize<index_t>{}, 0);
+
+        // rhs_lengthss_[ndim_rh_major_][max_ndim_rh_minor_]
+        static constexpr auto rhs_lengthss_ =
+            to_array_of_array(container_concat(make_tuple(rs_lengths_), hs_lengthss_));
+
+        // ys_lengths_
+        static constexpr auto ys_lengths_ = [] {
+            Array<index_t, NDimY> ys_lengths_tmp{-1};
+
+            for(index_t i = 0; i < NDimY; i++)
+            {
+                index_t rh_major = ys_to_rhs_major_[i];
+                index_t rh_minor = ys_to_rhs_minor_[i];
+
+                ys_lengths_tmp(i) = rhs_lengthss_[rh_major][rh_minor];
+            }
+
+            return ys_lengths_tmp;
+        }();
+
+        // rhs_major_minor_to_ys_[ndim_rh_majpr_][max_ndim_rh_minor_]
+        static constexpr auto rhs_major_minor_to_ys_ = [] {
+            Array<Array<index_t, max_ndim_rh_minor_>, NDimX + 1> rhs_major_minor_to_ys_tmp{{-1}};
+
+            static_for<0, NDimY, 1>{}([&](auto i) {
+                constexpr index_t rh_major = ys_to_rhs_major_[i];
+                constexpr index_t rh_minor = ys_to_rhs_minor_[i];
+
+                rhs_major_minor_to_ys_tmp(rh_major)(rh_minor) = i;
+            });
+
+            return rhs_major_minor_to_ys_tmp;
+        }();
+
+        // ndims_span_minor_[NDimY]
+        static constexpr auto ndims_span_minor_ = [] {
+            Array<index_t, NDimX> ndims_span_minor{0};
+
+            for(index_t i = 0; i < NDimY; i++)
+            {
+                const index_t span_major = ys_to_rhs_major_[i] - 1;
+
+                ndims_span_minor(span_major)++;
+            }
+
+            return ndims_span_minor;
+        }();
+
+        // max_ndim_span_minor_
+        static constexpr index_t max_ndim_span_minor_ =
+            container_reduce(ndims_span_minor_, math::maximize<index_t>{}, 0);
+
+        // rhs_major_minor_to_span_minor_ [ndim_rh_major_][max_ndim_rh_minor_]
+        static constexpr auto rhs_major_minor_to_span_minor_ = [] {
+            Array<Array<index_t, max_ndim_rh_minor_>, ndim_rh_major_> rhs_major_minor_to_span_minor{
+                {-1}};
+
+            static_for<0, ndim_rh_major_, 1>{}([&](auto rh_major) {
+                constexpr index_t ndim_rh_minor = ndims_rhs_minor_[rh_major];
+
+                index_t cnt_ndim_span_minor = 0;
+
+                static_for<0, ndim_rh_minor, 1>{}([&](auto rh_minor) {
+                    constexpr index_t idim_y = rhs_major_minor_to_ys_[rh_major][rh_minor];
+
+                    if(idim_y >= 0)
+                    {
+                        rhs_major_minor_to_span_minor(rh_major)(rh_minor) = cnt_ndim_span_minor;
+
+                        cnt_ndim_span_minor++;
+                    }
+                });
+            });
+
+            return rhs_major_minor_to_span_minor;
+        }();
+
+        // ys_to_span_major_[NDimY]
+        static constexpr auto ys_to_span_major_ =
+            generate_array([](auto i) { return ys_to_rhs_major_[i] - 1; }, Number<NDimY>{});
+
+        // ys_to_span_minor_[NDimY]
+        static constexpr auto ys_to_span_minor_ = generate_array(
+            [](auto i) {
+                return rhs_major_minor_to_span_minor_[ys_to_rhs_major_[i]][ys_to_rhs_minor_[i]];
+            },
+            Number<NDimY>{});
+
+        // distributed_spans_lengthss_[ndim_span_major_][max_ndim_span_minor_]
+        static constexpr auto distributed_spans_lengthss_ = [] {
+            Array<Array<index_t, max_ndim_span_minor_>, ndim_span_major_>
+                distributed_spans_lengthss{{-1}};
+
+            static_for<0, NDimY, 1>{}([&](auto i) {
+                const index_t rh_major = ys_to_rhs_major_[i];
+                const index_t rh_minor = ys_to_rhs_minor_[i];
+
+                const index_t h_length = hs_lengthss_[Number<rh_major - 1>{}][rh_minor];
+
+                const index_t span_major = rh_major - 1;
+                const index_t span_minor = rhs_major_minor_to_span_minor_[rh_major][rh_minor];
+
+                distributed_spans_lengthss(span_major)(span_minor) = h_length;
+            });
+
+            return distributed_spans_lengthss;
+        }();
+
+        // ndims_distributed_spans_minor_[ndim_span_major_]
+        static constexpr auto ndims_distributed_spans_minor_ = [] {
+            Array<index_t, ndim_span_major_> ndims_distributed_spans_minor{0};
+
+            static_for<0, NDimY, 1>{}([&](auto i) {
+                const index_t span_major = ys_to_rhs_major_[i] - 1;
+
+                ndims_distributed_spans_minor(span_major)++;
+            });
+
+            return ndims_distributed_spans_minor;
+        }();
+
+        // does_p_own_r_[NDimP][NDimR]
+        static constexpr auto does_p_own_r_ = [] {
+            if constexpr(NDimR > 0)
+            {
+                Array<Array<bool, NDimR>, NDimP> does_p_own_r{{false}};
+
+                static_for<0, NDimP, 1>{}([&](auto idim_p) {
+                    constexpr index_t ndim_low = ps_to_rhss_major_[idim_p].Size();
+
+                    static_for<0, ndim_low, 1>{}([&](auto idim_low) {
+                        constexpr index_t rh_major = ps_to_rhss_major_[idim_p][idim_low];
+                        constexpr index_t rh_minor = ps_to_rhss_minor_[idim_p][idim_low];
+
+                        if constexpr(rh_major == 0)
+                        {
+                            does_p_own_r(idim_p)(rh_minor) = true;
+                        }
+                    });
+                });
+
+                return does_p_own_r;
+            }
+            else
+            {
+                return Array<Array<bool, NDimR>, NDimP>{};
+            }
+        }();
+
+        // ps_over_rs_derivative_[NDimP][NDimR]
+        static constexpr auto ps_over_rs_derivative_ = [] {
+            if constexpr(NDimR > 0)
+            {
+                Array<Array<index_t, NDimR>, NDimP> ps_over_rs_derivative{{0}};
+
+                static_for<0, NDimP, 1>{}([&](auto idim_p) {
+                    constexpr index_t ndim_low = ps_to_rhss_major_[idim_p].Size();
+
+                    index_t p_over_rh_derivative = 1;
+
+                    static_for<ndim_low - 1, -1, -1>{}([&](auto idim_low) {
+                        constexpr index_t rh_major = ps_to_rhss_major_[idim_p][idim_low];
+                        constexpr index_t rh_minor = ps_to_rhss_minor_[idim_p][idim_low];
+
+                        constexpr index_t rh_length = rhs_lengthss_[rh_major][rh_minor];
+
+                        if constexpr(rh_major == 0)
+                        {
+                            ps_over_rs_derivative(idim_p)(rh_minor) = p_over_rh_derivative;
+                        }
+
+                        p_over_rh_derivative *= rh_length;
+                    });
+                });
+
+                return ps_over_rs_derivative;
+            }
+            else
+            {
+                return Array<Array<index_t, NDimR>, NDimP>{};
+            }
+        }();
+
+        // e.g. tuple<seq<1, 4, 32>, seq<4, 1, 4, 2, 4>> --> seq<3, 5> --> seq<0, 3, 8>
+        __host__ __device__ static constexpr auto GetHDimLengthsPrefixSum()
+        {
+            // <len_d0, len_d1, ...>
+            // e.g. tuple<seq<1, 4, 32>, seq<4, 1, 4, 2, 4>> --> seq<3, 5>
+            constexpr auto uniformed_h_dim_lengths = generate_sequence_v2(
+                [&](auto i) {
+                    constexpr index_t size = HsLengthss{}[i].Size();
+                    return Number<size>{};
+                },
+                Number<NDimX>{});
+
+            // <0, len_d0, len_d0+len_d1, ...>
+            // e.g. seq<3, 5> --> seq<0, 3, 8>
+            constexpr auto h_dim_prefix_sum = prefix_sum_sequence(uniformed_h_dim_lengths);
+
+            return h_dim_prefix_sum;
+        }
+
+        __host__ __device__ static constexpr auto GetUniformedIdxY2H()
+        {
+            constexpr auto all_ys_2_rhss = transform_sequences(
+                [](auto major, auto minor) constexpr {
+                    // <0, 0, len_d0, len_d0+len_d1, ...>
+                    constexpr auto x_dim_prefix_sum =
+                        merge_sequences(Sequence<0>{} /*for R dims*/, GetHDimLengthsPrefixSum());
+                    return x_dim_prefix_sum.At(major) + minor;
+                },
+                Ys2RHsMajor{},
+                Ys2RHsMinor{});
+
+            return all_ys_2_rhss;
+        }
+
+        // return tuple<sorted_dims, sorted_maps, sorted_prefix_sum>
+        template <typename IdxSeq, typename PrefixSumSeq>
+        __host__ __device__ static constexpr auto GetSortedInfo(IdxSeq, PrefixSumSeq)
+        {
+            using sorted_idx =
+                sequence_unique_sort<IdxSeq, math::less<index_t>, math::equal<index_t>>;
+
+            constexpr auto sorted_dims = typename sorted_idx::type{};
+            constexpr auto sorted_maps = typename sorted_idx::sorted2unsorted_map{};
+
+            constexpr auto sorted_histogram =
+                histogram_sorted_sequence(sorted_dims, PrefixSumSeq{});
+            constexpr auto sorted_prefix_sum = prefix_sum_sequence(sorted_histogram);
+
+            return make_tuple(sorted_dims, sorted_maps, sorted_prefix_sum);
+        }
+
+        __host__ __device__ static constexpr auto GetSortedYInfo()
+        {
+            return GetSortedInfo(GetUniformedIdxY2H(), GetHDimLengthsPrefixSum());
+        }
+
+        __host__ __device__ void Print() const
+        {
+            printf("StaticTileDistributionEncoding::Detail{");
+            //
+            printf("ndim_rh_major_: ");
+            print(ndim_rh_major_);
+            printf(", ");
+            //
+            printf("ndim_span_major_: ");
+            print(ndim_span_major_);
+            printf(", ");
+            //
+            printf("ndims_rhs_minor_: ");
+            print(ndims_rhs_minor_);
+            printf(", ");
+            //
+            printf("ndim_rh_major_: ");
+            print(ndim_rh_major_);
+            printf(", ");
+            //
+            printf("max_ndim_rh_minor_: ");
+            print(max_ndim_rh_minor_);
+            printf(", ");
+            //
+            printf("rhs_lengthss_: ");
+            print(rhs_lengthss_);
+            printf(", ");
+            //
+            printf("ys_lengths_: ");
+            print(ys_lengths_);
+            printf(", ");
+            //
+            printf("rhs_major_minor_to_ys_: ");
+            print(rhs_major_minor_to_ys_);
+            printf(", ");
+            //
+            printf("ndims_span_minor_: ");
+            print(ndims_span_minor_);
+            printf(", ");
+            //
+            printf("max_ndim_span_minor_: ");
+            print(max_ndim_span_minor_);
+            printf(", ");
+            //
+            printf("ys_to_span_major_: ");
+            print(ys_to_span_major_);
+            printf(", ");
+            //
+            printf("ys_to_span_minor_: ");
+            print(ys_to_span_minor_);
+            printf(", ");
+            //
+            printf("distributed_spans_lengthss_: ");
+            print(distributed_spans_lengthss_);
+            printf(", ");
+            //
+            printf("ndims_distributed_spans_minor_: ");
+            print(ndims_distributed_spans_minor_);
+            printf(", ");
+            //
+            printf("ps_over_rs_derivative_: ");
+            print(ps_over_rs_derivative_);
+            //
+            printf("}");
+        }
+    };
+
+    __host__ __device__ void Print() const
+    {
+        printf("StaticTileDistributionEncoding{");
+        //
+        printf("NDimX: %d, NDimP: %d, NDimY: %d, ", NDimX, NDimP, NDimY);
+        //
+        printf("rs_lengths_: ");
+        print(rs_lengths_);
+        printf(", ");
+        //
+        printf("hs_lengthss_: ");
+        print(hs_lengthss_);
+        printf(", ");
+        //
+        printf("ps_to_rhss_major_: ");
+        print(ps_to_rhss_major_);
+        printf(", ");
+        //
+        printf("ps_to_rhss_minor_: ");
+        print(ps_to_rhss_minor_);
+        printf(", ");
+        //
+        printf("ys_to_rhs_major_: ");
+        print(ys_to_rhs_major_);
+        printf(", ");
+        //
+        printf("ys_to_rhs_minor_: ");
+        print(ys_to_rhs_minor_);
+        printf(", ");
+        //
+        printf("Detail: ");
+        print(Detail{});
+        //
+        printf("}");
+    }
+};
+
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/static_tile_distribution_helper.hpp b/include/ck/tile_program/tile/static_tile_distribution_helper.hpp
new file mode 100644
index 0000000000..f3115422ed
--- /dev/null
+++ b/include/ck/tile_program/tile/static_tile_distribution_helper.hpp
@@ -0,0 +1,612 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace detail {
+
+template <typename Distribution>
+__host__ __device__ auto get_partition_index(Distribution)
+{
+    // only support warp-tile and block-tile
+    static_assert(Distribution::NDimP == 1 or Distribution::NDimP == 2, "wrong!");
+
+    if constexpr(Distribution::NDimP == 1)
+    {
+        return Array<index_t, 1>{get_lane_id()};
+    }
+    else if constexpr(Distribution::NDimP == 2)
+    {
+        return Array<index_t, 2>{get_warp_id(), get_lane_id()};
+    }
+}
+
+template <typename OuterDstr, typename InnerDstr>
+__host__ __device__ constexpr auto make_embed_tile_distribution_encoding(OuterDstr, InnerDstr)
+{
+    static_assert(OuterDstr::NDimX == InnerDstr::NDimX, "wrong!");
+
+    constexpr index_t NDimHMajor = OuterDstr::NDimX;
+
+    using RsLengths =
+        sequence_merge_t<typename OuterDstr::RsLengths, typename InnerDstr::RsLengths>;
+
+    constexpr auto hs_lengthss = generate_tuple(
+        [&](auto i) {
+            return merge_sequences(typename OuterDstr::HsLengthss{}[i],
+                                   typename InnerDstr::HsLengthss{}[i]);
+        },
+        Number<NDimHMajor>{});
+
+    //
+    constexpr auto rhs_major_2_ndim_outer_rhs_minor = [&]() {
+        Array<index_t, NDimHMajor + 1> rhs_major_2_ndim_outer_rhs_minor_;
+
+        // R dimension
+        rhs_major_2_ndim_outer_rhs_minor_(0) = OuterDstr::RsLengths::Size();
+
+        // Hs dimensions
+        static_for<0, NDimHMajor, 1>{}([&](auto i) {
+            rhs_major_2_ndim_outer_rhs_minor_(i + 1) = typename OuterDstr::HsLengthss{}[i].Size();
+        });
+
+        return rhs_major_2_ndim_outer_rhs_minor_;
+    }();
+
+    // Ps2RHssMinor
+    constexpr auto updated_inner_ps_2_rhss_minor = generate_tuple(
+        [&](auto p) {
+            constexpr auto inner_p_2_rhss_major = typename InnerDstr::Ps2RHssMajor{}[p];
+            constexpr auto inner_p_2_rhss_minor = typename InnerDstr::Ps2RHssMinor{}[p];
+
+            constexpr index_t ndim_tmp = inner_p_2_rhss_minor.Size();
+
+            constexpr auto updated_inner_p_2_rhss_minor = [&]() {
+                Array<index_t, ndim_tmp> updated_inner_p_2_rhss_minor_;
+
+                for(index_t i = 0; i < ndim_tmp; i++)
+                {
+                    index_t rh_major = inner_p_2_rhss_major[i];
+
+                    index_t ndim_outer_h_minor = rhs_major_2_ndim_outer_rhs_minor[rh_major];
+
+                    updated_inner_p_2_rhss_minor_(i) = inner_p_2_rhss_minor[i] + ndim_outer_h_minor;
+                }
+
+                return updated_inner_p_2_rhss_minor_;
+            }();
+
+            return TO_SEQUENCE(updated_inner_p_2_rhss_minor, ndim_tmp);
+        },
+        Number<InnerDstr::NDimP>{});
+
+    // Ys2RHsMinor
+    constexpr auto updated_inner_ys_2_rhs_minor = [&]() {
+        constexpr auto inner_ys_2_rhs_major = typename InnerDstr::Ys2RHsMajor{};
+        constexpr auto inner_ys_2_rhs_minor = typename InnerDstr::Ys2RHsMinor{};
+
+        constexpr index_t ndim_tmp = inner_ys_2_rhs_minor.Size();
+
+        constexpr auto updated_inner_ys_2_rhs_minor_ = [&]() {
+            Array<index_t, ndim_tmp> updated_inner_ys_2_rhs_minor__;
+
+            for(index_t i = 0; i < ndim_tmp; i++)
+            {
+                index_t rh_major = inner_ys_2_rhs_major[i];
+
+                index_t ndim_outer_h_minor = rhs_major_2_ndim_outer_rhs_minor[rh_major];
+
+                updated_inner_ys_2_rhs_minor__(i) = inner_ys_2_rhs_minor[i] + ndim_outer_h_minor;
+            }
+
+            return updated_inner_ys_2_rhs_minor__;
+        }();
+
+        return TO_SEQUENCE(updated_inner_ys_2_rhs_minor_, ndim_tmp);
+    }();
+
+    //
+    constexpr auto ps_2_rhss_major =
+        container_concat(typename OuterDstr::Ps2RHssMajor{}, typename InnerDstr::Ps2RHssMajor{});
+
+    constexpr auto ps_2_rhss_minor =
+        container_concat(typename OuterDstr::Ps2RHssMinor{}, updated_inner_ps_2_rhss_minor);
+
+    //
+    constexpr auto ys_2_rhs_major =
+        merge_sequences(typename OuterDstr::Ys2RHsMajor{}, typename InnerDstr::Ys2RHsMajor{});
+
+    constexpr auto ys_2_rhs_minor =
+        merge_sequences(typename OuterDstr::Ys2RHsMinor{}, updated_inner_ys_2_rhs_minor);
+
+    return StaticTileDistributionEncoding<RsLengths,
+                                          remove_cvref_t<decltype(hs_lengthss)>,
+                                          remove_cvref_t<decltype(ps_2_rhss_major)>,
+                                          remove_cvref_t<decltype(ps_2_rhss_minor)>,
+                                          remove_cvref_t<decltype(ys_2_rhs_major)>,
+                                          remove_cvref_t<decltype(ys_2_rhs_minor)>>{};
+}
+
+template <typename InDstr, index_t... InReduceDimXs>
+__host__ __device__ constexpr auto
+make_reduce_tile_distribution_encoding_impl(InDstr, Sequence<InReduceDimXs...> reduce_dim_xs_in)
+{
+    constexpr auto I1 = Number<1>{};
+
+    // FIXME: increase if fail
+    constexpr index_t max_ndim_r_out = 20;
+    constexpr index_t max_ndim_y_out = 20;
+
+    //
+    constexpr index_t ndim_p               = InDstr::NDimP;
+    constexpr index_t ndim_x_in            = InDstr::NDimX;
+    constexpr index_t ndim_y_in            = InDstr::NDimY;
+    constexpr index_t ndim_rh_major_in     = InDstr::NDimX + 1;
+    constexpr index_t ndim_x_out           = ndim_x_in - sizeof...(InReduceDimXs);
+    constexpr index_t max_ndim_rh_minor_in = InDstr::Detail::max_ndim_rh_minor_;
+
+    // ndims_ps_low
+    constexpr auto ndims_ps_low = generate_array(
+        [&](auto i) { return InDstr::ps_to_rhss_major_[i].Size(); }, Number<ndim_p>{});
+
+    // is_rh_major_in_for_reduce
+    Array<bool, ndim_rh_major_in> is_rh_major_in_for_reduce{false};
+
+    for(index_t i = 0; i < reduce_dim_xs_in.Size(); i++)
+    {
+        index_t rh_major = reduce_dim_xs_in[i] + 1;
+
+        is_rh_major_in_for_reduce(rh_major) = true;
+    }
+
+    // is_y_in_for_reduce
+    Array<bool, ndim_y_in> is_y_in_for_reduce{false};
+
+    for(index_t i = 0; i < ndim_y_in; i++)
+    {
+        index_t rh_major = InDstr::ys_to_rhs_major_[i];
+
+        if(is_rh_major_in_for_reduce[rh_major])
+        {
+            is_y_in_for_reduce(i) = true;
+        }
+    }
+
+    // is_rh_minor_in_for_y_reduce
+    Array<Array<bool, max_ndim_rh_minor_in>, ndim_rh_major_in> is_rh_minor_in_for_y_reduce{{false}};
+
+    static_for<0, ndim_y_in, 1>{}([&](auto i) {
+        index_t rh_major = InDstr::ys_to_rhs_major_[i];
+        index_t rh_minor = InDstr::ys_to_rhs_minor_[i];
+
+        if(is_y_in_for_reduce[i])
+        {
+            is_rh_minor_in_for_y_reduce(rh_major)(rh_minor) = true;
+        }
+    });
+
+    // in2out_rh_major
+    Array<index_t, ndim_rh_major_in> in2out_rh_major{-1};
+    index_t cnt_ndim_rh_major_out = 0;
+
+    for(index_t i = 0; i < ndim_rh_major_in; i++)
+    {
+        if(is_rh_major_in_for_reduce[i])
+        {
+            in2out_rh_major(i) = 0;
+        }
+        else
+        {
+            in2out_rh_major(i) = cnt_ndim_rh_major_out;
+
+            cnt_ndim_rh_major_out++;
+        }
+    }
+
+    // rs_lengths_out, in2out_rh_minor
+    Array<index_t, max_ndim_r_out> rs_lengths_out{-1};
+    Array<Array<index_t, max_ndim_rh_minor_in>, ndim_rh_major_in> in2out_rh_minor{{-1}};
+
+    // loop over input R dim
+    for(index_t i = 0; i < InDstr::rs_lengths_.Size(); i++)
+    {
+        // rs_lengths_out
+        rs_lengths_out(i) = InDstr::rs_lengths_[i];
+
+        // in2out_rh_minor
+        in2out_rh_minor(0)(i) = i;
+    }
+
+    // loop over input H Dim
+    index_t cnt_ndim_r_out = InDstr::rs_lengths_.Size();
+
+    static_for<1, ndim_rh_major_in, 1>{}([&](auto rh_major_in) {
+        constexpr auto h_major_in = rh_major_in - I1;
+
+        constexpr index_t ndim_rh_minor_in = InDstr::hs_lengthss_[h_major_in].Size();
+
+        if(is_rh_major_in_for_reduce[rh_major_in])
+        {
+            for(index_t rh_minor_in = 0; rh_minor_in < ndim_rh_minor_in; rh_minor_in++)
+            {
+                if(not is_rh_minor_in_for_y_reduce[rh_major_in][rh_minor_in])
+                {
+                    // rs_lengths_out
+                    rs_lengths_out(cnt_ndim_r_out) = InDstr::hs_lengthss_[h_major_in][rh_minor_in];
+
+                    // in2out_rh_minor
+                    in2out_rh_minor(rh_major_in)(rh_minor_in) = cnt_ndim_r_out;
+
+                    cnt_ndim_r_out++;
+                }
+            }
+        }
+        else
+        {
+            for(index_t rh_minor_in = 0; rh_minor_in < ndim_rh_minor_in; rh_minor_in++)
+            {
+                // in2out_rh_minor
+                in2out_rh_minor(rh_major_in)(rh_minor_in) = rh_minor_in;
+            }
+        }
+    });
+
+    // ndim_r_out
+    const index_t ndim_r_out = cnt_ndim_r_out;
+
+    // ndims_hs_minor_out, hs_lengthss_out
+    Array<index_t, ndim_x_out> ndims_hs_minor_out{-1};
+    Array<Array<index_t, max_ndim_rh_minor_in>, ndim_x_out> hs_lengthss_out{{-1}};
+
+    index_t cnt_ndim_x_out = 0;
+
+    static_for<0, ndim_x_in, 1>{}([&](auto i) {
+        if(not is_rh_major_in_for_reduce[i + I1])
+        {
+            // ndims_hs_minor_out
+            ndims_hs_minor_out(cnt_ndim_x_out) = InDstr::hs_lengthss_[i].Size();
+
+            // hs_lengthss_out
+            static_for<0, InDstr::hs_lengthss_[i].Size(), 1>{}(
+                [&](auto j) { hs_lengthss_out(cnt_ndim_x_out)(j) = InDstr::hs_lengthss_[i][j]; });
+
+            cnt_ndim_x_out++;
+        }
+    });
+
+    // ps_to_rhss_major_out, ps_to_rhss_minor_out
+    Array<Array<index_t, max_ndim_rh_minor_in>, ndim_p> ps_to_rhss_major_out{{-1}};
+    Array<Array<index_t, max_ndim_rh_minor_in>, ndim_p> ps_to_rhss_minor_out{{-1}};
+
+    static_for<0, ndim_p, 1>{}([&](auto idim_p) {
+        static_for<0, InDstr::ps_to_rhss_major_[idim_p].Size(), 1>{}([&](auto idim_low) {
+            index_t rh_major_in = InDstr::ps_to_rhss_major_[idim_p][idim_low];
+            index_t rh_minor_in = InDstr::ps_to_rhss_minor_[idim_p][idim_low];
+
+            ps_to_rhss_major_out(idim_p)(idim_low) = in2out_rh_major[rh_major_in];
+            ps_to_rhss_minor_out(idim_p)(idim_low) = in2out_rh_minor[rh_major_in][rh_minor_in];
+        });
+    });
+
+    // ys_to_rhs_major_out, ys_to_rhs_minor_out
+    Array<index_t, max_ndim_y_out> ys_to_rhs_major_out{-1};
+    Array<index_t, max_ndim_y_out> ys_to_rhs_minor_out{-1};
+
+    index_t cnt_ndim_y_out = 0;
+
+    static_for<0, ndim_y_in, 1>{}([&](auto i) {
+        if(not is_y_in_for_reduce[i])
+        {
+            index_t rh_major_in = InDstr::ys_to_rhs_major_[i];
+            index_t rh_minor_in = InDstr::ys_to_rhs_minor_[i];
+
+            ys_to_rhs_major_out(cnt_ndim_y_out) = in2out_rh_major[rh_major_in];
+            ys_to_rhs_minor_out(cnt_ndim_y_out) = in2out_rh_minor[rh_major_in][rh_minor_in];
+
+            cnt_ndim_y_out++;
+        }
+    });
+
+    // ndim_y_out
+    const index_t ndim_y_out = cnt_ndim_y_out;
+
+    //
+    return make_tuple(ndim_x_out,
+                      ndim_p,
+                      ndim_y_out,
+                      ndim_r_out,
+                      ndims_hs_minor_out,
+                      ndims_ps_low,
+                      rs_lengths_out,
+                      hs_lengthss_out,
+                      ps_to_rhss_major_out,
+                      ps_to_rhss_minor_out,
+                      ys_to_rhs_major_out,
+                      ys_to_rhs_minor_out);
+}
+
+template <typename InDstr, index_t... InReduceDimXs>
+__host__ __device__ constexpr auto
+make_reduce_tile_distribution_encoding(InDstr, Sequence<InReduceDimXs...> reduce_dim_xs_in)
+{
+    constexpr auto impl = make_reduce_tile_distribution_encoding_impl(InDstr{}, reduce_dim_xs_in);
+
+    constexpr index_t ndim_x             = impl.template At<0>();
+    constexpr index_t ndim_p             = impl.template At<1>();
+    constexpr index_t ndim_y             = impl.template At<2>();
+    constexpr index_t ndim_r             = impl.template At<3>();
+    constexpr auto ndims_hs_minor        = impl.template At<4>();
+    constexpr auto ndims_ps_low          = impl.template At<5>();
+    constexpr auto rs_lengths_impl       = impl.template At<6>();
+    constexpr auto hs_lengthss_impl      = impl.template At<7>();
+    constexpr auto ps_to_rhss_major_impl = impl.template At<8>();
+    constexpr auto ps_to_rhss_minor_impl = impl.template At<9>();
+    constexpr auto ys_to_rhs_major_impl  = impl.template At<10>();
+    constexpr auto ys_to_rhs_minor_impl  = impl.template At<11>();
+
+    constexpr auto rs_lengths  = TO_SEQUENCE(rs_lengths_impl, ndim_r);
+    constexpr auto hs_lengthss = TO_TUPLE_OF_SEQUENCE(hs_lengthss_impl, ndim_x, ndims_hs_minor);
+    constexpr auto ps_to_rhss_major =
+        TO_TUPLE_OF_SEQUENCE(ps_to_rhss_major_impl, ndim_p, ndims_ps_low);
+    constexpr auto ps_to_rhss_minor =
+        TO_TUPLE_OF_SEQUENCE(ps_to_rhss_minor_impl, ndim_p, ndims_ps_low);
+    constexpr auto ys_to_rhs_major = TO_SEQUENCE(ys_to_rhs_major_impl, ndim_y);
+    constexpr auto ys_to_rhs_minor = TO_SEQUENCE(ys_to_rhs_minor_impl, ndim_y);
+
+    return StaticTileDistributionEncoding<remove_cvref_t<decltype(rs_lengths)>,
+                                          remove_cvref_t<decltype(hs_lengthss)>,
+                                          remove_cvref_t<decltype(ps_to_rhss_major)>,
+                                          remove_cvref_t<decltype(ps_to_rhss_minor)>,
+                                          remove_cvref_t<decltype(ys_to_rhs_major)>,
+                                          remove_cvref_t<decltype(ys_to_rhs_minor)>>{};
+}
+
+template <typename, typename, typename, index_t>
+struct reverse_slice_sequence_impl;
+
+template <index_t x,
+          index_t... xs,
+          index_t m,
+          index_t... ms,
+          index_t id,
+          index_t... ids,
+          index_t SliceSize>
+struct reverse_slice_sequence_impl<Sequence<x, xs...>,
+                                   Sequence<m, ms...>,
+                                   Sequence<id, ids...>,
+                                   SliceSize>
+{
+    using old_scan =
+        reverse_slice_sequence_impl<Sequence<xs...>, Sequence<ms...>, Sequence<ids...>, SliceSize>;
+
+    static constexpr auto slice_size = old_scan::remaining_slice_sizes::Front().value;
+    static constexpr auto slice_length =
+        std::conditional_t<m, Number<math::gcd(x, slice_size)>, Number<x>>::value;
+
+    using dim_lengths =
+        typename sequence_merge<Sequence<slice_length>, typename old_scan::dim_lengths>::type;
+    using dim_slices =
+        typename sequence_merge<Sequence<x / slice_length>, typename old_scan::dim_slices>::type;
+    using remaining_slice_sizes = typename sequence_merge<
+        std::conditional_t<m, Sequence<slice_size / slice_length>, Sequence<slice_size>>,
+        typename old_scan::remaining_slice_sizes>::type;
+
+    // the first idx that sliced length not equal to original length
+    static constexpr index_t _flag =
+        slice_length != x && remaining_slice_sizes{}.Front().value == 1;
+    static constexpr index_t _split_flag = std::conditional_t<m, Number<_flag>, Number<0>>::value;
+    static constexpr index_t _split_idx =
+        std::conditional_t<_split_flag, Number<id>, Number<0>>::value;
+
+    static constexpr index_t split_flag = _split_flag || old_scan::split_flag;
+    static constexpr index_t split_idx  = std::
+        conditional_t<old_scan::split_flag, Number<old_scan::split_idx>, Number<_split_idx>>::value;
+};
+
+template <index_t x, index_t m, index_t id, index_t SliceSize>
+struct reverse_slice_sequence_impl<Sequence<x>, Sequence<m>, Sequence<id>, SliceSize>
+{
+    static constexpr auto slice_size = SliceSize;
+    static constexpr auto slice_length =
+        std::conditional_t<m, Number<math::gcd(x, slice_size)>, Number<x>>::value;
+
+    using dim_lengths = Sequence<slice_length>;
+    using dim_slices  = Sequence<x / slice_length>;
+    using remaining_slice_sizes =
+        std::conditional_t<m, Sequence<slice_size / slice_length>, Sequence<slice_size>>;
+
+    // the first idx that sliced length not equal to original length
+    static constexpr index_t _flag =
+        slice_length != x && remaining_slice_sizes{}.Front().value == 1;
+    static constexpr index_t split_flag = std::conditional_t<m, Number<_flag>, Number<0>>::value;
+    static constexpr index_t split_idx =
+        std::conditional_t<split_flag, Number<id>, Number<0>>::value;
+};
+
+// clang-format off
+// input a sequence(with optional mask), and the SliceSize : size per slice
+// output the sequence each slice, and Number of slices
+//
+// e.g. <2, 1, 4, 2>, 8     -> lengths:<1, 1, 4, 2>    , nums: <2, 1, 1, 1>    : 2 slices  , slice_idx: 0
+//      <4, 2, 4, 1, 2>, 4  -> lengths:<1, 1, 2, 1, 2> , nums: <4, 2, 2, 1, 1> : 16 slices , slice_idx: 2
+//      <4, 2, 4, 1, 6>, 4  -> lengths:<1, 1, 2, 1, 2> , nums: <4, 2, 2, 1, 3> : 48 slices , slice_idx: 2
+//      <4, 2, 5, 1, 2>, 10 -> lengths:<1, 1, 5, 1, 2> , nums: <4, 2, 1, 1, 1> : 8 slices  , slice_idx: 1
+//
+//      <4, 2, 8>, 64       -> lengths:<4, 2, 8>       , nums: <1, 1, 1>       : 1  slices , slice_idx: 0
+//      <4, 2, 8>, 32       -> lengths:<2, 2, 8>       , nums: <2, 1, 1>       : 2  slices , slice_idx: 0
+//      <4, 2, 8>, 16       -> lengths:<1, 2, 8>       , nums: <4, 1, 1>       : 4  slices , slice_idx: 0
+//      <4, 2, 8>, 8        -> lengths:<1, 1, 8>       , nums: <4, 2, 1>       : 8  slices , slice_idx: 1
+//      <4, 2, 8>, 4        -> lengths:<1, 1, 4>       , nums: <4, 2, 2>       : 16 slices , slice_idx: 2
+//      <4, 2, 8>, 2        -> lengths:<1, 1, 2>       , nums: <4, 2, 4>       : 32 slices , slice_idx: 2
+//      <4, 2, 8>, 1        -> lengths:<1, 1, 1>       , nums: <4, 2, 8>       : 64 slices , slice_idx: 2
+//
+//      <4, 2, 1, 4, 2> / 4 ->
+// mask:<1, 1, 1, 0, 1>,    -> lengths:<1, 2, 1, 4, 2> , nums: <4, 1, 1, 1, 1> : 8 slices  , slice_idx: 0
+//
+// return Tuple<slice_lengths, slice_nums, slice_index>, slice_index is at which index will start
+// have split slices (right -> left)
+//  or the first index that sliced length is different from the original length
+// clang-format on
+template <typename Seq,
+          index_t SliceSize,
+          typename Mask = typename uniform_sequence_gen<Seq::Size(), 1>::type>
+constexpr auto reverse_slice_sequence(Seq,
+                                      Number<SliceSize>,
+                                      Mask = typename uniform_sequence_gen<Seq::Size(), 1>::type{})
+{
+    static_assert(Seq::Size() == Mask::Size());
+    using sliced_type =
+        reverse_slice_sequence_impl<Seq,
+                                    Mask,
+                                    typename arithmetic_sequence_gen<0, Seq::Size(), 1>::type,
+                                    SliceSize>;
+    static_assert(sliced_type::remaining_slice_sizes::Front().value == 1,
+                  "can not evenly divide this sequence, please check");
+    return make_tuple(typename sliced_type::dim_lengths{},
+                      typename sliced_type::dim_slices{},
+                      Number<sliced_type::split_idx>{});
+}
+
+//
+// slice tensor from x_dim, result in split in y_dim, not p_dim.
+// We don't support slice cross p_dim (aka, slice different threads)
+// also, sliced along y_dim need be the first dim of current dim.
+// Multiply Y dim before sliced dim does not make sense
+//
+// e.g
+//       X0           X1
+//       <1, 4, 32> - <4, 1, 4, 2, 4>  | slice origin:<0, 0>, len:<0, 32>, (0 means all length)
+//        Y  P  P      Y  P  Y  P  Y
+//   =>  <1, 4, 32> - <1, 1, 4, 2, 4> -> OK
+//                     |--> slice along this Y dim, is the first dim of X1, totally 4 slices
+//
+//       X0           X1
+//       <1, 4, 32> - <4, 1, 4, 2, 4>  | slice origin:<0, 0>, len:<0, 8>, (0 means all length)
+//        Y  P  P      Y  P  Y  P  Y
+//   =>  <1, 4, 32> - <1, 1, 1, 2, 4> -> OK
+//                           |--> slice along this Y dim, the P dim is 1 in the left, so is OK
+//                                 totally 16 slices
+//
+//       X0           X1
+//       <1, 4, 32> - <4, 1, 4, 2, 4>  | slice origin:<0, 0>, len:<0, 4>, (0 means all length)
+//        Y  P  P      Y  P  Y  P  Y
+//   =>  <1, 4, 32> - <1, 1, 1, 1, 4> -> Fail
+//                              |--> slice along this P dim, will split threads, not supported
+//
+//       X0           X1
+//       <1, 4, 32> - <4, 1, 4, 2, 4>  | slice origin:<0, 0>, len:<0, 16>, (0 means all length)
+//        Y  P  P      Y  P  Y  P  Y
+//   =>  <1, 4, 32> - <1, 1, 2, 2, 4> -> OK
+//                           |--> slice along this Y dim, but this Y sim need to split into 2
+//                           subdime
+//                                the P dim in the left is 1, means actually not crossing P
+//
+template <typename Distribution, index_t... XSliceBegins, index_t... XSliceEnds>
+__host__ __device__ constexpr auto slice_distribution_from_x(
+    Distribution, Sequence<XSliceBegins...> x_slice_begins, Sequence<XSliceEnds...> x_slice_ends)
+{
+    // NOTE: this function need to be called under constexpr context,
+    // due to https://wg21.link/p2280r0 we have to use non-reference type for distribution
+    using Encoding = decltype(Distribution::GetStaticTileDistributionEncoding());
+
+    static_assert(sizeof...(XSliceBegins) == sizeof...(XSliceEnds));
+
+    constexpr auto x_slice_lengths = x_slice_ends - x_slice_begins;
+
+    constexpr auto src_h_prefix_sum = Encoding::Detail::GetHDimLengthsPrefixSum();
+    constexpr auto src_y_info       = Encoding::Detail::GetSortedYInfo();
+    constexpr auto src_y_dims       = src_y_info[Number<0>{}];
+    constexpr auto src_y_maps       = src_y_info[Number<1>{}];
+    constexpr auto src_y_prefix_sum = src_y_info[Number<2>{}];
+
+    constexpr auto sliced_hlen_yidx_ylen = [&]() constexpr
+    {
+        auto y_slice_sorted_origins = make_zero_multi_index<Encoding::NDimY>();
+        auto y_slice_lengths        = Encoding::Detail::ys_lengths_;
+
+        // This lambda will modify some value outside, so c++ will not treat return value as
+        // constexpr
+        // TODO: ugly
+        auto new_h_lengths = transform_tuples(
+            [&](auto h_len, auto id) {
+                constexpr auto sliced_h =
+                    reverse_slice_sequence(h_len, Number<x_slice_lengths[id]>{});
+
+                constexpr auto sliced_h_lens  = sliced_h[Number<0>{}];
+                constexpr auto sliced_h_index = sliced_h[Number<2>{}];
+
+                // update y_slice_lengths
+                constexpr auto uniformed_h_index = sliced_h_index + Number<src_h_prefix_sum[id]>{};
+                constexpr auto found_y_index     = container_find(src_y_dims, uniformed_h_index);
+
+                static_assert(found_y_index >= 0 && found_y_index < src_y_dims.Size(),
+                              "not sliced at y dim, please check");
+
+                static_for<0, sliced_h_index + 1, 1>{}([&](auto i) {
+                    y_slice_lengths(src_y_maps[found_y_index - i]) =
+                        sliced_h_lens[sliced_h_index - i];
+                });
+                // TODO: add validations not across p dim
+
+                // NOTE: this y_origin is for all dims, not only current dim
+                //       will later use pick to select target dim
+                constexpr auto y_origin = [&]() {
+                    constexpr auto h_trans = make_merge_transform_v3_division_mod(h_len);
+                    auto h_origin_         = make_zero_multi_index<h_trans.NDimLow>();
+                    h_trans.CalculateLowerIndex(h_origin_, Sequence<x_slice_begins[id].value>{});
+
+                    auto y_origin_ = make_zero_multi_index<Encoding::NDimY>();
+                    static_for<0, sliced_h_index + 1, 1>{}([&](auto i) {
+                        y_origin_(found_y_index - i) = h_origin_[sliced_h_index - i];
+                    });
+                    return y_origin_;
+                }();
+
+                constexpr auto y_picks = typename arithmetic_sequence_gen<src_y_prefix_sum[id],
+                                                                          src_y_prefix_sum[id + 1],
+                                                                          1>::type{};
+
+                set_container_subset(
+                    y_slice_sorted_origins, y_picks, get_container_subset(y_origin, y_picks));
+                return sliced_h_lens;
+            },
+            typename Encoding::HsLengthss{},
+            typename arithmetic_sequence_gen<0, Encoding::HsLengthss::Size(), 1>::type{});
+
+        auto y_slice_origins = container_reorder_given_old2new(y_slice_sorted_origins, src_y_maps);
+
+        return make_tuple(new_h_lengths, y_slice_origins, y_slice_lengths);
+    }
+    ();
+
+    constexpr auto sliced_h_lengths       = sliced_hlen_yidx_ylen[Number<0>{}];
+    constexpr auto sliced_y_origins_array = sliced_hlen_yidx_ylen[Number<1>{}];
+    constexpr auto sliced_y_origins_size  = sliced_y_origins_array.Size();
+    constexpr auto sliced_y_lengths_array = sliced_hlen_yidx_ylen[Number<2>{}];
+    constexpr auto sliced_y_lengths_size  = sliced_y_lengths_array.Size();
+
+    constexpr auto sliced_y_origins = TO_SEQUENCE(sliced_y_origins_array, sliced_y_origins_size);
+    constexpr auto sliced_y_lengths = TO_SEQUENCE(sliced_y_lengths_array, sliced_y_lengths_size);
+
+    return make_tuple(
+        make_static_tile_distribution(
+            StaticTileDistributionEncoding<typename Encoding::RsLengths,
+                                           decltype(sliced_h_lengths), // only need to change the
+                                                                       // h_lengths type
+                                           typename Encoding::Ps2RHssMajor,
+                                           typename Encoding::Ps2RHssMinor,
+                                           typename Encoding::Ys2RHsMajor,
+                                           typename Encoding::Ys2RHsMinor>{}),
+        sliced_y_origins,
+        sliced_y_lengths);
+}
+
+} // namespace detail
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/store_tile.hpp b/include/ck/tile_program/tile/store_tile.hpp
new file mode 100644
index 0000000000..974e8d0822
--- /dev/null
+++ b/include/ck/tile_program/tile/store_tile.hpp
@@ -0,0 +1,15 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/store_tile_impl_static_distribution.hpp"
+#include "ck/tile_program/tile/store_tile_impl_static_lengths.hpp"
+#include "ck/tile_program/tile/store_tile_impl_static_tensor.hpp"
diff --git a/include/ck/tile_program/tile/store_tile_impl_static_distribution.hpp b/include/ck/tile_program/tile/store_tile_impl_static_distribution.hpp
new file mode 100644
index 0000000000..a026efb0e8
--- /dev/null
+++ b/include/ck/tile_program/tile/store_tile_impl_static_distribution.hpp
@@ -0,0 +1,32 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+
+namespace ck {
+namespace tile_program {
+
+template <typename BottomTensorView_,
+          typename WindowLengths_,
+          typename TileDistribution_,
+          index_t NumCoord,
+          typename DataType_>
+__device__ void store_tile(TileWindowWithStaticDistribution<BottomTensorView_,
+                                                            WindowLengths_,
+                                                            TileDistribution_,
+                                                            NumCoord>& tile_window,
+                           const StaticDistributedTensor<DataType_, TileDistribution_>& dstr_tensor)
+{
+    tile_window.Store(dstr_tensor);
+}
+
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/store_tile_impl_static_lengths.hpp b/include/ck/tile_program/tile/store_tile_impl_static_lengths.hpp
new file mode 100644
index 0000000000..5d25c5e49b
--- /dev/null
+++ b/include/ck/tile_program/tile/store_tile_impl_static_lengths.hpp
@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_window.hpp"
+
+namespace ck {
+namespace tile_program {
+
+template <typename BottomTensorView_,
+          typename WindowLengths_,
+          typename TileDistribution_,
+          typename DataType_>
+__device__ void
+store_tile(TileWindowWithStaticLengths<BottomTensorView_, WindowLengths_>& tile_window_tmp,
+           const StaticDistributedTensor<DataType_, TileDistribution_>& dstr_tensor)
+{
+    using DataType = remove_cvref_t<typename BottomTensorView_::DataType>;
+    using TileDstr = remove_cvref_t<TileDistribution_>;
+
+    static_assert(is_same_v<remove_cvref_t<DataType_>, DataType>, "wrong!");
+
+    constexpr auto tile_dstr = TileDstr{};
+
+    auto tile_window = make_tile_window(tile_window_tmp.GetBottomTensorView(),
+                                        tile_window_tmp.GetWindowLengths(),
+                                        tile_window_tmp.GetWindowOrigin(),
+                                        tile_dstr);
+
+    tile_window.Store(dstr_tensor);
+}
+
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/store_tile_impl_static_tensor.hpp b/include/ck/tile_program/tile/store_tile_impl_static_tensor.hpp
new file mode 100644
index 0000000000..4407df6bd2
--- /dev/null
+++ b/include/ck/tile_program/tile/store_tile_impl_static_tensor.hpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_window.hpp"
+
+namespace ck {
+namespace tile_program {
+
+template <typename SrcTileDistribution_, typename DstTileDistribution_, typename DataType_>
+__device__ void
+store_tile(StaticDistributedTensor<DataType_, DstTileDistribution_>& dst_dstr_tensor,
+           const StaticDistributedTensor<DataType_, SrcTileDistribution_>& src_dstr_tensor)
+{
+    // static_assert(DstTileDistribution_==SrcTileDistribution_);
+    dst_dstr_tensor.GetThreadBuffer() = src_dstr_tensor.GetThreadBuffer();
+}
+
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/tile_distribution.hpp b/include/ck/tile_program/tile/tile_distribution.hpp
new file mode 100644
index 0000000000..f7109df47c
--- /dev/null
+++ b/include/ck/tile_program/tile/tile_distribution.hpp
@@ -0,0 +1,490 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_description/macro_func_tensor_adaptor_from_encoding.hpp"
+
+#include "ck/tile_program/tile/static_tile_distribution_encoding.hpp"
+
+namespace ck {
+namespace tile_program {
+
+// distributed span
+template <index_t... PartialHsLengths>
+struct TileDistributedSpan
+{
+    using Impl = Sequence<PartialHsLengths...>;
+
+    static constexpr auto impl_ = Impl{};
+
+    __host__ __device__ static constexpr bool IsStatic() { return true; }
+};
+
+// distributed index
+template <index_t... PartialHsIndices>
+struct TileDistributedIndex
+{
+    using Impl = Sequence<PartialHsIndices...>;
+
+    static constexpr auto impl_ = Impl{};
+
+    __host__ __device__ static constexpr bool IsStatic() { return true; }
+};
+
+namespace detail {
+
+template <index_t... Is>
+__host__ __device__ constexpr auto make_tile_distributed_span(Sequence<Is...>)
+{
+    return TileDistributedSpan<Is...>{};
+}
+
+template <index_t... Is>
+__host__ __device__ constexpr auto make_tile_distributed_index(Sequence<Is...>)
+{
+    return TileDistributedIndex<Is...>{};
+}
+
+} // namespace detail
+
+template <typename PsYs2XsAdaptor_,
+          typename Ys2DDescriptor_,
+          typename StaticTileDistributionEncoding_,
+          typename TileDistributionDetail_> // FIXME: this is for hold ad-hoc but useful info,
+                                            // should be more elegnat
+struct TileDistribution
+{
+    using PsYs2XsAdaptor = remove_cvref_t<PsYs2XsAdaptor_>;
+    using Ys2DDescriptor = remove_cvref_t<Ys2DDescriptor_>;
+    using DstrEncode     = remove_cvref_t<StaticTileDistributionEncoding_>;
+    using DstrDetail     = remove_cvref_t<TileDistributionDetail_>;
+
+    static_assert(PsYs2XsAdaptor::IsStatic() && Ys2DDescriptor::IsStatic(),
+                  "wrong! should be static");
+
+    static constexpr index_t NDimX = PsYs2XsAdaptor::GetNumOfBottomDimension();
+    static constexpr index_t NDimY = Ys2DDescriptor::GetNumOfTopDimension();
+    static constexpr index_t NDimP = PsYs2XsAdaptor::GetNumOfTopDimension() - NDimY;
+    static constexpr index_t NDimR = StaticTileDistributionEncoding_::NDimR;
+
+    PsYs2XsAdaptor ps_ys_to_xs_;
+    Ys2DDescriptor ys_to_d_;
+
+    __host__ __device__ static constexpr index_t GetNumOfDimensionX() { return NDimX; }
+    __host__ __device__ static constexpr index_t GetNumOfDimensionY() { return NDimY; }
+    __host__ __device__ static constexpr index_t GetNumOfDimensionP() { return NDimP; }
+    __host__ __device__ static constexpr index_t GetNumOfDimensionR() { return NDimR; }
+
+    __host__ __device__ static constexpr auto GetLengths()
+    {
+#if 0
+        // FIXME: TensorAdaptor::GetBottomDimensionLengths is wrong. re-enable this after it's fixed
+        ps_ys_to_xs_.GetBottomDimensionLengths();
+#else
+        return generate_tuple(
+            [&](auto i) {
+                constexpr index_t x_length =
+                    container_reduce(typename DstrEncode::HsLengthss{}[i], math::multiplies{}, 1);
+
+                return Number<x_length>{};
+            },
+            Number<NDimX>{});
+#endif
+    }
+
+    __host__ __device__ constexpr const auto& GetPsYs2XsAdaptor() const { return ps_ys_to_xs_; }
+
+    __host__ __device__ constexpr const auto& GetYs2DDescriptor() const { return ys_to_d_; }
+
+    __host__ __device__ static constexpr auto GetStaticTileDistributionEncoding()
+    {
+        return DstrEncode{};
+    }
+
+#if 1
+    // Calculate Replication index [R0, R1, ...] based on Partion index
+    // FIXME: very nasty implementation
+    template <typename PartitionIndex>
+    __host__ __device__ auto CalculateRsIndexFromPsIndex(const PartitionIndex& ps_idx) const
+    {
+        static_assert(PartitionIndex::Size() == NDimP, "wrong!");
+
+        const auto ps_ys_idx = container_concat(ps_idx, Array<index_t, NDimY>{0});
+
+        const auto dummy_adaptor_coord = make_tensor_adaptor_coordinate(ps_ys_to_xs_, ps_ys_idx);
+
+        Array<index_t, NDimR> rs_idx;
+
+        static_for<0, NDimP, 1>{}([&](auto idim_p) {
+            constexpr index_t ndim_low = DstrEncode::ps_to_rhss_major_[idim_p].Size();
+
+            static_for<0, ndim_low, 1>{}([&](auto i) {
+                constexpr index_t rh_major = DstrEncode::ps_to_rhss_major_[idim_p][i];
+                constexpr index_t rh_minor = DstrEncode::ps_to_rhss_minor_[idim_p][i];
+
+                // 0-th rh_major is the replicate dimension
+                if constexpr(rh_major == 0)
+                {
+                    constexpr index_t adaptor_hidden_id =
+                        DstrDetail::rh_major_minor_to_adaptor_hidden_idss_[rh_major][rh_minor];
+
+                    // fill in
+                    rs_idx(rh_minor) = dummy_adaptor_coord.GetHiddenIndex()[adaptor_hidden_id];
+                }
+            });
+        });
+
+        return rs_idx;
+    }
+#endif
+
+    __host__ __device__ static constexpr auto GetDistributedSpans()
+    {
+        constexpr auto distributed_spans_impl = DstrEncode::Detail::distributed_spans_lengthss_;
+        constexpr auto ndims_spans_minor      = DstrEncode::Detail::ndims_distributed_spans_minor_;
+
+        return generate_tuple(
+            [&](auto i) {
+                constexpr auto span_impl          = distributed_spans_impl[i];
+                constexpr index_t ndim_span_minor = ndims_spans_minor[i];
+
+                constexpr auto span = TO_SEQUENCE(span_impl, ndim_span_minor);
+
+                return detail::make_tile_distributed_span(span);
+            },
+            Number<NDimX>{});
+    }
+
+    // FIXME: it's hacky to get Y index from Distributed-Index
+    template <typename DistributedIndices>
+    __host__ __device__ static constexpr auto GetYIndicesFromDistributedIndices(DistributedIndices)
+    {
+        constexpr auto ys_idx_arr = [] {
+            Array<index_t, NDimY> ys_idx;
+
+            static_for<0, NDimY, 1>{}([&](auto i) {
+                constexpr index_t span_major = DstrEncode::Detail::ys_to_span_major_[i];
+                constexpr index_t span_minor = DstrEncode::Detail::ys_to_span_minor_[i];
+
+                constexpr auto dstr_index = DistributedIndices{}[Number<span_major>{}];
+
+                ys_idx(i) = dstr_index.impl_[span_minor];
+            });
+
+            return ys_idx;
+        }();
+
+        constexpr index_t ndim_y = NDimY;
+
+        return TO_SEQUENCE(ys_idx_arr, ndim_y);
+    }
+
+    __host__ __device__ static constexpr bool IsStatic()
+    {
+        return PsYs2XsAdaptor::IsStatic() && Ys2DDescriptor::IsStatic();
+    }
+
+    __host__ __device__ void Print() const
+    {
+        printf("TileDistribution{");
+        //
+        printf("StaticTileDistributionEncoding: ");
+        print(DstrEncode{});
+        printf(", ");
+        //
+        printf("ps_ys_to_xs_: ");
+        print(ps_ys_to_xs_);
+        printf(", ");
+        //
+        printf("ys_to_d_: ");
+        print(ys_to_d_);
+        //
+        printf("}");
+    }
+};
+
+namespace detail {
+
+template <index_t NDimMax>
+__host__ __device__ constexpr auto make_sequential_index(index_t ibegin, index_t iend)
+{
+    Array<index_t, NDimMax> arr{0};
+
+    for(index_t i = 0; i < iend - ibegin; ++i)
+    {
+        arr(i) = ibegin + i;
+    }
+
+    return arr;
+}
+
+// this returns a constexpr encoding of TileDistribution
+template <typename StaticTileDistributionEncoding_>
+__host__ __device__ constexpr auto
+    make_adaptor_encoding_for_tile_distribution(StaticTileDistributionEncoding_)
+{
+    using RsLengths    = typename StaticTileDistributionEncoding_::RsLengths;
+    using HsLengthss   = typename StaticTileDistributionEncoding_::HsLengthss;
+    using Ps2RHssMajor = typename StaticTileDistributionEncoding_::Ps2RHssMajor;
+    using Ps2RHssMinor = typename StaticTileDistributionEncoding_::Ps2RHssMinor;
+    using Ys2RHsMajor  = typename StaticTileDistributionEncoding_::Ys2RHsMajor;
+    using Ys2RHsMinor  = typename StaticTileDistributionEncoding_::Ys2RHsMinor;
+
+    // FIXME: increase max value if fail
+    constexpr index_t kMaxNumTransforms = 20;
+    constexpr index_t kMaxMetaDataSize  = 128;
+    constexpr index_t kMaxNumDim        = 10;
+
+    using Name     = IndexTransformEnum;
+    using MetaData = MetaDataBuffer<kMaxMetaDataSize>;
+    using NumDim   = index_t;
+    using Dims     = Array<index_t, kMaxNumDim>;
+    using Lengths  = Array<index_t, kMaxNumDim>;
+
+    // Tile Adaptor
+    //   bottom dims [x0, x1, x2, ...]
+    //   top dims [p0, p1, ..., y0, y1, ...]
+    constexpr index_t ndim_x = HsLengthss::Size();
+
+    // Dim Ids: [idim_x_major, idim_x_minor] to [idim_hidden]
+    Array<Array<index_t, kMaxNumDim>, ndim_x + 1> rh_major_minor_to_hidden_ids;
+    Array<Array<index_t, kMaxNumDim>, ndim_x + 1> rh_major_minor_to_hidden_lengths;
+
+    auto trans = Array<Tuple<Name, MetaData, NumDim, Dims, NumDim, Dims>, kMaxNumTransforms>{};
+
+    index_t num_tran       = 0;
+    index_t hidden_dim_cnt = ndim_x;
+
+    // this is Replicate transform
+    {
+        constexpr index_t ndim_r_minor = RsLengths::Size();
+
+        constexpr auto r_minor_lengths = RsLengths{};
+
+        trans(num_tran++) = {
+            IndexTransformEnum::Replicate,
+            MetaData{to_array<index_t, ndim_r_minor>(r_minor_lengths)},
+            NumDim{0},
+            Dims{},
+            NumDim{ndim_r_minor},
+            make_sequential_index<kMaxNumDim>(hidden_dim_cnt, hidden_dim_cnt + ndim_r_minor)};
+
+        for(index_t i = 0; i < ndim_r_minor; ++i)
+        {
+            rh_major_minor_to_hidden_ids(0)(i)     = hidden_dim_cnt;
+            rh_major_minor_to_hidden_lengths(0)(i) = r_minor_lengths[i];
+
+            hidden_dim_cnt++;
+        }
+    };
+
+    // these are Unmerge transforms for X dimesions
+    static_for<0, ndim_x, 1>{}([&trans,
+                                &num_tran,
+                                &hidden_dim_cnt,
+                                &rh_major_minor_to_hidden_ids,
+                                &rh_major_minor_to_hidden_lengths](auto idim_x) {
+        constexpr auto h_minor_lengths = tuple_element_t<idim_x, HsLengthss>{};
+
+        constexpr index_t ndim_h_minor = h_minor_lengths.Size();
+
+        trans(num_tran++) = {
+            IndexTransformEnum::UnMerge,
+            MetaData{to_array<index_t, ndim_h_minor>(h_minor_lengths)},
+            NumDim{1},
+            Dims{idim_x},
+            NumDim{ndim_h_minor},
+            make_sequential_index<kMaxNumDim>(hidden_dim_cnt, hidden_dim_cnt + ndim_h_minor)};
+
+        for(index_t i = 0; i < ndim_h_minor; ++i)
+        {
+            rh_major_minor_to_hidden_ids(idim_x + 1)(i)     = hidden_dim_cnt;
+            rh_major_minor_to_hidden_lengths(idim_x + 1)(i) = h_minor_lengths[i];
+
+            hidden_dim_cnt++;
+        }
+    });
+
+    // transform: P dimensions
+    constexpr index_t ndim_p = Ps2RHssMajor::Size();
+
+    Dims hidden_dim_id_ps;
+
+    static_for<0, ndim_p, 1>{}([&](auto iDimP) {
+        //
+        index_t hidden_dim_id_p = hidden_dim_cnt++;
+
+        hidden_dim_id_ps(iDimP) = hidden_dim_id_p;
+
+        constexpr auto p2RHsMajor = Ps2RHssMajor{}[iDimP];
+        constexpr auto p2RHsMinor = Ps2RHssMinor{}[iDimP];
+
+        static_assert(p2RHsMajor.Size() == p2RHsMinor.Size(), "wrong!");
+
+        constexpr index_t ndim_low = p2RHsMajor.Size();
+
+        Dims low_dims;
+        Lengths low_lengths;
+
+        for(index_t i = 0; i < ndim_low; ++i)
+        {
+            index_t rh_major = p2RHsMajor[i];
+            index_t rh_minor = p2RHsMinor[i];
+            low_dims(i)      = rh_major_minor_to_hidden_ids[rh_major][rh_minor];
+            low_lengths(i)   = rh_major_minor_to_hidden_lengths[rh_major][rh_minor];
+        }
+
+        trans(num_tran++) = {IndexTransformEnum::Merge,
+                             MetaData{to_array<index_t, ndim_low>(low_lengths)},
+                             NumDim{ndim_low},
+                             low_dims,
+                             NumDim{1},
+                             Dims{hidden_dim_id_p}};
+    });
+
+    constexpr index_t ndim_bottom = ndim_x;
+
+    constexpr auto bottom_dim_ids = make_sequential_index<kMaxNumDim>(0, ndim_bottom);
+
+    constexpr auto ys_to_rhs_major = Ys2RHsMajor{};
+    constexpr auto ys_to_rhs_minor = Ys2RHsMinor{};
+
+    constexpr index_t ndim_y   = Ys2RHsMajor::Size();
+    constexpr index_t ndim_top = ndim_p + ndim_y;
+
+    auto top_dim_ids = hidden_dim_id_ps;
+
+    {
+        for(index_t i = 0; i < ndim_y; ++i)
+        {
+            index_t rh_major        = ys_to_rhs_major[i];
+            index_t rh_minor        = ys_to_rhs_minor[i];
+            top_dim_ids(ndim_p + i) = rh_major_minor_to_hidden_ids[rh_major][rh_minor];
+        }
+    }
+
+    //
+    const auto ps_ys_to_xs_adaptor_encoding =
+        make_tuple(trans, num_tran, bottom_dim_ids, ndim_bottom, top_dim_ids, ndim_top);
+
+    // descriptor: [y0, y1, ...] to [d]
+    Lengths y_lengths;
+    index_t d_length = 1;
+
+    for(index_t i = 0; i < ndim_y; ++i)
+    {
+        index_t rh_major = ys_to_rhs_major[i];
+        index_t rh_minor = ys_to_rhs_minor[i];
+        index_t y_length = rh_major_minor_to_hidden_lengths[rh_major][rh_minor];
+        y_lengths(i)     = y_length;
+        d_length *= y_length;
+    }
+
+    auto tran = make_tuple(IndexTransformEnum::UnMerge,
+                           MetaData{to_array<index_t, ndim_y>(y_lengths)},
+                           NumDim{1},
+                           Dims{0},
+                           NumDim{ndim_y},
+                           make_sequential_index<kMaxNumDim>(1, ndim_y + 1));
+
+    const auto ys_to_d_adaptor_encoding = make_tuple(
+        make_tuple(tran), 1, Dims{0}, 1, make_sequential_index<kMaxNumDim>(1, ndim_y + 1), ndim_y);
+
+    return make_tuple(ps_ys_to_xs_adaptor_encoding,
+                      ys_to_d_adaptor_encoding,
+                      d_length,
+                      rh_major_minor_to_hidden_ids);
+}
+
+// FIXME: this is nasty. Move it inside TileDistributionEncoding::Detail
+template <typename RhMajorMinor2AdaptorHiddenIdss> // Tuple<Sequence<...>, ...>
+struct TileDistributionDetail
+{
+    static constexpr auto rh_major_minor_to_adaptor_hidden_idss_ =
+        to_array_of_array(RhMajorMinor2AdaptorHiddenIdss{});
+};
+
+} // namespace detail
+
+// this returns a constexpr TileDistribution
+template <typename StaticTileDistributionEncoding_>
+__host__ __device__ constexpr auto make_tile_distribution(StaticTileDistributionEncoding_)
+{
+    using DstrEncode = remove_cvref_t<StaticTileDistributionEncoding_>;
+
+    constexpr auto adaptor_impl =
+        detail::make_adaptor_encoding_for_tile_distribution(StaticTileDistributionEncoding_{});
+
+    constexpr auto ps_ys_to_xs_adaptor_impl          = adaptor_impl.template At<0>();
+    constexpr auto ys_to_d_adaptor_impl              = adaptor_impl.template At<1>();
+    constexpr index_t d_length                       = adaptor_impl.template At<2>();
+    constexpr auto rh_major_minor_to_hidden_ids_impl = adaptor_impl.template At<3>();
+
+    constexpr auto ps_ys_to_xs_adaptor =
+        CONSTRUCT_TENSOR_ADAPTOR_FROM_ENCODING(ps_ys_to_xs_adaptor_impl);
+
+    constexpr auto ys_to_d_adaptor = CONSTRUCT_TENSOR_ADAPTOR_FROM_ENCODING(ys_to_d_adaptor_impl);
+
+    constexpr auto ys_to_d_descriptor =
+        make_tensor_descriptor_from_adaptor(ys_to_d_adaptor, d_length);
+
+    //
+    constexpr index_t ndim_rh_major = DstrEncode::Detail::ndim_rh_major_;
+    constexpr auto ndims_rhs_minor  = DstrEncode::Detail::ndims_rhs_minor_;
+
+    constexpr auto rh_major_minor_to_hidden_ids =
+        TO_TUPLE_OF_SEQUENCE(rh_major_minor_to_hidden_ids_impl, ndim_rh_major, ndims_rhs_minor);
+
+    return TileDistribution<
+        remove_cvref_t<decltype(ps_ys_to_xs_adaptor)>,
+        remove_cvref_t<decltype(ys_to_d_descriptor)>,
+        remove_cvref_t<DstrEncode>,
+        detail::TileDistributionDetail<remove_cvref_t<decltype(rh_major_minor_to_hidden_ids)>>>{
+        ps_ys_to_xs_adaptor, ys_to_d_descriptor};
+}
+
+// this returns a static TileDistribution
+template <typename StaticTileDistributionEncoding_>
+__host__ __device__ constexpr auto make_static_tile_distribution(StaticTileDistributionEncoding_)
+{
+    using DstrEncode = remove_cvref_t<StaticTileDistributionEncoding_>;
+
+    constexpr auto adaptor_impl =
+        detail::make_adaptor_encoding_for_tile_distribution(StaticTileDistributionEncoding_{});
+
+    constexpr auto ps_ys_to_xs_adaptor_impl          = adaptor_impl.template At<0>();
+    constexpr auto ys_to_d_adaptor_impl              = adaptor_impl.template At<1>();
+    constexpr index_t d_length                       = adaptor_impl.template At<2>();
+    constexpr auto rh_major_minor_to_hidden_ids_impl = adaptor_impl.template At<3>();
+
+    constexpr auto ps_ys_to_xs_adaptor =
+        CONSTRUCT_STATIC_TENSOR_ADAPTOR_FROM_ENCODING(ps_ys_to_xs_adaptor_impl);
+
+    constexpr auto ys_to_d_adaptor =
+        CONSTRUCT_STATIC_TENSOR_ADAPTOR_FROM_ENCODING(ys_to_d_adaptor_impl);
+
+    constexpr auto ys_to_d_descriptor =
+        make_tensor_descriptor_from_adaptor(ys_to_d_adaptor, Number<d_length>{});
+
+    //
+    constexpr index_t ndim_rh_major = DstrEncode::Detail::ndim_rh_major_;
+    constexpr auto ndims_rhs_minor  = DstrEncode::Detail::ndims_rhs_minor_;
+
+    constexpr auto rh_major_minor_to_hidden_ids =
+        TO_TUPLE_OF_SEQUENCE(rh_major_minor_to_hidden_ids_impl, ndim_rh_major, ndims_rhs_minor);
+
+    return TileDistribution<
+        remove_cvref_t<decltype(ps_ys_to_xs_adaptor)>,
+        remove_cvref_t<decltype(ys_to_d_descriptor)>,
+        remove_cvref_t<DstrEncode>,
+        detail::TileDistributionDetail<remove_cvref_t<decltype(rh_major_minor_to_hidden_ids)>>>{
+        ps_ys_to_xs_adaptor, ys_to_d_descriptor};
+}
+
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/tile_elementwise.hpp b/include/ck/tile_program/tile/tile_elementwise.hpp
new file mode 100644
index 0000000000..dedfc59611
--- /dev/null
+++ b/include/ck/tile_program/tile/tile_elementwise.hpp
@@ -0,0 +1,56 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/static_distributed_tensor.hpp"
+
+namespace ck {
+namespace tile_program {
+
+// TODO: support tensors with different distribution
+template <typename InOutElementFunc, typename... InOutDstrTensors>
+__device__ void tile_elementwise_inout(const InOutElementFunc& inout_element_func,
+                                       InOutDstrTensors&... inout_dstr_tensors)
+{
+    // TODO: make sure all distributed tensors have same lengths and distribution
+    // static_assert(xxx);
+
+    constexpr index_t thread_buffer_size =
+        type_pack_element<0, InOutDstrTensors...>::GetThreadBufferSize();
+
+    static_for<0, thread_buffer_size, 1>{}(
+        [&](auto i) { inout_element_func(inout_dstr_tensors.GetThreadBuffer().At(i)...); });
+}
+
+template <typename InElementFunc, typename... InDstrTensors>
+__device__ auto tile_elementwise_in(const InElementFunc& in_element_func,
+                                    const InDstrTensors&... in_dstr_tensors)
+{
+    using OutDataType = decltype(in_element_func(typename InDstrTensors::DataType{}...));
+
+    // TODO: make sure all distributed tensors have same lengths and distribution
+    // static_assert(xxx);
+    constexpr auto in_tile_dstr = type_pack_element<0, InDstrTensors...>::GetTileDistribution();
+
+    constexpr index_t thread_buffer_size =
+        type_pack_element<0, InDstrTensors...>::GetThreadBufferSize();
+
+    auto out_dstr_tensor = make_static_distributed_tensor<OutDataType>(in_tile_dstr);
+
+    static_for<0, thread_buffer_size, 1>{}([&](auto i) {
+        out_dstr_tensor.GetThreadBuffer()(i) =
+            in_element_func(in_dstr_tensors.GetThreadBuffer()[i]...);
+    });
+
+    return out_dstr_tensor;
+}
+
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/tile_fmha_shape.hpp b/include/ck/tile_program/tile/tile_fmha_shape.hpp
new file mode 100644
index 0000000000..f26266e17b
--- /dev/null
+++ b/include/ck/tile_program/tile/tile_fmha_shape.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+
+namespace ck {
+namespace tile_program {
+
+template <index_t kM0PerTile_, // tile size along q seqlen
+          index_t kN0PerTile_, // tile size along k seqlen
+          index_t kK0PerTile_, // tile size along qk gemm unroll
+          index_t kN1PerTile_, // tile size along v head_dim
+          index_t kK1PerTile_  // tile size along kv gemm unroll
+          >
+struct TileFmhaShape
+{
+    static constexpr index_t kM0 = kM0PerTile_;
+    static constexpr index_t kN0 = kN0PerTile_;
+    static constexpr index_t kK0 = kK0PerTile_;
+    static constexpr index_t kN1 = kN1PerTile_;
+    static constexpr index_t kK1 = kK1PerTile_;
+};
+
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/tile_gemm_shape.hpp b/include/ck/tile_program/tile/tile_gemm_shape.hpp
new file mode 100644
index 0000000000..6dcfabd964
--- /dev/null
+++ b/include/ck/tile_program/tile/tile_gemm_shape.hpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+
+namespace ck {
+namespace tile_program {
+
+template <index_t kMPerTile, index_t kNPerTile, index_t kKPerTile>
+struct TileGemmShape
+{
+    static constexpr index_t kM = kMPerTile;
+    static constexpr index_t kN = kNPerTile;
+    static constexpr index_t kK = kKPerTile;
+};
+
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/tile_window.hpp b/include/ck/tile_program/tile/tile_window.hpp
new file mode 100644
index 0000000000..a07b52800f
--- /dev/null
+++ b/include/ck/tile_program/tile/tile_window.hpp
@@ -0,0 +1,12 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_description/tensor_adaptor_coordinate.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/tile_window_impl_static_distribution.hpp"
+#include "ck/tile_program/tile/tile_window_impl_static_lengths.hpp"
diff --git a/include/ck/tile_program/tile/tile_window_impl_static_distribution.hpp b/include/ck/tile_program/tile/tile_window_impl_static_distribution.hpp
new file mode 100644
index 0000000000..a4aea2e77c
--- /dev/null
+++ b/include/ck/tile_program/tile/tile_window_impl_static_distribution.hpp
@@ -0,0 +1,457 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_description/tensor_adaptor_coordinate.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/tile/static_tile_distribution_helper.hpp"
+#include "ck/tile_program/tile/static_distributed_tensor.hpp"
+
+namespace ck {
+namespace tile_program {
+
+template <typename BottomTensorView_,
+          typename WindowLengths_,
+          typename StaticTileDistribution_,
+          index_t NumCoord>
+struct TileWindowWithStaticDistribution
+{
+    using BottomTensorView = remove_reference_t<BottomTensorView_>;
+    using WindowLengths    = remove_cvref_t<WindowLengths_>;
+    using TileDstr         = remove_cvref_t<StaticTileDistribution_>;
+
+    using WindowAdaptor    = typename TileDstr::PsYs2XsAdaptor;
+    using BottomTensorDesc = typename BottomTensorView::TensorDesc;
+
+    using DataType = remove_cvref_t<typename BottomTensorView::DataType>;
+
+    static constexpr index_t NDimWindowAdaptorTop = WindowAdaptor::GetNumOfTopDimension();
+    static constexpr index_t NDimBottomTensor     = BottomTensorDesc::GetNumOfDimension();
+
+    static constexpr index_t NDimP = TileDstr::GetNumOfDimensionP();
+    static constexpr index_t NDimY = TileDstr::GetNumOfDimensionY();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    // TODO: check WindowLengths and StaticTileDistribution are consistent
+
+    static_assert(is_known_at_compile_time<WindowLengths>::value,
+                  "wrong! lengths should be static");
+    static_assert(TileDstr::IsStatic(), "wrong!");
+
+    static_assert(NDimBottomTensor == WindowAdaptor::GetNumOfBottomDimension(),
+                  "wrong! inconsistent # of diemsnions");
+
+    using AdaptorTopIndex   = Array<index_t, NDimWindowAdaptorTop>;
+    using BottomTensorIndex = Array<index_t, NDimBottomTensor>;
+
+    using WindowAdaptorCoord =
+        decltype(make_tensor_adaptor_coordinate(WindowAdaptor{}, AdaptorTopIndex{}));
+
+    using BottomTensorCoord =
+        decltype(make_tensor_coordinate(BottomTensorDesc{}, BottomTensorIndex{}));
+
+    struct LoadStoreTraits
+    {
+        private:
+        static constexpr auto GetVectorDimYScalarPerVector()
+        {
+            const auto [ys_vector_lengths, ys_vector_strides] =
+                TileWindowWithStaticDistribution::GetWindowAdaptorYsSafeVectorLengthStrides();
+
+            index_t VectorDimY_      = 0;
+            index_t ScalarPerVector_ = 1;
+
+            for(index_t i = 0; i < NDimY; ++i)
+            {
+                if(ys_vector_strides[i] == 1 && ys_vector_lengths[i] > ScalarPerVector_)
+                {
+                    ScalarPerVector_ = ys_vector_lengths[i];
+                    VectorDimY_      = i;
+                }
+            }
+
+            return make_tuple(VectorDimY_, ScalarPerVector_);
+        }
+
+        public:
+        static constexpr index_t VectorDimY      = GetVectorDimYScalarPerVector().template At<0>();
+        static constexpr index_t ScalarPerVector = GetVectorDimYScalarPerVector().template At<1>();
+
+        using vector_type_t = vector_type_maker_t<DataType, ScalarPerVector>;
+        using vector_t      = typename vector_type_t::type;
+
+        private:
+        static constexpr auto scalars_per_access_ = [] {
+            constexpr auto scalars_per_access_arr = generate_array(
+                [&](auto i) { return (i == VectorDimY) ? ScalarPerVector : 1; }, Number<NDimY>{});
+
+            /// TODO: add non-automatic storage argument support to macro TO_SEQUENCE()
+            constexpr auto NDimY_ = NDimY;
+
+            return TO_SEQUENCE(scalars_per_access_arr, NDimY_);
+        }();
+
+        static constexpr auto GetSpaceFillingCurve()
+        {
+            constexpr auto tile_dstr = TileDstr{};
+
+            constexpr auto thread_tensor_lengths_ys =
+                to_sequence(tile_dstr.GetYs2DDescriptor().GetLengths());
+
+            // FIXME: need logic to judge dim access order
+            using DimAccessOrder = typename arithmetic_sequence_gen<0, NDimY, 1>::type;
+
+            return SpaceFillingCurve<decltype(thread_tensor_lengths_ys),
+                                     DimAccessOrder,
+                                     decltype(scalars_per_access_)>{};
+        }
+
+        public:
+        using SFC_Ys = decltype(GetSpaceFillingCurve());
+
+        static constexpr index_t NumAccess = SFC_Ys::GetNumOfAccess();
+
+        static_assert(0 < NumAccess, "Wrong! NumAccess should be larger than 0");
+        static_assert(NumAccess % NumCoord == 0, "wrong! # of access is not divisible by NumCoord");
+    };
+
+    static constexpr index_t NumAccessPerCoord = LoadStoreTraits::NumAccess / NumCoord;
+
+    __device__ constexpr TileWindowWithStaticDistribution() = default;
+
+    __device__ constexpr TileWindowWithStaticDistribution(
+        const BottomTensorView& bottom_tensor_view,
+        const WindowLengths& window_lengths,
+        const BottomTensorIndex& window_origin,
+        const TileDstr& tile_distribution)
+        : bottom_tensor_view_{bottom_tensor_view},
+          window_lengths_{window_lengths},
+          window_origin_{window_origin},
+          tile_dstr_{tile_distribution},
+          pre_computed_coords_{}
+    {
+#if 0 // debug
+      // TODO: this use more register for FA, but less register for GEMM
+      // need investigation
+      // only support warp-tile and block-tile
+        static_assert(NDimP == 1 or NDimP == 2, "wrong!");
+
+        WindowAdaptorCoord window_adaptor_thread_coord_tmp;
+
+        if constexpr(NDimP == 1)
+        {
+            window_adaptor_thread_coord_tmp = make_tensor_adaptor_coordinate(
+                tile_distribution.GetPsYs2XsAdaptor(), AdaptorTopIndex{get_lane_id(), 0});
+        }
+        else if constexpr(NDimP == 2)
+        {
+            window_adaptor_thread_coord_tmp =
+                make_tensor_adaptor_coordinate(tile_distribution.GetPsYs2XsAdaptor(),
+                                               AdaptorTopIndex{get_warp_id(), get_lane_id(), 0});
+        }
+#else
+        // TODO: this use less register for FA, but more register for GEMM
+        // need investigation
+        const auto window_adaptor_thread_coord_tmp = make_tensor_adaptor_coordinate(
+            tile_distribution.GetPsYs2XsAdaptor(),
+            container_concat(detail::get_partition_index(tile_distribution),
+                             Array<index_t, NDimY>{0}));
+#endif
+
+        BottomTensorIndex bottom_tensor_thread_origin_idx_tmp =
+            window_origin + window_adaptor_thread_coord_tmp.GetBottomIndex();
+
+        const auto bottom_tensor_thread_coord_tmp = make_tensor_coordinate(
+            bottom_tensor_view_.GetTensorDescriptor(), bottom_tensor_thread_origin_idx_tmp);
+
+        // pre-compute NumCoord (WindowAdaptorCoord, BottomTensorCoord) bundles to speed up
+        // future Load/Store() calls (might allocate more registers)
+        using Traits = LoadStoreTraits;
+        using SFC_Ys = typename Traits::SFC_Ys;
+
+        static_for<0, NumCoord, 1>{}([&](auto iCoord) {
+            auto window_adaptor_thread_coord = window_adaptor_thread_coord_tmp;
+            auto bottom_tensor_thread_coord  = bottom_tensor_thread_coord_tmp;
+
+            constexpr auto idx_diff_ys =
+                SFC_Ys::GetStepBetween(Number<0>{}, Number<iCoord * NumAccessPerCoord>{});
+
+            constexpr auto idx_diff_ps_ys = container_concat(Array<index_t, NDimP>{0}, idx_diff_ys);
+
+            MoveWindowAdaptorAndBottomTensorThreadCoordinate(
+                window_adaptor_thread_coord, bottom_tensor_thread_coord, idx_diff_ps_ys);
+
+            pre_computed_coords_(iCoord) =
+                make_tuple(window_adaptor_thread_coord, bottom_tensor_thread_coord);
+        });
+    }
+
+    __device__ static constexpr index_t GetNumOfDimension() { return NDimBottomTensor; }
+
+    __device__ static constexpr bool HasStaticTileDistribution() { return TileDstr::IsStatic(); }
+
+    __device__ constexpr auto GetWindowLengths() const { return window_lengths_; }
+
+    __device__ constexpr auto GetTileDistribution() const { return tile_dstr_; }
+
+    __device__ constexpr auto GetBottomTensorView() const { return bottom_tensor_view_; }
+
+    __device__ constexpr auto GetWindowOrigin() const { return window_origin_; }
+
+    // move thread's window adaptor coordinate and bottom tensor coordinate
+    // [p0, p1, ..., y0, y1, ...] ==> [x0, x1, ...] ==> [x0', x1', ...] ==> [offset]
+    __device__ void MoveWindowAdaptorAndBottomTensorThreadCoordinate(
+        WindowAdaptorCoord& window_adaptor_thread_coord,
+        BottomTensorCoord& bottom_tensor_thread_coord,
+        const AdaptorTopIndex& idx_diff_adaptor_top) const
+    {
+        Array<index_t, NDimBottomTensor> idx_diff_adaptor_bottom;
+
+        move_tensor_adaptor_coordinate(tile_dstr_.GetPsYs2XsAdaptor(),
+                                       window_adaptor_thread_coord,
+                                       idx_diff_adaptor_top,
+                                       idx_diff_adaptor_bottom);
+
+        move_tensor_coordinate(bottom_tensor_view_.GetTensorDescriptor(),
+                               bottom_tensor_thread_coord,
+                               idx_diff_adaptor_bottom);
+    }
+
+    // return vector dimension among [y0, y1, ...]
+    __device__ static constexpr auto GetWindowAdaptorYsSafeVectorLengthStrides()
+    {
+        // bottom tensor top dimension vector lengths and strides
+        const auto [bottom_tensor_top_dim_vector_lengths, bottom_tensor_top_dim_vector_strides] =
+            BottomTensorDesc::GetTopDimensionSafeVectorLengthStrides();
+
+        // window vector lengths/strides
+        const auto window_adaptor_bottom_dim_vector_lengths = bottom_tensor_top_dim_vector_lengths;
+        const auto window_adaptor_bottom_dim_vector_strides = bottom_tensor_top_dim_vector_strides;
+
+        // window adaptor [p0, p1, ..., y0, y1, ...]
+        Array<index_t, WindowAdaptor::GetNumOfHiddenDimension()> window_adaptor_vector_lengths{-1};
+        Array<index_t, WindowAdaptor::GetNumOfHiddenDimension()> window_adaptor_vector_strides{-1};
+
+        constexpr auto window_adaptor_bottom_dims = WindowAdaptor::GetBottomDimensionHiddenIds();
+
+        set_container_subset(window_adaptor_vector_lengths,
+                             window_adaptor_bottom_dims,
+                             window_adaptor_bottom_dim_vector_lengths);
+        set_container_subset(window_adaptor_vector_strides,
+                             window_adaptor_bottom_dims,
+                             window_adaptor_bottom_dim_vector_strides);
+
+        const auto [window_adaptor_ps_ys_vector_lengths, window_adaptor_ps_ys_vector_strides] =
+            WindowAdaptor{}.GetTopDimensionSafeVectorLengthStrides(window_adaptor_vector_lengths,
+                                                                   window_adaptor_vector_strides);
+
+        // [y0, y1, ...]
+        constexpr auto y_dims = typename arithmetic_sequence_gen<TileDstr::GetNumOfDimensionP(),
+                                                                 NDimWindowAdaptorTop,
+                                                                 1>::type{};
+
+        return make_tuple(get_container_subset(window_adaptor_ps_ys_vector_lengths, y_dims),
+                          get_container_subset(window_adaptor_ps_ys_vector_strides, y_dims));
+    }
+
+    __device__ auto Load() const
+    {
+        using Traits = LoadStoreTraits;
+
+        using vector_type_t = typename Traits::vector_type_t;
+        using vector_t      = typename vector_type_t::type;
+        using SFC_Ys        = typename Traits::SFC_Ys;
+
+        constexpr auto tile_dstr = TileDstr{};
+
+        auto dst_tensor = make_static_distributed_tensor<DataType>(tile_dstr);
+
+        // loop over thread tensor space [y0, y1, ...]
+        static_for<0, NumCoord, 1>{}([&](auto iCoord) {
+            /// TODO: use structure binding (to be captured later) if compiled in C++20
+            auto window_adaptor_thread_coord = pre_computed_coords_[iCoord][I0];
+            auto bottom_tensor_thread_coord  = pre_computed_coords_[iCoord][I1];
+
+            static_for<0, NumAccessPerCoord, 1>{}([&](auto iCoordAccess) {
+                constexpr auto iAccess = Number<iCoord * NumAccessPerCoord + iCoordAccess>{};
+
+                // data index [y0, y1, ...]
+                constexpr auto idx_ys_start = SFC_Ys::GetIndex(iAccess);
+
+                // read from bottom tensor
+                const vector_t vec_value =
+                    GetBottomTensorView().template GetVectorizedElements<vector_t>(
+                        bottom_tensor_thread_coord);
+
+                const vector_type_t vec{vec_value};
+
+                // write into distributed tensor
+                static_for<0, Traits::ScalarPerVector, 1>{}([&](auto j) {
+                    constexpr auto idx_ys = generate_array(
+                        [&](auto jj) {
+                            return jj == Traits::VectorDimY ? (idx_ys_start[jj] + j)
+                                                            : idx_ys_start[jj];
+                        },
+                        Number<NDimY>{});
+
+                    constexpr index_t d = tile_dstr.GetYs2DDescriptor().CalculateOffset(idx_ys);
+
+                    dst_tensor.GetThreadBuffer().template At<d>() =
+                        vec.template AsType<DataType>()[j];
+                });
+
+                // move thread coordinate
+                if constexpr(iCoordAccess != (NumAccessPerCoord - 1))
+                {
+                    constexpr auto idx_diff_ys = SFC_Ys::GetForwardStep(iAccess);
+
+                    constexpr auto idx_diff_ps_ys =
+                        container_concat(Array<index_t, NDimP>{0}, idx_diff_ys);
+
+                    MoveWindowAdaptorAndBottomTensorThreadCoordinate(
+                        window_adaptor_thread_coord, bottom_tensor_thread_coord, idx_diff_ps_ys);
+                }
+            });
+        });
+
+        return dst_tensor;
+    }
+
+    __device__ void Store(const StaticDistributedTensor<DataType, TileDstr>& dstr_tensor) const
+    {
+        using Traits = LoadStoreTraits;
+
+        using vector_type_t = typename Traits::vector_type_t;
+        using vector_t      = typename vector_type_t::type;
+        using SFC_Ys        = typename Traits::SFC_Ys;
+
+        constexpr auto tile_dstr = TileDstr{};
+
+        // loop over thread tensor space [y0, y1, ...]
+        static_for<0, NumCoord, 1>{}([&](auto iCoord) {
+            /// TODO: use structure binding (to be captured later) if compiled in C++20
+            auto window_adaptor_thread_coord = pre_computed_coords_[iCoord][I0];
+            auto bottom_tensor_thread_coord  = pre_computed_coords_[iCoord][I1];
+
+            static_for<0, NumAccessPerCoord, 1>{}([&](auto iCoordAccess) {
+                constexpr auto iAccess = Number<iCoord * NumAccessPerCoord + iCoordAccess>{};
+
+                // data index [y0, y1, ...]
+                constexpr auto idx_ys_start = SFC_Ys::GetIndex(iAccess);
+
+                // read from distributed tensor
+                vector_type_t vec;
+
+                static_for<0, Traits::ScalarPerVector, 1>{}([&](auto j) {
+                    constexpr auto idx_ys = generate_array(
+                        [&](auto jj) {
+                            return jj == Traits::VectorDimY ? (idx_ys_start[jj] + j)
+                                                            : idx_ys_start[jj];
+                        },
+                        Number<NDimY>{});
+
+                    constexpr index_t d = tile_dstr.GetYs2DDescriptor().CalculateOffset(idx_ys);
+
+                    vec.template AsType<DataType>()(j) =
+                        dstr_tensor.GetThreadBuffer().template At<d>();
+                });
+
+                const vector_t vec_value = vec.template AsType<vector_t>().template At<0>();
+
+                // write into bottom tensor
+                GetBottomTensorView().template SetVectorizedElements<vector_t>(
+                    bottom_tensor_thread_coord, vec_value);
+
+                // move thread coordinate
+                if constexpr(iCoordAccess != (NumAccessPerCoord - 1))
+                {
+                    constexpr auto idx_diff_ys = SFC_Ys::GetForwardStep(iAccess);
+
+                    constexpr auto idx_diff_ps_ys =
+                        container_concat(Array<index_t, NDimP>{0}, idx_diff_ys);
+
+                    MoveWindowAdaptorAndBottomTensorThreadCoordinate(
+                        window_adaptor_thread_coord, bottom_tensor_thread_coord, idx_diff_ps_ys);
+                }
+            });
+        });
+    }
+
+    // move thread's botom tensor coordiante
+    // [x0', x1', ... ] ==> [offset]
+    // also move window-origin
+    __device__ void Move(const BottomTensorIndex& step)
+    {
+        window_origin_ += step;
+
+        static_for<0, NumCoord, 1>{}([&](auto iCoord) {
+            move_tensor_coordinate(
+                bottom_tensor_view_.GetTensorDescriptor(), pre_computed_coords_(iCoord)(I1), step);
+        });
+    }
+
+    // this is the bottom tensor view
+    // [x0', x1', ...] ==> [offset]
+    BottomTensorView bottom_tensor_view_;
+
+    //
+    WindowLengths window_lengths_;
+
+    // origin ([x0', x1', ...]) of window on bottom tensor
+    BottomTensorIndex window_origin_;
+
+    // Tile tensor distribution, which contains:
+    //   1. adaptor for window: [p0, p1, ..., y0, y1, ...] ==> [x0, x1, ...]
+    //   2. thread descriptor for thread tensor in register: [y0, y1, ...] ==> [d]
+    TileDstr tile_dstr_;
+
+    // this contains:
+    //   per-thread coordinate for window adaptor
+    //   per-thread coordinate for bottom tensor
+    Array<Tuple<WindowAdaptorCoord, BottomTensorCoord>, NumCoord> pre_computed_coords_;
+};
+
+// TODO: use strategy
+template <typename TensorView_,
+          typename WindowLengths_,
+          typename StaticTileDistribution_,
+          index_t NumCoord = 1>
+__device__ constexpr auto
+make_tile_window(const TensorView_& tensor_view,
+                 const WindowLengths_& window_lengths,
+                 const MultiIndex<TensorView_::GetNumOfDimension()>& origin,
+                 const StaticTileDistribution_& tile_distribution,
+                 Number<NumCoord> = {})
+{
+    return TileWindowWithStaticDistribution<remove_cvref_t<TensorView_>,
+                                            remove_cvref_t<WindowLengths_>,
+                                            remove_cvref_t<StaticTileDistribution_>,
+                                            NumCoord>{
+        tensor_view, window_lengths, origin, tile_distribution};
+}
+
+template <typename TensorView_,
+          typename WindowLengths_,
+          typename StaticTileDistribution_,
+          index_t NumCoord>
+__device__ void
+move_tile_window(TileWindowWithStaticDistribution<TensorView_,
+                                                  WindowLengths_,
+                                                  StaticTileDistribution_,
+                                                  NumCoord>& window,
+                 const typename TileWindowWithStaticDistribution<TensorView_,
+                                                                 WindowLengths_,
+                                                                 StaticTileDistribution_,
+                                                                 NumCoord>::BottomTensorIndex& step)
+{
+    window.Move(step);
+}
+
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/tile/tile_window_impl_static_lengths.hpp b/include/ck/tile_program/tile/tile_window_impl_static_lengths.hpp
new file mode 100644
index 0000000000..3b0f6b2c50
--- /dev/null
+++ b/include/ck/tile_program/tile/tile_window_impl_static_lengths.hpp
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+#include "ck/tensor_description/tensor_adaptor_coordinate.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+
+namespace ck {
+namespace tile_program {
+
+template <typename BottomTensorView_, typename WindowLengths_>
+struct TileWindowWithStaticLengths
+{
+    using BottomTensorView = remove_reference_t<BottomTensorView_>;
+    using WindowLengths    = remove_cvref_t<WindowLengths_>;
+    using BottomTensorDesc = typename BottomTensorView::TensorDesc;
+    using DataType         = typename BottomTensorView::DataType;
+
+    static constexpr index_t NDimBottomTensor = BottomTensorDesc::GetNumOfDimension();
+
+    static_assert(is_known_at_compile_time<WindowLengths>::value,
+                  "wrong! lengths should be static");
+
+    using BottomTensorIndex = Array<index_t, NDimBottomTensor>;
+
+    __device__ constexpr TileWindowWithStaticLengths() = default;
+
+    __device__ constexpr TileWindowWithStaticLengths(const BottomTensorView& bottom_tensor_view,
+                                                     const WindowLengths& window_lengths,
+                                                     const BottomTensorIndex& window_origin)
+        : bottom_tensor_view_{bottom_tensor_view},
+          window_lengths_{window_lengths},
+          window_origin_{window_origin}
+    {
+    }
+
+    __device__ static constexpr index_t GetNumOfDimension() { return NDimBottomTensor; }
+
+    __device__ constexpr auto GetWindowLengths() const { return window_lengths_; }
+
+    __device__ constexpr auto GetBottomTensorView() const { return bottom_tensor_view_; }
+
+    __device__ constexpr auto GetWindowOrigin() const { return window_origin_; }
+
+    // move window-origin
+    __device__ void Move(const BottomTensorIndex& step) { window_origin_ += step; }
+
+    // this is the bottom tensor view
+    // [x0', x1', ...] ==> [offset]
+    BottomTensorView bottom_tensor_view_;
+
+    //
+    WindowLengths window_lengths_;
+
+    // origin ([x0', x1', ...]) of window on bottom tensor
+    BottomTensorIndex window_origin_;
+};
+
+template <typename TensorView_, typename WindowLengths_>
+__device__ constexpr auto
+make_tile_window(const TensorView_& tensor_view,
+                 const WindowLengths_& window_lengths,
+                 const MultiIndex<TensorView_::GetNumOfDimension()>& origin)
+{
+    static_assert(is_known_at_compile_time<WindowLengths_>::value,
+                  "wrong! lengths should be static");
+
+    return TileWindowWithStaticLengths<remove_cvref_t<TensorView_>, remove_cvref_t<WindowLengths_>>{
+        tensor_view, window_lengths, origin};
+}
+
+template <typename TensorView_, typename WindowLengths_>
+__device__ void move_tile_window(
+    TileWindowWithStaticLengths<TensorView_, WindowLengths_>& window,
+    const typename TileWindowWithStaticLengths<TensorView_, WindowLengths_>::BottomTensorIndex&
+        step)
+{
+    window.Move(step);
+}
+
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/warp_tile/warp_gemm.hpp b/include/ck/tile_program/warp_tile/warp_gemm.hpp
new file mode 100644
index 0000000000..21ff0bfc77
--- /dev/null
+++ b/include/ck/tile_program/warp_tile/warp_gemm.hpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+
+#include "ck/tile_program/warp_tile/warp_gemm_impl.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm_attribute_mfma.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm_attribute_mfma_impl.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace warp {
+
+using WarpGemmMfmaF16F16F32M32N32K8 =
+    WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M32N32K8>>;
+
+using WarpGemmMfmaF16F16F32M16N16K16 =
+    WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M16N16K16>>;
+
+using WarpGemmMfmaF16F16F32M32N32K16 =
+    WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<WarpGemmAttributeMfmaImplF16F16F32M32N32K8, 2>>;
+
+using WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution = WarpGemmImpl<
+    WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImplF16F16F32M32N32K8>>;
+
+using WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution =
+    WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
+        WarpGemmAttributeMfmaImplF16F16F32M32N32K8,
+        2>>;
+
+using WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution =
+    WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
+        WarpGemmAttributeMfmaImplF16F16F32M16N16K16,
+        2>>;
+
+} // namespace warp
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/warp_tile/warp_gemm_attribute_mfma.hpp b/include/ck/tile_program/warp_tile/warp_gemm_attribute_mfma.hpp
new file mode 100644
index 0000000000..f87f7c77f1
--- /dev/null
+++ b/include/ck/tile_program/warp_tile/warp_gemm_attribute_mfma.hpp
@@ -0,0 +1,292 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+#include "ck/tile_program/warp_tile/warp_gemm_attribute_mfma_impl.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace warp {
+
+template <typename WarpGemmAttributeMfmaImpl_>
+struct WarpGemmAtrributeMfma
+{
+    using Impl = remove_cvref_t<WarpGemmAttributeMfmaImpl_>;
+
+    using ADataType = typename Impl::ADataType;
+    using BDataType = typename Impl::BDataType;
+    using CDataType = typename Impl::CDataType;
+
+    using AVecType = typename Impl::AVecType;
+    using BVecType = typename Impl::BVecType;
+    using CVecType = typename Impl::CVecType;
+
+    static constexpr index_t kM = Impl::kM;
+    static constexpr index_t kN = Impl::kN;
+    static constexpr index_t kK = Impl::kK;
+
+    using AWarpDstrEncoding = StaticTileDistributionEncoding<
+        Sequence<>,
+        Tuple<Sequence<Impl::kAMLane>, Sequence<Impl::kABKLane, Impl::kABKPerLane>>,
+        Tuple<Sequence<2, 1>>,
+        Tuple<Sequence<0, 0>>,
+        Sequence<2>,
+        Sequence<1>>;
+
+    using BWarpDstrEncoding = StaticTileDistributionEncoding<
+        Sequence<>,
+        Tuple<Sequence<Impl::kBNLane>, Sequence<Impl::kABKLane, Impl::kABKPerLane>>,
+        Tuple<Sequence<2, 1>>,
+        Tuple<Sequence<0, 0>>,
+        Sequence<2>,
+        Sequence<1>>;
+
+    using CWarpDstrEncoding = StaticTileDistributionEncoding<
+        Sequence<>,
+        Tuple<Sequence<Impl::kCM0PerLane, Impl::kCMLane, Impl::kCM1PerLane>,
+              Sequence<Impl::kCNLane>>,
+        Tuple<Sequence<1, 2>>,
+        Tuple<Sequence<1, 0>>,
+        Sequence<1, 1>,
+        Sequence<0, 2>>;
+
+    // c_vec += a_vec * b_vec
+    __device__ void operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    {
+        Impl{}(c_vec, a_vec, b_vec);
+    }
+
+    // c_vec = a_vec * b_vec
+    __device__ CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
+    {
+        return Impl{}(a_vec, b_vec);
+    }
+};
+
+template <typename WarpGemmAttributeMfmaImpl_, index_t kKIter>
+struct WarpGemmAtrributeMfmaIterateK
+{
+    static_assert(kKIter > 0, "wrong!");
+
+    using Impl = remove_cvref_t<WarpGemmAttributeMfmaImpl_>;
+
+    using ADataType = typename Impl::ADataType;
+    using BDataType = typename Impl::BDataType;
+    using CDataType = typename Impl::CDataType;
+
+    using AVecType = typename vector_type_maker<typename Impl::AVecType, kKIter>::type::type;
+    using BVecType = typename vector_type_maker<typename Impl::BVecType, kKIter>::type::type;
+    using CVecType = typename Impl::CVecType;
+
+    static constexpr index_t kM = Impl::kM;
+    static constexpr index_t kN = Impl::kN;
+    static constexpr index_t kK = Impl::kK * kKIter;
+
+    using AWarpDstrEncoding = StaticTileDistributionEncoding<
+        Sequence<>,
+        Tuple<Sequence<Impl::kAMLane>, Sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+        Tuple<Sequence<2, 1>>,
+        Tuple<Sequence<0, 0>>,
+        Sequence<2>,
+        Sequence<1>>;
+
+    using BWarpDstrEncoding = StaticTileDistributionEncoding<
+        Sequence<>,
+        Tuple<Sequence<Impl::kBNLane>, Sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+        Tuple<Sequence<2, 1>>,
+        Tuple<Sequence<0, 0>>,
+        Sequence<2>,
+        Sequence<1>>;
+
+    using CWarpDstrEncoding = StaticTileDistributionEncoding<
+        Sequence<>,
+        Tuple<Sequence<Impl::kCM0PerLane, Impl::kCMLane, Impl::kCM1PerLane>,
+              Sequence<Impl::kCNLane>>,
+        Tuple<Sequence<1, 2>>,
+        Tuple<Sequence<1, 0>>,
+        Sequence<1, 1>,
+        Sequence<0, 2>>;
+
+    // c_vec += a_vec * b_vec
+    __device__ void operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    {
+        const auto a_vector = typename vector_type_maker<AVecType, 1>::type{a_vec};
+        const auto b_vector = typename vector_type_maker<BVecType, 1>::type{b_vec};
+
+        static_for<0, kKIter, 1>{}([&](auto iKIter) {
+            Impl{}(c_vec,
+                   a_vector.template AsType<typename Impl::AVecType>()[iKIter],
+                   b_vector.template AsType<typename Impl::BVecType>()[iKIter]);
+        });
+    }
+
+    // c_vec = a_vec * b_vec
+    __device__ CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
+    {
+        const auto a_vector = typename vector_type_maker<AVecType, 1>::type{a_vec};
+        const auto b_vector = typename vector_type_maker<BVecType, 1>::type{b_vec};
+
+        constexpr auto I0 = Number<0>{};
+
+        // c = a * b
+        auto c_vec = Impl{}(a_vector.template AsType<typename Impl::AVecType>()[I0],
+                            b_vector.template AsType<typename Impl::BVecType>()[I0]);
+
+        // c += a * b
+        static_for<1, kKIter, 1>{}([&](auto iKIter) {
+            Impl{}(c_vec,
+                   a_vector.template AsType<typename Impl::AVecType>()[iKIter],
+                   b_vector.template AsType<typename Impl::BVecType>()[iKIter]);
+        });
+
+        return c_vec;
+    }
+};
+
+template <typename WarpGemmAttributeMfmaImpl_>
+struct WarpGemmAtrributeMfmaTransposedCDistribution
+{
+    using Impl = remove_cvref_t<WarpGemmAttributeMfmaImpl_>;
+
+    using ADataType = typename Impl::BDataType;
+    using BDataType = typename Impl::ADataType;
+    using CDataType = typename Impl::CDataType;
+
+    using AVecType = typename Impl::BVecType;
+    using BVecType = typename Impl::AVecType;
+    using CVecType = typename Impl::CVecType;
+
+    static constexpr index_t kM = Impl::kN;
+    static constexpr index_t kN = Impl::kM;
+    static constexpr index_t kK = Impl::kK;
+
+    using AWarpDstrEncoding = StaticTileDistributionEncoding<
+        Sequence<>,
+        Tuple<Sequence<Impl::kBNLane>, Sequence<Impl::kABKLane, Impl::kABKPerLane>>,
+        Tuple<Sequence<2, 1>>,
+        Tuple<Sequence<0, 0>>,
+        Sequence<2>,
+        Sequence<1>>;
+
+    using BWarpDstrEncoding = StaticTileDistributionEncoding<
+        Sequence<>,
+        Tuple<Sequence<Impl::kAMLane>, Sequence<Impl::kABKLane, Impl::kABKPerLane>>,
+        Tuple<Sequence<2, 1>>,
+        Tuple<Sequence<0, 0>>,
+        Sequence<2>,
+        Sequence<1>>;
+
+    using CWarpDstrEncoding = StaticTileDistributionEncoding<
+        Sequence<>,
+        Tuple<Sequence<Impl::kCNLane>,
+              Sequence<Impl::kCM0PerLane, Impl::kCMLane, Impl::kCM1PerLane>>,
+        Tuple<Sequence<2, 1>>,
+        Tuple<Sequence<1, 0>>,
+        Sequence<2, 2>,
+        Sequence<0, 2>>;
+
+    // c_vec += a_vec * b_vec
+    __device__ void operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    {
+        // swap A and B
+        Impl{}(c_vec, b_vec, a_vec);
+    }
+
+    // c_vec = a_vec * b_vec
+    __device__ CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
+    {
+        // swap A and B
+        return Impl{}(b_vec, a_vec);
+    }
+};
+
+template <typename WarpGemmAttributeMfmaImpl_, index_t kKIter>
+struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution
+{
+    using Impl = remove_cvref_t<WarpGemmAttributeMfmaImpl_>;
+
+    // swap A and B
+    using ADataType = typename Impl::BDataType;
+    using BDataType = typename Impl::ADataType;
+    using CDataType = typename Impl::CDataType;
+
+    using AVecType = typename vector_type_maker<typename Impl::BVecType, kKIter>::type::type;
+    using BVecType = typename vector_type_maker<typename Impl::AVecType, kKIter>::type::type;
+    using CVecType = typename Impl::CVecType;
+
+    static constexpr index_t kM = Impl::kN;
+    static constexpr index_t kN = Impl::kM;
+    static constexpr index_t kK = Impl::kK * kKIter;
+
+    using AWarpDstrEncoding = StaticTileDistributionEncoding<
+        Sequence<>,
+        Tuple<Sequence<Impl::kBNLane>, Sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+        Tuple<Sequence<2, 1>>,
+        Tuple<Sequence<0, 0>>,
+        Sequence<2>,
+        Sequence<1>>;
+
+    using BWarpDstrEncoding = StaticTileDistributionEncoding<
+        Sequence<>,
+        Tuple<Sequence<Impl::kAMLane>, Sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+        Tuple<Sequence<2, 1>>,
+        Tuple<Sequence<0, 0>>,
+        Sequence<2>,
+        Sequence<1>>;
+
+    using CWarpDstrEncoding = StaticTileDistributionEncoding<
+        Sequence<>,
+        Tuple<Sequence<Impl::kCNLane>,
+              Sequence<Impl::kCM0PerLane, Impl::kCMLane, Impl::kCM1PerLane>>,
+        Tuple<Sequence<2, 1>>,
+        Tuple<Sequence<1, 0>>,
+        Sequence<2, 2>,
+        Sequence<0, 2>>;
+
+    // c_vec += a_vec * b_vec
+    __device__ void operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    {
+        const auto a_vector = typename vector_type_maker<AVecType, 1>::type{a_vec};
+
+        const auto b_vector = typename vector_type_maker<BVecType, 1>::type{b_vec};
+
+        // swap A and B, value and type
+        static_for<0, kKIter, 1>{}([&](auto iKIter) {
+            Impl{}(c_vec,
+                   b_vector.template AsType<typename Impl::AVecType>()[iKIter],
+                   a_vector.template AsType<typename Impl::BVecType>()[iKIter]);
+        });
+    }
+
+    // c_vec = a_vec * b_vec
+    __device__ CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
+    {
+        const auto a_vector = typename vector_type_maker<AVecType, 1>::type{a_vec};
+        const auto b_vector = typename vector_type_maker<BVecType, 1>::type{b_vec};
+
+        constexpr auto I0 = Number<0>{};
+
+        // swap A and B, value and type
+        auto c_vec = Impl{}(b_vector.template AsType<typename Impl::AVecType>()[I0],
+                            a_vector.template AsType<typename Impl::BVecType>()[I0]);
+
+        static_for<1, kKIter, 1>{}([&](auto iKIter) {
+            Impl{}(c_vec,
+                   b_vector.template AsType<typename Impl::AVecType>()[iKIter],
+                   a_vector.template AsType<typename Impl::BVecType>()[iKIter]);
+        });
+
+        return c_vec;
+    }
+};
+
+} // namespace warp
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/warp_tile/warp_gemm_attribute_mfma_impl.hpp b/include/ck/tile_program/warp_tile/warp_gemm_attribute_mfma_impl.hpp
new file mode 100644
index 0000000000..68c721bfcf
--- /dev/null
+++ b/include/ck/tile_program/warp_tile/warp_gemm_attribute_mfma_impl.hpp
@@ -0,0 +1,90 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace warp {
+
+struct WarpGemmAttributeMfmaImplF16F16F32M32N32K8
+{
+    using ADataType = half_t;
+    using BDataType = half_t;
+    using CDataType = float;
+
+    using AVecType = typename vector_type<half_t, 4>::type;
+    using BVecType = typename vector_type<half_t, 4>::type;
+    using CVecType = typename vector_type<float, 16>::type;
+
+    static constexpr index_t kM = 32;
+    static constexpr index_t kN = 32;
+    static constexpr index_t kK = 8;
+
+    static constexpr index_t kAMLane     = 32;
+    static constexpr index_t kBNLane     = 32;
+    static constexpr index_t kABKLane    = 2;
+    static constexpr index_t kABKPerLane = 4;
+
+    static constexpr index_t kCMLane     = 2;
+    static constexpr index_t kCNLane     = 32;
+    static constexpr index_t kCM0PerLane = 4;
+    static constexpr index_t kCM1PerLane = 4;
+
+    // c_vec += a_vec * b_vec
+    __device__ void operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    {
+        c_vec = __builtin_amdgcn_mfma_f32_32x32x8f16(a_vec, b_vec, c_vec, 0, 0, 0);
+    }
+
+    // c_vec = a_vec * b_vec
+    __device__ CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
+    {
+        // FIXME: Is this correct?
+        return __builtin_amdgcn_mfma_f32_32x32x8f16(a_vec, b_vec, CVecType{0.f}, 0, 0, 0);
+    }
+};
+
+struct WarpGemmAttributeMfmaImplF16F16F32M16N16K16
+{
+    using ADataType = half_t;
+    using BDataType = half_t;
+    using CDataType = float;
+
+    using AVecType = typename vector_type<half_t, 4>::type;
+    using BVecType = typename vector_type<half_t, 4>::type;
+    using CVecType = typename vector_type<float, 4>::type;
+
+    static constexpr index_t kM = 16;
+    static constexpr index_t kN = 16;
+    static constexpr index_t kK = 16;
+
+    static constexpr index_t kAMLane     = 16;
+    static constexpr index_t kBNLane     = 16;
+    static constexpr index_t kABKLane    = 4;
+    static constexpr index_t kABKPerLane = 4;
+
+    static constexpr index_t kCMLane     = 4;
+    static constexpr index_t kCNLane     = 16;
+    static constexpr index_t kCM0PerLane = 1;
+    static constexpr index_t kCM1PerLane = 4;
+
+    // c_vec += a_vec * b_vec
+    __device__ void operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    {
+        c_vec = __builtin_amdgcn_mfma_f32_16x16x16f16(a_vec, b_vec, c_vec, 0, 0, 0);
+    }
+
+    // c_vec = a_vec * b_vec
+    __device__ CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
+    {
+        // FIXME: Is this correct?
+        return __builtin_amdgcn_mfma_f32_16x16x16f16(a_vec, b_vec, CVecType{0.f}, 0, 0, 0);
+    }
+};
+
+} // namespace warp
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/tile_program/warp_tile/warp_gemm_impl.hpp b/include/ck/tile_program/warp_tile/warp_gemm_impl.hpp
new file mode 100644
index 0000000000..b27c783df8
--- /dev/null
+++ b/include/ck/tile_program/warp_tile/warp_gemm_impl.hpp
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+#include "ck/tile_program/tile/tile_distribution.hpp"
+
+namespace ck {
+namespace tile_program {
+namespace warp {
+
+template <typename WarpGemmAttribute_>
+struct WarpGemmImpl
+{
+    using WarpGemmAttribute = remove_cvref_t<WarpGemmAttribute_>;
+
+    static constexpr index_t kM = WarpGemmAttribute::kM;
+    static constexpr index_t kN = WarpGemmAttribute::kN;
+    static constexpr index_t kK = WarpGemmAttribute::kK;
+
+    using ADataType = typename WarpGemmAttribute::ADataType;
+    using BDataType = typename WarpGemmAttribute::BDataType;
+    using CDataType = typename WarpGemmAttribute::CDataType;
+
+    using AWarpDstrEncoding = typename WarpGemmAttribute::AWarpDstrEncoding;
+    using BWarpDstrEncoding = typename WarpGemmAttribute::BWarpDstrEncoding;
+    using CWarpDstrEncoding = typename WarpGemmAttribute::CWarpDstrEncoding;
+
+    using AWarpDstr = remove_cvref_t<decltype(make_static_tile_distribution(AWarpDstrEncoding{}))>;
+    using BWarpDstr = remove_cvref_t<decltype(make_static_tile_distribution(BWarpDstrEncoding{}))>;
+    using CWarpDstr = remove_cvref_t<decltype(make_static_tile_distribution(CWarpDstrEncoding{}))>;
+
+    using AWarpTensor = StaticDistributedTensor<ADataType, AWarpDstr>;
+    using BWarpTensor = StaticDistributedTensor<BDataType, BWarpDstr>;
+    using CWarpTensor = StaticDistributedTensor<CDataType, CWarpDstr>;
+
+    __device__ void operator()(CWarpTensor& c, const AWarpTensor& a, const BWarpTensor& b) const
+    {
+        using AVec = typename vector_type<ADataType, AWarpTensor::GetThreadBufferSize()>::type;
+        using BVec = typename vector_type<BDataType, BWarpTensor::GetThreadBufferSize()>::type;
+        using CVec = typename vector_type<CDataType, CWarpTensor::GetThreadBufferSize()>::type;
+
+        constexpr auto I0 = Number<0>{};
+
+        const auto a_vec = a.GetThreadBuffer().template GetAsType<AVec>(I0);
+        const auto b_vec = b.GetThreadBuffer().template GetAsType<BVec>(I0);
+        auto c_vec       = c.GetThreadBuffer().template GetAsType<CVec>(I0);
+
+        // c_vec += a_vec * b_vec
+        WarpGemmAttribute{}(c_vec, a_vec, b_vec);
+
+        c.GetThreadBuffer().template SetAsType<CVec>(I0, c_vec);
+    }
+
+    __device__ auto operator()(const AWarpTensor& a, const BWarpTensor& b) const
+    {
+        CWarpTensor c;
+
+        using AVec = typename vector_type<ADataType, AWarpTensor::GetThreadBufferSize()>::type;
+        using BVec = typename vector_type<BDataType, BWarpTensor::GetThreadBufferSize()>::type;
+        using CVec = typename vector_type<CDataType, CWarpTensor::GetThreadBufferSize()>::type;
+
+        constexpr auto I0 = Number<0>{};
+
+        const auto a_vec = a.GetThreadBuffer().template GetAsType<AVec>(I0);
+        const auto b_vec = b.GetThreadBuffer().template GetAsType<BVec>(I0);
+
+        // c_vec = a_vec * b_vec
+        auto c_vec = WarpGemmAttribute{}(a_vec, b_vec);
+
+        c.GetThreadBuffer().template SetAsType<CVec>(I0, c_vec);
+
+        return c;
+    }
+};
+
+} // namespace warp
+} // namespace tile_program
+} // namespace ck
diff --git a/include/ck/utility/amd_address_space.hpp b/include/ck/utility/amd_address_space.hpp
new file mode 100644
index 0000000000..d54f70e750
--- /dev/null
+++ b/include/ck/utility/amd_address_space.hpp
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "c_style_pointer_cast.hpp"
+
+// Address Space for AMDGCN
+// https://llvm.org/docs/AMDGPUUsage.html#address-space
+
+namespace ck {
+
+enum struct AddressSpaceEnum
+{
+    Generic,
+    Global,
+    Lds,
+    Sgpr,
+    Vgpr,
+};
+
+template <typename T>
+__device__ T* cast_pointer_to_generic_address_space(T CK_CONSTANT_ADDRESS_SPACE* p)
+{
+    // cast a pointer in "Constant" address space (4) to "Generic" address space (0)
+    // only c-style pointer cast seems be able to be compiled
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wold-style-cast"
+    return (T*)p; // NOLINT(old-style-cast)
+#pragma clang diagnostic pop
+}
+
+template <typename T>
+__host__ __device__ T CK_CONSTANT_ADDRESS_SPACE* cast_pointer_to_constant_address_space(T* p)
+{
+    // cast a pointer in "Generic" address space (0) to "Constant" address space (4)
+    // only c-style pointer cast seems be able to be compiled
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wold-style-cast"
+    return (T CK_CONSTANT_ADDRESS_SPACE*)p; // NOLINT(old-style-cast)
+#pragma clang diagnostic pop
+}
+
+} // namespace ck
diff --git a/include/ck/utility/amd_buffer_addressing.hpp b/include/ck/utility/amd_buffer_addressing.hpp
new file mode 100644
index 0000000000..897cb4f249
--- /dev/null
+++ b/include/ck/utility/amd_buffer_addressing.hpp
@@ -0,0 +1,1314 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+#include "data_type.hpp"
+
+namespace ck {
+
+template <typename T>
+union BufferResource
+{
+    __device__ constexpr BufferResource() : content{} {}
+
+    // 128 bit SGPRs to supply buffer resource in buffer instructions
+    // https://rocm-documentation.readthedocs.io/en/latest/GCN_ISA_Manuals/testdocbook.html#vector-memory-buffer-instructions
+    int32x4_t content;
+    StaticallyIndexedArray<T*, 2> address;
+    StaticallyIndexedArray<int32_t, 4> range;
+    StaticallyIndexedArray<int32_t, 4> config;
+};
+
+template <typename T>
+__device__ int32x4_t make_wave_buffer_resource(T* p_wave, index_t element_space_size)
+{
+    BufferResource<T> wave_buffer_resource;
+
+    // wavewise base address (64 bit)
+    wave_buffer_resource.address(Number<0>{}) = const_cast<remove_cv_t<T>*>(p_wave);
+    // wavewise range (32 bit)
+    wave_buffer_resource.range(Number<2>{}) = element_space_size * sizeof(T);
+    // wavewise setting (32 bit)
+    wave_buffer_resource.config(Number<3>{}) = CK_BUFFER_RESOURCE_3RD_DWORD;
+
+    return wave_buffer_resource.content;
+}
+
+template <typename T>
+__device__ int32x4_t make_wave_buffer_resource_with_default_range(T* p_wave)
+{
+    BufferResource<T> wave_buffer_resource;
+
+    // wavewise base address (64 bit)
+    wave_buffer_resource.address(Number<0>{}) = const_cast<remove_cv_t<T>*>(p_wave);
+    // wavewise range (32 bit)
+    wave_buffer_resource.range(Number<2>{}) = 0xffffffff; // max possible range
+    // wavewise setting (32 bit)
+    wave_buffer_resource.config(Number<3>{}) = CK_BUFFER_RESOURCE_3RD_DWORD;
+
+    return wave_buffer_resource.content;
+}
+
+// buffer load i8
+__device__ int8_t
+llvm_amdgcn_raw_buffer_load_i8(int32x4_t srsrc,
+                               index_t voffset,
+                               index_t soffset,
+                               index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.i8");
+
+__device__ int8x2_t
+llvm_amdgcn_raw_buffer_load_i8x2(int32x4_t srsrc,
+                                 index_t voffset,
+                                 index_t soffset,
+                                 index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2i8");
+
+__device__ int8x4_t
+llvm_amdgcn_raw_buffer_load_i8x4(int32x4_t srsrc,
+                                 index_t voffset,
+                                 index_t soffset,
+                                 index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4i8");
+
+// buffer load i16
+__device__ bhalf_t
+llvm_amdgcn_raw_buffer_load_i16(int32x4_t srsrc,
+                                index_t voffset,
+                                index_t soffset,
+                                index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.i16");
+
+__device__ bhalf2_t
+llvm_amdgcn_raw_buffer_load_i16x2(int32x4_t srsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2i16");
+
+__device__ bhalf4_t
+llvm_amdgcn_raw_buffer_load_i16x4(int32x4_t srsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4i16");
+
+// buffer load i32
+__device__ int32_t
+llvm_amdgcn_raw_buffer_load_i32(int32x4_t srsrc,
+                                index_t voffset,
+                                index_t soffset,
+                                index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.i32");
+
+__device__ int32x2_t
+llvm_amdgcn_raw_buffer_load_i32x2(int32x4_t srsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2i32");
+
+__device__ int32x4_t
+llvm_amdgcn_raw_buffer_load_i32x4(int32x4_t srsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4i32");
+
+// buffer load fp16
+__device__ half_t
+llvm_amdgcn_raw_buffer_load_fp16(int32x4_t srsrc,
+                                 index_t voffset,
+                                 index_t soffset,
+                                 index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.f16");
+
+__device__ half2_t
+llvm_amdgcn_raw_buffer_load_fp16x2(int32x4_t srsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2f16");
+
+__device__ half4_t
+llvm_amdgcn_raw_buffer_load_fp16x4(int32x4_t srsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4f16");
+
+// buffer load fp32
+__device__ float
+llvm_amdgcn_raw_buffer_load_fp32(int32x4_t srsrc,
+                                 index_t voffset,
+                                 index_t soffset,
+                                 index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.f32");
+
+__device__ float2_t
+llvm_amdgcn_raw_buffer_load_fp32x2(int32x4_t srsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v2f32");
+
+__device__ float4_t
+llvm_amdgcn_raw_buffer_load_fp32x4(int32x4_t srsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4f32");
+
+// buffer store i8
+__device__ void
+llvm_amdgcn_raw_buffer_store_i8(int8_t vdata,
+                                int32x4_t rsrc,
+                                index_t voffset,
+                                index_t soffset,
+                                index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i8");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_i8x2(int8x2_t vdata,
+                                  int32x4_t rsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i8");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_i8x4(int8x4_t vdata,
+                                  int32x4_t rsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4i8");
+
+// buffer store i16
+__device__ void
+llvm_amdgcn_raw_buffer_store_i16(bhalf_t vdata,
+                                 int32x4_t rsrc,
+                                 index_t voffset,
+                                 index_t soffset,
+                                 index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i16");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_i16x2(bhalf2_t vdata,
+                                   int32x4_t rsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i16");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_i16x4(bhalf4_t vdata,
+                                   int32x4_t rsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4i16");
+
+// buffer store i32
+__device__ void
+llvm_amdgcn_raw_buffer_store_i32(int32_t vdata,
+                                 int32x4_t rsrc,
+                                 index_t voffset,
+                                 index_t soffset,
+                                 index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.i32");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_i32x2(int32x2_t vdata,
+                                   int32x4_t rsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2i32");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_i32x4(int32x4_t vdata,
+                                   int32x4_t rsrc,
+                                   index_t voffset,
+                                   index_t soffset,
+                                   index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4i32");
+
+// buffer store fp16
+__device__ void
+llvm_amdgcn_raw_buffer_store_fp16(half_t vdata,
+                                  int32x4_t rsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.f16");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_fp16x2(half2_t vdata,
+                                    int32x4_t rsrc,
+                                    index_t voffset,
+                                    index_t soffset,
+                                    index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2f16");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_fp16x4(half4_t vdata,
+                                    int32x4_t rsrc,
+                                    index_t voffset,
+                                    index_t soffset,
+                                    index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4f16");
+
+// buffer store fp32
+__device__ void
+llvm_amdgcn_raw_buffer_store_fp32(float vdata,
+                                  int32x4_t rsrc,
+                                  index_t voffset,
+                                  index_t soffset,
+                                  index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.f32");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_fp32x2(float2_t vdata,
+                                    int32x4_t rsrc,
+                                    index_t voffset,
+                                    index_t soffset,
+                                    index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v2f32");
+
+__device__ void
+llvm_amdgcn_raw_buffer_store_fp32x4(float4_t vdata,
+                                    int32x4_t rsrc,
+                                    index_t voffset,
+                                    index_t soffset,
+                                    index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4f32");
+
+// buffer atomic-add fp16
+__device__ half2_t llvm_amdgcn_raw_buffer_atomic_add_fp16x2(
+    half2_t vdata,
+    int32x4_t rsrc,
+    index_t voffset,
+    index_t soffset,
+    index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fadd.v2f16");
+
+// buffer atomic-add i32
+__device__ int32_t llvm_amdgcn_raw_buffer_atomic_add_i32(
+    int32_t vdata,
+    int32x4_t rsrc,
+    index_t voffset,
+    index_t soffset,
+    index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.add.i32");
+
+// buffer atomic-add fp32
+__device__ float llvm_amdgcn_raw_buffer_atomic_add_fp32(
+    float vdata,
+    int32x4_t rsrc,
+    index_t voffset,
+    index_t soffset,
+    index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fadd.f32");
+
+// buffer atomic-add fp32
+__device__ double
+llvm_amdgcn_raw_buffer_atomic_max_fp64(double vdata,
+                                       int32x4_t rsrc, // dst_wave_buffer_resource
+                                       int voffset,    // dst_thread_addr_offset
+                                       int soffset,    // dst_wave_addr_offset
+                                       int glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fmax.f64");
+
+// memory coherency bit for buffer store/load instruction
+// check ISA manual for each GFX target
+// e.g. for
+// https://www.amd.com/system/files/TechDocs/instinct-mi200-cdna2-instruction-set-architecture.pdf,
+// page 67~68
+enum struct AmdBufferCoherenceEnum
+{
+    DefaultCoherence = 0, // default value
+    GLC              = 1,
+    SLC              = 2,
+    GLC_SLC          = 3,
+};
+
+template <typename T,
+          index_t N,
+          AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence>
+__device__ typename vector_type<T, N>::type amd_buffer_load_impl(int32x4_t src_wave_buffer_resource,
+                                                                 index_t src_thread_addr_offset,
+                                                                 index_t src_wave_addr_offset)
+{
+    static_assert(
+        (is_same<T, double>::value && (N == 1 || N == 2 || N == 4)) ||
+            (is_same<T, float>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
+            (is_same<T, half_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
+            (is_same<T, bhalf_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
+            (is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
+            (is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)),
+        "wrong! not implemented");
+
+    if constexpr(is_same<T, double>::value)
+    {
+        // use fp32 load to mimic fp64 load
+        if constexpr(N == 1)
+        {
+            const float2_t tmp =
+                llvm_amdgcn_raw_buffer_load_fp32x2(src_wave_buffer_resource,
+                                                   src_thread_addr_offset,
+                                                   src_wave_addr_offset,
+                                                   static_cast<index_t>(coherence));
+
+            return bit_cast<double>(tmp);
+        }
+        else if constexpr(N == 2)
+        {
+            const float4_t tmp =
+                llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource,
+                                                   src_thread_addr_offset,
+                                                   src_wave_addr_offset,
+                                                   static_cast<index_t>(coherence));
+
+            return bit_cast<double2_t>(tmp);
+        }
+        else if constexpr(N == 4)
+        {
+            const float4_t f32_0 =
+                llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource,
+                                                   src_thread_addr_offset,
+                                                   src_wave_addr_offset,
+                                                   static_cast<index_t>(coherence));
+
+            const float4_t f32_1 =
+                llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource,
+                                                   src_thread_addr_offset,
+                                                   src_wave_addr_offset + 4 * sizeof(float),
+                                                   static_cast<index_t>(coherence));
+            vector_type<double, 4> tmp;
+
+            tmp.AsType<double2_t>()(Number<0>{}) = bit_cast<double2_t>(f32_0);
+            tmp.AsType<double2_t>()(Number<1>{}) = bit_cast<double2_t>(f32_1);
+
+            return tmp.AsType<double4_t>()(Number<0>{});
+        }
+    }
+    else if constexpr(is_same<T, float>::value)
+    {
+        if constexpr(N == 1)
+        {
+            return llvm_amdgcn_raw_buffer_load_fp32(src_wave_buffer_resource,
+                                                    src_thread_addr_offset,
+                                                    src_wave_addr_offset,
+                                                    static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 2)
+        {
+            return llvm_amdgcn_raw_buffer_load_fp32x2(src_wave_buffer_resource,
+                                                      src_thread_addr_offset,
+                                                      src_wave_addr_offset,
+                                                      static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 4)
+        {
+            return llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource,
+                                                      src_thread_addr_offset,
+                                                      src_wave_addr_offset,
+                                                      static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 8)
+        {
+            vector_type<float, 8> tmp;
+
+            tmp.AsType<float4_t>()(Number<0>{}) =
+                llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource,
+                                                   src_thread_addr_offset,
+                                                   src_wave_addr_offset,
+                                                   static_cast<index_t>(coherence));
+
+            tmp.AsType<float4_t>()(Number<1>{}) =
+                llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource,
+                                                   src_thread_addr_offset,
+                                                   src_wave_addr_offset + 4 * sizeof(float),
+                                                   static_cast<index_t>(coherence));
+
+            return tmp.AsType<float8_t>()(Number<0>{});
+        }
+    }
+    else if constexpr(is_same<T, half_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            return llvm_amdgcn_raw_buffer_load_fp16(src_wave_buffer_resource,
+                                                    src_thread_addr_offset,
+                                                    src_wave_addr_offset,
+                                                    static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 2)
+        {
+            return llvm_amdgcn_raw_buffer_load_fp16x2(src_wave_buffer_resource,
+                                                      src_thread_addr_offset,
+                                                      src_wave_addr_offset,
+                                                      static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 4)
+        {
+            return llvm_amdgcn_raw_buffer_load_fp16x4(src_wave_buffer_resource,
+                                                      src_thread_addr_offset,
+                                                      src_wave_addr_offset,
+                                                      static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 8)
+        {
+            // use fp32 load to mimic fp16 load
+            float4_t tmp = llvm_amdgcn_raw_buffer_load_fp32x4(src_wave_buffer_resource,
+                                                              src_thread_addr_offset,
+                                                              src_wave_addr_offset,
+                                                              static_cast<index_t>(coherence));
+
+            return bit_cast<half8_t>(tmp);
+        }
+    }
+    else if constexpr(is_same<T, bhalf_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            return llvm_amdgcn_raw_buffer_load_i16(src_wave_buffer_resource,
+                                                   src_thread_addr_offset,
+                                                   src_wave_addr_offset,
+                                                   static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 2)
+        {
+            return llvm_amdgcn_raw_buffer_load_i16x2(src_wave_buffer_resource,
+                                                     src_thread_addr_offset,
+                                                     src_wave_addr_offset,
+                                                     static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 4)
+        {
+            return llvm_amdgcn_raw_buffer_load_i16x4(src_wave_buffer_resource,
+                                                     src_thread_addr_offset,
+                                                     src_wave_addr_offset,
+                                                     static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 8)
+        {
+            int32x4_t tmp = llvm_amdgcn_raw_buffer_load_i32x4(src_wave_buffer_resource,
+                                                              src_thread_addr_offset,
+                                                              src_wave_addr_offset,
+                                                              static_cast<index_t>(coherence));
+
+            return bit_cast<bhalf8_t>(tmp);
+        }
+    }
+    else if constexpr(is_same<T, int32_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            return llvm_amdgcn_raw_buffer_load_i32(src_wave_buffer_resource,
+                                                   src_thread_addr_offset,
+                                                   src_wave_addr_offset,
+                                                   static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 2)
+        {
+            return llvm_amdgcn_raw_buffer_load_i32x2(src_wave_buffer_resource,
+                                                     src_thread_addr_offset,
+                                                     src_wave_addr_offset,
+                                                     static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 4)
+        {
+            return llvm_amdgcn_raw_buffer_load_i32x4(src_wave_buffer_resource,
+                                                     src_thread_addr_offset,
+                                                     src_wave_addr_offset,
+                                                     static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 8)
+        {
+            vector_type<int32_t, 8> tmp;
+
+            tmp.AsType<int32x4_t>()(Number<0>{}) =
+                llvm_amdgcn_raw_buffer_load_i32x4(src_wave_buffer_resource,
+                                                  src_thread_addr_offset,
+                                                  src_wave_addr_offset,
+                                                  static_cast<index_t>(coherence));
+
+            tmp.AsType<int32x4_t>()(Number<1>{}) =
+                llvm_amdgcn_raw_buffer_load_i32x4(src_wave_buffer_resource,
+                                                  src_thread_addr_offset,
+                                                  src_wave_addr_offset + 4 * sizeof(int32_t),
+                                                  static_cast<index_t>(coherence));
+            return tmp.AsType<int32x8_t>()(Number<0>{});
+        }
+    }
+    else if constexpr(is_same<T, int8_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            return llvm_amdgcn_raw_buffer_load_i8(src_wave_buffer_resource,
+                                                  src_thread_addr_offset,
+                                                  src_wave_addr_offset,
+                                                  static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 2)
+        {
+#if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE
+            return llvm_amdgcn_raw_buffer_load_i8x2(src_wave_buffer_resource,
+                                                    src_thread_addr_offset,
+                                                    src_wave_addr_offset,
+                                                    static_cast<index_t>(coherence));
+#else
+            int16_t tmp = llvm_amdgcn_raw_buffer_load_i16(src_wave_buffer_resource,
+                                                          src_thread_addr_offset,
+                                                          src_wave_addr_offset,
+                                                          static_cast<index_t>(coherence));
+
+            return bit_cast<int8x2_t>(tmp);
+#endif
+        }
+        else if constexpr(N == 4)
+        {
+#if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE
+            return llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource,
+                                                    src_thread_addr_offset,
+                                                    src_wave_addr_offset,
+                                                    static_cast<index_t>(coherence));
+#else
+            int32_t tmp = llvm_amdgcn_raw_buffer_load_i32(src_wave_buffer_resource,
+                                                          src_thread_addr_offset,
+                                                          src_wave_addr_offset,
+                                                          static_cast<index_t>(coherence));
+
+            return bit_cast<int8x4_t>(tmp);
+#endif
+        }
+        else if constexpr(N == 8)
+        {
+#if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE
+            vector_type<int8_t, 8> tmp;
+
+            tmp.AsType<int8x4_t>()(Number<0>{}) =
+                llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource,
+                                                 src_thread_addr_offset,
+                                                 src_wave_addr_offset,
+                                                 static_cast<index_t>(coherence));
+
+            tmp.AsType<int8x4_t>()(Number<1>{}) =
+                llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource,
+                                                 src_thread_addr_offset,
+                                                 src_wave_addr_offset + 4 * sizeof(int8_t),
+                                                 static_cast<index_t>(coherence));
+
+            return tmp.AsType<int8x8_t>()(Number<0>{});
+#else
+            int32x2_t tmp = llvm_amdgcn_raw_buffer_load_i32x2(src_wave_buffer_resource,
+                                                              src_thread_addr_offset,
+                                                              src_wave_addr_offset,
+                                                              static_cast<index_t>(coherence));
+
+            return bit_cast<int8x8_t>(tmp);
+#endif
+        }
+        else if constexpr(N == 16)
+        {
+#if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE
+            vector_type<int8_t, 16> tmp;
+
+            tmp.AsType<int8x4_t>()(Number<0>{}) =
+                llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource,
+                                                 src_thread_addr_offset,
+                                                 src_wave_addr_offset,
+                                                 static_cast<index_t>(coherence));
+
+            tmp.AsType<int8x4_t>()(Number<1>{}) =
+                llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource,
+                                                 src_thread_addr_offset,
+                                                 src_wave_addr_offset + 4 * sizeof(int8_t),
+                                                 static_cast<index_t>(coherence));
+
+            tmp.AsType<int8x4_t>()(Number<2>{}) =
+                llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource,
+                                                 src_thread_addr_offset,
+                                                 src_wave_addr_offset + 8 * sizeof(int8_t),
+                                                 static_cast<index_t>(coherence));
+
+            tmp.AsType<int8x4_t>()(Number<3>{}) =
+                llvm_amdgcn_raw_buffer_load_i8x4(src_wave_buffer_resource,
+                                                 src_thread_addr_offset,
+                                                 src_wave_addr_offset + 12 * sizeof(int8_t),
+                                                 static_cast<index_t>(coherence));
+
+            return tmp.AsType<int8x16_t>()(Number<0>{});
+#else
+            int32x4_t tmp = llvm_amdgcn_raw_buffer_load_i32x4(src_wave_buffer_resource,
+                                                              src_thread_addr_offset,
+                                                              src_wave_addr_offset,
+                                                              static_cast<index_t>(coherence));
+
+            return bit_cast<int8x16_t>(tmp);
+#endif
+        }
+    }
+}
+
+template <typename T,
+          index_t N,
+          AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence>
+__device__ void amd_buffer_store_impl(const typename vector_type<T, N>::type src_thread_data,
+                                      int32x4_t dst_wave_buffer_resource,
+                                      index_t dst_thread_addr_offset,
+                                      index_t dst_wave_addr_offset)
+{
+    static_assert(
+        (is_same<T, double>::value && (N == 1 || N == 2)) ||
+            (is_same<T, float>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
+            (is_same<T, half_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
+            (is_same<T, bhalf_t>::value && (N == 1 || N == 2 || N == 4 || N == 8)) ||
+            (is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4)) ||
+            (is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)),
+        "wrong! not implemented");
+
+    if constexpr(is_same<T, double>::value)
+    {
+        // use fp32 store to mimic fp64 store
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_store_fp32x2(bit_cast<float2_t>(src_thread_data),
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 2)
+        {
+            llvm_amdgcn_raw_buffer_store_fp32x4(bit_cast<float4_t>(src_thread_data),
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                static_cast<index_t>(coherence));
+        }
+    }
+    else if constexpr(is_same<T, float>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_store_fp32(src_thread_data,
+                                              dst_wave_buffer_resource,
+                                              dst_thread_addr_offset,
+                                              dst_wave_addr_offset,
+                                              static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 2)
+        {
+            llvm_amdgcn_raw_buffer_store_fp32x2(src_thread_data,
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 4)
+        {
+            llvm_amdgcn_raw_buffer_store_fp32x4(src_thread_data,
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 8)
+        {
+            vector_type<float, 8> tmp{src_thread_data};
+            llvm_amdgcn_raw_buffer_store_fp32x4(tmp.AsType<float4_t>()[Number<0>{}],
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                static_cast<index_t>(coherence));
+            llvm_amdgcn_raw_buffer_store_fp32x4(tmp.AsType<float4_t>()[Number<1>{}],
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset + 4 * sizeof(float),
+                                                static_cast<index_t>(coherence));
+        }
+    }
+    else if constexpr(is_same<T, half_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_store_fp16(src_thread_data,
+                                              dst_wave_buffer_resource,
+                                              dst_thread_addr_offset,
+                                              dst_wave_addr_offset,
+                                              static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 2)
+        {
+            llvm_amdgcn_raw_buffer_store_fp16x2(src_thread_data,
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 4)
+        {
+            llvm_amdgcn_raw_buffer_store_fp16x4(src_thread_data,
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 8)
+        {
+#if 0
+            vector_type<half_t, 8> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_store_fp16x4(tmp.AsType<half4_t>()[Number<0>{}],
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                static_cast<index_t>(coherence));
+
+            llvm_amdgcn_raw_buffer_store_fp16x4(tmp.AsType<half4_t>()[Number<1>{}],
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset + 4 * sizeof(half_t),
+                                                static_cast<index_t>(coherence));
+#else
+            llvm_amdgcn_raw_buffer_store_fp32x4(bit_cast<float4_t>(src_thread_data),
+                                                dst_wave_buffer_resource,
+                                                dst_thread_addr_offset,
+                                                dst_wave_addr_offset,
+                                                static_cast<index_t>(coherence));
+#endif
+        }
+    }
+    else if constexpr(is_same<T, bhalf_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_store_i16(src_thread_data,
+                                             dst_wave_buffer_resource,
+                                             dst_thread_addr_offset,
+                                             dst_wave_addr_offset,
+                                             static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 2)
+        {
+            llvm_amdgcn_raw_buffer_store_i16x2(src_thread_data,
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset,
+                                               static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 4)
+        {
+            llvm_amdgcn_raw_buffer_store_i16x4(src_thread_data,
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset,
+                                               static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 8)
+        {
+            vector_type<bhalf_t, 8> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_store_i16x4(tmp.AsType<bhalf4_t>()[Number<0>{}],
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset,
+                                               static_cast<index_t>(coherence));
+
+            llvm_amdgcn_raw_buffer_store_i16x4(tmp.AsType<bhalf4_t>()[Number<1>{}],
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset + 4 * sizeof(bhalf_t),
+                                               static_cast<index_t>(coherence));
+        }
+    }
+    else if constexpr(is_same<T, int32_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_store_i32(src_thread_data,
+                                             dst_wave_buffer_resource,
+                                             dst_thread_addr_offset,
+                                             dst_wave_addr_offset,
+                                             static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 2)
+        {
+            llvm_amdgcn_raw_buffer_store_i32x2(src_thread_data,
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset,
+                                               static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 4)
+        {
+            llvm_amdgcn_raw_buffer_store_i32x4(src_thread_data,
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset,
+                                               static_cast<index_t>(coherence));
+        }
+    }
+    else if constexpr(is_same<T, int8_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_store_i8(src_thread_data,
+                                            dst_wave_buffer_resource,
+                                            dst_thread_addr_offset,
+                                            dst_wave_addr_offset,
+                                            static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 2)
+        {
+#if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE
+            llvm_amdgcn_raw_buffer_store_i8x2(src_thread_data,
+                                              dst_wave_buffer_resource,
+                                              dst_thread_addr_offset,
+                                              dst_wave_addr_offset,
+                                              static_cast<index_t>(coherence));
+#else
+            llvm_amdgcn_raw_buffer_store_i16(bit_cast<int16_t>(src_thread_data),
+                                             dst_wave_buffer_resource,
+                                             dst_thread_addr_offset,
+                                             dst_wave_addr_offset,
+                                             static_cast<index_t>(coherence));
+#endif
+        }
+        else if constexpr(N == 4)
+        {
+#if !CK_WORKAROUND_SWDEV_XXXXXX_INT8_BUFFER_LOAD_STORE_ISSUE
+            llvm_amdgcn_raw_buffer_store_i8x4(src_thread_data,
+                                              dst_wave_buffer_resource,
+                                              dst_thread_addr_offset,
+                                              dst_wave_addr_offset,
+                                              static_cast<index_t>(coherence));
+#else
+            llvm_amdgcn_raw_buffer_store_i32(bit_cast<int32_t>(src_thread_data),
+                                             dst_wave_buffer_resource,
+                                             dst_thread_addr_offset,
+                                             dst_wave_addr_offset,
+                                             static_cast<index_t>(coherence));
+#endif
+        }
+        else if constexpr(N == 8)
+        {
+            llvm_amdgcn_raw_buffer_store_i32x2(bit_cast<int32x2_t>(src_thread_data),
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset,
+                                               static_cast<index_t>(coherence));
+        }
+        else if constexpr(N == 16)
+        {
+            llvm_amdgcn_raw_buffer_store_i32x4(bit_cast<int32x4_t>(src_thread_data),
+                                               dst_wave_buffer_resource,
+                                               dst_thread_addr_offset,
+                                               dst_wave_addr_offset,
+                                               static_cast<index_t>(coherence));
+        }
+    }
+}
+
+template <typename T, index_t N>
+__device__ void amd_buffer_atomic_add_impl(const typename vector_type<T, N>::type src_thread_data,
+                                           int32x4_t dst_wave_buffer_resource,
+                                           index_t dst_thread_addr_offset,
+                                           index_t dst_wave_addr_offset)
+{
+    static_assert((is_same<T, float>::value && (N == 1 || N == 2 || N == 4)) ||
+                      (is_same<T, half_t>::value && (N == 2 || N == 4 || N == 8)) ||
+                      (is_same<T, int32_t>::value && (N == 1 || N == 2 || N == 4)),
+                  "wrong! not implemented");
+
+    if constexpr(is_same<T, float>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_atomic_add_fp32(src_thread_data,
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset,
+                                                   0);
+        }
+        else if constexpr(N == 2)
+        {
+            vector_type<float, 2> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_atomic_add_fp32(tmp.AsType<float>()[Number<0>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset,
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_fp32(tmp.AsType<float>()[Number<1>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + sizeof(float),
+                                                   0);
+        }
+        else if constexpr(N == 4)
+        {
+            vector_type<float, 4> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_atomic_add_fp32(tmp.AsType<float>()[Number<0>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset,
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_fp32(tmp.AsType<float>()[Number<1>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + sizeof(float),
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_fp32(tmp.AsType<float>()[Number<2>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + 2 * sizeof(float),
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_fp32(tmp.AsType<float>()[Number<3>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + 3 * sizeof(float),
+                                                   0);
+        }
+    }
+    else if constexpr(is_same<T, half_t>::value)
+    {
+        if constexpr(N == 2)
+        {
+            llvm_amdgcn_raw_buffer_atomic_add_fp16x2(src_thread_data,
+                                                     dst_wave_buffer_resource,
+                                                     dst_thread_addr_offset,
+                                                     dst_wave_addr_offset,
+                                                     0);
+        }
+        else if constexpr(N == 4)
+        {
+            vector_type<half_t, 4> tmp{src_thread_data};
+
+            static_for<0, 2, 1>{}([&](auto i) {
+                llvm_amdgcn_raw_buffer_atomic_add_fp16x2(tmp.AsType<half2_t>()[i],
+                                                         dst_wave_buffer_resource,
+                                                         dst_thread_addr_offset,
+                                                         dst_wave_addr_offset + i * sizeof(half2_t),
+                                                         0);
+            });
+        }
+        else if constexpr(N == 8)
+        {
+            vector_type<half_t, 8> tmp{src_thread_data};
+
+            static_for<0, 4, 1>{}([&](auto i) {
+                llvm_amdgcn_raw_buffer_atomic_add_fp16x2(tmp.AsType<half2_t>()[i],
+                                                         dst_wave_buffer_resource,
+                                                         dst_thread_addr_offset,
+                                                         dst_wave_addr_offset + i * sizeof(half2_t),
+                                                         0);
+            });
+        }
+    }
+    else if constexpr(is_same<T, int32_t>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_atomic_add_i32(src_thread_data,
+                                                  dst_wave_buffer_resource,
+                                                  dst_thread_addr_offset,
+                                                  dst_wave_addr_offset,
+                                                  0);
+        }
+        else if constexpr(N == 2)
+        {
+            vector_type<int32_t, 2> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_atomic_add_i32(tmp.AsType<int32_t>()[Number<0>{}],
+                                                  dst_wave_buffer_resource,
+                                                  dst_thread_addr_offset,
+                                                  dst_wave_addr_offset,
+                                                  0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_i32(tmp.AsType<int32_t>()[Number<1>{}],
+                                                  dst_wave_buffer_resource,
+                                                  dst_thread_addr_offset,
+                                                  dst_wave_addr_offset + sizeof(int32_t),
+                                                  0);
+        }
+        else if constexpr(N == 4)
+        {
+            vector_type<int32_t, 4> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_atomic_add_i32(tmp.AsType<int32_t>()[Number<0>{}],
+                                                  dst_wave_buffer_resource,
+                                                  dst_thread_addr_offset,
+                                                  dst_wave_addr_offset,
+                                                  0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_i32(tmp.AsType<int32_t>()[Number<1>{}],
+                                                  dst_wave_buffer_resource,
+                                                  dst_thread_addr_offset,
+                                                  dst_wave_addr_offset + sizeof(int32_t),
+                                                  0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_i32(tmp.AsType<int32_t>()[Number<2>{}],
+                                                  dst_wave_buffer_resource,
+                                                  dst_thread_addr_offset,
+                                                  dst_wave_addr_offset + 2 * sizeof(int32_t),
+                                                  0);
+
+            llvm_amdgcn_raw_buffer_atomic_add_i32(tmp.AsType<int32_t>()[Number<3>{}],
+                                                  dst_wave_buffer_resource,
+                                                  dst_thread_addr_offset,
+                                                  dst_wave_addr_offset + 3 * sizeof(int32_t),
+                                                  0);
+        }
+    }
+}
+
+template <typename T, index_t N>
+__device__ void amd_buffer_atomic_max_impl(const typename vector_type<T, N>::type src_thread_data,
+                                           int32x4_t dst_wave_buffer_resource,
+                                           index_t dst_thread_addr_offset,
+                                           index_t dst_wave_addr_offset)
+{
+    static_assert((is_same<T, double>::value && (N == 1 || N == 2 || N == 4)),
+                  "wrong! not implemented");
+    if constexpr(is_same<T, double>::value)
+    {
+        if constexpr(N == 1)
+        {
+            llvm_amdgcn_raw_buffer_atomic_max_fp64(src_thread_data,
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset,
+                                                   0);
+        }
+        else if constexpr(N == 2)
+        {
+            vector_type<double, 2> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType<double>()[Number<0>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset,
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType<double>()[Number<1>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + sizeof(double),
+                                                   0);
+        }
+        else if constexpr(N == 4)
+        {
+            vector_type<double, 4> tmp{src_thread_data};
+
+            llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType<double>()[Number<0>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset,
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType<double>()[Number<1>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + sizeof(double),
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType<double>()[Number<2>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + 2 * sizeof(double),
+                                                   0);
+
+            llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType<double>()[Number<3>{}],
+                                                   dst_wave_buffer_resource,
+                                                   dst_thread_addr_offset,
+                                                   dst_wave_addr_offset + 3 * sizeof(double),
+                                                   0);
+        }
+    }
+}
+
+// buffer_load requires:
+//   1) p_src_wave must point to global memory space
+//   2) p_src_wave must be a wavewise pointer.
+// It is user's responsibility to make sure that is true.
+template <typename T,
+          index_t N,
+          AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence>
+__device__ typename vector_type_maker<T, N>::type::type
+amd_buffer_load_invalid_element_return_zero(const T* p_src_wave,
+                                            index_t src_thread_element_offset,
+                                            bool src_thread_element_valid,
+                                            index_t src_element_space_size)
+{
+    const int32x4_t src_wave_buffer_resource =
+        make_wave_buffer_resource(p_src_wave, src_element_space_size);
+
+    index_t src_thread_addr_offset = src_thread_element_offset * sizeof(T);
+
+    using vector_t = typename vector_type_maker<T, N>::type::type;
+    using scalar_t = typename scalar_type<vector_t>::type;
+
+    constexpr index_t vector_size = scalar_type<vector_t>::vector_size;
+
+#if CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK
+    uint32_t src_addr_shift = src_thread_element_valid ? 0 : 0x80000000;
+
+    if constexpr(is_same<scalar_t, f8_t>::value)
+    {
+        auto tmp = amd_buffer_load_impl<int8_t, vector_size, coherence>(
+            src_wave_buffer_resource, src_addr_shift + src_thread_addr_offset, 0);
+        return bit_cast<vector_t>(tmp);
+    }
+    else
+    {
+        return amd_buffer_load_impl<scalar_t, vector_size, coherence>(
+            src_wave_buffer_resource, src_addr_shift + src_thread_addr_offset, 0);
+    }
+#else
+    if constexpr(is_same<scalar_t, f8_t>::value)
+    {
+        auto tmp = amd_buffer_load_impl<int8_t, vector_size, coherence>(
+            src_wave_buffer_resource, src_thread_addr_offset, 0);
+        return src_thread_element_valid ? bit_cast<vector_t>(tmp) : vector_t(0);
+    }
+    else
+    {
+        vector_t tmp = amd_buffer_load_impl<scalar_t, vector_size, coherence>(
+            src_wave_buffer_resource, src_thread_addr_offset, 0);
+        return src_thread_element_valid ? tmp : vector_t(0);
+    }
+#endif
+}
+
+// buffer_load requires:
+//   1) p_src_wave must point to global memory space
+//   2) p_src_wave must be a wavewise pointer.
+// It is user's responsibility to make sure that is true.
+template <typename T,
+          index_t N,
+          AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence>
+__device__ typename vector_type_maker<T, N>::type::type
+amd_buffer_load_invalid_element_return_customized_value(const T* p_src_wave,
+                                                        index_t src_thread_element_offset,
+                                                        bool src_thread_element_valid,
+                                                        index_t src_element_space_size,
+                                                        T customized_value)
+{
+    const int32x4_t src_wave_buffer_resource =
+        make_wave_buffer_resource(p_src_wave, src_element_space_size);
+
+    index_t src_thread_addr_offset = src_thread_element_offset * sizeof(T);
+
+    using vector_t = typename vector_type_maker<T, N>::type::type;
+    using scalar_t = typename scalar_type<vector_t>::type;
+
+    constexpr index_t vector_size = scalar_type<vector_t>::vector_size;
+
+    vector_t tmp = amd_buffer_load_impl<scalar_t, vector_size, coherence>(
+        src_wave_buffer_resource, src_thread_addr_offset, 0);
+
+    return src_thread_element_valid ? tmp : vector_t(customized_value);
+}
+
+// buffer_store requires:
+//   1) p_dst_wave must point to global memory
+//   2) p_dst_wave must be a wavewise pointer.
+// It is user's responsibility to make sure that is true.
+template <typename T,
+          index_t N,
+          AmdBufferCoherenceEnum coherence = AmdBufferCoherenceEnum::DefaultCoherence>
+__device__ void amd_buffer_store(const typename vector_type_maker<T, N>::type::type src_thread_data,
+                                 T* p_dst_wave,
+                                 const index_t dst_thread_element_offset,
+                                 const bool dst_thread_element_valid,
+                                 const index_t dst_element_space_size)
+{
+    const int32x4_t dst_wave_buffer_resource =
+        make_wave_buffer_resource(p_dst_wave, dst_element_space_size);
+
+    index_t dst_thread_addr_offset = dst_thread_element_offset * sizeof(T);
+
+    using vector_t                = typename vector_type_maker<T, N>::type::type;
+    using scalar_t                = typename scalar_type<vector_t>::type;
+    constexpr index_t vector_size = scalar_type<vector_t>::vector_size;
+
+#if CK_EXPERIMENTAL_USE_BUFFER_STORE_OOB_CHECK_OFFSET_TRICK
+    uint32_t dst_addr_shift = dst_thread_element_valid ? 0 : 0x80000000;
+
+    if constexpr(is_same<scalar_t, f8_t>::value)
+    {
+        auto tmp =
+            bit_cast<typename vector_type_maker<int8_t, vector_size>::type::type>(src_thread_data);
+        amd_buffer_store_impl<int8_t, vector_size, coherence>(
+            tmp, dst_wave_buffer_resource, dst_addr_shift + dst_thread_addr_offset, 0);
+    }
+    else
+    {
+        amd_buffer_store_impl<scalar_t, vector_size, coherence>(
+            src_thread_data, dst_wave_buffer_resource, dst_addr_shift + dst_thread_addr_offset, 0);
+    }
+#else
+    if(dst_thread_element_valid)
+    {
+        if constexpr(is_same<scalar_t, f8_t>::value)
+        {
+            auto tmp = bit_cast<typename vector_type_maker<int8_t, vector_size>::type::type>(
+                src_thread_data);
+            amd_buffer_store_impl<int8_t, vector_size, coherence>(
+                tmp, dst_wave_buffer_resource, dst_thread_addr_offset, 0);
+        }
+        else
+        {
+            amd_buffer_store_impl<scalar_t, vector_size, coherence>(
+                src_thread_data, dst_wave_buffer_resource, dst_thread_addr_offset, 0);
+        }
+    }
+#endif
+}
+
+// buffer_atomic_add requires:
+//   1) p_dst_wave must point to global memory
+//   2) p_dst_wave must be a wavewise pointer.
+// It is user's responsibility to make sure that is true.
+template <typename T, index_t N>
+__device__ void
+amd_buffer_atomic_add(const typename vector_type_maker<T, N>::type::type src_thread_data,
+                      T* p_dst_wave,
+                      const index_t dst_thread_element_offset,
+                      const bool dst_thread_element_valid,
+                      const index_t dst_element_space_size)
+{
+    const int32x4_t dst_wave_buffer_resource =
+        make_wave_buffer_resource(p_dst_wave, dst_element_space_size);
+
+    index_t dst_thread_addr_offset = dst_thread_element_offset * sizeof(T);
+
+    using vector_t                = typename vector_type_maker<T, N>::type::type;
+    using scalar_t                = typename scalar_type<vector_t>::type;
+    constexpr index_t vector_size = scalar_type<vector_t>::vector_size;
+
+#if CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_ADD_OOB_CHECK_OFFSET_TRICK
+    uint32_t dst_addr_shift = dst_thread_element_valid ? 0 : 0x80000000;
+
+    amd_buffer_atomic_add_impl<scalar_t, vector_size>(
+        src_thread_data, dst_wave_buffer_resource, dst_addr_shift + dst_thread_addr_offset, 0);
+#else
+    if(dst_thread_element_valid)
+    {
+        amd_buffer_atomic_add_impl<scalar_t, vector_size>(
+            src_thread_data, dst_wave_buffer_resource, dst_thread_addr_offset, 0);
+    }
+#endif
+}
+
+// buffer_atomic_max requires:
+//   1) p_dst_wave must point to global memory
+//   2) p_dst_wave must be a wavewise pointer.
+// It is user's responsibility to make sure that is true.
+template <typename T, index_t N>
+__device__ void
+amd_buffer_atomic_max(const typename vector_type_maker<T, N>::type::type src_thread_data,
+                      T* p_dst_wave,
+                      const index_t dst_thread_element_offset,
+                      const bool dst_thread_element_valid,
+                      const index_t dst_element_space_size)
+{
+    const int32x4_t dst_wave_buffer_resource =
+        make_wave_buffer_resource(p_dst_wave, dst_element_space_size);
+
+    index_t dst_thread_addr_offset = dst_thread_element_offset * sizeof(T);
+
+    using vector_t                = typename vector_type_maker<T, N>::type::type;
+    using scalar_t                = typename scalar_type<vector_t>::type;
+    constexpr index_t vector_size = scalar_type<vector_t>::vector_size;
+
+#if CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_MAX_OOB_CHECK_OFFSET_TRICK
+    uint32_t dst_addr_shift = dst_thread_element_valid ? 0 : 0x80000000;
+
+    amd_buffer_atomic_max_impl<scalar_t, vector_size>(
+        src_thread_data, dst_wave_buffer_resource, dst_addr_shift + dst_thread_addr_offset, 0);
+#else
+    if(dst_thread_element_valid)
+    {
+        amd_buffer_atomic_max_impl<scalar_t, vector_size>(
+            src_thread_data, dst_wave_buffer_resource, dst_thread_addr_offset, 0);
+    }
+#endif
+}
+
+} // namespace ck
diff --git a/include/ck/utility/amd_gemm_dpp.hpp b/include/ck/utility/amd_gemm_dpp.hpp
new file mode 100644
index 0000000000..8d6c7eede9
--- /dev/null
+++ b/include/ck/utility/amd_gemm_dpp.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/math.hpp"
+#include "ck/utility/amd_gemm_dpp.hpp"
+
+namespace ck {
+
+namespace dpp8 {
+
+/// Number of lanes that can share data using DPP8 modifiers.
+constexpr index_t lane_group_size = 8;
+
+__device__ index_t get_lane_group_local_idx() { return threadIdx.x / lane_group_size; }
+__device__ index_t get_thread_idx_in_lane_group() { return threadIdx.x % lane_group_size; }
+
+} // namespace dpp8
+
+} // namespace ck
diff --git a/include/ck/utility/amd_inline_asm.hpp b/include/ck/utility/amd_inline_asm.hpp
new file mode 100644
index 0000000000..43baa817d3
--- /dev/null
+++ b/include/ck/utility/amd_inline_asm.hpp
@@ -0,0 +1,371 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_AMD_INLINE_ASM_HPP
+#define CK_AMD_INLINE_ASM_HPP
+
+#include "data_type.hpp"
+#include "c_style_pointer_cast.hpp"
+
+// TODO: deprecate all amd_assembly_outer_product_xxx
+
+namespace ck {
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+__device__ void amd_assembly_outer_product_1x2(float a, float b0, float b1, float& c0, float& c1)
+{
+    asm volatile("\n \
+            v_fmac_f32 %0, %2, %3 \n \
+            v_fmac_f32 %1, %2, %4 \n \
+            "
+                 : "=v"(c0), "=v"(c1)
+                 : "v"(a), "v"(b0), "v"(b1), "0"(c0), "1"(c1));
+}
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+// c2 += inner_product(a, b2)
+// c3 += inner_product(a, b3)
+__device__ void amd_assembly_outer_product_1x4(
+    float a, float b0, float b1, float b2, float b3, float& c0, float& c1, float& c2, float& c3)
+{
+    asm volatile("\n \
+            v_fmac_f32 %0, %4, %5 \n \
+            v_fmac_f32 %1, %4, %6 \n \
+            v_fmac_f32 %2, %4, %7 \n \
+            v_fmac_f32 %3, %4, %8 \n \
+            "
+                 : "=v"(c0), "=v"(c1), "=v"(c2), "=v"(c3)
+                 : "v"(a), "v"(b0), "v"(b1), "v"(b2), "v"(b3), "0"(c0), "1"(c1), "2"(c2), "3"(c3));
+}
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+__device__ void
+amd_assembly_outer_product_1x2(half2_t a, half2_t b0, half2_t b1, float& c0, float& c1)
+{
+    asm volatile("\n \
+            v_dot2_f32_f16 %0, %2, %3, %0\n \
+            v_dot2_f32_f16 %1, %2, %4, %1\n \
+            "
+                 : "=v"(c0), "=v"(c1)
+                 : "v"(a), "v"(b0), "v"(b1), "0"(c0), "1"(c1));
+}
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+__device__ void
+amd_assembly_outer_product_1x2(half4_t a, half4_t b0, half4_t b1, float& c0, float& c1)
+{
+    // TODO remove pointer casting
+    const half2_t* p_a_half2  = c_style_pointer_cast<const half2_t*>(&a);
+    const half2_t* p_b0_half2 = c_style_pointer_cast<const half2_t*>(&b0);
+    const half2_t* p_b1_half2 = c_style_pointer_cast<const half2_t*>(&b1);
+
+    // do dot2 two times
+    asm volatile("\n \
+            v_dot2_f32_f16 %0, %2, %4, %0\n \
+            v_dot2_f32_f16 %1, %2, %6, %1\n \
+            v_dot2_f32_f16 %0, %3, %5, %0\n \
+            v_dot2_f32_f16 %1, %3, %7, %1\n \
+            "
+                 : "=v"(c0), "=v"(c1)
+                 : "v"(p_a_half2[0]),
+                   "v"(p_a_half2[1]),
+                   "v"(p_b0_half2[0]),
+                   "v"(p_b0_half2[1]),
+                   "v"(p_b1_half2[0]),
+                   "v"(p_b1_half2[1]),
+                   "0"(c0),
+                   "1"(c1));
+}
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+// c2 += inner_product(a, b2)
+// c3 += inner_product(a, b3)
+__device__ void amd_assembly_outer_product_1x4(half2_t a,
+                                               half2_t b0,
+                                               half2_t b1,
+                                               half2_t b2,
+                                               half2_t b3,
+                                               float& c0,
+                                               float& c1,
+                                               float& c2,
+                                               float& c3)
+{
+    asm volatile("\n \
+            v_dot2_f32_f16 %0, %4, %5, %0\n \
+            v_dot2_f32_f16 %1, %4, %6, %1\n \
+            v_dot2_f32_f16 %2, %4, %7, %2\n \
+            v_dot2_f32_f16 %3, %4, %8, %3\n \
+            "
+                 : "=v"(c0), "=v"(c1), "=v"(c2), "=v"(c3)
+                 : "v"(a), "v"(b0), "v"(b1), "v"(b2), "v"(b3), "0"(c0), "1"(c1), "2"(c2), "3"(c3));
+}
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+// c2 += inner_product(a, b2)
+// c3 += inner_product(a, b3)
+__device__ void amd_assembly_outer_product_1x4(half4_t a,
+                                               half4_t b0,
+                                               half4_t b1,
+                                               half4_t b2,
+                                               half4_t b3,
+                                               float& c0,
+                                               float& c1,
+                                               float& c2,
+                                               float& c3)
+{
+    // TODO remove pointer casting
+    const half2_t* p_a_half2  = c_style_pointer_cast<const half2_t*>(&a);
+    const half2_t* p_b0_half2 = c_style_pointer_cast<const half2_t*>(&b0);
+    const half2_t* p_b1_half2 = c_style_pointer_cast<const half2_t*>(&b1);
+    const half2_t* p_b2_half2 = c_style_pointer_cast<const half2_t*>(&b2);
+    const half2_t* p_b3_half2 = c_style_pointer_cast<const half2_t*>(&b3);
+
+    // do dot2 two times
+    asm volatile("\n \
+            v_dot2_f32_f16 %0, %4, %6,  %0\n \
+            v_dot2_f32_f16 %1, %4, %8,  %1\n \
+            v_dot2_f32_f16 %2, %4, %10, %2\n \
+            v_dot2_f32_f16 %3, %4, %12, %3\n \
+            v_dot2_f32_f16 %0, %5, %7,  %0\n \
+            v_dot2_f32_f16 %1, %5, %9,  %1\n \
+            v_dot2_f32_f16 %2, %5, %11, %2\n \
+            v_dot2_f32_f16 %3, %5, %13, %3\n \
+            "
+                 : "=v"(c0), "=v"(c1), "=v"(c2), "=v"(c3)
+                 : "v"(p_a_half2[0]),
+                   "v"(p_a_half2[1]),
+                   "v"(p_b0_half2[0]),
+                   "v"(p_b0_half2[1]),
+                   "v"(p_b1_half2[0]),
+                   "v"(p_b1_half2[1]),
+                   "v"(p_b2_half2[0]),
+                   "v"(p_b2_half2[1]),
+                   "v"(p_b3_half2[0]),
+                   "v"(p_b3_half2[1]),
+                   "0"(c0),
+                   "1"(c1),
+                   "2"(c2),
+                   "3"(c3));
+}
+
+__device__ void amd_assembly_outer_product_1x4(half8_t a,
+                                               half8_t b0,
+                                               half8_t b1,
+                                               half8_t b2,
+                                               half8_t b3,
+                                               float& c0,
+                                               float& c1,
+                                               float& c2,
+                                               float& c3)
+{
+
+    // TODO remove pointer casting
+    const half4_t* p_a_half4  = c_style_pointer_cast<const half4_t*>(&a);
+    const half4_t* p_b0_half4 = c_style_pointer_cast<const half4_t*>(&b0);
+    const half4_t* p_b1_half4 = c_style_pointer_cast<const half4_t*>(&b1);
+    const half4_t* p_b2_half4 = c_style_pointer_cast<const half4_t*>(&b2);
+    const half4_t* p_b3_half4 = c_style_pointer_cast<const half4_t*>(&b3);
+
+    amd_assembly_outer_product_1x4(
+        p_a_half4[0], p_b0_half4[0], p_b1_half4[0], p_b2_half4[0], p_b3_half4[0], c0, c1, c2, c3);
+
+    amd_assembly_outer_product_1x4(
+        p_a_half4[1], p_b0_half4[1], p_b1_half4[1], p_b2_half4[1], p_b3_half4[1], c0, c1, c2, c3);
+}
+
+__device__ void amd_assembly_outer_product_1x4(half16_t a,
+                                               half16_t b0,
+                                               half16_t b1,
+                                               half16_t b2,
+                                               half16_t b3,
+                                               float& c0,
+                                               float& c1,
+                                               float& c2,
+                                               float& c3)
+{
+    // TODO remove pointer casting
+    const half8_t* p_a_half8  = c_style_pointer_cast<const half8_t*>(&a);
+    const half8_t* p_b0_half8 = c_style_pointer_cast<const half8_t*>(&b0);
+    const half8_t* p_b1_half8 = c_style_pointer_cast<const half8_t*>(&b1);
+    const half8_t* p_b2_half8 = c_style_pointer_cast<const half8_t*>(&b2);
+    const half8_t* p_b3_half8 = c_style_pointer_cast<const half8_t*>(&b3);
+
+    amd_assembly_outer_product_1x4(
+        p_a_half8[0], p_b0_half8[0], p_b1_half8[0], p_b2_half8[0], p_b3_half8[0], c0, c1, c2, c3);
+
+    amd_assembly_outer_product_1x4(
+        p_a_half8[1], p_b0_half8[1], p_b1_half8[1], p_b2_half8[1], p_b3_half8[1], c0, c1, c2, c3);
+}
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+__device__ void
+amd_assembly_outer_product_1x2(int8x4_t a, int8x4_t b0, int8x4_t b1, int32_t& c0, int32_t& c1)
+{
+#if 1
+    asm volatile("\n \
+            v_dot4_i32_i8 %0, %2, %3, %0\n \
+            v_dot4_i32_i8 %1, %2, %4, %1\n \
+            "
+                 : "=v"(c0), "=v"(c1)
+                 : "v"(bit_cast<int32_t>(a)),
+                   "v"(bit_cast<int32_t>(b0)),
+                   "v"(bit_cast<int32_t>(b1)),
+                   "0"(c0),
+                   "1"(c1));
+#else
+    c0     = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b0), c0, false);
+    c1     = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b1), c1, false);
+#endif
+}
+
+// c0 += inner_product(a, b0)
+// c1 += inner_product(a, b1)
+// c2 += inner_product(a, b2)
+// c3 += inner_product(a, b3)
+__device__ void amd_assembly_outer_product_1x4(int8x4_t a,
+                                               int8x4_t b0,
+                                               int8x4_t b1,
+                                               int8x4_t b2,
+                                               int8x4_t b3,
+                                               int32_t& c0,
+                                               int32_t& c1,
+                                               int32_t& c2,
+                                               int32_t& c3)
+{
+#if 1
+    asm volatile("\n \
+            v_dot4_i32_i8 %0, %4, %5, %0\n \
+            v_dot4_i32_i8 %1, %4, %6, %1\n \
+            v_dot4_i32_i8 %2, %4, %7, %2\n \
+            v_dot4_i32_i8 %3, %4, %8, %3\n \
+            "
+                 : "=v"(c0), "=v"(c1), "=v"(c2), "=v"(c3)
+                 : "v"(bit_cast<int32_t>(a)),
+                   "v"(bit_cast<int32_t>(b0)),
+                   "v"(bit_cast<int32_t>(b1)),
+                   "v"(bit_cast<int32_t>(b2)),
+                   "v"(bit_cast<int32_t>(b3)),
+                   "0"(c0),
+                   "1"(c1),
+                   "2"(c2),
+                   "3"(c3));
+#else
+    c0     = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b0), c0, false);
+    c1     = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b1), c1, false);
+    c2     = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b2), c2, false);
+    c3     = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b3), c3, false);
+#endif
+}
+
+__device__ void amd_assembly_outer_product_1x4(int8x8_t a,
+                                               int8x8_t b0,
+                                               int8x8_t b1,
+                                               int8x8_t b2,
+                                               int8x8_t b3,
+                                               int32_t& c0,
+                                               int32_t& c1,
+                                               int32_t& c2,
+                                               int32_t& c3)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    amd_assembly_outer_product_1x4(vector_type<int8_t, 8>{a}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 8>{b0}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 8>{b1}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 8>{b2}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 8>{b3}.AsType<int8x4_t>()[I0],
+                                   c0,
+                                   c1,
+                                   c2,
+                                   c3);
+
+    amd_assembly_outer_product_1x4(vector_type<int8_t, 8>{a}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 8>{b0}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 8>{b1}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 8>{b2}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 8>{b3}.AsType<int8x4_t>()[I1],
+                                   c0,
+                                   c1,
+                                   c2,
+                                   c3);
+}
+
+__device__ void amd_assembly_outer_product_1x4(int8x16_t a,
+                                               int8x16_t b0,
+                                               int8x16_t b1,
+                                               int8x16_t b2,
+                                               int8x16_t b3,
+                                               int32_t& c0,
+                                               int32_t& c1,
+                                               int32_t& c2,
+                                               int32_t& c3)
+
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    amd_assembly_outer_product_1x4(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 16>{b0}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 16>{b1}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 16>{b2}.AsType<int8x4_t>()[I0],
+                                   vector_type<int8_t, 16>{b3}.AsType<int8x4_t>()[I0],
+                                   c0,
+                                   c1,
+                                   c2,
+                                   c3);
+
+    amd_assembly_outer_product_1x4(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 16>{b0}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 16>{b1}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 16>{b2}.AsType<int8x4_t>()[I1],
+                                   vector_type<int8_t, 16>{b3}.AsType<int8x4_t>()[I1],
+                                   c0,
+                                   c1,
+                                   c2,
+                                   c3);
+
+    amd_assembly_outer_product_1x4(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I2],
+                                   vector_type<int8_t, 16>{b0}.AsType<int8x4_t>()[I2],
+                                   vector_type<int8_t, 16>{b1}.AsType<int8x4_t>()[I2],
+                                   vector_type<int8_t, 16>{b2}.AsType<int8x4_t>()[I2],
+                                   vector_type<int8_t, 16>{b3}.AsType<int8x4_t>()[I2],
+                                   c0,
+                                   c1,
+                                   c2,
+                                   c3);
+
+    amd_assembly_outer_product_1x4(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I3],
+                                   vector_type<int8_t, 16>{b0}.AsType<int8x4_t>()[I3],
+                                   vector_type<int8_t, 16>{b1}.AsType<int8x4_t>()[I3],
+                                   vector_type<int8_t, 16>{b2}.AsType<int8x4_t>()[I3],
+                                   vector_type<int8_t, 16>{b3}.AsType<int8x4_t>()[I3],
+                                   c0,
+                                   c1,
+                                   c2,
+                                   c3);
+}
+
+// Ranged input operand
+__device__ void amd_assembly_wmma_f32_16x16x16_f16_w32(half16_t a, half16_t b, float8_t& c)
+{
+#if defined(__gfx11__)
+    asm volatile("v_wmma_f32_16x16x16_f16 %0, %1, %2, %0" : "=v"(c) : "v"(a), "v"(b), "0"(c));
+#else
+    ignore = a;
+    ignore = b;
+    ignore = c;
+#endif
+}
+
+} // namespace ck
+#endif
diff --git a/include/ck/utility/amd_warp_shuffle.hpp b/include/ck/utility/amd_warp_shuffle.hpp
new file mode 100644
index 0000000000..f79a82e2f7
--- /dev/null
+++ b/include/ck/utility/amd_warp_shuffle.hpp
@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+
+namespace ck {
+
+template <typename T>
+__device__ T warp_shuffle_up(const T& v_local, uint32_t lane_delta)
+{
+#if 0
+    return  __shfl_up(v_local, lane_delta);
+#elif 1
+    static_assert(sizeof(T) == sizeof(int32_t), "wrong!");
+
+    const uint32_t wrap_around_lane_delta = warpSize - lane_delta;
+
+    const int32_t v_remote_tmp = __builtin_amdgcn_ds_bpermute(
+        (__lane_id() << 2) + (wrap_around_lane_delta << 2), bit_cast<int32_t>(v_local));
+
+    return bit_cast<T>(v_remote_tmp);
+#endif
+}
+
+template <typename T>
+__device__ T warp_shuffle_down(const T& v_local, uint32_t lane_delta)
+{
+#if 0
+    return  __shfl_down(v_local, lane_delta);
+#elif 1
+    static_assert(sizeof(T) == sizeof(int32_t), "wrong!");
+
+    const int32_t v_remote_tmp = __builtin_amdgcn_ds_bpermute(
+        (__lane_id() << 2) + (lane_delta << 2), bit_cast<int32_t>(v_local));
+
+    return bit_cast<T>(v_remote_tmp);
+#endif
+}
+
+} // namespace ck
diff --git a/include/ck/utility/amd_wave_read_first_lane.hpp b/include/ck/utility/amd_wave_read_first_lane.hpp
new file mode 100644
index 0000000000..741b2975af
--- /dev/null
+++ b/include/ck/utility/amd_wave_read_first_lane.hpp
@@ -0,0 +1,138 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/functional2.hpp"
+#include "ck/utility/math.hpp"
+
+#include <array>
+#include <cstddef>
+#include <cstdint>
+#include <type_traits>
+
+namespace ck {
+namespace detail {
+
+template <unsigned SizeInBytes>
+struct get_carrier;
+
+template <>
+struct get_carrier<1>
+{
+    using type = uint8_t;
+};
+
+template <>
+struct get_carrier<2>
+{
+    using type = uint16_t;
+};
+
+template <>
+struct get_carrier<3>
+{
+    using type = class carrier
+    {
+        using value_type = uint32_t;
+
+        std::array<std::byte, 3> bytes;
+        static_assert(sizeof(bytes) <= sizeof(value_type));
+
+        // replacement of host std::copy_n()
+        template <typename InputIterator, typename Size, typename OutputIterator>
+        __device__ static OutputIterator copy_n(InputIterator from, Size size, OutputIterator to)
+        {
+            if(0 < size)
+            {
+                *to = *from;
+                ++to;
+                for(Size count = 1; count < size; ++count)
+                {
+                    *to = *++from;
+                    ++to;
+                }
+            }
+
+            return to;
+        }
+
+        // method to trigger template substitution failure
+        __device__ carrier(const carrier& other) noexcept
+        {
+            copy_n(other.bytes.begin(), bytes.size(), bytes.begin());
+        }
+
+        public:
+        __device__ carrier& operator=(value_type value) noexcept
+        {
+            copy_n(reinterpret_cast<const std::byte*>(&value), bytes.size(), bytes.begin());
+
+            return *this;
+        }
+
+        __device__ operator value_type() const noexcept
+        {
+            std::byte result[sizeof(value_type)];
+
+            copy_n(bytes.begin(), bytes.size(), result);
+
+            return *reinterpret_cast<const value_type*>(result);
+        }
+    };
+};
+static_assert(sizeof(get_carrier<3>::type) == 3);
+
+template <>
+struct get_carrier<4>
+{
+    using type = uint32_t;
+};
+
+template <unsigned SizeInBytes>
+using get_carrier_t = typename get_carrier<SizeInBytes>::type;
+
+} // namespace detail
+
+__device__ inline int32_t amd_wave_read_first_lane(int32_t value)
+{
+    return __builtin_amdgcn_readfirstlane(value);
+}
+
+template <
+    typename Object,
+    typename = std::enable_if_t<std::is_class_v<Object> && std::is_trivially_copyable_v<Object>>>
+__device__ auto amd_wave_read_first_lane(const Object& obj)
+{
+    using Size                = unsigned;
+    constexpr Size SgprSize   = 4;
+    constexpr Size ObjectSize = sizeof(Object);
+
+    auto* const from_obj = reinterpret_cast<const std::byte*>(&obj);
+    alignas(Object) std::byte to_obj[ObjectSize];
+
+    constexpr Size RemainedSize             = ObjectSize % SgprSize;
+    constexpr Size CompleteSgprCopyBoundary = ObjectSize - RemainedSize;
+    for(Size offset = 0; offset < CompleteSgprCopyBoundary; offset += SgprSize)
+    {
+        using Sgpr = detail::get_carrier_t<SgprSize>;
+
+        *reinterpret_cast<Sgpr*>(to_obj + offset) =
+            amd_wave_read_first_lane(*reinterpret_cast<const Sgpr*>(from_obj + offset));
+    }
+
+    if constexpr(0 < RemainedSize)
+    {
+        using Carrier = detail::get_carrier_t<RemainedSize>;
+
+        *reinterpret_cast<Carrier*>(to_obj + CompleteSgprCopyBoundary) = amd_wave_read_first_lane(
+            *reinterpret_cast<const Carrier*>(from_obj + CompleteSgprCopyBoundary));
+    }
+
+    /// NOTE: Implicitly start object lifetime. It's better to use std::start_lifetime_at() in this
+    /// scenario
+    return *reinterpret_cast<Object*>(to_obj);
+}
+
+} // namespace ck
diff --git a/include/ck/utility/amd_wmma.hpp b/include/ck/utility/amd_wmma.hpp
new file mode 100644
index 0000000000..dd7f0b770a
--- /dev/null
+++ b/include/ck/utility/amd_wmma.hpp
@@ -0,0 +1,258 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_AMD_WMMA_HPP
+#define CK_AMD_WMMA_HPP
+
+#include "ck/utility/amd_inline_asm.hpp"
+#include "data_type.hpp"
+// TODO: Add arch limitation
+namespace ck {
+
+/********************************WAVE32 MODE***********************************************/
+
+// src: fp16, dst: fp32
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_wmma_f32_16x16x16_f16_w32;
+
+template <>
+struct intrin_wmma_f32_16x16x16_f16_w32<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const half16_t& reg_a, const half16_t& reg_b, FloatC& reg_c)
+    {
+        // * Inline assembly need to elimate the duplicated data load, compiler won't help you
+        // delete them.
+        // amd_assembly_wmma_f32_16x16x16_f16_w32(
+        //     reg_a, reg_b, reg_c.template AsType<float8_t>()(Number<0>{}));
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__)
+        reg_c.template AsType<float8_t>()(Number<0>{}) = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32(
+            reg_a, reg_b, reg_c.template AsType<float8_t>()[Number<0>{}]);
+#else
+        ignore = reg_a;
+        ignore = reg_b;
+        ignore = reg_c;
+#endif
+    }
+};
+
+// src: bf16, dst: fp32
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_wmma_f32_16x16x16_bf16_w32;
+
+template <>
+struct intrin_wmma_f32_16x16x16_bf16_w32<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf16_t& reg_a, const bhalf16_t& reg_b, FloatC& reg_c)
+    {
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__)
+        reg_c.template AsType<float8_t>()(Number<0>{}) =
+            __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32(
+                reg_a, reg_b, reg_c.template AsType<float8_t>()[Number<0>{}]);
+#else
+        ignore = reg_a;
+        ignore = reg_b;
+        ignore = reg_c;
+#endif
+    }
+};
+
+// src: fp16, dst: fp16
+template <index_t MPerWave, index_t NPerWave, index_t Opsel>
+struct intrin_wmma_f16_16x16x16_f16_w32;
+
+template <index_t Opsel>
+struct intrin_wmma_f16_16x16x16_f16_w32<16, 16, Opsel>
+{
+    template <class FloatC>
+    __device__ static void Run(const half16_t& reg_a, const half16_t& reg_b, FloatC& reg_c)
+    {
+        // opsel usage
+        // false: D0.[0:15] = result
+        // true : D0.[16:31]= result
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__)
+        reg_c.template AsType<half16_t>()(Number<0>{}) = __builtin_amdgcn_wmma_f16_16x16x16_f16_w32(
+            reg_a, reg_b, reg_c.template AsType<half16_t>()[Number<0>{}], Opsel);
+#else
+        ignore = reg_a;
+        ignore = reg_b;
+        ignore = reg_c;
+#endif
+    }
+};
+
+// src: bf16, dst: bf16
+template <index_t MPerWave, index_t NPerWave, index_t Opsel>
+struct intrin_wmma_bf16_16x16x16_bf16_w32;
+
+template <index_t Opsel>
+struct intrin_wmma_bf16_16x16x16_bf16_w32<16, 16, Opsel>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf16_t& reg_a, const bhalf16_t& reg_b, FloatC& reg_c)
+    {
+        // opsel usage
+        // false: D0.[0:15] = result
+        // true : D0.[16:31]= result
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__)
+        reg_c.template AsType<bhalf16_t>()(Number<0>{}) =
+            __builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32(
+                reg_a, reg_b, reg_c.template AsType<bhalf16_t>()[Number<0>{}], Opsel);
+#else
+        ignore = reg_a;
+        ignore = reg_b;
+        ignore = reg_c;
+#endif
+    }
+};
+
+// src: iu8, dst: i32
+template <index_t MPerWave, index_t NPerWave, bool neg_a, bool neg_b, bool clamp>
+struct intrin_wmma_i32_16x16x16_iu8_w32;
+
+template <bool neg_a, bool neg_b, bool clamp>
+struct intrin_wmma_i32_16x16x16_iu8_w32<16, 16, neg_a, neg_b, clamp>
+{
+    template <class FloatC>
+    __device__ static void Run(const int8x16_t& reg_a, const int8x16_t& reg_b, FloatC& reg_c)
+    {
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__)
+        reg_c.template AsType<int32x8_t>()(Number<0>{}) =
+            __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
+                neg_a,
+                bit_cast<int32x4_t>(reg_a),
+                neg_b,
+                bit_cast<int32x4_t>(reg_b),
+                reg_c.template AsType<int32x8_t>()[Number<0>{}],
+                clamp);
+#else
+        ignore = reg_a;
+        ignore = reg_b;
+        ignore = reg_c;
+#endif
+    }
+};
+
+/********************************WAVE64 MODE***********************************************/
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_wmma_f32_16x16x16_f16_w64;
+
+template <>
+struct intrin_wmma_f32_16x16x16_f16_w64<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const half16_t& reg_a, const half16_t& reg_b, FloatC& reg_c)
+    {
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__)
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_wmma_f32_16x16x16_f16_w64(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}]);
+#else
+        ignore = reg_a;
+        ignore = reg_b;
+        ignore = reg_c;
+#endif
+    }
+};
+
+// src: bf16, dst: fp32
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_wmma_f32_16x16x16_bf16_w64;
+
+template <>
+struct intrin_wmma_f32_16x16x16_bf16_w64<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf16_t& reg_a, const bhalf16_t& reg_b, FloatC& reg_c)
+    {
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__)
+        reg_c.template AsType<float4_t>()(Number<0>{}) =
+            __builtin_amdgcn_wmma_f32_16x16x16_bf16_w64(
+                reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}]);
+#else
+        ignore = reg_a;
+        ignore = reg_b;
+        ignore = reg_c;
+#endif
+    }
+};
+
+// src: fp16, dst: fp16
+template <index_t MPerWave, index_t NPerWave, index_t Opsel>
+struct intrin_wmma_f16_16x16x16_f16_w64;
+
+template <index_t Opsel>
+struct intrin_wmma_f16_16x16x16_f16_w64<16, 16, Opsel>
+{
+    template <class FloatC>
+    __device__ static void Run(const half16_t& reg_a, const half16_t& reg_b, FloatC& reg_c)
+    {
+        // opsel usage
+        // false: D0.[0:15] = result
+        // true : D0.[16:31]= result
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__)
+        reg_c.template AsType<half8_t>()(Number<0>{}) = __builtin_amdgcn_wmma_f16_16x16x16_f16_w64(
+            reg_a, reg_b, reg_c.template AsType<half8_t>()[Number<0>{}], Opsel);
+#else
+        ignore = reg_a;
+        ignore = reg_b;
+        ignore = reg_c;
+#endif
+    }
+};
+
+// src: bf16, dst: bf16
+template <index_t MPerWave, index_t NPerWave, index_t Opsel>
+struct intrin_wmma_bf16_16x16x16_bf16_w64;
+
+template <index_t Opsel>
+struct intrin_wmma_bf16_16x16x16_bf16_w64<16, 16, Opsel>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf16_t& reg_a, const bhalf16_t& reg_b, FloatC& reg_c)
+    {
+        // opsel usage
+        // false: D0.[0:15] = result
+        // true : D0.[16:31]= result
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__)
+        reg_c.template AsType<bhalf8_t>()(Number<0>{}) =
+            __builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64(
+                reg_a, reg_b, reg_c.template AsType<bhalf8_t>()[Number<0>{}], Opsel);
+#else
+        ignore = reg_a;
+        ignore = reg_b;
+        ignore = reg_c;
+#endif
+    }
+};
+
+// src: iu8, dst: i32
+template <index_t MPerWave, index_t NPerWave, bool neg_a, bool neg_b, bool clamp>
+struct intrin_wmma_i32_16x16x16_iu8_w64;
+
+template <bool neg_a, bool neg_b, bool clamp>
+struct intrin_wmma_i32_16x16x16_iu8_w64<16, 16, neg_a, neg_b, clamp>
+{
+    template <class FloatC>
+    __device__ static void Run(const int8x16_t& reg_a, const int8x16_t& reg_b, FloatC& reg_c)
+    {
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__)
+        reg_c.template AsType<int32x4_t>()(Number<0>{}) =
+            __builtin_amdgcn_wmma_i32_16x16x16_iu8_w64(
+                neg_a,
+                bit_cast<int32x4_t>(reg_a),
+                neg_b,
+                bit_cast<int32x4_t>(reg_b),
+                reg_c.template AsType<int32x4_t>()[Number<0>{}],
+                clamp);
+#else
+        ignore = reg_a;
+        ignore = reg_b;
+        ignore = reg_c;
+#endif
+    }
+};
+
+} // namespace ck
+#endif
diff --git a/include/ck/utility/amd_xdlops.hpp b/include/ck/utility/amd_xdlops.hpp
new file mode 100644
index 0000000000..ea7755036f
--- /dev/null
+++ b/include/ck/utility/amd_xdlops.hpp
@@ -0,0 +1,421 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_AMD_XDLOPS_HPP
+#define CK_AMD_XDLOPS_HPP
+
+#include "data_type.hpp"
+
+namespace ck {
+
+// fp32
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_32x32x1f32;
+
+template <>
+struct intrin_mfma_f32_32x32x1f32<64, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const float& reg_a, const float& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float32_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x1f32(
+            reg_a, reg_b, reg_c.template AsType<float32_t>()[Number<0>{}], 1, 0, 0);
+        reg_c.template AsType<float32_t>()(Number<1>{}) = __builtin_amdgcn_mfma_f32_32x32x1f32(
+            reg_a, reg_b, reg_c.template AsType<float32_t>()[Number<1>{}], 1, 1, 0);
+    }
+};
+
+template <>
+struct intrin_mfma_f32_32x32x1f32<32, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const float& reg_a, const float& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float32_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x1f32(
+            reg_a, reg_b, reg_c.template AsType<float32_t>()[Number<0>{}], 1, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_32x32x2f32;
+
+template <>
+struct intrin_mfma_f32_32x32x2f32<32, 32>
+{
+    template <class FloatC>
+    __device__ static void Run(const float& reg_a, const float& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x2f32(
+            reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_16x16x4f32;
+
+template <>
+struct intrin_mfma_f32_16x16x4f32<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const float& reg_a, const float& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x4f32(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_16x16x1f32;
+
+template <>
+struct intrin_mfma_f32_16x16x1f32<16, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const float& reg_a, const float& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x1f32(
+            reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 2, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_4x4x1f32;
+
+template <>
+struct intrin_mfma_f32_4x4x1f32<4, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const float& reg_a, const float& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_4x4x1f32(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 4, 0, 0);
+    }
+};
+
+template <>
+struct intrin_mfma_f32_4x4x1f32<8, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const float& reg_a, const float& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_4x4x1f32(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 4, 0, 0);
+        reg_c.template AsType<float4_t>()(Number<1>{}) = __builtin_amdgcn_mfma_f32_4x4x1f32(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<1>{}], 4, 1, 0);
+    }
+};
+
+// fp16
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_32x32x4f16;
+
+template <>
+struct intrin_mfma_f32_32x32x4f16<64, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const half4_t& reg_a, const half4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float32_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x4f16(
+            reg_a, reg_b, reg_c.template AsType<float32_t>()[Number<0>{}], 1, 0, 0);
+        reg_c.template AsType<float32_t>()(Number<1>{}) = __builtin_amdgcn_mfma_f32_32x32x4f16(
+            reg_a, reg_b, reg_c.template AsType<float32_t>()[Number<1>{}], 1, 1, 0);
+    }
+};
+
+template <>
+struct intrin_mfma_f32_32x32x4f16<32, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const half4_t& reg_a, const half4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float32_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x4f16(
+            reg_a, reg_b, reg_c.template AsType<float32_t>()[Number<0>{}], 1, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_32x32x8f16;
+
+template <>
+struct intrin_mfma_f32_32x32x8f16<32, 32>
+{
+    template <class FloatC>
+    __device__ static void Run(const half4_t& reg_a, const half4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x8f16(
+            reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_16x16x16f16;
+
+template <>
+struct intrin_mfma_f32_16x16x16f16<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const half4_t& reg_a, const half4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x16f16(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_16x16x4f16;
+
+template <>
+struct intrin_mfma_f32_16x16x4f16<16, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const half4_t& reg_a, const half4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x4f16(
+            reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 2, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_4x4x4f16;
+
+template <>
+struct intrin_mfma_f32_4x4x4f16<4, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const half4_t& reg_a, const half4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_4x4x4f16(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 4, 0, 0);
+    }
+};
+
+template <>
+struct intrin_mfma_f32_4x4x4f16<8, 64>
+{
+    template <class FloatC>
+    __device__ static void Run(const half4_t& reg_a, const half4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_4x4x4f16(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 4, 0, 0);
+        reg_c.template AsType<float4_t>()(Number<1>{}) = __builtin_amdgcn_mfma_f32_4x4x4f16(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<1>{}], 4, 1, 0);
+    }
+};
+
+// bfp16
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_32x32x8bf16_1k;
+
+template <>
+struct intrin_mfma_f32_32x32x8bf16_1k<32, 32>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf4_t& reg_a, const bhalf4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x8bf16_1k(
+            reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_16x16x16bf16_1k;
+
+template <>
+struct intrin_mfma_f32_16x16x16bf16_1k<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf4_t& reg_a, const bhalf4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_32x32x4bf16;
+
+template <>
+struct intrin_mfma_f32_32x32x4bf16<32, 32>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf2_t& reg_a, const bhalf2_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float16_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_32x32x4bf16(
+            reg_a, reg_b, reg_c.template AsType<float16_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_16x16x8bf16;
+
+template <>
+struct intrin_mfma_f32_16x16x8bf16<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const bhalf2_t& reg_a, const bhalf2_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x8bf16(
+            reg_a, reg_b, reg_c.template AsType<float4_t>()[Number<0>{}], 0, 0, 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_i32_32x32x8i8;
+
+template <>
+struct intrin_mfma_i32_32x32x8i8<32, 32>
+{
+    template <class FloatC>
+    __device__ static void Run(const int8x4_t& reg_a, const int8x4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<int32x16_t>()(Number<0>{}) =
+            __builtin_amdgcn_mfma_i32_32x32x8i8(bit_cast<int32_t>(reg_a),
+                                                bit_cast<int32_t>(reg_b),
+                                                reg_c.template AsType<int32x16_t>()[Number<0>{}],
+                                                0,
+                                                0,
+                                                0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_i32_16x16x16i8;
+
+template <>
+struct intrin_mfma_i32_16x16x16i8<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const int8x4_t& reg_a, const int8x4_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<int32x4_t>()(Number<0>{}) =
+            __builtin_amdgcn_mfma_i32_16x16x16i8(bit_cast<int32_t>(reg_a),
+                                                 bit_cast<int32_t>(reg_b),
+                                                 reg_c.template AsType<int32x4_t>()[Number<0>{}],
+                                                 0,
+                                                 0,
+                                                 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_i32_32x32x16i8;
+
+template <>
+struct intrin_mfma_i32_32x32x16i8<32, 32>
+{
+    template <class FloatC>
+    __device__ static void Run(const int8x8_t& reg_a, const int8x8_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<int32x16_t>()(Number<0>{}) =
+            __builtin_amdgcn_mfma_i32_32x32x16_i8(bit_cast<int64_t>(reg_a),
+                                                  bit_cast<int64_t>(reg_b),
+                                                  reg_c.template AsType<int32x16_t>()[Number<0>{}],
+                                                  0,
+                                                  0,
+                                                  0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_i32_16x16x32i8;
+
+template <>
+struct intrin_mfma_i32_16x16x32i8<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const int8x8_t& reg_a, const int8x8_t& reg_b, FloatC& reg_c)
+    {
+        reg_c.template AsType<int32x4_t>()(Number<0>{}) =
+            __builtin_amdgcn_mfma_i32_16x16x32i8(bit_cast<int64_t>(reg_a),
+                                                 bit_cast<int64_t>(reg_b),
+                                                 reg_c.template AsType<int32x4_t>()[Number<0>{}],
+                                                 0,
+                                                 0,
+                                                 0);
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f64_16x16x4f64;
+
+template <>
+struct intrin_mfma_f64_16x16x4f64<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const double& reg_a, const double& reg_b, FloatC& reg_c)
+    {
+#if defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+        reg_c.template AsType<double4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f64_16x16x4f64(
+            reg_a, reg_b, reg_c.template AsType<double4_t>()[Number<0>{}], 0, 0, 0);
+#else
+        ignore = reg_a;
+        ignore = reg_b;
+        ignore = reg_c;
+#endif
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_32x32x16f8f8;
+
+template <>
+struct intrin_mfma_f32_32x32x16f8f8<32, 32>
+{
+    template <class FloatC>
+    __device__ static void Run(const f8x8_t& reg_a, const f8x8_t& reg_b, FloatC& reg_c)
+    {
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+        reg_c.template AsType<float16_t>()(Number<0>{}) =
+            __builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8(
+                bit_cast<long>(reg_a),
+                bit_cast<long>(reg_b),
+                reg_c.template AsType<float16_t>()[Number<0>{}],
+                0,
+                0,
+                0);
+#else
+        vector_type<f8_t, 8> reg_a_v(reg_a);
+        vector_type<f8_t, 8> reg_b_v(reg_b);
+
+        static_for<0, 8, 1>{}([&](auto k) {
+            float reg_a_f32 = type_convert<float>(reg_a_v.template AsType<f8_t>()[Number<k>{}]);
+            float reg_b_f32 = type_convert<float>(reg_b_v.template AsType<f8_t>()[Number<k>{}]);
+
+            intrin_mfma_f32_32x32x2f32<32, 32>::Run(reg_a_f32, reg_b_f32, reg_c);
+        });
+#endif
+    }
+};
+
+template <index_t MPerWave, index_t NPerWave>
+struct intrin_mfma_f32_16x16x32f8f8;
+
+template <>
+struct intrin_mfma_f32_16x16x32f8f8<16, 16>
+{
+    template <class FloatC>
+    __device__ static void Run(const f8x8_t& reg_a, const f8x8_t& reg_b, FloatC& reg_c)
+    {
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+        reg_c.template AsType<float4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(
+            bit_cast<long>(reg_a),
+            bit_cast<long>(reg_b),
+            reg_c.template AsType<float4_t>()[Number<0>{}],
+            0,
+            0,
+            0);
+#else
+        vector_type<f8_t, 8> reg_a_v(reg_a);
+        vector_type<f8_t, 8> reg_b_v(reg_b);
+
+        static_for<0, 8, 1>{}([&](auto k) {
+            float reg_a_f32 = type_convert<float>(reg_a_v.template AsType<f8_t>()[Number<k>{}]);
+            float reg_b_f32 = type_convert<float>(reg_b_v.template AsType<f8_t>()[Number<k>{}]);
+
+            intrin_mfma_f32_16x16x4f32<16, 16>::Run(reg_a_f32, reg_b_f32, reg_c);
+        });
+#endif
+    }
+};
+} // namespace ck
+#endif
diff --git a/include/ck/utility/array.hpp b/include/ck/utility/array.hpp
new file mode 100644
index 0000000000..15ee175346
--- /dev/null
+++ b/include/ck/utility/array.hpp
@@ -0,0 +1,170 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <initializer_list>
+
+#include "ck/utility/functional2.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+
+template <typename TData, index_t NSize>
+struct Array
+{
+    using type      = Array;
+    using data_type = TData;
+
+    TData mData[NSize];
+
+    __host__ __device__ constexpr Array() : mData{} {}
+
+    __host__ __device__ constexpr Array(std::initializer_list<TData> ilist)
+    {
+        constexpr index_t list_size = std::initializer_list<TData>{}.size();
+
+        static_assert(list_size <= NSize, "out of bound");
+
+        index_t i   = 0;
+        TData vlast = TData{};
+
+        for(const TData& val : ilist)
+        {
+            mData[i] = val;
+            vlast    = val;
+            ++i;
+        }
+
+        for(; i < NSize; ++i)
+        {
+            mData[i] = vlast;
+        }
+    }
+
+    __host__ __device__ static constexpr index_t Size() { return NSize; }
+
+    template <index_t I>
+    __host__ __device__ constexpr const TData& At() const
+    {
+        return mData[I];
+    }
+
+    template <index_t I>
+    __host__ __device__ constexpr TData& At()
+    {
+        return mData[I];
+    }
+
+    __host__ __device__ constexpr const TData& At(index_t i) const { return mData[i]; }
+
+    __host__ __device__ constexpr TData& At(index_t i) { return mData[i]; }
+
+    __host__ __device__ constexpr const TData& operator[](index_t i) const { return mData[i]; }
+
+    __host__ __device__ constexpr TData& operator()(index_t i) { return mData[i]; }
+
+    template <typename T>
+    __host__ __device__ constexpr auto operator=(const T& a)
+    {
+        static_assert(T::Size() == Size(), "wrong! size not the same");
+
+        for(index_t i = 0; i < NSize; ++i)
+        {
+            mData[i] = a[i];
+        }
+
+        return *this;
+    }
+
+    __host__ __device__ static constexpr bool IsStatic() { return is_static_v<TData>; };
+
+    __host__ __device__ void Print() const
+    {
+        printf("Array{size: %d, data: [", NSize);
+
+        for(index_t i = 0; i < NSize; i++)
+        {
+            print(mData[i]);
+
+            if(i < NSize - 1)
+            {
+                printf(", ");
+            }
+        }
+
+        printf("]}");
+    }
+};
+
+// empty Array
+template <typename TData>
+struct Array<TData, 0>
+{
+    using type      = Array;
+    using data_type = TData;
+
+    __host__ __device__ constexpr Array() {}
+
+    __host__ __device__ static constexpr index_t Size() { return 0; }
+
+    __host__ __device__ static constexpr bool IsStatic() { return is_static_v<TData>; };
+
+    __host__ __device__ void Print() const { printf("Array{size: 0, data: []}"); }
+};
+
+template <typename TData, index_t NSize>
+__host__ __device__ constexpr bool operator==(const Array<TData, NSize>& a,
+                                              const Array<TData, NSize>& b)
+{
+    bool same = true;
+
+    for(index_t i = 0; i < NSize; ++i)
+    {
+        if(a[i] != b[i])
+        {
+            same = false;
+            break;
+        }
+    }
+
+    return same;
+}
+
+template <typename TData, index_t NSize>
+__host__ __device__ constexpr bool operator!=(const Array<TData, NSize>& a,
+                                              const Array<TData, NSize>& b)
+{
+    return !(a == b);
+}
+
+template <typename T, typename... Xs>
+__host__ __device__ constexpr auto make_array(Xs&&... xs)
+{
+    using data_type = remove_cvref_t<T>;
+
+    return Array<data_type, sizeof...(Xs)>{std::forward<Xs>(xs)...};
+}
+
+template <typename F, index_t N>
+__host__ __device__ constexpr auto generate_array(F&& f, Number<N>)
+{
+    using T = remove_cvref_t<decltype(f(Number<0>{}))>;
+
+    return unpack([&f](auto&&... is) { return Array<T, N>{f(is)...}; },
+                  typename arithmetic_sequence_gen<0, N, 1>::type{});
+}
+
+template <typename T, index_t N, typename X>
+__host__ __device__ constexpr auto to_array(const X& x)
+{
+    STATIC_ASSERT(N <= X::Size(), "");
+
+    Array<T, N> arr;
+
+    static_for<0, N, 1>{}([&x, &arr](auto i) { arr(i) = x[i]; });
+
+    return arr;
+}
+
+} // namespace ck
diff --git a/include/ck/utility/array_multi_index.hpp b/include/ck/utility/array_multi_index.hpp
new file mode 100644
index 0000000000..8c0f5ff769
--- /dev/null
+++ b/include/ck/utility/array_multi_index.hpp
@@ -0,0 +1,78 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "common_header.hpp"
+
+namespace ck {
+
+template <index_t N>
+using MultiIndex = Array<index_t, N>;
+
+template <typename... Xs>
+__host__ __device__ constexpr auto make_multi_index(Xs&&... xs)
+{
+    return make_array<index_t>(index_t{xs}...);
+}
+
+template <index_t NSize>
+__host__ __device__ constexpr auto make_zero_multi_index()
+{
+    return unpack([](auto... xs) { return make_multi_index(xs...); },
+                  typename uniform_sequence_gen<NSize, 0>::type{});
+}
+
+template <typename T>
+__host__ __device__ constexpr auto to_multi_index(const T& x)
+{
+    return unpack([](auto... ys) { return make_multi_index(ys...); }, x);
+}
+
+template <index_t NSize, typename X>
+__host__ __device__ constexpr auto operator+=(MultiIndex<NSize>& y, const X& x)
+{
+    static_assert(X::Size() == NSize, "wrong! size not the same");
+    static_for<0, NSize, 1>{}([&](auto i) { y(i) += x[i]; });
+    return y;
+}
+
+template <index_t NSize, typename X>
+__host__ __device__ constexpr auto operator-=(MultiIndex<NSize>& y, const X& x)
+{
+    static_assert(X::Size() == NSize, "wrong! size not the same");
+    static_for<0, NSize, 1>{}([&](auto i) { y(i) -= x[i]; });
+    return y;
+}
+
+template <index_t NSize, typename T>
+__host__ __device__ constexpr auto operator+(const MultiIndex<NSize>& a, const T& b)
+{
+    using type = MultiIndex<NSize>;
+    static_assert(T::Size() == NSize, "wrong! size not the same");
+    type r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = a[i] + b[i]; });
+    return r;
+}
+
+template <index_t NSize, typename T>
+__host__ __device__ constexpr auto operator-(const MultiIndex<NSize>& a, const T& b)
+{
+    using type = MultiIndex<NSize>;
+    static_assert(T::Size() == NSize, "wrong! size not the same");
+    type r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = a[i] - b[i]; });
+    return r;
+}
+
+template <index_t NSize, typename T>
+__host__ __device__ constexpr auto operator*(const MultiIndex<NSize>& a, const T& b)
+{
+    using type = MultiIndex<NSize>;
+    static_assert(T::Size() == NSize, "wrong! size not the same");
+    type r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = a[i] * b[i]; });
+    return r;
+}
+
+} // namespace ck
diff --git a/include/ck/utility/bit_cast.hpp b/include/ck/utility/bit_cast.hpp
new file mode 100644
index 0000000000..c6d729e025
--- /dev/null
+++ b/include/ck/utility/bit_cast.hpp
@@ -0,0 +1,19 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/enable_if.hpp"
+
+namespace ck {
+
+template <typename Y, typename X, typename enable_if<sizeof(X) == sizeof(Y), bool>::type = false>
+__host__ __device__ constexpr Y bit_cast(const X& x)
+{
+    static_assert(__has_builtin(__builtin_bit_cast), "");
+
+    return __builtin_bit_cast(Y, x);
+}
+
+} // namespace ck
diff --git a/include/ck/utility/buffer_view.hpp b/include/ck/utility/buffer_view.hpp
new file mode 100644
index 0000000000..f5eadb1b50
--- /dev/null
+++ b/include/ck/utility/buffer_view.hpp
@@ -0,0 +1,38 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/buffer_view_declare.hpp"
+#include "ck/utility/buffer_view_impl_generic.hpp"
+#include "ck/utility/buffer_view_impl_global.hpp"
+#include "ck/utility/buffer_view_impl_lds.hpp"
+#include "ck/utility/buffer_view_impl_vgpr.hpp"
+
+namespace ck {
+
+template <AddressSpaceEnum BufferAddressSpace,
+          AmdBufferCoherenceEnum Coherence = AmdBufferCoherenceEnum::DefaultCoherence,
+          typename T,
+          typename BufferSizeType>
+__host__ __device__ constexpr auto make_buffer_view(T* p, BufferSizeType buffer_size)
+{
+    return BufferView<BufferAddressSpace, T, BufferSizeType, true, Coherence>{p, buffer_size};
+}
+
+template <
+    AddressSpaceEnum BufferAddressSpace,
+    AmdBufferCoherenceEnum Coherence = AmdBufferCoherenceEnum::DefaultCoherence,
+    typename T,
+    typename BufferSizeType,
+    typename X,
+    typename enable_if<is_same<remove_cvref_t<T>, remove_cvref_t<X>>::value, bool>::type = false>
+__host__ __device__ constexpr auto
+make_buffer_view(T* p, BufferSizeType buffer_size, X invalid_element_value)
+{
+    return BufferView<BufferAddressSpace, T, BufferSizeType, false, Coherence>{
+        p, buffer_size, invalid_element_value};
+}
+
+} // namespace ck
diff --git a/include/ck/utility/buffer_view_declare.hpp b/include/ck/utility/buffer_view_declare.hpp
new file mode 100644
index 0000000000..747f1ab63b
--- /dev/null
+++ b/include/ck/utility/buffer_view_declare.hpp
@@ -0,0 +1,29 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/enable_if.hpp"
+#include "ck/utility/c_style_pointer_cast.hpp"
+
+namespace ck {
+
+// T may be scalar or vector
+// X may be scalar or vector
+// T and X have same scalar type
+// X contains multiple T
+// FIXME: InvalidElementUseNumericalZeroValue and invalid_element_value_ should be a property of
+//        transforms of TensorView/Tensor
+// FIXME: AmdBufferCoherenceEnum is only meaningful for buffer addressing. Need to split BufferView
+// definition for different memory address space (Global/GenericLds/Vgpr)
+template <AddressSpaceEnum BufferAddressSpace,
+          typename T,
+          typename BufferSizeType,
+          bool InvalidElementUseNumericalZeroValue,
+          AmdBufferCoherenceEnum Coherence = AmdBufferCoherenceEnum::DefaultCoherence>
+struct BufferView;
+
+} // namespace ck
diff --git a/include/ck/utility/buffer_view_impl_generic.hpp b/include/ck/utility/buffer_view_impl_generic.hpp
new file mode 100644
index 0000000000..78c7b8e9a0
--- /dev/null
+++ b/include/ck/utility/buffer_view_impl_generic.hpp
@@ -0,0 +1,177 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/buffer_view_declare.hpp"
+#include "ck/utility/generic_memory_space_atomic.hpp"
+
+namespace ck {
+
+// Address Space: Generic
+// T may be scalar or vector
+// X may be scalar or vector
+// T and X have same scalar type
+// X contains multiple T
+// FIXME: InvalidElementUseNumericalZeroValue and invalid_element_value_ should be a property of
+//        transforms of TensorView/Tensor
+template <typename T, typename BufferSizeType, bool InvalidElementUseNumericalZeroValue>
+struct BufferView<AddressSpaceEnum::Generic,
+                  T,
+                  BufferSizeType,
+                  InvalidElementUseNumericalZeroValue,
+                  AmdBufferCoherenceEnum::DefaultCoherence>
+{
+    using type = T;
+
+    T* p_data_ = nullptr;
+    BufferSizeType buffer_size_;
+    remove_cvref_t<T> invalid_element_value_ = T{0};
+
+    __host__ __device__ constexpr BufferView() : p_data_{}, buffer_size_{}, invalid_element_value_{}
+    {
+    }
+
+    __host__ __device__ constexpr BufferView(T* p_data, BufferSizeType buffer_size)
+        : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{0}
+    {
+    }
+
+    __host__ __device__ constexpr BufferView(T* p_data,
+                                             BufferSizeType buffer_size,
+                                             T invalid_element_value)
+        : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{invalid_element_value}
+    {
+    }
+
+    __device__ static constexpr AddressSpaceEnum GetAddressSpace()
+    {
+        return AddressSpaceEnum::Generic;
+    }
+
+    // i is offset of T
+    // FIXME: doesn't do is_valid check
+    __device__ constexpr const T& operator[](index_t i) const { return p_data_[i]; }
+
+    // i is offset of T
+    // FIXME: doesn't do is_valid check
+    __device__ constexpr T& operator()(index_t i) { return p_data_[i]; }
+
+    // i is offset of T, not X. i should be aligned to X
+    template <typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __device__ constexpr auto Get(index_t i, bool is_valid_element) const
+    {
+        // X contains multiple T
+        constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
+
+        constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
+
+        static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
+                      "wrong! X should contain multiple T");
+
+        if(is_valid_element)
+        {
+#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
+            X tmp;
+
+            __builtin_memcpy(&tmp, &(p_data_[i]), sizeof(X));
+
+            return tmp;
+#else
+            return *c_style_pointer_cast<const X*>(&p_data_[i]);
+#endif
+        }
+        else
+        {
+            if constexpr(InvalidElementUseNumericalZeroValue)
+            {
+                return X{0};
+            }
+            else
+            {
+                return X{invalid_element_value_};
+            }
+        }
+    }
+
+    // i is offset of T, not X. i should be aligned to X
+    template <InMemoryDataOperationEnum Op,
+              typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __device__ void Update(index_t i, bool is_valid_element, const X& x)
+    {
+        if constexpr(Op == InMemoryDataOperationEnum::Set)
+        {
+            this->template Set<X>(i, is_valid_element, x);
+        }
+        // FIXME: remove InMemoryDataOperationEnum::Add
+        else if constexpr(Op == InMemoryDataOperationEnum::Add)
+        {
+            auto tmp = this->template Get<X>(i, is_valid_element);
+            this->template Set<X>(i, is_valid_element, x + tmp);
+        }
+    }
+
+    // i is offset of T, not X. i should be aligned to X
+    template <typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __device__ void Set(index_t i, bool is_valid_element, const X& x)
+    {
+        // X contains multiple T
+        constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
+
+        constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
+
+        static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
+                      "wrong! X should contain multiple T");
+
+        if(is_valid_element)
+        {
+#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
+            X tmp = x;
+
+            __builtin_memcpy(&(p_data_[i]), &tmp, sizeof(X));
+#else
+            *c_style_pointer_cast<X*>(&p_data_[i]) = x;
+#endif
+        }
+    }
+
+    // FIXME: remove
+    __device__ static constexpr bool IsStaticBuffer() { return false; }
+
+    // FIXME: remove
+    __device__ static constexpr bool IsDynamicBuffer() { return true; }
+
+    __host__ __device__ void Print() const
+    {
+        printf("BufferView{");
+
+        // AddressSpace
+        printf("AddressSpace: Generic, ");
+
+        // p_data_
+        printf("p_data_: %p, ", static_cast<void*>(const_cast<remove_cvref_t<T>*>(p_data_)));
+
+        // buffer_size_
+        printf("buffer_size_: ");
+        print(buffer_size_);
+        printf(", ");
+
+        // invalid_element_value_
+        printf("invalid_element_value_: ");
+        print(invalid_element_value_);
+
+        printf("}");
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/utility/buffer_view_impl_global.hpp b/include/ck/utility/buffer_view_impl_global.hpp
new file mode 100644
index 0000000000..92ca62d15d
--- /dev/null
+++ b/include/ck/utility/buffer_view_impl_global.hpp
@@ -0,0 +1,319 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/buffer_view.hpp"
+#include "ck/utility/amd_buffer_addressing.hpp"
+
+namespace ck {
+
+// Address Space: Global
+// T may be scalar or vector
+// X may be scalar or vector
+// T and X have same scalar type
+// X contains multiple T
+// FIXME: InvalidElementUseNumericalZeroValue and invalid_element_value_ should be a property of
+//        transforms of TensorView/Tensor
+template <typename T,
+          typename BufferSizeType,
+          bool InvalidElementUseNumericalZeroValue,
+          AmdBufferCoherenceEnum Coherence>
+struct BufferView<AddressSpaceEnum::Global,
+                  T,
+                  BufferSizeType,
+                  InvalidElementUseNumericalZeroValue,
+                  Coherence>
+{
+    using type = T;
+
+    T* p_data_ = nullptr;
+    BufferSizeType buffer_size_;
+    remove_cvref_t<T> invalid_element_value_ = T{0};
+
+    __host__ __device__ constexpr BufferView() : p_data_{}, buffer_size_{}, invalid_element_value_{}
+    {
+    }
+
+    __host__ __device__ constexpr BufferView(T* p_data, BufferSizeType buffer_size)
+        : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{0}
+    {
+    }
+
+    __host__ __device__ constexpr BufferView(T* p_data,
+                                             BufferSizeType buffer_size,
+                                             T invalid_element_value)
+        : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{invalid_element_value}
+    {
+    }
+
+    __device__ static constexpr AddressSpaceEnum GetAddressSpace()
+    {
+        return AddressSpaceEnum::Global;
+    }
+
+    // i is offset of T
+    // FIXME: doesn't do is_valid check
+    __device__ constexpr const T& operator[](index_t i) const { return p_data_[i]; }
+
+    // i is offset of T
+    // FIXME: doesn't do is_valid check
+    __device__ constexpr T& operator()(index_t i) { return p_data_[i]; }
+
+    // i is offset of T, not X. i should be aligned to X
+    template <typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __device__ constexpr auto Get(index_t i, bool is_valid_element) const
+    {
+        // X contains multiple T
+        constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
+
+        constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
+
+        static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
+                      "wrong! X should contain multiple T");
+
+#if CK_USE_AMD_BUFFER_LOAD
+        bool constexpr use_amd_buffer_addressing = true;
+#else
+        bool constexpr use_amd_buffer_addressing = false;
+#endif
+
+        if constexpr(use_amd_buffer_addressing)
+        {
+            constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
+
+            if constexpr(InvalidElementUseNumericalZeroValue)
+            {
+                return amd_buffer_load_invalid_element_return_zero<remove_cvref_t<T>,
+                                                                   t_per_x,
+                                                                   Coherence>(
+                    p_data_, i, is_valid_element, buffer_size_);
+            }
+            else
+            {
+                return amd_buffer_load_invalid_element_return_customized_value<remove_cvref_t<T>,
+                                                                               t_per_x,
+                                                                               Coherence>(
+                    p_data_, i, is_valid_element, buffer_size_, invalid_element_value_);
+            }
+        }
+        else
+        {
+            if(is_valid_element)
+            {
+#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
+                X tmp;
+
+                __builtin_memcpy(&tmp, &(p_data_[i]), sizeof(X));
+
+                return tmp;
+#else
+                return *c_style_pointer_cast<const X*>(&p_data_[i]);
+#endif
+            }
+            else
+            {
+                if constexpr(InvalidElementUseNumericalZeroValue)
+                {
+                    return X{0};
+                }
+                else
+                {
+                    return X{invalid_element_value_};
+                }
+            }
+        }
+    }
+
+    // i is offset of T, not X. i should be aligned to X
+    template <InMemoryDataOperationEnum Op,
+              typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __device__ void Update(index_t i, bool is_valid_element, const X& x)
+    {
+        if constexpr(Op == InMemoryDataOperationEnum::Set)
+        {
+            this->template Set<X>(i, is_valid_element, x);
+        }
+        else if constexpr(Op == InMemoryDataOperationEnum::AtomicAdd)
+        {
+            this->template AtomicAdd<X>(i, is_valid_element, x);
+        }
+        else if constexpr(Op == InMemoryDataOperationEnum::AtomicMax)
+        {
+            this->template AtomicMax<X>(i, is_valid_element, x);
+        }
+        // FIXME: remove InMemoryDataOperationEnum::Add
+        else if constexpr(Op == InMemoryDataOperationEnum::Add)
+        {
+            auto tmp = this->template Get<X>(i, is_valid_element);
+            this->template Set<X>(i, is_valid_element, x + tmp);
+            // tmp += x;
+            // this->template Set<X>(i, is_valid_element, tmp);
+        }
+    }
+
+    // i is offset of T, not X. i should be aligned to X
+    template <typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __device__ void Set(index_t i, bool is_valid_element, const X& x)
+    {
+        // X contains multiple T
+        constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
+
+        constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
+
+        static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
+                      "wrong! X should contain multiple T");
+
+#if CK_USE_AMD_BUFFER_STORE
+        bool constexpr use_amd_buffer_addressing = true;
+#else
+        bool constexpr use_amd_buffer_addressing = false;
+#endif
+
+        if constexpr(use_amd_buffer_addressing)
+        {
+            constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
+
+            amd_buffer_store<remove_cvref_t<T>, t_per_x, Coherence>(
+                x, p_data_, i, is_valid_element, buffer_size_);
+        }
+        else
+        {
+            if(is_valid_element)
+            {
+#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
+                X tmp = x;
+
+                __builtin_memcpy(&(p_data_[i]), &tmp, sizeof(X));
+#else
+                *c_style_pointer_cast<X*>(&p_data_[i]) = x;
+#endif
+            }
+        }
+    }
+
+    template <typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __device__ void AtomicAdd(index_t i, bool is_valid_element, const X& x)
+    {
+        using scalar_t = typename scalar_type<remove_cvref_t<T>>::type;
+
+        // X contains multiple T
+        constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
+
+        constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
+
+        static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
+                      "wrong! X should contain multiple T");
+
+        static_assert(GetAddressSpace() == AddressSpaceEnum::Global, "only support global mem");
+
+#if CK_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER && CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT
+        bool constexpr use_amd_buffer_addressing =
+            is_same_v<remove_cvref_t<scalar_t>, int32_t> ||
+            is_same_v<remove_cvref_t<scalar_t>, float> ||
+            (is_same_v<remove_cvref_t<scalar_t>, half_t> && scalar_per_x_vector % 2 == 0);
+#elif CK_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER && (!CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT)
+        bool constexpr use_amd_buffer_addressing = is_same_v<remove_cvref_t<scalar_t>, int32_t>;
+#elif(!CK_USE_AMD_BUFFER_ATOMIC_ADD_INTEGER) && CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT
+        bool constexpr use_amd_buffer_addressing =
+            is_same_v<remove_cvref_t<scalar_t>, float> ||
+            (is_same_v<remove_cvref_t<scalar_t>, half_t> && scalar_per_x_vector % 2 == 0);
+#else
+        bool constexpr use_amd_buffer_addressing = false;
+#endif
+
+        if constexpr(use_amd_buffer_addressing)
+        {
+            constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
+
+            amd_buffer_atomic_add<remove_cvref_t<T>, t_per_x, Coherence>(
+                x, p_data_, i, is_valid_element, buffer_size_);
+        }
+        else
+        {
+            if(is_valid_element)
+            {
+                atomic_add<X>(c_style_pointer_cast<X*>(&p_data_[i]), x);
+            }
+        }
+    }
+
+    template <typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __device__ void AtomicMax(index_t i, bool is_valid_element, const X& x)
+    {
+        // X contains multiple T
+        constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
+
+        constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
+
+        static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
+                      "wrong! X should contain multiple T");
+
+        static_assert(GetAddressSpace() == AddressSpaceEnum::Global, "only support global mem");
+
+#if CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64
+        using scalar_t                           = typename scalar_type<remove_cvref_t<T>>::type;
+        bool constexpr use_amd_buffer_addressing = is_same_v<remove_cvref_t<scalar_t>, double>;
+#else
+        bool constexpr use_amd_buffer_addressing = false;
+#endif
+
+        if constexpr(use_amd_buffer_addressing)
+        {
+            constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector;
+
+            amd_buffer_atomic_max<remove_cvref_t<T>, t_per_x>(
+                x, p_data_, i, is_valid_element, buffer_size_);
+        }
+        else if(is_valid_element)
+        {
+            atomic_max<X>(c_style_pointer_cast<X*>(&p_data_[i]), x);
+        }
+    }
+
+    // FIXME: remove
+    __device__ static constexpr bool IsStaticBuffer() { return false; }
+
+    // FIXME: remove
+    __device__ static constexpr bool IsDynamicBuffer() { return true; }
+
+    __host__ __device__ void Print() const
+    {
+        printf("BufferView{");
+
+        // AddressSpace
+        printf("AddressSpace: Global, ");
+
+        // p_data_
+        printf("p_data_: %p, ", static_cast<void*>(const_cast<remove_cvref_t<T>*>(p_data_)));
+
+        // buffer_size_
+        printf("buffer_size_: ");
+        print(buffer_size_);
+        printf(", ");
+
+        // invalid_element_value_
+        printf("invalid_element_value_: ");
+        print(invalid_element_value_);
+
+        printf("}");
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/utility/buffer_view_impl_lds.hpp b/include/ck/utility/buffer_view_impl_lds.hpp
new file mode 100644
index 0000000000..d8923d9daa
--- /dev/null
+++ b/include/ck/utility/buffer_view_impl_lds.hpp
@@ -0,0 +1,276 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/buffer_view.hpp"
+
+namespace ck {
+
+// Address Space: LDS
+// T may be scalar or vector
+// X may be scalar or vector
+// T and X have same scalar type
+// X contains multiple T
+// FIXME: InvalidElementUseNumericalZeroValue and invalid_element_value_ should be a property of
+//        transforms of TensorView/Tensor
+template <typename T, typename BufferSizeType, bool InvalidElementUseNumericalZeroValue>
+struct BufferView<AddressSpaceEnum::Lds,
+                  T,
+                  BufferSizeType,
+                  InvalidElementUseNumericalZeroValue,
+                  AmdBufferCoherenceEnum::DefaultCoherence>
+{
+    using type = T;
+
+    T* p_data_ = nullptr;
+    BufferSizeType buffer_size_;
+    remove_cvref_t<T> invalid_element_value_ = T{0};
+
+    __host__ __device__ constexpr BufferView() : p_data_{}, buffer_size_{}, invalid_element_value_{}
+    {
+    }
+
+    __host__ __device__ constexpr BufferView(T* p_data, BufferSizeType buffer_size)
+        : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{0}
+    {
+    }
+
+    __host__ __device__ constexpr BufferView(T* p_data,
+                                             BufferSizeType buffer_size,
+                                             T invalid_element_value)
+        : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{invalid_element_value}
+    {
+    }
+
+    __device__ static constexpr AddressSpaceEnum GetAddressSpace() { return AddressSpaceEnum::Lds; }
+
+    // i is offset of T
+    // FIXME: doesn't do is_valid check
+    __device__ constexpr const T& operator[](index_t i) const { return p_data_[i]; }
+
+    // i is offset of T
+    // FIXME: doesn't do is_valid check
+    __device__ constexpr T& operator()(index_t i) { return p_data_[i]; }
+
+    // i is offset of T, not X. i should be aligned to X
+    template <typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __device__ constexpr auto Get(index_t i, bool is_valid_element) const
+    {
+        // X contains multiple T
+        constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
+
+        constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
+
+        static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
+                      "wrong! X should contain multiple T");
+
+        if(is_valid_element)
+        {
+#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
+            X tmp;
+
+            __builtin_memcpy(&tmp, &(p_data_[i]), sizeof(X));
+
+            return tmp;
+#else
+            return *c_style_pointer_cast<const X*>(&p_data_[i]);
+#endif
+        }
+        else
+        {
+            if constexpr(InvalidElementUseNumericalZeroValue)
+            {
+                return X{0};
+            }
+            else
+            {
+                return X{invalid_element_value_};
+            }
+        }
+    }
+
+    // i is offset of T, not X. i should be aligned to X
+    template <InMemoryDataOperationEnum Op,
+              typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __device__ void Update(index_t i, bool is_valid_element, const X& x)
+    {
+        if constexpr(Op == InMemoryDataOperationEnum::Set)
+        {
+            this->template Set<X>(i, is_valid_element, x);
+        }
+        // FIXME: remove InMemoryDataOperationEnum::Add
+        else if constexpr(Op == InMemoryDataOperationEnum::Add)
+        {
+            auto tmp = this->template Get<X>(i, is_valid_element);
+            this->template Set<X>(i, is_valid_element, x + tmp);
+        }
+    }
+
+    // i is offset of T, not X. i should be aligned to X
+    template <typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __device__ void Set(index_t i, bool is_valid_element, const X& x)
+    {
+        // X contains multiple T
+        constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
+
+        constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
+
+        static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
+                      "wrong! X should contain multiple T");
+
+#if CK_WORKAROUND_SWDEV_XXXXXX_INT8_DS_WRITE_ISSUE
+        bool constexpr workaround_int8_ds_write_issue = true;
+#else
+        bool constexpr workaround_int8_ds_write_issue = false;
+#endif
+
+        if constexpr(is_same<typename scalar_type<remove_cvref_t<T>>::type, int8_t>::value &&
+                     workaround_int8_ds_write_issue)
+        {
+            if(is_valid_element)
+            {
+                // HACK: compiler would lower IR "store<i8, 16> address_space(3)" into inefficient
+                // ISA, so I try to let compiler emit IR "store<i32, 4>" which would be lower to
+                // ds_write_b128
+                // TODO: remove this after compiler fix
+                static_assert((is_same<remove_cvref_t<T>, int8_t>::value &&
+                               is_same<remove_cvref_t<X>, int8_t>::value) ||
+                                  (is_same<remove_cvref_t<T>, int8_t>::value &&
+                                   is_same<remove_cvref_t<X>, int8x2_t>::value) ||
+                                  (is_same<remove_cvref_t<T>, int8_t>::value &&
+                                   is_same<remove_cvref_t<X>, int8x4_t>::value) ||
+                                  (is_same<remove_cvref_t<T>, int8_t>::value &&
+                                   is_same<remove_cvref_t<X>, int8x8_t>::value) ||
+                                  (is_same<remove_cvref_t<T>, int8_t>::value &&
+                                   is_same<remove_cvref_t<X>, int8x16_t>::value) ||
+                                  (is_same<remove_cvref_t<T>, int8x4_t>::value &&
+                                   is_same<remove_cvref_t<X>, int8x4_t>::value) ||
+                                  (is_same<remove_cvref_t<T>, int8x8_t>::value &&
+                                   is_same<remove_cvref_t<X>, int8x8_t>::value) ||
+                                  (is_same<remove_cvref_t<T>, int8x16_t>::value &&
+                                   is_same<remove_cvref_t<X>, int8x16_t>::value),
+                              "wrong! not implemented for this combination, please add "
+                              "implementation");
+
+                if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
+                             is_same<remove_cvref_t<X>, int8_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int8_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int8_t*>(&x);
+                }
+                else if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
+                                  is_same<remove_cvref_t<X>, int8x2_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int16_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int16_t*>(&x);
+                }
+                else if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
+                                  is_same<remove_cvref_t<X>, int8x4_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int32_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int32_t*>(&x);
+                }
+                else if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
+                                  is_same<remove_cvref_t<X>, int8x8_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int32x2_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int32x2_t*>(&x);
+                }
+                else if constexpr(is_same<remove_cvref_t<T>, int8_t>::value &&
+                                  is_same<remove_cvref_t<X>, int8x16_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int32x4_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int32x4_t*>(&x);
+                }
+                else if constexpr(is_same<remove_cvref_t<T>, int8x4_t>::value &&
+                                  is_same<remove_cvref_t<X>, int8x4_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int32_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int32_t*>(&x);
+                }
+                else if constexpr(is_same<remove_cvref_t<T>, int8x8_t>::value &&
+                                  is_same<remove_cvref_t<X>, int8x8_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int32x2_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int32x2_t*>(&x);
+                }
+                else if constexpr(is_same<remove_cvref_t<T>, int8x16_t>::value &&
+                                  is_same<remove_cvref_t<X>, int8x16_t>::value)
+                {
+                    // HACK: cast pointer of x is bad
+                    // TODO: remove this after compiler fix
+                    *c_style_pointer_cast<int32x4_t*>(&p_data_[i]) =
+                        *c_style_pointer_cast<const int32x4_t*>(&x);
+                }
+            }
+        }
+        else
+        {
+            if(is_valid_element)
+            {
+#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
+                X tmp = x;
+
+                __builtin_memcpy(&(p_data_[i]), &tmp, sizeof(X));
+#else
+                *c_style_pointer_cast<X*>(&p_data_[i]) = x;
+#endif
+            }
+        }
+    }
+
+    // FIXME: remove
+    __device__ static constexpr bool IsStaticBuffer() { return false; }
+
+    // FIXME: remove
+    __device__ static constexpr bool IsDynamicBuffer() { return true; }
+
+    __host__ __device__ void Print() const
+    {
+        printf("BufferView{");
+
+        // AddressSpace
+        printf("AddressSpace: Lds, ");
+
+        // p_data_
+        printf("p_data_: %p, ", static_cast<void*>(const_cast<remove_cvref_t<T>*>(p_data_)));
+
+        // buffer_size_
+        printf("buffer_size_: ");
+        print(buffer_size_);
+        printf(", ");
+
+        // invalid_element_value_
+        printf("invalid_element_value_: ");
+        print(invalid_element_value_);
+
+        printf("}");
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/utility/buffer_view_impl_vgpr.hpp b/include/ck/utility/buffer_view_impl_vgpr.hpp
new file mode 100644
index 0000000000..4c3e948849
--- /dev/null
+++ b/include/ck/utility/buffer_view_impl_vgpr.hpp
@@ -0,0 +1,177 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/buffer_view_declare.hpp"
+#include "ck/utility/generic_memory_space_atomic.hpp"
+
+namespace ck {
+
+// Address Space: Vgpr
+// T may be scalar or vector
+// X may be scalar or vector
+// T and X have same scalar type
+// X contains multiple T
+// FIXME: InvalidElementUseNumericalZeroValue and invalid_element_value_ should be a property of
+//        transforms of TensorView/Tensor
+template <typename T, typename BufferSizeType, bool InvalidElementUseNumericalZeroValue>
+struct BufferView<AddressSpaceEnum::Vgpr,
+                  T,
+                  BufferSizeType,
+                  InvalidElementUseNumericalZeroValue,
+                  AmdBufferCoherenceEnum::DefaultCoherence>
+{
+    using type = T;
+
+    T* p_data_ = nullptr;
+    BufferSizeType buffer_size_;
+    remove_cvref_t<T> invalid_element_value_ = T{0};
+
+    __host__ __device__ constexpr BufferView() : p_data_{}, buffer_size_{}, invalid_element_value_{}
+    {
+    }
+
+    __host__ __device__ constexpr BufferView(T* p_data, BufferSizeType buffer_size)
+        : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{0}
+    {
+    }
+
+    __host__ __device__ constexpr BufferView(T* p_data,
+                                             BufferSizeType buffer_size,
+                                             T invalid_element_value)
+        : p_data_{p_data}, buffer_size_{buffer_size}, invalid_element_value_{invalid_element_value}
+    {
+    }
+
+    __device__ static constexpr AddressSpaceEnum GetAddressSpace()
+    {
+        return AddressSpaceEnum::Vgpr;
+    }
+
+    // i is offset of T
+    // FIXME: doesn't do is_valid check
+    __device__ constexpr const T& operator[](index_t i) const { return p_data_[i]; }
+
+    // i is offset of T
+    // FIXME: doesn't do is_valid check
+    __device__ constexpr T& operator()(index_t i) { return p_data_[i]; }
+
+    // i is offset of T, not X. i should be aligned to X
+    template <typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __device__ constexpr auto Get(index_t i, bool is_valid_element) const
+    {
+        // X contains multiple T
+        constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
+
+        constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
+
+        static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
+                      "wrong! X should contain multiple T");
+
+        if(is_valid_element)
+        {
+#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
+            X tmp;
+
+            __builtin_memcpy(&tmp, &(p_data_[i]), sizeof(X));
+
+            return tmp;
+#else
+            return *c_style_pointer_cast<const X*>(&p_data_[i]);
+#endif
+        }
+        else
+        {
+            if constexpr(InvalidElementUseNumericalZeroValue)
+            {
+                return X{0};
+            }
+            else
+            {
+                return X{invalid_element_value_};
+            }
+        }
+    }
+
+    // i is offset of T, not X. i should be aligned to X
+    template <InMemoryDataOperationEnum Op,
+              typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __device__ void Update(index_t i, bool is_valid_element, const X& x)
+    {
+        if constexpr(Op == InMemoryDataOperationEnum::Set)
+        {
+            this->template Set<X>(i, is_valid_element, x);
+        }
+        // FIXME: remove InMemoryDataOperationEnum::Add
+        else if constexpr(Op == InMemoryDataOperationEnum::Add)
+        {
+            auto tmp = this->template Get<X>(i, is_valid_element);
+            this->template Set<X>(i, is_valid_element, x + tmp);
+        }
+    }
+
+    // i is offset of T, not X. i should be aligned to X
+    template <typename X,
+              typename enable_if<is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                         typename scalar_type<remove_cvref_t<T>>::type>::value,
+                                 bool>::type = false>
+    __device__ void Set(index_t i, bool is_valid_element, const X& x)
+    {
+        // X contains multiple T
+        constexpr index_t scalar_per_t_vector = scalar_type<remove_cvref_t<T>>::vector_size;
+
+        constexpr index_t scalar_per_x_vector = scalar_type<remove_cvref_t<X>>::vector_size;
+
+        static_assert(scalar_per_x_vector % scalar_per_t_vector == 0,
+                      "wrong! X should contain multiple T");
+
+        if(is_valid_element)
+        {
+#if CK_EXPERIMENTAL_USE_MEMCPY_FOR_VECTOR_ACCESS
+            X tmp = x;
+
+            __builtin_memcpy(&(p_data_[i]), &tmp, sizeof(X));
+#else
+            *c_style_pointer_cast<X*>(&p_data_[i]) = x;
+#endif
+        }
+    }
+
+    // FIXME: remove
+    __device__ static constexpr bool IsStaticBuffer() { return false; }
+
+    // FIXME: remove
+    __device__ static constexpr bool IsDynamicBuffer() { return true; }
+
+    __host__ __device__ void Print() const
+    {
+        printf("BufferView{");
+
+        // AddressSpace
+        printf("AddressSpace: Vgpr, ");
+
+        // p_data_
+        printf("p_data_: %p, ", static_cast<void*>(const_cast<remove_cvref_t<T>*>(p_data_)));
+
+        // buffer_size_
+        printf("buffer_size_: ");
+        print(buffer_size_);
+        printf(", ");
+
+        // invalid_element_value_
+        printf("invalid_element_value_: ");
+        print(invalid_element_value_);
+
+        printf("}");
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/utility/c_style_pointer_cast.hpp b/include/ck/utility/c_style_pointer_cast.hpp
new file mode 100644
index 0000000000..610e393a77
--- /dev/null
+++ b/include/ck/utility/c_style_pointer_cast.hpp
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_C_STYLE_POINTER_CAST_HPP
+#define CK_C_STYLE_POINTER_CAST_HPP
+
+#include "type.hpp"
+#include "enable_if.hpp"
+
+namespace ck {
+
+template <typename PY,
+          typename PX,
+          typename enable_if<is_pointer_v<PY> && is_pointer_v<PX>, bool>::type = false>
+__host__ __device__ PY c_style_pointer_cast(PX p_x)
+{
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wold-style-cast"
+#pragma clang diagnostic ignored "-Wcast-align"
+    return (PY)p_x; // NOLINT(old-style-cast, cast-align)
+#pragma clang diagnostic pop
+}
+
+} // namespace ck
+#endif
diff --git a/include/ck/utility/common_header.hpp b/include/ck/utility/common_header.hpp
new file mode 100644
index 0000000000..58ebac9d80
--- /dev/null
+++ b/include/ck/utility/common_header.hpp
@@ -0,0 +1,64 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/static_assert.hpp"
+#include "ck/utility/remove_cvref.hpp"
+#include "ck/utility/is_static.hpp"
+#include "ck/utility/bit_cast.hpp"
+#include "ck/utility/print.hpp"
+#include "ck/utility/array.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/sequence_helper.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/utility/tuple_helper.hpp"
+#include "ck/utility/map.hpp"
+#include "ck/utility/container_helper.hpp"
+#include "ck/utility/statically_indexed_array.hpp"
+#include "ck/utility/multi_index.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/functional.hpp"
+#include "ck/utility/functional2.hpp"
+#include "ck/utility/functional3.hpp"
+#include "ck/utility/functional4.hpp"
+#include "ck/utility/enable_if.hpp"
+#include "ck/utility/ignore.hpp"
+#include "ck/utility/integral_constant.hpp"
+#include "ck/utility/math.hpp"
+#include "ck/utility/math_v2.hpp"
+#include "ck/utility/math_ext.hpp"
+#include "ck/utility/number.hpp"
+#include "ck/utility/tuple_of_sequence_to_array_of_array.hpp"
+#include "ck/utility/macro_func_array_to_sequence.hpp"
+#include "ck/utility/macro_func_array_of_array_to_tuple_of_sequence.hpp"
+#include "ck/utility/type.hpp"
+#include "ck/utility/type_convert.hpp"
+#include "ck/utility/magic_division.hpp"
+#include "ck/utility/c_style_pointer_cast.hpp"
+#include "ck/utility/transpose_vectors.hpp"
+#include "ck/utility/inner_product.hpp"
+#include "ck/utility/thread_group.hpp"
+#include "ck/utility/meta_data_buffer.hpp"
+#include "ck/utility/debug.hpp"
+
+#include "ck/utility/amd_buffer_addressing.hpp"
+#include "ck/utility/amd_wave_read_first_lane.hpp"
+#include "ck/utility/amd_warp_shuffle.hpp"
+#include "ck/utility/generic_memory_space_atomic.hpp"
+#include "ck/utility/get_id.hpp"
+#include "ck/utility/thread_group.hpp"
+#include "ck/utility/synchronization.hpp"
+#include "ck/utility/amd_address_space.hpp"
+#include "ck/utility/static_buffer.hpp"
+#include "ck/utility/dynamic_buffer.hpp"
+
+// TODO: remove this
+#if CK_USE_AMD_INLINE_ASM
+#include "ck/utility/amd_inline_asm.hpp"
+#endif
+
+#ifdef CK_USE_AMD_MFMA
+#include "ck/utility/amd_xdlops.hpp"
+#endif
diff --git a/include/ck/utility/container_element_picker.hpp b/include/ck/utility/container_element_picker.hpp
new file mode 100644
index 0000000000..838147e420
--- /dev/null
+++ b/include/ck/utility/container_element_picker.hpp
@@ -0,0 +1,158 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef CK_CONTAINER_ELEMENT_PICKER_HPP
+#define CK_CONTAINER_ELEMENT_PICKER_HPP
+
+#include "functional2.hpp"
+#include "sequence.hpp"
+
+namespace ck {
+
+// Arr: Array or StaticallyIndexedArray
+// Picks: Sequence<...>
+template <typename Arr, typename Picks>
+struct ContainerElementPicker
+{
+    using type = ContainerElementPicker;
+#if 0
+    using data_type = typename Arr::data_type;
+#endif
+
+    __host__ __device__ constexpr ContainerElementPicker() = delete;
+
+    __host__ __device__ constexpr ContainerElementPicker(Arr& array) : mArray{array}
+    {
+        constexpr index_t imax =
+            reduce_on_sequence(Picks{}, math::maximize<index_t>{}, Number<0>{});
+
+        static_assert(imax < Arr::Size(), "wrong! exceeding # array element");
+    }
+
+    __host__ __device__ static constexpr auto Size() { return Picks::Size(); }
+
+    template <index_t I>
+    __host__ __device__ constexpr const auto& At(Number<I> i) const
+    {
+        static_assert(I < Size(), "wrong!");
+
+        constexpr auto IP = Picks{}[i];
+        return mArray[IP];
+    }
+
+    template <index_t I>
+    __host__ __device__ constexpr auto& At(Number<I> i)
+    {
+        static_assert(I < Size(), "wrong!");
+
+        constexpr auto IP = Picks{}[i];
+        return mArray(IP);
+    }
+
+    template <index_t I>
+    __host__ __device__ constexpr const auto& operator[](Number<I> i) const
+    {
+        return At(i);
+    }
+
+    template <index_t I>
+    __host__ __device__ constexpr auto& operator()(Number<I> i)
+    {
+        return At(i);
+    }
+
+    template <typename T>
+    __host__ __device__ constexpr auto operator=(const T& a)
+    {
+        static_assert(T::Size() == Size(), "wrong! size not the same");
+
+        static_for<0, Size(), 1>{}([&](auto i) { operator()(i) = a[i]; });
+
+        return *this;
+    }
+
+    private:
+    Arr& mArray;
+};
+
+// Arr: Array or StaticallyIndexedArray
+// Picks: Sequence<...>
+template <typename Arr, typename Picks>
+struct ConstantContainerElementPicker
+{
+    using type = ConstantContainerElementPicker;
+#if 0
+    using data_type = typename Arr::data_type;
+#endif
+
+    __host__ __device__ constexpr ConstantContainerElementPicker() = delete;
+
+    __host__ __device__ constexpr ConstantContainerElementPicker(const Arr& array) : mArray{array}
+    {
+        constexpr index_t imax =
+            reduce_on_sequence(Picks{}, math::maximize<index_t>{}, Number<0>{});
+
+        static_assert(imax < Arr::Size(), "wrong! exceeding # array element");
+    }
+
+    __host__ __device__ static constexpr auto Size() { return Picks::Size(); }
+
+    template <index_t I>
+    __host__ __device__ constexpr const auto& At(Number<I> i) const
+    {
+        static_assert(I < Size(), "wrong!");
+
+        constexpr auto IP = Picks{}[i];
+        return mArray[IP];
+    }
+
+    template <index_t I>
+    __host__ __device__ constexpr const auto& operator[](Number<I> i) const
+    {
+        return At(i);
+    }
+
+    private:
+    const Arr& mArray;
+};
+
+template <typename Arr, typename Picks, typename X>
+__host__ __device__ constexpr auto operator+=(ContainerElementPicker<Arr, Picks>& y, const X& x)
+{
+    using Y                 = ContainerElementPicker<Arr, Picks>;
+    constexpr index_t nsize = Y::Size();
+
+    static_assert(nsize == X::Size(), "wrong! size not the same");
+
+    static_for<0, nsize, 1>{}([&](auto i) { y(i) += x[i]; });
+
+    return y;
+}
+
+template <typename Arr, typename Picks, typename X>
+__host__ __device__ constexpr auto operator-=(ContainerElementPicker<Arr, Picks>& y, const X& x)
+{
+    using Y                 = ContainerElementPicker<Arr, Picks>;
+    constexpr index_t nsize = Y::Size();
+
+    static_assert(nsize == X::Size(), "wrong! size not the same");
+
+    static_for<0, nsize, 1>{}([&](auto i) { y(i) -= x[i]; });
+
+    return y;
+}
+
+template <typename Arr, typename Picks>
+__host__ __device__ constexpr auto pick_container_element(Arr& a, Picks)
+{
+    return ContainerElementPicker<Arr, Picks>(a);
+}
+
+template <typename Arr, typename Picks>
+__host__ __device__ constexpr auto pick_container_element(const Arr& a, Picks)
+{
+    return ConstantContainerElementPicker<Arr, Picks>(a);
+}
+
+} // namespace ck
+#endif
diff --git a/include/ck/utility/container_helper.hpp b/include/ck/utility/container_helper.hpp
new file mode 100644
index 0000000000..e92bc40e59
--- /dev/null
+++ b/include/ck/utility/container_helper.hpp
@@ -0,0 +1,477 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/sequence_helper.hpp"
+#include "ck/utility/array.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/utility/tuple_helper.hpp"
+#include "ck/utility/statically_indexed_array.hpp"
+#include "ck/utility/map.hpp"
+
+namespace ck {
+
+template <typename TData, index_t NSize>
+__host__ __device__ constexpr auto container_push_back(const Array<TData, NSize>& a, const TData& x)
+{
+    Array<TData, NSize + 1> r;
+
+    static_for<0, NSize, 1>{}([&r, &a ](auto i) constexpr { r(i) = a[i]; });
+
+    r(Number<NSize>{}) = x;
+
+    return r;
+}
+
+template <typename... Ts, typename T>
+__host__ __device__ constexpr auto container_push_front(const Tuple<Ts...>& a, const T& x)
+{
+    return container_concat(make_tuple(x), a);
+}
+
+template <typename... Ts, typename T>
+__host__ __device__ constexpr auto container_push_back(const Tuple<Ts...>& a, const T& x)
+{
+    return container_concat(a, make_tuple(x));
+}
+
+// reorder Array
+template <typename TData, index_t NSize, index_t... IRs>
+__host__ __device__ constexpr auto
+container_reorder_given_new2old(const Array<TData, NSize>& old_array, Sequence<IRs...> /*new2old*/)
+{
+    static_assert(NSize == sizeof...(IRs), "wrong! size not consistent");
+
+    static_assert(is_valid_sequence_map<Sequence<IRs...>>{}, "wrong! invalid reorder map");
+
+    return make_array<remove_cvref_t<TData>>(old_array[IRs]...);
+}
+
+template <typename TData, index_t NSize, index_t... IRs>
+__host__ __device__ constexpr auto
+container_reorder_given_old2new(const Array<TData, NSize>& old_array, Sequence<IRs...> old2new)
+{
+    return container_reorder_given_new2old(
+        old_array, typename sequence_map_inverse<decltype(old2new)>::type{});
+}
+
+// reorder Array
+template <typename TData, index_t NSize>
+__host__ __device__ constexpr auto
+container_reorder_given_new2old(const Array<TData, NSize>& old_array,
+                                const Map<index_t, index_t>& new2old)
+{
+    Array<TData, NSize> new_array;
+
+    for(const auto& [new_pos, old_pos] : new2old)
+    {
+        new_array(new_pos) = old_array[old_pos];
+    }
+
+    return new_array;
+}
+
+template <typename TData, index_t NSize>
+__host__ __device__ constexpr auto
+container_reorder_given_old2new(const Array<TData, NSize>& old_array,
+                                const Map<index_t, index_t>& old2new)
+{
+    Array<TData, NSize> new_array;
+
+    for(const auto& [old_pos, new_pos] : old2new)
+    {
+        new_array(new_pos) = old_array[old_pos];
+    }
+
+    return new_array;
+}
+
+// reorder Tuple
+template <typename... Ts, index_t... IRs>
+__host__ __device__ constexpr auto container_reorder_given_new2old(const Tuple<Ts...>& old_tuple,
+                                                                   Sequence<IRs...> /*new2old*/)
+{
+    static_assert(sizeof...(Ts) == sizeof...(IRs), "wrong! size not consistent");
+
+    static_assert(is_valid_sequence_map<Sequence<IRs...>>{}, "wrong! invalid reorder map");
+
+    return make_tuple(old_tuple[Number<IRs>{}]...);
+}
+
+template <typename... Ts, index_t... IRs>
+__host__ __device__ constexpr auto container_reorder_given_old2new(const Tuple<Ts...>& old_tuple,
+                                                                   Sequence<IRs...> old2new)
+{
+    return container_reorder_given_new2old(
+        old_tuple, typename sequence_map_inverse<decltype(old2new)>::type{});
+}
+
+// reorder Sequence
+template <index_t... Is, index_t... IRs>
+__host__ __device__ constexpr auto container_reorder_given_new2old(Sequence<Is...> /* old_seq */,
+                                                                   Sequence<IRs...> /*new2old*/)
+{
+    static_assert(sizeof...(Is) == sizeof...(IRs), "wrong! size not consistent");
+
+    static_assert(is_valid_sequence_map<Sequence<IRs...>>{}, "wrong! invalid reorder map");
+
+    return Sequence<Sequence<Is...>::At(Number<IRs>{})...>{};
+}
+
+template <index_t... Is, index_t... IRs>
+__host__ __device__ constexpr auto container_reorder_given_old2new(Sequence<Is...> old_seq,
+                                                                   Sequence<IRs...> /* old2new */)
+{
+    static_assert(sizeof...(Is) == sizeof...(IRs), "wrong! size not consistent");
+
+    static_assert(is_valid_sequence_map<Sequence<IRs...>>{}, "wrong! invalid reorder map");
+
+    constexpr auto new2old = typename sequence_map_inverse<Sequence<IRs...>>::type{};
+
+    return container_reorder_given_new2old(old_seq, new2old);
+}
+
+#if !CK_WORKAROUND_SWDEV_275126
+// rocm-4.1 compiler would crash for recursive lambda
+template <typename Container,
+          typename Reduce,
+          typename Init,
+          index_t IBegin = 0,
+          index_t IEnd   = Container::Size(),
+          index_t IStep  = 1>
+__host__ __device__ constexpr auto container_reduce(const Container& x,
+                                                    Reduce reduce,
+                                                    Init init,
+                                                    Number<IBegin> = Number<0>{},
+                                                    Number<IEnd>   = Number<Container::Size()>{},
+                                                    Number<IStep>  = Number<1>{})
+{
+    static_assert((IEnd - IBegin) % IStep == 0, "wrong!");
+
+    // f is recursive function, fs is a dummy of f
+    // i is index, y_old is current scan, r_old is current reduction
+    auto f = [&](auto fs, auto i, auto r_old) {
+        auto r_new = reduce(x[i], r_old);
+
+        if constexpr(i.value < IEnd - IStep)
+        {
+            // recursively call f/fs
+            return fs(fs, i + Number<IStep>{}, r_new);
+        }
+        else
+        {
+            return r_new;
+        }
+    };
+
+    // start recursion
+    return f(f, Number<IBegin>{}, init);
+}
+#else
+// i is index, y_old is current scan, r_old is current reduction
+template <typename Container,
+          typename Reduce,
+          typename ROld,
+          index_t I,
+          index_t IEnd,
+          index_t IStep>
+__host__ __device__ constexpr auto container_reduce_impl(
+    const Container& x, Reduce reduce, ROld r_old, Number<I> i, Number<IEnd>, Number<IStep>)
+{
+    auto r_new = reduce(x[i], r_old);
+
+    if constexpr(i.value < IEnd - IStep)
+    {
+        return container_reduce_impl(
+            x, reduce, r_new, i + Number<IStep>{}, Number<IEnd>{}, Number<IStep>{});
+    }
+    else
+    {
+        return r_new;
+    }
+}
+
+// rocm-4.1 compiler would crash for recursive lambda
+// container reduce with initial value
+template <typename Container,
+          typename Reduce,
+          typename Init,
+          index_t IBegin = 0,
+          index_t IEnd   = Container::Size(),
+          index_t IStep  = 1>
+__host__ __device__ constexpr auto container_reduce(const Container& x,
+                                                    Reduce reduce,
+                                                    Init init,
+                                                    Number<IBegin> = Number<0>{},
+                                                    Number<IEnd>   = Number<Container::Size()>{},
+                                                    Number<IStep>  = Number<1>{})
+{
+    static_assert((IEnd - IBegin) % IStep == 0, "wrong!");
+
+    if constexpr(IEnd > IBegin)
+    {
+        return container_reduce_impl(
+            x, reduce, init, Number<IBegin>{}, Number<IEnd>{}, Number<IStep>{});
+    }
+    else
+    {
+        return init;
+    }
+}
+#endif
+
+template <typename TData, index_t NSize, typename Reduce>
+__host__ __device__ constexpr auto
+container_reverse_inclusive_scan(const Array<TData, NSize>& x, Reduce f, TData init)
+{
+    Array<TData, NSize> y;
+
+    TData r = init;
+
+    static_for<NSize - 1, 0, -1>{}([&](auto i) {
+        r    = f(r, x[i]);
+        y(i) = r;
+    });
+
+    r              = f(r, x[Number<0>{}]);
+    y(Number<0>{}) = r;
+
+    return y;
+}
+
+template <typename TData, index_t NSize, typename Reduce, typename Init>
+__host__ __device__ constexpr auto
+container_reverse_exclusive_scan(const Array<TData, NSize>& x, Reduce f, Init init)
+{
+#if 0
+    Array<TData, NSize> y;
+
+    TData r = init;
+
+    static_for<NSize - 1, 0, -1>{}([&](auto i) {
+        y(i) = r;
+        r    = f(r, x[i]);
+    });
+
+    y(Number<0>{}) = r;
+
+    return y;
+#else
+    Array<TData, NSize> y;
+
+    TData r = init;
+
+    for(index_t i = NSize - 1; i > 0; --i)
+    {
+        y(i) = r;
+        r    = f(r, x[i]);
+    }
+
+    y(0) = r;
+
+    return y;
+#endif
+}
+
+template <index_t... Is, typename Reduce, index_t Init>
+__host__ __device__ constexpr auto
+container_reverse_exclusive_scan(const Sequence<Is...>& seq, Reduce f, Number<Init>)
+{
+    return reverse_exclusive_scan_sequence(seq, f, Number<Init>{});
+}
+
+#if !CK_WORKAROUND_SWDEV_275126
+// rocm4.1 compiler would crash with recursive lambda
+template <typename... Xs, typename Reduce, typename Init>
+__host__ __device__ constexpr auto
+container_reverse_exclusive_scan(const Tuple<Xs...>& x, Reduce reduce, Init init)
+{
+    constexpr index_t NSize = sizeof...(Xs);
+
+    // f is recursive function, fs is a dummy of f
+    // i is index, y_old is current scan, r_old is current reduction
+    auto f = [&](auto fs, auto i, auto y_old, auto r_old) {
+        auto r_new = reduce(x[i], r_old);
+
+        auto y_new = container_push_front(y_old, r_new);
+
+        if constexpr(i.value > 1)
+        {
+            // recursively call f/fs
+            return fs(fs, i - Number<1>{}, y_new, r_new);
+        }
+        else
+        {
+            return y_new;
+        }
+    };
+
+    // start recursion
+    return f(f, Number<NSize - 1>{}, make_tuple(init), init);
+}
+#else
+// i is index, y_old is current scan, r_old is current reduction
+template <typename... Xs, typename Reduce, index_t I, typename YOld, typename ROld>
+__host__ __device__ constexpr auto container_reverse_exclusive_scan_impl(
+    const Tuple<Xs...>& x, Reduce reduce, Number<I> i, YOld y_old, ROld r_old)
+{
+    auto r_new = reduce(x[i], r_old);
+
+    auto y_new = container_push_front(y_old, r_new);
+
+    if constexpr(i.value > 1)
+    {
+        // recursively call f/fs
+        return container_reverse_exclusive_scan_impl(x, reduce, i - Number<1>{}, y_new, r_new);
+    }
+    else
+    {
+        return y_new;
+    }
+}
+
+template <typename... Xs, typename Reduce, typename Init>
+__host__ __device__ constexpr auto
+container_reverse_exclusive_scan(const Tuple<Xs...>& x, Reduce reduce, Init init)
+{
+    constexpr index_t NSize = sizeof...(Xs);
+
+    return container_reverse_exclusive_scan_impl(
+        x, reduce, Number<NSize - 1>{}, make_tuple(init), init);
+}
+#endif
+
+// TODO: update to like container_reverse_exclusive_scan to deal with Tuple of Numebr<>
+template <typename... Xs, typename Reduce, typename TData>
+__host__ __device__ constexpr auto
+container_reverse_inclusive_scan(const Tuple<Xs...>& x, Reduce f, TData init)
+{
+    constexpr index_t NSize = sizeof...(Xs);
+
+    Tuple<Xs...> y;
+
+    TData r = init;
+
+    static_for<NSize - 1, 0, -1>{}([&](auto i) {
+        r    = f(r, x[i]);
+        y(i) = r;
+    });
+
+    r              = f(r, x[Number<0>{}]);
+    y(Number<0>{}) = r;
+
+    return y;
+}
+
+template <typename X, typename... Ys>
+__host__ __device__ constexpr auto container_concat(const X& x, const Ys&... ys)
+{
+    return container_concat(x, container_concat(ys...));
+}
+
+template <typename T, index_t NX, index_t NY>
+__host__ __device__ constexpr auto container_concat(const Array<T, NX>& ax, const Array<T, NY>& ay)
+{
+    return unpack2(
+        [&](auto&&... zs) { return make_array<T>(std::forward<decltype(zs)>(zs)...); }, ax, ay);
+}
+
+template <typename... X, typename... Y>
+__host__ __device__ constexpr auto container_concat(const Tuple<X...>& tx, const Tuple<Y...>& ty)
+{
+    return unpack2(
+        [&](auto&&... zs) { return make_tuple(std::forward<decltype(zs)>(zs)...); }, tx, ty);
+}
+
+template <typename Container>
+__host__ __device__ constexpr auto container_concat(const Container& x)
+{
+    return x;
+}
+
+template <typename T, index_t N, index_t... Is>
+__host__ __device__ constexpr auto get_container_subset(const Array<T, N>& arr, Sequence<Is...>)
+{
+    STATIC_ASSERT(N >= sizeof...(Is), "wrong! size");
+
+    if constexpr(sizeof...(Is) > 0)
+    {
+        return make_array<T>(arr[Is]...);
+    }
+    else
+    {
+        return Array<T, 0>{};
+    }
+}
+
+template <typename... Ts, index_t... Is>
+__host__ __device__ constexpr auto get_container_subset(const Tuple<Ts...>& tup, Sequence<Is...>)
+{
+    STATIC_ASSERT(sizeof...(Ts) >= sizeof...(Is), "wrong! size");
+
+    if constexpr(sizeof...(Is) > 0)
+    {
+        return make_tuple(tup[Number<Is>{}]...);
+    }
+    else
+    {
+        return Tuple<>{};
+    }
+}
+
+template <typename T, index_t N, index_t... Is>
+__host__ __device__ constexpr void
+set_container_subset(Array<T, N>& y, Sequence<Is...> picks, const Array<T, sizeof...(Is)>& x)
+{
+    STATIC_ASSERT(N >= sizeof...(Is), "wrong! size");
+
+    if constexpr(sizeof...(Is) > 0)
+    {
+        for(index_t i = 0; i < picks.Size(); ++i)
+        {
+            y(picks[i]) = x[i];
+        }
+    }
+}
+
+template <typename Y, typename X, index_t... Is>
+__host__ __device__ constexpr void set_container_subset(Y& y, Sequence<Is...> picks, const X& x)
+{
+    STATIC_ASSERT(Y::Size() >= sizeof...(Is) && X::Size() == sizeof...(Is), "wrong! size");
+
+    if constexpr(sizeof...(Is) > 0)
+    {
+        static_for<0, sizeof...(Is), 1>{}([&](auto i) { y(picks[i]) = x[i]; });
+    }
+}
+
+// return the index of first occurance in the sequence.
+// return seq.Size(), if not found
+template <index_t... Is>
+constexpr index_t container_find(Sequence<Is...> seq, index_t value)
+{
+    for(auto i = 0; i < seq.Size(); i++)
+    {
+        if(seq[i] == value)
+            return i;
+    }
+
+    return seq.Size();
+}
+
+template <index_t... Is>
+__host__ __device__ constexpr auto sequence_to_tuple_of_number(Sequence<Is...>)
+{
+    using Seq = Sequence<Is...>;
+
+    return generate_tuple(
+        [&](auto i) {
+            constexpr index_t tmp = Seq::At(i);
+            return Number<tmp>{};
+        },
+        Seq::Size());
+}
+
+} // namespace ck
diff --git a/include/ck/utility/data_type.hpp b/include/ck/utility/data_type.hpp
new file mode 100644
index 0000000000..253fa74837
--- /dev/null
+++ b/include/ck/utility/data_type.hpp
@@ -0,0 +1,1028 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/bit_cast.hpp"
+#include "ck/utility/statically_indexed_array.hpp"
+
+namespace ck {
+
+using bhalf_t = ushort;
+using half_t  = _Float16;
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+using int4_t = _BitInt(4);
+#endif
+using f8_t = uint8_t;
+
+// vector_type
+template <typename T, index_t N>
+struct vector_type;
+
+// Caution: DO NOT REMOVE
+// intentionally have only declaration but no definition to cause compilation failure when trying to
+// instantiate this template. The purpose is to catch user's mistake when trying to make "vector of
+// vectors"
+template <typename T, index_t V, index_t N>
+struct vector_type<T __attribute__((ext_vector_type(V))), N>;
+
+// Caution: DO NOT REMOVE
+// intentionally have only declaration but no definition to cause compilation failure when trying to
+// instantiate this template. The purpose is to catch user's mistake when trying to make "vector of
+// vectors"
+template <typename T, index_t V, index_t N>
+struct vector_type<vector_type<T, V>, N>;
+
+// vector_type_maker
+// This is the right way to handle "vector of vectors": making a bigger vector instead
+template <typename T, index_t N>
+struct vector_type_maker
+{
+    using type = vector_type<T, N>;
+};
+
+template <typename T, index_t N0, index_t N1>
+struct vector_type_maker<T __attribute__((ext_vector_type(N1))), N0>
+{
+    using type = vector_type<T, N0 * N1>;
+};
+
+template <typename T, index_t N0, index_t N1>
+struct vector_type_maker<vector_type<T, N1>, N0>
+{
+    using type = vector_type<T, N0 * N1>;
+};
+
+template <typename T, index_t N>
+using vector_type_maker_t = typename vector_type_maker<T, N>::type;
+
+template <typename T, index_t N>
+__host__ __device__ constexpr auto make_vector_type(Number<N>)
+{
+    return typename vector_type_maker<T, N>::type{};
+}
+
+// scalar_type
+template <typename TV>
+struct scalar_type;
+
+// is_scalar_type
+template <typename TV>
+struct is_scalar_type
+{
+    static constexpr bool value = (scalar_type<remove_cvref_t<TV>>::vector_size == 1);
+};
+
+// has_same_scalar_type
+template <typename X, typename Y>
+using has_same_scalar_type = is_same<typename scalar_type<remove_cvref_t<X>>::type,
+                                     typename scalar_type<remove_cvref_t<Y>>::type>;
+
+template <typename T, index_t N>
+struct scalar_type<T __attribute__((ext_vector_type(N)))>
+{
+    using type                           = T;
+    static constexpr index_t vector_size = N;
+};
+
+template <typename T, index_t N>
+struct scalar_type<vector_type<T, N>>
+{
+    using type                           = T;
+    static constexpr index_t vector_size = N;
+};
+
+//
+template <>
+struct scalar_type<double>
+{
+    using type                           = double;
+    static constexpr index_t vector_size = 1;
+};
+
+template <>
+struct scalar_type<float>
+{
+    using type                           = float;
+    static constexpr index_t vector_size = 1;
+};
+
+template <>
+struct scalar_type<half_t>
+{
+    using type                           = half_t;
+    static constexpr index_t vector_size = 1;
+};
+
+template <>
+struct scalar_type<bhalf_t>
+{
+    using type                           = bhalf_t;
+    static constexpr index_t vector_size = 1;
+};
+
+template <>
+struct scalar_type<int32_t>
+{
+    using type                           = int32_t;
+    static constexpr index_t vector_size = 1;
+};
+
+template <>
+struct scalar_type<int8_t>
+{
+    using type                           = int8_t;
+    static constexpr index_t vector_size = 1;
+};
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+template <>
+struct scalar_type<int4_t>
+{
+    using type                           = int4_t;
+    static constexpr index_t vector_size = 1;
+};
+#endif
+
+template <>
+struct scalar_type<f8_t>
+{
+    using type                           = f8_t;
+    static constexpr index_t vector_size = 1;
+};
+
+//
+template <typename T>
+struct vector_type<T, 1>
+{
+    using d1_t = T;
+    using type = d1_t;
+
+    union
+    {
+        T d1_;
+        StaticallyIndexedArray<T, 1> d1x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value, "wrong!");
+
+        return data_.d1x1_;
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value, "wrong!");
+
+        return data_.d1x1_;
+    }
+};
+
+template <typename T>
+struct vector_type<T, 2>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+
+    using type = d2_t;
+
+    union
+    {
+        d2_t d2_;
+        StaticallyIndexedArray<d1_t, 2> d1x2_;
+        StaticallyIndexedArray<d2_t, 1> d2x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value, "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x2_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value, "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x2_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x1_;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 4>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+
+    using type = d4_t;
+
+    union
+    {
+        d4_t d4_;
+        StaticallyIndexedArray<d1_t, 4> d1x4_;
+        StaticallyIndexedArray<d2_t, 2> d2x2_;
+        StaticallyIndexedArray<d4_t, 1> d4x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x4_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x2_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x4_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x2_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x1_;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 8>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+    typedef T d8_t __attribute__((ext_vector_type(8)));
+
+    using type = d8_t;
+
+    union
+    {
+        d8_t d8_;
+        StaticallyIndexedArray<d1_t, 8> d1x8_;
+        StaticallyIndexedArray<d2_t, 4> d2x4_;
+        StaticallyIndexedArray<d4_t, 2> d4x2_;
+        StaticallyIndexedArray<d8_t, 1> d8x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x8_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x4_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x2_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x8_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x4_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x2_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x1_;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 16>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+    typedef T d8_t __attribute__((ext_vector_type(8)));
+    typedef T d16_t __attribute__((ext_vector_type(16)));
+
+    using type = d16_t;
+
+    union
+    {
+        d16_t d16_;
+        StaticallyIndexedArray<d1_t, 16> d1x16_;
+        StaticallyIndexedArray<d2_t, 8> d2x8_;
+        StaticallyIndexedArray<d4_t, 4> d4x4_;
+        StaticallyIndexedArray<d8_t, 2> d8x2_;
+        StaticallyIndexedArray<d16_t, 1> d16x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x16_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x8_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x4_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x2_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x16_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x8_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x4_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x2_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x1_;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 32>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+    typedef T d8_t __attribute__((ext_vector_type(8)));
+    typedef T d16_t __attribute__((ext_vector_type(16)));
+    typedef T d32_t __attribute__((ext_vector_type(32)));
+
+    using type = d32_t;
+
+    union
+    {
+        d32_t d32_;
+        StaticallyIndexedArray<d1_t, 32> d1x32_;
+        StaticallyIndexedArray<d2_t, 16> d2x16_;
+        StaticallyIndexedArray<d4_t, 8> d4x8_;
+        StaticallyIndexedArray<d8_t, 4> d8x4_;
+        StaticallyIndexedArray<d16_t, 2> d16x2_;
+        StaticallyIndexedArray<d32_t, 1> d32x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x32_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x16_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x8_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x4_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x2_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x32_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x16_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x8_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x4_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x2_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x1_;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 64>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+    typedef T d8_t __attribute__((ext_vector_type(8)));
+    typedef T d16_t __attribute__((ext_vector_type(16)));
+    typedef T d32_t __attribute__((ext_vector_type(32)));
+    typedef T d64_t __attribute__((ext_vector_type(64)));
+
+    using type = d64_t;
+
+    union
+    {
+        d64_t d64_;
+        StaticallyIndexedArray<d1_t, 64> d1x64_;
+        StaticallyIndexedArray<d2_t, 32> d2x32_;
+        StaticallyIndexedArray<d4_t, 16> d4x16_;
+        StaticallyIndexedArray<d8_t, 8> d8x8_;
+        StaticallyIndexedArray<d16_t, 4> d16x4_;
+        StaticallyIndexedArray<d32_t, 2> d32x2_;
+        StaticallyIndexedArray<d64_t, 1> d64x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
+                          is_same<X, d64_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x64_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x32_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x16_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x8_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x4_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x2_;
+        }
+        else if constexpr(is_same<X, d64_t>::value)
+        {
+            return data_.d64x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
+                          is_same<X, d64_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x64_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x32_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x16_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x8_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x4_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x2_;
+        }
+        else if constexpr(is_same<X, d64_t>::value)
+        {
+            return data_.d64x1_;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 128>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+    typedef T d8_t __attribute__((ext_vector_type(8)));
+    typedef T d16_t __attribute__((ext_vector_type(16)));
+    typedef T d32_t __attribute__((ext_vector_type(32)));
+    typedef T d64_t __attribute__((ext_vector_type(64)));
+    typedef T d128_t __attribute__((ext_vector_type(128)));
+
+    using type = d128_t;
+
+    union
+    {
+        d128_t d128_;
+        StaticallyIndexedArray<d1_t, 128> d1x128_;
+        StaticallyIndexedArray<d2_t, 64> d2x64_;
+        StaticallyIndexedArray<d4_t, 32> d4x32_;
+        StaticallyIndexedArray<d8_t, 16> d8x16_;
+        StaticallyIndexedArray<d16_t, 8> d16x8_;
+        StaticallyIndexedArray<d32_t, 4> d32x4_;
+        StaticallyIndexedArray<d64_t, 2> d64x2_;
+        StaticallyIndexedArray<d128_t, 1> d128x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
+                          is_same<X, d64_t>::value || is_same<X, d128_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x128_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x64_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x32_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x16_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x8_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x4_;
+        }
+        else if constexpr(is_same<X, d64_t>::value)
+        {
+            return data_.d64x2_;
+        }
+        else if constexpr(is_same<X, d128_t>::value)
+        {
+            return data_.d128x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
+                          is_same<X, d64_t>::value || is_same<X, d128_t>::value,
+                      "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x128_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x64_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x32_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x16_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x8_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x4_;
+        }
+        else if constexpr(is_same<X, d64_t>::value)
+        {
+            return data_.d64x2_;
+        }
+        else if constexpr(is_same<X, d128_t>::value)
+        {
+            return data_.d128x1_;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 256>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+    typedef T d8_t __attribute__((ext_vector_type(8)));
+    typedef T d16_t __attribute__((ext_vector_type(16)));
+    typedef T d32_t __attribute__((ext_vector_type(32)));
+    typedef T d64_t __attribute__((ext_vector_type(64)));
+    typedef T d128_t __attribute__((ext_vector_type(128)));
+    typedef T d256_t __attribute__((ext_vector_type(256)));
+
+    using type = d256_t;
+
+    union
+    {
+        d256_t d256_;
+        StaticallyIndexedArray<d1_t, 256> d1x256_;
+        StaticallyIndexedArray<d2_t, 128> d2x128_;
+        StaticallyIndexedArray<d4_t, 64> d4x64_;
+        StaticallyIndexedArray<d8_t, 32> d8x32_;
+        StaticallyIndexedArray<d16_t, 16> d16x16_;
+        StaticallyIndexedArray<d32_t, 8> d32x8_;
+        StaticallyIndexedArray<d64_t, 4> d64x4_;
+        StaticallyIndexedArray<d128_t, 2> d128x2_;
+        StaticallyIndexedArray<d256_t, 1> d256x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(
+            is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value ||
+                is_same<X, d8_t>::value || is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
+                is_same<X, d64_t>::value || is_same<X, d128_t>::value || is_same<X, d256_t>::value,
+            "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x256_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x128_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x64_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x32_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x16_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x8_;
+        }
+        else if constexpr(is_same<X, d64_t>::value)
+        {
+            return data_.d64x4_;
+        }
+        else if constexpr(is_same<X, d128_t>::value)
+        {
+            return data_.d128x2_;
+        }
+        else if constexpr(is_same<X, d256_t>::value)
+        {
+            return data_.d256x1_;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(
+            is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value ||
+                is_same<X, d8_t>::value || is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
+                is_same<X, d64_t>::value || is_same<X, d128_t>::value || is_same<X, d256_t>::value,
+            "wrong!");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x256_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x128_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x64_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x32_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x16_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x8_;
+        }
+        else if constexpr(is_same<X, d64_t>::value)
+        {
+            return data_.d64x4_;
+        }
+        else if constexpr(is_same<X, d128_t>::value)
+        {
+            return data_.d128x2_;
+        }
+        else if constexpr(is_same<X, d256_t>::value)
+        {
+            return data_.d256x1_;
+        }
+    }
+};
+
+using int64_t = long;
+
+// fp64
+using double2_t = typename vector_type<double, 2>::type;
+using double4_t = typename vector_type<double, 4>::type;
+
+// fp32
+using float2_t  = typename vector_type<float, 2>::type;
+using float4_t  = typename vector_type<float, 4>::type;
+using float8_t  = typename vector_type<float, 8>::type;
+using float16_t = typename vector_type<float, 16>::type;
+using float32_t = typename vector_type<float, 32>::type;
+using float64_t = typename vector_type<float, 64>::type;
+
+// fp16
+using half2_t  = typename vector_type<half_t, 2>::type;
+using half4_t  = typename vector_type<half_t, 4>::type;
+using half8_t  = typename vector_type<half_t, 8>::type;
+using half16_t = typename vector_type<half_t, 16>::type;
+using half32_t = typename vector_type<half_t, 32>::type;
+using half64_t = typename vector_type<half_t, 64>::type;
+
+// bfp16
+using bhalf2_t  = typename vector_type<bhalf_t, 2>::type;
+using bhalf4_t  = typename vector_type<bhalf_t, 4>::type;
+using bhalf8_t  = typename vector_type<bhalf_t, 8>::type;
+using bhalf16_t = typename vector_type<bhalf_t, 16>::type;
+using bhalf32_t = typename vector_type<bhalf_t, 32>::type;
+using bhalf64_t = typename vector_type<bhalf_t, 64>::type;
+
+// i32
+using int32x2_t  = typename vector_type<int32_t, 2>::type;
+using int32x4_t  = typename vector_type<int32_t, 4>::type;
+using int32x8_t  = typename vector_type<int32_t, 8>::type;
+using int32x16_t = typename vector_type<int32_t, 16>::type;
+using int32x32_t = typename vector_type<int32_t, 32>::type;
+using int32x64_t = typename vector_type<int32_t, 64>::type;
+
+// i8
+using int8x2_t  = typename vector_type<int8_t, 2>::type;
+using int8x4_t  = typename vector_type<int8_t, 4>::type;
+using int8x8_t  = typename vector_type<int8_t, 8>::type;
+using int8x16_t = typename vector_type<int8_t, 16>::type;
+using int8x32_t = typename vector_type<int8_t, 32>::type;
+using int8x64_t = typename vector_type<int8_t, 64>::type;
+
+// f8
+using f8x2_t  = typename vector_type<f8_t, 2>::type;
+using f8x4_t  = typename vector_type<f8_t, 4>::type;
+using f8x8_t  = typename vector_type<f8_t, 8>::type;
+using f8x16_t = typename vector_type<f8_t, 16>::type;
+using f8x32_t = typename vector_type<f8_t, 32>::type;
+using f8x64_t = typename vector_type<f8_t, 64>::type;
+
+template <typename T>
+struct NumericLimits
+{
+    __host__ __device__ static constexpr T Min() { return std::numeric_limits<T>::min(); }
+
+    __host__ __device__ static constexpr T Max() { return std::numeric_limits<T>::max(); }
+
+    __host__ __device__ static constexpr T Lowest() { return std::numeric_limits<T>::lowest(); }
+
+    __host__ __device__ static constexpr T QuietNaN()
+    {
+        return std::numeric_limits<T>::quiet_NaN();
+    }
+
+    __host__ __device__ static constexpr T Infinity() { return std::numeric_limits<T>::infinity(); }
+};
+
+template <>
+struct NumericLimits<half_t>
+{
+    static constexpr unsigned short binary_min    = 0x0400;
+    static constexpr unsigned short binary_max    = 0x7BFF;
+    static constexpr unsigned short binary_lowest = 0xFBFF;
+    static constexpr unsigned short binary_qnan   = 0x7FFF;
+
+    __host__ __device__ static constexpr half_t Min() { return bit_cast<half_t>(binary_min); }
+
+    __host__ __device__ static constexpr half_t Max() { return bit_cast<half_t>(binary_max); }
+
+    __host__ __device__ static constexpr half_t Lowest() { return bit_cast<half_t>(binary_lowest); }
+
+    __host__ __device__ static constexpr half_t QuietNaN() { return bit_cast<half_t>(binary_qnan); }
+};
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+template <>
+struct NumericLimits<int4_t>
+{
+    __host__ __device__ static constexpr int4_t Min() { return int4_t(-8); }
+
+    __host__ __device__ static constexpr int4_t Max() { return int4_t(7); }
+
+    __host__ __device__ static constexpr int4_t Lowest() { return int4_t(-8); }
+};
+#endif // CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+
+template <>
+struct NumericLimits<f8_t>
+{
+    static constexpr uint8_t binary_min    = 0x08; // 0b00001000
+    static constexpr uint8_t binary_max    = 0x77; // 0b01110111
+    static constexpr uint8_t binary_lowest = 0xF7; // 0b11110111
+    static constexpr uint8_t binary_qnan   = 0x80; // 0b10000000
+
+    __host__ __device__ static constexpr f8_t Min() { return bit_cast<f8_t>(binary_min); }
+
+    __host__ __device__ static constexpr f8_t Max() { return bit_cast<f8_t>(binary_max); }
+
+    __host__ __device__ static constexpr f8_t Lowest() { return bit_cast<f8_t>(binary_lowest); }
+
+    __host__ __device__ static constexpr f8_t QuietNaN() { return bit_cast<f8_t>(binary_qnan); }
+};
+
+} // namespace ck
diff --git a/include/ck/utility/debug.hpp b/include/ck/utility/debug.hpp
new file mode 100644
index 0000000000..80346f0d9f
--- /dev/null
+++ b/include/ck/utility/debug.hpp
@@ -0,0 +1,85 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef UTILITY_DEBUG_HPP
+#define UTILITY_DEBUG_HPP
+
+namespace ck {
+namespace debug {
+
+namespace detail {
+template <typename T, typename Enable = void>
+struct PrintAsType;
+
+template <typename T>
+struct PrintAsType<T, typename std::enable_if<std::is_floating_point<T>::value>::type>
+{
+    using type = float;
+    __host__ __device__ static void Print(const T& p) { printf("%.3f ", static_cast<type>(p)); }
+};
+
+template <>
+struct PrintAsType<ck::half_t, void>
+{
+    using type = float;
+    __host__ __device__ static void Print(const ck::half_t& p)
+    {
+        printf("%.3f ", static_cast<type>(p));
+    }
+};
+
+template <typename T>
+struct PrintAsType<T, typename std::enable_if<std::is_integral<T>::value>::type>
+{
+    using type = int;
+    __host__ __device__ static void Print(const T& p) { printf("%d ", static_cast<type>(p)); }
+};
+} // namespace detail
+
+// Print at runtime the data in shared memory in 128 bytes per row format given shared mem pointer
+// and the number of elements. Can optionally specify strides between elements and how many bytes'
+// worth of data per row.
+//
+// Usage example:
+//
+//   debug::print_shared(a_block_buf.p_data_, index_t(a_block_desc_k0_m_k1.GetElementSpaceSize()));
+//
+template <typename T, index_t element_stride = 1, index_t row_bytes = 128>
+__device__ void print_shared(T const* p_shared, index_t num_elements)
+{
+    constexpr index_t row_elements = row_bytes / sizeof(T);
+    static_assert((element_stride >= 1 && element_stride <= row_elements),
+                  "element_stride should between [1, row_elements]");
+
+    index_t wgid = blockIdx.x + blockIdx.y * gridDim.x + gridDim.x * gridDim.y * blockIdx.z;
+    index_t tid =
+        (threadIdx.z * (blockDim.x * blockDim.y)) + (threadIdx.y * blockDim.x) + threadIdx.x;
+
+    __syncthreads();
+
+    if(tid == 0)
+    {
+        printf("\nWorkgroup id %d, bytes per row %d, element stride %d\n\n",
+               wgid,
+               row_bytes,
+               element_stride);
+        for(index_t i = 0; i < num_elements; i += row_elements)
+        {
+            printf("elem %5d: ", i);
+            for(index_t j = 0; j < row_elements; j += element_stride)
+            {
+                detail::PrintAsType<T>::Print(p_shared[i + j]);
+            }
+
+            printf("\n");
+        }
+        printf("\n");
+    }
+
+    __syncthreads();
+}
+
+} // namespace debug
+} // namespace ck
+
+#endif // UTILITY_DEBUG_HPP
diff --git a/include/ck/utility/dynamic_buffer.hpp b/include/ck/utility/dynamic_buffer.hpp
new file mode 100644
index 0000000000..52e59bcb50
--- /dev/null
+++ b/include/ck/utility/dynamic_buffer.hpp
@@ -0,0 +1,50 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "buffer_view.hpp"
+
+// FIXME: deprecate DynamicBuffer, use BufferView instead
+
+namespace ck {
+
+// FIXME: deprecate DynamicBuffer, use BufferView instead
+template <AddressSpaceEnum BufferAddressSpace,
+          typename T,
+          typename ElementSpaceSize,
+          bool InvalidElementUseNumericalZeroValue,
+          AmdBufferCoherenceEnum Coherence = AmdBufferCoherenceEnum::DefaultCoherence>
+using DynamicBuffer = BufferView<BufferAddressSpace,
+                                 T,
+                                 ElementSpaceSize,
+                                 InvalidElementUseNumericalZeroValue,
+                                 Coherence>;
+
+// FIXME: deprecate make_dynamic_buffer, use make_buffer_view instead
+template <AddressSpaceEnum BufferAddressSpace,
+          AmdBufferCoherenceEnum Coherence = AmdBufferCoherenceEnum::DefaultCoherence,
+          typename T,
+          typename ElementSpaceSize>
+__host__ __device__ constexpr auto make_dynamic_buffer(T* p, ElementSpaceSize element_space_size)
+{
+    return make_buffer_view<BufferAddressSpace, Coherence, T, ElementSpaceSize>(p,
+                                                                                element_space_size);
+}
+
+// FIXME: deprecate make_dynamic_buffer, use make_buffer_view instead
+template <
+    AddressSpaceEnum BufferAddressSpace,
+    AmdBufferCoherenceEnum Coherence = AmdBufferCoherenceEnum::DefaultCoherence,
+    typename T,
+    typename ElementSpaceSize,
+    typename X,
+    typename enable_if<is_same<remove_cvref_t<T>, remove_cvref_t<X>>::value, bool>::type = false>
+__host__ __device__ constexpr auto
+make_dynamic_buffer(T* p, ElementSpaceSize element_space_size, X invalid_element_value)
+{
+    return make_buffer_view<BufferAddressSpace, Coherence, T, ElementSpaceSize>(
+        p, element_space_size, invalid_element_value);
+}
+
+} // namespace ck
diff --git a/include/ck/utility/enable_if.hpp b/include/ck/utility/enable_if.hpp
new file mode 100644
index 0000000000..c0a3c99f1f
--- /dev/null
+++ b/include/ck/utility/enable_if.hpp
@@ -0,0 +1,14 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+template <bool B, typename T = void>
+using enable_if = std::enable_if<B, T>;
+
+template <bool B, typename T = void>
+using enable_if_t = typename std::enable_if<B, T>::type;
+
+} // namespace ck
diff --git a/include/ck/utility/f8_utils.hpp b/include/ck/utility/f8_utils.hpp
new file mode 100644
index 0000000000..bb13f98154
--- /dev/null
+++ b/include/ck/utility/f8_utils.hpp
@@ -0,0 +1,250 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+
+namespace ck {
+
+// fp8 rounding modes
+// use standard for rounding to nearest, the faster one
+// use stochastic for stochastic rounding, helps to avoid error accumulation
+enum class f8_rounding_mode
+{
+    standard,
+    stochastic
+};
+
+} // namespace ck
+
+namespace ck::utils {
+
+namespace {
+
+template <typename T, bool negative_zero_nan, bool clip, bool stoch>
+__host__ __device__ f8_t run_cast_to_f8(T x, uint32_t rng)
+{
+    // check data type
+    constexpr bool is_half  = std::is_same<T, half_t>::value;
+    constexpr bool is_float = std::is_same<T, float>::value;
+
+    // fp8 exponent/mantissa layout
+    constexpr int f8_exp  = 4;
+    constexpr int f8_mant = 3;
+
+    // resulting type exponent/mantissa layout
+    constexpr int type_exp  = is_half ? 5 : 8;
+    constexpr int type_mant = is_half ? 10 : 23;
+
+    int exponent;
+    uint32_t head, mantissa, sign;
+    // nan code is same for float and half
+    constexpr uint8_t nan_code  = 0x80;
+    constexpr uint32_t nan_mask = is_half ? 0x7C00 : 0x7F800000;
+
+    // convert to bitwise
+    typedef typename std::conditional<std::is_same<T, half_t>::value, uint16_t, uint32_t>::type
+        T_bitwise;
+    T_bitwise x_bitwise = *(reinterpret_cast<T_bitwise*>(&x));
+
+    // unpack the input, depends on datatype
+    if constexpr(is_float)
+    {
+        head     = x_bitwise & 0xFF800000;
+        mantissa = x_bitwise & 0x7FFFFF;
+        exponent = (head >> type_mant) & 0xFF;
+        sign     = head >> (type_exp + type_mant);
+    }
+    else if constexpr(is_half)
+    {
+        head     = x_bitwise & 0xFC00;
+        mantissa = x_bitwise & 0x3FF;
+        exponent = (head >> type_mant) & 0x1F;
+        sign     = head >> (type_exp + type_mant);
+    }
+
+    uint32_t signed_inf   = (sign << (type_exp + type_mant)) + (((1 << type_exp) - 1) << type_mant);
+    uint32_t drop_mask    = (1 << (type_mant - f8_mant)) - 1;
+    constexpr int max_exp = (1 << f8_exp) - (negative_zero_nan ? 1 : 2);
+    constexpr int exp_low_cutoff =
+        (1 << (type_exp - 1)) - (1 << (f8_exp - 1)) + 1 - (negative_zero_nan ? 1 : 0);
+
+    if constexpr(negative_zero_nan)
+    {
+        if((x_bitwise & nan_mask) == nan_mask)
+            return nan_code;
+    }
+    else
+    {
+        if((x_bitwise & nan_mask) == nan_mask)
+            return signed_inf + (mantissa != 0 ? 1 : 0);
+    }
+
+    // check if x is 0.0
+    if(x_bitwise == 0)
+        return 0;
+
+    exponent -= exp_low_cutoff - 1;
+    if(exponent <= 0)
+        drop_mask = (1 << (type_mant - f8_mant + 1 - exponent)) - 1;
+    mantissa += 1 << type_mant;
+    // apply random number if needed
+    mantissa += (stoch ? rng : mantissa) & drop_mask;
+    if(mantissa >= (2 << type_mant))
+    {
+        mantissa >>= 1;
+        exponent++;
+    }
+    mantissa >>= (type_mant - f8_mant);
+
+    // check negative exponent
+    if(exponent <= 0)
+    {
+        if(x_bitwise == 0)
+            return 0;
+        else
+        {
+            // subnormal range; represented by a subnormal float8 (exponent 0)
+            // and involves loss of accuracy
+            mantissa >>= 1 - exponent;
+            exponent = 0;
+        }
+    }
+    // above range: quantize to maximum possible float of the same sign
+    else if(exponent > max_exp)
+    {
+        if(clip)
+        {
+            mantissa = (1 << f8_mant) - 1;
+            exponent = max_exp;
+        }
+        else
+        {
+            return signed_inf;
+        }
+    }
+
+    // check if x is 0.0 or -0.0
+    if(exponent == 0 && mantissa == 0)
+        return negative_zero_nan ? 0 : (sign << (f8_exp + f8_mant));
+    mantissa &= (1 << f8_mant) - 1;
+    return (sign << (f8_exp + f8_mant)) | (exponent << f8_mant) | mantissa;
+}
+
+template <typename T, bool negative_zero_nan>
+__host__ __device__ T run_cast_from_f8(f8_t x)
+{
+    // check data type
+    constexpr bool is_half  = std::is_same<T, half_t>::value;
+    constexpr bool is_float = std::is_same<T, float>::value;
+
+    // fp8 exponent/mantissa layout
+    constexpr int f8_exp  = 4;
+    constexpr int f8_mant = 3;
+
+    // resulting type exponent/mantissa layout
+    constexpr int type_exp  = is_half ? 5 : 8;
+    constexpr int type_mant = is_half ? 10 : 23;
+
+    // prepare the codes
+    constexpr uint8_t nan_code = 0x80;
+    T fInf, fNegInf, fNaN, fNeg0;
+    if constexpr(is_half)
+    {
+        constexpr uint16_t ihInf    = 0x7C00;
+        constexpr uint16_t ihNegInf = 0xFC00;
+        constexpr uint16_t ihNaN    = 0x7C01;
+        constexpr uint16_t ihNeg0   = 0x8000;
+        fInf                        = *(reinterpret_cast<const half_t*>(&ihInf));
+        fNegInf                     = *(reinterpret_cast<const half_t*>(&ihNegInf));
+        fNaN                        = *(reinterpret_cast<const half_t*>(&ihNaN));
+        fNeg0                       = *(reinterpret_cast<const half_t*>(&ihNeg0));
+    }
+    else if constexpr(is_float)
+    {
+        constexpr uint32_t ifInf    = 0x7F800000;
+        constexpr uint32_t ifNegInf = 0xFF800000;
+        constexpr uint32_t ifNaN    = 0x7F800001;
+        constexpr uint32_t ifNeg0   = 0x80000000;
+        fInf                        = *(reinterpret_cast<const float*>(&ifInf));
+        fNegInf                     = *(reinterpret_cast<const float*>(&ifNegInf));
+        fNaN                        = *(reinterpret_cast<const float*>(&ifNaN));
+        fNeg0                       = *(reinterpret_cast<const float*>(&ifNeg0));
+    }
+
+    // unpack the input
+    uint32_t sign     = x >> (f8_exp + f8_mant);
+    uint32_t mantissa = x & ((1 << f8_mant) - 1);
+    int exponent      = (x & 0x7F) >> f8_mant;
+
+    constexpr int exp_low_cutoff =
+        (1 << (type_exp - 1)) - (1 << (f8_exp - 1)) + 1 - (negative_zero_nan ? 1 : 0);
+    typename std::conditional<std::is_same<T, half_t>::value, uint16_t, uint32_t>::type retval;
+
+    if constexpr(negative_zero_nan)
+    {
+        if(x == nan_code)
+            return fNaN;
+    }
+    else
+    {
+        if(x == nan_code)
+            return fNeg0;
+        if(exponent == ((1 << f8_exp) - 1))
+            return (mantissa == 0) ? (sign ? fNegInf : fInf) : fNaN;
+    }
+
+    // subnormal input
+    if(exponent == 0)
+    {
+        // guaranteed mantissa!=0 since cases 0x0 and 0x80 are handled above
+        int sh = 1 + __builtin_clz(mantissa) - ((1 + type_exp + type_mant) - f8_mant);
+        mantissa <<= sh;
+        mantissa &= ((1 << f8_mant) - 1);
+        exponent += 1 - sh;
+    }
+    exponent += exp_low_cutoff - 1;
+    mantissa <<= type_mant - f8_mant;
+
+    // subnormal output (occurs when T=half, we=5, negative_zero_nan=true)
+    if(exponent <= 0)
+    {
+        mantissa |= 1 << type_mant;
+        mantissa >>= 1 - exponent;
+        exponent = 0;
+    }
+
+    retval = (sign << (type_exp + type_mant)) | (exponent << type_mant) | mantissa;
+    return *(reinterpret_cast<const T*>(&retval));
+}
+
+} // namespace
+
+template <typename T, bool negative_zero_nan, bool clip, bool stoch>
+__host__ __device__ f8_t cast_to_f8(T x, uint32_t rng)
+{
+    // check datatype
+    constexpr bool is_half  = std::is_same<T, half_t>::value;
+    constexpr bool is_float = std::is_same<T, float>::value;
+    static_assert(is_half || is_float, "Only half and float can be casted to f8.");
+
+    return run_cast_to_f8<T, negative_zero_nan, clip, stoch>(x, rng);
+}
+
+template <typename T, bool negative_zero_nan>
+__host__ __device__ T cast_from_f8(f8_t x)
+{
+    // check datatype
+    constexpr bool is_half  = std::is_same<T, half_t>::value;
+    constexpr bool is_float = std::is_same<T, float>::value;
+    static_assert(is_half || is_float, "only half and float are supported.");
+
+    // check if x is 0.0
+    if(x == 0)
+        return static_cast<T>(0);
+
+    return run_cast_from_f8<T, negative_zero_nan>(x);
+}
+
+} // namespace ck::utils
diff --git a/include/ck/utility/functional.hpp b/include/ck/utility/functional.hpp
new file mode 100644
index 0000000000..91797d2409
--- /dev/null
+++ b/include/ck/utility/functional.hpp
@@ -0,0 +1,131 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/integral_constant.hpp"
+#include "ck/utility/type.hpp"
+
+namespace ck {
+
+// TODO: right? wrong?
+struct forwarder
+{
+    template <typename T>
+    __host__ __device__ constexpr T&& operator()(T&& x) const
+    {
+        return static_cast<T&&>(x);
+    }
+};
+
+struct swallow
+{
+    template <typename... Ts>
+    __host__ __device__ constexpr swallow(Ts&&...)
+    {
+    }
+};
+
+template <typename T>
+struct logical_and
+{
+    constexpr bool operator()(const T& x, const T& y) const { return x && y; }
+};
+
+template <typename T>
+struct logical_or
+{
+    constexpr bool operator()(const T& x, const T& y) const { return x || y; }
+};
+
+template <typename T>
+struct logical_not
+{
+    constexpr bool operator()(const T& x) const { return !x; }
+};
+
+// Emulate if constexpr
+template <bool>
+struct static_if;
+
+template <>
+struct static_if<true>
+{
+    using Type = static_if<true>;
+
+    template <typename F>
+    __host__ __device__ constexpr auto operator()(F f) const
+    {
+        // This is a trick for compiler:
+        //   Pass forwarder to lambda "f" as "auto" argument, and make sure "f" will
+        //   use it,
+        //   this will make "f" a generic lambda, so that "f" won't be compiled
+        //   until being
+        //   instantiated here
+        f(forwarder{});
+        return Type{};
+    }
+
+    template <typename F>
+    __host__ __device__ static void Else(F)
+    {
+    }
+};
+
+template <>
+struct static_if<false>
+{
+    using Type = static_if<false>;
+
+    template <typename F>
+    __host__ __device__ constexpr auto operator()(F) const
+    {
+        return Type{};
+    }
+
+    template <typename F>
+    __host__ __device__ static void Else(F f)
+    {
+        // This is a trick for compiler:
+        //   Pass forwarder to lambda "f" as "auto" argument, and make sure "f" will
+        //   use it,
+        //   this will make "f" a generic lambda, so that "f" won't be compiled
+        //   until being
+        //   instantiated here
+        f(forwarder{});
+    }
+};
+
+template <bool predicate, class X, class Y>
+struct conditional;
+
+template <class X, class Y>
+struct conditional<true, X, Y>
+{
+    using type = X;
+};
+
+template <class X, class Y>
+struct conditional<false, X, Y>
+{
+    using type = Y;
+};
+
+template <bool predicate, class X, class Y>
+using conditional_t = typename conditional<predicate, X, Y>::type;
+
+// z = predicate ? x : y
+template <bool predicate, typename X, typename Y>
+constexpr auto conditional_expr(X&& x, Y&& y)
+{
+    if constexpr(predicate)
+    {
+        return std::forward<X>(x);
+    }
+    else
+    {
+        return std::forward<Y>(y);
+    }
+}
+
+} // namespace ck
diff --git a/include/ck/utility/functional2.hpp b/include/ck/utility/functional2.hpp
new file mode 100644
index 0000000000..99c65f4eb8
--- /dev/null
+++ b/include/ck/utility/functional2.hpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/functional.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+
+namespace detail {
+
+template <class>
+struct static_for_impl;
+
+template <index_t... Is>
+struct static_for_impl<Sequence<Is...>>
+{
+    template <class F>
+    __host__ __device__ constexpr void operator()(F f) const
+    {
+        swallow{(f(Number<Is>{}), 0)...};
+    }
+};
+
+} // namespace detail
+
+// F signature: F(Number<Iter>)
+template <index_t NBegin, index_t NEnd, index_t Increment>
+struct static_for
+{
+    __host__ __device__ constexpr static_for()
+    {
+        static_assert(Increment != 0 && (NEnd - NBegin) % Increment == 0,
+                      "Wrong! should satisfy (NEnd - NBegin) % Increment == 0");
+        static_assert((Increment > 0 && NBegin <= NEnd) || (Increment < 0 && NBegin >= NEnd),
+                      "wrongs! should (Increment > 0 && NBegin <= NEnd) || (Increment < 0 && "
+                      "NBegin >= NEnd)");
+    }
+
+    template <class F>
+    __host__ __device__ constexpr void operator()(F f) const
+    {
+        detail::static_for_impl<typename arithmetic_sequence_gen<NBegin, NEnd, Increment>::type>{}(
+            f);
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/utility/functional3.hpp b/include/ck/utility/functional3.hpp
new file mode 100644
index 0000000000..fd81420260
--- /dev/null
+++ b/include/ck/utility/functional3.hpp
@@ -0,0 +1,144 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/functional.hpp"
+#include "ck/utility/functional2.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/multi_index.hpp"
+
+namespace ck {
+
+namespace detail {
+
+// RemainLengths: Sequence<...>
+// Orders: Sequence<...>
+template <class RemainLengths, class Orders>
+struct static_ford_impl
+{
+    __host__ __device__ constexpr static_ford_impl()
+    {
+        static_assert(RemainLengths::Size() > 0, "wrong! should not get here");
+    }
+
+    // F signature: F(Sequence<...>)
+    // CurrentOrderedId: Sequence<...>
+    template <class F, class CurrentOrderedId>
+    __host__ __device__ constexpr void operator()(F f, CurrentOrderedId) const
+    {
+        static_for<0, RemainLengths::Front(), 1>{}([=](auto I) {
+            static_ford_impl<decltype(RemainLengths::PopFront()), Orders>{}(
+                f, CurrentOrderedId::PushBack(I));
+        });
+    }
+};
+
+template <class Orders>
+struct static_ford_impl<Sequence<>, Orders>
+{
+    // F signature: F(Sequence<...>)
+    // OrderedId: Sequence<...>
+    template <class F, class OrderedId>
+    __host__ __device__ constexpr void operator()(F f, OrderedId) const
+    {
+        // retrive unordered Id
+        f(OrderedId::ReorderGivenOld2New(Orders{}));
+    }
+};
+
+// RemainLengths: Sequence<...>
+// Orders: Sequence<...>
+template <class RemainLengths, class Orders>
+struct ford_impl
+{
+    __host__ __device__ constexpr ford_impl()
+    {
+        static_assert(RemainLengths::Size() > 0, "wrong! should not get here");
+    }
+
+    // F signature: F(Array<...> multi_id)
+    // CurrentOrderdId: Array<...>
+    template <class F, class CurrentOrderedId>
+    __host__ __device__ constexpr void operator()(F f, CurrentOrderedId current_ordered_id) const
+    {
+        for(index_t i = 0; i < RemainLengths::Front(); ++i)
+        {
+            ford_impl<decltype(RemainLengths::PopFront()), Orders>{}(
+                f, container_push_back(current_ordered_id, i));
+        }
+    }
+};
+
+template <class Orders>
+struct ford_impl<Sequence<>, Orders>
+{
+    // F signature: F(Array<...> multi_id)
+    // CurrentOrderdId: Array<...>
+    template <class F, class CurrentOrderedId>
+    __host__ __device__ constexpr void operator()(F f, CurrentOrderedId current_ordered_id) const
+    {
+        // retrive unordered Id
+        f(container_reorder_given_old2new(current_ordered_id, Orders{}));
+    }
+};
+
+} // namespace detail
+
+// Lengths is Sequence<...>, it is the length of each dimension for
+// N-dimensional loop
+// Orders is Sequence<...>, it is the order of dimension in which static_ford
+// will loop over each
+// dimension
+template <class Lengths,
+          class Orders = typename arithmetic_sequence_gen<0, Lengths::Size(), 1>::type>
+struct static_ford
+{
+    __host__ __device__ constexpr static_ford()
+    {
+        static_assert(Lengths::Size() > 0, "wrong! Lengths is empty");
+        static_assert(Lengths::Size() == Orders::Size(), "wrong! inconsistent size");
+    }
+
+    // F signature: F(Sequence<...> multi_id)
+    // multi_id is the unordered multi-index
+    template <class F>
+    __host__ __device__ constexpr void operator()(F f) const
+    {
+        constexpr auto ordered_lengths = Lengths::ReorderGivenNew2Old(Orders{});
+        detail::static_ford_impl<decltype(ordered_lengths), Orders>{}(f, Sequence<>{});
+    }
+};
+
+// Lengths is Sequence<...>, it is the length of each dimension for
+// N-dimensional loop
+// Orders is Sequence<...>, it is the order of dimension in which ford will loop
+// over each
+// dimension
+template <class Lengths,
+          class Orders = typename arithmetic_sequence_gen<0, Lengths::Size(), 1>::type>
+struct ford
+{
+    __host__ __device__ constexpr ford()
+    {
+        static_assert(Lengths::Size() > 0, "wrong! Lengths is empty");
+        static_assert(Lengths::Size() == Orders::Size(), "wrong! inconsistent size");
+    }
+
+    // F signature: F(Array<...> multi_id)
+    // multi_id is the unordered multi-index
+    template <class F>
+    __host__ __device__ constexpr void operator()(F f) const
+    {
+        constexpr auto ordered_lengths = Lengths::ReorderGivenNew2Old(Orders{});
+
+        for(index_t i = 0; i < ordered_lengths.Front(); ++i)
+        {
+            detail::ford_impl<decltype(ordered_lengths.PopFront()), Orders>{}(f,
+                                                                              make_multi_index(i));
+        }
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/utility/functional4.hpp b/include/ck/utility/functional4.hpp
new file mode 100644
index 0000000000..66c7ef753e
--- /dev/null
+++ b/include/ck/utility/functional4.hpp
@@ -0,0 +1,72 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "sequence.hpp"
+#include "tuple.hpp"
+#include "array.hpp"
+
+namespace ck {
+
+namespace detail {
+
+template <typename Indices>
+struct unpack_impl;
+
+template <index_t... Is>
+struct unpack_impl<Sequence<Is...>>
+{
+    template <typename F, typename X>
+    __host__ __device__ constexpr auto operator()(F&& f, X&& x) const
+    {
+#if 0
+        return std::forward<F>(f)(std::forward<X>(x).At(Number<Is>{})...);
+#else
+        return std::forward<F>(f)(std::forward<X>(x).template At<Is>()...);
+#endif
+    }
+};
+
+template <typename Seq0, typename Seq1>
+struct unpack2_impl;
+
+// TODO: remove this, after properly implementing unpack that takes any number of containers
+template <index_t... Is, index_t... Js>
+struct unpack2_impl<Sequence<Is...>, Sequence<Js...>>
+{
+    template <typename F, typename X, typename Y>
+    __host__ __device__ constexpr auto operator()(F&& f, X&& x, Y&& y) const
+    {
+#if 0
+        return std::forward<F>(f)(std::forward<X>(x).At(Number<Is>{})...,
+                                  std::forward<Y>(y).At(Number<Js>{})...);
+#else
+        return std::forward<F>(f)(std::forward<X>(x).template At<Is>()...,
+                                  std::forward<Y>(y).template At<Js>()...);
+#endif
+    }
+};
+
+} // namespace detail
+
+template <typename F, typename X>
+__host__ __device__ constexpr auto unpack(F&& f, X&& x)
+{
+    using X_ = remove_reference_t<X>;
+    return detail::unpack_impl<typename arithmetic_sequence_gen<0, X_::Size(), 1>::type>{}(
+        std::forward<F>(f), std::forward<X>(x));
+}
+
+// TODO: properly implement unpack that takes any number of containers
+template <typename F, typename X, typename Y>
+__host__ __device__ constexpr auto unpack2(F&& f, X&& x, Y&& y)
+{
+    using X_ = remove_reference_t<X>;
+    using Y_ = remove_reference_t<Y>;
+    return detail::unpack2_impl<typename arithmetic_sequence_gen<0, X_::Size(), 1>::type,
+                                typename arithmetic_sequence_gen<0, Y_::Size(), 1>::type>{}(
+        std::forward<F>(f), std::forward<X>(x), std::forward<Y>(y));
+}
+
+} // namespace ck
diff --git a/include/ck/utility/generic_memory_space_atomic.hpp b/include/ck/utility/generic_memory_space_atomic.hpp
new file mode 100644
index 0000000000..98f40a4363
--- /dev/null
+++ b/include/ck/utility/generic_memory_space_atomic.hpp
@@ -0,0 +1,123 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+#include "data_type.hpp"
+
+namespace ck {
+
+// Caution: DO NOT REMOVE
+// intentionally have only declaration but no definition to cause compilation failure when trying to
+// instantiate this template. The purpose is to make the implementation of atomic_add explicit for
+// each datatype.
+template <typename X>
+__device__ X atomic_add(X* p_dst, const X& x);
+
+template <>
+__device__ int32_t atomic_add<int32_t>(int32_t* p_dst, const int32_t& x)
+{
+    return atomicAdd(p_dst, x);
+}
+
+template <>
+__device__ uint32_t atomic_add<uint32_t>(uint32_t* p_dst, const uint32_t& x)
+{
+    return atomicAdd(p_dst, x);
+}
+
+template <>
+__device__ float atomic_add<float>(float* p_dst, const float& x)
+{
+    return atomicAdd(p_dst, x);
+}
+
+template <>
+__device__ double atomic_add<double>(double* p_dst, const double& x)
+{
+    return atomicAdd(p_dst, x);
+}
+
+template <>
+__device__ float2_t atomic_add<float2_t>(float2_t* p_dst, const float2_t& x)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    const vector_type<float, 2> vx{x};
+    vector_type<float, 2> vy{0};
+
+    vy.template AsType<float>()(I0) =
+        atomicAdd(c_style_pointer_cast<float*>(p_dst), vx.template AsType<float>()[I0]);
+    vy.template AsType<float>()(I1) =
+        atomicAdd(c_style_pointer_cast<float*>(p_dst) + 1, vx.template AsType<float>()[I1]);
+
+    return vy.template AsType<float2_t>()[I0];
+}
+
+template <>
+__device__ double2_t atomic_add<double2_t>(double2_t* p_dst, const double2_t& x)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    const vector_type<double, 2> vx{x};
+    vector_type<double, 2> vy{0};
+
+    vy.template AsType<double>()(I0) =
+        atomicAdd(c_style_pointer_cast<double*>(p_dst), vx.template AsType<double>()[I0]);
+    vy.template AsType<double>()(I1) =
+        atomicAdd(c_style_pointer_cast<double*>(p_dst) + 1, vx.template AsType<double>()[I1]);
+
+    return vy.template AsType<double2_t>()[I0];
+}
+
+// Caution: DO NOT REMOVE
+// intentionally have only declaration but no definition to cause compilation failure when trying to
+// instantiate this template. The purpose is to make the implementation of atomic_max explicit for
+// each datatype.
+
+template <typename X>
+__device__ X atomic_max(X* p_dst, const X& x);
+
+template <>
+__device__ int32_t atomic_max<int32_t>(int32_t* p_dst, const int32_t& x)
+{
+    return atomicMax(p_dst, x);
+}
+
+template <>
+__device__ uint32_t atomic_max<uint32_t>(uint32_t* p_dst, const uint32_t& x)
+{
+    return atomicMax(p_dst, x);
+}
+
+template <>
+__device__ float atomic_max<float>(float* p_dst, const float& x)
+{
+    return atomicMax(p_dst, x);
+}
+
+template <>
+__device__ double atomic_max<double>(double* p_dst, const double& x)
+{
+    return atomicMax(p_dst, x);
+}
+
+template <>
+__device__ float2_t atomic_max<float2_t>(float2_t* p_dst, const float2_t& x)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    const vector_type<float, 2> vx{x};
+    vector_type<float, 2> vy{0};
+
+    vy.template AsType<float>()(I0) =
+        atomicMax(c_style_pointer_cast<float*>(p_dst), vx.template AsType<float>()[I0]);
+    vy.template AsType<float>()(I1) =
+        atomicMax(c_style_pointer_cast<float*>(p_dst) + 1, vx.template AsType<float>()[I1]);
+
+    return vy.template AsType<float2_t>()[I0];
+}
+
+} // namespace ck
diff --git a/include/ck/utility/get_id.hpp b/include/ck/utility/get_id.hpp
new file mode 100644
index 0000000000..c5e7179acd
--- /dev/null
+++ b/include/ck/utility/get_id.hpp
@@ -0,0 +1,39 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+
+namespace ck {
+
+__host__ __device__ constexpr index_t get_warp_size()
+{
+    // warpSize is defined by HIP
+    return warpSize;
+}
+
+__device__ index_t get_grid_size() { return gridDim.x; }
+
+__device__ index_t get_block_size() { return blockDim.x; }
+
+// TODO: deprecate these
+__device__ index_t get_thread_local_1d_id() { return threadIdx.x; }
+
+__device__ index_t get_thread_global_1d_id() { return blockIdx.x * blockDim.x + threadIdx.x; }
+
+__device__ index_t get_block_1d_id() { return blockIdx.x; }
+
+// Use these instead
+__device__ index_t get_lane_id() { return __lane_id(); }
+
+__device__ index_t get_warp_id()
+{
+    return __builtin_amdgcn_readfirstlane(threadIdx.x / get_warp_size());
+}
+
+__device__ index_t get_thread_id() { return threadIdx.x; }
+
+__device__ index_t get_block_id() { return blockIdx.x; }
+
+} // namespace ck
diff --git a/include/ck/utility/get_shift.hpp b/include/ck/utility/get_shift.hpp
new file mode 100644
index 0000000000..0a93081cfd
--- /dev/null
+++ b/include/ck/utility/get_shift.hpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+template <index_t N>
+static constexpr __device__ index_t get_shift()
+{
+    return (get_shift<N / 2>() + 1);
+};
+
+template <>
+constexpr __device__ index_t get_shift<1>()
+{
+    return (0);
+}
+
+} // namespace ck
diff --git a/include/ck/utility/ignore.hpp b/include/ck/utility/ignore.hpp
new file mode 100644
index 0000000000..f70a182fd4
--- /dev/null
+++ b/include/ck/utility/ignore.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+// https://en.cppreference.com/w/cpp/utility/tuple/ignore
+
+namespace ck {
+
+namespace detail {
+struct ignore_t
+{
+    template <typename T>
+    constexpr void operator=(T&&) const noexcept
+    {
+    }
+};
+} // namespace detail
+
+inline constexpr detail::ignore_t ignore;
+
+} // namespace ck
diff --git a/include/ck/utility/inner_product.hpp b/include/ck/utility/inner_product.hpp
new file mode 100644
index 0000000000..b58b2b3319
--- /dev/null
+++ b/include/ck/utility/inner_product.hpp
@@ -0,0 +1,236 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+#include "data_type.hpp"
+#include "type_convert.hpp"
+
+namespace ck {
+
+template <typename TA, typename TB, typename TC>
+__device__ void inner_product(const TA& a, const TB& b, TC& c);
+
+template <>
+__device__ void inner_product<float, float, float>(const float& a, const float& b, float& c)
+{
+#if CK_USE_AMD_V_MAC_INLINE_ASM && defined(CK_USE_AMD_V_MAC_F32)
+    asm volatile("\n \
+            v_mac_f32 %0, %1, %2 \n \
+            "
+                 : "=v"(c)
+                 : "v"(a), "v"(b), "0"(c));
+#elif CK_USE_AMD_V_MAC_INLINE_ASM && defined(CK_USE_AMD_V_FMAC_F32)
+    asm volatile("\n \
+            v_fmac_f32 %0, %1, %2 \n \
+            "
+                 : "=v"(c)
+                 : "v"(a), "v"(b), "0"(c));
+#else
+    c += a * b;
+#endif
+}
+
+template <>
+__device__ void
+inner_product<float2_t, float2_t, float>(const float2_t& a, const float2_t& b, float& c)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    inner_product(vector_type<float, 2>{a}.AsType<float>()[I0],
+                  vector_type<float, 2>{b}.AsType<float>()[I0],
+                  c);
+
+    inner_product(vector_type<float, 2>{a}.AsType<float>()[I1],
+                  vector_type<float, 2>{b}.AsType<float>()[I1],
+                  c);
+}
+
+template <>
+__device__ void
+inner_product<float4_t, float4_t, float>(const float4_t& a, const float4_t& b, float& c)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    inner_product(vector_type<float, 4>{a}.AsType<float>()[I0],
+                  vector_type<float, 4>{b}.AsType<float>()[I0],
+                  c);
+
+    inner_product(vector_type<float, 4>{a}.AsType<float>()[I1],
+                  vector_type<float, 4>{b}.AsType<float>()[I1],
+                  c);
+
+    inner_product(vector_type<float, 4>{a}.AsType<float>()[I2],
+                  vector_type<float, 4>{b}.AsType<float>()[I2],
+                  c);
+
+    inner_product(vector_type<float, 4>{a}.AsType<float>()[I3],
+                  vector_type<float, 4>{b}.AsType<float>()[I3],
+                  c);
+}
+
+template <>
+__device__ void inner_product<half2_t, half2_t, float>(const half2_t& a, const half2_t& b, float& c)
+{
+#if defined(CK_USE_AMD_V_DOT2_F32_F16)
+#if CK_USE_AMD_V_DOT_INLINE_ASM
+    // Use 3 x s_nop to avoid hazard (mi200 cdna2 isa page 47
+    // https://www.amd.com/system/files/TechDocs/instinct-mi200-cdna2-instruction-set-architecture.pdf
+    // ) s_nop with parameter 2 is equal to 3 x s_nop
+    asm volatile("\n \
+            v_dot2_f32_f16 %0, %1, %2, %0\n \
+            s_nop 2 \n \
+            "
+                 : "=v"(c)
+                 : "v"(a), "v"(b), "0"(c));
+#else
+    c = __builtin_amdgcn_fdot2(a, b, c, false);
+#endif
+#else
+    const vector_type<half_t, 2> a_vector{a};
+    const vector_type<half_t, 2> b_vector{b};
+
+    static_for<0, 2, 1>{}([&](auto i) {
+        c += type_convert<float>(a_vector.AsType<half_t>()[i]) *
+             type_convert<float>(b_vector.AsType<half_t>()[i]);
+    });
+#endif
+}
+
+template <>
+__device__ void inner_product<half4_t, half4_t, float>(const half4_t& a, const half4_t& b, float& c)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    inner_product(vector_type<half_t, 4>{a}.AsType<half2_t>()[I0],
+                  vector_type<half_t, 4>{b}.AsType<half2_t>()[I0],
+                  c);
+
+    inner_product(vector_type<half_t, 4>{a}.AsType<half2_t>()[I1],
+                  vector_type<half_t, 4>{b}.AsType<half2_t>()[I1],
+                  c);
+}
+
+template <>
+__device__ void inner_product<half8_t, half8_t, float>(const half8_t& a, const half8_t& b, float& c)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    inner_product(vector_type<half_t, 8>{a}.AsType<half2_t>()[I0],
+                  vector_type<half_t, 8>{b}.AsType<half2_t>()[I0],
+                  c);
+
+    inner_product(vector_type<half_t, 8>{a}.AsType<half2_t>()[I1],
+                  vector_type<half_t, 8>{b}.AsType<half2_t>()[I1],
+                  c);
+
+    inner_product(vector_type<half_t, 8>{a}.AsType<half2_t>()[I2],
+                  vector_type<half_t, 8>{b}.AsType<half2_t>()[I2],
+                  c);
+
+    inner_product(vector_type<half_t, 8>{a}.AsType<half2_t>()[I3],
+                  vector_type<half_t, 8>{b}.AsType<half2_t>()[I3],
+                  c);
+}
+
+template <>
+__device__ void inner_product<int8_t, int8_t, int32_t>(const int8_t& a, const int8_t& b, int32_t& c)
+{
+    c += type_convert<int32_t>(a) * type_convert<int32_t>(b);
+}
+
+template <>
+__device__ void
+inner_product<int8x2_t, int8x2_t, int32_t>(const int8x2_t& a, const int8x2_t& b, int32_t& c)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    inner_product(vector_type<int8_t, 2>{a}.AsType<int8_t>()[I0],
+                  vector_type<int8_t, 2>{b}.AsType<int8_t>()[I0],
+                  c);
+
+    inner_product(vector_type<int8_t, 2>{a}.AsType<int8_t>()[I1],
+                  vector_type<int8_t, 2>{b}.AsType<int8_t>()[I1],
+                  c);
+}
+
+template <>
+__device__ void
+inner_product<int8x4_t, int8x4_t, int32_t>(const int8x4_t& a, const int8x4_t& b, int32_t& c)
+{
+#if defined(CK_USE_AMD_V_DOT4_I32_I8)
+#if CK_USE_AMD_V_DOT_INLINE_ASM
+    // Use 3 x s_nop to avoid hazard (mi200 cdna2 isa page 47
+    // https://www.amd.com/system/files/TechDocs/instinct-mi200-cdna2-instruction-set-architecture.pdf
+    // ) s_nop with parameter 2 is equal to 3 x s_nop
+    asm volatile("\n \
+            v_dot4_i32_i8 %0, %1, %2, %0\n \
+            s_nop 2 \n \
+            "
+                 : "=v"(c)
+                 : "v"(bit_cast<int32_t>(a)), "v"(bit_cast<int32_t>(b)), "0"(c));
+#else
+    c = __builtin_amdgcn_sdot4(bit_cast<int32_t>(a), bit_cast<int32_t>(b), c, false);
+#endif
+#else
+    const vector_type<int8_t, 4> a_vector{a};
+    const vector_type<int8_t, 4> b_vector{b};
+
+    static_for<0, 4, 1>{}([&](auto i) {
+        c += type_convert<int32_t>(a_vector.AsType<int8_t>()[i]) *
+             type_convert<int32_t>(b_vector.AsType<int8_t>()[i]);
+    });
+#endif
+}
+
+template <>
+__device__ void
+inner_product<int8x8_t, int8x8_t, int32_t>(const int8x8_t& a, const int8x8_t& b, int32_t& c)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+
+    inner_product(vector_type<int8_t, 8>{a}.AsType<int8x4_t>()[I0],
+                  vector_type<int8_t, 8>{b}.AsType<int8x4_t>()[I0],
+                  c);
+
+    inner_product(vector_type<int8_t, 8>{a}.AsType<int8x4_t>()[I1],
+                  vector_type<int8_t, 8>{b}.AsType<int8x4_t>()[I1],
+                  c);
+}
+
+template <>
+__device__ void
+inner_product<int8x16_t, int8x16_t, int32_t>(const int8x16_t& a, const int8x16_t& b, int32_t& c)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    inner_product(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I0],
+                  vector_type<int8_t, 16>{b}.AsType<int8x4_t>()[I0],
+                  c);
+
+    inner_product(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I1],
+                  vector_type<int8_t, 16>{b}.AsType<int8x4_t>()[I1],
+                  c);
+
+    inner_product(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I2],
+                  vector_type<int8_t, 16>{b}.AsType<int8x4_t>()[I2],
+                  c);
+
+    inner_product(vector_type<int8_t, 16>{a}.AsType<int8x4_t>()[I3],
+                  vector_type<int8_t, 16>{b}.AsType<int8x4_t>()[I3],
+                  c);
+}
+
+} // namespace ck
diff --git a/include/ck/utility/inner_product_dpp8.hpp b/include/ck/utility/inner_product_dpp8.hpp
new file mode 100644
index 0000000000..ccd7a4e628
--- /dev/null
+++ b/include/ck/utility/inner_product_dpp8.hpp
@@ -0,0 +1,142 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+#include "amd_gemm_dpp.hpp"
+#include "data_type.hpp"
+#include "type_convert.hpp"
+
+namespace ck {
+
+namespace dpp8 {
+
+template <int SrcLaneIdx>
+__device__ void inline_v_dot2c_dpp8_instr(const half2_t& a, const half2_t& b, float& c);
+
+// clang-format off
+template <>
+__device__ void inline_v_dot2c_dpp8_instr<0>(const half2_t& a, const half2_t& b, float& c){
+    asm volatile("\n v_dot2c_f32_f16_dpp %0, %1, %2 dpp8:[0, 0, 0, 0, 0, 0, 0, 0]" : "=v"(c) : "v"(a), "v"(b), "0"(c));
+}
+template <>
+__device__ void inline_v_dot2c_dpp8_instr<1>(const half2_t& a, const half2_t& b, float& c){
+    asm volatile("\n v_dot2c_f32_f16_dpp %0, %1, %2 dpp8:[1, 1, 1, 1, 1, 1, 1, 1]" : "=v"(c) : "v"(a), "v"(b), "0"(c));
+}
+template <>
+__device__ void inline_v_dot2c_dpp8_instr<2>(const half2_t& a, const half2_t& b, float& c){
+    asm volatile("\n v_dot2c_f32_f16_dpp %0, %1, %2 dpp8:[2, 2, 2, 2, 2, 2, 2, 2]" : "=v"(c) : "v"(a), "v"(b), "0"(c));
+}
+template <>
+__device__ void inline_v_dot2c_dpp8_instr<3>(const half2_t& a, const half2_t& b, float& c){
+    asm volatile("\n v_dot2c_f32_f16_dpp %0, %1, %2 dpp8:[3, 3, 3, 3, 3, 3, 3, 3]" : "=v"(c) : "v"(a), "v"(b), "0"(c));
+}
+template <>
+__device__ void inline_v_dot2c_dpp8_instr<4>(const half2_t& a, const half2_t& b, float& c){
+    asm volatile("\n v_dot2c_f32_f16_dpp %0, %1, %2 dpp8:[4, 4, 4, 4, 4, 4, 4, 4]" : "=v"(c) : "v"(a), "v"(b), "0"(c));
+}
+template <>
+__device__ void inline_v_dot2c_dpp8_instr<5>(const half2_t& a, const half2_t& b, float& c){
+    asm volatile("\n v_dot2c_f32_f16_dpp %0, %1, %2 dpp8:[5, 5, 5, 5, 5, 5, 5, 5]" : "=v"(c) : "v"(a), "v"(b), "0"(c));
+}
+template <>
+__device__ void inline_v_dot2c_dpp8_instr<6>(const half2_t& a, const half2_t& b, float& c){
+    asm volatile("\n v_dot2c_f32_f16_dpp %0, %1, %2 dpp8:[6, 6, 6, 6, 6, 6, 6, 6]" : "=v"(c) : "v"(a), "v"(b), "0"(c));
+}
+template <>
+__device__ void inline_v_dot2c_dpp8_instr<7>(const half2_t& a, const half2_t& b, float& c){
+    asm volatile("\n v_dot2c_f32_f16_dpp %0, %1, %2 dpp8:[7, 7, 7, 7, 7, 7, 7, 7]" : "=v"(c) : "v"(a), "v"(b), "0"(c));
+}
+// clang-format on
+
+/**
+ * Dot product of two vectors using `v_dot` instruction with DPP8 submitted as inline assembly.
+ */
+template <int SrcLaneIdx, bool ShareA>
+__device__ void inline_v_dot2c_dpp8(const half2_t& a, const half2_t& b, float& c)
+{
+    static_assert(SrcLaneIdx >= 0 && SrcLaneIdx < dpp8::lane_group_size,
+                  "DPP8 src broadcast lane out of range <0, 7>.");
+    if constexpr(ShareA)
+    {
+        inline_v_dot2c_dpp8_instr<SrcLaneIdx>(a, b, c);
+    }
+    else
+    {
+        inline_v_dot2c_dpp8_instr<SrcLaneIdx>(b, a, c);
+    }
+}
+
+/**
+ * DPP8 instrinsics expects to get an integer mask, hardcoding integers for specific broadcast
+ * patters.
+ */
+constexpr std::array<int, dpp8::lane_group_size> IntrinsicMaskDpp8 = {
+    0,        // 0, 0, 0, 0, 0, 0, 0, 0
+    2396745,  // 1, 1, 1, 1, 1, 1, 1, 1
+    4793490,  // 2, 2, 2, 2, 2, 2, 2, 2
+    7190235,  // 3, 3, 3, 3, 3, 3, 3, 3
+    9586980,  // 4, 4, 4, 4, 4, 4, 4, 4
+    11983725, // 5, 5, 5, 5, 5, 5, 5, 5
+    14380470, // 6, 6, 6, 6, 6, 6, 6, 6
+    16777215, // 7, 7, 7, 7, 7, 7, 7, 7
+};
+
+/**
+ * Returns DPP8 sel modifier as an integer required for the intrinsic instruction.
+ */
+template <int SrcLaneIdx>
+constexpr int get_dpp_sel_mask_broadcast()
+{
+    static_assert(SrcLaneIdx >= 0 && SrcLaneIdx < dpp8::lane_group_size,
+                  "DPP8 src broadcast lane out of range <0, 7>.");
+    return IntrinsicMaskDpp8[SrcLaneIdx];
+}
+
+template <int SrcLaneIdx>
+__device__ void intrinsic_fdot2_impl(const half2_t& a, const half2_t& b, float& c)
+{
+    constexpr int sel_mask = get_dpp_sel_mask_broadcast<SrcLaneIdx>();
+    const half2_t val_from_other_lane =
+        bit_cast<half2_t>(__builtin_amdgcn_mov_dpp8(bit_cast<int>(a), sel_mask));
+    c = __builtin_amdgcn_fdot2(val_from_other_lane, b, c, false);
+}
+
+/**
+ * Dot product of two vectors using `v_dot` instruction with DPP8 submitted using intrinsics.
+ */
+template <int SrcLaneIdx, bool ShareA>
+__device__ void intrinsic_fdot2(const half2_t& a, const half2_t& b, float& c)
+{
+    if constexpr(ShareA)
+    {
+        intrinsic_fdot2_impl<SrcLaneIdx>(a, b, c);
+    }
+    else
+    {
+        intrinsic_fdot2_impl<SrcLaneIdx>(b, a, c);
+    }
+}
+
+/**
+ * Dot product of two input vectors `a`, `b` using `v_dot` instructions with DPP modifier.
+ *
+ * DPP modifier allows us to share one of the vectors between lanes in a lane group.
+ * When `ShareA` is set, instruction uses vector `a` from lane `SrcLaneIdx` from the same
+ * lane group (8 lanes per lane group in DPP8). When `ShareA` is not set, vector `b` is shared.
+ * Note that all the threads in a lane group uses the same vector - broadcast pattern.
+ *
+ * `SrcLaneIdx` must be in range from 0 to 7.
+ */
+template <typename TA, typename TB, typename TC, int SrcLaneIdx, bool ShareA>
+__device__ void inner_product_dpp(const TA& a, const TB& b, TC& c)
+{
+#if CK_USE_AMD_V_DOT_DPP8_INLINE_ASM
+    inline_v_dot2c_dpp8<SrcLaneIdx, ShareA>(a, b, c);
+#else
+    intrinsic_fdot2<SrcLaneIdx, ShareA>(a, b, c);
+#endif
+}
+
+} // namespace dpp8
+
+} // namespace ck
diff --git a/include/ck/utility/integral_constant.hpp b/include/ck/utility/integral_constant.hpp
new file mode 100644
index 0000000000..04c07bc4cd
--- /dev/null
+++ b/include/ck/utility/integral_constant.hpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+template <class T, T v>
+struct integral_constant
+{
+    static constexpr T value = v;
+    typedef T value_type;
+    typedef integral_constant type;
+    __host__ __device__ constexpr operator value_type() const noexcept { return value; }
+    __host__ __device__ constexpr value_type operator()() const noexcept { return value; }
+    __host__ __device__ static constexpr bool IsStatic() { return true; };
+    __host__ __device__ void Print() const { print(v); }
+};
+
+template <typename TX, TX X, typename TY, TY Y>
+__host__ __device__ constexpr auto operator+(integral_constant<TX, X>, integral_constant<TY, Y>)
+{
+    return integral_constant<decltype(X + Y), X + Y>{};
+}
+
+template <typename TX, TX X, typename TY, TY Y>
+__host__ __device__ constexpr auto operator-(integral_constant<TX, X>, integral_constant<TY, Y>)
+{
+    static_assert(Y <= X, "wrong!");
+    return integral_constant<decltype(X - Y), X - Y>{};
+}
+
+template <typename TX, TX X, typename TY, TY Y>
+__host__ __device__ constexpr auto operator*(integral_constant<TX, X>, integral_constant<TY, Y>)
+{
+    return integral_constant<decltype(X * Y), X * Y>{};
+}
+
+template <typename TX, TX X, typename TY, TY Y>
+__host__ __device__ constexpr auto operator/(integral_constant<TX, X>, integral_constant<TY, Y>)
+{
+    static_assert(Y > 0, "wrong!");
+    return integral_constant<decltype(X / Y), X / Y>{};
+}
+
+template <typename TX, TX X, typename TY, TY Y>
+__host__ __device__ constexpr auto operator%(integral_constant<TX, X>, integral_constant<TY, Y>)
+{
+    static_assert(Y > 0, "wrong!");
+    return integral_constant<decltype(X % Y), X % Y>{};
+}
+
+} // namespace ck
diff --git a/include/ck/utility/is_static.hpp b/include/ck/utility/is_static.hpp
new file mode 100644
index 0000000000..382993e821
--- /dev/null
+++ b/include/ck/utility/is_static.hpp
@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/remove_cvref.hpp"
+
+namespace ck {
+
+namespace detail {
+
+template <typename T>
+struct is_static_impl
+{
+    static constexpr bool value = T::IsStatic();
+};
+
+template <>
+struct is_static_impl<int32_t>
+{
+    static constexpr bool value = false;
+};
+
+template <>
+struct is_static_impl<int64_t>
+{
+    static constexpr bool value = false;
+};
+
+} // namespace detail
+
+template <typename T>
+using is_static = detail::is_static_impl<remove_cvref_t<T>>;
+
+template <typename T>
+inline constexpr bool is_static_v = is_static<T>::value;
+
+// TODO: deprecate this
+template <typename T>
+using is_known_at_compile_time = is_static<T>;
+
+} // namespace ck
diff --git a/include/ck/utility/macro_func_array_of_array_to_tuple_of_sequence.hpp b/include/ck/utility/macro_func_array_of_array_to_tuple_of_sequence.hpp
new file mode 100644
index 0000000000..491f7d7f7f
--- /dev/null
+++ b/include/ck/utility/macro_func_array_of_array_to_tuple_of_sequence.hpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/macro_func_array_to_sequence.hpp"
+
+// Macro function
+// convert constexpr Array<Array<index_t, xxx>, xxx> to Tuple<Sequence<...>, ...>
+//   Input:
+//     1. a_of_b_impl: constexpr Array<Array<index_t, xxx>, xxx>
+//     2. a_size: constexper index_t
+//     3. bs_sizes: constexpr Array<index_t, xxx>
+//   Output:
+//     Tuple<Sequence<...>, ...>
+#define TO_TUPLE_OF_SEQUENCE(a_of_b_impl, a_size, bs_sizes)             \
+    [a_of_b_impl, a_size, bs_sizes] {                                   \
+        return ck::generate_tuple(                                      \
+            [=](auto i) {                                               \
+                constexpr auto b_impl    = a_of_b_impl[i];              \
+                constexpr index_t b_size = bs_sizes[i];                 \
+                constexpr auto b         = TO_SEQUENCE(b_impl, b_size); \
+                return b;                                               \
+            },                                                          \
+            ck::Number<a_size>{});                                      \
+    }()
diff --git a/include/ck/utility/macro_func_array_to_sequence.hpp b/include/ck/utility/macro_func_array_to_sequence.hpp
new file mode 100644
index 0000000000..c4721dfe2d
--- /dev/null
+++ b/include/ck/utility/macro_func_array_to_sequence.hpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+// Macro function
+// convert constexpr Array to Sequence
+#define TO_SEQUENCE(a, n)                                                                      \
+    [a, n] {                                                                                   \
+        static_assert(a.Size() >= n, "wrong! out of bound");                                   \
+                                                                                               \
+        static_assert(n <= 10, "not implemented");                                             \
+                                                                                               \
+        if constexpr(n == 0)                                                                   \
+        {                                                                                      \
+            return ck::Sequence<>{};                                                           \
+        }                                                                                      \
+        else if constexpr(n == 1)                                                              \
+        {                                                                                      \
+            return ck::Sequence<a[0]>{};                                                       \
+        }                                                                                      \
+        else if constexpr(n == 2)                                                              \
+        {                                                                                      \
+            return ck::Sequence<a[0], a[1]>{};                                                 \
+        }                                                                                      \
+        else if constexpr(n == 3)                                                              \
+        {                                                                                      \
+            return ck::Sequence<a[0], a[1], a[2]>{};                                           \
+        }                                                                                      \
+        else if constexpr(n == 4)                                                              \
+        {                                                                                      \
+            return ck::Sequence<a[0], a[1], a[2], a[3]>{};                                     \
+        }                                                                                      \
+        else if constexpr(n == 5)                                                              \
+        {                                                                                      \
+            return ck::Sequence<a[0], a[1], a[2], a[3], a[4]>{};                               \
+        }                                                                                      \
+        else if constexpr(n == 6)                                                              \
+        {                                                                                      \
+            return ck::Sequence<a[0], a[1], a[2], a[3], a[4], a[5]>{};                         \
+        }                                                                                      \
+        else if constexpr(n == 7)                                                              \
+        {                                                                                      \
+            return ck::Sequence<a[0], a[1], a[2], a[3], a[4], a[5], a[6]>{};                   \
+        }                                                                                      \
+        else if constexpr(n == 8)                                                              \
+        {                                                                                      \
+            return ck::Sequence<a[0], a[1], a[2], a[3], a[4], a[5], a[6], a[7]>{};             \
+        }                                                                                      \
+        else if constexpr(n == 9)                                                              \
+        {                                                                                      \
+            return ck::Sequence<a[0], a[1], a[2], a[3], a[4], a[5], a[6], a[7], a[8]>{};       \
+        }                                                                                      \
+        else if constexpr(n == 10)                                                             \
+        {                                                                                      \
+            return ck::Sequence<a[0], a[1], a[2], a[3], a[4], a[5], a[6], a[7], a[8], a[9]>{}; \
+        }                                                                                      \
+    }()
diff --git a/include/ck/utility/magic_division.hpp b/include/ck/utility/magic_division.hpp
new file mode 100644
index 0000000000..a0f4e11930
--- /dev/null
+++ b/include/ck/utility/magic_division.hpp
@@ -0,0 +1,260 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "integral_constant.hpp"
+#include "number.hpp"
+#include "type.hpp"
+#include "tuple.hpp"
+
+namespace ck {
+
+// magic number division
+// Caution:
+//   1. For uint32_t as dividend: magic number division implementation being used would produce
+//   correct result if the dividend is uint32_t and its value is within 31-bit value range.
+//   2. For int32_t as dividendd: magic number division for int32_t dividened has not been
+//   implemented, the int32_t dividend would be bit-wise interpreted as uint32_t and magic number
+//   division implementation for uint32_t is then used. Therefore, dividend value need to be
+//   non-negative.
+// TODO:
+//   1. Implement magic number divison for int32_t
+//   2. Implement magic number divison for unit32_t with 32-bit value range
+struct MagicDivision32BitRange
+{
+    // uint32_t
+    __host__ __device__ static constexpr auto CalculateMagicNumbers(uint32_t divisor)
+    {
+        // WARNING: magic division is only valid for division inside this range.
+        // assert(divisor >= 1 && divisor <= INT32_MAX)
+
+        uint32_t shift_u32 = 0;
+
+        while((1U << shift_u32) < divisor)
+        {
+            shift_u32++;
+        };
+
+        uint64_t tmp_u64        = ((1UL << shift_u32) - divisor) << 32;
+        uint32_t multiplier_u32 = tmp_u64 / divisor + 1;
+
+        return make_tuple(multiplier_u32, shift_u32);
+    }
+
+    // integral_constant<uint32_t, .>
+    template <uint32_t Divisor>
+    __host__ __device__ static constexpr auto
+        CalculateMagicNumbers(integral_constant<uint32_t, Divisor>)
+    {
+        constexpr auto tmp = CalculateMagicNumbers(uint32_t{Divisor});
+
+        constexpr uint32_t multiplier = tmp[Number<0>{}];
+        constexpr uint32_t shift      = tmp[Number<1>{}];
+
+        return make_tuple(integral_constant<uint32_t, multiplier>{},
+                          integral_constant<uint32_t, shift>{});
+    }
+
+    // integral_constant<int32_t, .>
+    template <int32_t Divisor>
+    __host__ __device__ static constexpr auto
+        CalculateMagicNumbers(integral_constant<int32_t, Divisor>)
+    {
+        return CalculateMagicNumbers(integral_constant<uint32_t, Divisor>{});
+    }
+
+    // magic division for uint32_t
+    __device__ static constexpr uint32_t
+    DoMagicDivision(uint32_t dividend, uint32_t multiplier, uint32_t shift)
+    {
+        uint32_t tmp = __umulhi(dividend, multiplier);
+        return (tmp + dividend) >> shift;
+    }
+
+    __host__ static constexpr uint32_t
+    DoMagicDivision(uint32_t dividend, uint32_t multiplier, uint32_t shift)
+    {
+        uint32_t tmp = (static_cast<uint64_t>(dividend) * multiplier) >> 32;
+        return (tmp + dividend) >> shift;
+    }
+
+    // magic division for int32_t
+    // HACK: use dividend_i32 as if it's uint32_t, dividend_i32 need to be
+    // non-negative for result to be correct
+    // TODO: figure out how to do magic number divison for int32_t as dividended
+    __device__ static constexpr int32_t
+    DoMagicDivision(int32_t dividend_i32, uint32_t multiplier, uint32_t shift)
+    {
+        uint32_t dividend_u32 = bit_cast<uint32_t>(dividend_i32);
+        uint32_t tmp          = __umulhi(dividend_u32, multiplier);
+        return (tmp + dividend_u32) >> shift;
+    }
+
+    __host__ static constexpr int32_t
+    DoMagicDivision(int32_t dividend_i32, uint32_t multiplier, uint32_t shift)
+    {
+        uint32_t dividend_u32 = bit_cast<uint32_t>(dividend_i32);
+        uint32_t tmp          = (static_cast<uint64_t>(dividend_u32) * multiplier) >> 32;
+        return (tmp + dividend_u32) >> shift;
+    }
+};
+
+// magic number division
+// This version on works for divisor and dividended between [0, 1 << 16]
+struct MagicDivision16BitRange
+{
+    // uint32_t
+    __host__ __device__ static constexpr auto CalculateMagicNumbers(uint32_t divisor)
+    {
+        // WARNING: magic division is only valid for division inside this range.
+        // assert(divisor >= 1 && divisor <= (1U << 16));
+
+        uint32_t shift_u32 = 0;
+
+        while((1U << shift_u32) < divisor)
+        {
+            shift_u32++;
+        };
+
+        uint32_t one            = 1;
+        uint32_t multiplier_u32 = ((one << 16) * ((one << shift_u32) - divisor)) / divisor + 1;
+
+        return make_tuple(multiplier_u32, shift_u32);
+    }
+
+    // integral_constant<uint32_t, .>
+    template <uint32_t Divisor>
+    __host__ __device__ static constexpr auto
+        CalculateMagicNumbers(integral_constant<uint32_t, Divisor>)
+    {
+        constexpr auto tmp = CalculateMagicNumbers(uint32_t{Divisor});
+
+        constexpr uint32_t multiplier = tmp[Number<0>{}];
+        constexpr uint32_t shift      = tmp[Number<1>{}];
+
+        return make_tuple(integral_constant<uint32_t, multiplier>{},
+                          integral_constant<uint32_t, shift>{});
+    }
+
+    // integral_constant<int32_t, .>
+    template <int32_t Divisor>
+    __host__ __device__ static constexpr auto
+        CalculateMagicNumbers(integral_constant<int32_t, Divisor>)
+    {
+        return CalculateMagicNumbers(integral_constant<uint32_t, Divisor>{});
+    }
+
+    // magic division for uint32_t
+    __device__ static constexpr uint32_t
+    DoMagicDivision(uint32_t dividend, uint32_t multiplier, uint32_t shift)
+    {
+        uint32_t tmp = (dividend * multiplier) >> 16;
+        return (tmp + dividend) >> shift;
+    }
+
+    __host__ static constexpr uint32_t
+    DoMagicDivision(uint32_t dividend, uint32_t multiplier, uint32_t shift)
+    {
+        uint32_t tmp = (dividend * multiplier) >> 16;
+        return (tmp + dividend) >> shift;
+    }
+
+    // magic division for int32_t
+    // HACK: use dividend_i32 as if it's uint32_t, dividend_i32 need to be
+    // non-negative for result to be correct
+    // TODO: figure out how to do magic number divison for int32_t as dividended
+    __device__ static constexpr int32_t
+    DoMagicDivision(int32_t dividend_i32, uint32_t multiplier, uint32_t shift)
+    {
+        uint32_t dividend_u32 = bit_cast<uint32_t>(dividend_i32);
+        uint32_t tmp          = (dividend_u32 * multiplier) >> 16;
+        return (tmp + dividend_u32) >> shift;
+    }
+
+    __host__ static constexpr int32_t
+    DoMagicDivision(int32_t dividend_i32, uint32_t multiplier, uint32_t shift)
+    {
+        uint32_t dividend_u32 = bit_cast<uint32_t>(dividend_i32);
+        uint32_t tmp          = (dividend_u32 * multiplier) >> 16;
+        return (tmp + dividend_u32) >> shift;
+    }
+};
+
+// use 32bit version
+using MagicDivision = MagicDivision32BitRange;
+
+struct MDiv
+{
+    // 1 dword -> 3 dword storage
+    uint32_t divisor;
+    uint32_t multiplier;
+    uint32_t shift; // TODO: 8 bit is enough
+
+    // prefer construct on host
+    __host__ __device__ MDiv(uint32_t divisor_) : divisor(divisor_)
+    {
+        auto tmp = MagicDivision::CalculateMagicNumbers(divisor_);
+
+        multiplier = tmp[Number<0>{}];
+        shift      = tmp[Number<1>{}];
+    }
+
+    __host__ __device__ MDiv() : divisor(0), multiplier(0), shift(0) {}
+
+    __host__ __device__ void update(uint32_t divisor_)
+    {
+        divisor  = divisor_;
+        auto tmp = MagicDivision::CalculateMagicNumbers(divisor_);
+
+        multiplier = tmp[Number<0>{}];
+        shift      = tmp[Number<1>{}];
+    }
+
+    __host__ __device__ uint32_t div(uint32_t dividend_) const
+    {
+        return MagicDivision::DoMagicDivision(dividend_, multiplier, shift);
+    }
+
+    __host__ __device__ void
+    divmod(uint32_t dividend_, uint32_t& quotient_, uint32_t& remainder_) const
+    {
+        quotient_  = div(dividend_);
+        remainder_ = dividend_ - (quotient_ * divisor);
+    }
+
+    __host__ __device__ uint32_t get() const { return divisor; }
+};
+
+struct MDiv2
+{
+    // 1 dword -> 2 dword storage, divisor need compute from runtime
+    uint32_t multiplier;
+    uint32_t shift; // TODO: 8 bit is enough
+
+    // prefer construct on host
+    __host__ __device__ MDiv2(uint32_t divisor_)
+    {
+        auto tmp = MagicDivision::CalculateMagicNumbers(divisor_);
+
+        multiplier = tmp[Number<0>{}];
+        shift      = tmp[Number<1>{}];
+    }
+
+    __host__ __device__ MDiv2() : multiplier(0), shift(0) {}
+
+    __host__ __device__ uint32_t div(uint32_t dividend_) const
+    {
+        return MagicDivision::DoMagicDivision(dividend_, multiplier, shift);
+    }
+
+    __host__ __device__ void
+    divmod(uint32_t dividend_, uint32_t divisor_, uint32_t& quotient_, uint32_t& remainder_) const
+    {
+        quotient_  = div(dividend_);
+        remainder_ = dividend_ - (quotient_ * divisor_);
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/utility/map.hpp b/include/ck/utility/map.hpp
new file mode 100644
index 0000000000..3783b6918e
--- /dev/null
+++ b/include/ck/utility/map.hpp
@@ -0,0 +1,158 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/array.hpp"
+#include "ck/utility/tuple.hpp"
+
+namespace ck {
+
+// naive Map
+template <typename Key, typename Data, index_t MaxSize = 128>
+struct Map
+{
+    using Pair = Tuple<Key, Data>;
+    using Impl = Array<Pair, MaxSize>;
+
+    Impl impl_;
+    index_t size_;
+
+    struct Iterator
+    {
+        Impl& impl_;
+        index_t pos_;
+
+        __host__ __device__ constexpr Iterator(Impl& impl, index_t pos) : impl_{impl}, pos_{pos} {}
+
+        __host__ __device__ constexpr Iterator& operator++()
+        {
+            pos_++;
+
+            return *this;
+        }
+
+        __host__ __device__ constexpr bool operator!=(const Iterator& other) const
+        {
+            return other.pos_ != pos_;
+        }
+
+        __host__ __device__ constexpr Pair& operator*() { return impl_.At(pos_); }
+    };
+
+    struct ConstIterator
+    {
+        const Impl& impl_;
+        index_t pos_;
+
+        __host__ __device__ constexpr ConstIterator(const Impl& impl, index_t pos)
+            : impl_{impl}, pos_{pos}
+        {
+        }
+
+        __host__ __device__ constexpr ConstIterator& operator++()
+        {
+            pos_++;
+
+            return *this;
+        }
+
+        __host__ __device__ constexpr bool operator!=(const ConstIterator& other) const
+        {
+            return other.pos_ != pos_;
+        }
+
+        __host__ __device__ constexpr const Pair& operator*() const { return impl_.At(pos_); }
+    };
+
+    __host__ __device__ constexpr Map() : impl_{}, size_{0} {}
+
+    __host__ __device__ constexpr index_t Size() const { return size_; }
+
+    __host__ __device__ void Clear() { size_ = 0; }
+
+    __host__ __device__ constexpr index_t FindPosition(const Key& key) const
+    {
+        for(index_t i = 0; i < Size(); i++)
+        {
+            if(impl_[i].template At<0>() == key)
+            {
+                return i;
+            }
+        }
+
+        return size_;
+    }
+
+    __host__ __device__ constexpr ConstIterator Find(const Key& key) const
+    {
+        return ConstIterator{impl_, FindPosition(key)};
+    }
+
+    __host__ __device__ constexpr Iterator Find(const Key& key)
+    {
+        return Iterator{impl_, FindPosition(key)};
+    }
+
+    __host__ __device__ constexpr const Data& operator[](const Key& key) const
+    {
+        const auto it = Find(key);
+
+        // FIXME
+        assert(pos < Size());
+
+        return impl_[it.pos_].template At<1>();
+    }
+
+    __host__ __device__ constexpr Data& operator()(const Key& key)
+    {
+        auto it = Find(key);
+
+        // if entry not found
+        if(it.pos_ == Size())
+        {
+            impl_(it.pos_).template At<0>() = key;
+            size_++;
+        }
+
+        // FIXME
+        assert(size_ <= MaxSize);
+
+        return impl_(it.pos_).template At<1>();
+    }
+
+    // WARNING: needed by compiler for C++ range-based for loop only, don't use this function!
+    __host__ __device__ constexpr ConstIterator begin() const { return ConstIterator{impl_, 0}; }
+
+    // WARNING: needed by compiler for C++ range-based for loop only, don't use this function!
+    __host__ __device__ constexpr ConstIterator end() const { return ConstIterator{impl_, size_}; }
+
+    // WARNING: needed by compiler for C++ range-based for loop only, don't use this function!
+    __host__ __device__ constexpr Iterator begin() { return Iterator{impl_, 0}; }
+
+    // WARNING: needed by compiler for C++ range-based for loop only, don't use this function!
+    __host__ __device__ constexpr Iterator end() { return Iterator{impl_, size_}; }
+
+    __host__ __device__ void Print() const
+    {
+        printf("Map{size_: %d, ", size_);
+        //
+        printf("impl_: [");
+        //
+        for(const auto& [key, data] : *this)
+        {
+            printf("{key: ");
+            print(key);
+            printf(", data: ");
+            print(data);
+            printf("}, ");
+        }
+        //
+        printf("]");
+        //
+        printf("}");
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/utility/math.hpp b/include/ck/utility/math.hpp
new file mode 100644
index 0000000000..89a3d54dc3
--- /dev/null
+++ b/include/ck/utility/math.hpp
@@ -0,0 +1,288 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "integral_constant.hpp"
+#include "number.hpp"
+#include "type.hpp"
+#include "enable_if.hpp"
+
+namespace ck {
+namespace math {
+
+template <typename T, T s>
+struct scales
+{
+    __host__ __device__ constexpr T operator()(T a) const { return s * a; }
+};
+
+template <typename T>
+struct plus
+{
+    __host__ __device__ constexpr T operator()(T a, T b) const { return a + b; }
+};
+
+template <typename T>
+struct minus
+{
+    __host__ __device__ constexpr T operator()(T a, T b) const { return a - b; }
+};
+
+struct multiplies
+{
+    template <typename A, typename B>
+    __host__ __device__ constexpr auto operator()(const A& a, const B& b) const
+    {
+        return a * b;
+    }
+};
+
+template <typename T>
+struct maximize
+{
+    __host__ __device__ constexpr T operator()(T a, T b) const { return a >= b ? a : b; }
+};
+
+template <typename T>
+struct minimize
+{
+    __host__ __device__ constexpr T operator()(T a, T b) const { return a <= b ? a : b; }
+};
+
+template <typename T>
+struct integer_divide_ceiler
+{
+    __host__ __device__ constexpr T operator()(T a, T b) const
+    {
+        static_assert(is_same<T, index_t>{} || is_same<T, int>{}, "wrong type");
+
+        return (a + b - Number<1>{}) / b;
+    }
+};
+
+template <typename X, typename Y>
+__host__ __device__ constexpr auto integer_divide_floor(X x, Y y)
+{
+    return x / y;
+}
+
+template <typename X, typename Y>
+__host__ __device__ constexpr auto integer_divide_ceil(X x, Y y)
+{
+    return (x + y - Number<1>{}) / y;
+}
+
+template <typename X, typename Y>
+__host__ __device__ constexpr auto integer_least_multiple(X x, Y y)
+{
+    return y * integer_divide_ceil(x, y);
+}
+
+template <typename T>
+__host__ __device__ constexpr T max(T x)
+{
+    return x;
+}
+
+template <typename T>
+__host__ __device__ constexpr T max(T x, T y)
+{
+    return x > y ? x : y;
+}
+
+template <index_t X>
+__host__ __device__ constexpr index_t max(Number<X>, index_t y)
+{
+    return X > y ? X : y;
+}
+
+template <index_t Y>
+__host__ __device__ constexpr index_t max(index_t x, Number<Y>)
+{
+    return x > Y ? x : Y;
+}
+
+template <typename X, typename... Ys>
+__host__ __device__ constexpr auto max(X x, Ys... ys)
+{
+    static_assert(sizeof...(Ys) > 0, "not enough argument");
+
+    return max(x, max(ys...));
+}
+
+template <typename T>
+__host__ __device__ constexpr T min(T x)
+{
+    return x;
+}
+
+template <typename T>
+__host__ __device__ constexpr T min(T x, T y)
+{
+    return x < y ? x : y;
+}
+
+template <index_t X>
+__host__ __device__ constexpr index_t min(Number<X>, index_t y)
+{
+    return X < y ? X : y;
+}
+
+template <index_t Y>
+__host__ __device__ constexpr index_t min(index_t x, Number<Y>)
+{
+    return x < Y ? x : Y;
+}
+
+template <typename X, typename... Ys>
+__host__ __device__ constexpr auto min(X x, Ys... ys)
+{
+    static_assert(sizeof...(Ys) > 0, "not enough argument");
+
+    return min(x, min(ys...));
+}
+
+template <typename T>
+__host__ __device__ constexpr T clamp(const T& x, const T& lowerbound, const T& upperbound)
+{
+    return min(max(x, lowerbound), upperbound);
+}
+
+// prevent implicit type casting
+template <typename T>
+__host__ T exp(T x);
+
+template <typename T>
+__device__ T exp(T x);
+
+// TODO: add f16 support using v_exp_f16
+
+template <>
+inline __device__ float exp<float>(float x)
+{
+    return __expf(x);
+}
+
+template <>
+inline __device__ double exp<double>(double x)
+{
+    return exp(x);
+}
+
+template <>
+inline __host__ float exp<float>(float x)
+{
+    return std::expf(x);
+}
+
+template <>
+inline __host__ double exp<double>(double x)
+{
+    return std::exp(x);
+}
+
+// greatest common divisor, aka highest common factor
+__host__ __device__ constexpr index_t gcd(index_t x, index_t y)
+{
+    if(x < 0)
+    {
+        return gcd(-x, y);
+    }
+    else if(y < 0)
+    {
+        return gcd(x, -y);
+    }
+    else if(x == y || x == 0)
+    {
+        return y;
+    }
+    else if(y == 0)
+    {
+        return x;
+    }
+    else if(x > y)
+    {
+        return gcd(x % y, y);
+    }
+    else
+    {
+        return gcd(x, y % x);
+    }
+}
+
+template <index_t X, index_t Y>
+__host__ __device__ constexpr auto gcd(Number<X>, Number<Y>)
+{
+    constexpr auto r = gcd(X, Y);
+
+    return Number<r>{};
+}
+
+template <typename X, typename... Ys, typename enable_if<sizeof...(Ys) >= 2, bool>::type = false>
+__host__ __device__ constexpr auto gcd(X x, Ys... ys)
+{
+    return gcd(x, gcd(ys...));
+}
+
+// least common multiple
+template <typename X, typename Y>
+__host__ __device__ constexpr auto lcm(X x, Y y)
+{
+    return (x * y) / gcd(x, y);
+}
+
+template <typename X, typename... Ys, typename enable_if<sizeof...(Ys) >= 2, bool>::type = false>
+__host__ __device__ constexpr auto lcm(X x, Ys... ys)
+{
+    return lcm(x, lcm(ys...));
+}
+
+template <typename T>
+struct equal
+{
+    __host__ __device__ constexpr bool operator()(T x, T y) const { return x == y; }
+};
+
+template <typename T>
+struct less
+{
+    __host__ __device__ constexpr bool operator()(T x, T y) const { return x < y; }
+};
+
+__host__ __device__ constexpr int32_t next_power_of_two(int32_t x)
+{
+    // TODO: x need to be 2 ~ 0x7fffffff. 0, 1, or larger than 0x7fffffff will compile fail
+    return 1 << (32 - __builtin_clz(x - 1));
+}
+
+template <index_t X>
+__host__ __device__ constexpr auto next_power_of_two()
+{
+    constexpr index_t y = next_power_of_two(X);
+    return Number<y>{};
+}
+
+template <index_t X>
+__host__ __device__ constexpr auto next_power_of_two(Number<X>)
+{
+    constexpr index_t y = next_power_of_two(X);
+    return Number<y>{};
+}
+
+__host__ __device__ constexpr int32_t integer_log2_floor(int32_t x)
+{
+    // TODO: x need to be 1 ~ 0x7fffffff
+    // __builtin_clz will produce unexpected result if x is 0;
+    return 31 - __builtin_clz(x);
+}
+
+__host__ __device__ constexpr bool is_power_of_two_integer(int32_t x)
+{
+    // TODO: x need to be 1 ~ 0x7fffffff
+    return x == (1 << integer_log2_floor(x));
+}
+
+} // namespace math
+} // namespace ck
diff --git a/include/ck/utility/math_ext.hpp b/include/ck/utility/math_ext.hpp
new file mode 100644
index 0000000000..d77fc744cf
--- /dev/null
+++ b/include/ck/utility/math_ext.hpp
@@ -0,0 +1,134 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+// Here should use MultiIndex<NSize>, instead of Tuple<Ys...>, although the former
+// is the alias of the latter. This is because compiler cannot infer the NSize if
+// using MultiIndex<NSize>
+// TODO: how to fix this?
+template <
+    typename... Ys,
+    typename X,
+    enable_if_t<!std::is_integral<X>::value && !std::is_floating_point<X>::value, bool> = false>
+__host__ __device__ constexpr auto operator+=(Tuple<Ys...>& y, const X& x)
+{
+    static_assert(X::Size() == sizeof...(Ys), "wrong! size not the same");
+    constexpr index_t NSize = sizeof...(Ys);
+    static_for<0, NSize, 1>{}([&](auto i) { y(i) += x[i]; });
+    return y;
+}
+
+template <
+    typename... Ys,
+    typename X,
+    enable_if_t<!std::is_integral<X>::value && !std::is_floating_point<X>::value, bool> = false>
+__host__ __device__ constexpr auto operator-=(Tuple<Ys...>& y, const X& x)
+{
+    static_assert(X::Size() == sizeof...(Ys), "wrong! size not the same");
+    constexpr index_t NSize = sizeof...(Ys);
+    static_for<0, NSize, 1>{}([&](auto i) { y(i) -= x[i]; });
+    return y;
+}
+
+template <
+    typename... Xs,
+    typename Y,
+    enable_if_t<!std::is_integral<Y>::value && !std::is_floating_point<Y>::value, bool> = false>
+__host__ __device__ constexpr auto operator+(const Tuple<Xs...>& x, const Y& y)
+{
+    static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same");
+    constexpr index_t NSize = sizeof...(Xs);
+
+    Tuple<Xs...> r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = x[i] + y[i]; });
+    return r;
+}
+
+template <
+    typename... Xs,
+    typename Y,
+    enable_if_t<!std::is_integral<Y>::value && !std::is_floating_point<Y>::value, bool> = false>
+__host__ __device__ constexpr auto operator-(const Tuple<Xs...>& x, const Y& y)
+{
+    static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same");
+    constexpr index_t NSize = sizeof...(Xs);
+
+    Tuple<Xs...> r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = x[i] - y[i]; });
+    return r;
+}
+
+template <
+    typename... Xs,
+    typename Y,
+    enable_if_t<!std::is_integral<Y>::value && !std::is_floating_point<Y>::value, bool> = false>
+__host__ __device__ constexpr auto operator*(const Tuple<Xs...>& x, const Y& y)
+{
+    static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same");
+    constexpr index_t NSize = sizeof...(Xs);
+
+    Tuple<Xs...> r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = x[i] * y[i]; });
+    return r;
+}
+
+// MultiIndex = scalar * MultiIndex
+template <typename... Xs,
+          typename Y,
+          enable_if_t<std::is_integral<Y>::value || std::is_floating_point<Y>::value, bool> = false>
+__host__ __device__ constexpr auto operator*(Y a, const Tuple<Xs...>& x)
+{
+    constexpr index_t NSize = sizeof...(Xs);
+
+    Tuple<Xs...> r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = a * x[i]; });
+    return r;
+}
+
+// MultiIndex = MultiIndex * scalar
+template <typename... Xs,
+          typename Y,
+          enable_if_t<std::is_integral<Y>::value || std::is_floating_point<Y>::value, bool> = false>
+__host__ __device__ constexpr auto operator*(const Tuple<Xs...>& x, Y a)
+{
+    return a * x;
+}
+
+template <typename... Xs, typename... Ys>
+__host__ __device__ constexpr auto operator/(const Tuple<Xs...>& x, const Tuple<Ys...>& y)
+{
+    static_assert(sizeof...(Xs) == sizeof...(Ys), "wrong!");
+
+    constexpr index_t NSize = sizeof...(Xs);
+
+    return generate_tuple([&](auto i) { return x[i] / y[i]; }, Number<NSize>{});
+}
+
+namespace mathext {
+
+template <typename... Xs>
+__host__ __device__ constexpr auto exp(const Tuple<Xs...>& x)
+{
+    constexpr index_t NSize = sizeof...(Xs);
+
+    Tuple<Xs...> r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = math::exp(x[i]); });
+    return r;
+}
+
+template <typename... Xs, typename Y>
+__host__ __device__ constexpr auto max(const Tuple<Xs...>& x, const Y& y)
+{
+    static_assert(Y::Size() == sizeof...(Xs), "wrong! size not the same");
+    constexpr index_t NSize = sizeof...(Xs);
+
+    Tuple<Xs...> r;
+    static_for<0, NSize, 1>{}([&](auto i) { r(i) = math::max(x[i], y[i]); });
+    return r;
+}
+
+} // namespace mathext
+} // namespace ck
diff --git a/include/ck/utility/math_v2.hpp b/include/ck/utility/math_v2.hpp
new file mode 100644
index 0000000000..5682bc2453
--- /dev/null
+++ b/include/ck/utility/math_v2.hpp
@@ -0,0 +1,194 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#ifndef __HIP_DEVICE_COMPILE__
+#include <cmath>
+#endif
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/type.hpp"
+
+namespace ck {
+namespace math {
+
+// math functions for the host,  some are implemented by calling C++ std functions
+
+static inline __host__ float abs(float x) { return std::abs(x); };
+
+static inline __host__ double abs(double x) { return std::abs(x); };
+
+static inline __host__ int8_t abs(int8_t x)
+{
+    int8_t sgn = x >> (8 - 1);
+
+    return (x ^ sgn) - sgn;
+};
+
+static inline __host__ int32_t abs(int32_t x)
+{
+    int32_t sgn = x >> (32 - 1);
+
+    return (x ^ sgn) - sgn;
+};
+
+static inline __host__ half_t abs(half_t x)
+{
+    uint16_t xx = ck::bit_cast<uint16_t>(x);
+
+    uint16_t abs_xx = xx & 0x7fff;
+
+    half_t abs_x = ck::bit_cast<half_t>(abs_xx);
+
+    return abs_x;
+};
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+static inline __host__ int4_t abs(int4_t x)
+{
+    int4_t sgn = x >> (4 - 1);
+    return (x ^ sgn) - sgn;
+}
+#endif
+
+static inline __host__ bool isnan(float x) { return std::isnan(x); };
+
+static inline __host__ bool isnan(double x) { return std::isnan(x); };
+
+static inline __host__ bool isnan(int8_t x)
+{
+    (void)x;
+    return false;
+};
+
+static inline __host__ bool isnan(int32_t x)
+{
+    (void)x;
+    return false;
+};
+
+static inline __host__ bool isnan(half_t x)
+{
+    uint16_t xx = ck::bit_cast<uint16_t>(x);
+
+    return (xx & 0x7FFF) > 0x7C00;
+};
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+static inline __host__ bool isnan(int4_t x)
+{
+    (void)x;
+    return false;
+};
+#endif
+
+static inline __host__ half_t sqrt(half_t x)
+{
+    return static_cast<half_t>(std::sqrt(static_cast<float>(x)));
+};
+
+static inline __host__ float sqrt(float x) { return std::sqrt(x); };
+
+static inline __host__ double sqrt(double x) { return std::sqrt(x); };
+
+static inline __host__ half_t tanh(half_t x)
+{
+    return static_cast<half_t>(std::tanh(static_cast<float>(x)));
+};
+
+static inline __host__ float tanh(float x) { return std::tanh(x); };
+
+static inline __host__ double tanh(double x) { return std::tanh(x); };
+
+// math functions for the HIP kernel,  some are implemented by calling hip builtin functions
+
+static inline __device__ float abs(float x) { return ::abs(x); };
+
+static inline __device__ double abs(double x) { return ::abs(x); };
+
+static inline __device__ int8_t abs(int8_t x)
+{
+    int8_t sgn = x >> (8 - 1);
+
+    return (x ^ sgn) - sgn;
+};
+
+static inline __device__ constexpr int32_t abs(int32_t x)
+{
+    int32_t sgn = x >> (32 - 1);
+
+    return (x ^ sgn) - sgn;
+};
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+static inline __device__ int4_t abs(int4_t x)
+{
+    int4_t sgn = x >> (4 - 1);
+
+    return (x ^ sgn) - sgn;
+};
+#endif
+
+static inline __device__ half_t abs(half_t x)
+{
+    uint16_t xx = ck::bit_cast<uint16_t>(x);
+
+    uint16_t abs_xx = xx & 0x7fff;
+
+    half_t abs_x = ck::bit_cast<half_t>(abs_xx);
+
+    return abs_x;
+};
+
+static inline __device__ bool isnan(float x) { return ::isnan(x); };
+
+static inline __device__ bool isnan(double x) { return ::isnan(x); };
+
+static inline __device__ bool isnan(int8_t x)
+{
+    (void)x;
+    return false;
+};
+
+static inline __device__ bool isnan(int32_t x)
+{
+    (void)x;
+    return false;
+};
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+static inline __device__ bool isnan(int4_t x)
+{
+    (void)x;
+    return false;
+};
+#endif
+
+static inline __device__ bool isnan(half_t x)
+{
+    uint16_t xx = ck::bit_cast<uint16_t>(x);
+
+    return (xx & 0x7FFF) > 0x7C00;
+};
+
+static inline __device__ half_t sqrt(half_t x)
+{
+    return static_cast<half_t>(__builtin_amdgcn_sqrtf(static_cast<float>(x)));
+};
+
+static inline __device__ float sqrt(float x) { return __builtin_amdgcn_sqrtf(x); };
+
+static inline __device__ double sqrt(double x) { return __builtin_amdgcn_sqrt(x); };
+
+static inline __device__ half_t tanh(half_t x)
+{
+    return static_cast<half_t>(::tanhf(static_cast<float>(x)));
+};
+
+static inline __device__ float tanh(float x) { return ::tanhf(x); };
+
+static inline __device__ double tanh(double x) { return ::tanh(x); };
+
+} // namespace math
+} // namespace ck
diff --git a/include/ck/utility/meta_data_buffer.hpp b/include/ck/utility/meta_data_buffer.hpp
new file mode 100644
index 0000000000..ccaea78627
--- /dev/null
+++ b/include/ck/utility/meta_data_buffer.hpp
@@ -0,0 +1,92 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+template <index_t MaxSize>
+struct MetaDataBuffer
+{
+    __host__ __device__ constexpr MetaDataBuffer() : buffer_{}, size_{0} {}
+
+    template <typename X, typename... Xs>
+    __host__ __device__ constexpr MetaDataBuffer(const X& x, const Xs&... xs) : buffer_{}, size_{0}
+    {
+        Push(x, xs...);
+    }
+
+    template <typename T>
+    __host__ __device__ constexpr void Push(const T& data)
+    {
+        if constexpr(!is_empty_v<T>)
+        {
+            constexpr index_t size = sizeof(T);
+
+            auto tmp = bit_cast<Array<std::byte, size>>(data);
+
+            for(int i = 0; i < size; i++)
+            {
+                buffer_(size_) = tmp[i];
+
+                size_++;
+            }
+        }
+    }
+
+    template <typename X, typename... Xs>
+    __host__ __device__ constexpr void Push(const X& x, const Xs&... xs)
+    {
+        Push(x);
+        Push(xs...);
+    }
+
+    template <typename T>
+    __host__ __device__ constexpr T Pop(index_t& pos) const
+    {
+        T data;
+
+        if constexpr(!is_empty_v<T>)
+        {
+            constexpr index_t size = sizeof(T);
+
+            Array<std::byte, size> tmp;
+
+            for(int i = 0; i < size; i++)
+            {
+                tmp(i) = buffer_[pos];
+
+                pos++;
+            }
+
+            data = bit_cast<T>(tmp);
+        }
+
+        return data;
+    }
+
+    template <typename T>
+    __host__ __device__ constexpr T Get(index_t pos) const
+    {
+        constexpr index_t size = sizeof(T);
+
+        Array<std::byte, size> tmp;
+
+        for(int i = 0; i < size; i++)
+        {
+            tmp(i) = buffer_[pos];
+
+            pos++;
+        }
+
+        auto data = bit_cast<T>(tmp);
+
+        return data;
+    }
+
+    //
+    Array<std::byte, MaxSize> buffer_;
+    index_t size_ = 0;
+};
+
+} // namespace ck
diff --git a/include/ck/utility/multi_index.hpp b/include/ck/utility/multi_index.hpp
new file mode 100644
index 0000000000..9f7ba8bff6
--- /dev/null
+++ b/include/ck/utility/multi_index.hpp
@@ -0,0 +1,12 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "common_header.hpp"
+
+#if CK_EXPERIMENTAL_USE_DYNAMICALLY_INDEXED_MULTI_INDEX
+#include "array_multi_index.hpp"
+#else
+#include "statically_indexed_array_multi_index.hpp"
+#endif
diff --git a/include/ck/utility/number.hpp b/include/ck/utility/number.hpp
new file mode 100644
index 0000000000..2fdef5dcd5
--- /dev/null
+++ b/include/ck/utility/number.hpp
@@ -0,0 +1,16 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/integral_constant.hpp"
+
+namespace ck {
+
+template <index_t N>
+using Number = integral_constant<index_t, N>;
+
+template <index_t N>
+using LongNumber = integral_constant<long_index_t, N>;
+
+} // namespace ck
diff --git a/include/ck/utility/print.hpp b/include/ck/utility/print.hpp
new file mode 100644
index 0000000000..3e6e41cbbd
--- /dev/null
+++ b/include/ck/utility/print.hpp
@@ -0,0 +1,44 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+template <typename T>
+__host__ __device__ inline void print(T t)
+{
+    t.Print();
+}
+
+template <>
+__host__ __device__ inline void print(bool v)
+{
+    printf("%d", static_cast<int32_t>(v));
+}
+
+template <>
+__host__ __device__ inline void print(int32_t v)
+{
+    printf("%d", v);
+}
+
+template <>
+__host__ __device__ inline void print(int64_t v)
+{
+    printf("%ld", v);
+}
+
+template <>
+__host__ __device__ inline void print(float v)
+{
+    printf("%f", v);
+}
+
+template <>
+__host__ __device__ inline void print(_Float16 v)
+{
+    printf("%f", static_cast<float>(v));
+}
+
+} // namespace ck
diff --git a/include/ck/utility/random_gen.hpp b/include/ck/utility/random_gen.hpp
new file mode 100644
index 0000000000..b7edf26507
--- /dev/null
+++ b/include/ck/utility/random_gen.hpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+// Pseudo random number generator
+// version for fp32
+template <typename T, uint32_t seed_t, std::enable_if_t<std::is_same<float, T>{}, bool> = false>
+__host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed = seed_t)
+{
+    uint32_t x         = *(reinterpret_cast<uint32_t*>(&val));
+    uint32_t drop_bits = uint32_t(x) & 0xFFFFu;
+    drop_bits ^= x >> 16;
+    drop_bits = ((drop_bits & 31) << 11) | (drop_bits >> 5);
+    drop_bits *= 0x7000149;
+    // NOTE: If id is in 64 bit, we are only using lower 32 bit.
+    //       So, it can have an effect of using same id for multiple elements when the id is very
+    //       large!
+    uint32_t rng = (drop_bits ^ 0x13371337 ^ (id * 229791) ^ seed);
+    return rng;
+}
+
+// version for fp16
+template <typename T, uint32_t seed_t, std::enable_if_t<std::is_same<half_t, T>{}, bool> = false>
+__host__ __device__ uint32_t prand_generator(index_t id, T val, uint32_t seed = seed_t)
+{
+    uint16_t x         = *(reinterpret_cast<uint16_t*>(&val));
+    uint32_t drop_bits = uint32_t(x) & 0xFFFFu;
+    drop_bits          = ((drop_bits & 31) << 11) | (drop_bits >> 5);
+    drop_bits *= 0x7000149;
+    // NOTE: If id is in 64 bit, we are only using lower 32 bit.
+    //       So, it can have an effect of using same id for multiple elements when the id is very
+    //       large!
+    uint32_t rng = (drop_bits ^ 0x13371337 ^ (id * 229791) ^ seed);
+    return rng;
+}
+
+// return 0 if data is not fp16 or fp32
+template <typename T,
+          uint32_t seed_t,
+          std::enable_if_t<!(std::is_same<float, T>{} || std::is_same<half_t, T>{}), bool> = false>
+__host__ __device__ uint32_t prand_generator(int id, T val, uint32_t seed = seed_t)
+{
+    std::ignore = id;
+    std::ignore = val;
+    std::ignore = seed;
+
+    return 0;
+}
+
+} // namespace ck
diff --git a/include/ck/utility/reduction_common.hpp b/include/ck/utility/reduction_common.hpp
new file mode 100644
index 0000000000..75fdd85825
--- /dev/null
+++ b/include/ck/utility/reduction_common.hpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_enums.hpp"
+
+namespace ck {
+
+struct float_equal_one
+{
+    template <class T>
+    __host__ __device__ inline bool operator()(T x)
+    {
+        return x <= static_cast<T>(1.0f) and x >= static_cast<T>(1.0f);
+    };
+};
+
+struct float_equal_zero
+{
+    template <class T>
+    __host__ __device__ inline bool operator()(T x)
+    {
+        return x <= static_cast<T>(0.0f) and x >= static_cast<T>(0.0f);
+    };
+};
+
+} // namespace ck
diff --git a/include/ck/utility/reduction_enums.hpp b/include/ck/utility/reduction_enums.hpp
new file mode 100644
index 0000000000..23b7149f8e
--- /dev/null
+++ b/include/ck/utility/reduction_enums.hpp
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+enum struct ReduceTensorOp
+{
+    ADD   = 0,
+    MUL   = 1,
+    MIN   = 2,
+    MAX   = 3,
+    AMAX  = 4,
+    AVG   = 5,
+    NORM1 = 6,
+    NORM2 = 7,
+    // MUL_NO_ZEROS = 8,
+};
+
+enum struct NanPropagation
+{
+    NOT_PROPAGATE_NAN = 0,
+    PROPAGATE_NAN     = 1,
+};
+
+enum struct ReduceTensorIndices
+{
+    NO_INDICES        = 0,
+    FLATTENED_INDICES = 1,
+};
+
+enum struct IndicesType
+{
+    INDICES_32BIT = 0,
+    INDICES_64BIT = 1,
+    INDICES_16BIT = 2,
+    INDICES_8BIT  = 3,
+};
+
+} // namespace ck
diff --git a/include/ck/utility/reduction_functions_accumulate.hpp b/include/ck/utility/reduction_functions_accumulate.hpp
new file mode 100644
index 0000000000..b9765ff0d2
--- /dev/null
+++ b/include/ck/utility/reduction_functions_accumulate.hpp
@@ -0,0 +1,115 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math_v2.hpp"
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_operator.hpp"
+
+namespace ck {
+namespace detail {
+
+// Check for NaN; guarantee NaNs are NOT propagated to result (i.e., ignore NaNs)
+template <typename ReduceOperation, typename AccDataType>
+struct AccumulateWithNanIgnore
+{
+    __device__ static inline void Calculate(AccDataType& accuVal, AccDataType currVal)
+    {
+        if(!ck::math::isnan(currVal))
+        {
+            ReduceOperation{}(accuVal, currVal);
+        }
+    };
+};
+
+template <bool PropagateNan, typename ReduceOperation, typename AccDataType>
+struct AccumulateWithNanCheck;
+
+// Does not check for NaN; does not guarantee NaNs be propagated to result
+// e.g., given that max(a, b) = a > b ? a : b
+// then  max(NaN, 1) returns 1
+//       max(1, NaN) returns NaN
+// since any comparison involving NaNs returns false
+template <typename ReduceOperation, typename AccDataType>
+struct AccumulateWithNanCheck<false, ReduceOperation, AccDataType>
+{
+    // cppcheck-suppress constParameter
+    __host__ __device__ static inline void Calculate(AccDataType& accuVal, AccDataType currVal)
+    {
+        ReduceOperation{}(accuVal, currVal);
+    };
+};
+
+// Check for NaN; guarantees NaNs be propagated to result
+template <typename ReduceOperation, typename AccDataType>
+struct AccumulateWithNanCheck<true, ReduceOperation, AccDataType>
+{
+    __host__ __device__ static inline void Calculate(AccDataType& accuVal, AccDataType currVal)
+    {
+        using ck::math::isnan;
+
+        if(isnan(currVal))
+        {
+            accuVal = currVal;
+        }
+        else
+        {
+            ReduceOperation{}(accuVal, currVal);
+        };
+    };
+};
+
+template <bool PropagateNan, typename ReduceOperation, typename AccDataType, typename IndexDataType>
+struct AccumulateWithIndexAndNanCheck;
+
+template <typename ReduceOperation, typename AccDataType, typename IndexDataType>
+struct AccumulateWithIndexAndNanCheck<false, ReduceOperation, AccDataType, IndexDataType>
+{
+    __host__ __device__ static inline void
+    // cppcheck-suppress constParameter
+    Calculate(AccDataType& accuVal,
+              AccDataType currVal,
+              IndexDataType& accuIndex,
+              IndexDataType currIndex)
+    {
+        bool changed = false;
+
+        ReduceOperation{}(accuVal, currVal, changed);
+
+        if(changed)
+            accuIndex = currIndex;
+    };
+};
+
+template <typename ReduceOperation, typename AccDataType, typename IndexDataType>
+struct AccumulateWithIndexAndNanCheck<true, ReduceOperation, AccDataType, IndexDataType>
+{
+    // The method is called when the ReduceOperation is indexable and the user asked for indices
+    __host__ __device__ static inline void Calculate(AccDataType& accuVal,
+                                                     AccDataType currVal,
+                                                     IndexDataType& accuIndex,
+                                                     IndexDataType currIndex)
+    {
+        using ck::math::isnan;
+
+        if(isnan(currVal))
+        {
+            accuVal   = currVal;
+            accuIndex = currIndex;
+        }
+        else
+        {
+            bool changed = false;
+
+            ReduceOperation{}(accuVal, currVal, changed);
+
+            if(changed)
+                accuIndex = currIndex;
+        }
+    };
+};
+
+} // namespace detail
+} // namespace ck
diff --git a/include/ck/utility/reduction_operator.hpp b/include/ck/utility/reduction_operator.hpp
new file mode 100644
index 0000000000..5480a98409
--- /dev/null
+++ b/include/ck/utility/reduction_operator.hpp
@@ -0,0 +1,352 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/type.hpp"
+#include "ck/utility/type_convert.hpp"
+
+namespace ck {
+
+namespace reduce {
+
+// Every binary operator used in reduction is represented by a templated functor class. Each functor
+// class must provide at least
+// three members:
+// 1) GetIdentityValue() -- the interface to return the "identity element" for the binary
+// operator, "identity element" is the unique
+//                    element in the algebraic space that doesn't affect the value of other elements
+//                    when operated against them, and the concept is similar to zero vector in
+//                    vector space
+//                    (http://pages.cs.wisc.edu/~matthewb/pages/notes/pdf/linearalgebra/VectorSpaces.pdf).
+// 2) IsCompatibleInMemoryDataOperation() -- return true if the reduction task corresponding to this
+// operator can use the InMemoryDataOperation to finalize, or else it return false
+// 3) operator() -- the first argument of the operator must be both an input & output, and the
+//                  corresponding variable usually stores
+//                  the accumulated result of many operator() calls; the second argument is only an
+//                  input. For indexable binary
+//                  operator, the second version of operator() has third argument (which is an
+//                  output) to indicate whether the
+//                  accumulated value (the first argument) has changed, in which case the recorded
+//                  accumulated index also need be
+//                  changed.
+
+struct Add
+{
+    template <typename T>
+    __host__ __device__ static constexpr T GetIdentityValue()
+    {
+        return type_convert<T>(0.0f);
+    };
+
+    __host__ __device__ static constexpr bool
+    IsCompatibleInMemoryDataOperation(InMemoryDataOperationEnum operation)
+    {
+        return operation == InMemoryDataOperationEnum::AtomicAdd ||
+               operation == InMemoryDataOperationEnum::Set;
+    };
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, int32_t>::value,
+                      "The data type is not supported by the Add accumulator!");
+
+        a = a + b;
+    }
+};
+
+struct SquaredAdd
+{
+    template <class T>
+    __host__ __device__ static constexpr T GetIdentityValue()
+    {
+        return type_convert<T>(0.0f);
+    };
+
+    __host__ __device__ static constexpr bool
+    IsCompatibleInMemoryDataOperation(InMemoryDataOperationEnum operation)
+    {
+        return operation == InMemoryDataOperationEnum::AtomicAdd ||
+               operation == InMemoryDataOperationEnum::Set;
+    };
+
+    template <class T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "The data type is not supported by the SquaredAdd accumulator!");
+
+        a = a + b * b;
+    }
+};
+
+struct Mul
+{
+    template <typename T>
+    __host__ __device__ static constexpr T GetIdentityValue()
+    {
+        return type_convert<T>(1.0f);
+    };
+
+    __host__ __device__ static constexpr bool
+    IsCompatibleInMemoryDataOperation(InMemoryDataOperationEnum operation)
+    {
+        return operation == InMemoryDataOperationEnum::Set;
+    };
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, int32_t>::value,
+                      "The data type is not supported by the Mul accumulator!");
+
+        a = a * b;
+    }
+};
+
+struct Max
+{
+    template <typename T>
+    __host__ __device__ static constexpr T GetIdentityValue()
+    {
+        if constexpr(is_same_v<T, bhalf_t>)
+        {
+            float val = NumericLimits<float>::Lowest();
+            return type_convert<bhalf_t>(val);
+        }
+        else
+        {
+            return NumericLimits<T>::Lowest();
+        }
+    };
+
+    __host__ __device__ static constexpr bool
+    IsCompatibleInMemoryDataOperation(InMemoryDataOperationEnum operation)
+    {
+        // ToChange: atomic_max to be added
+        return operation == InMemoryDataOperationEnum::Set;
+    };
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "The data type is not supported by the Max accumulator!");
+
+        if(a < b)
+            a = b;
+    }
+
+    __host__ __device__ inline constexpr void operator()(bhalf_t& a, bhalf_t b) const
+    {
+        float a_ = type_convert<float>(a);
+        float b_ = type_convert<float>(b);
+
+        if(a_ < b_)
+            a = b;
+    }
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b, bool& changed) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "The data type is not supported by the Max accumulator!");
+
+        if(a < b)
+        {
+            a       = b;
+            changed = true;
+        }
+    }
+
+    __host__ __device__ inline constexpr void operator()(bhalf_t& a, bhalf_t b, bool& changed) const
+    {
+        float a_ = type_convert<float>(a);
+        float b_ = type_convert<float>(b);
+
+        if(a_ < b_)
+        {
+            a       = b;
+            changed = true;
+        }
+    }
+};
+
+struct Min
+{
+    template <typename T>
+    __host__ __device__ static constexpr T GetIdentityValue()
+    {
+        if constexpr(is_same_v<T, bhalf_t>)
+        {
+            float val = NumericLimits<float>::Max();
+            return type_convert<bhalf_t>(val);
+        }
+        else
+        {
+            return NumericLimits<T>::Max();
+        }
+        return NumericLimits<T>::Max();
+    };
+
+    __host__ __device__ static constexpr bool
+    IsCompatibleInMemoryDataOperation(InMemoryDataOperationEnum operation)
+    {
+        // ToChange: atomic_min to be added
+        return operation == InMemoryDataOperationEnum::Set;
+    };
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "The data type is not supported by the Min accumulator!");
+
+        if(a > b)
+            a = b;
+    }
+
+    __host__ __device__ inline constexpr void operator()(bhalf_t& a, bhalf_t b) const
+    {
+        float a_ = type_convert<float>(a);
+        float b_ = type_convert<float>(b);
+
+        if(a_ > b_)
+            a = b;
+    }
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b, bool& changed) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "The data type is not supported by the Min accumulator!");
+
+        if(a > b)
+        {
+            a       = b;
+            changed = true;
+        }
+    }
+
+    __host__ __device__ inline constexpr void operator()(bhalf_t& a, bhalf_t b, bool& changed) const
+    {
+        float a_ = type_convert<float>(a);
+        float b_ = type_convert<float>(b);
+
+        if(a_ > b_)
+        {
+            a       = b;
+            changed = true;
+        }
+    }
+};
+
+struct AMax
+{
+    template <typename T>
+    __host__ __device__ static constexpr T GetIdentityValue()
+    {
+        return type_convert<T>(0.0f);
+    };
+
+    __host__ __device__ static constexpr bool
+    IsCompatibleInMemoryDataOperation(InMemoryDataOperationEnum operation)
+    {
+        // ToChange: atomic_max to be added
+        return operation == InMemoryDataOperationEnum::Set;
+    };
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "The data type is not supported by the AMax accumulator!");
+
+        if(a < b)
+            a = b;
+    }
+
+    template <typename T>
+    __host__ __device__ inline constexpr void operator()(T& a, T b, bool& changed) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "The data type is not supported by the AMax accumulator!");
+
+        if(a < b)
+        {
+            a       = b;
+            changed = true;
+        }
+    }
+};
+
+template <typename T>
+constexpr T GetIdentityValueForInMemoryDataOperation(InMemoryDataOperationEnum operation)
+{
+    T result = ck::type_convert<T>(0.0f);
+
+    if(operation == InMemoryDataOperationEnum::AtomicMax)
+        result = ck::NumericLimits<T>::Lowest();
+
+    return (result);
+};
+
+template <InMemoryDataOperationEnum Operation, typename DataType>
+struct InMemoryDataOperationSupportedOnDataType
+{
+    static constexpr bool value = false;
+};
+
+template <typename DataType>
+struct InMemoryDataOperationSupportedOnDataType<InMemoryDataOperationEnum::AtomicAdd, DataType>
+{
+    static constexpr bool value =
+        is_same<DataType, float>::value || is_same<DataType, double>::value;
+};
+
+template <typename DataType>
+struct InMemoryDataOperationSupportedOnDataType<InMemoryDataOperationEnum::AtomicMax, DataType>
+{
+    static constexpr bool value =
+        is_same<DataType, float>::value || is_same<DataType, double>::value;
+};
+
+template <typename DataType>
+struct InMemoryDataOperationSupportedOnDataType<InMemoryDataOperationEnum::Set, DataType>
+{
+    static constexpr bool value =
+        is_same<DataType, float>::value || is_same<DataType, double>::value ||
+        is_same<DataType, half_t>::value || is_same<DataType, bhalf_t>::value ||
+        is_same<DataType, int8_t>::value || is_same<DataType, int32_t>::value;
+};
+
+template <typename DataType>
+struct InMemoryDataOperationSupportedOnDataType<InMemoryDataOperationEnum::Add, DataType>
+{
+    static constexpr bool value =
+        is_same<DataType, float>::value || is_same<DataType, double>::value ||
+        is_same<DataType, half_t>::value || is_same<DataType, int8_t>::value ||
+        is_same<DataType, int32_t>::value;
+};
+
+} // namespace reduce
+} // namespace ck
diff --git a/include/ck/utility/remove_cvref.hpp b/include/ck/utility/remove_cvref.hpp
new file mode 100644
index 0000000000..155c42dc53
--- /dev/null
+++ b/include/ck/utility/remove_cvref.hpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+
+namespace ck {
+
+// remove_cvref_t
+template <typename T>
+using remove_reference_t = typename std::remove_reference<T>::type;
+
+template <typename T>
+using remove_cv_t = typename std::remove_cv<T>::type;
+
+template <typename T>
+using remove_cvref_t = remove_cv_t<std::remove_reference_t<T>>;
+
+template <typename T>
+using remove_pointer_t = typename std::remove_pointer<T>::type;
+
+} // namespace ck
diff --git a/include/ck/utility/sequence.hpp b/include/ck/utility/sequence.hpp
new file mode 100644
index 0000000000..ec6745a011
--- /dev/null
+++ b/include/ck/utility/sequence.hpp
@@ -0,0 +1,965 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/integral_constant.hpp"
+#include "ck/utility/type.hpp"
+#include "ck/utility/functional.hpp"
+#include "ck/utility/math.hpp"
+
+namespace ck {
+
+template <index_t, index_t, index_t>
+struct static_for;
+
+template <index_t...>
+struct Sequence;
+
+template <typename Seq, index_t I>
+struct sequence_split;
+
+template <typename>
+struct sequence_reverse;
+
+template <typename>
+struct sequence_map_inverse;
+
+template <typename>
+struct is_valid_sequence_map;
+
+template <index_t I, index_t... Is>
+__host__ __device__ constexpr auto sequence_pop_front(Sequence<I, Is...>);
+
+template <typename Seq>
+__host__ __device__ constexpr auto sequence_pop_back(Seq);
+
+template <index_t... Is>
+struct Sequence
+{
+    using Type      = Sequence;
+    using data_type = index_t;
+
+    static constexpr index_t mSize = sizeof...(Is);
+
+    __host__ __device__ static constexpr index_t Size() { return mSize; }
+
+    // TODO: deprecate
+    __host__ __device__ static constexpr auto GetSize() { return Size(); }
+
+    __host__ __device__ static constexpr index_t At(index_t I)
+    {
+        // the last dummy element is to prevent compiler complain about empty array, when mSize = 0
+        const index_t mData[mSize + 1] = {Is..., 0};
+        return mData[I];
+    }
+
+    template <index_t I>
+    __host__ __device__ static constexpr auto At(Number<I>)
+    {
+        static_assert(I < mSize, "wrong! I too large");
+
+        return Number<At(I)>{};
+    }
+
+    template <index_t I>
+    __host__ __device__ static constexpr auto At()
+    {
+        static_assert(I < mSize, "wrong! I too large");
+
+        return Number<At(I)>{};
+    }
+
+    template <index_t I>
+    __host__ __device__ static constexpr auto Get(Number<I>)
+    {
+        return At(I);
+    }
+
+    template <typename I>
+    __host__ __device__ constexpr auto operator[](I i) const
+    {
+        return At(i);
+    }
+
+    template <index_t... IRs>
+    __host__ __device__ static constexpr auto ReorderGivenNew2Old(Sequence<IRs...> /*new2old*/)
+    {
+        static_assert(sizeof...(Is) == sizeof...(IRs),
+                      "wrong! reorder map should have the same size as Sequence to be rerodered");
+
+        static_assert(is_valid_sequence_map<Sequence<IRs...>>::value, "wrong! invalid reorder map");
+
+        return Sequence<Type::At(Number<IRs>{})...>{};
+    }
+
+    // MapOld2New is Sequence<...>
+    template <typename MapOld2New>
+    __host__ __device__ static constexpr auto ReorderGivenOld2New(MapOld2New)
+    {
+        static_assert(MapOld2New::Size() == Size(),
+                      "wrong! reorder map should have the same size as Sequence to be rerodered");
+
+        static_assert(is_valid_sequence_map<MapOld2New>::value, "wrong! invalid reorder map");
+
+        return ReorderGivenNew2Old(typename sequence_map_inverse<MapOld2New>::type{});
+    }
+
+    __host__ __device__ static constexpr auto Reverse()
+    {
+        return typename sequence_reverse<Type>::type{};
+    }
+
+    __host__ __device__ static constexpr auto Front()
+    {
+        static_assert(mSize > 0, "wrong!");
+        return At(Number<0>{});
+    }
+
+    __host__ __device__ static constexpr auto Back()
+    {
+        static_assert(mSize > 0, "wrong!");
+        return At(Number<mSize - 1>{});
+    }
+
+    __host__ __device__ static constexpr auto PopFront() { return sequence_pop_front(Type{}); }
+
+    __host__ __device__ static constexpr auto PopBack() { return sequence_pop_back(Type{}); }
+
+    template <index_t... Xs>
+    __host__ __device__ static constexpr auto PushFront(Sequence<Xs...>)
+    {
+        return Sequence<Xs..., Is...>{};
+    }
+
+    template <index_t... Xs>
+    __host__ __device__ static constexpr auto PushFront(Number<Xs>...)
+    {
+        return Sequence<Xs..., Is...>{};
+    }
+
+    template <index_t... Xs>
+    __host__ __device__ static constexpr auto PushBack(Sequence<Xs...>)
+    {
+        return Sequence<Is..., Xs...>{};
+    }
+
+    template <index_t... Xs>
+    __host__ __device__ static constexpr auto PushBack(Number<Xs>...)
+    {
+        return Sequence<Is..., Xs...>{};
+    }
+
+    template <index_t... Ns>
+    __host__ __device__ static constexpr auto Extract(Number<Ns>...)
+    {
+        return Sequence<Type::At(Number<Ns>{})...>{};
+    }
+
+    template <index_t... Ns>
+    __host__ __device__ static constexpr auto Extract(Sequence<Ns...>)
+    {
+        return Sequence<Type::At(Number<Ns>{})...>{};
+    }
+
+    template <index_t I, index_t X>
+    __host__ __device__ static constexpr auto Modify(Number<I>, Number<X>)
+    {
+        static_assert(I < Size(), "wrong!");
+
+        using seq_split          = sequence_split<Type, I>;
+        constexpr auto seq_left  = typename seq_split::left_type{};
+        constexpr auto seq_right = typename seq_split::right_type{}.PopFront();
+
+        return seq_left.PushBack(Number<X>{}).PushBack(seq_right);
+    }
+
+    template <typename F>
+    __host__ __device__ static constexpr auto Transform(F f)
+    {
+        return Sequence<f(Is)...>{};
+    }
+
+    __host__ __device__ static constexpr bool IsStatic() { return true; };
+
+    __host__ __device__ static void Print()
+    {
+        printf("Sequence{size: %d, data: [", Size());
+
+        for(index_t i = 0; i < Size(); i++)
+        {
+            print(At(i));
+
+            if(i < Size() - 1)
+            {
+                printf(", ");
+            }
+        }
+
+        printf("]}");
+    }
+};
+
+// merge sequence
+template <typename Seq, typename... Seqs>
+struct sequence_merge
+{
+    using type = typename sequence_merge<Seq, typename sequence_merge<Seqs...>::type>::type;
+};
+
+template <index_t... Xs, index_t... Ys>
+struct sequence_merge<Sequence<Xs...>, Sequence<Ys...>>
+{
+    using type = Sequence<Xs..., Ys...>;
+};
+
+template <typename Seq>
+struct sequence_merge<Seq>
+{
+    using type = Seq;
+};
+
+// generate sequence
+template <index_t NSize, typename F>
+struct sequence_gen
+{
+    template <index_t IBegin, index_t NRemain, typename G>
+    struct sequence_gen_impl
+    {
+        static constexpr index_t NRemainLeft  = NRemain / 2;
+        static constexpr index_t NRemainRight = NRemain - NRemainLeft;
+        static constexpr index_t IMiddle      = IBegin + NRemainLeft;
+
+        using type = typename sequence_merge<
+            typename sequence_gen_impl<IBegin, NRemainLeft, G>::type,
+            typename sequence_gen_impl<IMiddle, NRemainRight, G>::type>::type;
+    };
+
+    template <index_t I, typename G>
+    struct sequence_gen_impl<I, 1, G>
+    {
+        static constexpr index_t Is = G{}(Number<I>{});
+        using type                  = Sequence<Is>;
+    };
+
+    template <index_t I, typename G>
+    struct sequence_gen_impl<I, 0, G>
+    {
+        using type = Sequence<>;
+    };
+
+    using type = typename sequence_gen_impl<0, NSize, F>::type;
+};
+
+// arithmetic sequence
+template <index_t IBegin, index_t IEnd, index_t Increment>
+struct arithmetic_sequence_gen
+{
+    struct F
+    {
+        __host__ __device__ constexpr index_t operator()(index_t i) const
+        {
+            return i * Increment + IBegin;
+        }
+    };
+
+    using type0 = typename sequence_gen<(IEnd - IBegin) / Increment, F>::type;
+    using type1 = Sequence<>;
+
+    static constexpr bool kHasContent =
+        (Increment > 0 && IBegin < IEnd) || (Increment < 0 && IBegin > IEnd);
+
+    using type = typename conditional<kHasContent, type0, type1>::type;
+};
+
+// uniform sequence
+template <index_t NSize, index_t I>
+struct uniform_sequence_gen
+{
+    struct F
+    {
+        __host__ __device__ constexpr index_t operator()(index_t) const { return I; }
+    };
+
+    using type = typename sequence_gen<NSize, F>::type;
+};
+
+// reverse inclusive scan (with init) sequence
+template <typename, typename, index_t>
+struct sequence_reverse_inclusive_scan;
+
+template <index_t I, index_t... Is, typename Reduce, index_t Init>
+struct sequence_reverse_inclusive_scan<Sequence<I, Is...>, Reduce, Init>
+{
+    using old_scan = typename sequence_reverse_inclusive_scan<Sequence<Is...>, Reduce, Init>::type;
+
+    static constexpr index_t new_reduce = Reduce{}(I, old_scan{}.Front());
+
+    using type = typename sequence_merge<Sequence<new_reduce>, old_scan>::type;
+};
+
+template <index_t I, typename Reduce, index_t Init>
+struct sequence_reverse_inclusive_scan<Sequence<I>, Reduce, Init>
+{
+    using type = Sequence<Reduce{}(I, Init)>;
+};
+
+template <typename Reduce, index_t Init>
+struct sequence_reverse_inclusive_scan<Sequence<>, Reduce, Init>
+{
+    using type = Sequence<>;
+};
+
+// split sequence
+template <typename Seq, index_t I>
+struct sequence_split
+{
+    static constexpr index_t NSize = Seq{}.Size();
+
+    using range0 = typename arithmetic_sequence_gen<0, I, 1>::type;
+    using range1 = typename arithmetic_sequence_gen<I, NSize, 1>::type;
+
+    using left_type  = decltype(Seq::Extract(range0{}));
+    using right_type = decltype(Seq::Extract(range1{}));
+};
+
+// reverse sequence
+template <typename Seq>
+struct sequence_reverse
+{
+    static constexpr index_t NSize = Seq{}.Size();
+
+    using seq_split = sequence_split<Seq, NSize / 2>;
+    using type      = typename sequence_merge<
+        typename sequence_reverse<typename seq_split::right_type>::type,
+        typename sequence_reverse<typename seq_split::left_type>::type>::type;
+};
+
+template <index_t I>
+struct sequence_reverse<Sequence<I>>
+{
+    using type = Sequence<I>;
+};
+
+template <index_t I0, index_t I1>
+struct sequence_reverse<Sequence<I0, I1>>
+{
+    using type = Sequence<I1, I0>;
+};
+
+#if 1
+template <typename Reduce, typename Seq, typename... Seqs>
+struct sequence_reduce
+{
+    using type = typename sequence_reduce<Reduce,
+                                          Seq,
+                                          typename sequence_reduce<Reduce, Seqs...>::type>::type;
+};
+
+template <typename Reduce, index_t... Xs, index_t... Ys>
+struct sequence_reduce<Reduce, Sequence<Xs...>, Sequence<Ys...>>
+{
+    using type = Sequence<Reduce{}(Xs, Ys)...>;
+};
+
+template <typename Reduce, typename Seq>
+struct sequence_reduce<Reduce, Seq>
+{
+    using type = Seq;
+};
+#endif
+
+template <typename Values, typename Ids, typename Compare>
+struct sequence_sort_impl
+{
+    template <typename LeftValues,
+              typename LeftIds,
+              typename RightValues,
+              typename RightIds,
+              typename MergedValues,
+              typename MergedIds,
+              typename Comp>
+    struct sorted_sequence_merge_impl
+    {
+        static constexpr bool choose_left = LeftValues::Front() < RightValues::Front();
+
+        static constexpr index_t chosen_value =
+            choose_left ? LeftValues::Front() : RightValues::Front();
+        static constexpr index_t chosen_id = choose_left ? LeftIds::Front() : RightIds::Front();
+
+        using new_merged_values = decltype(MergedValues::PushBack(Number<chosen_value>{}));
+        using new_merged_ids    = decltype(MergedIds::PushBack(Number<chosen_id>{}));
+
+        using new_left_values =
+            typename conditional<choose_left, decltype(LeftValues::PopFront()), LeftValues>::type;
+        using new_left_ids =
+            typename conditional<choose_left, decltype(LeftIds::PopFront()), LeftIds>::type;
+
+        using new_right_values =
+            typename conditional<choose_left, RightValues, decltype(RightValues::PopFront())>::type;
+        using new_right_ids =
+            typename conditional<choose_left, RightIds, decltype(RightIds::PopFront())>::type;
+
+        using merge = sorted_sequence_merge_impl<new_left_values,
+                                                 new_left_ids,
+                                                 new_right_values,
+                                                 new_right_ids,
+                                                 new_merged_values,
+                                                 new_merged_ids,
+                                                 Comp>;
+        // this is output
+        using merged_values = typename merge::merged_values;
+        using merged_ids    = typename merge::merged_ids;
+    };
+
+    template <typename LeftValues,
+              typename LeftIds,
+              typename MergedValues,
+              typename MergedIds,
+              typename Comp>
+    struct sorted_sequence_merge_impl<LeftValues,
+                                      LeftIds,
+                                      Sequence<>,
+                                      Sequence<>,
+                                      MergedValues,
+                                      MergedIds,
+                                      Comp>
+    {
+        using merged_values = typename sequence_merge<MergedValues, LeftValues>::type;
+        using merged_ids    = typename sequence_merge<MergedIds, LeftIds>::type;
+    };
+
+    template <typename RightValues,
+              typename RightIds,
+              typename MergedValues,
+              typename MergedIds,
+              typename Comp>
+    struct sorted_sequence_merge_impl<Sequence<>,
+                                      Sequence<>,
+                                      RightValues,
+                                      RightIds,
+                                      MergedValues,
+                                      MergedIds,
+                                      Comp>
+    {
+        using merged_values = typename sequence_merge<MergedValues, RightValues>::type;
+        using merged_ids    = typename sequence_merge<MergedIds, RightIds>::type;
+    };
+
+    template <typename LeftValues,
+              typename LeftIds,
+              typename RightValues,
+              typename RightIds,
+              typename Comp>
+    struct sorted_sequence_merge
+    {
+        using merge = sorted_sequence_merge_impl<LeftValues,
+                                                 LeftIds,
+                                                 RightValues,
+                                                 RightIds,
+                                                 Sequence<>,
+                                                 Sequence<>,
+                                                 Comp>;
+
+        using merged_values = typename merge::merged_values;
+        using merged_ids    = typename merge::merged_ids;
+    };
+
+    static constexpr index_t nsize = Values::Size();
+
+    using split_unsorted_values = sequence_split<Values, nsize / 2>;
+    using split_unsorted_ids    = sequence_split<Ids, nsize / 2>;
+
+    using left_unsorted_values = typename split_unsorted_values::left_type;
+    using left_unsorted_ids    = typename split_unsorted_ids::left_type;
+    using left_sort          = sequence_sort_impl<left_unsorted_values, left_unsorted_ids, Compare>;
+    using left_sorted_values = typename left_sort::sorted_values;
+    using left_sorted_ids    = typename left_sort::sorted_ids;
+
+    using right_unsorted_values = typename split_unsorted_values::right_type;
+    using right_unsorted_ids    = typename split_unsorted_ids::right_type;
+    using right_sort = sequence_sort_impl<right_unsorted_values, right_unsorted_ids, Compare>;
+    using right_sorted_values = typename right_sort::sorted_values;
+    using right_sorted_ids    = typename right_sort::sorted_ids;
+
+    using merged_sorted = sorted_sequence_merge<left_sorted_values,
+                                                left_sorted_ids,
+                                                right_sorted_values,
+                                                right_sorted_ids,
+                                                Compare>;
+
+    using sorted_values = typename merged_sorted::merged_values;
+    using sorted_ids    = typename merged_sorted::merged_ids;
+};
+
+template <index_t ValueX, index_t ValueY, index_t IdX, index_t IdY, typename Compare>
+struct sequence_sort_impl<Sequence<ValueX, ValueY>, Sequence<IdX, IdY>, Compare>
+{
+    static constexpr bool choose_x = Compare{}(ValueX, ValueY);
+
+    using sorted_values =
+        typename conditional<choose_x, Sequence<ValueX, ValueY>, Sequence<ValueY, ValueX>>::type;
+    using sorted_ids = typename conditional<choose_x, Sequence<IdX, IdY>, Sequence<IdY, IdX>>::type;
+};
+
+template <index_t Value, index_t Id, typename Compare>
+struct sequence_sort_impl<Sequence<Value>, Sequence<Id>, Compare>
+{
+    using sorted_values = Sequence<Value>;
+    using sorted_ids    = Sequence<Id>;
+};
+
+template <typename Compare>
+struct sequence_sort_impl<Sequence<>, Sequence<>, Compare>
+{
+    using sorted_values = Sequence<>;
+    using sorted_ids    = Sequence<>;
+};
+
+template <typename Values, typename Compare>
+struct sequence_sort
+{
+    using unsorted_ids = typename arithmetic_sequence_gen<0, Values::Size(), 1>::type;
+    using sort         = sequence_sort_impl<Values, unsorted_ids, Compare>;
+
+    // this is output
+    using type                = typename sort::sorted_values;
+    using sorted2unsorted_map = typename sort::sorted_ids;
+};
+
+template <typename Values, typename Less, typename Equal>
+struct sequence_unique_sort
+{
+    template <typename RemainValues,
+              typename RemainIds,
+              typename UniquifiedValues,
+              typename UniquifiedIds,
+              typename Eq>
+    struct sorted_sequence_uniquify_impl
+    {
+        static constexpr index_t current_value = RemainValues::Front();
+        static constexpr index_t current_id    = RemainIds::Front();
+
+        static constexpr bool is_unique_value = (current_value != UniquifiedValues::Back());
+
+        using new_remain_values = decltype(RemainValues::PopFront());
+        using new_remain_ids    = decltype(RemainIds::PopFront());
+
+        using new_uniquified_values =
+            typename conditional<is_unique_value,
+                                 decltype(UniquifiedValues::PushBack(Number<current_value>{})),
+                                 UniquifiedValues>::type;
+
+        using new_uniquified_ids =
+            typename conditional<is_unique_value,
+                                 decltype(UniquifiedIds::PushBack(Number<current_id>{})),
+                                 UniquifiedIds>::type;
+
+        using uniquify = sorted_sequence_uniquify_impl<new_remain_values,
+                                                       new_remain_ids,
+                                                       new_uniquified_values,
+                                                       new_uniquified_ids,
+                                                       Eq>;
+
+        // this is output
+        using uniquified_values = typename uniquify::uniquified_values;
+        using uniquified_ids    = typename uniquify::uniquified_ids;
+    };
+
+    template <typename UniquifiedValues, typename UniquifiedIds, typename Eq>
+    struct sorted_sequence_uniquify_impl<Sequence<>,
+                                         Sequence<>,
+                                         UniquifiedValues,
+                                         UniquifiedIds,
+                                         Eq>
+    {
+        using uniquified_values = UniquifiedValues;
+        using uniquified_ids    = UniquifiedIds;
+    };
+
+    template <typename SortedValues, typename SortedIds, typename Eq>
+    struct sorted_sequence_uniquify
+    {
+        using uniquify = sorted_sequence_uniquify_impl<decltype(SortedValues::PopFront()),
+                                                       decltype(SortedIds::PopFront()),
+                                                       Sequence<SortedValues::Front()>,
+                                                       Sequence<SortedIds::Front()>,
+                                                       Eq>;
+
+        using uniquified_values = typename uniquify::uniquified_values;
+        using uniquified_ids    = typename uniquify::uniquified_ids;
+    };
+
+    using sort          = sequence_sort<Values, Less>;
+    using sorted_values = typename sort::type;
+    using sorted_ids    = typename sort::sorted2unsorted_map;
+
+    using uniquify = sorted_sequence_uniquify<sorted_values, sorted_ids, Equal>;
+
+    // this is output
+    using type                = typename uniquify::uniquified_values;
+    using sorted2unsorted_map = typename uniquify::uniquified_ids;
+};
+
+template <typename SeqMap>
+struct is_valid_sequence_map : is_same<typename arithmetic_sequence_gen<0, SeqMap::Size(), 1>::type,
+                                       typename sequence_sort<SeqMap, math::less<index_t>>::type>
+{
+};
+
+template <typename SeqMap>
+struct sequence_map_inverse
+{
+    template <typename X2Y, typename WorkingY2X, index_t XBegin, index_t XRemain>
+    struct sequence_map_inverse_impl
+    {
+        static constexpr auto new_y2x =
+            WorkingY2X::Modify(X2Y::At(Number<XBegin>{}), Number<XBegin>{});
+
+        using type =
+            typename sequence_map_inverse_impl<X2Y, decltype(new_y2x), XBegin + 1, XRemain - 1>::
+                type;
+    };
+
+    template <typename X2Y, typename WorkingY2X, index_t XBegin>
+    struct sequence_map_inverse_impl<X2Y, WorkingY2X, XBegin, 0>
+    {
+        using type = WorkingY2X;
+    };
+
+    using type =
+        typename sequence_map_inverse_impl<SeqMap,
+                                           typename uniform_sequence_gen<SeqMap::Size(), 0>::type,
+                                           0,
+                                           SeqMap::Size()>::type;
+};
+
+template <index_t... Xs, index_t... Ys>
+__host__ __device__ constexpr bool operator==(Sequence<Xs...>, Sequence<Ys...>)
+{
+    return ((Xs == Ys) && ...);
+}
+
+template <index_t... Xs, index_t... Ys>
+__host__ __device__ constexpr bool operator!=(Sequence<Xs...> x, Sequence<Ys...> y)
+{
+    return !(x == y);
+}
+
+template <index_t... Xs, index_t... Ys>
+__host__ __device__ constexpr auto operator+(Sequence<Xs...>, Sequence<Ys...>)
+{
+    static_assert(sizeof...(Xs) == sizeof...(Ys), "wrong! inconsistent size");
+
+    return Sequence<(Xs + Ys)...>{};
+}
+
+template <index_t... Xs, index_t... Ys>
+__host__ __device__ constexpr auto operator-(Sequence<Xs...>, Sequence<Ys...>)
+{
+    static_assert(sizeof...(Xs) == sizeof...(Ys), "wrong! inconsistent size");
+
+    return Sequence<(Xs - Ys)...>{};
+}
+
+template <index_t... Xs, index_t... Ys>
+__host__ __device__ constexpr auto operator*(Sequence<Xs...>, Sequence<Ys...>)
+{
+    static_assert(sizeof...(Xs) == sizeof...(Ys), "wrong! inconsistent size");
+
+    return Sequence<(Xs * Ys)...>{};
+}
+
+template <index_t... Xs, index_t... Ys>
+__host__ __device__ constexpr auto operator/(Sequence<Xs...>, Sequence<Ys...>)
+{
+    static_assert(sizeof...(Xs) == sizeof...(Ys), "wrong! inconsistent size");
+
+    return Sequence<(Xs / Ys)...>{};
+}
+
+template <index_t... Xs, index_t... Ys>
+__host__ __device__ constexpr auto operator%(Sequence<Xs...>, Sequence<Ys...>)
+{
+    static_assert(sizeof...(Xs) == sizeof...(Ys), "wrong! inconsistent size");
+
+    return Sequence<(Xs % Ys)...>{};
+}
+
+template <index_t... Xs, index_t Y>
+__host__ __device__ constexpr auto operator+(Sequence<Xs...>, Number<Y>)
+{
+    return Sequence<(Xs + Y)...>{};
+}
+
+template <index_t... Xs, index_t Y>
+__host__ __device__ constexpr auto operator-(Sequence<Xs...>, Number<Y>)
+{
+    return Sequence<(Xs - Y)...>{};
+}
+
+template <index_t... Xs, index_t Y>
+__host__ __device__ constexpr auto operator*(Sequence<Xs...>, Number<Y>)
+{
+    return Sequence<(Xs * Y)...>{};
+}
+
+template <index_t... Xs, index_t Y>
+__host__ __device__ constexpr auto operator/(Sequence<Xs...>, Number<Y>)
+{
+    return Sequence<(Xs / Y)...>{};
+}
+
+template <index_t... Xs, index_t Y>
+__host__ __device__ constexpr auto operator%(Sequence<Xs...>, Number<Y>)
+{
+    return Sequence<(Xs % Y)...>{};
+}
+
+template <index_t Y, index_t... Xs>
+__host__ __device__ constexpr auto operator+(Number<Y>, Sequence<Xs...>)
+{
+    return Sequence<(Y + Xs)...>{};
+}
+
+template <index_t Y, index_t... Xs>
+__host__ __device__ constexpr auto operator-(Number<Y>, Sequence<Xs...>)
+{
+    return Sequence<(Y - Xs)...>{};
+}
+
+template <index_t Y, index_t... Xs>
+__host__ __device__ constexpr auto operator*(Number<Y>, Sequence<Xs...>)
+{
+    return Sequence<(Y * Xs)...>{};
+}
+
+template <index_t Y, index_t... Xs>
+__host__ __device__ constexpr auto operator/(Number<Y>, Sequence<Xs...>)
+{
+    return Sequence<(Y / Xs)...>{};
+}
+
+template <index_t Y, index_t... Xs>
+__host__ __device__ constexpr auto operator%(Number<Y>, Sequence<Xs...>)
+{
+    return Sequence<(Y % Xs)...>{};
+}
+
+template <index_t I, index_t... Is>
+__host__ __device__ constexpr auto sequence_pop_front(Sequence<I, Is...>)
+{
+    return Sequence<Is...>{};
+}
+
+template <typename Seq>
+__host__ __device__ constexpr auto sequence_pop_back(Seq)
+{
+    static_assert(Seq::Size() > 0, "wrong! cannot pop an empty Sequence!");
+    return sequence_pop_front(Seq::Reverse()).Reverse();
+}
+
+template <typename... Seqs>
+__host__ __device__ constexpr auto merge_sequences(Seqs...)
+{
+    return typename sequence_merge<Seqs...>::type{};
+}
+
+template <typename F, index_t... Xs>
+__host__ __device__ constexpr auto transform_sequences(F f, Sequence<Xs...>)
+{
+    return Sequence<f(Xs)...>{};
+}
+
+template <typename F, index_t... Xs, index_t... Ys>
+__host__ __device__ constexpr auto transform_sequences(F f, Sequence<Xs...>, Sequence<Ys...>)
+{
+    static_assert(Sequence<Xs...>::mSize == Sequence<Ys...>::mSize, "Dim not the same");
+
+    return Sequence<f(Xs, Ys)...>{};
+}
+
+template <typename F, index_t... Xs, index_t... Ys, index_t... Zs>
+__host__ __device__ constexpr auto
+transform_sequences(F f, Sequence<Xs...>, Sequence<Ys...>, Sequence<Zs...>)
+{
+    static_assert(Sequence<Xs...>::mSize == Sequence<Ys...>::mSize &&
+                      Sequence<Xs...>::mSize == Sequence<Zs...>::mSize,
+                  "Dim not the same");
+
+    return Sequence<f(Xs, Ys, Zs)...>{};
+}
+
+template <typename Seq, typename Reduce, index_t Init>
+__host__ __device__ constexpr auto reverse_inclusive_scan_sequence(Seq, Reduce, Number<Init>)
+{
+    return typename sequence_reverse_inclusive_scan<Seq, Reduce, Init>::type{};
+}
+
+template <typename Seq, typename Reduce, index_t Init>
+__host__ __device__ constexpr auto reverse_exclusive_scan_sequence(Seq, Reduce, Number<Init>)
+{
+    return reverse_inclusive_scan_sequence(Seq::PopFront(), Reduce{}, Number<Init>{})
+        .PushBack(Number<Init>{});
+}
+
+template <typename Seq, typename Reduce, index_t Init>
+__host__ __device__ constexpr auto inclusive_scan_sequence(Seq, Reduce, Number<Init>)
+{
+    return reverse_inclusive_scan_sequence(Seq{}.Reverse(), Reduce{}, Number<Init>{}).Reverse();
+}
+
+// e.g. Seq<2, 3, 4> --> Seq<0, 2, 5>, Init=0, Reduce=Add
+//      ResultSeq  TargetSeq  Reduce
+template <typename, typename, typename>
+struct sequence_exclusive_scan;
+
+template <index_t... Xs, index_t Y, index_t... Ys, typename Reduce>
+struct sequence_exclusive_scan<Sequence<Xs...>, Sequence<Y, Ys...>, Reduce>
+{
+    using old_scan = typename sequence_merge<Sequence<Xs...>,
+                                             Sequence<Reduce{}(Y, Sequence<Xs...>{}.Back())>>::type;
+    using type     = typename sequence_exclusive_scan<old_scan, Sequence<Ys...>, Reduce>::type;
+};
+
+template <index_t... Xs, index_t Y, typename Reduce>
+struct sequence_exclusive_scan<Sequence<Xs...>, Sequence<Y>, Reduce>
+{
+    using type = Sequence<Xs...>;
+};
+
+template <index_t... Xs, typename Reduce>
+struct sequence_exclusive_scan<Sequence<Xs...>, Sequence<>, Reduce>
+{
+    using type = Sequence<Xs...>;
+};
+
+template <typename Seq, typename Reduce, index_t Init>
+constexpr auto exclusive_scan_sequence(Seq, Reduce, Number<Init>)
+{
+    // TODO: c++20 and later can pass in Reduce with a lambda expression
+    return typename sequence_exclusive_scan<Sequence<Init>, Seq, Reduce>::type{};
+}
+
+template <typename Seq>
+constexpr auto prefix_sum_sequence(Seq)
+{
+    return typename sequence_exclusive_scan<Sequence<0>,
+                                            typename sequence_merge<Seq, Sequence<0>>::type,
+                                            math::plus<index_t>>::type{};
+}
+
+template <typename Seq, index_t... Is>
+__host__ __device__ constexpr auto pick_sequence_elements_by_ids(Seq, Sequence<Is...> /* ids */)
+{
+    return Sequence<Seq::At(Number<Is>{})...>{};
+}
+
+#if 1
+namespace detail {
+template <typename WorkSeq, typename RemainSeq, typename RemainMask>
+struct pick_sequence_elements_by_mask_impl
+{
+    using new_work_seq = typename conditional<RemainMask::Front(),
+                                              decltype(WorkSeq::PushBack(RemainSeq::Front())),
+                                              WorkSeq>::type;
+
+    using type =
+        typename pick_sequence_elements_by_mask_impl<new_work_seq,
+                                                     decltype(RemainSeq::PopFront()),
+                                                     decltype(RemainMask::PopFront())>::type;
+};
+
+template <typename WorkSeq>
+struct pick_sequence_elements_by_mask_impl<WorkSeq, Sequence<>, Sequence<>>
+{
+    using type = WorkSeq;
+};
+
+} // namespace detail
+
+template <typename Seq, typename Mask>
+__host__ __device__ constexpr auto pick_sequence_elements_by_mask(Seq, Mask)
+{
+    static_assert(Seq::Size() == Mask::Size(), "wrong!");
+
+    return typename detail::pick_sequence_elements_by_mask_impl<Sequence<>, Seq, Mask>::type{};
+}
+
+namespace detail {
+template <typename WorkSeq, typename RemainValues, typename RemainIds>
+struct modify_sequence_elements_by_ids_impl
+{
+    using new_work_seq = decltype(WorkSeq::Modify(RemainIds::Front(), RemainValues::Front()));
+
+    using type =
+        typename modify_sequence_elements_by_ids_impl<new_work_seq,
+                                                      decltype(RemainValues::PopFront()),
+                                                      decltype(RemainIds::PopFront())>::type;
+};
+
+template <typename WorkSeq>
+struct modify_sequence_elements_by_ids_impl<WorkSeq, Sequence<>, Sequence<>>
+{
+    using type = WorkSeq;
+};
+} // namespace detail
+
+template <typename Seq, typename Values, typename Ids>
+__host__ __device__ constexpr auto modify_sequence_elements_by_ids(Seq, Values, Ids)
+{
+    static_assert(Values::Size() == Ids::Size() && Seq::Size() >= Values::Size(), "wrong!");
+
+    return typename detail::modify_sequence_elements_by_ids_impl<Seq, Values, Ids>::type{};
+}
+#endif
+
+template <typename Seq, typename Reduce, index_t Init>
+__host__ __device__ constexpr index_t
+reduce_on_sequence(Seq, Reduce f, Number<Init> /*initial_value*/)
+{
+    index_t result = Init;
+
+    for(index_t i = 0; i < Seq::Size(); ++i)
+    {
+        result = f(result, Seq::At(i));
+    }
+
+    return result;
+}
+
+// TODO: a generic any_of for any container
+template <typename Seq, typename F>
+__host__ __device__ constexpr bool sequence_any_of(Seq, F f)
+{
+    bool flag = false;
+
+    for(index_t i = 0; i < Seq::Size(); ++i)
+    {
+        flag = flag || f(Seq::At(i));
+    }
+
+    return flag;
+}
+
+// TODO: a generic all_of for any container
+template <typename Seq, typename F>
+__host__ __device__ constexpr bool sequence_all_of(Seq, F f)
+{
+    bool flag = true;
+
+    for(index_t i = 0; i < Seq::Size(); ++i)
+    {
+        flag = flag && f(Seq::At(i));
+    }
+
+    return flag;
+}
+
+template <typename... Seqs>
+using sequence_merge_t = typename sequence_merge<Seqs...>::type;
+
+template <index_t NSize, index_t I>
+using uniform_sequence_gen_t = typename uniform_sequence_gen<NSize, I>::type;
+
+} // namespace ck
diff --git a/include/ck/utility/sequence_helper.hpp b/include/ck/utility/sequence_helper.hpp
new file mode 100644
index 0000000000..3734b9c326
--- /dev/null
+++ b/include/ck/utility/sequence_helper.hpp
@@ -0,0 +1,93 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/array.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/utility/macro_func_array_to_sequence.hpp"
+
+namespace ck {
+
+template <index_t... Is>
+__host__ __device__ constexpr auto make_sequence(Number<Is>...)
+{
+    return Sequence<Is...>{};
+}
+
+// F() returns index_t
+// F use default constructor, so F cannot be lambda function
+template <typename F, index_t N>
+__host__ __device__ constexpr auto generate_sequence(F, Number<N>)
+{
+    return typename sequence_gen<N, F>::type{};
+}
+
+// F() returns Number<>
+// F could be lambda function
+template <typename F, index_t N>
+__host__ __device__ constexpr auto generate_sequence_v2(F&& f, Number<N>)
+{
+    return unpack([&f](auto&&... xs) { return make_sequence(f(xs)...); },
+                  typename arithmetic_sequence_gen<0, N, 1>::type{});
+}
+
+template <index_t... Is>
+__host__ __device__ constexpr auto to_sequence(Tuple<Number<Is>...>)
+{
+    return Sequence<Is...>{};
+}
+
+namespace detail {
+template <index_t h_idx, typename SeqSortedSamples, typename SeqRange>
+struct sorted_sequence_histogram;
+
+template <index_t h_idx, index_t x, index_t... xs, index_t r, index_t... rs>
+struct sorted_sequence_histogram<h_idx, Sequence<x, xs...>, Sequence<r, rs...>>
+{
+    template <typename Histogram>
+    constexpr auto operator()(Histogram& h)
+    {
+        if constexpr(x < r)
+        {
+            h.template At<h_idx>() += 1;
+            sorted_sequence_histogram<h_idx, Sequence<xs...>, Sequence<r, rs...>>{}(h);
+        }
+        else
+        {
+            h.template At<h_idx + 1>() = 1;
+            sorted_sequence_histogram<h_idx + 1, Sequence<xs...>, Sequence<rs...>>{}(h);
+        }
+    }
+};
+
+template <index_t h_idx, index_t x, index_t r, index_t... rs>
+struct sorted_sequence_histogram<h_idx, Sequence<x>, Sequence<r, rs...>>
+{
+    template <typename Histogram>
+    constexpr auto operator()(Histogram& h)
+    {
+        if constexpr(x < r)
+        {
+            h.template At<h_idx>() += 1;
+        }
+    }
+};
+} // namespace detail
+
+// SeqSortedSamples: <0, 2, 3, 5, 7>, SeqRange: <0, 3, 6, 9> -> SeqHistogram : <2, 2, 1>
+template <typename SeqSortedSamples, index_t r, index_t... rs>
+constexpr auto histogram_sorted_sequence(SeqSortedSamples, Sequence<r, rs...>)
+{
+    constexpr auto bins      = sizeof...(rs); // or categories
+    constexpr auto histogram = [&]() {
+        Array<index_t, bins> h{0}; // make sure this can clear all element to zero
+        detail::sorted_sequence_histogram<0, SeqSortedSamples, Sequence<rs...>>{}(h);
+        return h;
+    }();
+
+    return TO_SEQUENCE(histogram, bins);
+}
+
+} // namespace ck
diff --git a/include/ck/utility/span.hpp b/include/ck/utility/span.hpp
new file mode 100644
index 0000000000..5e7567a847
--- /dev/null
+++ b/include/ck/utility/span.hpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstddef>
+#include <array>
+#include <type_traits>
+
+namespace ck {
+
+template <typename T>
+class span
+{
+    public:
+    using element_type    = T;
+    using value_type      = std::remove_cv_t<element_type>;
+    using size_type       = std::size_t;
+    using difference_type = std::ptrdiff_t;
+    using pointer         = element_type*;
+    using const_pointer   = const element_type*;
+    using reference       = element_type&;
+    using const_reference = const element_type&;
+    using iterator        = pointer;
+    using const_iterator  = pointer;
+
+    constexpr span() : span(nullptr, size_type{0}) {}
+
+    constexpr span(pointer first, size_type count) : ptr_(first), size_(count) {}
+
+    constexpr span(pointer first, pointer last) : span(first, last - first) {}
+
+    template <std::size_t N>
+    constexpr span(element_type (&arr)[N]) noexcept : span(arr, N)
+    {
+    }
+
+    template <std::size_t N>
+    constexpr span(std::array<value_type, N>& arr) noexcept : span(arr.data(), N)
+    {
+    }
+
+    template <typename Container>
+    constexpr span(const Container& container) : span(container.data(), container.size())
+    {
+    }
+
+    constexpr iterator begin() const noexcept { return ptr_; }
+    constexpr const_iterator cbegin() const noexcept { return begin(); }
+
+    constexpr iterator end() const noexcept { return begin() + size(); }
+    constexpr const_iterator cend() const noexcept { return end(); }
+
+    constexpr reference front() const { return *begin(); }
+    constexpr reference back() const { return *(--end()); }
+
+    constexpr reference operator[](size_type idx) const { return *(begin() + idx); }
+    constexpr pointer data() const noexcept { return ptr_; }
+
+    constexpr size_type size() const noexcept { return size_; }
+
+    private:
+    pointer ptr_;
+    size_type size_;
+};
+
+} // namespace ck
diff --git a/include/ck/utility/static_assert.hpp b/include/ck/utility/static_assert.hpp
new file mode 100644
index 0000000000..0e671632ab
--- /dev/null
+++ b/include/ck/utility/static_assert.hpp
@@ -0,0 +1,12 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#ifndef STATIC_ASSERT
+#ifndef NDEBUG
+#define STATIC_ASSERT(...) static_assert(__VA_ARGS__)
+#else
+#define STATIC_ASSERT(...)
+#endif
+#endif
diff --git a/include/ck/utility/static_buffer.hpp b/include/ck/utility/static_buffer.hpp
new file mode 100644
index 0000000000..0ccebd4766
--- /dev/null
+++ b/include/ck/utility/static_buffer.hpp
@@ -0,0 +1,241 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "statically_indexed_array.hpp"
+
+namespace ck {
+
+// static buffer for scalar
+template <AddressSpaceEnum AddressSpace,
+          typename S_,
+          index_t N,
+          bool InvalidElementUseNumericalZeroValue> // TODO remove this bool, no longer needed
+struct StaticBuffer : public StaticallyIndexedArray<remove_cvref_t<S_>, N>
+{
+    using S    = remove_cvref_t<S_>;
+    using type = S;
+    using base = StaticallyIndexedArray<S, N>;
+
+    __host__ __device__ constexpr StaticBuffer() : base{} {}
+
+    template <typename... Ys>
+    __host__ __device__ constexpr StaticBuffer& operator=(const Tuple<Ys...>& y)
+    {
+        static_assert(base::Size() == sizeof...(Ys), "wrong! size not the same");
+        StaticBuffer& x = *this;
+        static_for<0, base::Size(), 1>{}([&](auto i) { x(i) = y[i]; });
+        return x;
+    }
+
+    __host__ __device__ static constexpr AddressSpaceEnum GetAddressSpace() { return AddressSpace; }
+
+    __host__ __device__ static constexpr index_t Size() { return N; }
+
+    __host__ __device__ static constexpr bool IsStaticBuffer() { return true; }
+
+    __host__ __device__ static constexpr bool IsDynamicBuffer() { return false; }
+
+    // read access to scalar
+    template <index_t I>
+    __host__ __device__ constexpr const S& operator[](Number<I> i) const
+    {
+        return base::operator[](i);
+    }
+
+    // write access to scalar
+    template <index_t I>
+    __host__ __device__ constexpr S& operator()(Number<I> i)
+    {
+        return base::operator()(i);
+    }
+
+    // Get a vector (type X)
+    // "is" is offset of S, not X.
+    // "is" should be aligned to X
+    template <typename X_,
+              index_t Is,
+              typename enable_if<has_same_scalar_type<S, X_>::value, bool>::type = false>
+    __host__ __device__ constexpr remove_reference_t<X_> GetAsType(Number<Is> is) const
+    {
+        using X = remove_cvref_t<X_>;
+
+        constexpr index_t kSPerX = scalar_type<X>::vector_size;
+
+        static_assert(Is % kSPerX == 0, "wrong! \"Is\" should be aligned to X");
+
+        vector_type<S, kSPerX> vx;
+
+        static_for<0, kSPerX, 1>{}(
+            [&](auto j) { vx.template AsType<S>()(j) = base::operator[](is + j); });
+
+        return vx.template AsType<X>().template At<0>();
+    }
+
+    // Set a vector (type X)
+    // "is" is offset of S, not X.
+    // "is" should be aligned to X
+    template <typename X_,
+              index_t Is,
+              typename enable_if<has_same_scalar_type<S, X_>::value, bool>::type = false>
+    __host__ __device__ constexpr void SetAsType(Number<Is> is, X_ x)
+    {
+        using X = remove_cvref_t<X_>;
+
+        constexpr index_t kSPerX = scalar_type<X>::vector_size;
+
+        static_assert(Is % kSPerX == 0, "wrong! \"Is\" should be aligned to X");
+
+        const vector_type<S, kSPerX> vx{x};
+
+        static_for<0, kSPerX, 1>{}(
+            [&](auto j) { base::operator()(is + j) = vx.template AsType<S>()[j]; });
+    }
+
+    __host__ __device__ void Initialize(const S& x)
+    {
+        static_for<0, N, 1>{}([&](auto i) { operator()(i) = S{x}; });
+    }
+
+    // FIXME: deprecated
+    __host__ __device__ void Clear() { Initialize(0); }
+
+    // FIXME: deprecated
+    __host__ __device__ constexpr StaticBuffer& operator=(const S& v)
+    {
+        Initialize(v);
+        return *this;
+    }
+};
+
+// static buffer for vector
+template <AddressSpaceEnum AddressSpace,
+          typename S,
+          index_t NumOfVector,
+          index_t ScalarPerVector,
+          bool InvalidElementUseNumericalZeroValue, // TODO remove this bool, no longer needed,
+          typename enable_if<is_scalar_type<S>::value, bool>::type = false>
+struct StaticBufferTupleOfVector
+    : public StaticallyIndexedArray<vector_type<S, ScalarPerVector>, NumOfVector>
+{
+    using V    = typename vector_type<S, ScalarPerVector>::type;
+    using base = StaticallyIndexedArray<vector_type<S, ScalarPerVector>, NumOfVector>;
+
+    static constexpr auto s_per_v   = Number<ScalarPerVector>{};
+    static constexpr auto num_of_v_ = Number<NumOfVector>{};
+    static constexpr auto s_per_buf = s_per_v * num_of_v_;
+
+    __host__ __device__ constexpr StaticBufferTupleOfVector() : base{} {}
+
+    __host__ __device__ static constexpr AddressSpaceEnum GetAddressSpace() { return AddressSpace; }
+
+    __host__ __device__ static constexpr bool IsStaticBuffer() { return true; }
+
+    __host__ __device__ static constexpr bool IsDynamicBuffer() { return false; }
+
+    __host__ __device__ static constexpr index_t Size() { return s_per_buf; };
+
+    // Get S
+    // i is offset of S
+    template <index_t I>
+    __host__ __device__ constexpr const S& operator[](Number<I> i) const
+    {
+        constexpr auto i_v = i / s_per_v;
+        constexpr auto i_s = i % s_per_v;
+
+        return base::operator[](i_v).template AsType<S>()[i_s];
+    }
+
+    // Set S
+    // i is offset of S
+    template <index_t I>
+    __host__ __device__ constexpr S& operator()(Number<I> i)
+    {
+        constexpr auto i_v = i / s_per_v;
+        constexpr auto i_s = i % s_per_v;
+
+        return base::operator()(i_v).template AsType<S>()(i_s);
+    }
+
+    // Get X
+    // i is offset of S, not X. i should be aligned to X
+    template <typename X,
+              index_t I,
+              typename enable_if<has_same_scalar_type<S, X>::value, bool>::type = false>
+    __host__ __device__ constexpr auto GetAsType(Number<I> i) const
+    {
+        constexpr auto s_per_x = Number<scalar_type<remove_cvref_t<X>>::vector_size>{};
+
+        static_assert(s_per_v % s_per_x == 0, "wrong! V must  one or multiple X");
+        static_assert(i % s_per_x == 0, "wrong!");
+
+        constexpr auto i_v = i / s_per_v;
+        constexpr auto i_x = (i % s_per_v) / s_per_x;
+
+        return base::operator[](i_v).template AsType<X>()[i_x];
+    }
+
+    // Set X
+    // i is offset of S, not X. i should be aligned to X
+    template <typename X,
+              index_t I,
+              typename enable_if<has_same_scalar_type<S, X>::value, bool>::type = false>
+    __host__ __device__ constexpr void SetAsType(Number<I> i, X x)
+    {
+        constexpr auto s_per_x = Number<scalar_type<remove_cvref_t<X>>::vector_size>{};
+
+        static_assert(s_per_v % s_per_x == 0, "wrong! V must contain one or multiple X");
+        static_assert(i % s_per_x == 0, "wrong!");
+
+        constexpr auto i_v = i / s_per_v;
+        constexpr auto i_x = (i % s_per_v) / s_per_x;
+
+        base::operator()(i_v).template AsType<X>()(i_x) = x;
+    }
+
+    // Get read access to vector_type V
+    // i is offset of S, not V. i should be aligned to V
+    template <index_t I>
+    __host__ __device__ constexpr const auto& GetVectorTypeReference(Number<I> i) const
+    {
+        static_assert(i % s_per_v == 0, "wrong!");
+
+        constexpr auto i_v = i / s_per_v;
+
+        return base::operator[](i_v);
+    }
+
+    // Get write access to vector_type V
+    // i is offset of S, not V. i should be aligned to V
+    template <index_t I>
+    __host__ __device__ constexpr auto& GetVectorTypeReference(Number<I> i)
+    {
+        static_assert(i % s_per_v == 0, "wrong!");
+
+        constexpr auto i_v = i / s_per_v;
+
+        return base::operator()(i_v);
+    }
+
+    __host__ __device__ void Clear()
+    {
+        constexpr index_t NumScalars = NumOfVector * ScalarPerVector;
+
+        static_for<0, NumScalars, 1>{}([&](auto i) { SetAsType(i, S{0}); });
+    }
+};
+
+template <AddressSpaceEnum AddressSpace, typename T, index_t N>
+__host__ __device__ constexpr auto make_static_buffer(Number<N>)
+{
+    return StaticBuffer<AddressSpace, T, N, true>{};
+}
+
+template <AddressSpaceEnum AddressSpace, typename T, long_index_t N>
+__host__ __device__ constexpr auto make_static_buffer(LongNumber<N>)
+{
+    return StaticBuffer<AddressSpace, T, N, true>{};
+}
+
+} // namespace ck
diff --git a/include/ck/utility/statically_indexed_array.hpp b/include/ck/utility/statically_indexed_array.hpp
new file mode 100644
index 0000000000..b1bf919e77
--- /dev/null
+++ b/include/ck/utility/statically_indexed_array.hpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "functional2.hpp"
+#include "sequence.hpp"
+#include "tuple.hpp"
+
+namespace ck {
+
+namespace detail {
+template <typename X, typename Y>
+struct tuple_concat;
+
+template <typename... Xs, typename... Ys>
+struct tuple_concat<Tuple<Xs...>, Tuple<Ys...>>
+{
+    using type = Tuple<Xs..., Ys...>;
+};
+
+template <typename T, index_t N>
+struct StaticallyIndexedArrayImpl
+{
+    using type =
+        typename tuple_concat<typename StaticallyIndexedArrayImpl<T, N / 2>::type,
+                              typename StaticallyIndexedArrayImpl<T, N - N / 2>::type>::type;
+};
+
+template <typename T>
+struct StaticallyIndexedArrayImpl<T, 0>
+{
+    using type = Tuple<>;
+};
+
+template <typename T>
+struct StaticallyIndexedArrayImpl<T, 1>
+{
+    using type = Tuple<T>;
+};
+} // namespace detail
+
+template <typename T, index_t N>
+using StaticallyIndexedArray = typename detail::StaticallyIndexedArrayImpl<T, N>::type;
+
+template <typename X, typename... Xs>
+__host__ __device__ constexpr auto make_statically_indexed_array(const X& x, const Xs&... xs)
+{
+    return StaticallyIndexedArray<X, sizeof...(Xs) + 1>(x, static_cast<X>(xs)...);
+}
+
+// make empty StaticallyIndexedArray
+template <typename X>
+__host__ __device__ constexpr auto make_statically_indexed_array()
+{
+    return StaticallyIndexedArray<X, 0>();
+}
+
+} // namespace ck
diff --git a/include/ck/utility/statically_indexed_array_multi_index.hpp b/include/ck/utility/statically_indexed_array_multi_index.hpp
new file mode 100644
index 0000000000..d695e8cc64
--- /dev/null
+++ b/include/ck/utility/statically_indexed_array_multi_index.hpp
@@ -0,0 +1,36 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "common_header.hpp"
+
+namespace ck {
+
+#if 0
+template <index_t N>
+using MultiIndex = StaticallyIndexedArray<index_t, N>;
+
+template <typename... Xs>
+__host__ __device__ constexpr auto make_multi_index(Xs&&... xs)
+{
+    static_assert(false, "wrong! deprecated");
+
+    return make_statically_indexed_array<index_t>(index_t{xs}...);
+}
+
+template <index_t NSize>
+__host__ __device__ constexpr auto make_zero_multi_index()
+{
+    return unpack([](auto... xs) { return make_multi_index(xs...); },
+                  typename uniform_sequence_gen<NSize, 0>::type{});
+}
+
+template <typename T>
+__host__ __device__ constexpr auto to_multi_index(const T& x)
+{
+    return unpack([](auto... ys) { return make_multi_index(ys...); }, x);
+}
+#endif
+
+} // namespace ck
diff --git a/include/ck/utility/synchronization.hpp b/include/ck/utility/synchronization.hpp
new file mode 100644
index 0000000000..775e7ac3a3
--- /dev/null
+++ b/include/ck/utility/synchronization.hpp
@@ -0,0 +1,33 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+
+namespace ck {
+
+__device__ void block_sync_lds()
+{
+#if CK_EXPERIMENTAL_BLOCK_SYNC_LDS_WITHOUT_SYNC_VMEM
+    asm volatile("\
+    s_waitcnt lgkmcnt(0) \n \
+    s_barrier \
+    " ::);
+#else
+    __syncthreads();
+#endif
+}
+
+__device__ void s_nop()
+{
+#if 1
+    asm volatile("\
+    s_nop 0 \n \
+    " ::);
+#else
+    __builtin_amdgcn_sched_barrier(0);
+#endif
+}
+
+} // namespace ck
diff --git a/include/ck/utility/thread_group.hpp b/include/ck/utility/thread_group.hpp
new file mode 100644
index 0000000000..1cd6b2f3ce
--- /dev/null
+++ b/include/ck/utility/thread_group.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "get_id.hpp"
+
+namespace ck {
+
+template <index_t ThreadPerBlock>
+struct ThisThreadBlock
+{
+    static constexpr index_t kNumThread_ = ThreadPerBlock;
+
+    __device__ static constexpr index_t GetNumOfThread() { return kNumThread_; }
+
+    __device__ static constexpr bool IsBelong() { return true; }
+
+    __device__ static index_t GetThreadId() { return get_thread_local_1d_id(); }
+};
+
+} // namespace ck
diff --git a/include/ck/utility/transpose_vectors.hpp b/include/ck/utility/transpose_vectors.hpp
new file mode 100644
index 0000000000..503832ab4f
--- /dev/null
+++ b/include/ck/utility/transpose_vectors.hpp
@@ -0,0 +1,194 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "statically_indexed_array.hpp"
+#include "data_type.hpp"
+
+namespace ck {
+
+// S: scalar type
+// NX: # of vector before transpose
+// NY: # of vector after transpose
+template <typename S,
+          index_t NX,
+          index_t NY,
+          typename enable_if<is_scalar_type<S>::value, bool>::type = false>
+struct transpose_vectors;
+
+// transpose fp16 2x2
+__device__ void transpose_fp16_2x2(const half2_t& x0, const half2_t& x1, half2_t& y0, half2_t& y1)
+{
+#if 0
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    const vector_type<half_t, 2> vx0{x0}, vx1{x1};
+    vector_type<half_t, 2> vy0, vy1;
+
+    vy0.template AsType<half_t>()(I0) = vx0.template AsType<half_t>()[I0];
+    vy0.template AsType<half_t>()(I1) = vx1.template AsType<half_t>()[I0];
+
+    vy1.template AsType<half_t>()(I0) = vx0.template AsType<half_t>()[I1];
+    vy1.template AsType<half_t>()(I1) = vx1.template AsType<half_t>()[I1];
+
+    y0 = vy0.template AsType<half2_t>()[I0];
+    y1 = vy1.template AsType<half2_t>()[I0];
+#else
+    constexpr int32_t m0 = 0x05040100;
+    constexpr int32_t m1 = 0x07060302;
+
+    // ex: v_perm_b32(0x 11 22 33 44, 0x 55 66 77 88, 0x 05 01 04 00) -> 0x33774488
+    //                   -- -- -- --     -- -- -- --      -  -  -  -
+    //             index  7  6  5  4      3  2  1  0     33 77 44 88
+    // index is reversed because of little endianness (least significant bits first)
+    y0 = bit_cast<half2_t>(__builtin_amdgcn_perm(bit_cast<int32_t>(x1), bit_cast<int32_t>(x0), m0));
+    y1 = bit_cast<half2_t>(__builtin_amdgcn_perm(bit_cast<int32_t>(x1), bit_cast<int32_t>(x0), m1));
+#endif
+}
+
+template <index_t NX, index_t NY>
+struct transpose_vectors<half_t, NX, NY>
+{
+    // we got [NY * NX] amount of S data to be transposed
+    static constexpr index_t s_per_x = NY;
+    static constexpr index_t s_per_y = NX;
+
+    using S  = half_t;
+    using VX = vector_type<half_t, s_per_x>;
+    using VY = vector_type<half_t, s_per_y>;
+
+    __device__ void operator()(const StaticallyIndexedArray<const VX&, NX>& vx_tuple,
+                               StaticallyIndexedArray<VY&, NY>& vy_tuple)
+    {
+        static constexpr auto I1 = Number<1>{};
+        static constexpr auto I2 = Number<2>{};
+
+        static_assert((NX % 2 == 0 && NY % 2 == 0), "wrong!");
+
+        // loop over 2x2 tile and transpose data from vx_tuple into vy_tuple
+        static_for<0, NY, 2>{}([&](auto iy) {
+            static_for<0, NX, 2>{}([&](auto ix) {
+                // reference to 2 half2_t data from vx_tuple
+                const auto& x_s2_0 = vx_tuple[ix].template AsType<half2_t>()[iy / I2];
+                const auto& x_s2_1 = vx_tuple[ix + I1].template AsType<half2_t>()[iy / I2];
+
+                // reference to 2 half2_t data from vy_tuple
+                auto& y_s2_0 = vy_tuple(iy).template AsType<half2_t>()(ix / I2);
+                auto& y_s2_1 = vy_tuple(iy + I1).template AsType<half2_t>()(ix / I2);
+
+                // transpose
+                transpose_fp16_2x2(x_s2_0, x_s2_1, y_s2_0, y_s2_1);
+            });
+        });
+    }
+
+    // FIXME: duplicated code
+    __device__ void operator()(const StaticallyIndexedArray<VX, NX>& vx_tuple,
+                               StaticallyIndexedArray<VY, NY>& vy_tuple)
+    {
+        static constexpr auto I1 = Number<1>{};
+        static constexpr auto I2 = Number<2>{};
+
+        static_assert((NX % 2 == 0 && NY % 2 == 0), "wrong!");
+
+        // loop over 2x2 tile and transpose data from vx_tuple into vy_tuple
+        static_for<0, NY, 2>{}([&](auto iy) {
+            static_for<0, NX, 2>{}([&](auto ix) {
+                // reference to 2 half2_t data from vx_tuple
+                const auto& x_s2_0 = vx_tuple[ix].template AsType<half2_t>()[iy / I2];
+                const auto& x_s2_1 = vx_tuple[ix + I1].template AsType<half2_t>()[iy / I2];
+
+                // reference to 2 half2_t data from vy_tuple
+                auto& y_s2_0 = vy_tuple(iy).template AsType<half2_t>()(ix / I2);
+                auto& y_s2_1 = vy_tuple(iy + I1).template AsType<half2_t>()(ix / I2);
+
+                // transpose
+                transpose_fp16_2x2(x_s2_0, x_s2_1, y_s2_0, y_s2_1);
+            });
+        });
+    }
+};
+
+// transpose int8 4x4
+__device__ void transpose_int8_4x4(const int8x4_t& x0,
+                                   const int8x4_t& x1,
+                                   const int8x4_t& x2,
+                                   const int8x4_t& x3,
+                                   int8x4_t& y0,
+                                   int8x4_t& y1,
+                                   int8x4_t& y2,
+                                   int8x4_t& y3)
+{
+    int32_t t0, t1;
+    int32_t z0, z1, z2, z3;
+    constexpr int32_t m0 = 0x05010400;
+    constexpr int32_t m1 = 0x05040100;
+    constexpr int32_t m2 = 0x07060302;
+    constexpr int32_t m3 = 0x07030602;
+
+    // ex: v_perm_b32(0x 11 22 33 44, 0x 55 66 77 88, 0x 05 01 04 00) -> 0x33774488
+    //                   -- -- -- --     -- -- -- --      -  -  -  -
+    //             index  7  6  5  4      3  2  1  0     33 77 44 88
+    // index is reversed because of little endianness (least significant bits first)
+    t0 = __builtin_amdgcn_perm(bit_cast<int32_t>(x1), bit_cast<int32_t>(x0), m0);
+    t1 = __builtin_amdgcn_perm(bit_cast<int32_t>(x3), bit_cast<int32_t>(x2), m0);
+    z0 = __builtin_amdgcn_perm(bit_cast<int32_t>(t1), bit_cast<int32_t>(t0), m1);
+    z1 = __builtin_amdgcn_perm(bit_cast<int32_t>(t1), bit_cast<int32_t>(t0), m2);
+    t0 = __builtin_amdgcn_perm(bit_cast<int32_t>(x1), bit_cast<int32_t>(x0), m3);
+    t1 = __builtin_amdgcn_perm(bit_cast<int32_t>(x3), bit_cast<int32_t>(x2), m3);
+    z2 = __builtin_amdgcn_perm(bit_cast<int32_t>(t1), bit_cast<int32_t>(t0), m1);
+    z3 = __builtin_amdgcn_perm(bit_cast<int32_t>(t1), bit_cast<int32_t>(t0), m2);
+
+    y0 = bit_cast<int8x4_t>(z0);
+    y1 = bit_cast<int8x4_t>(z1);
+    y2 = bit_cast<int8x4_t>(z2);
+    y3 = bit_cast<int8x4_t>(z3);
+}
+
+template <index_t NX, index_t NY>
+struct transpose_vectors<int8_t, NX, NY>
+{
+    // we got [NY * NX] amount of S data to be transposed
+    static constexpr index_t s_per_x = NY;
+    static constexpr index_t s_per_y = NX;
+
+    using S  = int8_t;
+    using VX = vector_type<int8_t, s_per_x>;
+    using VY = vector_type<int8_t, s_per_y>;
+
+    __device__ void operator()(const StaticallyIndexedArray<const VX&, NX>& vx_tuple,
+                               StaticallyIndexedArray<VY&, NY>& vy_tuple)
+    {
+        static constexpr auto I1 = Number<1>{};
+        static constexpr auto I2 = Number<2>{};
+        static constexpr auto I3 = Number<3>{};
+        static constexpr auto I4 = Number<4>{};
+
+        static_assert((NX % 4 == 0 && NY % 4 == 0), "wrong!");
+
+        // loop over 4x4 tile and transpose data from vx_tuple into vy_tuple
+        static_for<0, NY, 4>{}([&](auto iy) {
+            static_for<0, NX, 4>{}([&](auto ix) {
+                // reference to 4 int8 data from vx_tuple
+                const auto& x_s4_0 = vx_tuple[ix].template AsType<int8x4_t>()[iy / I4];
+                const auto& x_s4_1 = vx_tuple[ix + I1].template AsType<int8x4_t>()[iy / I4];
+                const auto& x_s4_2 = vx_tuple[ix + I2].template AsType<int8x4_t>()[iy / I4];
+                const auto& x_s4_3 = vx_tuple[ix + I3].template AsType<int8x4_t>()[iy / I4];
+
+                // reference to 4 int8 data from vy_tuple
+                auto& y_s4_0 = vy_tuple(iy).template AsType<int8x4_t>()(ix / I4);
+                auto& y_s4_1 = vy_tuple(iy + I1).template AsType<int8x4_t>()(ix / I4);
+                auto& y_s4_2 = vy_tuple(iy + I2).template AsType<int8x4_t>()(ix / I4);
+                auto& y_s4_3 = vy_tuple(iy + I3).template AsType<int8x4_t>()(ix / I4);
+
+                // transpose
+                transpose_int8_4x4(x_s4_0, x_s4_1, x_s4_2, x_s4_3, y_s4_0, y_s4_1, y_s4_2, y_s4_3);
+            });
+        });
+    }
+};
+
+} // namespace ck
diff --git a/include/ck/utility/tuple.hpp b/include/ck/utility/tuple.hpp
new file mode 100644
index 0000000000..558d94e089
--- /dev/null
+++ b/include/ck/utility/tuple.hpp
@@ -0,0 +1,323 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/is_static.hpp"
+#include "ck/utility/print.hpp"
+#include "ck/utility/integral_constant.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/type.hpp"
+#include "ck/utility/enable_if.hpp"
+
+namespace ck {
+
+namespace detail {
+
+template <index_t>
+struct TupleElementKey
+{
+    __host__ __device__ constexpr TupleElementKey() = default;
+};
+
+template <typename Key, typename Data>
+struct TupleElementKeyData
+{
+    using DataType = Data;
+
+#if 0 // workaround compiler complaint about implicitly-deleted default constructor
+    __host__ __device__ constexpr TupleElementKeyData() = default;
+#else
+    __host__ __device__ constexpr TupleElementKeyData() : mData{} {}
+#endif
+
+    template <typename T,
+              typename enable_if<!is_same<remove_cvref_t<T>, TupleElementKeyData>::value,
+                                 bool>::type = false>
+    __host__ __device__ constexpr TupleElementKeyData(T&& v) : mData(std::forward<T>(v))
+    {
+    }
+
+    DataType mData;
+};
+
+// for read access of tuple element
+template <typename Key, typename Data>
+__host__ __device__ constexpr const Data&
+get_tuple_element_data_reference(const TupleElementKeyData<Key, Data>& x)
+{
+    return static_cast<const Data&>(x.mData);
+}
+
+// for write access of tuple element
+template <typename Key, typename Data>
+__host__ __device__ constexpr Data&
+get_tuple_element_data_reference(TupleElementKeyData<Key, Data>& x)
+{
+    return x.mData;
+}
+
+// TODO: not sure the use of reference is correct
+template <typename Key, typename Data>
+__host__ __device__ constexpr Data&&
+get_tuple_element_data_reference(TupleElementKeyData<Key, Data>&& x)
+{
+    return static_cast<Data&&>(x.mData);
+}
+
+// for infering type of tuple element
+template <typename Key, typename Data>
+__host__ __device__ constexpr Data get_tuple_element_data(const TupleElementKeyData<Key, Data>& x)
+{
+    return std::forward(x.mData);
+}
+
+template <typename Indices, typename... Xs>
+struct TupleImpl;
+
+template <index_t... Is, typename... Xs>
+struct TupleImpl<Sequence<Is...>, Xs...> : TupleElementKeyData<TupleElementKey<Is>, Xs>...
+{
+    __host__ __device__ constexpr TupleImpl() = default;
+
+    template <typename Y,
+              typename enable_if<sizeof...(Is) == 1 && sizeof...(Xs) == 1 &&
+                                     !is_same<remove_cvref_t<Y>, TupleImpl>::value,
+                                 bool>::type = false>
+    __host__ __device__ constexpr TupleImpl(Y&& y)
+        : TupleElementKeyData<TupleElementKey<Is>, Xs>(std::forward<Y>(y))...
+    {
+    }
+
+    template <typename... Ys, typename enable_if<sizeof...(Ys) >= 2, bool>::type = false>
+    __host__ __device__ constexpr TupleImpl(Ys&&... ys)
+        : TupleElementKeyData<TupleElementKey<Is>, Xs>(std::forward<Ys>(ys))...
+    {
+        static_assert(sizeof...(Is) == sizeof...(Xs) && sizeof...(Is) == sizeof...(Ys),
+                      "wrong! inconsistent size");
+    }
+
+    __host__ __device__ static constexpr index_t Size() { return sizeof...(Xs); }
+
+    template <index_t I>
+    __host__ __device__ constexpr const auto& GetElementDataByKey(TupleElementKey<I>) const
+    {
+        return get_tuple_element_data_reference<TupleElementKey<I>>(*this);
+    }
+
+    template <index_t I>
+    __host__ __device__ constexpr auto& GetElementDataByKey(TupleElementKey<I>)
+    {
+        return get_tuple_element_data_reference<TupleElementKey<I>>(*this);
+    }
+};
+
+} // namespace detail
+
+template <typename... Xs>
+struct Tuple : detail::TupleImpl<typename arithmetic_sequence_gen<0, sizeof...(Xs), 1>::type, Xs...>
+{
+    using base =
+        detail::TupleImpl<typename arithmetic_sequence_gen<0, sizeof...(Xs), 1>::type, Xs...>;
+
+    __host__ __device__ constexpr Tuple() = default;
+
+    template <typename Y,
+              typename enable_if<sizeof...(Xs) == 1 && !is_same<remove_cvref_t<Y>, Tuple>::value,
+                                 bool>::type = false>
+    __host__ __device__ constexpr Tuple(Y&& y) : base(std::forward<Y>(y))
+    {
+    }
+
+    template <typename... Ys,
+              typename enable_if<sizeof...(Ys) == sizeof...(Xs) && sizeof...(Ys) >= 2, bool>::type =
+                  false>
+    __host__ __device__ constexpr Tuple(Ys&&... ys) : base(std::forward<Ys>(ys)...)
+    {
+    }
+
+    __host__ __device__ static constexpr index_t Size() { return sizeof...(Xs); }
+
+    // read access
+    template <index_t I>
+    __host__ __device__ constexpr const auto& At() const
+    {
+        static_assert(I < base::Size(), "wrong! out of range");
+        return base::GetElementDataByKey(detail::TupleElementKey<I>{});
+    }
+
+    // write access
+    template <index_t I>
+    __host__ __device__ constexpr auto& At()
+    {
+        static_assert(I < base::Size(), "wrong! out of range");
+        return base::GetElementDataByKey(detail::TupleElementKey<I>{});
+    }
+
+    // read access
+    template <index_t I>
+    __host__ __device__ constexpr const auto& At(Number<I>) const
+    {
+        static_assert(I < base::Size(), "wrong! out of range");
+        return base::GetElementDataByKey(detail::TupleElementKey<I>{});
+    }
+
+    // write access
+    template <index_t I>
+    __host__ __device__ constexpr auto& At(Number<I>)
+    {
+        static_assert(I < base::Size(), "wrong! out of range");
+        return base::GetElementDataByKey(detail::TupleElementKey<I>{});
+    }
+
+    // read access
+    template <index_t I>
+    __host__ __device__ constexpr const auto& operator[](Number<I> i) const
+    {
+        return At(i);
+    }
+
+    // write access
+    template <index_t I>
+    __host__ __device__ constexpr auto& operator()(Number<I> i)
+    {
+        return At(i);
+    }
+
+    // WARNING: needed by compiler for C++ structured binding support only, don't use this function!
+    template <std::size_t I>
+    __host__ __device__ constexpr const auto& get() const
+    {
+        return this->template At<I>();
+    }
+
+    // WARNING: needed bu compiler for C++ structured binding support only, don't use this function!
+    template <std::size_t I>
+    __host__ __device__ constexpr auto& get()
+    {
+        return this->template At<I>();
+    }
+
+    template <typename T>
+    __host__ __device__ constexpr auto operator=(const T& a)
+    {
+        static_assert(T::Size() == Size(), "wrong! size not the same");
+
+        static_for<0, Size(), 1>{}([&](auto i) { operator()(i) = a[i]; });
+
+        return *this;
+    }
+
+    __host__ __device__ static constexpr bool IsStatic()
+    {
+        bool flag = true;
+
+        static_for<0, sizeof...(Xs), 1>{}([&flag](auto i) {
+            flag &= is_static_v<remove_cvref_t<type_pack_element<i.value, Xs...>>>;
+        });
+
+        return flag;
+    }
+
+    // FIXME: remove
+    __host__ __device__ static constexpr bool IsStaticBuffer() { return true; }
+
+    __host__ __device__ void Print() const
+    {
+        printf("Tuple{size: %d, data: [", static_cast<index_t>(Size()));
+
+        static_for<0, Size(), 1>{}([&](auto i) {
+            print(At(i));
+
+            if(i < Size() - 1)
+            {
+                printf(", ");
+            }
+        });
+
+        printf("]}");
+    }
+};
+
+template <>
+struct Tuple<>
+{
+    __host__ __device__ constexpr Tuple() = default;
+
+    __host__ __device__ static constexpr index_t Size() { return 0; }
+
+    template <typename T>
+    __host__ __device__ constexpr auto operator=(const T&)
+    {
+        return *this;
+    }
+
+    __host__ __device__ static constexpr bool IsStatic() { return true; }
+
+    // FIXME: remove
+    __host__ __device__ static constexpr bool IsStaticBuffer() { return true; }
+
+    __host__ __device__ void Print() const { printf("Tuple{size: 0, data: []}"); }
+};
+
+template <typename... Xs>
+__host__ __device__ constexpr bool operator==(const Tuple<Xs...>& a, const Tuple<Xs...>& b)
+{
+    bool same = true;
+
+    static_for<0, sizeof...(Xs), 1>{}([&](auto i) {
+        if(a[i] != b[i])
+        {
+            same = false;
+        }
+    });
+
+    return same;
+}
+
+template <typename... Xs>
+__host__ __device__ constexpr bool operator!=(const Tuple<Xs...>& a, const Tuple<Xs...>& b)
+{
+    return !(a == b);
+}
+
+template <index_t I, typename TTuple>
+struct tuple_element
+{
+    // type should keep the cv/ref qualifier of original tuple element
+    using type = decltype(detail::get_tuple_element_data<detail::TupleElementKey<I>>(TTuple{}));
+};
+
+template <index_t I, typename TTuple>
+using tuple_element_t = typename tuple_element<I, TTuple>::type;
+
+template <typename... Xs>
+__host__ __device__ constexpr auto make_tuple(Xs&&... xs)
+{
+    return Tuple<remove_cvref_t<Xs>...>(std::forward<Xs>(xs)...);
+}
+
+// https://en.cppreference.com/w/cpp/utility/tuple/tie
+template <typename... Args>
+constexpr Tuple<Args&...> tie(Args&... args) noexcept
+{
+    return {args...};
+}
+
+} // namespace ck
+
+namespace std {
+
+// WARNING: needed by compiler for C++ structured binding support only, don't use this
+template <typename... Ts>
+struct tuple_size<ck::Tuple<Ts...>> : std::integral_constant<std::size_t, sizeof...(Ts)>
+{
+};
+
+// WARNING: needed by compiler for C++ structured binding support only, don't use this
+template <std::size_t I, typename... Ts>
+struct tuple_element<I, ck::Tuple<Ts...>> : ck::tuple_element<I, ck::Tuple<Ts...>>
+{
+};
+
+} // namespace std
diff --git a/include/ck/utility/tuple_helper.hpp b/include/ck/utility/tuple_helper.hpp
new file mode 100644
index 0000000000..165d43b389
--- /dev/null
+++ b/include/ck/utility/tuple_helper.hpp
@@ -0,0 +1,128 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "functional4.hpp"
+#include "tuple.hpp"
+
+namespace ck {
+
+template <typename F, index_t N>
+__host__ __device__ constexpr auto generate_tuple(F&& f, Number<N>)
+{
+    return unpack([&f](auto&&... is) { return make_tuple(f(is)...); },
+                  typename arithmetic_sequence_gen<0, N, 1>::type{});
+}
+
+template <typename F, index_t N>
+__host__ __device__ constexpr auto generate_tie(F&& f, Number<N>)
+{
+    return unpack([&f](auto&&... is) { return tie(f(is)...); },
+                  typename arithmetic_sequence_gen<0, N, 1>::type{});
+}
+
+// tx and ty are tuple of references, return type of will tuple of referennce (not rvalue)
+template <typename... X, typename... Y>
+__host__ __device__ constexpr auto concat_tuple_of_reference(const Tuple<X&...>& tx,
+                                                             const Tuple<Y&...>& ty)
+{
+    return unpack2(
+        [&](auto&&... zs) { return Tuple<decltype(zs)...>{std::forward<decltype(zs)>(zs)...}; },
+        tx,
+        ty);
+}
+
+namespace detail {
+
+template <typename F, typename X, index_t... Is>
+__host__ __device__ constexpr auto transform_tuples_impl(F f, const X& x, Sequence<Is...>)
+{
+    return make_tuple(f(x.At(Number<Is>{}))...);
+}
+
+template <typename F, typename X, typename Y, index_t... Is>
+__host__ __device__ constexpr auto
+transform_tuples_impl(F f, const X& x, const Y& y, Sequence<Is...>)
+{
+    return make_tuple(f(x.At(Number<Is>{}), y.At(Number<Is>{}))...);
+}
+
+template <typename F, typename X, typename Y, typename Z, index_t... Is>
+__host__ __device__ constexpr auto
+transform_tuples_impl(F f, const X& x, const Y& y, const Z& z, Sequence<Is...>)
+{
+    return make_tuple(f(x.At(Number<Is>{}), y.At(Number<Is>{}), z.At(Number<Is>{}))...);
+}
+
+} // namespace detail
+
+template <typename F, typename X>
+__host__ __device__ constexpr auto transform_tuples(F f, const X& x)
+{
+    return detail::transform_tuples_impl(
+        f, x, typename arithmetic_sequence_gen<0, X::Size(), 1>::type{});
+}
+
+template <typename F, typename X, typename Y>
+__host__ __device__ constexpr auto transform_tuples(F f, const X& x, const Y& y)
+{
+    return detail::transform_tuples_impl(
+        f, x, y, typename arithmetic_sequence_gen<0, X::Size(), 1>::type{});
+}
+
+template <typename F, typename X, typename Y, typename Z>
+__host__ __device__ constexpr auto transform_tuples(F f, const X& x, const Y& y, const Z& z)
+{
+    return detail::transform_tuples_impl(
+        f, x, y, z, typename arithmetic_sequence_gen<0, X::Size(), 1>::type{});
+}
+
+} // namespace ck
+
+// Macro function
+// convert constexpr Array to Tuple of Number
+#define TO_TUPLE_OF_NUMBER(arr, n)                                                              \
+    [&arr, &n] {                                                                                \
+        static_assert(arr.Size() >= n, "wrong! out of bound");                                  \
+                                                                                                \
+        static_assert(n < 7, "not implemented");                                                \
+                                                                                                \
+        if constexpr(n == 0)                                                                    \
+        {                                                                                       \
+            return ck::Tuple<>{};                                                               \
+        }                                                                                       \
+        else if constexpr(n == 1)                                                               \
+        {                                                                                       \
+            return ck::Tuple<Number<arr[0]>>{};                                                 \
+        }                                                                                       \
+        else if constexpr(n == 2)                                                               \
+        {                                                                                       \
+            return ck::Tuple<Number<arr[0]>, Number<arr[1]>>{};                                 \
+        }                                                                                       \
+        else if constexpr(n == 3)                                                               \
+        {                                                                                       \
+            return ck::Tuple<Number<arr[0]>, Number<arr[1]>, Number<arr[2]>>{};                 \
+        }                                                                                       \
+        else if constexpr(n == 4)                                                               \
+        {                                                                                       \
+            return ck::Tuple<Number<arr[0]>, Number<arr[1]>, Number<arr[2]>, Number<arr[3]>>{}; \
+        }                                                                                       \
+        else if constexpr(n == 5)                                                               \
+        {                                                                                       \
+            return ck::Tuple<Number<arr[0]>,                                                    \
+                             Number<arr[1]>,                                                    \
+                             Number<arr[2]>,                                                    \
+                             Number<arr[3]>,                                                    \
+                             Number<arr[4]>>{};                                                 \
+        }                                                                                       \
+        else if constexpr(n == 6)                                                               \
+        {                                                                                       \
+            return ck::Tuple<Number<arr[0]>,                                                    \
+                             Number<arr[1]>,                                                    \
+                             Number<arr[2]>,                                                    \
+                             Number<arr[3]>,                                                    \
+                             Number<arr[4]>,                                                    \
+                             Number<arr[5]>>{};                                                 \
+        }                                                                                       \
+    }()
diff --git a/include/ck/utility/tuple_of_sequence_to_array_of_array.hpp b/include/ck/utility/tuple_of_sequence_to_array_of_array.hpp
new file mode 100644
index 0000000000..48ce41a5f5
--- /dev/null
+++ b/include/ck/utility/tuple_of_sequence_to_array_of_array.hpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "sequence.hpp"
+#include "array.hpp"
+#include "tuple.hpp"
+
+namespace ck {
+
+template <typename... Seqs>
+__host__ __device__ constexpr auto to_array_of_array(Tuple<Seqs...> t_of_s)
+{
+    constexpr index_t n0 = sizeof...(Seqs);
+
+    constexpr index_t max_n1 = [&] {
+        index_t max_n1_ = 0;
+
+        static_for<0, n0, 1>{}([&](auto i0) {
+            constexpr index_t n1 = t_of_s[i0].Size();
+
+            max_n1_ = max_n1_ < n1 ? n1 : max_n1_;
+        });
+
+        return max_n1_;
+    }();
+
+    Array<Array<index_t, max_n1>, n0> a_of_a{{-1}};
+
+    static_for<0, n0, 1>{}([&](auto i0) {
+        constexpr index_t n1 = t_of_s[i0].Size();
+
+        static_for<0, n1, 1>{}([&](auto i1) { a_of_a(i0)(i1) = t_of_s[i0][i1]; });
+    });
+
+    return a_of_a;
+}
+
+} // namespace ck
diff --git a/include/ck/utility/type.hpp b/include/ck/utility/type.hpp
new file mode 100644
index 0000000000..fda96a248f
--- /dev/null
+++ b/include/ck/utility/type.hpp
@@ -0,0 +1,35 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/integral_constant.hpp"
+#include "ck/utility/enable_if.hpp"
+
+namespace ck {
+
+// is_same
+static_assert(__has_builtin(__is_same), "");
+
+template <typename X, typename Y>
+using is_same = integral_constant<bool, __is_same(X, Y)>;
+
+template <typename X, typename Y>
+inline constexpr bool is_same_v = is_same<X, Y>::value;
+
+static_assert(__has_builtin(__type_pack_element), "");
+
+// type_pack_element
+template <index_t I, typename... Ts>
+using type_pack_element = __type_pack_element<I, Ts...>;
+
+// is_pointer
+template <typename T>
+inline constexpr bool is_pointer_v = std::is_pointer<T>::value;
+
+// is_empty
+template <typename T>
+inline constexpr bool is_empty_v = std::is_empty<T>::value;
+
+} // namespace ck
diff --git a/include/ck/utility/type_convert.hpp b/include/ck/utility/type_convert.hpp
new file mode 100644
index 0000000000..65d8940377
--- /dev/null
+++ b/include/ck/utility/type_convert.hpp
@@ -0,0 +1,212 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/f8_utils.hpp"
+#include "ck/utility/random_gen.hpp"
+
+namespace ck {
+
+// Convert X to Y
+template <typename Y, typename X>
+__host__ __device__ constexpr Y type_convert(X x)
+{
+    static_assert(!std::is_reference_v<Y> && !std::is_reference_v<X>);
+
+    return static_cast<Y>(x);
+}
+
+// convert bfp16 to fp32
+template <>
+inline __host__ __device__ constexpr float type_convert<float, bhalf_t>(bhalf_t x)
+{
+    union
+    {
+        uint32_t int32;
+        float fp32;
+    } u = {uint32_t(x) << 16};
+
+    return u.fp32;
+}
+
+// convert fp32 to bfp16
+template <>
+inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, float>(float x)
+{
+    union
+    {
+        float fp32;
+        uint32_t int32;
+    } u = {x};
+
+    return uint16_t(u.int32 >> 16);
+}
+
+// convert bfp16 to fp16 via fp32
+template <>
+inline __host__ __device__ constexpr half_t type_convert<half_t, bhalf_t>(bhalf_t x)
+{
+    float x_fp32 = type_convert<float>(x);
+
+    return static_cast<half_t>(x_fp32);
+}
+
+// convert fp16 to bfp16 via fp32
+template <>
+inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, half_t>(half_t x)
+{
+    float x_fp32 = static_cast<float>(x);
+
+    return type_convert<bhalf_t>(x_fp32);
+}
+
+// convert bfp16 to int8 via fp32
+template <>
+inline __host__ __device__ constexpr int8_t type_convert<int8_t, bhalf_t>(bhalf_t x)
+{
+    float x_fp32 = type_convert<float>(x);
+
+    return static_cast<int8_t>(x_fp32);
+}
+
+// convert int8 to bfp16 via fp32
+template <>
+inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int8_t>(int8_t x)
+{
+    float x_fp32 = static_cast<float>(x);
+
+    return type_convert<bhalf_t>(x_fp32);
+}
+
+// convert fp32 to fp8
+template <>
+inline __host__ __device__ f8_t type_convert<f8_t, float>(float x)
+{
+    constexpr bool negative_zero_nan = true;
+    constexpr bool clip              = true;
+    constexpr f8_rounding_mode rm    = f8_rounding_mode::standard;
+    constexpr uint32_t rng           = 0;
+    return utils::cast_to_f8<float, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
+        x, rng);
+}
+
+// convert fp8 to fp32
+template <>
+inline __host__ __device__ float type_convert<float, f8_t>(f8_t x)
+{
+    constexpr bool negative_zero_nan = true;
+    return utils::cast_from_f8<float, negative_zero_nan>(x);
+}
+
+// convert fp16 to fp8
+template <>
+inline __host__ __device__ f8_t type_convert<f8_t, half_t>(half_t x)
+{
+    constexpr bool negative_zero_nan = true;
+    constexpr bool clip              = true;
+    constexpr f8_rounding_mode rm    = f8_rounding_mode::standard;
+    constexpr uint32_t rng           = 0;
+    return utils::cast_to_f8<half_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
+        x, rng);
+}
+
+// convert fp8 to fp16
+template <>
+inline __host__ __device__ half_t type_convert<half_t, f8_t>(f8_t x)
+{
+    constexpr bool negative_zero_nan = true;
+    return utils::cast_from_f8<half_t, negative_zero_nan>(x);
+}
+
+// Declare a template function for bf16 conversion using RTN
+template <typename Y, typename X>
+__host__ __device__ constexpr Y bf16_convert_rtn(X x);
+
+// Convert fp32 to bf16 with RTN if higher precision is needed
+template <>
+inline __host__ __device__ constexpr bhalf_t bf16_convert_rtn<bhalf_t, float>(float x)
+{
+    union
+    {
+        float fp32;
+        uint32_t int32;
+    } u = {x};
+
+    // When the exponent bits are not all 1s, then the value is zero, normal,
+    // or subnormal. We round the bfloat16 mantissa up by adding 0x7FFF, plus
+    // 1 if the least significant bit of the bfloat16 mantissa is 1 (odd).
+    // This causes the bfloat16's mantissa to be incremented by 1 if the 16
+    // least significant bits of the float mantissa are greater than 0x8000,
+    // or if they are equal to 0x8000 and the least significant bit of the
+    // bfloat16 mantissa is 1 (odd). This causes it to be rounded to even when
+    // the lower 16 bits are exactly 0x8000. If the bfloat16 mantissa already
+    // has the value 0x7f, then incrementing it causes it to become 0x00 and
+    // the exponent is incremented by one, which is the next higher FP value
+    // to the unrounded bfloat16 value. When the bfloat16 value is subnormal
+    // with an exponent of 0x00 and a mantissa of 0x7f, it may be rounded up
+    // to a normal value with an exponent of 0x01 and a mantissa of 0x00.
+    // When the bfloat16 value has an exponent of 0xFE and a mantissa of 0x7F,
+    // incrementing it causes it to become an exponent of 0xFF and a mantissa
+    // of 0x00, which is Inf, the next higher value to the unrounded value.
+    bool flag0 = ~u.int32 & 0x7f800000;
+
+    // When all of the exponent bits are 1, the value is Inf or NaN.
+    // Inf is indicated by a zero mantissa. NaN is indicated by any nonzero
+    // mantissa bit. Quiet NaN is indicated by the most significant mantissa
+    // bit being 1. Signaling NaN is indicated by the most significant
+    // mantissa bit being 0 but some other bit(s) being 1. If any of the
+    // lower 16 bits of the mantissa are 1, we set the least significant bit
+    // of the bfloat16 mantissa, in order to preserve signaling NaN in case
+    // the bfloat16's mantissa bits are all 0.
+    bool flag1 = !flag0 && (u.int32 & 0xffff);
+
+    u.int32 += flag0 ? 0x7fff + ((u.int32 >> 16) & 1) : 0; // Round to nearest, round to even
+    u.int32 |= flag1 ? 0x10000 : 0x0;                      // Preserve signaling NaN
+
+    return uint16_t(u.int32 >> 16);
+}
+
+// convert fp16 to bfp16 via fp32 with RTN if higher precision is needed
+template <>
+inline __host__ __device__ constexpr bhalf_t bf16_convert_rtn<bhalf_t, half_t>(half_t x)
+{
+    float x_fp32 = static_cast<float>(x);
+
+    return bf16_convert_rtn<bhalf_t>(x_fp32);
+}
+
+// Declare a template function for fp8 conversion using SR
+template <typename Y, typename X>
+__host__ __device__ constexpr Y f8_convert_sr(X x);
+
+// convert fp32 to fp8 with stochastic rounding
+template <>
+inline __host__ __device__ f8_t f8_convert_sr<f8_t, float>(float x)
+{
+    constexpr bool negative_zero_nan = true;
+    constexpr bool clip              = true;
+    constexpr f8_rounding_mode rm    = f8_rounding_mode::stochastic;
+    constexpr int seed               = 42;
+    // as thread id is not available on host, use 0 for prn generation
+    uint32_t rng = prand_generator<float, seed>(reinterpret_cast<uintptr_t>(&x), x);
+    return utils::cast_to_f8<float, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
+        x, rng);
+}
+
+// convert fp16 to fp8 with stochastic rounding
+template <>
+inline __host__ __device__ f8_t f8_convert_sr<f8_t, half_t>(half_t x)
+{
+    constexpr bool negative_zero_nan = true;
+    constexpr bool clip              = true;
+    constexpr f8_rounding_mode rm    = f8_rounding_mode::stochastic;
+    constexpr int seed               = 42;
+    // as thread id is not available on host, use 0 for prn generation
+    uint32_t rng = prand_generator<half_t, seed>(reinterpret_cast<uintptr_t>(&x), x);
+    return utils::cast_to_f8<half_t, negative_zero_nan, clip, (rm == f8_rounding_mode::stochastic)>(
+        x, rng);
+}
+
+} // namespace ck
diff --git a/include/ck/utility/workgroup_barrier.hpp b/include/ck/utility/workgroup_barrier.hpp
new file mode 100644
index 0000000000..ec9151fd1b
--- /dev/null
+++ b/include/ck/utility/workgroup_barrier.hpp
@@ -0,0 +1,73 @@
+#pragma once
+#include <hip/hip_runtime.h>
+#include <stdint.h>
+
+namespace ck {
+struct workgroup_barrier
+{
+    __device__ workgroup_barrier(uint32_t* ptr) : base_ptr(ptr) {}
+
+    __device__ uint32_t ld(uint32_t offset)
+    {
+#if 0
+        float d = llvm_amdgcn_raw_buffer_load_fp32(
+                        amdgcn_make_buffer_resource(base_ptr),
+                        0,
+                        offset,
+                        AMDGCN_BUFFER_GLC);
+        union cvt {
+            float f32;
+            uint32_t u32;
+        };
+        cvt x;
+        x.f32 = d;
+        return x.u32;
+#endif
+        return __atomic_load_n(base_ptr + offset, __ATOMIC_RELAXED);
+    }
+
+    __device__ void wait_eq(uint32_t offset, uint32_t value)
+    {
+        if(threadIdx.x == 0)
+        {
+            while(ld(offset) != value) {}
+        }
+        __syncthreads();
+    }
+
+    __device__ void wait_lt(uint32_t offset, uint32_t value)
+    {
+        if(threadIdx.x == 0)
+        {
+            while(ld(offset) < value) {}
+        }
+        __syncthreads();
+    }
+
+    __device__ void wait_set(uint32_t offset, uint32_t compare, uint32_t value)
+    {
+        if(threadIdx.x == 0)
+        {
+            while(atomicCAS(base_ptr + offset, compare, value) != compare) {}
+        }
+        __syncthreads();
+    }
+
+    // enter critical zoon, assume buffer is zero when launch kernel
+    __device__ void aquire(uint32_t offset) { wait_set(offset, 0, 1); }
+
+    // exit critical zoon, assume buffer is zero when launch kernel
+    __device__ void release(uint32_t offset) { wait_set(offset, 1, 0); }
+
+    __device__ void inc(uint32_t offset)
+    {
+        __syncthreads();
+        if(threadIdx.x == 0)
+        {
+            atomicAdd(base_ptr + offset, 1);
+        }
+    }
+
+    uint32_t* base_ptr;
+};
+} // namespace ck
diff --git a/include/ck/utility/workgroup_synchronization.hpp b/include/ck/utility/workgroup_synchronization.hpp
new file mode 100644
index 0000000000..24858fdbdc
--- /dev/null
+++ b/include/ck/utility/workgroup_synchronization.hpp
@@ -0,0 +1,74 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+#include "ck/host_utility/hip_check_error.hpp"
+
+namespace ck {
+
+// Initialization flag of Barrier object, can be any value except for zero
+static constexpr int BarrierInitFlag = 0x7856;
+
+// 1) only the first thread-block in the synchronizaton group is supposed to call this function. It
+// is the responsibility of the user to ensure the two integer values in p_control_bits are zeros
+// before calling gms_init().
+// 2) Aftercalling gms_reset(), the two integer values in p_control_bits will be zeros, so no
+// repetitious initialization of p_control_bits buffer is required
+static __device__ void gms_init(int NumWarps, int* p_control_bits)
+{
+    union
+    {
+        int two32[2];
+        unsigned long one64;
+    } regs;
+
+    regs.two32[0] = BarrierInitFlag;
+    regs.two32[1] = NumWarps;
+
+    if(threadIdx.x == 0)
+        atomicCAS(reinterpret_cast<unsigned long*>(p_control_bits), 0, regs.one64);
+};
+
+// all the workgroups in the synchronization group is supposed to call this function
+static __device__ void gms_barrier(int* p_control_bits)
+{
+    constexpr int mask = warpSize - 1;
+
+    if((threadIdx.x & mask) == 0)
+    {
+        // ensure the barrier object is initialized
+        do
+        {
+            const int r0 = __atomic_load_n(&p_control_bits[0], __ATOMIC_RELAXED);
+
+            if(r0 == BarrierInitFlag)
+                break;
+
+        } while(true);
+
+        // go ahead toward the barrier line
+        atomicSub(&p_control_bits[1], 1);
+
+        // wait until all warps have arrived
+        do
+        {
+            const int r1 = __atomic_load_n(&p_control_bits[1], __ATOMIC_RELAXED);
+
+            if(r1 == 0)
+                break;
+
+        } while(true);
+    };
+};
+
+// 1) Only the first thread-block in the synchronizaton group is supposed to call this function.
+// 2) Aftercalling gms_reset(), the two integer values in p_control_bits will be zeros, so no
+// repetitious initialization of p_control_bits buffer is required
+static __device__ void gms_reset(int* p_control_bits)
+{
+    // reset the barrier object
+    if(threadIdx.x == 0)
+        (void)atomicCAS(&p_control_bits[0], BarrierInitFlag, 0);
+};
+
+} // namespace ck
diff --git a/include/ck/version.h.in b/include/ck/version.h.in
new file mode 100644
index 0000000000..0d6a6512fb
--- /dev/null
+++ b/include/ck/version.h.in
@@ -0,0 +1,40 @@
+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+
+/* the configured version and settings for miopen- Composable Kernel */
+
+#ifndef CK_VERSION_H_
+#define CK_VERSION_H_
+
+// clang-format off
+#define CK_VERSION @CMAKE_PROJECT_VERSION@
+#define CK_VERSION_MAJOR @CMAKE_PROJECT_VERSION_MAJOR@
+#define CK_VERSION_MINOR @CMAKE_PROJECT_VERSION_MINOR@
+#define CK_VERSION_PATCH @CMAKE_PROJECT_VERSION_PATCH@
+#define CK_COMMIT_ID @COMMIT_ID@
+// clang-format on
+
+#endif
diff --git a/library/CMakeLists.txt b/library/CMakeLists.txt
new file mode 100644
index 0000000000..90873fdd14
--- /dev/null
+++ b/library/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_subdirectory(src/tensor_operation_instance/gpu)
+add_subdirectory(src/utility)
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp
new file mode 100644
index 0000000000..fa06e77560
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp
@@ -0,0 +1,354 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// dinput descriptor in [N, C, Do, Ho, Wo] order
+// doutput descriptor in [N, C, Di, Hi, Wi] order
+// phyiscal layout is irrelavent
+template <ck::index_t NDimSpatial,
+          typename DInDataType,
+          typename DOutDataType,
+          typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
+struct ReferenceAvgPoolBwd : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(Tensor<DInDataType>& dinput,
+                 const Tensor<DOutDataType>& doutput,
+                 std::vector<ck::index_t> window_spatial_lengths,
+                 std::vector<ck::index_t> window_strides,
+                 std::vector<ck::index_t> window_dilations,
+                 std::vector<ck::index_t> dinput_left_pads,
+                 std::vector<ck::index_t> dinput_right_pads)
+            : dinput_{dinput},
+              doutput_{doutput},
+              window_spatial_lengths_{window_spatial_lengths},
+              window_strides_{window_strides},
+              window_dilations_{window_dilations},
+              in_left_pads_{dinput_left_pads},
+              in_right_pads_{dinput_right_pads}
+        {
+        }
+
+        Tensor<DInDataType>& dinput_;
+        const Tensor<DOutDataType>& doutput_;
+
+        std::vector<ck::index_t> window_spatial_lengths_;
+        std::vector<index_t> window_strides_;
+        std::vector<index_t> window_dilations_;
+        std::vector<index_t> in_left_pads_;
+        std::vector<index_t> in_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceAvgPoolBwd::Argument;
+
+        template <ck::index_t NDimSpatial_,
+                  typename std::enable_if<NDimSpatial_ == 1, bool>::type = false>
+        float RunAvgPoolBwd(const Argument& arg)
+        {
+            // Let input = x, outpu = y
+            // shape of x = [10], y = [6]
+            // window_size = 5, pad = 0, stride = 1, dilation = 1
+            // Forward:
+            // y0 = 1/5 * (x0 + x1 + x2 + x3 + x4)
+            // y1 = 1/5 * (x1 + x2 + x3 + x4 + x5)
+            // ...
+            // y5 = 1/5 * (x5 + x6 + x7 + x8 + x9)
+            // y6 = 1/5 * (x6 + x7 + x8 + x9)
+            // ...
+            // y9 = 1/5 * (x9)
+
+            // Backward:
+            // shape of dy = [6], dx = [10]
+            // dx0 = 1/5 * dy0
+            // dx1 = 1/5 * (dy0 + dy1)
+            // dx2 = 1/5 * (dy0 + dy1 + dy2)
+            // ...
+            // dx4 = 1/5 * (dy0 + dy1 + dy2 + dy3 + dy4)
+            // dx5 = 1/5 * (dy1 + dy2 + dy3 + dy4 + dy5)
+            // ...
+            // dx9 = 1/5 * (dy5 + dy6 + dy7 + dy8 + dy9)
+
+            auto f_ncw = [&](auto n, auto c, auto wi) {
+                std::size_t X  = arg.window_spatial_lengths_[0];
+                std::size_t Wo = arg.doutput_.GetLengths()[2];
+
+                float v_acc = 0;
+
+                for(std::size_t x = 0; x < X; ++x)
+                {
+                    // Out_Position = (In_Position + pad - x * dilation) / stride
+                    auto w_tmp = static_cast<ck::long_index_t>(wi) +
+                                 static_cast<ck::long_index_t>(arg.in_left_pads_[0]) -
+                                 static_cast<ck::long_index_t>(x * arg.window_dilations_[0]);
+
+                    // Check the input pixel validity (in perspective of being affected by some
+                    // doutput pixel)
+                    if(w_tmp % arg.window_strides_[0] == 0)
+                    {
+                        auto wo = static_cast<ck::long_index_t>(w_tmp) /
+                                  static_cast<ck::long_index_t>(arg.window_strides_[0]);
+
+                        // Get the doutput pixel in valid range to accumulate the gradients for this
+                        // input pixel
+                        if(wo >= 0 && ck::type_convert<std::size_t>(wo) < Wo)
+                        {
+                            v_acc += ck::type_convert<float>(arg.doutput_(n, c, wo));
+                        }
+                    }
+                }
+
+                v_acc /= ck::type_convert<float>(X);
+                arg.dinput_(n, c, wi) = ck::type_convert<DInDataType>(v_acc);
+            };
+
+            make_ParallelTensorFunctor(f_ncw,
+                                       arg.dinput_.GetLengths()[0],
+                                       arg.dinput_.GetLengths()[1],
+                                       arg.dinput_.GetLengths()[2])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        template <ck::index_t NDimSpatial_,
+                  typename std::enable_if<NDimSpatial_ == 2, bool>::type = false>
+        float RunAvgPoolBwd(const Argument& arg)
+        {
+            auto f_nchw = [&](auto n, auto c, auto hi, auto wi) {
+                std::size_t Y = arg.window_spatial_lengths_[0];
+                std::size_t X = arg.window_spatial_lengths_[1];
+
+                std::size_t Ho = arg.doutput_.GetLengths()[2];
+                std::size_t Wo = arg.doutput_.GetLengths()[3];
+
+                float v_acc = 0;
+
+                for(std::size_t y = 0; y < Y; ++y)
+                {
+                    // Out_Position = (In_Position + pad - x * dilation) / stride
+                    auto h_tmp = static_cast<ck::long_index_t>(hi) +
+                                 static_cast<ck::long_index_t>(arg.in_left_pads_[0]) -
+                                 static_cast<ck::long_index_t>(y * arg.window_dilations_[0]);
+
+                    // Check the input pixel validity (in perspective of being affected by some
+                    // doutput pixel)
+                    if(h_tmp % arg.window_strides_[0] == 0)
+                    {
+                        auto ho = static_cast<ck::long_index_t>(h_tmp) /
+                                  static_cast<ck::long_index_t>(arg.window_strides_[0]);
+
+                        // Get the doutput pixel in valid range to accumulate the gradients for this
+                        // input pixel
+                        if(ho >= 0 && ck::type_convert<std::size_t>(ho) < Ho)
+                        {
+                            for(std::size_t x = 0; x < X; ++x)
+                            {
+                                auto w_tmp =
+                                    static_cast<ck::long_index_t>(wi) +
+                                    static_cast<ck::long_index_t>(arg.in_left_pads_[1]) -
+                                    static_cast<ck::long_index_t>(x * arg.window_dilations_[1]);
+                                if(w_tmp % arg.window_strides_[1] == 0)
+                                {
+                                    auto wo = static_cast<ck::long_index_t>(w_tmp) /
+                                              static_cast<ck::long_index_t>(arg.window_strides_[1]);
+                                    if(wo >= 0 && ck::type_convert<std::size_t>(wo) < Wo)
+                                    {
+                                        v_acc +=
+                                            ck::type_convert<float>(arg.doutput_(n, c, ho, wo));
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+
+                v_acc /= ck::type_convert<float>(Y * X);
+                arg.dinput_(n, c, hi, wi) = ck::type_convert<DInDataType>(v_acc);
+            };
+
+            make_ParallelTensorFunctor(f_nchw,
+                                       arg.dinput_.GetLengths()[0],
+                                       arg.dinput_.GetLengths()[1],
+                                       arg.dinput_.GetLengths()[2],
+                                       arg.dinput_.GetLengths()[3])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        template <ck::index_t NDimSpatial_,
+                  typename std::enable_if<NDimSpatial_ == 3, bool>::type = false>
+        float RunAvgPoolBwd(const Argument& arg)
+        {
+            auto f_ncdhw = [&](auto n, auto c, auto di, auto hi, auto wi) {
+                std::size_t Z = arg.window_spatial_lengths_[0];
+                std::size_t Y = arg.window_spatial_lengths_[1];
+                std::size_t X = arg.window_spatial_lengths_[2];
+
+                std::size_t Do = arg.doutput_.GetLengths()[2];
+                std::size_t Ho = arg.doutput_.GetLengths()[3];
+                std::size_t Wo = arg.doutput_.GetLengths()[4];
+
+                float v_acc = 0;
+
+                for(std::size_t z = 0; z < Z; ++z)
+                {
+                    // Out_Position = (In_Position + pad - x * dilation) / stride
+                    auto d_tmp = static_cast<ck::long_index_t>(di) +
+                                 static_cast<ck::long_index_t>(arg.in_left_pads_[0]) -
+                                 static_cast<ck::long_index_t>(z * arg.window_dilations_[0]);
+
+                    // Check the input pixel validity (in perspective of being affected by some
+                    // doutput pixel)
+                    if(d_tmp % arg.window_strides_[0] == 0)
+                    {
+                        auto do_ = static_cast<ck::long_index_t>(d_tmp) /
+                                   static_cast<ck::long_index_t>(arg.window_strides_[0]);
+
+                        // Get the doutput pixel in valid range to accumulate the gradients for this
+                        // input pixel
+                        if(do_ >= 0 && ck::type_convert<std::size_t>(do_) < Do)
+                        {
+                            for(std::size_t y = 0; y < Y; ++y)
+                            {
+                                auto h_tmp =
+                                    static_cast<ck::long_index_t>(hi) +
+                                    static_cast<ck::long_index_t>(arg.in_left_pads_[1]) -
+                                    static_cast<ck::long_index_t>(y * arg.window_dilations_[1]);
+                                if(h_tmp % arg.window_strides_[1] == 0)
+                                {
+                                    auto ho = static_cast<ck::long_index_t>(h_tmp) /
+                                              static_cast<ck::long_index_t>(arg.window_strides_[1]);
+                                    if(ho >= 0 && ck::type_convert<std::size_t>(ho) < Ho)
+                                    {
+                                        for(std::size_t x = 0; x < X; ++x)
+                                        {
+                                            auto w_tmp = static_cast<ck::long_index_t>(wi) +
+                                                         static_cast<ck::long_index_t>(
+                                                             arg.in_left_pads_[2]) -
+                                                         static_cast<ck::long_index_t>(
+                                                             x * arg.window_dilations_[2]);
+
+                                            if(w_tmp % arg.window_strides_[2] == 0)
+                                            {
+                                                auto wo = static_cast<ck::long_index_t>(w_tmp) /
+                                                          static_cast<ck::long_index_t>(
+                                                              arg.window_strides_[2]);
+                                                if(wo >= 0 &&
+                                                   ck::type_convert<std::size_t>(wo) < Wo)
+                                                {
+                                                    v_acc += ck::type_convert<float>(
+                                                        arg.doutput_(n, c, do_, ho, wo));
+                                                }
+                                            }
+                                        }
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+
+                v_acc /= ck::type_convert<float>(Z * Y * X);
+                arg.dinput_(n, c, di, hi, wi) = ck::type_convert<DInDataType>(v_acc);
+            };
+
+            make_ParallelTensorFunctor(f_ncdhw,
+                                       arg.dinput_.GetLengths()[0],
+                                       arg.dinput_.GetLengths()[1],
+                                       arg.dinput_.GetLengths()[2],
+                                       arg.dinput_.GetLengths()[3],
+                                       arg.dinput_.GetLengths()[4])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const Argument& arg)
+        {
+            if(!(arg.dinput_.GetNumOfDimension() == NDimSpatial + 2 &&
+                 arg.doutput_.GetNumOfDimension() == NDimSpatial + 2))
+            {
+                throw std::runtime_error("wrong! inconsistent dimension");
+            }
+
+            return RunAvgPoolBwd<NDimSpatial>(arg);
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(Tensor<DInDataType>& dinput,
+                             const Tensor<DOutDataType>& doutput,
+                             std::vector<ck::index_t> window_spatial_lengths,
+                             std::vector<ck::index_t> window_strides,
+                             std::vector<ck::index_t> window_dilations,
+                             std::vector<ck::index_t> dinput_left_pads,
+                             std::vector<ck::index_t> dinput_right_pads)
+    {
+        if(window_spatial_lengths.size() != NDimSpatial || window_strides.size() != NDimSpatial ||
+           window_dilations.size() != NDimSpatial || dinput_left_pads.size() != NDimSpatial ||
+           dinput_right_pads.size() != NDimSpatial)
+            throw std::runtime_error("dimension is incorrect");
+
+        return Argument{dinput,
+                        doutput,
+                        window_spatial_lengths,
+                        window_strides,
+                        window_dilations,
+                        dinput_left_pads,
+                        dinput_right_pads};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceAvgPoolBwd"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp
new file mode 100644
index 0000000000..a1b1e0d91b
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp
@@ -0,0 +1,138 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct ReferenceBatchedGemm : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_g_m_k,
+                 const Tensor<BDataType>& b_g_k_n,
+                 Tensor<CDataType>& c_g_m_n,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : a_g_m_k_{a_g_m_k},
+              b_g_k_n_{b_g_k_n},
+              c_g_m_n_{c_g_m_n},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_g_m_k_;
+        const Tensor<BDataType>& b_g_k_n_;
+        Tensor<CDataType>& c_g_m_n_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceBatchedGemm::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_gmk_gkn_gmn = [&](auto g, auto m, auto n) {
+                const int K = arg.a_g_m_k_.mDesc.GetLengths()[2];
+
+                AccDataType v_acc = 0;
+
+                for(int k = 0; k < K; ++k)
+                {
+                    ADataType v_a;
+                    BDataType v_b;
+
+                    arg.a_element_op_(v_a, arg.a_g_m_k_(g, m, k));
+                    arg.b_element_op_(v_b, arg.b_g_k_n_(g, k, n));
+
+                    v_acc +=
+                        ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
+                }
+
+                AccDataType v_c;
+
+                arg.c_element_op_(v_c, v_acc);
+
+                arg.c_g_m_n_(g, m, n) = ck::type_convert<CDataType>(v_c);
+            };
+
+            make_ParallelTensorFunctor(f_gmk_gkn_gmn,
+                                       arg.c_g_m_n_.mDesc.GetLengths()[0],
+                                       arg.c_g_m_n_.mDesc.GetLengths()[1],
+                                       arg.c_g_m_n_.mDesc.GetLengths()[2])(
+                std::thread::hardware_concurrency());
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_g_m_k,
+                             const Tensor<BDataType>& b_g_k_n,
+                             Tensor<CDataType>& c_g_m_n,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{a_g_m_k, b_g_k_n, c_g_m_n, a_element_op, b_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceBatchedGemm"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp
new file mode 100644
index 0000000000..a2eabdf5c1
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp
@@ -0,0 +1,412 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <array>
+#include <algorithm>
+#include <thread>
+
+#include "ck/utility/math_v2.hpp"
+#include "ck/utility/ignore.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename XDataType,
+          typename DxDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+struct ReferenceBatchNormBwd : public device::DeviceBatchNormBwd<XDataType,
+                                                                 DxDataType,
+                                                                 DyDataType,
+                                                                 AccDataType,
+                                                                 ScaleDataType,
+                                                                 DscaleDbiasDataType,
+                                                                 MeanVarDataType,
+                                                                 DyElementwiseOp,
+                                                                 Rank,
+                                                                 NumBatchNormReduceDim>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> xyLengths,
+                 const std::array<index_t, Rank> xStrides,
+                 const std::array<index_t, Rank> dxStrides,
+                 const std::array<index_t, Rank> dyStrides,
+                 const std::array<int, NumBatchNormReduceDim> reduceDims,
+                 const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                 const std::array<index_t, NumInvariantDim> bnScaleStrides,
+                 const std::array<index_t, NumInvariantDim> bnDscaleDbiasStrides,
+                 const std::array<index_t, NumInvariantDim> bnMeanVarStrides,
+                 const XDataType* p_x,
+                 const DyDataType* p_dy,
+                 const ScaleDataType* p_scale,
+                 const MeanVarDataType* p_savedMean,
+                 const MeanVarDataType* p_savedInvVar,
+                 double epsilon,
+                 const DyElementwiseOp dy_elementwise_op,
+                 DxDataType* p_dx,
+                 DscaleDbiasDataType* p_dscale,
+                 DscaleDbiasDataType* p_dbias)
+            : reduceDims_(reduceDims),
+              bnScaleBiasMeanVarLengths_(bnScaleBiasMeanVarLengths),
+              bnScaleStrides_(bnScaleStrides),
+              bnDscaleDbiasStrides_(bnDscaleDbiasStrides),
+              bnMeanVarStrides_(bnMeanVarStrides),
+              p_x_(p_x),
+              p_dy_(p_dy),
+              p_scale_(p_scale),
+              p_savedMean_(p_savedMean),
+              p_savedInvVar_(p_savedInvVar),
+              dy_elementwise_op_(dy_elementwise_op),
+              p_dx_(p_dx),
+              p_dscale_(p_dscale),
+              p_dbias_(p_dbias)
+        {
+            using ck::host_common::get_index_set;
+
+            if(std::any_of(
+                   reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+                throw std::runtime_error("Invalid reduce dimensions!");
+
+            // get invariant_dims[] and invariant_lengths[]
+            for(int dim = 0, i = 0; dim < Rank; dim++)
+                if(std::none_of(
+                       reduceDims.begin(), reduceDims.end(), [&](int d) { return d == dim; }))
+                {
+                    invariantDims_[i]     = dim;
+                    invariant_lengths_[i] = xyLengths[dim];
+                    i++;
+                };
+
+            // get reduce_lengths_[]
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim              = reduceDims[j];
+                reduce_lengths_[i++] = xyLengths[dim];
+            };
+
+            for(int i = 0; i < NumInvariantDim; i++)
+                if(invariant_lengths_[i] != bnScaleBiasMeanVarLengths_[i])
+                    throw std::runtime_error("Invalid lengths parameters!");
+
+            for(int j = 0, i = 0; j < NumInvariantDim; j++)
+            {
+                int dim                  = invariantDims_[j];
+                x_invariant_strides_[i]  = xStrides[dim];
+                dy_invariant_strides_[i] = dyStrides[dim];
+                dx_invariant_strides_[i] = dxStrides[dim];
+                i++;
+            };
+
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim               = reduceDims_[j];
+                x_reduce_strides_[i]  = xStrides[dim];
+                dy_reduce_strides_[i] = dyStrides[dim];
+                dx_reduce_strides_[i] = dxStrides[dim];
+                i++;
+            };
+
+            reduceSize_ = std::accumulate(
+                reduce_lengths_.begin(), reduce_lengths_.end(), 1, std::multiplies<size_t>{});
+
+            invariant_index_set_ = get_index_set<NumInvariantDim>(invariant_lengths_);
+            reduce_index_set_    = get_index_set<NumBatchNormReduceDim>(reduce_lengths_);
+
+            epsilon_ = type_convert<AccDataType>(epsilon);
+
+            haveSavedMeanInvVar_ = (p_savedMean != nullptr && p_savedInvVar != nullptr);
+        }
+
+        std::array<int, NumBatchNormReduceDim> reduceDims_;
+        std::array<int, NumInvariantDim> invariantDims_;
+        std::array<index_t, NumInvariantDim> invariant_lengths_;
+        std::array<index_t, NumBatchNormReduceDim> reduce_lengths_;
+
+        const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths_;
+        const std::array<index_t, NumInvariantDim> bnScaleStrides_;
+        const std::array<index_t, NumInvariantDim> bnDscaleDbiasStrides_;
+        const std::array<index_t, NumInvariantDim> bnMeanVarStrides_;
+
+        std::array<index_t, NumInvariantDim> x_invariant_strides_;
+        std::array<index_t, NumInvariantDim> dy_invariant_strides_;
+        std::array<index_t, NumInvariantDim> dx_invariant_strides_;
+        std::array<index_t, NumBatchNormReduceDim> x_reduce_strides_;
+        std::array<index_t, NumBatchNormReduceDim> dy_reduce_strides_;
+        std::array<index_t, NumBatchNormReduceDim> dx_reduce_strides_;
+
+        const XDataType* p_x_;
+        const DyDataType* p_dy_;
+        const ScaleDataType* p_scale_;
+        const MeanVarDataType* p_savedMean_;
+        const MeanVarDataType* p_savedInvVar_;
+        const DyElementwiseOp dy_elementwise_op_;
+
+        DxDataType* p_dx_;
+        DscaleDbiasDataType* p_dscale_;
+        DscaleDbiasDataType* p_dbias_;
+
+        bool haveSavedMeanInvVar_;
+
+        std::vector<std::array<index_t, NumInvariantDim>> invariant_index_set_;
+        std::vector<std::array<index_t, NumBatchNormReduceDim>> reduce_index_set_;
+
+        AccDataType epsilon_;
+        size_t reduceSize_;
+    };
+
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            using ck::host_common::get_offset_from_index;
+
+            auto thread_reduce_func = [&](auto invariant_index) {
+                size_t x_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.x_invariant_strides_, invariant_index);
+                size_t dy_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.dy_invariant_strides_, invariant_index);
+                size_t dx_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.dx_invariant_strides_, invariant_index);
+
+                AccDataType mean     = type_convert<AccDataType>(0.0f);
+                AccDataType variance = type_convert<AccDataType>(0.0f);
+                AccDataType invVar;
+                int32_t curr_count = 0;
+
+                if(arg.haveSavedMeanInvVar_)
+                {
+                    size_t mean_invVar_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                        arg.bnMeanVarStrides_, invariant_index);
+
+                    mean =
+                        type_convert<AccDataType>(arg.p_savedMean_[mean_invVar_invariant_offset]);
+                    invVar =
+                        type_convert<AccDataType>(arg.p_savedInvVar_[mean_invVar_invariant_offset]);
+                }
+                else
+                {
+                    // compute mean, variance using welford method
+                    for(const auto& reduce_index : arg.reduce_index_set_)
+                    {
+                        size_t x_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                            arg.x_reduce_strides_, reduce_index);
+
+                        auto x_offset = x_invariant_offset + x_reduce_offset;
+
+                        curr_count++;
+
+                        AccDataType x = type_convert<AccDataType>(arg.p_x_[x_offset]);
+
+                        AccDataType delta = x - mean;
+
+                        mean += delta / curr_count;
+
+                        AccDataType delta2 = x - mean;
+
+                        variance += delta * delta2;
+                    };
+
+                    // actual variance
+                    variance = variance / curr_count;
+
+                    // inv-variance defined as 1/sqrt(epsilon+variance)
+                    invVar =
+                        type_convert<AccDataType>(1.0f) / ck::math::sqrt(arg.epsilon_ + variance);
+                };
+
+                AccDataType dbias =
+                    type_convert<AccDataType>(0.0f); // Sum on reduced dimensions of dy
+                AccDataType dscale =
+                    type_convert<AccDataType>(0.0f); // Sum on reduced dimensions of dy * norm_x
+
+                // 1) calculate dy * (x - mean) * inv-variance
+                // 2) calculate sum(dy) on reduced dimensions
+                // 3) calculate sum(dy * norm_x) on reduced dimensions
+                for(const auto& reduce_index : arg.reduce_index_set_)
+                {
+                    size_t x_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.x_reduce_strides_, reduce_index);
+                    size_t dy_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.dy_reduce_strides_, reduce_index);
+
+                    auto x_offset  = x_invariant_offset + x_reduce_offset;
+                    auto dy_offset = dy_invariant_offset + dy_reduce_offset;
+
+                    AccDataType x = type_convert<AccDataType>(arg.p_x_[x_offset]);
+
+                    AccDataType norm_x = (x - mean) * invVar;
+                    AccDataType dy     = type_convert<AccDataType>(arg.p_dy_[dy_offset]);
+
+                    arg.dy_elementwise_op_(dy, dy);
+
+                    dbias += dy;
+                    dscale += norm_x * dy;
+                };
+
+                size_t dscale_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.bnDscaleDbiasStrides_, invariant_index);
+                size_t dbias_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.bnDscaleDbiasStrides_, invariant_index);
+
+                arg.p_dscale_[dscale_offset] = type_convert<DscaleDbiasDataType>(dscale);
+                arg.p_dbias_[dbias_offset]   = type_convert<DscaleDbiasDataType>(dbias);
+
+                size_t scale_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnScaleStrides_, invariant_index);
+
+                AccDataType scale = type_convert<AccDataType>(arg.p_scale_[scale_offset]);
+
+                AccDataType multiplier = type_convert<AccDataType>(1.0f) /
+                                         type_convert<AccDataType>(arg.reduceSize_) * invVar *
+                                         scale;
+
+                // 1) calculate tmp = dscale * (x - mean) * inv-variance
+                // 2) calculate dx = 1/reduceSize * inv-variance * scale * (reduceSize * dy - dbias
+                // - tmp)
+                for(const auto& reduce_index : arg.reduce_index_set_)
+                {
+                    size_t x_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.x_reduce_strides_, reduce_index);
+                    size_t dy_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.dy_reduce_strides_, reduce_index);
+                    size_t dx_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.dx_reduce_strides_, reduce_index);
+
+                    auto x_offset  = x_invariant_offset + x_reduce_offset;
+                    auto dy_offset = dy_invariant_offset + dy_reduce_offset;
+                    auto dx_offset = dx_invariant_offset + dx_reduce_offset;
+
+                    AccDataType x = type_convert<AccDataType>(arg.p_x_[x_offset]);
+
+                    AccDataType norm_x = (x - mean) * invVar;
+                    AccDataType dy     = type_convert<AccDataType>(arg.p_dy_[dy_offset]);
+
+                    arg.dy_elementwise_op_(dy, dy);
+
+                    AccDataType tmpVal = norm_x * dscale;
+
+                    AccDataType dx = multiplier * (type_convert<AccDataType>(arg.reduceSize_) * dy -
+                                                   dbias - tmpVal);
+
+                    arg.p_dx_[dx_offset] = type_convert<DxDataType>(dx);
+                };
+            };
+
+            std::size_t num_thread = std::thread::hardware_concurrency();
+            std::size_t work_per_thread =
+                (arg.invariant_index_set_.size() + num_thread - 1) / num_thread;
+
+            std::vector<joinable_thread> threads(num_thread);
+
+            for(std::size_t it = 0; it < num_thread; ++it)
+            {
+                std::size_t i_begin = it * work_per_thread;
+                std::size_t i_end   = std::min(static_cast<size_t>((it + 1) * work_per_thread),
+                                             arg.invariant_index_set_.size());
+
+                auto f = [=] {
+                    for(std::size_t i = i_begin; i < i_end; ++i)
+                    {
+                        thread_reduce_func(arg.invariant_index_set_[i]);
+                    }
+                };
+
+                threads[it] = joinable_thread(f);
+            }
+
+            return (0.0f);
+        };
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        };
+    };
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        (void)p_arg;
+
+        return (true);
+    };
+
+    std::unique_ptr<device::BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> xyLengths,
+                        const std::array<index_t, Rank> xStrides,
+                        const std::array<index_t, Rank> dxStrides,
+                        const std::array<index_t, Rank> dyStrides,
+                        const std::array<int, NumBatchNormReduceDim> reduceDims,
+                        const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                        const std::array<index_t, NumInvariantDim> bnScaleStrides,
+                        const std::array<index_t, NumInvariantDim> bnDscaleDbiasStrides,
+                        const std::array<index_t, NumInvariantDim> bnMeanVarStrides,
+                        const void* p_x,
+                        const void* p_dy,
+                        const void* p_scale,
+                        const void* p_savedMean,
+                        const void* p_savedInvVar,
+                        double epsilon,
+                        const DyElementwiseOp dy_elementwise_op,
+                        void* p_dx,
+                        void* p_dscale,
+                        void* p_dbias) override
+    {
+        return std::make_unique<Argument>(xyLengths,
+                                          xStrides,
+                                          dxStrides,
+                                          dyStrides,
+                                          reduceDims,
+                                          bnScaleBiasMeanVarLengths,
+                                          bnScaleStrides,
+                                          bnDscaleDbiasStrides,
+                                          bnMeanVarStrides,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const DyDataType*>(p_dy),
+                                          static_cast<const ScaleDataType*>(p_scale),
+                                          static_cast<const MeanVarDataType*>(p_savedMean),
+                                          static_cast<const MeanVarDataType*>(p_savedInvVar),
+                                          epsilon,
+                                          dy_elementwise_op,
+                                          static_cast<DxDataType*>(p_dx),
+                                          static_cast<DscaleDbiasDataType*>(p_dscale),
+                                          static_cast<DscaleDbiasDataType*>(p_dbias));
+    };
+
+    std::unique_ptr<device::BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "Reference_BatchNorm_Backward" << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp
new file mode 100644
index 0000000000..20c1fcd736
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp
@@ -0,0 +1,368 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <array>
+#include <algorithm>
+#include <thread>
+
+#include "ck/utility/math_v2.hpp"
+#include "ck/utility/ignore.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+struct ReferenceBatchNormFwd : public device::DeviceBatchNormFwd<XDataType,
+                                                                 YDataType,
+                                                                 AccDataType,
+                                                                 ScaleDataType,
+                                                                 BiasDataType,
+                                                                 MeanVarDataType,
+                                                                 YElementwiseOp,
+                                                                 Rank,
+                                                                 NumBatchNormReduceDim>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> xyLengths,
+                 const std::array<index_t, Rank> xStrides,
+                 const std::array<index_t, Rank> yStrides,
+                 const std::array<int, NumBatchNormReduceDim> reduceDims,
+                 const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                 const std::array<index_t, NumInvariantDim> bnScaleStrides,
+                 const std::array<index_t, NumInvariantDim> bnBiasStrides,
+                 const std::array<index_t, NumInvariantDim> bnMeanVarStrides,
+                 const XDataType* p_x,
+                 const ScaleDataType* bnScale,
+                 const BiasDataType* bnBias,
+                 double epsilon,
+                 const YElementwiseOp y_elementwise_op,
+                 YDataType* p_y,
+                 MeanVarDataType* resultSaveMean,
+                 MeanVarDataType* resultSaveInvVariance,
+                 double averageFactor,
+                 MeanVarDataType* resultRunningMean,
+                 MeanVarDataType* resultRunningVariance)
+            : reduceDims_(reduceDims),
+              bnScaleBiasMeanVarLengths_(bnScaleBiasMeanVarLengths),
+              bnScaleStrides_(bnScaleStrides),
+              bnBiasStrides_(bnBiasStrides),
+              bnMeanVarStrides_(bnMeanVarStrides),
+              p_x_(p_x),
+              bnScale_(bnScale),
+              bnBias_(bnBias),
+              y_elementwise_op_(y_elementwise_op),
+              p_y_(p_y),
+              resultSaveMean_(resultSaveMean),
+              resultSaveInvVariance_(resultSaveInvVariance),
+              resultRunningMean_(resultRunningMean),
+              resultRunningVariance_(resultRunningVariance)
+        {
+            using ck::host_common::get_index_set;
+
+            if(std::any_of(
+                   reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+                throw std::runtime_error("Invalid reduce dimensions!");
+
+            // get invariant_dims[] and invariant_lengths[]
+            for(int dim = 0, i = 0; dim < Rank; dim++)
+                if(std::none_of(
+                       reduceDims.begin(), reduceDims.end(), [&](int d) { return d == dim; }))
+                {
+                    invariantDims_[i]     = dim;
+                    invariant_lengths_[i] = xyLengths[dim];
+                    i++;
+                };
+
+            // get reduce_lengths_[]
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim              = reduceDims[j];
+                reduce_lengths_[i++] = xyLengths[dim];
+            };
+
+            for(int i = 0; i < NumInvariantDim; i++)
+                if(invariant_lengths_[i] != bnScaleBiasMeanVarLengths_[i])
+                    throw std::runtime_error("Invalid lengths parameters!");
+
+            for(int j = 0, i = 0; j < NumInvariantDim; j++)
+            {
+                int dim                 = invariantDims_[j];
+                x_invariant_strides_[i] = xStrides[dim];
+                y_invariant_strides_[i] = yStrides[dim];
+                i++;
+            };
+
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim              = reduceDims_[j];
+                x_reduce_strides_[i] = xStrides[dim];
+                y_reduce_strides_[i] = yStrides[dim];
+                i++;
+            };
+
+            invariant_index_set_ = get_index_set<NumInvariantDim>(invariant_lengths_);
+            reduce_index_set_    = get_index_set<NumBatchNormReduceDim>(reduce_lengths_);
+
+            epsilon_       = type_convert<AccDataType>(epsilon);
+            averageFactor_ = type_convert<AccDataType>(averageFactor);
+
+            resultSave    = (resultSaveMean != nullptr && resultSaveInvVariance != nullptr);
+            resultRunning = (resultRunningMean != nullptr && resultRunningVariance != nullptr);
+        }
+
+        std::array<int, NumBatchNormReduceDim> reduceDims_;
+        std::array<int, NumInvariantDim> invariantDims_;
+        std::array<index_t, NumInvariantDim> invariant_lengths_;
+        std::array<index_t, NumBatchNormReduceDim> reduce_lengths_;
+
+        const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths_;
+        const std::array<index_t, NumInvariantDim> bnScaleStrides_;
+        const std::array<index_t, NumInvariantDim> bnBiasStrides_;
+        const std::array<index_t, NumInvariantDim> bnMeanVarStrides_;
+
+        std::array<index_t, NumInvariantDim> x_invariant_strides_;
+        std::array<index_t, NumInvariantDim> y_invariant_strides_;
+        std::array<index_t, NumBatchNormReduceDim> x_reduce_strides_;
+        std::array<index_t, NumBatchNormReduceDim> y_reduce_strides_;
+
+        const XDataType* p_x_;
+        const ScaleDataType* bnScale_;
+        const BiasDataType* bnBias_;
+        const YElementwiseOp y_elementwise_op_;
+        YDataType* p_y_;
+
+        MeanVarDataType* resultSaveMean_;
+        MeanVarDataType* resultSaveInvVariance_;
+        MeanVarDataType* resultRunningMean_;
+        MeanVarDataType* resultRunningVariance_;
+
+        bool resultSave, resultRunning;
+
+        std::vector<std::array<index_t, NumInvariantDim>> invariant_index_set_;
+        std::vector<std::array<index_t, NumBatchNormReduceDim>> reduce_index_set_;
+
+        AccDataType averageFactor_;
+        AccDataType epsilon_;
+    };
+
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            using ck::host_common::get_offset_from_index;
+
+            auto thread_reduce_func = [&](auto invariant_index) {
+                size_t x_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.x_invariant_strides_, invariant_index);
+                size_t y_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.y_invariant_strides_, invariant_index);
+                AccDataType mean     = type_convert<AccDataType>(0.0f);
+                AccDataType variance = type_convert<AccDataType>(0.0f);
+                int32_t curr_count   = 0;
+
+                // compute mean, variance using welford method
+                for(const auto& reduce_index : arg.reduce_index_set_)
+                {
+                    size_t x_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.x_reduce_strides_, reduce_index);
+
+                    auto x_offset = x_invariant_offset + x_reduce_offset;
+
+                    curr_count++;
+
+                    AccDataType x = type_convert<AccDataType>(arg.p_x_[x_offset]);
+
+                    AccDataType delta = x - mean;
+
+                    mean += delta / curr_count;
+
+                    AccDataType delta2 = x - mean;
+
+                    variance += delta * delta2;
+                };
+
+                // actual variance
+                variance = variance / curr_count;
+
+                // inv-variance defined as 1/sqrt(epsilon+variance)
+                AccDataType invVariance =
+                    type_convert<AccDataType>(1.0f) / ck::math::sqrt(arg.epsilon_ + variance);
+
+                // save the mean/inv-variance if required
+                if(arg.resultSave)
+                {
+                    size_t offset = get_offset_from_index<NumInvariantDim>(arg.bnMeanVarStrides_,
+                                                                           invariant_index);
+
+                    arg.resultSaveMean_[offset]        = type_convert<MeanVarDataType>(mean);
+                    arg.resultSaveInvVariance_[offset] = type_convert<MeanVarDataType>(invVariance);
+                };
+
+                // update the moving average if required
+                if(arg.resultRunning)
+                {
+                    size_t offset = get_offset_from_index<NumInvariantDim>(arg.bnMeanVarStrides_,
+                                                                           invariant_index);
+
+                    AccDataType oneMinusAverageFactor =
+                        type_convert<AccDataType>(1.0) - arg.averageFactor_;
+                    arg.resultRunningMean_[offset] = type_convert<MeanVarDataType>(
+                        type_convert<AccDataType>(arg.resultRunningMean_[offset]) *
+                            oneMinusAverageFactor +
+                        mean * arg.averageFactor_);
+                    arg.resultRunningVariance_[offset] = type_convert<MeanVarDataType>(
+                        arg.resultRunningVariance_[offset] * oneMinusAverageFactor +
+                        variance * arg.averageFactor_);
+                };
+
+                size_t scale_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnScaleStrides_, invariant_index);
+                size_t bias_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnBiasStrides_, invariant_index);
+
+                AccDataType scale = type_convert<AccDataType>(arg.bnScale_[scale_offset]);
+                AccDataType bias  = type_convert<AccDataType>(arg.bnBias_[bias_offset]);
+
+                // Normalization
+                for(const auto& reduce_index : arg.reduce_index_set_)
+                {
+                    size_t x_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.x_reduce_strides_, reduce_index);
+                    size_t y_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.y_reduce_strides_, reduce_index);
+
+                    auto x_offset = x_invariant_offset + x_reduce_offset;
+                    auto y_offset = y_invariant_offset + y_reduce_offset;
+
+                    AccDataType x = type_convert<AccDataType>(arg.p_x_[x_offset]);
+
+                    AccDataType norm_x = (x - mean) * invVariance;
+
+                    AccDataType y = scale * norm_x + bias;
+
+                    arg.y_elementwise_op_(y, y);
+
+                    arg.p_y_[y_offset] = type_convert<YDataType>(y);
+                };
+            };
+
+            std::size_t num_thread = std::thread::hardware_concurrency();
+            std::size_t work_per_thread =
+                (arg.invariant_index_set_.size() + num_thread - 1) / num_thread;
+
+            std::vector<joinable_thread> threads(num_thread);
+
+            for(std::size_t it = 0; it < num_thread; ++it)
+            {
+                std::size_t i_begin = it * work_per_thread;
+                std::size_t i_end   = std::min(static_cast<size_t>((it + 1) * work_per_thread),
+                                             arg.invariant_index_set_.size());
+
+                auto f = [=] {
+                    for(std::size_t i = i_begin; i < i_end; ++i)
+                    {
+                        thread_reduce_func(arg.invariant_index_set_[i]);
+                    }
+                };
+
+                threads[it] = joinable_thread(f);
+            }
+
+            return (0.0f);
+        };
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        };
+    };
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        (void)p_arg;
+
+        return (true);
+    };
+
+    std::unique_ptr<device::BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, 4> xyLengths,
+                        const std::array<index_t, 4> xStrides,
+                        const std::array<index_t, 4> yStrides,
+                        const std::array<int, 3> reduceDims,
+                        const std::array<index_t, 1> bnScaleBiasMeanVarLengths,
+                        const std::array<index_t, 1> bnScaleStrides,
+                        const std::array<index_t, 1> bnBiasStrides,
+                        const std::array<index_t, 1> bnMeanVarStrides,
+                        const void* p_x,
+                        const void* bnScale,
+                        const void* bnBias,
+                        double epsilon,
+                        const YElementwiseOp y_elementwise_op,
+                        void* p_y,
+                        void* resultSaveMean,
+                        void* resultSaveInvVariance,
+                        double averageFactor,
+                        void* resultRunningMean,
+                        void* resultRunningVariance) override
+    {
+        return std::make_unique<Argument>(xyLengths,
+                                          xStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          bnScaleBiasMeanVarLengths,
+                                          bnScaleStrides,
+                                          bnBiasStrides,
+                                          bnMeanVarStrides,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const ScaleDataType*>(bnScale),
+                                          static_cast<const BiasDataType*>(bnBias),
+                                          epsilon,
+                                          y_elementwise_op,
+                                          static_cast<YDataType*>(p_y),
+                                          static_cast<MeanVarDataType*>(resultSaveMean),
+                                          static_cast<MeanVarDataType*>(resultSaveInvVariance),
+                                          averageFactor,
+                                          static_cast<MeanVarDataType*>(resultRunningMean),
+                                          static_cast<MeanVarDataType*>(resultRunningVariance));
+    };
+
+    std::unique_ptr<device::BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "Reference_BatchNorm_Forward" << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp
new file mode 100644
index 0000000000..7d652fe4c4
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp
@@ -0,0 +1,300 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+#include <array>
+#include <algorithm>
+
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+struct ReferenceBatchNormInfer : public device::DeviceBatchNormInfer<XDataType,
+                                                                     YDataType,
+                                                                     AccDataType,
+                                                                     ScaleDataType,
+                                                                     BiasDataType,
+                                                                     MeanVarDataType,
+                                                                     YElementwiseOp,
+                                                                     Rank,
+                                                                     NumBatchNormReduceDim>
+{
+    static_assert(Rank <= 6, "Bigger Rank size is not supported!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumBatchNormReduceDim;
+
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> xyLengths,
+                 const std::array<index_t, Rank> xStrides,
+                 const std::array<index_t, Rank> yStrides,
+                 const std::array<int, NumBatchNormReduceDim> reduceDims,
+                 const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                 const std::array<index_t, NumInvariantDim> bnScaleStrides,
+                 const std::array<index_t, NumInvariantDim> bnBiasStrides,
+                 const std::array<index_t, NumInvariantDim> bnMeanVarStrides,
+                 const XDataType* p_x,
+                 const ScaleDataType* bnScale,
+                 const BiasDataType* bnBias,
+                 double epsilon,
+                 const YElementwiseOp y_elementwise_op,
+                 const MeanVarDataType* estimatedMean,
+                 const MeanVarDataType* estimatedVariance,
+                 YDataType* p_y)
+            : reduceDims_(reduceDims),
+              bnScaleBiasMeanVarLengths_(bnScaleBiasMeanVarLengths),
+              bnScaleStrides_(bnScaleStrides),
+              bnBiasStrides_(bnBiasStrides),
+              bnMeanVarStrides_(bnMeanVarStrides),
+              p_x_(p_x),
+              bnScale_(bnScale),
+              bnBias_(bnBias),
+              y_elementwise_op_(y_elementwise_op),
+              estimatedMean_(estimatedMean),
+              estimatedVariance_(estimatedVariance),
+              p_y_(p_y)
+        {
+            using ck::host_common::get_index_set;
+
+            if(std::any_of(
+                   reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+                throw std::runtime_error("Invalid reduce dimensions!");
+
+            // get invariant_dims[] and invariant_lengths[]
+            for(int dim = 0, i = 0; dim < Rank; dim++)
+                if(std::none_of(
+                       reduceDims.begin(), reduceDims.end(), [&](int d) { return d == dim; }))
+                {
+                    invariantDims_[i]     = dim;
+                    invariant_lengths_[i] = xyLengths[dim];
+                    i++;
+                };
+
+            // get reduce_lengths_[]
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim              = reduceDims[j];
+                reduce_lengths_[i++] = xyLengths[dim];
+            };
+
+            // check invariant_lengths_ and bnScaleBiasMeanVarLengths
+            for(int i = 0; i < NumInvariantDim; i++)
+                if(invariant_lengths_[i] != bnScaleBiasMeanVarLengths_[i])
+                    throw std::runtime_error("Invalid lengths parameters!");
+
+            for(int j = 0, i = 0; j < NumInvariantDim; j++)
+            {
+                int dim                 = invariantDims_[j];
+                x_invariant_strides_[i] = xStrides[dim];
+                y_invariant_strides_[i] = yStrides[dim];
+                i++;
+            };
+
+            for(int j = 0, i = 0; j < NumBatchNormReduceDim; j++)
+            {
+                int dim              = reduceDims_[j];
+                x_reduce_strides_[i] = xStrides[dim];
+                y_reduce_strides_[i] = yStrides[dim];
+                i++;
+            };
+
+            invariant_index_set_ = get_index_set<NumInvariantDim>(invariant_lengths_);
+            reduce_index_set_    = get_index_set<NumBatchNormReduceDim>(reduce_lengths_);
+
+            epsilon_ = type_convert<AccDataType>(epsilon);
+        }
+
+        std::array<int, NumBatchNormReduceDim> reduceDims_;
+        std::array<int, NumInvariantDim> invariantDims_;
+        std::array<index_t, NumInvariantDim> invariant_lengths_;
+        std::array<index_t, NumBatchNormReduceDim> reduce_lengths_;
+
+        const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths_;
+        const std::array<index_t, NumInvariantDim> bnScaleStrides_;
+        const std::array<index_t, NumInvariantDim> bnBiasStrides_;
+        const std::array<index_t, NumInvariantDim> bnMeanVarStrides_;
+
+        std::array<index_t, NumInvariantDim> x_invariant_strides_;
+        std::array<index_t, NumInvariantDim> y_invariant_strides_;
+        std::array<index_t, NumBatchNormReduceDim> x_reduce_strides_;
+        std::array<index_t, NumBatchNormReduceDim> y_reduce_strides_;
+
+        const XDataType* p_x_;
+        const ScaleDataType* bnScale_;
+        const BiasDataType* bnBias_;
+        const YElementwiseOp y_elementwise_op_;
+
+        const MeanVarDataType* estimatedMean_;
+        const MeanVarDataType* estimatedVariance_;
+
+        YDataType* p_y_;
+
+        std::vector<std::array<index_t, NumInvariantDim>> invariant_index_set_;
+        std::vector<std::array<index_t, NumBatchNormReduceDim>> reduce_index_set_;
+
+        AccDataType epsilon_;
+    };
+
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            using ck::host_common::get_offset_from_index;
+
+            auto thread_reduce_func = [&](auto invariant_index) {
+                size_t x_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.x_invariant_strides_, invariant_index);
+                size_t y_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                    arg.y_invariant_strides_, invariant_index);
+
+                size_t mean_variance_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnMeanVarStrides_, invariant_index);
+
+                AccDataType mean     = arg.estimatedMean_[mean_variance_offset];
+                AccDataType variance = arg.estimatedVariance_[mean_variance_offset];
+
+                // inv-variance defined as 1/sqrt(epsilon+variance)
+                AccDataType invVariance =
+                    type_convert<AccDataType>(1.0f) / std::sqrt(arg.epsilon_ + variance);
+
+                size_t scale_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnScaleStrides_, invariant_index);
+                size_t bias_offset =
+                    get_offset_from_index<NumInvariantDim>(arg.bnBiasStrides_, invariant_index);
+
+                AccDataType scale = type_convert<AccDataType>(arg.bnScale_[scale_offset]);
+                AccDataType bias  = type_convert<AccDataType>(arg.bnBias_[bias_offset]);
+
+                // normalization
+                for(const auto& reduce_index : arg.reduce_index_set_)
+                {
+                    size_t x_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.x_reduce_strides_, reduce_index);
+                    size_t y_reduce_offset = get_offset_from_index<NumBatchNormReduceDim>(
+                        arg.y_reduce_strides_, reduce_index);
+
+                    auto x_offset = x_invariant_offset + x_reduce_offset;
+                    auto y_offset = y_invariant_offset + y_reduce_offset;
+
+                    AccDataType x = type_convert<AccDataType>(arg.p_x_[x_offset]);
+
+                    AccDataType norm_x = (x - mean) * invVariance;
+
+                    AccDataType y = scale * norm_x + bias;
+
+                    arg.y_elementwise_op_(y, y);
+
+                    arg.p_y_[y_offset] = type_convert<YDataType>(y);
+                };
+            };
+
+            std::size_t num_thread = std::thread::hardware_concurrency();
+            std::size_t work_per_thread =
+                (arg.invariant_index_set_.size() + num_thread - 1) / num_thread;
+
+            std::vector<joinable_thread> threads(num_thread);
+
+            for(std::size_t it = 0; it < num_thread; ++it)
+            {
+                std::size_t i_begin = it * work_per_thread;
+                std::size_t i_end   = std::min(static_cast<size_t>((it + 1) * work_per_thread),
+                                             arg.invariant_index_set_.size());
+
+                auto f = [=] {
+                    for(std::size_t i = i_begin; i < i_end; ++i)
+                    {
+                        thread_reduce_func(arg.invariant_index_set_[i]);
+                    }
+                };
+
+                threads[it] = joinable_thread(f);
+            }
+
+            return (0.0f);
+        };
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        };
+    };
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        (void)p_arg;
+
+        return (true);
+    };
+
+    std::unique_ptr<device::BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> xyLengths,
+                        const std::array<index_t, Rank> xStrides,
+                        const std::array<index_t, Rank> yStrides,
+                        const std::array<int, NumBatchNormReduceDim> reduceDims,
+                        const std::array<index_t, NumInvariantDim> bnScaleBiasMeanVarLengths,
+                        const std::array<index_t, NumInvariantDim> bnScaleStrides,
+                        const std::array<index_t, NumInvariantDim> bnBiasStrides,
+                        const std::array<index_t, NumInvariantDim> bnMeanVarStrides,
+                        const void* p_x,
+                        const void* bnScale,
+                        const void* bnBias,
+                        double epsilon,
+                        const YElementwiseOp y_elementwise_op,
+                        const void* estimatedMean,
+                        const void* estimatedVariance,
+                        void* p_y) override
+    {
+        return std::make_unique<Argument>(xyLengths,
+                                          xStrides,
+                                          yStrides,
+                                          reduceDims,
+                                          bnScaleBiasMeanVarLengths,
+                                          bnScaleStrides,
+                                          bnBiasStrides,
+                                          bnMeanVarStrides,
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const ScaleDataType*>(bnScale),
+                                          static_cast<const BiasDataType*>(bnBias),
+                                          epsilon,
+                                          y_elementwise_op,
+                                          static_cast<const MeanVarDataType*>(estimatedMean),
+                                          static_cast<const MeanVarDataType*>(estimatedVariance),
+                                          static_cast<YDataType*>(p_y));
+    };
+
+    std::unique_ptr<device::BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "Reference_BatchNorm_Infer<" << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp
new file mode 100644
index 0000000000..24f754e598
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp
@@ -0,0 +1,184 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// FIXME: support arbitrary elementwise operation for A/B/C
+template <
+    typename ADataType,
+    typename BDataType,
+    typename CDataType,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    enable_if_t<
+        is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<CElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+        bool> = false>
+struct ReferenceCGemm : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_m_k_real,
+                 const Tensor<ADataType>& a_m_k_imag,
+                 const Tensor<BDataType>& b_k_n_real,
+                 const Tensor<BDataType>& b_k_n_imag,
+                 Tensor<CDataType>& c_m_n_real,
+                 Tensor<CDataType>& c_m_n_imag,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : a_m_k_real_{a_m_k_real},
+              a_m_k_imag_{a_m_k_imag},
+              b_k_n_real_{b_k_n_real},
+              b_k_n_imag_{b_k_n_imag},
+              c_m_n_real_{c_m_n_real},
+              c_m_n_imag_{c_m_n_imag},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_m_k_real_;
+        const Tensor<ADataType>& a_m_k_imag_;
+        const Tensor<BDataType>& b_k_n_real_;
+        const Tensor<BDataType>& b_k_n_imag_;
+        Tensor<CDataType>& c_m_n_real_;
+        Tensor<CDataType>& c_m_n_imag_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceCGemm::Argument;
+
+        float Run(const Argument& arg)
+        {
+            const std::size_t K = arg.a_m_k_real_.mDesc.GetLengths()[1];
+
+            if(K != arg.a_m_k_imag_.mDesc.GetLengths()[1])
+            {
+                throw std::runtime_error("wrong! Incompatible real and imag sizes in CGEMM");
+            }
+
+            auto f_mk_kn_mn_real = [&](auto m, auto n) {
+                float v_c_real = 0;
+
+                for(std::size_t k = 0; k < K; ++k)
+                {
+                    float v_a_real = ck::type_convert<float>(arg.a_m_k_real_(m, k));
+                    float v_a_imag = ck::type_convert<float>(arg.a_m_k_imag_(m, k));
+                    float v_b_real = ck::type_convert<float>(arg.b_k_n_real_(k, n));
+                    float v_b_imag = ck::type_convert<float>(arg.b_k_n_imag_(k, n));
+
+                    v_c_real += v_a_real * v_b_real - v_a_imag * v_b_imag;
+                }
+
+                arg.c_m_n_real_(m, n) = ck::type_convert<CDataType>(v_c_real);
+            };
+
+            auto f_mk_kn_mn_imag = [&](auto m, auto n) {
+                float v_c_imag = 0;
+
+                for(std::size_t k = 0; k < K; ++k)
+                {
+                    float v_a_real = ck::type_convert<float>(arg.a_m_k_real_(m, k));
+                    float v_a_imag = ck::type_convert<float>(arg.a_m_k_imag_(m, k));
+                    float v_b_real = ck::type_convert<float>(arg.b_k_n_real_(k, n));
+                    float v_b_imag = ck::type_convert<float>(arg.b_k_n_imag_(k, n));
+
+                    v_c_imag += v_a_real * v_b_imag + v_a_imag * v_b_real;
+                }
+
+                arg.c_m_n_imag_(m, n) = ck::type_convert<CDataType>(v_c_imag);
+            };
+
+            make_ParallelTensorFunctor(f_mk_kn_mn_real,
+                                       arg.c_m_n_real_.mDesc.GetLengths()[0],
+                                       arg.c_m_n_real_.mDesc.GetLengths()[1])(
+                std::thread::hardware_concurrency());
+            make_ParallelTensorFunctor(f_mk_kn_mn_imag,
+                                       arg.c_m_n_imag_.mDesc.GetLengths()[0],
+                                       arg.c_m_n_imag_.mDesc.GetLengths()[1])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_m_k_real,
+                             const Tensor<ADataType>& a_m_k_imag,
+                             const Tensor<BDataType>& b_k_n_real,
+                             const Tensor<BDataType>& b_k_n_imag,
+                             Tensor<CDataType>& c_m_n_real,
+                             Tensor<CDataType>& c_m_n_imag,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{a_m_k_real,
+                        a_m_k_imag,
+                        b_k_n_real,
+                        b_k_n_imag,
+                        c_m_n_real,
+                        c_m_n_imag,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceCGemm"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_contraction.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_contraction.hpp
new file mode 100644
index 0000000000..92a8c82a6e
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_contraction.hpp
@@ -0,0 +1,146 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// hardcoded for NumDimM == NumDimN == NumDimK == 2
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          ck::enable_if_t<NumDimM == 2 && NumDimN == 2 && NumDimK == 2, bool> = false>
+struct ReferenceContraction_M2_N2_K2 : public ck::tensor_operation::device::BaseOperator
+{
+    // Argument
+    struct Argument : public ck::tensor_operation::device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_ms_ks,
+                 const Tensor<BDataType>& b_ns_ks,
+                 Tensor<CDataType>& c_ms_ns,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op)
+            : a_ms_ks_{a_ms_ks},
+              b_ns_ks_{b_ns_ks},
+              c_ms_ns_{c_ms_ns},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_ms_ks_;
+        const Tensor<BDataType>& b_ns_ks_;
+        Tensor<CDataType>& c_ms_ns_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public ck::tensor_operation::device::BaseInvoker
+    {
+        using Argument = ReferenceContraction_M2_N2_K2::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_ms_ns = [&](auto m0, auto m1, auto n0, auto n1) {
+                const ck::index_t K0 = arg.a_ms_ks_.mDesc.GetLengths()[2];
+                const ck::index_t K1 = arg.a_ms_ks_.mDesc.GetLengths()[3];
+
+                AccDataType v_acc = 0;
+
+                for(ck::index_t k0 = 0; k0 < K0; ++k0)
+                {
+                    for(ck::index_t k1 = 0; k1 < K1; ++k1)
+                    {
+                        AccDataType v_a;
+                        AccDataType v_b;
+
+                        arg.a_element_op_(
+                            v_a, ck::type_convert<const AccDataType>(arg.a_ms_ks_(m0, m1, k0, k1)));
+                        arg.b_element_op_(
+                            v_b, ck::type_convert<const AccDataType>(arg.b_ns_ks_(n0, n1, k0, k1)));
+
+                        v_acc += v_a * v_b;
+                    }
+                }
+
+                arg.c_ms_ns_(m0, m1, n0, n1) = v_acc;
+            };
+
+            make_ParallelTensorFunctor(f_ms_ns,
+                                       arg.c_ms_ns_.mDesc.GetLengths()[0],
+                                       arg.c_ms_ns_.mDesc.GetLengths()[1],
+                                       arg.c_ms_ns_.mDesc.GetLengths()[2],
+                                       arg.c_ms_ns_.mDesc.GetLengths()[3])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
+    {
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_ms_ks,
+                             const Tensor<BDataType>& b_ns_ks,
+                             Tensor<CDataType>& c_ms_ns,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op)
+    {
+        return Argument{a_ms_ks, b_ns_ks, c_ms_ns, a_element_op, b_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceContraction_M2_N2_K2"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
new file mode 100644
index 0000000000..50040a2441
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
@@ -0,0 +1,378 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// input descriptor in [G, N, C, Do, Ho, Wo] order
+// weight descriptor in [G, K, C, Z, Y, X] order
+// output descriptor in [G, N, K, Di, Hi, Wi] order
+// phyiscal layout is irrelavent
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
+struct ReferenceConvBwdData : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(Tensor<InDataType>& input,
+                 const Tensor<WeiDataType>& weight,
+                 const Tensor<OutDataType>& output,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : input_{input},
+              weight_{weight},
+              output_{output},
+              conv_strides_{conv_filter_strides},
+              conv_dilations_{conv_filter_dilations},
+              in_left_pads_{input_left_pads},
+              in_right_pads_{input_right_pads},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+        {
+        }
+
+        Tensor<InDataType>& input_;
+        const Tensor<WeiDataType>& weight_;
+        const Tensor<OutDataType>& output_;
+
+        std::vector<index_t> conv_strides_;
+        std::vector<index_t> conv_dilations_;
+        std::vector<index_t> in_left_pads_;
+        std::vector<index_t> in_right_pads_;
+
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceConvBwdData::Argument;
+
+        float Run(const Argument& arg)
+        {
+            if(!(arg.input_.GetNumOfDimension() == NDimSpatial + 3 &&
+                 arg.weight_.GetNumOfDimension() == NDimSpatial + 3 &&
+                 arg.output_.GetNumOfDimension() == NDimSpatial + 3))
+            {
+                throw std::runtime_error("wrong! inconsistent dimension");
+            }
+
+            if constexpr(NDimSpatial == 1)
+            {
+                auto f_ncw = [&](auto g, auto n, auto c, auto wi) {
+                    std::size_t K  = arg.weight_.GetLengths()[1];
+                    std::size_t X  = arg.weight_.GetLengths()[3];
+                    std::size_t Wo = arg.output_.GetLengths()[3];
+
+                    float v_acc = 0;
+
+                    for(std::size_t x = 0; x < X; ++x)
+                    {
+                        auto w_tmp = static_cast<ck::long_index_t>(wi) +
+                                     static_cast<ck::long_index_t>(arg.in_left_pads_[0]) -
+                                     static_cast<ck::long_index_t>(x * arg.conv_dilations_[0]);
+
+                        if(w_tmp % arg.conv_strides_[0] == 0)
+                        {
+                            auto wo = static_cast<ck::long_index_t>(w_tmp) /
+                                      static_cast<ck::long_index_t>(arg.conv_strides_[0]);
+
+                            if(wo >= 0 && ck::type_convert<std::size_t>(wo) < Wo)
+                            {
+                                for(std::size_t k = 0; k < K; ++k)
+                                {
+                                    float v_out = 0;
+                                    float v_wei = 0;
+
+                                    arg.out_element_op_(
+                                        v_out, ck::type_convert<float>(arg.output_(g, n, k, wo)));
+
+                                    arg.wei_element_op_(
+                                        v_wei, ck::type_convert<float>(arg.weight_(g, k, c, x)));
+
+                                    v_acc += v_out * v_wei;
+                                }
+                            }
+                        }
+                    }
+
+                    float v_in;
+
+                    arg.in_element_op_(v_in, v_acc);
+
+                    arg.input_(g, n, c, wi) = ck::type_convert<InDataType>(v_in);
+                };
+
+                make_ParallelTensorFunctor(f_ncw,
+                                           arg.input_.GetLengths()[0],
+                                           arg.input_.GetLengths()[1],
+                                           arg.input_.GetLengths()[2],
+                                           arg.input_.GetLengths()[3])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 2)
+            {
+                auto f_nchw = [&](auto g, auto n, auto c, auto hi, auto wi) {
+                    std::size_t K = arg.weight_.GetLengths()[1];
+                    std::size_t Y = arg.weight_.GetLengths()[3];
+                    std::size_t X = arg.weight_.GetLengths()[4];
+
+                    std::size_t Ho = arg.output_.GetLengths()[3];
+                    std::size_t Wo = arg.output_.GetLengths()[4];
+
+                    float v_acc = 0;
+
+                    for(std::size_t y = 0; y < Y; ++y)
+                    {
+                        auto h_tmp = static_cast<ck::long_index_t>(hi) +
+                                     static_cast<ck::long_index_t>(arg.in_left_pads_[0]) -
+                                     static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]);
+                        if(h_tmp % arg.conv_strides_[0] == 0)
+                        {
+                            auto ho = static_cast<ck::long_index_t>(h_tmp) /
+                                      static_cast<ck::long_index_t>(arg.conv_strides_[0]);
+                            if(ho >= 0 && ck::type_convert<std::size_t>(ho) < Ho)
+                            {
+                                for(std::size_t x = 0; x < X; ++x)
+                                {
+                                    auto w_tmp =
+                                        static_cast<ck::long_index_t>(wi) +
+                                        static_cast<ck::long_index_t>(arg.in_left_pads_[1]) -
+                                        static_cast<ck::long_index_t>(x * arg.conv_dilations_[1]);
+                                    if(w_tmp % arg.conv_strides_[1] == 0)
+                                    {
+                                        auto wo =
+                                            static_cast<ck::long_index_t>(w_tmp) /
+                                            static_cast<ck::long_index_t>(arg.conv_strides_[1]);
+                                        if(wo >= 0 && ck::type_convert<std::size_t>(wo) < Wo)
+                                        {
+                                            for(std::size_t k = 0; k < K; ++k)
+                                            {
+                                                float v_out = 0;
+                                                float v_wei = 0;
+
+                                                arg.out_element_op_(
+                                                    v_out,
+                                                    ck::type_convert<float>(
+                                                        arg.output_(g, n, k, ho, wo)));
+
+                                                arg.wei_element_op_(
+                                                    v_wei,
+                                                    ck::type_convert<float>(
+                                                        arg.weight_(g, k, c, y, x)));
+
+                                                v_acc += v_out * v_wei;
+                                            }
+                                        }
+                                    }
+                                }
+                            }
+                        }
+                    }
+
+                    float v_in;
+
+                    arg.in_element_op_(v_in, v_acc);
+
+                    arg.input_(g, n, c, hi, wi) = ck::type_convert<InDataType>(v_in);
+                };
+
+                make_ParallelTensorFunctor(f_nchw,
+                                           arg.input_.GetLengths()[0],
+                                           arg.input_.GetLengths()[1],
+                                           arg.input_.GetLengths()[2],
+                                           arg.input_.GetLengths()[3],
+                                           arg.input_.GetLengths()[4])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+                auto f_ncdhw = [&](auto g, auto n, auto c, auto di, auto hi, auto wi) {
+                    std::size_t K = arg.weight_.GetLengths()[1];
+                    std::size_t Z = arg.weight_.GetLengths()[3];
+                    std::size_t Y = arg.weight_.GetLengths()[4];
+                    std::size_t X = arg.weight_.GetLengths()[5];
+
+                    std::size_t Do = arg.output_.GetLengths()[3];
+                    std::size_t Ho = arg.output_.GetLengths()[4];
+                    std::size_t Wo = arg.output_.GetLengths()[5];
+
+                    float v_acc = 0;
+
+                    for(std::size_t z = 0; z < Z; ++z)
+                    {
+                        auto d_tmp = static_cast<ck::long_index_t>(di) +
+                                     static_cast<ck::long_index_t>(arg.in_left_pads_[0]) -
+                                     static_cast<ck::long_index_t>(z * arg.conv_dilations_[0]);
+                        if(d_tmp % arg.conv_strides_[0] == 0)
+                        {
+                            auto do_ = static_cast<ck::long_index_t>(d_tmp) /
+                                       static_cast<ck::long_index_t>(arg.conv_strides_[0]);
+                            if(do_ >= 0 && ck::type_convert<std::size_t>(do_) < Do)
+                            {
+                                for(std::size_t y = 0; y < Y; ++y)
+                                {
+                                    auto h_tmp =
+                                        static_cast<ck::long_index_t>(hi) +
+                                        static_cast<ck::long_index_t>(arg.in_left_pads_[1]) -
+                                        static_cast<ck::long_index_t>(y * arg.conv_dilations_[1]);
+                                    if(h_tmp % arg.conv_strides_[1] == 0)
+                                    {
+                                        auto ho =
+                                            static_cast<ck::long_index_t>(h_tmp) /
+                                            static_cast<ck::long_index_t>(arg.conv_strides_[1]);
+                                        if(ho >= 0 && ck::type_convert<std::size_t>(ho) < Ho)
+                                        {
+                                            for(std::size_t x = 0; x < X; ++x)
+                                            {
+                                                auto w_tmp = static_cast<ck::long_index_t>(wi) +
+                                                             static_cast<ck::long_index_t>(
+                                                                 arg.in_left_pads_[2]) -
+                                                             static_cast<ck::long_index_t>(
+                                                                 x * arg.conv_dilations_[2]);
+
+                                                if(w_tmp % arg.conv_strides_[2] == 0)
+                                                {
+                                                    auto wo = static_cast<ck::long_index_t>(w_tmp) /
+                                                              static_cast<ck::long_index_t>(
+                                                                  arg.conv_strides_[2]);
+                                                    if(wo >= 0 &&
+                                                       ck::type_convert<std::size_t>(wo) < Wo)
+                                                    {
+                                                        for(std::size_t k = 0; k < K; ++k)
+                                                        {
+                                                            float v_out = 0;
+                                                            float v_wei = 0;
+
+                                                            arg.out_element_op_(
+                                                                v_out,
+                                                                ck::type_convert<float>(arg.output_(
+                                                                    g, n, k, do_, ho, wo)));
+
+                                                            arg.wei_element_op_(
+                                                                v_wei,
+                                                                ck::type_convert<float>(
+                                                                    arg.weight_(g, k, c, z, y, x)));
+
+                                                            v_acc += v_out * v_wei;
+                                                        }
+                                                    }
+                                                }
+                                            }
+                                        }
+                                    }
+                                }
+                            }
+                        }
+                    }
+
+                    float v_in;
+
+                    arg.in_element_op_(v_in, v_acc);
+
+                    arg.input_(g, n, c, di, hi, wi) = ck::type_convert<InDataType>(v_in);
+                };
+
+                make_ParallelTensorFunctor(f_ncdhw,
+                                           arg.input_.GetLengths()[0],
+                                           arg.input_.GetLengths()[1],
+                                           arg.input_.GetLengths()[2],
+                                           arg.input_.GetLengths()[3],
+                                           arg.input_.GetLengths()[4],
+                                           arg.input_.GetLengths()[5])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(Tensor<InDataType>& input,
+                             const Tensor<WeiDataType>& weight,
+                             const Tensor<OutDataType>& output,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{input,
+                        weight,
+                        output,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceConvBwdData"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp
new file mode 100644
index 0000000000..ec5df238ab
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp
@@ -0,0 +1,327 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// input descriptor in [G, N, C, Do, Ho, Wo] order
+// weight descriptor in [G, K, C, Z, Y, X] order
+// output descriptor in [G, N, K, Di, Hi, Wi] order
+// phyiscal layout is irrelavent
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
+struct ReferenceConvBwdWeight : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<InDataType>& in_n_c_hi_wi,
+                 Tensor<WeiDataType>& wei_k_c_y_x,
+                 const Tensor<OutDataType>& out_n_k_ho_wo,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : input_{in_n_c_hi_wi},
+              weight_{wei_k_c_y_x},
+              output_{out_n_k_ho_wo},
+              conv_strides_{conv_filter_strides},
+              conv_dilations_{conv_filter_dilations},
+              in_left_pads_{input_left_pads},
+              in_right_pads_{input_right_pads},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+        {
+        }
+
+        const Tensor<InDataType>& input_;
+        Tensor<WeiDataType>& weight_;
+        const Tensor<OutDataType>& output_;
+
+        std::vector<index_t> conv_strides_;
+        std::vector<index_t> conv_dilations_;
+        std::vector<index_t> in_left_pads_;
+        std::vector<index_t> in_right_pads_;
+
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceConvBwdWeight::Argument;
+
+        float Run(const Argument& arg)
+        {
+            if(!(arg.input_.GetNumOfDimension() == NDimSpatial + 3 &&
+                 arg.weight_.GetNumOfDimension() == NDimSpatial + 3 &&
+                 arg.output_.GetNumOfDimension() == NDimSpatial + 3))
+            {
+                throw std::runtime_error("wrong! inconsistent dimension");
+            }
+
+            if constexpr(NDimSpatial == 1)
+            {
+                auto f_kcx = [&](auto g, auto k, auto c, auto x) {
+                    float v_acc = 0;
+
+                    for(std::size_t n = 0; n < arg.output_.GetLengths()[1]; ++n)
+                    {
+                        for(std::size_t wo = 0; wo < arg.output_.GetLengths()[3]; ++wo)
+                        {
+                            auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[0]) +
+                                      static_cast<ck::long_index_t>(x * arg.conv_dilations_[0]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+
+                            if(wi >= 0 &&
+                               ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3])
+                            {
+                                float v_out;
+                                float v_in;
+
+                                arg.out_element_op_(
+                                    v_out, ck::type_convert<float>(arg.output_(g, n, k, wo)));
+
+                                arg.in_element_op_(
+                                    v_in, ck::type_convert<float>(arg.input_(g, n, c, wi)));
+
+                                v_acc += v_out * v_in;
+                            }
+                        }
+                    }
+
+                    float v_wei;
+
+                    arg.wei_element_op_(v_wei, v_acc);
+
+                    arg.weight_(g, k, c, x) = ck::type_convert<WeiDataType>(v_wei);
+                };
+
+                make_ParallelTensorFunctor(f_kcx,
+                                           arg.weight_.GetLengths()[0],
+                                           arg.weight_.GetLengths()[1],
+                                           arg.weight_.GetLengths()[2],
+                                           arg.weight_.GetLengths()[3])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 2)
+            {
+                auto f_kcyx = [&](auto g, auto k, auto c, auto y, auto x) {
+                    std::size_t N = arg.output_.GetLengths()[1];
+
+                    std::size_t Ho = arg.output_.GetLengths()[3];
+                    std::size_t Wo = arg.output_.GetLengths()[4];
+
+                    float v_acc = 0;
+
+                    for(std::size_t n = 0; n < N; ++n)
+                    {
+                        for(std::size_t ho = 0; ho < Ho; ++ho)
+                        {
+                            auto hi = static_cast<ck::long_index_t>(ho * arg.conv_strides_[0]) +
+                                      static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+
+                            for(std::size_t wo = 0; wo < Wo; ++wo)
+                            {
+                                auto wi =
+                                    static_cast<ck::long_index_t>(wo * arg.conv_strides_[1]) +
+                                    static_cast<ck::long_index_t>(x * arg.conv_dilations_[1]) -
+                                    static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
+
+                                if(hi >= 0 &&
+                                   ck::type_convert<std::size_t>(hi) < arg.input_.GetLengths()[3] &&
+                                   wi >= 0 &&
+                                   ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[4])
+                                {
+                                    float v_out;
+                                    float v_in;
+
+                                    arg.out_element_op_(
+                                        v_out,
+                                        ck::type_convert<float>(arg.output_(g, n, k, ho, wo)));
+
+                                    arg.in_element_op_(
+                                        v_in, ck::type_convert<float>(arg.input_(g, n, c, hi, wi)));
+
+                                    v_acc += v_out * v_in;
+                                }
+                            }
+                        }
+                    }
+
+                    float v_wei;
+
+                    arg.wei_element_op_(v_wei, v_acc);
+
+                    arg.weight_(g, k, c, y, x) = ck::type_convert<WeiDataType>(v_wei);
+                };
+
+                make_ParallelTensorFunctor(f_kcyx,
+                                           arg.weight_.GetLengths()[0],
+                                           arg.weight_.GetLengths()[1],
+                                           arg.weight_.GetLengths()[2],
+                                           arg.weight_.GetLengths()[3],
+                                           arg.weight_.GetLengths()[4])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+                auto f_kczyx = [&](auto g, auto k, auto c, auto z, auto y, auto x) {
+                    float v_acc = 0;
+
+                    for(std::size_t n = 0; n < arg.output_.GetLengths()[1]; ++n)
+                    {
+                        for(std::size_t do_ = 0; do_ < arg.output_.GetLengths()[3]; ++do_)
+                        {
+                            auto di = static_cast<ck::long_index_t>(do_ * arg.conv_strides_[0]) +
+                                      static_cast<ck::long_index_t>(z * arg.conv_dilations_[0]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+                            for(std::size_t ho = 0; ho < arg.output_.GetLengths()[4]; ++ho)
+                            {
+                                auto hi =
+                                    static_cast<ck::long_index_t>(ho * arg.conv_strides_[1]) +
+                                    static_cast<ck::long_index_t>(y * arg.conv_dilations_[1]) -
+                                    static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
+                                for(std::size_t wo = 0; wo < arg.output_.GetLengths()[5]; ++wo)
+                                {
+                                    auto wi =
+                                        static_cast<ck::long_index_t>(wo * arg.conv_strides_[2]) +
+                                        static_cast<ck::long_index_t>(x * arg.conv_dilations_[2]) -
+                                        static_cast<ck::long_index_t>(arg.in_left_pads_[2]);
+
+                                    if(di >= 0 &&
+                                       ck::type_convert<std::size_t>(di) <
+                                           arg.input_.GetLengths()[3] &&
+                                       hi >= 0 &&
+                                       ck::type_convert<std::size_t>(hi) <
+                                           arg.input_.GetLengths()[4] &&
+                                       wi >= 0 &&
+                                       ck::type_convert<std::size_t>(wi) <
+                                           arg.input_.GetLengths()[5])
+                                    {
+                                        float v_out;
+                                        float v_in;
+
+                                        arg.out_element_op_(v_out,
+                                                            ck::type_convert<float>(
+                                                                arg.output_(g, n, k, do_, ho, wo)));
+
+                                        arg.in_element_op_(v_in,
+                                                           ck::type_convert<float>(
+                                                               arg.input_(g, n, c, di, hi, wi)));
+
+                                        v_acc += v_out * v_in;
+                                    }
+                                }
+                            }
+                        }
+                    }
+
+                    float v_wei;
+
+                    arg.wei_element_op_(v_wei, v_acc);
+
+                    arg.weight_(g, k, c, z, y, x) = ck::type_convert<WeiDataType>(v_wei);
+                };
+
+                make_ParallelTensorFunctor(f_kczyx,
+                                           arg.weight_.GetLengths()[0],
+                                           arg.weight_.GetLengths()[1],
+                                           arg.weight_.GetLengths()[2],
+                                           arg.weight_.GetLengths()[3],
+                                           arg.weight_.GetLengths()[4],
+                                           arg.weight_.GetLengths()[5])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<InDataType>& in_n_c_hi_wi,
+                             Tensor<WeiDataType>& wei_k_c_y_x,
+                             const Tensor<OutDataType>& out_n_k_ho_wo,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{in_n_c_hi_wi,
+                        wei_k_c_y_x,
+                        out_n_k_ho_wo,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceConvBwdWeight"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
new file mode 100644
index 0000000000..8f4182a231
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
@@ -0,0 +1,339 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <type_traits>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+//
+// @brief      Reference implementation for forward convolution.
+//
+// @paragraph
+//             Tensor descriptor in GNCHW/GKCXY/GNKHW dimensional order
+//             Supports both GNCHW/NGCHW as well as GNHWC/NHWGC physical layout
+//             as long as dimensions in tensor descriptor is in GNCHW order
+//
+// @tparam     InDataType               Input tensor data type.
+// @tparam     WeiDataType              Weights tensor data type.
+// @tparam     OutDataType              Output tensor data type.
+// @tparam     InElementwiseOperation   Functor for input tensor elementwise
+//                                      operation.
+// @tparam     WeiElementwiseOperation  Functor for weights tensor elementwise
+//                                      operation.
+// @tparam     NDimSpatial  Number of spatial dimensions.
+//
+// input descriptor in [G, N, C, Do, Ho, Wo] order
+// weight descriptor in [G, K, C, Z, Y, X] order
+// output descriptor in [G, N, K, Di, Hi, Wi] order
+// phyiscal layout is irrelavent
+template <ck::index_t NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation,
+          typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
+struct ReferenceConvFwd : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<InDataType>& input,
+                 const Tensor<WeiDataType>& weight,
+                 Tensor<OutDataType>& output,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : input_{input},
+              weight_{weight},
+              output_{output},
+              conv_strides_{conv_filter_strides},
+              conv_dilations_{conv_filter_dilations},
+              in_left_pads_{input_left_pads},
+              in_right_pads_{input_right_pads},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+        {
+        }
+
+        const Tensor<InDataType>& input_;
+        const Tensor<WeiDataType>& weight_;
+        Tensor<OutDataType>& output_;
+
+        std::vector<index_t> conv_strides_;
+        std::vector<index_t> conv_dilations_;
+        std::vector<index_t> in_left_pads_;
+        std::vector<index_t> in_right_pads_;
+
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceConvFwd::Argument;
+
+        float Run(const Argument& arg)
+        {
+            if(!(arg.input_.GetNumOfDimension() == NDimSpatial + 3 &&
+                 arg.weight_.GetNumOfDimension() == NDimSpatial + 3 &&
+                 arg.output_.GetNumOfDimension() == NDimSpatial + 3))
+            {
+                throw std::runtime_error("wrong! inconsistent dimension");
+            }
+
+            if constexpr(NDimSpatial == 1)
+            {
+                auto func = [&](auto g, auto n, auto k, auto wo) {
+                    float v_acc = 0;
+
+                    for(std::size_t c = 0; c < arg.weight_.GetLengths()[2]; ++c)
+                    {
+                        for(std::size_t x = 0; x < arg.weight_.GetLengths()[3]; ++x)
+                        {
+                            auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[0]) +
+                                      static_cast<ck::long_index_t>(x * arg.conv_dilations_[0]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+
+                            if(wi >= 0 &&
+                               ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3])
+                            {
+                                float v_in;
+                                float v_wei;
+
+                                arg.in_element_op_(
+                                    v_in, ck::type_convert<float>(arg.input_(g, n, c, wi)));
+
+                                arg.wei_element_op_(
+                                    v_wei, ck::type_convert<float>(arg.weight_(g, k, c, x)));
+
+                                v_acc += v_in * v_wei;
+                            }
+                        }
+                    }
+
+                    float v_out;
+
+                    arg.out_element_op_(v_out, v_acc);
+
+                    arg.output_(g, n, k, wo) = ck::type_convert<OutDataType>(v_out);
+                };
+
+                make_ParallelTensorFunctor(func,
+                                           arg.output_.GetLengths()[0],
+                                           arg.output_.GetLengths()[1],
+                                           arg.output_.GetLengths()[2],
+                                           arg.output_.GetLengths()[3])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 2)
+            {
+                auto func = [&](auto g, auto n, auto k, auto ho, auto wo) {
+                    float v_acc = 0;
+
+                    for(std::size_t c = 0; c < arg.weight_.GetLengths()[2]; ++c)
+                    {
+                        for(std::size_t y = 0; y < arg.weight_.GetLengths()[3]; ++y)
+                        {
+                            auto hi = static_cast<ck::long_index_t>(ho * arg.conv_strides_[0]) +
+                                      static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+
+                            for(std::size_t x = 0; x < arg.weight_.GetLengths()[4]; ++x)
+                            {
+                                auto wi =
+                                    static_cast<ck::long_index_t>(wo * arg.conv_strides_[1]) +
+                                    static_cast<ck::long_index_t>(x * arg.conv_dilations_[1]) -
+                                    static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
+
+                                if(hi >= 0 &&
+                                   ck::type_convert<std::size_t>(hi) < arg.input_.GetLengths()[3] &&
+                                   wi >= 0 &&
+                                   ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[4])
+                                {
+                                    float v_in;
+                                    float v_wei;
+
+                                    arg.in_element_op_(
+                                        v_in, ck::type_convert<float>(arg.input_(g, n, c, hi, wi)));
+
+                                    arg.wei_element_op_(
+                                        v_wei, ck::type_convert<float>(arg.weight_(g, k, c, y, x)));
+
+                                    v_acc += v_in * v_wei;
+                                }
+                            }
+                        }
+                    }
+
+                    float v_out;
+
+                    arg.out_element_op_(v_out, v_acc);
+
+                    arg.output_(g, n, k, ho, wo) = ck::type_convert<OutDataType>(v_out);
+                };
+
+                make_ParallelTensorFunctor(func,
+                                           arg.output_.GetLengths()[0],
+                                           arg.output_.GetLengths()[1],
+                                           arg.output_.GetLengths()[2],
+                                           arg.output_.GetLengths()[3],
+                                           arg.output_.GetLengths()[4])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+                auto func = [&](auto g, auto n, auto k, auto d_o, auto ho, auto wo) {
+                    float v_acc = 0;
+
+                    for(std::size_t c = 0; c < arg.weight_.GetLengths()[2]; ++c)
+                    {
+                        for(std::size_t z = 0; z < arg.weight_.GetLengths()[3]; ++z)
+                        {
+                            auto di = static_cast<ck::long_index_t>(d_o * arg.conv_strides_[0]) +
+                                      static_cast<ck::long_index_t>(z * arg.conv_dilations_[0]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+                            for(std::size_t y = 0; y < arg.weight_.GetLengths()[4]; ++y)
+                            {
+                                auto hi =
+                                    static_cast<ck::long_index_t>(ho * arg.conv_strides_[1]) +
+                                    static_cast<ck::long_index_t>(y * arg.conv_dilations_[1]) -
+                                    static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
+                                for(std::size_t x = 0; x < arg.weight_.GetLengths()[5]; ++x)
+                                {
+                                    auto wi =
+                                        static_cast<ck::long_index_t>(wo * arg.conv_strides_[2]) +
+                                        static_cast<ck::long_index_t>(x * arg.conv_dilations_[2]) -
+                                        static_cast<ck::long_index_t>(arg.in_left_pads_[2]);
+                                    if(di >= 0 &&
+                                       ck::type_convert<std::size_t>(di) <
+                                           arg.input_.GetLengths()[3] &&
+                                       hi >= 0 &&
+                                       ck::type_convert<std::size_t>(hi) <
+                                           arg.input_.GetLengths()[4] &&
+                                       wi >= 0 &&
+                                       ck::type_convert<std::size_t>(wi) <
+                                           arg.input_.GetLengths()[5])
+                                    {
+                                        float v_in;
+                                        float v_wei;
+
+                                        arg.in_element_op_(v_in,
+                                                           ck::type_convert<float>(
+                                                               arg.input_(g, n, c, di, hi, wi)));
+
+                                        arg.wei_element_op_(
+                                            v_wei,
+                                            ck::type_convert<float>(arg.weight_(g, k, c, z, y, x)));
+
+                                        v_acc += v_in * v_wei;
+                                    }
+                                }
+                            }
+                        }
+                    }
+
+                    float v_out;
+
+                    arg.out_element_op_(v_out, v_acc);
+
+                    arg.output_(g, n, k, d_o, ho, wo) = ck::type_convert<OutDataType>(v_out);
+                };
+
+                make_ParallelTensorFunctor(func,
+                                           arg.output_.GetLengths()[0],
+                                           arg.output_.GetLengths()[1],
+                                           arg.output_.GetLengths()[2],
+                                           arg.output_.GetLengths()[3],
+                                           arg.output_.GetLengths()[4],
+                                           arg.output_.GetLengths()[5])(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override
+    {
+        return NDimSpatial >= 1 && NDimSpatial <= 3;
+    }
+
+    static auto MakeArgument(const Tensor<InDataType>& input,
+                             const Tensor<WeiDataType>& weight,
+                             Tensor<OutDataType>& output,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{input,
+                        weight,
+                        output,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceConvFwd"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp
new file mode 100644
index 0000000000..71c84a1f5c
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp
@@ -0,0 +1,192 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// out[N, Ho, Wo, K] =
+//     activate(in[N, Hi, Wi, C] * wei[K, Y, X, C] + bias[K])
+template <typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct ReferenceConvFwd_Bias_Activation : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<InDataType>& in_n_c_hi_wi,
+                 const Tensor<WeiDataType>& wei_k_c_y_x,
+                 Tensor<OutDataType>& out_n_k_ho_wo,
+                 const Tensor<OutDataType>& bias_k,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : in_n_c_hi_wi_{in_n_c_hi_wi},
+              wei_k_c_y_x_{wei_k_c_y_x},
+              out_n_k_ho_wo_{out_n_k_ho_wo},
+              bias_k_{bias_k},
+              conv_strides_{conv_filter_strides},
+              conv_dilations_{conv_filter_dilations},
+              in_left_pads_{input_left_pads},
+              in_right_pads_{input_right_pads},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+        {
+        }
+
+        const Tensor<InDataType>& in_n_c_hi_wi_;
+        const Tensor<WeiDataType>& wei_k_c_y_x_;
+        Tensor<OutDataType>& out_n_k_ho_wo_;
+        const Tensor<OutDataType>& bias_k_;
+
+        std::vector<index_t> conv_strides_;
+        std::vector<index_t> conv_dilations_;
+        std::vector<index_t> in_left_pads_;
+        std::vector<index_t> in_right_pads_;
+
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceConvFwd_Bias_Activation::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
+                float v_acc = 0;
+
+                for(std::size_t c = 0; c < arg.wei_k_c_y_x_.mDesc.GetLengths()[1]; ++c)
+                {
+                    for(std::size_t y = 0; y < arg.wei_k_c_y_x_.mDesc.GetLengths()[2]; ++y)
+                    {
+                        auto hi = ck::type_convert<ck::long_index_t>(ho * arg.conv_strides_[0]) +
+                                  ck::type_convert<ck::long_index_t>(y * arg.conv_dilations_[0]) -
+                                  ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]);
+                        for(std::size_t x = 0; x < arg.wei_k_c_y_x_.mDesc.GetLengths()[3]; ++x)
+                        {
+                            auto wi =
+                                ck::type_convert<ck::long_index_t>(wo * arg.conv_strides_[1]) +
+                                ck::type_convert<ck::long_index_t>(x * arg.conv_dilations_[1]) -
+                                ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]);
+                            if(hi >= 0 &&
+                               ck::type_convert<std::size_t>(hi) <
+                                   arg.in_n_c_hi_wi_.mDesc.GetLengths()[2] &&
+                               wi >= 0 &&
+                               ck::type_convert<std::size_t>(wi) <
+                                   arg.in_n_c_hi_wi_.mDesc.GetLengths()[3])
+                            {
+                                float v_in;
+                                float v_wei;
+
+                                arg.in_element_op_(
+                                    v_in,
+                                    static_cast<const float>(arg.in_n_c_hi_wi_(n, c, hi, wi)));
+                                arg.wei_element_op_(
+                                    v_wei, static_cast<const float>(arg.wei_k_c_y_x_(k, c, y, x)));
+
+                                v_acc += v_in * v_wei;
+                            }
+                        }
+                    }
+                }
+
+                float v_out;
+
+                arg.out_element_op_(v_out, v_acc, static_cast<float>(arg.bias_k_(k)));
+
+                arg.out_n_k_ho_wo_(n, k, ho, wo) = v_out;
+            };
+
+            make_ParallelTensorFunctor(f_nchw,
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[0],
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[1],
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[2],
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[3])(
+                std::thread::hardware_concurrency());
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<InDataType>& in_n_c_hi_wi,
+                             const Tensor<WeiDataType>& wei_k_c_y_x,
+                             Tensor<OutDataType>& out_n_k_ho_wo,
+                             const Tensor<OutDataType>& bias_k,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{in_n_c_hi_wi,
+                        wei_k_c_y_x,
+                        out_n_k_ho_wo,
+                        bias_k,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceConvFwd_Bias_Activation"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp
new file mode 100644
index 0000000000..0b90b4b50e
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp
@@ -0,0 +1,200 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// out[N, Ho, Wo, K] =
+//     activate(in[N, Hi, Wi, C] * wei[K, Y, X, C] + bias[K]) + residual[N, Ho, Wo, K]
+template <typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct ReferenceConvFwd_Bias_Activation_Add : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<InDataType>& in_n_c_hi_wi,
+                 const Tensor<WeiDataType>& wei_k_c_y_x,
+                 Tensor<OutDataType>& out_n_k_ho_wo,
+                 const Tensor<OutDataType>& bias_k,
+                 const Tensor<OutDataType>& resi_n_k_ho_wo,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : in_n_c_hi_wi_{in_n_c_hi_wi},
+              wei_k_c_y_x_{wei_k_c_y_x},
+              out_n_k_ho_wo_{out_n_k_ho_wo},
+              bias_k_{bias_k},
+              resi_n_k_ho_wo_{resi_n_k_ho_wo},
+              conv_strides_{conv_filter_strides},
+              conv_dilations_{conv_filter_dilations},
+              in_left_pads_{input_left_pads},
+              in_right_pads_{input_right_pads},
+              in_element_op_{in_element_op},
+              wei_element_op_{wei_element_op},
+              out_element_op_{out_element_op}
+        {
+        }
+
+        const Tensor<InDataType>& in_n_c_hi_wi_;
+        const Tensor<WeiDataType>& wei_k_c_y_x_;
+        Tensor<OutDataType>& out_n_k_ho_wo_;
+        const Tensor<OutDataType>& bias_k_;
+        const Tensor<OutDataType>& resi_n_k_ho_wo_;
+
+        std::vector<index_t> conv_strides_;
+        std::vector<index_t> conv_dilations_;
+        std::vector<index_t> in_left_pads_;
+        std::vector<index_t> in_right_pads_;
+
+        InElementwiseOperation in_element_op_;
+        WeiElementwiseOperation wei_element_op_;
+        OutElementwiseOperation out_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceConvFwd_Bias_Activation_Add::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
+                float v_acc = 0;
+
+                for(std::size_t c = 0; c < arg.wei_k_c_y_x_.mDesc.GetLengths()[1]; ++c)
+                {
+                    for(std::size_t y = 0; y < arg.wei_k_c_y_x_.mDesc.GetLengths()[2]; ++y)
+                    {
+                        auto hi = ck::type_convert<ck::long_index_t>(ho * arg.conv_strides_[0]) +
+                                  ck::type_convert<ck::long_index_t>(y * arg.conv_dilations_[0]) -
+                                  ck::type_convert<ck::long_index_t>(arg.in_left_pads_[0]);
+                        for(std::size_t x = 0; x < arg.wei_k_c_y_x_.mDesc.GetLengths()[3]; ++x)
+                        {
+                            auto wi =
+                                ck::type_convert<ck::long_index_t>(wo * arg.conv_strides_[1]) +
+                                ck::type_convert<ck::long_index_t>(x * arg.conv_dilations_[1]) -
+                                ck::type_convert<ck::long_index_t>(arg.in_left_pads_[1]);
+                            if(hi >= 0 &&
+                               ck::type_convert<std::size_t>(hi) <
+                                   arg.in_n_c_hi_wi_.mDesc.GetLengths()[2] &&
+                               wi >= 0 &&
+                               ck::type_convert<std::size_t>(wi) <
+                                   arg.in_n_c_hi_wi_.mDesc.GetLengths()[3])
+                            {
+                                float v_in;
+                                float v_wei;
+
+                                arg.in_element_op_(
+                                    v_in,
+                                    static_cast<const float>(arg.in_n_c_hi_wi_(n, c, hi, wi)));
+                                arg.wei_element_op_(
+                                    v_wei, static_cast<const float>(arg.wei_k_c_y_x_(k, c, y, x)));
+
+                                v_acc += v_in * v_wei;
+                            }
+                        }
+                    }
+                }
+
+                float v_out;
+
+                arg.out_element_op_(v_out,
+                                    v_acc,
+                                    static_cast<const float>(arg.bias_k_(k)),
+                                    static_cast<const float>(arg.resi_n_k_ho_wo_(n, k, ho, wo)));
+
+                arg.out_n_k_ho_wo_(n, k, ho, wo) = v_out;
+            };
+
+            make_ParallelTensorFunctor(f_nchw,
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[0],
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[1],
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[2],
+                                       arg.out_n_k_ho_wo_.mDesc.GetLengths()[3])(
+                std::thread::hardware_concurrency());
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<InDataType>& in_n_c_hi_wi,
+                             const Tensor<WeiDataType>& wei_k_c_y_x,
+                             Tensor<OutDataType>& out_n_k_ho_wo,
+                             const Tensor<OutDataType>& bias_k,
+                             const Tensor<OutDataType>& resi_n_k_ho_wo,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        return Argument{in_n_c_hi_wi,
+                        wei_k_c_y_x,
+                        out_n_k_ho_wo,
+                        bias_k,
+                        resi_n_k_ho_wo,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceConvFwd_Bias_Activation_Add"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
new file mode 100644
index 0000000000..309b4afad8
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
@@ -0,0 +1,156 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct ReferenceGemm : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_m_k,
+                 const Tensor<BDataType>& b_k_n,
+                 Tensor<CDataType>& c_m_n,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : a_m_k_{a_m_k},
+              b_k_n_{b_k_n},
+              c_m_n_{c_m_n},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_m_k_;
+        const Tensor<BDataType>& b_k_n_;
+        Tensor<CDataType>& c_m_n_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceGemm::Argument;
+
+        float Run(const Argument& arg)
+        {
+            auto f_mk_kn_mn = [&](auto m, auto n) {
+                const int K = arg.a_m_k_.mDesc.GetLengths()[1];
+
+                AccDataType v_acc = 0;
+
+                for(int k = 0; k < K; ++k)
+                {
+                    ADataType v_a;
+                    BDataType v_b;
+
+                    // use PassThrough instead of ConvertBF16RTN for reference calculation
+                    if constexpr(is_same_v<AElementwiseOperation,
+                                           ck::tensor_operation::element_wise::ConvertBF16RTN>)
+                    {
+                        ck::tensor_operation::element_wise::PassThrough{}(v_a, arg.a_m_k_(m, k));
+                    }
+                    else
+                    {
+                        arg.a_element_op_(v_a, arg.a_m_k_(m, k));
+                    }
+                    // same for B matrix
+                    if constexpr(is_same_v<BElementwiseOperation,
+                                           ck::tensor_operation::element_wise::ConvertBF16RTN>)
+                    {
+                        ck::tensor_operation::element_wise::PassThrough{}(v_b, arg.b_k_n_(k, n));
+                    }
+                    else
+                    {
+                        arg.b_element_op_(v_b, arg.b_k_n_(k, n));
+                    }
+
+                    v_acc +=
+                        ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
+                }
+
+                CDataType v_c;
+
+                arg.c_element_op_(v_c, v_acc);
+
+                arg.c_m_n_(m, n) = v_c;
+            };
+
+            make_ParallelTensorFunctor(
+                f_mk_kn_mn, arg.c_m_n_.mDesc.GetLengths()[0], arg.c_m_n_.mDesc.GetLengths()[1])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_m_k,
+                             const Tensor<BDataType>& b_k_n,
+                             Tensor<CDataType>& c_m_n,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceGemm"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp
new file mode 100644
index 0000000000..ce2a83da61
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp
@@ -0,0 +1,236 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// D = Layernorm(acc_element_op(A * B + broadcast(bias)) + add) * broadcast(gamma) + broadcast(beta)
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename C0DataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename AccElementwiseOperation,
+          typename CElementwiseOperation>
+struct ReferenceGemmLayernorm : public device::BaseOperator
+{
+    using ReferenceGemmInstance = ReferenceGemm<ADataType,
+                                                BDataType,
+                                                AccDataType,
+                                                AccDataType,
+                                                AElementwiseOperation,
+                                                BElementwiseOperation,
+                                                element_wise::PassThrough>;
+
+    template <typename InDataType, typename OutDataType, typename ComputeDataType>
+    static void RunLayernorm(Tensor<OutDataType>& result,
+                             const Tensor<ComputeDataType>& acc, // MxN
+                             const Tensor<InDataType>& gamma,    // 1xN
+                             const Tensor<InDataType>& beta,     // 1xN
+                             const InDataType epsilon = 1e-5)
+    {
+        assert(acc.mDesc.GetLengths()[1] == gamma.mDesc.GetLengths()[0] &&
+               acc.mDesc.GetLengths()[1] == beta.mDesc.GetLengths()[0]);
+
+        size_t M = acc.mDesc.GetLengths()[0];
+        size_t N = acc.mDesc.GetLengths()[1];
+
+        Tensor<ComputeDataType> avg_acc_sq({M});
+        Tensor<ComputeDataType> avg_acc({M});
+        Tensor<ComputeDataType> acc_layernorm(acc);
+
+        // reduce N dim
+        for(size_t i = 0; i < M; i++)
+        {
+            ComputeDataType sum_acc_sq = 0;
+            ComputeDataType sum_acc    = 0;
+            for(size_t j = 0; j < N; j++)
+            {
+                sum_acc_sq += acc_layernorm(i, j) * acc_layernorm(i, j);
+                sum_acc += acc_layernorm(i, j);
+            }
+            avg_acc_sq(i) = sum_acc_sq / N;
+            avg_acc(i)    = sum_acc / N;
+        }
+
+        // normalize
+        acc_layernorm.ForEach([&](auto& self, auto idx) {
+            self(idx[0], idx[1]) =
+                (self(idx[0], idx[1]) - avg_acc(idx[0])) /
+                sqrt(avg_acc_sq(idx[0]) - avg_acc(idx[0]) * avg_acc(idx[0]) + epsilon);
+        });
+
+        // affine
+        acc_layernorm.ForEach([&](auto& self, auto idx) {
+            self(idx[0], idx[1]) = self(idx[0], idx[1]) * gamma(idx[1]) + beta(idx[1]);
+        });
+
+        // cast
+        result = acc_layernorm.template CopyAsType<OutDataType>();
+    }
+
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_m_k,
+                 const Tensor<BDataType>& b_k_n,
+                 Tensor<CDataType>& c_m_n,
+                 const Tensor<C0DataType>& c0_n_bias,  // 1xN
+                 const Tensor<C0DataType>& c0_m_n_add, // MxN
+                 const Tensor<C0DataType>& c0_n_gamma, // 1xN
+                 const Tensor<C0DataType>& c0_n_beta,  // 1xN
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 AccElementwiseOperation acc_element_op,
+                 CElementwiseOperation c_element_op,
+                 const CDataType epsilon = 1e-5)
+            : a_m_k_{a_m_k},
+              b_k_n_{b_k_n},
+              c_m_n_{c_m_n},
+              c0_n_bias_{c0_n_bias},
+              c0_m_n_add_{c0_m_n_add},
+              c0_n_gamma_{c0_n_gamma},
+              c0_n_beta_{c0_n_beta},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              acc_element_op_{acc_element_op},
+              c_element_op_{c_element_op},
+              epsilon_{epsilon}
+        {
+        }
+
+        const Tensor<ADataType>& a_m_k_;
+        const Tensor<BDataType>& b_k_n_;
+        Tensor<CDataType>& c_m_n_;
+        const Tensor<C0DataType>& c0_n_bias_;
+        const Tensor<C0DataType>& c0_m_n_add_;
+        const Tensor<C0DataType>& c0_n_gamma_;
+        const Tensor<C0DataType>& c0_n_beta_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        AccElementwiseOperation acc_element_op_;
+        CElementwiseOperation c_element_op_;
+
+        const CDataType epsilon_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        // using Argument = ReferenceGemm::Argument;
+
+        float Run(const Argument& arg)
+        {
+            Tensor<AccDataType> acc_m_n(arg.c_m_n_.mDesc);
+            acc_m_n.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0});
+
+            auto ref_gemm     = ReferenceGemmInstance{};
+            auto ref_invoker  = ref_gemm.MakeInvoker();
+            auto ref_argument = ref_gemm.MakeArgument(arg.a_m_k_,
+                                                      arg.b_k_n_,
+                                                      acc_m_n,
+                                                      arg.a_element_op_,
+                                                      arg.b_element_op_,
+                                                      element_wise::PassThrough{});
+
+            // gemm
+            ref_invoker.Run(ref_argument);
+
+            // activation(acc + bias)
+            acc_m_n.ForEach([&](auto& self, auto idx) {
+                AccDataType out;
+                arg.acc_element_op_(out, acc_m_n(idx[0], idx[1]) + arg.c0_n_bias_(idx[1]));
+                self(idx[0], idx[1]) = out;
+            });
+
+            // add from other layers
+            acc_m_n.ForEach([&](auto& self, auto idx) {
+                self(idx[0], idx[1]) += arg.c0_m_n_add_(idx[0], idx[1]);
+            });
+
+            // layernorm
+            RunLayernorm(arg.c_m_n_, acc_m_n, arg.c0_n_gamma_, arg.c0_n_beta_);
+
+            // elementwise op
+            arg.c_m_n_.ForEach([&](auto& self, auto idx) {
+                arg.c_element_op_(self(idx[0], idx[1]), self(idx[0], idx[1]));
+            });
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_m_k,
+                             const Tensor<BDataType>& b_k_n,
+                             Tensor<CDataType>& c_m_n,
+                             const Tensor<C0DataType>& c0_n_bias,  // 1xN
+                             const Tensor<C0DataType>& c0_m_n_add, // 1xN
+                             const Tensor<C0DataType>& c0_n_gamma, // 1xN
+                             const Tensor<C0DataType>& c0_n_beta,  // 1xN
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             AccElementwiseOperation acc_element_op,
+                             CElementwiseOperation c_element_op,
+                             const CDataType epsilon = 1e-5)
+    {
+        return Argument{a_m_k,
+                        b_k_n,
+                        c_m_n,
+                        c0_n_bias,
+                        c0_m_n_add,
+                        c0_n_gamma,
+                        c0_n_beta,
+                        a_element_op,
+                        b_element_op,
+                        acc_element_op,
+                        c_element_op,
+                        epsilon};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceGemmLayernorm"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp
new file mode 100644
index 0000000000..6a48528c54
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp
@@ -0,0 +1,191 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <vector>
+#include <algorithm>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename AccElementwiseOperation>
+struct ReferenceGroupnorm : public device::BaseOperator
+{
+    // x = [N, H, W, G, C]
+    // y = [N, H, W, G, C]
+    // reduce dim [H, W, C], mean, var = [N, G]
+    // gamma, beta = [G, C]
+    // beta: [G, C]
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<XDataType>& x,
+                 const Tensor<GammaDataType>& gamma,
+                 const Tensor<BetaDataType>& beta,
+                 Tensor<YDataType>& y,
+                 AccElementwiseOperation acc_elementwise_op,
+                 const std::vector<index_t> lengths,
+                 AccDataType epsilon)
+            : x_(x),
+              gamma_(gamma),
+              beta_(beta),
+              y_(y),
+              acc_elementwise_op_(acc_elementwise_op),
+              lengths_(lengths),
+              epsilon_(epsilon)
+        {
+        }
+
+        const Tensor<XDataType> x_;
+        const Tensor<XDataType> gamma_;
+        const Tensor<XDataType> beta_;
+        Tensor<YDataType>& y_;
+        AccElementwiseOperation acc_elementwise_op_;
+        std::vector<index_t> lengths_;
+        AccDataType epsilon_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            int N = arg.lengths_[0];
+            int H = arg.lengths_[1];
+            int W = arg.lengths_[2];
+            int G = arg.lengths_[3];
+            int C = arg.lengths_[4];
+
+            Tensor<AccDataType> mean({N, G});
+            Tensor<AccDataType> var({N, G});
+
+            // Compute mean & var in [H, W, C] by Welford Algorithm
+            // TODO - parallel for each HWC
+            // TODO - address calculation
+            for(int n = 0; n < N; ++n)
+            {
+                for(int g = 0; g < G; ++g)
+                {
+                    AccDataType mean_val = type_convert<AccDataType>(0.0f);
+                    AccDataType var_val  = type_convert<AccDataType>(0.0f);
+                    int32_t curr_count   = 0;
+
+                    for(int h = 0; h < H; ++h)
+                    {
+                        for(int w = 0; w < W; ++w)
+                        {
+                            for(int c = 0; c < C; ++c)
+                            {
+                                curr_count++;
+                                AccDataType x = type_convert<AccDataType>(arg.x_(n, h, w, g, c));
+                                AccDataType delta = x - mean_val;
+                                mean_val += delta / curr_count;
+                                AccDataType delta2 = x - mean_val;
+                                var_val += delta * delta2;
+                            }
+                        }
+                    }
+
+                    mean(n, g) = mean_val;
+                    var(n, g)  = var_val / curr_count;
+                }
+            }
+
+            // Normalization
+            for(int n = 0; n < N; ++n)
+            {
+                for(int h = 0; h < H; ++h)
+                {
+                    for(int w = 0; w < W; ++w)
+                    {
+                        for(int g = 0; g < G; ++g)
+                        {
+                            for(int c = 0; c < C; ++c)
+                            {
+                                AccDataType x = type_convert<AccDataType>(arg.x_(n, h, w, g, c));
+                                AccDataType gamma    = type_convert<AccDataType>(arg.gamma_(g, c));
+                                AccDataType beta     = type_convert<AccDataType>(arg.beta_(g, c));
+                                AccDataType mean_val = type_convert<AccDataType>(mean(n, g));
+                                AccDataType var_val  = type_convert<AccDataType>(var(n, g));
+                                AccDataType y        = gamma * (x - mean_val) /
+                                                    ck::math::sqrt(arg.epsilon_ + var_val) +
+                                                beta;
+                                arg.acc_elementwise_op_(y, y);
+                                arg.y_(n, h, w, g, c) = type_convert<YDataType>(y);
+                            }
+                        }
+                    }
+                }
+            }
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
+        if(p_arg_->lengths_.size() != 5)
+            return false;
+
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<XDataType>& x,
+                             const Tensor<GammaDataType>& gamma,
+                             const Tensor<BetaDataType>& beta,
+                             Tensor<YDataType>& y,
+                             AccElementwiseOperation acc_elementwise_op,
+                             const std::vector<index_t> lengths,
+                             AccDataType epsilon)
+    {
+        return Argument{x, gamma, beta, y, acc_elementwise_op, lengths, epsilon};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceLayernorm"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp
new file mode 100644
index 0000000000..3f50ab88b3
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp
@@ -0,0 +1,325 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <type_traits>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+/**
+ * \brief Reference implementation for image to column.
+ *
+ * Tensor descriptor has [G, N, C, Di, Hi, Wi] data layout.
+ * G must be equal to 1. Memory layout is [G, N, Di, Hi, Wi, C].
+ *
+ * \tparam NDimSpatial Number of spatial dimensions.
+ * \tparam InputLayout Input Layout.
+ * \tparam InDataType Input Data Type.
+ * \tparam OutDataType Output Data Type.
+ */
+template <ck::index_t NDimSpatial,
+          typename InputLayout,
+          typename InDataType,
+          typename OutDataType,
+          typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
+struct ReferenceImageToColumn : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        public:
+        Argument(const Tensor<InDataType>& input,
+                 Tensor<OutDataType>& output,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads)
+            : input_{input},
+              output_{output},
+              conv_strides_{conv_filter_strides},
+              conv_dilations_{conv_filter_dilations},
+              in_left_pads_{input_left_pads},
+              in_right_pads_{input_right_pads},
+              filter_spatial_lengths_{filter_spatial_lengths}
+        {
+            initOutputSpatialLengths();
+        }
+
+        const Tensor<InDataType>& input_;
+        Tensor<OutDataType>& output_;
+
+        std::vector<index_t> conv_strides_;
+        std::vector<index_t> conv_dilations_;
+        std::vector<index_t> in_left_pads_;
+        std::vector<index_t> in_right_pads_;
+
+        std::vector<index_t> filter_spatial_lengths_;
+        std::vector<index_t> output_spatial_lengths_;
+
+        private:
+        void initOutputSpatialLengths()
+        {
+            constexpr auto input_offset_to_spatial = 3;
+
+            for(ck::index_t i = 0; i < NDimSpatial; ++i)
+            {
+                // XEff = (X - 1) * conv_dilation_w + 1;
+                // Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
+                const ck::index_t x_eff = (filter_spatial_lengths_[i] - 1) * conv_dilations_[i] + 1;
+
+                output_spatial_lengths_.push_back(
+                    (input_.GetLengths()[i + input_offset_to_spatial] + in_left_pads_[i] +
+                     in_right_pads_[i] - x_eff) /
+                        conv_strides_[i] +
+                    1);
+            }
+        }
+    };
+
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceImageToColumn::Argument;
+
+        float Run(const Argument& arg)
+        {
+            if(!(arg.input_.GetNumOfDimension() == NDimSpatial + 3 &&
+                 arg.output_.GetNumOfDimension() == 2))
+            {
+                throw std::runtime_error("wrong! inconsistent dimension");
+            }
+
+            const index_t N = arg.input_.GetLengths()[1];
+            const index_t C = arg.input_.GetLengths()[2];
+
+            if constexpr(NDimSpatial == 1)
+            {
+                const index_t Wo = arg.output_spatial_lengths_[0];
+                auto func        = [&](auto n, auto wo) {
+                    index_t row    = n * Wo + wo;
+                    index_t column = 0;
+
+                    for(index_t x = 0; x < arg.filter_spatial_lengths_[0]; ++x)
+                    {
+                        auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[0]) +
+                                  static_cast<ck::long_index_t>(x * arg.conv_dilations_[0]) -
+                                  static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+
+                        for(index_t c = 0; c < C; ++c)
+                        {
+                            if(wi >= 0 &&
+                               ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3])
+                            {
+                                InDataType v_in          = arg.input_(0, n, c, wi);
+                                arg.output_(row, column) = ck::type_convert<OutDataType>(v_in);
+                            }
+                            column++;
+                        }
+                    }
+                };
+
+                make_ParallelTensorFunctor(func, N, Wo)(std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 2)
+            {
+                const index_t Ho = arg.output_spatial_lengths_[0];
+                const index_t Wo = arg.output_spatial_lengths_[1];
+
+                auto func = [&](auto n, auto ho, auto wo) {
+                    index_t row    = n * Ho * Wo + ho * Wo + wo;
+                    index_t column = 0;
+
+                    for(index_t y = 0; y < arg.filter_spatial_lengths_[0]; ++y)
+                    {
+                        auto hi = static_cast<ck::long_index_t>(ho * arg.conv_strides_[0]) +
+                                  static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]) -
+                                  static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+
+                        for(index_t x = 0; x < arg.filter_spatial_lengths_[1]; ++x)
+                        {
+                            auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[1]) +
+                                      static_cast<ck::long_index_t>(x * arg.conv_dilations_[1]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
+
+                            for(index_t c = 0; c < C; ++c)
+                            {
+
+                                if(hi >= 0 &&
+                                   ck::type_convert<std::size_t>(hi) < arg.input_.GetLengths()[3] &&
+                                   wi >= 0 &&
+                                   ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[4])
+                                {
+                                    InDataType v_in          = arg.input_(0, n, c, hi, wi);
+                                    arg.output_(row, column) = ck::type_convert<OutDataType>(v_in);
+                                }
+                                column++;
+                            }
+                        }
+                    }
+                };
+
+                make_ParallelTensorFunctor(func, N, Ho, Wo)(std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+                const index_t Do = arg.output_spatial_lengths_[0];
+                const index_t Ho = arg.output_spatial_lengths_[1];
+                const index_t Wo = arg.output_spatial_lengths_[2];
+
+                auto func = [&](auto n, auto d_o, auto ho, auto wo) {
+                    index_t row    = n * Do * Ho * Wo + d_o * Ho * Wo + ho * Wo + wo;
+                    index_t column = 0;
+
+                    for(index_t z = 0; z < arg.filter_spatial_lengths_[0]; ++z)
+                    {
+                        auto di = static_cast<ck::long_index_t>(d_o * arg.conv_strides_[0]) +
+                                  static_cast<ck::long_index_t>(z * arg.conv_dilations_[0]) -
+                                  static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+                        for(index_t y = 0; y < arg.filter_spatial_lengths_[1]; ++y)
+                        {
+                            auto hi = static_cast<ck::long_index_t>(ho * arg.conv_strides_[1]) +
+                                      static_cast<ck::long_index_t>(y * arg.conv_dilations_[1]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
+                            for(index_t x = 0; x < arg.filter_spatial_lengths_[2]; ++x)
+                            {
+                                auto wi =
+                                    static_cast<ck::long_index_t>(wo * arg.conv_strides_[2]) +
+                                    static_cast<ck::long_index_t>(x * arg.conv_dilations_[2]) -
+                                    static_cast<ck::long_index_t>(arg.in_left_pads_[2]);
+                                for(index_t c = 0; c < C; ++c)
+                                {
+                                    if(di >= 0 &&
+                                       ck::type_convert<std::size_t>(di) <
+                                           arg.input_.GetLengths()[3] &&
+                                       hi >= 0 &&
+                                       ck::type_convert<std::size_t>(hi) <
+                                           arg.input_.GetLengths()[4] &&
+                                       wi >= 0 &&
+                                       ck::type_convert<std::size_t>(wi) <
+                                           arg.input_.GetLengths()[5])
+                                    {
+                                        InDataType v_in = arg.input_(0, n, c, di, hi, wi);
+                                        arg.output_(row, column) =
+                                            ck::type_convert<OutDataType>(v_in);
+                                    }
+                                    column++;
+                                }
+                            }
+                        }
+                    }
+                };
+
+                make_ParallelTensorFunctor(func, N, Do, Ho, Wo)(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        using namespace tensor_layout::convolution;
+
+        if constexpr(!(std::is_same_v<InputLayout, GNWC> || std::is_same_v<InputLayout, GNHWC> ||
+                       std::is_same_v<InputLayout, GNDHWC>))
+        {
+            return false;
+        }
+        if constexpr(!(NDimSpatial >= 1 && NDimSpatial <= 3))
+        {
+            return false;
+        }
+        return true;
+    }
+
+    bool IsSupportedArgument(const Argument& arg)
+    {
+        const ck::index_t G = arg.input_.GetLengths()[0];
+        const ck::index_t N = arg.input_.GetLengths()[1];
+        const ck::index_t C = arg.input_.GetLengths()[2];
+
+        const index_t NDoHoWo =
+            N * ck::accumulate_n<index_t>(
+                    arg.output_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+        const index_t CZYX =
+            C * ck::accumulate_n<index_t>(
+                    arg.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+
+        if(!(arg.output_.GetLengths()[0] == static_cast<std::size_t>(NDoHoWo) &&
+             arg.output_.GetLengths()[1] == static_cast<std::size_t>(CZYX)))
+        {
+            return false;
+        }
+
+        if(G != 1)
+        {
+            return false;
+        }
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const Tensor<InDataType>& input,
+                             Tensor<OutDataType>& output,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads)
+    {
+        return Argument{input,
+                        output,
+                        filter_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceImageToColumn"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp
new file mode 100644
index 0000000000..9994a2f9f7
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp
@@ -0,0 +1,174 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <vector>
+#include <algorithm>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename AccElementwiseOperation,
+          index_t Rank,
+          index_t NumReduceDim>
+struct ReferenceLayernorm : public device::BaseOperator
+{
+    // TODO - support generic layernorm
+    static_assert((Rank == 2 && NumReduceDim == 1), "Only support 2D version so far");
+
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<XDataType>& x_m_n,
+                 const Tensor<GammaDataType>& gamma_n,
+                 const Tensor<BetaDataType>& beta_n,
+                 Tensor<YDataType>& y_m_n,
+                 AccElementwiseOperation acc_elementwise_op,
+                 const std::vector<index_t> lengths,
+                 const std::vector<index_t> reduceDims,
+                 AccDataType epsilon)
+            : x_m_n_(x_m_n),
+              gamma_n_(gamma_n),
+              beta_n_(beta_n),
+              y_m_n_(y_m_n),
+              acc_elementwise_op_(acc_elementwise_op),
+              lengths_(lengths),
+              reduceDims_(reduceDims),
+              epsilon_(epsilon)
+        {
+        }
+
+        const Tensor<XDataType> x_m_n_;
+        const Tensor<XDataType> gamma_n_;
+        const Tensor<XDataType> beta_n_;
+        Tensor<YDataType>& y_m_n_;
+        AccElementwiseOperation acc_elementwise_op_;
+        std::vector<index_t> lengths_;
+        std::vector<index_t> reduceDims_;
+        AccDataType epsilon_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            int M = arg.lengths_[0];
+            int N = arg.lengths_[1];
+
+            Tensor<AccDataType> mean({M});
+            Tensor<AccDataType> var({M});
+
+            for(int m = 0; m < M; ++m)
+            {
+                mean(m) = 0;
+                var(m)  = 0;
+
+                for(int n = 0; n < N; ++n)
+                {
+                    auto x_val = ck::type_convert<AccDataType>(arg.x_m_n_(m, n));
+                    mean(m) += x_val;
+                    var(m) += x_val * x_val;
+                }
+
+                mean(m) = mean(m) / N;
+                var(m)  = (var(m) / N) - (mean(m) * mean(m));
+            }
+
+            for(int m = 0; m < M; ++m)
+            {
+                AccDataType divisor =
+                    static_cast<AccDataType>(1) / ck::math::sqrt(var(m) + arg.epsilon_);
+
+                for(int n = 0; n < N; ++n)
+                {
+                    auto x_val = ck::type_convert<AccDataType>(arg.x_m_n_(m, n));
+                    auto y_val = (x_val - mean(m)) * divisor;
+                    y_val      = (y_val * arg.gamma_n_(n)) + arg.beta_n_(n);
+                    arg.acc_elementwise_op_(y_val, y_val);
+                    arg.y_m_n_(m, n) = ck::type_convert<YDataType>(y_val);
+                }
+            }
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
+
+        // TODO - support generic layernorm
+        if(p_arg_->lengths_.size() != 2)
+            return false;
+
+        if(p_arg_->reduceDims_.size() != 1)
+            return false;
+
+        if(p_arg_->reduceDims_[0] != 1)
+            return false;
+
+        return true;
+    }
+
+    static auto MakeArgument(const Tensor<XDataType>& x_m_n,
+                             const Tensor<GammaDataType>& gamma_n,
+                             const Tensor<BetaDataType>& beta_n,
+                             Tensor<YDataType>& y_m_n,
+                             AccElementwiseOperation acc_elementwise_op,
+                             const std::vector<index_t> lengths,
+                             const std::vector<index_t> reduceDims,
+                             AccDataType epsilon)
+    {
+        return Argument{
+            x_m_n, gamma_n, beta_n, y_m_n, acc_elementwise_op, lengths, reduceDims, epsilon};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceLayernorm"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp
new file mode 100644
index 0000000000..60c74fbf14
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp
@@ -0,0 +1,112 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+using namespace std;
+
+template <typename DOutDataType,
+          typename IndexDataType,
+          typename ConputeDataType,
+          typename DInDataType,
+          typename ElementwiseOperation>
+struct ReferenceMaxPoolBwd : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<DOutDataType>& dout,
+                 const Tensor<IndexDataType>& indices,
+                 Tensor<DInDataType>& din,
+                 ElementwiseOperation elementwise_op)
+            : dout_(dout), indices_(indices), din_(din), elementwise_op_(elementwise_op)
+        {
+        }
+
+        const Tensor<DOutDataType>& dout_;
+        const Tensor<IndexDataType>& indices_;
+        Tensor<DInDataType>& din_;
+        ElementwiseOperation elementwise_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            int din_length  = arg.din_.GetElementSpaceSize();
+            int dout_length = arg.dout_.GetElementSpaceSize();
+            std::vector<ConputeDataType> buf(din_length, 0);
+
+            for(int i = 0; i < dout_length; ++i)
+            {
+                int index = arg.indices_.mData[i];
+                if(index >= 0 && index < din_length)
+                {
+                    if constexpr(is_same_v<ConputeDataType, bhalf_t>)
+                    {
+                        float buf_val = ck::type_convert<float>(buf[index]);
+                        buf_val += ck::type_convert<float>(arg.dout_.mData[i]);
+                        buf[index] = ck::type_convert<ConputeDataType>(buf_val);
+                    }
+                    else
+                        buf[index] += ck::type_convert<ConputeDataType>(arg.dout_.mData[i]);
+                }
+            }
+
+            for(int i = 0; i < din_length; ++i)
+                arg.din_.mData[i] = ck::type_convert<DInDataType>(buf[i]);
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<DOutDataType>& dout,
+                             const Tensor<IndexDataType>& indices,
+                             Tensor<DInDataType>& din,
+                             ElementwiseOperation elementwise_op)
+    {
+        return Argument{dout, indices, din, elementwise_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceMaxPoolBwd"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp
new file mode 100644
index 0000000000..cf241ac1b1
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp
@@ -0,0 +1,360 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <vector>
+#include <algorithm>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <index_t InOutRank,
+          index_t WindowRank,
+          typename InDataType,
+          typename OutDataType,
+          typename ComputeDataType,
+          typename IndexDataType,
+          ck::ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool OutputIndex>
+struct ReferencePoolingFwd : public device::BaseOperator
+{
+    using ReduceOperation = typename ck::reduce_binary_operator<ReduceOpId>::opType;
+
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<InDataType>& in,
+                 Tensor<OutDataType>& out,
+                 Tensor<IndexDataType>& out_indices,
+                 const std::vector<ck::index_t>& window_spatial_lengths,
+                 const std::vector<ck::index_t>& window_strides,
+                 const std::vector<ck::index_t>& window_dilations,
+                 const std::vector<ck::index_t>& in_left_pads,
+                 const std::vector<ck::index_t>& /*in_right_pads*/)
+            : in_(in),
+              out_(out),
+              out_indices_(out_indices),
+              window_spatial_lengths_(window_spatial_lengths),
+              window_strides_(window_strides),
+              window_dilations_(window_dilations),
+              in_left_pads_(in_left_pads),
+              reduceLength_(1)
+        {
+            static_for<0, WindowRank, 1>{}(
+                [&](auto I) { reduceLength_ *= window_spatial_lengths[I]; });
+        }
+
+        const Tensor<InDataType>& in_;
+        Tensor<OutDataType>& out_;
+        Tensor<IndexDataType>& out_indices_;
+        const std::vector<ck::index_t>& window_spatial_lengths_;
+        const std::vector<ck::index_t>& window_strides_;
+        const std::vector<ck::index_t>& window_dilations_;
+        const std::vector<ck::index_t>& in_left_pads_;
+        int reduceLength_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        float RunPooling3dFwd(const Argument& arg)
+        {
+
+            auto elementwise_ops =
+                ck::reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+                    arg.reduceLength_);
+
+            auto in_elementwise_op  = std::get<0>(elementwise_ops);
+            auto acc_elementwise_op = std::get<1>(elementwise_ops);
+
+            if constexpr(!OutputIndex)
+            {
+                using Accumulation = ck::detail::
+                    AccumulateWithNanCheck<PropagateNan, ReduceOperation, ComputeDataType>;
+
+                auto f_ncdhw = [&](auto n, auto c, auto do_, auto ho, auto wo) {
+                    auto accuVal = ReduceOperation::template GetIdentityValue<ComputeDataType>();
+
+                    for(ck::index_t z = 0; z < arg.window_spatial_lengths_[0]; ++z)
+                    {
+                        ck::index_t di = do_ * arg.window_strides_[0] +
+                                         z * arg.window_dilations_[0] - arg.in_left_pads_[0];
+                        for(ck::index_t y = 0; y < arg.window_spatial_lengths_[1]; ++y)
+                        {
+                            ck::index_t hi = ho * arg.window_strides_[1] +
+                                             y * arg.window_dilations_[1] - arg.in_left_pads_[1];
+                            for(ck::index_t x = 0; x < arg.window_spatial_lengths_[2]; ++x)
+                            {
+                                ck::index_t wi = wo * arg.window_strides_[2] +
+                                                 x * arg.window_dilations_[2] -
+                                                 arg.in_left_pads_[2];
+                                if(di >= 0 &&
+                                   di < static_cast<ck::index_t>(arg.in_.mDesc.GetLengths()[2]) &&
+                                   hi >= 0 &&
+                                   hi < static_cast<ck::index_t>(arg.in_.mDesc.GetLengths()[3]) &&
+                                   wi >= 0 &&
+                                   wi < static_cast<ck::index_t>(arg.in_.mDesc.GetLengths()[4]))
+                                {
+                                    ComputeDataType currVal = ck::type_convert<ComputeDataType>(
+                                        arg.in_(n, c, di, hi, wi));
+
+                                    in_elementwise_op(currVal, currVal);
+
+                                    Accumulation::Calculate(accuVal, currVal);
+                                }
+                            }
+                        }
+                    }
+                    acc_elementwise_op(accuVal, accuVal);
+
+                    arg.out_(n, c, do_, ho, wo) = ck::type_convert<OutDataType>(accuVal);
+                };
+
+                make_ParallelTensorFunctor(f_ncdhw,
+                                           arg.out_.mDesc.GetLengths()[0],
+                                           arg.out_.mDesc.GetLengths()[1],
+                                           arg.out_.mDesc.GetLengths()[2],
+                                           arg.out_.mDesc.GetLengths()[3],
+                                           arg.out_.mDesc.GetLengths()[4])(
+                    std::thread::hardware_concurrency());
+            }
+            else
+            {
+                using Accumulation = ck::detail::AccumulateWithIndexAndNanCheck<PropagateNan,
+                                                                                ReduceOperation,
+                                                                                ComputeDataType,
+                                                                                IndexDataType>;
+
+                auto f_ncdhw = [&](auto n, auto c, auto do_, auto ho, auto wo) {
+                    auto accuVal = ReduceOperation::template GetIdentityValue<ComputeDataType>();
+                    IndexDataType accuIndex = 0;
+
+                    for(ck::index_t z = 0; z < arg.window_spatial_lengths_[0]; ++z)
+                    {
+                        ck::index_t di = do_ * arg.window_strides_[0] +
+                                         z * arg.window_dilations_[0] - arg.in_left_pads_[0];
+                        for(ck::index_t y = 0; y < arg.window_spatial_lengths_[1]; ++y)
+                        {
+                            ck::index_t hi = ho * arg.window_strides_[1] +
+                                             y * arg.window_dilations_[1] - arg.in_left_pads_[1];
+                            for(ck::index_t x = 0; x < arg.window_spatial_lengths_[2]; ++x)
+                            {
+                                ck::index_t wi = wo * arg.window_strides_[2] +
+                                                 x * arg.window_dilations_[2] -
+                                                 arg.in_left_pads_[2];
+                                if(di >= 0 &&
+                                   di < static_cast<ck::index_t>(arg.in_.mDesc.GetLengths()[2]) &&
+                                   hi >= 0 &&
+                                   hi < static_cast<ck::index_t>(arg.in_.mDesc.GetLengths()[3]) &&
+                                   wi >= 0 &&
+                                   wi < static_cast<ck::index_t>(arg.in_.mDesc.GetLengths()[4]))
+                                {
+                                    ComputeDataType currVal = ck::type_convert<ComputeDataType>(
+                                        arg.in_(n, c, di, hi, wi));
+                                    IndexDataType currIndex =
+                                        arg.in_.GetOffsetFromMultiIndex(n, c, di, hi, wi);
+
+                                    in_elementwise_op(currVal, currVal);
+
+                                    Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
+                                }
+                            }
+                        }
+                    }
+
+                    acc_elementwise_op(accuVal, accuVal);
+
+                    arg.out_(n, c, do_, ho, wo)         = ck::type_convert<OutDataType>(accuVal);
+                    arg.out_indices_(n, c, do_, ho, wo) = accuIndex;
+                };
+
+                make_ParallelTensorFunctor(f_ncdhw,
+                                           arg.out_.mDesc.GetLengths()[0],
+                                           arg.out_.mDesc.GetLengths()[1],
+                                           arg.out_.mDesc.GetLengths()[2],
+                                           arg.out_.mDesc.GetLengths()[3],
+                                           arg.out_.mDesc.GetLengths()[4])(
+                    std::thread::hardware_concurrency());
+            };
+
+            return 0;
+        }
+
+        float RunPooling2dFwd(const Argument& arg)
+        {
+
+            auto elementwise_ops =
+                ck::reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+                    arg.reduceLength_);
+
+            auto in_elementwise_op  = std::get<0>(elementwise_ops);
+            auto acc_elementwise_op = std::get<1>(elementwise_ops);
+
+            if constexpr(!OutputIndex)
+            {
+                using Accumulation = ck::detail::
+                    AccumulateWithNanCheck<PropagateNan, ReduceOperation, ComputeDataType>;
+
+                auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
+                    auto accuVal = ReduceOperation::template GetIdentityValue<ComputeDataType>();
+
+                    for(ck::index_t y = 0; y < arg.window_spatial_lengths_[0]; ++y)
+                    {
+                        ck::index_t hi = ho * arg.window_strides_[0] +
+                                         y * arg.window_dilations_[0] - arg.in_left_pads_[0];
+                        for(ck::index_t x = 0; x < arg.window_spatial_lengths_[1]; ++x)
+                        {
+                            ck::index_t wi = wo * arg.window_strides_[1] +
+                                             x * arg.window_dilations_[1] - arg.in_left_pads_[1];
+                            if(hi >= 0 &&
+                               hi < static_cast<ck::index_t>(arg.in_.mDesc.GetLengths()[2]) &&
+                               wi >= 0 &&
+                               wi < static_cast<ck::index_t>(arg.in_.mDesc.GetLengths()[3]))
+                            {
+                                ComputeDataType currVal =
+                                    ck::type_convert<ComputeDataType>(arg.in_(n, c, hi, wi));
+
+                                in_elementwise_op(currVal, currVal);
+
+                                Accumulation::Calculate(accuVal, currVal);
+                            }
+                        }
+                    }
+
+                    acc_elementwise_op(accuVal, accuVal);
+                    arg.out_(n, c, ho, wo) = ck::type_convert<OutDataType>(accuVal);
+                };
+
+                make_ParallelTensorFunctor(f_nchw,
+                                           arg.out_.mDesc.GetLengths()[0],
+                                           arg.out_.mDesc.GetLengths()[1],
+                                           arg.out_.mDesc.GetLengths()[2],
+                                           arg.out_.mDesc.GetLengths()[3])(
+                    std::thread::hardware_concurrency());
+            }
+            else
+            {
+                using Accumulation = ck::detail::AccumulateWithIndexAndNanCheck<PropagateNan,
+                                                                                ReduceOperation,
+                                                                                ComputeDataType,
+                                                                                IndexDataType>;
+
+                auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
+                    auto accuVal = ReduceOperation::template GetIdentityValue<ComputeDataType>();
+                    IndexDataType accuIndex = 0;
+
+                    for(ck::index_t y = 0; y < arg.window_spatial_lengths_[0]; ++y)
+                    {
+                        ck::index_t hi = ho * arg.window_strides_[0] +
+                                         y * arg.window_dilations_[0] - arg.in_left_pads_[0];
+                        for(ck::index_t x = 0; x < arg.window_spatial_lengths_[1]; ++x)
+                        {
+                            ck::index_t wi = wo * arg.window_strides_[1] +
+                                             x * arg.window_dilations_[1] - arg.in_left_pads_[1];
+                            if(hi >= 0 &&
+                               hi < static_cast<ck::index_t>(arg.in_.mDesc.GetLengths()[2]) &&
+                               wi >= 0 &&
+                               wi < static_cast<ck::index_t>(arg.in_.mDesc.GetLengths()[3]))
+                            {
+                                ComputeDataType currVal =
+                                    ck::type_convert<ComputeDataType>(arg.in_(n, c, hi, wi));
+
+                                IndexDataType currIndex =
+                                    arg.in_.GetOffsetFromMultiIndex(n, c, hi, wi);
+
+                                in_elementwise_op(currVal, currVal);
+
+                                Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
+                            }
+                        }
+                    }
+
+                    acc_elementwise_op(accuVal, accuVal);
+                    arg.out_(n, c, ho, wo)         = ck::type_convert<OutDataType>(accuVal);
+                    arg.out_indices_(n, c, ho, wo) = accuIndex;
+                };
+
+                make_ParallelTensorFunctor(f_nchw,
+                                           arg.out_.mDesc.GetLengths()[0],
+                                           arg.out_.mDesc.GetLengths()[1],
+                                           arg.out_.mDesc.GetLengths()[2],
+                                           arg.out_.mDesc.GetLengths()[3])(
+                    std::thread::hardware_concurrency());
+            };
+
+            return 0;
+        }
+
+        float Run(const Argument& arg)
+        {
+            // TODO - support generic pooling
+            if constexpr(InOutRank == 5 && WindowRank == 3)
+                return RunPooling3dFwd(arg);
+            else if constexpr(InOutRank == 4 && WindowRank == 2)
+                return RunPooling2dFwd(arg);
+            else
+                throw std::runtime_error("Only support pooling3d or pooling2d so far");
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<InDataType>& in,
+                             Tensor<OutDataType>& out,
+                             Tensor<IndexDataType>& out_indices,
+                             const std::vector<ck::index_t>& window_spatial_lengths,
+                             const std::vector<ck::index_t>& window_strides,
+                             const std::vector<ck::index_t>& window_dilations,
+                             const std::vector<ck::index_t>& in_left_pads,
+                             const std::vector<ck::index_t>& in_right_pads)
+    {
+        return Argument{in,
+                        out,
+                        out_indices,
+                        window_spatial_lengths,
+                        window_strides,
+                        window_dilations,
+                        in_left_pads,
+                        in_right_pads};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferencePoolingFwd"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_reduce.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_reduce.hpp
new file mode 100644
index 0000000000..944f34007e
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_reduce.hpp
@@ -0,0 +1,435 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <array>
+#include <algorithm>
+#include <thread>
+
+#include "ck/ck.hpp"
+#include "ck/utility/ignore.hpp"
+#include "ck/utility/reduction_common.hpp"
+#include "ck/utility/reduction_functions_accumulate.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOperation,
+          typename AccElementwiseOperation,
+          bool PropagateNan,
+          bool OutputIndex>
+struct ReferenceReduce : public device::DeviceReduce<InDataType,
+                                                     AccDataType,
+                                                     OutDataType,
+                                                     Rank,
+                                                     NumReduceDim,
+                                                     ReduceOperation,
+                                                     InElementwiseOperation,
+                                                     AccElementwiseOperation,
+                                                     PropagateNan,
+                                                     OutputIndex>
+{
+    using IndexDataType = int32_t;
+
+    static constexpr int NumInvariantDim = Rank - NumReduceDim;
+
+    static constexpr index_t NumSrcDim = Rank;
+    static constexpr index_t NumDstDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
+    static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const std::array<index_t, Rank> inLengths,
+                 const std::array<index_t, Rank> inStrides,
+                 const std::array<index_t, NumDstDim> outLengths,
+                 const std::array<index_t, NumDstDim> outStrides,
+                 const std::array<int, NumReduceDim> reduceDims,
+                 double alpha,
+                 double beta,
+                 const InDataType* in_host,
+                 OutDataType* out_host,
+                 IndexDataType* out_index_host,
+                 const InElementwiseOperation in_elementwise_op,
+                 const AccElementwiseOperation acc_elementwise_op)
+            : reduceDims_(reduceDims),
+              outLengths_(outLengths),
+              outStrides_(outStrides),
+              in_host_(in_host),
+              out_host_(out_host),
+              out_index_host_(out_index_host),
+              in_elementwise_op_(in_elementwise_op),
+              acc_elementwise_op_(acc_elementwise_op)
+        {
+            using ck::host_common::get_index_set;
+
+            if(std::any_of(
+                   reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+                throw std::runtime_error("Invalid reduce dimensions!");
+
+            if constexpr(NumInvariantDim > 0)
+            {
+                // get invariant_dims[] and invariant_lengths[]
+                for(int dim = 0, i = 0; dim < Rank; dim++)
+                    if(std::none_of(
+                           reduceDims.begin(), reduceDims.end(), [&](int d) { return d == dim; }))
+                    {
+                        invariantDims_[i]     = dim;
+                        invariant_lengths_[i] = inLengths[dim];
+                        i++;
+                    };
+            };
+
+            // get reduce_lengths_[]
+            for(int j = 0, i = 0; j < NumReduceDim; j++)
+            {
+                int dim              = reduceDims[j];
+                reduce_lengths_[i++] = inLengths[dim];
+            };
+
+            if constexpr(NumInvariantDim > 0)
+            {
+                // check invariant_lengths_ and outLengths
+                for(int i = 0; i < NumInvariantDim; i++)
+                    if(invariant_lengths_[i] != outLengths_[i])
+                        throw std::runtime_error("Invalid lengths parameters!");
+            }
+
+            if constexpr(NumInvariantDim > 0)
+            {
+                for(int j = 0, i = 0; j < NumInvariantDim; j++)
+                {
+                    int dim                  = invariantDims_[j];
+                    in_invariant_strides_[i] = inStrides[dim];
+                    i++;
+                };
+            };
+
+            for(int j = 0, i = 0; j < NumReduceDim; j++)
+            {
+                int dim               = reduceDims_[j];
+                in_reduce_strides_[i] = inStrides[dim];
+                i++;
+            };
+
+            if constexpr(NumInvariantDim > 0)
+                invariant_index_set_ = get_index_set<NumInvariantDim>(invariant_lengths_);
+
+            reduce_index_set_ = get_index_set<NumReduceDim>(reduce_lengths_);
+
+            alpha_ = type_convert<AccDataType>(alpha);
+            beta_  = type_convert<AccDataType>(beta);
+        };
+
+        const std::array<int, NumReduceDim> reduceDims_;
+        std::array<int, NumInvariantDim> invariantDims_;
+        std::array<index_t, NumInvariantDim> invariant_lengths_;
+        std::array<index_t, NumReduceDim> reduce_lengths_;
+
+        const std::array<index_t, NumDstDim> outLengths_;
+        const std::array<index_t, NumDstDim> outStrides_;
+
+        std::array<index_t, NumInvariantDim> in_invariant_strides_;
+        std::array<index_t, NumReduceDim> in_reduce_strides_;
+
+        const InDataType* in_host_;
+        OutDataType* out_host_;
+        IndexDataType* out_index_host_;
+        const InElementwiseOperation in_elementwise_op_;
+        const AccElementwiseOperation acc_elementwise_op_;
+
+        AccDataType alpha_;
+        AccDataType beta_;
+
+        std::vector<std::array<index_t, NumInvariantDim>> invariant_index_set_;
+        std::vector<std::array<index_t, NumReduceDim>> reduce_index_set_;
+    };
+
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            ignore = stream_config;
+
+            using ck::float_equal_one;
+            using ck::float_equal_zero;
+            using ck::type_convert;
+            using ck::host_common::get_index_set;
+            using ck::host_common::get_offset_from_index;
+
+            if constexpr(OutputIndex)
+            {
+                using Accumulation = ck::detail::AccumulateWithIndexAndNanCheck<PropagateNan,
+                                                                                ReduceOperation,
+                                                                                AccDataType,
+                                                                                IndexDataType>;
+
+                if constexpr(NumInvariantDim == 0)
+                {
+                    AccDataType accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+                    IndexDataType accuIndex = 0;
+
+                    for(std::size_t i = 0; i < arg.reduce_index_set_.size(); i++)
+                    {
+                        auto in_offset = get_offset_from_index<NumReduceDim>(
+                            arg.in_reduce_strides_, arg.reduce_index_set_[i]);
+
+                        auto currVal = type_convert<AccDataType>(arg.in_host_[in_offset]);
+
+                        arg.in_elementwise_op_(currVal, currVal);
+
+                        auto currIndex = static_cast<IndexDataType>(i);
+
+                        Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
+                    };
+
+                    arg.acc_elementwise_op_(accuVal, accuVal);
+
+                    if(!float_equal_one{}(arg.alpha_))
+                        accuVal *= type_convert<AccDataType>(arg.alpha_);
+
+                    if(!float_equal_zero{}(arg.beta_))
+                        accuVal += type_convert<AccDataType>(arg.out_host_[0]) *
+                                   type_convert<AccDataType>(arg.beta_);
+
+                    arg.out_host_[0]       = type_convert<OutDataType>(accuVal);
+                    arg.out_index_host_[0] = accuIndex;
+                }
+                else
+                {
+                    auto thread_reduce_func = [&](auto invariant_index) {
+                        AccDataType accuVal =
+                            ReduceOperation::template GetIdentityValue<AccDataType>();
+                        IndexDataType accuIndex = 0;
+
+                        auto in_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                            arg.in_invariant_strides_, invariant_index);
+
+                        for(std::size_t i = 0; i < arg.reduce_index_set_.size(); i++)
+                        {
+                            auto in_reduce_offset = get_offset_from_index<NumReduceDim>(
+                                arg.in_reduce_strides_, arg.reduce_index_set_[i]);
+
+                            auto currVal = type_convert<AccDataType>(
+                                arg.in_host_[in_invariant_offset + in_reduce_offset]);
+
+                            arg.in_elementwise_op_(currVal, currVal);
+
+                            auto currIndex = static_cast<IndexDataType>(i);
+
+                            Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
+                        };
+
+                        arg.acc_elementwise_op_(accuVal, accuVal);
+
+                        if(!float_equal_one{}(arg.alpha_))
+                            accuVal *= type_convert<AccDataType>(arg.alpha_);
+
+                        auto dst_offset = get_offset_from_index<NumInvariantDim>(arg.outStrides_,
+                                                                                 invariant_index);
+
+                        if(!float_equal_zero{}(arg.beta_))
+                            accuVal += type_convert<AccDataType>(arg.out_host_[dst_offset]) *
+                                       type_convert<AccDataType>(arg.beta_);
+
+                        arg.out_host_[dst_offset]       = type_convert<OutDataType>(accuVal);
+                        arg.out_index_host_[dst_offset] = accuIndex;
+                    };
+
+                    std::size_t num_thread = std::thread::hardware_concurrency();
+
+                    std::size_t work_per_thread =
+                        (arg.invariant_index_set_.size() + num_thread - 1) / num_thread;
+
+                    std::vector<joinable_thread> threads(num_thread);
+
+                    for(std::size_t it = 0; it < num_thread; ++it)
+                    {
+                        std::size_t i_begin = it * work_per_thread;
+                        std::size_t i_end =
+                            std::min((it + 1) * work_per_thread, arg.invariant_index_set_.size());
+
+                        auto f = [=] {
+                            for(std::size_t i = i_begin; i < i_end; i++)
+                            {
+                                thread_reduce_func(arg.invariant_index_set_[i]);
+                            }
+                        };
+
+                        threads[it] = joinable_thread(f);
+                    }
+                };
+            }
+            else
+            {
+                using Accumulation =
+                    ck::detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
+
+                if constexpr(NumInvariantDim == 0)
+                {
+                    AccDataType accuVal = ReduceOperation::template GetIdentityValue<AccDataType>();
+
+                    for(const auto& reduce_index : arg.reduce_index_set_)
+                    {
+                        auto in_offset = get_offset_from_index<NumReduceDim>(arg.in_reduce_strides_,
+                                                                             reduce_index);
+
+                        auto currVal = type_convert<AccDataType>(arg.in_host_[in_offset]);
+
+                        arg.in_elementwise_op_(currVal, currVal);
+
+                        Accumulation::Calculate(accuVal, currVal);
+                    };
+
+                    arg.acc_elementwise_op_(accuVal, accuVal);
+
+                    if(!float_equal_one{}(arg.alpha_))
+                        accuVal *= type_convert<AccDataType>(arg.alpha_);
+
+                    if(!float_equal_zero{}(arg.beta_))
+                        accuVal += type_convert<AccDataType>(arg.out_host_[0]) *
+                                   type_convert<AccDataType>(arg.beta_);
+
+                    arg.out_host_[0] = type_convert<OutDataType>(accuVal);
+                }
+                else
+                {
+                    auto thread_reduce_func = [&](auto invariant_index) {
+                        AccDataType accuVal =
+                            ReduceOperation::template GetIdentityValue<AccDataType>();
+
+                        auto in_invariant_offset = get_offset_from_index<NumInvariantDim>(
+                            arg.in_invariant_strides_, invariant_index);
+
+                        for(const auto& reduce_index : arg.reduce_index_set_)
+                        {
+                            auto in_reduce_offset = get_offset_from_index<NumReduceDim>(
+                                arg.in_reduce_strides_, reduce_index);
+
+                            auto currVal = type_convert<AccDataType>(
+                                arg.in_host_[in_invariant_offset + in_reduce_offset]);
+
+                            arg.in_elementwise_op_(currVal, currVal);
+
+                            Accumulation::Calculate(accuVal, currVal);
+                        };
+
+                        arg.acc_elementwise_op_(accuVal, accuVal);
+
+                        if(!float_equal_one{}(arg.alpha_))
+                            accuVal *= type_convert<AccDataType>(arg.alpha_);
+
+                        auto dst_offset = get_offset_from_index<NumInvariantDim>(arg.outStrides_,
+                                                                                 invariant_index);
+
+                        if(!float_equal_zero{}(arg.beta_))
+                            accuVal += type_convert<AccDataType>(arg.out_host_[dst_offset]) *
+                                       type_convert<AccDataType>(arg.beta_);
+
+                        arg.out_host_[dst_offset] = type_convert<OutDataType>(accuVal);
+                    };
+
+                    std::size_t num_thread = std::thread::hardware_concurrency();
+
+                    std::size_t work_per_thread =
+                        (arg.invariant_index_set_.size() + num_thread - 1) / num_thread;
+
+                    std::vector<joinable_thread> threads(num_thread);
+
+                    for(std::size_t it = 0; it < num_thread; ++it)
+                    {
+                        std::size_t i_begin = it * work_per_thread;
+                        std::size_t i_end =
+                            std::min((it + 1) * work_per_thread, arg.invariant_index_set_.size());
+
+                        auto f = [=] {
+                            for(std::size_t i = i_begin; i < i_end; i++)
+                            {
+                                thread_reduce_func(arg.invariant_index_set_[i]);
+                            }
+                        };
+
+                        threads[it] = joinable_thread(f);
+                    }
+                };
+            };
+
+            return (0.0f);
+        };
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        };
+    };
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        ignore = p_arg;
+
+        return true;
+    };
+
+    std::unique_ptr<device::BaseArgument>
+    MakeArgumentPointer(const std::array<index_t, Rank> inLengths,
+                        const std::array<index_t, Rank> inStrides,
+                        const std::array<index_t, NumDstDim> outLengths,
+                        const std::array<index_t, NumDstDim> outStrides,
+                        const std::array<int, NumReduceDim> reduceDims,
+                        double alpha,
+                        double beta,
+                        const void* in_host,
+                        const void* in_index_host,
+                        void* out_host,
+                        void* out_index_host,
+                        const InElementwiseOperation in_elementwise_op,
+                        const AccElementwiseOperation acc_elementwise_op) override
+    {
+        ignore = in_index_host;
+
+        return std::make_unique<Argument>(inLengths,
+                                          inStrides,
+                                          outLengths,
+                                          outStrides,
+                                          reduceDims,
+                                          alpha,
+                                          beta,
+                                          static_cast<const InDataType*>(in_host),
+                                          static_cast<OutDataType*>(out_host),
+                                          static_cast<IndexDataType*>(out_index_host),
+                                          in_elementwise_op,
+                                          acc_elementwise_op);
+    };
+
+    std::unique_ptr<device::BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    };
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "Reference_Reduce<" << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp
new file mode 100644
index 0000000000..9916a03b9c
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp
@@ -0,0 +1,178 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <vector>
+#include <algorithm>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename InDataType, typename OutDataType, typename AccDataType>
+struct ReferenceSoftmax : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<InDataType>& in,
+                 Tensor<OutDataType>& out,
+                 double alpha,
+                 double beta,
+                 const std::vector<index_t> sm_reduce_dims)
+            : in_(in), out_(out), sm_reduce_dims_(sm_reduce_dims)
+        {
+            alpha_ = static_cast<AccDataType>(alpha);
+            beta_  = static_cast<AccDataType>(beta);
+
+            // std::cout << "debug: scalar dims: ";
+            for(size_t i = 0; i < in.mDesc.GetNumOfDimension(); i++)
+            {
+                if(std::find(sm_reduce_dims.begin(), sm_reduce_dims.end(), i) ==
+                   sm_reduce_dims.end())
+                {
+                    sm_scalar_dims_.push_back(i);
+                    // std::cout << i << ", ";
+                }
+            }
+            // std::cout << std::endl;
+        }
+
+        const Tensor<InDataType>& in_;
+        Tensor<OutDataType>& out_;
+        AccDataType alpha_;
+        AccDataType beta_;
+        std::vector<index_t> sm_reduce_dims_;
+        std::vector<index_t> sm_scalar_dims_; // dim after internal max/sum reduction
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            std::vector<size_t> scalar_lengths;
+            for(index_t dim : arg.sm_scalar_dims_)
+            {
+                scalar_lengths.push_back(arg.in_.mDesc.GetLengths()[dim]);
+            }
+            // max and sum reduction with final reduced values of dim=0 is a scalar so give it
+            // appropriate lengths of {1}
+            if(arg.sm_scalar_dims_.size() == 0)
+            {
+                scalar_lengths.push_back(1);
+            }
+
+            Tensor<AccDataType> reduce_max(scalar_lengths);
+            reduce_max.GenerateTensorValue(
+                GeneratorTensor_1<AccDataType>{std::numeric_limits<AccDataType>::lowest()});
+            Tensor<AccDataType> reduce_sum(scalar_lengths);
+            reduce_sum.GenerateTensorValue(GeneratorTensor_1<AccDataType>{0});
+
+            // when final reduced values is of dim=0, the index will be transformed into empty
+            // std::vector which is actually a valid input for Tensor::operator(std::vector) and
+            // internally accesses 0'th element
+            auto to_sm_scalar_idx = [&](auto idx) {
+                std::vector<size_t> sm_scalar_idx;
+                for(index_t dim : arg.sm_scalar_dims_)
+                {
+                    sm_scalar_idx.push_back(idx[dim]);
+                }
+                return sm_scalar_idx;
+            };
+
+            arg.in_.ForEach([&](auto& self, auto idx) {
+                reduce_max(to_sm_scalar_idx(idx)) = std::max(
+                    reduce_max(to_sm_scalar_idx(idx)), ck::type_convert<AccDataType>(self(idx)));
+            });
+
+            // LogRangeAsType<float>(std::cout << "reduce_max: ", reduce_max.mData, ",") <<
+            // std::endl;
+
+            Tensor<AccDataType> in_stable(arg.in_.mDesc);
+            in_stable.ForEach([&](auto& self, auto idx) {
+                // numerator = exp(x - max(x))
+                self(idx) = std::exp(ck::type_convert<AccDataType>(arg.in_(idx)) -
+                                     reduce_max(to_sm_scalar_idx(idx)));
+            });
+
+            // LogRangeAsType<float>(std::cout << "in_stable: ", in_stable.mData, ",") << std::endl;
+
+            in_stable.ForEach([&](auto& self, auto idx) {
+                // denominator = sum(exp(x - max(x)))
+                reduce_sum(to_sm_scalar_idx(idx)) += self(idx);
+            });
+
+            // LogRangeAsType<float>(std::cout << "reduce_sum: ", reduce_sum.mData, ",") <<
+            // std::endl;
+
+            arg.out_.ForEach([&](auto& self, auto idx) {
+                AccDataType temp_result =
+                    arg.alpha_ * in_stable(idx) / reduce_sum(to_sm_scalar_idx(idx)) +
+                    arg.beta_ * self(idx);
+                self(idx) = ck::type_convert<OutDataType>(temp_result);
+            });
+
+            // LogRangeAsType<float>(std::cout << "out: ", arg.out_.mData, ",") << std::endl;
+            // reduction along reduce dims
+            // LogRangeAsType<float>(std::cout << "reduce_max: ", reduce_max.mData, ",") <<
+            // std::endl; LogRangeAsType<float>(std::cout << "reduce_sum: ", reduce_sum.mData, ",")
+            // << std::endl;
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<InDataType>& in,
+                             Tensor<OutDataType>& out,
+                             double alpha,
+                             double beta,
+                             const std::vector<index_t> sm_reduce_dims)
+    {
+        return Argument{in, out, alpha, beta, sm_reduce_dims};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceSoftmax"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp
new file mode 100644
index 0000000000..f949260ca4
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp
@@ -0,0 +1,205 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <vector>
+#include <algorithm>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename EmbType,
+          typename IndexType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename OutType>
+struct ReferenceSparseEmbedding3ForwardLayernorm : public device::BaseOperator
+{
+    struct Argument : public device::BaseArgument
+    {
+        Argument(Tensor<OutType>& output,
+                 const Tensor<EmbType>& emb_a,
+                 const Tensor<EmbType>& emb_b,
+                 const Tensor<EmbType>& emb_c,
+                 const Tensor<IndexType>& index_a,
+                 const Tensor<IndexType>& index_b,
+                 const Tensor<IndexType>& index_c,
+                 const Tensor<GammaDataType>& gamma,
+                 const Tensor<BetaDataType>& beta,
+                 ck::index_t NumRows,
+                 ck::index_t EmbeddingDim,
+                 ck::index_t IndexLength,
+                 AccDataType epsilon)
+            : output_(output),
+              emb_a_(emb_a),
+              emb_b_(emb_b),
+              emb_c_(emb_c),
+              index_a_(index_a),
+              index_b_(index_b),
+              index_c_(index_c),
+              gamma_(gamma),
+              beta_(beta),
+              NumRows_(NumRows),
+              EmbeddingDim_(EmbeddingDim),
+              IndexLength_(IndexLength),
+              epsilon_(epsilon)
+        {
+        }
+        Tensor<OutType>& output_;
+        const Tensor<EmbType> emb_a_;
+        const Tensor<EmbType> emb_b_;
+        const Tensor<EmbType> emb_c_;
+        const Tensor<IndexType> index_a_;
+        const Tensor<IndexType> index_b_;
+        const Tensor<IndexType> index_c_;
+        const Tensor<GammaDataType> gamma_;
+        const Tensor<BetaDataType> beta_;
+        ck::index_t NumRows_;
+        ck::index_t EmbeddingDim_;
+        ck::index_t IndexLength_;
+        AccDataType epsilon_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        float Run(const Argument& arg)
+        {
+            ck::index_t D = arg.EmbeddingDim_;
+            ck::index_t L = arg.IndexLength_;
+            ck::index_t E = arg.NumRows_;
+
+            Tensor<AccDataType> accumulator({L, D});
+
+            Tensor<AccDataType> mean({L});
+            Tensor<AccDataType> var({L});
+
+            accumulator.SetZero();
+
+            auto f_emb_per_row = [&](auto idx) {
+                IndexType idx_a = arg.index_a_(idx);
+                IndexType idx_b = arg.index_b_(idx);
+                IndexType idx_c = arg.index_c_(idx);
+
+                if(!((idx_a < E) && (idx_b < E) && (idx_c < E)))
+                {
+                    throw(std::runtime_error("wrong! out of range"));
+                }
+
+                for(auto d = 0; d < D; d++)
+                {
+                    auto v_a = ck::type_convert<AccDataType>(arg.emb_a_(idx_a, d));
+                    auto v_b = ck::type_convert<AccDataType>(arg.emb_b_(idx_b, d));
+                    auto v_c = ck::type_convert<AccDataType>(arg.emb_c_(idx_c, d));
+
+                    accumulator(idx, d) += v_a + v_b + v_c;
+                }
+            };
+            make_ParallelTensorFunctor(f_emb_per_row, L)(std::thread::hardware_concurrency());
+
+            // layernorm
+            for(auto idx = 0; idx < L; ++idx)
+            {
+                mean(idx) = 0;
+                var(idx)  = 0;
+
+                for(auto d = 0; d < D; ++d)
+                {
+                    auto x_val = accumulator(idx, d);
+                    mean(idx) += x_val;
+                    var(idx) += x_val * x_val;
+                }
+
+                mean(idx) = mean(idx) / D;
+                var(idx)  = (var(idx) / D) - (mean(idx) * mean(idx));
+            }
+
+            for(auto idx = 0; idx < L; ++idx)
+            {
+                for(auto d = 0; d < D; ++d)
+                {
+                    auto x_val          = accumulator(idx, d);
+                    auto y_val          = (x_val - mean(idx)) / sqrt(var(idx) + arg.epsilon_);
+                    y_val               = (y_val * arg.gamma_(d)) + arg.beta_(d);
+                    arg.output_(idx, d) = ck::type_convert<OutType>(y_val);
+                }
+            }
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(Tensor<OutType>& output,
+                             const Tensor<EmbType>& emb_a,
+                             const Tensor<EmbType>& emb_b,
+                             const Tensor<EmbType>& emb_c,
+                             const Tensor<IndexType>& index_a,
+                             const Tensor<IndexType>& index_b,
+                             const Tensor<IndexType>& index_c,
+                             const Tensor<GammaDataType>& gamma,
+                             const Tensor<BetaDataType>& beta,
+                             ck::index_t NumRows,
+                             ck::index_t EmbeddingDim,
+                             ck::index_t IndexLength,
+                             AccDataType epsilon)
+    {
+        return Argument(output,
+                        emb_a,
+                        emb_b,
+                        emb_c,
+                        index_a,
+                        index_b,
+                        index_c,
+                        gamma,
+                        beta,
+                        NumRows,
+                        EmbeddingDim,
+                        IndexLength,
+                        epsilon);
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceSparseEmbedding3ForwardLayernorm"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp b/library/include/ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp
new file mode 100644
index 0000000000..0b7887efb6
--- /dev/null
+++ b/library/include/ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp
@@ -0,0 +1,125 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef NAIVE_CONV_FWD_HPP
+#define NAIVE_CONV_FWD_HPP
+
+namespace ck {
+namespace ref {
+
+/*
+ * \brief naive implementation of 3D convolution. Layout is (NDHWC, KZYXC, NDHWK).
+ *
+ * \param N number of batches
+ * \param K number of filters
+ * \param C number of channels of weight
+ * \param (Di, Hi, Wi) depth, height and width dimension of data
+ * \param (Z, Y, X) depth, height and width dimensions of weights
+ * \param (Do, Ho, Wo) depth, height and width dimension of output
+ * \param (stride_z, stride_y, stride_x) strides
+ * \param (dilation_z, dilation_y, dilation_x) dilations
+ * \param (pad_z, pad_y, pad_x) pads
+ */
+template <typename TIn,
+          typename TWei,
+          typename TOut,
+          typename TAcc,
+          typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+__global__ void naive_conv_fwd_ndhwc_kzyxc_ndhwk(const TIn* __restrict__ p_in,
+                                                 const TWei* __restrict__ p_wei,
+                                                 TOut* __restrict__ p_out,
+                                                 index_t N,
+                                                 index_t K,
+                                                 index_t C,
+                                                 index_t Di,
+                                                 index_t Hi,
+                                                 index_t Wi,
+                                                 index_t Z,
+                                                 index_t Y,
+                                                 index_t X,
+                                                 index_t Do,
+                                                 index_t Ho,
+                                                 index_t Wo,
+                                                 index_t stride_z,
+                                                 index_t stride_y,
+                                                 index_t stride_x,
+                                                 index_t dilation_z,
+                                                 index_t dilation_y,
+                                                 index_t dilation_x,
+                                                 index_t pad_z,
+                                                 index_t pad_y,
+                                                 index_t pad_x)
+{
+    const index_t tid                = blockIdx.x * blockDim.x + threadIdx.x;
+    const index_t num_threads        = blockDim.x * gridDim.x;
+    const long_index_t output_length = N * Do * Ho * Wo * K;
+
+    const index_t out_strides[] = {Do * Ho * Wo * K, Ho * Wo * K, Wo * K, K};
+    const index_t in_strides[]  = {Di * Hi * Wi * C, Hi * Wi * C, Wi * C, C};
+    const index_t wei_strides[] = {Z * Y * X * C, Y * X * C, X * C, C};
+
+    constexpr auto in_op  = InElementwiseOperation{};
+    constexpr auto wei_op = WeiElementwiseOperation{};
+    constexpr auto out_op = OutElementwiseOperation{};
+
+    TIn in_val;
+    TWei wei_val;
+    TOut out_val;
+
+    for(long_index_t ii = tid; ii < output_length; ii += num_threads)
+    {
+        const index_t n  = ii / out_strides[0];
+        index_t k        = ii - n * out_strides[0];
+        const index_t dO = k / out_strides[1];
+        k -= dO * out_strides[1];
+        const index_t ho = k / out_strides[2];
+        k -= ho * out_strides[2];
+        const index_t wo = k / out_strides[3];
+        k -= wo * out_strides[3];
+
+        TAcc acc = static_cast<TAcc>(0);
+
+        const TIn* in_n   = p_in + static_cast<long_index_t>(n) * in_strides[0];
+        const TWei* wei_k = p_wei + static_cast<long_index_t>(k) * wei_strides[0];
+
+        for(index_t z = 0; z < Z; ++z)
+        {
+            index_t di          = stride_z * dO - pad_z + dilation_z * z;
+            const TIn* in_n_di  = in_n + di * in_strides[1];
+            const TWei* wei_k_z = wei_k + z * wei_strides[1];
+
+            for(index_t y = 0; y < Y; ++y)
+            {
+                index_t hi            = stride_y * ho - pad_y + dilation_y * y;
+                const TIn* in_n_di_hi = in_n_di + hi * in_strides[2];
+                const TWei* wei_k_z_y = wei_k_z + y * wei_strides[2];
+
+                for(index_t x = 0; x < X; ++x)
+                {
+                    index_t wi               = stride_x * wo - pad_x + dilation_x * x;
+                    const TIn* in_n_di_hi_wi = in_n_di_hi + wi * in_strides[3];
+                    const TWei* wei_k_z_y_x  = wei_k_z_y + x * wei_strides[3];
+
+                    if(di >= 0 && di < Di && hi >= 0 && hi < Hi && wi >= 0 && wi < Wi)
+                    {
+                        for(index_t c = 0; c < C; ++c)
+                        {
+                            in_op(in_val, in_n_di_hi_wi[c]);
+                            wei_op(wei_val, wei_k_z_y_x[c]);
+                            acc += in_val * wei_val;
+                        }
+                    }
+                }
+            }
+        }
+
+        out_op(out_val, static_cast<TOut>(acc));
+        p_out[ii] = out_val;
+    }
+}
+} // namespace ref
+} // namespace ck
+
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/add_device_operation_instance.hpp b/library/include/ck/library/tensor_operation_instance/add_device_operation_instance.hpp
new file mode 100644
index 0000000000..f57fed9c07
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/add_device_operation_instance.hpp
@@ -0,0 +1,35 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <type_traits>
+
+#include "ck/utility/functional2.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <typename BaseOp, typename NewOpInstances>
+void add_device_operation_instances(std::vector<std::unique_ptr<BaseOp>>& op_instances,
+                                    const NewOpInstances& new_op_instances)
+{
+    ck::static_for<0, std::tuple_size_v<NewOpInstances>, 1>{}([&](auto i) {
+        const auto new_op_instance = std::get<i>(new_op_instances);
+
+        using NewOpInstance = remove_cvref_t<decltype(new_op_instance)>;
+
+        static_assert(std::is_base_of_v<BaseOp, NewOpInstance>,
+                      "wrong! NewOpInstance should be derived from BaseOp");
+
+        op_instances.push_back(std::make_unique<NewOpInstance>(new_op_instance));
+    });
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp b/library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
new file mode 100644
index 0000000000..5096f5c30b
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
@@ -0,0 +1,134 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/tuple.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// aliasing, for commonly used data type
+using F64  = double;
+using F32  = float;
+using F16  = ck::half_t;
+using F8   = ck::f8_t;
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using I32  = int32_t;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using F16_Tuple     = ck::Tuple<F16>;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using F64_Tuple     = ck::Tuple<F64>;
+using F32_Tuple     = ck::Tuple<F32>;
+using I32_Tuple     = ck::Tuple<I32>;
+using I32_F32_Tuple = ck::Tuple<I32, F32>;
+
+using F32_F32_Tuple = ck::Tuple<F32, F32>;
+
+// GEMM layout
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using Row_Tuple     = ck::Tuple<Row>;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+// Conv layout
+//
+using NWC   = ck::tensor_layout::convolution::NWC;
+using NHWC  = ck::tensor_layout::convolution::NHWC;
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+using KXC   = ck::tensor_layout::convolution::KXC;
+using KYXC  = ck::tensor_layout::convolution::KYXC;
+using KZYXC = ck::tensor_layout::convolution::KZYXC;
+
+using NWK   = ck::tensor_layout::convolution::NWK;
+using NHWK  = ck::tensor_layout::convolution::NHWK;
+using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+//
+using GNWC   = ck::tensor_layout::convolution::GNWC;
+using GNHWC  = ck::tensor_layout::convolution::GNHWC;
+using GNDHWC = ck::tensor_layout::convolution::GNDHWC;
+
+using GKXC   = ck::tensor_layout::convolution::GKXC;
+using GKYXC  = ck::tensor_layout::convolution::GKYXC;
+using GKZYXC = ck::tensor_layout::convolution::GKZYXC;
+
+using GNWK   = ck::tensor_layout::convolution::GNWK;
+using GNHWK  = ck::tensor_layout::convolution::GNHWK;
+using GNDHWK = ck::tensor_layout::convolution::GNDHWK;
+
+//
+using NWGC   = ck::tensor_layout::convolution::NWGC;
+using NHWGC  = ck::tensor_layout::convolution::NHWGC;
+using NDHWGC = ck::tensor_layout::convolution::NDHWGC;
+
+using KXGC   = ck::tensor_layout::convolution::KXGC;
+using KYXGC  = ck::tensor_layout::convolution::KYXGC;
+using KZYXGC = ck::tensor_layout::convolution::KZYXGC;
+
+using NWGK   = ck::tensor_layout::convolution::NWGK;
+using NHWGK  = ck::tensor_layout::convolution::NHWGK;
+using NDHWGK = ck::tensor_layout::convolution::NDHWGK;
+
+//
+using GK          = ck::tensor_layout::convolution::G_K;
+using GK_Tuple    = ck::Tuple<GK>;
+using GK_GK_Tuple = ck::Tuple<GK, GK>;
+
+// pointwise functor
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using Relu           = ck::tensor_operation::element_wise::Relu;
+using TanH           = ck::tensor_operation::element_wise::TanH;
+using Scale          = ck::tensor_operation::element_wise::Scale;
+using Bilinear       = ck::tensor_operation::element_wise::Bilinear;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+using AddFastGelu    = ck::tensor_operation::element_wise::AddFastGelu;
+using AddReluAdd     = ck::tensor_operation::element_wise::AddReluAdd;
+using FastGelu       = ck::tensor_operation::element_wise::FastGelu;
+using AddMultiply    = ck::tensor_operation::element_wise::AddMultiply;
+using MultiplyAdd    = ck::tensor_operation::element_wise::MultiplyAdd;
+using ScaleAdd       = ck::tensor_operation::element_wise::ScaleAdd;
+using Gelu           = ck::tensor_operation::element_wise::Gelu;
+using Swish          = ck::tensor_operation::element_wise::Swish;
+using Add            = ck::tensor_operation::element_wise::Add;
+
+template <typename Activation>
+using Activation_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Activation>;
+
+template <typename Activation>
+using Add_Activation_Mul_Clamp =
+    ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<Activation>;
+
+template <typename Activation>
+using Add_Mul_Activation_Mul_Clamp =
+    ck::tensor_operation::element_wise::Add_Mul_Activation_Mul_Clamp<Activation>;
+
+template <typename Activation>
+using Activation_Mul2_Clamp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<Activation>;
+
+template <typename Activation>
+using Add_Activation_Mul2_Clamp =
+    ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<Activation>;
+
+template <typename Activation>
+using Add_Mul2_Activation_Mul_Clamp =
+    ck::tensor_operation::element_wise::Add_Mul2_Activation_Mul_Clamp<Activation>;
+
+template <typename DeviceOp, typename Tag = void>
+struct DeviceOperationInstanceFactory;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp b/library/include/ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp
new file mode 100644
index 0000000000..949e1d2dd0
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+#ifdef CK_ENABLE_FP16
+void add_device_avgpool_bwd_ndhwc_f16_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, F16, F16, NDHWC, NDHWC>>>&);
+#endif
+#ifdef CK_ENABLE_BF16
+void add_device_avgpool_bwd_ndhwc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, BF16, BF16, NDHWC, NDHWC>>>&);
+#endif
+#ifdef CK_ENABLE_FP32
+void add_device_avgpool_bwd_ndhwc_f32_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, F32, F32, NDHWC, NDHWC>>>&);
+#endif
+template <typename DOutDataType, typename DInDataType, typename InLayout, typename OutLayout>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::
+        DeviceAvgPoolBwd<3, DOutDataType, DInDataType, InLayout, OutLayout>>
+{
+    using DeviceOp = DeviceAvgPoolBwd<3, DOutDataType, DInDataType, InLayout, OutLayout>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+        if constexpr(is_same_v<InLayout, NDHWC> && is_same_v<OutLayout, NDHWC>)
+        {
+#ifdef CK_ENABLE_FP16
+            if constexpr(is_same_v<DOutDataType, F16> && is_same_v<DInDataType, F16>)
+                add_device_avgpool_bwd_ndhwc_f16_instances(op_ptrs);
+#endif
+#ifdef CK_ENABLE_BF16
+            else if constexpr(is_same_v<DOutDataType, BF16> && is_same_v<DInDataType, BF16>)
+                add_device_avgpool_bwd_ndhwc_bf16_instances(op_ptrs);
+#endif
+#ifdef CK_ENABLE_FP32
+            else if constexpr(is_same_v<DOutDataType, F32> && is_same_v<DInDataType, F32>)
+                add_device_avgpool_bwd_ndhwc_f32_instances(op_ptrs);
+#endif
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp
new file mode 100644
index 0000000000..8f15e80794
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp
@@ -0,0 +1,268 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef CK_ENABLE_BF16
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
+#ifdef CK_ENABLE_FP16
+void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
+#ifdef CK_ENABLE_FP32
+void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
+#ifdef CK_ENABLE_INT8
+void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Col,
+                                                  Row,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Col,
+                                                  Col,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Row,
+                                                  Row,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Row,
+                                                  Col,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+#endif
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatchedGemm<
+    ALayout,
+    BLayout,
+    CLayout,
+    ADataType,
+    BDataType,
+    CDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceBatchedGemm<ALayout,
+                                       BLayout,
+                                       CLayout,
+                                       ADataType,
+                                       BDataType,
+                                       CDataType,
+                                       ck::tensor_operation::element_wise::PassThrough,
+                                       ck::tensor_operation::element_wise::PassThrough,
+                                       ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+#ifdef CK_ENABLE_FP32
+        if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
+                     is_same_v<CDataType, float>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(op_ptrs);
+            }
+        }
+#endif
+#ifdef CK_ENABLE_FP16
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(op_ptrs);
+            }
+        }
+#endif
+#ifdef CK_ENABLE_BF16
+        if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, bhalf_t> &&
+                     is_same_v<CDataType, bhalf_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(op_ptrs);
+            }
+        }
+#endif
+#ifdef CK_ENABLE_INT8
+        if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
+                     is_same_v<CDataType, int8_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(op_ptrs);
+            }
+        }
+#endif
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp
new file mode 100644
index 0000000000..462b7f1102
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp
@@ -0,0 +1,140 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+using CDE0ElementOp = ck::tensor_operation::element_wise::AddRelu;
+using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
+#ifdef CK_ENABLE_FP16
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultipleDGemmMultipleD<Row,
+                                                                        Col,
+                                                                        ck::Tuple<Row>,
+                                                                        Row,
+                                                                        ck::Tuple<Row>,
+                                                                        Row,
+                                                                        F16,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        CDE0ElementOp,
+                                                                        PassThrough,
+                                                                        CDE1ElementOp>>>&
+        instances);
+
+void add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultipleDGemmMultipleD<Row,
+                                                                        Col,
+                                                                        ck::Tuple<Row>,
+                                                                        Col,
+                                                                        ck::Tuple<Row>,
+                                                                        Row,
+                                                                        F16,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        CDE0ElementOp,
+                                                                        PassThrough,
+                                                                        CDE1ElementOp>>>&
+        instances);
+
+template <typename A0Layout,
+          typename B0Layout,
+          typename D0sLayout,
+          typename B1Layout,
+          typename D1sLayout,
+          typename E1Layout,
+          typename A0DataType,
+          typename B0DataType,
+          typename D0sDataType,
+          typename B1DataType,
+          typename D1sDataType,
+          typename E1DataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedGemmMultipleDGemmMultipleD<A0Layout,
+                                                                          B0Layout,
+                                                                          D0sLayout,
+                                                                          B1Layout,
+                                                                          D1sLayout,
+                                                                          E1Layout,
+                                                                          A0DataType,
+                                                                          B0DataType,
+                                                                          D0sDataType,
+                                                                          B1DataType,
+                                                                          D1sDataType,
+                                                                          E1DataType,
+                                                                          PassThrough,
+                                                                          PassThrough,
+                                                                          CDE0ElementOp,
+                                                                          PassThrough,
+                                                                          CDE1ElementOp>>
+{
+    using DeviceOp = DeviceBatchedGemmMultipleDGemmMultipleD<A0Layout,
+                                                             B0Layout,
+                                                             D0sLayout,
+                                                             B1Layout,
+                                                             D1sLayout,
+                                                             E1Layout,
+                                                             A0DataType,
+                                                             B0DataType,
+                                                             D0sDataType,
+                                                             B1DataType,
+                                                             D1sDataType,
+                                                             E1DataType,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             CDE0ElementOp,
+                                                             PassThrough,
+                                                             CDE1ElementOp>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<A0DataType, half_t> && is_same_v<B0DataType, half_t> &&
+                     is_same_v<B1DataType, half_t> && is_same_v<E1DataType, half_t>)
+        {
+            if constexpr(is_same_v<A0Layout, Row> && is_same_v<B0Layout, Col> &&
+                         is_same_v<B1Layout, Row> && is_same_v<E1Layout, Row>)
+            {
+                add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<A0Layout, Row> && is_same_v<B0Layout, Col> &&
+                              is_same_v<B1Layout, Col> && is_same_v<E1Layout, Row>)
+            {
+                add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance(
+                    op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp
new file mode 100644
index 0000000000..b471268765
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp
@@ -0,0 +1,94 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#ifdef CK_ENABLE_FP16
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_batched_contraction_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_mnnm_instance(
+    std::vector<std::unique_ptr<
+        DeviceBatchedContractionMultipleD<1,
+                                          2,
+                                          3,
+                                          1,
+                                          F16,
+                                          F16,
+                                          F16_Tuple,
+                                          F16,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::Add>>>& instances);
+
+// Contraction + add
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedContractionMultipleD<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        ck::Tuple<DDataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Add>>
+{
+    using DeviceOp =
+        DeviceBatchedContractionMultipleD<NumDimG,
+                                          NumDimM,
+                                          NumDimN,
+                                          NumDimK,
+                                          ADataType,
+                                          BDataType,
+                                          ck::Tuple<DDataType>,
+                                          EDataType,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::Add>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, ck::half_t> && is_same_v<BDataType, ck::half_t> &&
+                     is_same_v<DDataType, ck::half_t> && is_same_v<EDataType, ck::half_t>)
+        {
+            if constexpr(NumDimG == 1 && NumDimM == 2 && NumDimN == 3 && NumDimK == 1)
+            {
+                add_device_batched_contraction_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_mnnm_instance(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp
new file mode 100644
index 0000000000..161085e4d0
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp
@@ -0,0 +1,194 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef CK_ENABLE_FP16
+void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            ck::Tuple<F16>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            ScaleAdd,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances);
+
+void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            ck::Tuple<F16>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            ScaleAdd,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances);
+#endif
+#ifdef CK_ENABLE_BF16
+void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            ck::Tuple<BF16>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            ScaleAdd,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances);
+
+void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            ck::Tuple<BF16>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            ScaleAdd,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances);
+#endif
+template <typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasDataType,
+          MaskingSpecialization MaskingSpec>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      ADataType,
+                                                                      B0DataType,
+                                                                      B1DataType,
+                                                                      CDataType,
+                                                                      Acc0BiasDataType,
+                                                                      ck::Tuple<>,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      ScaleAdd,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      MaskingSpec>>
+{
+    using DeviceOp = DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                         1,
+                                                         1,
+                                                         1,
+                                                         1,
+                                                         ADataType,
+                                                         B0DataType,
+                                                         B1DataType,
+                                                         CDataType,
+                                                         Acc0BiasDataType,
+                                                         ck::Tuple<>,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         ScaleAdd,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         MaskingSpec>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+#ifdef CK_ENABLE_FP16
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
+                     is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t> &&
+                     Acc0BiasDataType::Size() == 1 &&
+                     is_same_v<tuple_element_t<0, Acc0BiasDataType>, half_t>)
+        {
+            if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+            {
+                add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+            else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
+            {
+                add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+        }
+#endif
+#ifdef CK_ENABLE_BF16
+        else if constexpr(is_same_v<ADataType, BF16> && is_same_v<B0DataType, BF16> &&
+                          is_same_v<B1DataType, BF16> && is_same_v<CDataType, BF16> &&
+                          Acc0BiasDataType::Size() == 1 &&
+                          is_same_v<tuple_element_t<0, Acc0BiasDataType>, BF16>)
+        {
+            if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+            {
+                add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+            else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
+            {
+                add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+        }
+#endif
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp
new file mode 100644
index 0000000000..77ad36b97b
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp
@@ -0,0 +1,114 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef CK_ENABLE_FP16
+void add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmGemm<Row,
+                                                      Col,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+
+void add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmGemm<Row,
+                                                      Col,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances);
+template <typename ALayout,
+          typename B0Layout,
+          typename B1Layout,
+          typename CLayout,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedGemmGemm<ALayout,
+                                                        B0Layout,
+                                                        B1Layout,
+                                                        CLayout,
+                                                        ADataType,
+                                                        B0DataType,
+                                                        B1DataType,
+                                                        CDataType,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>
+{
+    using DeviceOp = DeviceBatchedGemmGemm<ALayout,
+                                           B0Layout,
+                                           B1Layout,
+                                           CLayout,
+                                           ADataType,
+                                           B0DataType,
+                                           B1DataType,
+                                           CDataType,
+                                           PassThrough,
+                                           PassThrough,
+                                           PassThrough,
+                                           PassThrough,
+                                           PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
+                     is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<B0Layout, Col> &&
+                         is_same_v<B1Layout, Row> && is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<B0Layout, Col> &&
+                              is_same_v<B1Layout, Col> && is_same_v<CLayout, Row>)
+            {
+                add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance(
+                    op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp
new file mode 100644
index 0000000000..c9535e828b
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp
@@ -0,0 +1,342 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+#ifdef DL_KERNELS
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef CK_ENABLE_FP16
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+#endif
+#ifdef CK_ENABLE_INT8
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances);
+#endif
+template <typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatchedGemmMultiD<
+    ALayout,
+    BLayout,
+    Empty_Tuple,
+    ELayout,
+    ADataType,
+    BDataType,
+    Empty_Tuple,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceBatchedGemmMultiD<ALayout,
+                                             BLayout,
+                                             Empty_Tuple,
+                                             ELayout,
+                                             ADataType,
+                                             BDataType,
+                                             Empty_Tuple,
+                                             EDataType,
+                                             ck::tensor_operation::element_wise::PassThrough,
+                                             ck::tensor_operation::element_wise::PassThrough,
+                                             ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+#ifdef CK_ENABLE_FP16
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instances(op_ptrs);
+                add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instances(op_ptrs);
+                add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances(op_ptrs);
+                add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instances(op_ptrs);
+                add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instances(
+                    op_ptrs);
+            }
+        }
+#endif
+#ifdef CK_ENABLE_INT8
+        else if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
+                          is_same_v<EDataType, int8_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instances(op_ptrs);
+                add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instances(op_ptrs);
+                add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances(op_ptrs);
+                add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instances(op_ptrs);
+                add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instances(
+                    op_ptrs);
+            }
+        }
+#endif
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp
new file mode 100644
index 0000000000..88bdaea23a
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#ifdef CK_ENABLE_FP16
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmSoftmaxGemm<Row,
+                                                             Col,
+                                                             Row,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             Scale,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             false>>>& instances);
+
+void add_device_batched_gemm_masking_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmSoftmaxGemm<Row,
+                                                             Col,
+                                                             Row,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             Scale,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             true>>>& instances);
+
+template <typename ALayout,
+          typename B0Layout,
+          typename B1Layout,
+          typename CLayout,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          bool MaskOutUpperTriangle>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm<ALayout,
+                                                               B0Layout,
+                                                               B1Layout,
+                                                               CLayout,
+                                                               ADataType,
+                                                               B0DataType,
+                                                               B1DataType,
+                                                               CDataType,
+                                                               PassThrough,
+                                                               PassThrough,
+                                                               Scale,
+                                                               PassThrough,
+                                                               PassThrough,
+                                                               MaskOutUpperTriangle>>
+{
+    using DeviceOp = DeviceBatchedGemmSoftmaxGemm<ALayout,
+                                                  B0Layout,
+                                                  B1Layout,
+                                                  CLayout,
+                                                  ADataType,
+                                                  B0DataType,
+                                                  B1DataType,
+                                                  CDataType,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  Scale,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  MaskOutUpperTriangle>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
+                     is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<B0Layout, Col> &&
+                         is_same_v<B1Layout, Row> && is_same_v<CLayout, Row>)
+            {
+                if constexpr(MaskOutUpperTriangle)
+                {
+                    add_device_batched_gemm_masking_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+                        op_ptrs);
+                }
+                else
+                {
+                    add_device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+                        op_ptrs);
+                }
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp
new file mode 100644
index 0000000000..1814cfabb4
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp
@@ -0,0 +1,190 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef CK_ENABLE_FP16
+void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            ck::Tuple<>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            Scale,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances);
+
+void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            ck::Tuple<>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            Scale,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances);
+#endif
+#ifdef CK_ENABLE_BF16
+void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            ck::Tuple<>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            Scale,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances);
+
+void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            ck::Tuple<>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            Scale,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances);
+#endif
+
+template <typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          MaskingSpecialization MaskingSpec>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      ADataType,
+                                                                      B0DataType,
+                                                                      B1DataType,
+                                                                      CDataType,
+                                                                      ck::Tuple<>,
+                                                                      ck::Tuple<>,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      Scale,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      MaskingSpec>>
+{
+    using DeviceOp = DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                         1,
+                                                         1,
+                                                         1,
+                                                         1,
+                                                         ADataType,
+                                                         B0DataType,
+                                                         B1DataType,
+                                                         CDataType,
+                                                         ck::Tuple<>,
+                                                         ck::Tuple<>,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         Scale,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         MaskingSpec>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+#ifdef CK_ENABLE_FP16
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
+                     is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t>)
+        {
+            if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+            {
+                add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+            else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
+            {
+                add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+        }
+#endif
+#ifdef CK_ENABLE_BF16
+        else if constexpr(is_same_v<ADataType, BF16> && is_same_v<B0DataType, BF16> &&
+                          is_same_v<B1DataType, BF16> && is_same_v<CDataType, BF16>)
+        {
+            if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+            {
+                add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+            else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
+            {
+                add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+        }
+#endif
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp
new file mode 100644
index 0000000000..0e1f6f04e8
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp
@@ -0,0 +1,124 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// FP16
+void add_device_batchnorm_backward_rank_4_3_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchNormBwd<F16, F32, F32, F32, F16, F32, F32, PassThrough, 4, 3>>>&);
+
+// FP32
+void add_device_batchnorm_backward_rank_4_3_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchNormBwd<F32, F32, F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&);
+
+// BF16
+void add_device_batchnorm_backward_rank_4_3_bf16_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchNormBwd<BF16, F32, F32, F32, BF16, F32, F32, PassThrough, 4, 3>>>&);
+
+// FP64
+void add_device_batchnorm_backward_rank_4_3_f64_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchNormBwd<F64, F64, F64, F64, F64, F64, F64, PassThrough, 4, 3>>>&);
+
+template <typename XDataType,
+          typename DxDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          typename DyElementwiseOp,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchNormBwd<XDataType,
+                                                     DxDataType,
+                                                     DyDataType,
+                                                     AccDataType,
+                                                     ScaleDataType,
+                                                     DscaleDbiasDataType,
+                                                     MeanVarDataType,
+                                                     DyElementwiseOp,
+                                                     Rank,
+                                                     NumReduceDim>>
+{
+    using DeviceOp = DeviceBatchNormBwd<XDataType,
+                                        DxDataType,
+                                        DyDataType,
+                                        AccDataType,
+                                        ScaleDataType,
+                                        DscaleDbiasDataType,
+                                        MeanVarDataType,
+                                        DyElementwiseOp,
+                                        Rank,
+                                        NumReduceDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<XDataType, F16> && is_same_v<DxDataType, F32> &&
+                     is_same_v<DyDataType, F32> && is_same_v<AccDataType, F32> &&
+                     is_same_v<ScaleDataType, F16> && is_same_v<DscaleDbiasDataType, F32> &&
+                     is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_backward_rank_4_3_f16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F32> && is_same_v<DxDataType, F32> &&
+                          is_same_v<DyDataType, F32> && is_same_v<AccDataType, F32> &&
+                          is_same_v<ScaleDataType, F32> && is_same_v<DscaleDbiasDataType, F32> &&
+                          is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_backward_rank_4_3_f32_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, BF16> && is_same_v<DxDataType, F32> &&
+                          is_same_v<DyDataType, F32> && is_same_v<AccDataType, F32> &&
+                          is_same_v<ScaleDataType, BF16> && is_same_v<DscaleDbiasDataType, F32> &&
+                          is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_backward_rank_4_3_bf16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F64> && is_same_v<DxDataType, F64> &&
+                          is_same_v<DyDataType, F64> && is_same_v<AccDataType, F64> &&
+                          is_same_v<ScaleDataType, F64> && is_same_v<DscaleDbiasDataType, F64> &&
+                          is_same_v<MeanVarDataType, F64>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<DyElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_backward_rank_4_3_f64_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp
new file mode 100644
index 0000000000..8fd1c7665d
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp
@@ -0,0 +1,117 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// FP16
+void add_device_batchnorm_forward_rank_4_3_f16_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<F16, F16, F32, F16, F16, F32, PassThrough, 4, 3>>>&);
+
+// FP32
+void add_device_batchnorm_forward_rank_4_3_f32_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<F32, F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&);
+
+// BF16
+void add_device_batchnorm_forward_rank_4_3_bf16_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<BF16, BF16, F32, BF16, BF16, F32, PassThrough, 4, 3>>>&);
+
+// FP64
+void add_device_batchnorm_forward_rank_4_3_f64_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<F64, F64, F64, F64, F64, F64, PassThrough, 4, 3>>>&);
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          typename YElementwiseOp,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchNormFwd<XDataType,
+                                                     YDataType,
+                                                     AccDataType,
+                                                     ScaleDataType,
+                                                     BiasDataType,
+                                                     MeanVarDataType,
+                                                     YElementwiseOp,
+                                                     Rank,
+                                                     NumReduceDim>>
+{
+    using DeviceOp = DeviceBatchNormFwd<XDataType,
+                                        YDataType,
+                                        AccDataType,
+                                        ScaleDataType,
+                                        BiasDataType,
+                                        MeanVarDataType,
+                                        YElementwiseOp,
+                                        Rank,
+                                        NumReduceDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<XDataType, F16> && is_same_v<YDataType, F16> &&
+                     is_same_v<AccDataType, F32> && is_same_v<ScaleDataType, F16> &&
+                     is_same_v<BiasDataType, F16> && is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_forward_rank_4_3_f16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F32> && is_same_v<YDataType, F32> &&
+                          is_same_v<AccDataType, F32> && is_same_v<ScaleDataType, F32> &&
+                          is_same_v<BiasDataType, F32> && is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_forward_rank_4_3_f32_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, BF16> && is_same_v<YDataType, BF16> &&
+                          is_same_v<AccDataType, F32> && is_same_v<ScaleDataType, BF16> &&
+                          is_same_v<BiasDataType, BF16> && is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_forward_rank_4_3_bf16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F64> && is_same_v<YDataType, F64> &&
+                          is_same_v<AccDataType, F64> && is_same_v<ScaleDataType, F64> &&
+                          is_same_v<BiasDataType, F64> && is_same_v<MeanVarDataType, F64>)
+        {
+            if constexpr(Rank == 4 && NumReduceDim == 3 && is_same_v<YElementwiseOp, PassThrough>)
+            {
+                add_device_batchnorm_forward_rank_4_3_f64_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp
new file mode 100644
index 0000000000..f6f4df7e2e
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp
@@ -0,0 +1,117 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// FP16
+void add_device_batchnorm_infer_rank_4_f16_instances(
+    std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<F16, F32, F32, F16, F16>,
+        ck::Tuple<F16>,
+        ck::tensor_operation::element_wise::NormalizeInInfer,
+        4>>>&);
+
+// FP32
+void add_device_batchnorm_infer_rank_4_f32_instances(
+    std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<F32, F32, F32, F32, F32>,
+        ck::Tuple<F32>,
+        ck::tensor_operation::element_wise::NormalizeInInfer,
+        4>>>&);
+
+// BF16
+void add_device_batchnorm_infer_rank_4_bf16_instances(
+    std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<BF16, F32, F32, BF16, BF16>,
+        ck::Tuple<BF16>,
+        ck::tensor_operation::element_wise::NormalizeInInfer,
+        4>>>&);
+
+// FP64
+void add_device_batchnorm_infer_rank_4_f64_instances(
+    std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<F64, F64, F64, F64, F64>,
+        ck::Tuple<F64>,
+        ck::tensor_operation::element_wise::NormalizeInInfer,
+        4>>>&);
+
+template <typename XDataType,
+          typename YDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          index_t Rank>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceElementwise<
+    ck::Tuple<XDataType, MeanVarDataType, MeanVarDataType, ScaleDataType, BiasDataType>,
+    ck::Tuple<YDataType>,
+    ck::tensor_operation::element_wise::NormalizeInInfer,
+    Rank>>
+{
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<XDataType, MeanVarDataType, MeanVarDataType, ScaleDataType, BiasDataType>,
+        ck::Tuple<YDataType>,
+        ck::tensor_operation::element_wise::NormalizeInInfer,
+        Rank>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<XDataType, F16> && is_same_v<YDataType, F16> &&
+                     is_same_v<ScaleDataType, F16> && is_same_v<BiasDataType, F16> &&
+                     is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4)
+            {
+                add_device_batchnorm_infer_rank_4_f16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F32> && is_same_v<YDataType, F32> &&
+                          is_same_v<ScaleDataType, F32> && is_same_v<BiasDataType, F32> &&
+                          is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4)
+            {
+                add_device_batchnorm_infer_rank_4_f32_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, BF16> && is_same_v<YDataType, BF16> &&
+                          is_same_v<ScaleDataType, BF16> && is_same_v<BiasDataType, BF16> &&
+                          is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4)
+            {
+                add_device_batchnorm_infer_rank_4_bf16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F64> && is_same_v<YDataType, F64> &&
+                          is_same_v<ScaleDataType, F64> && is_same_v<BiasDataType, F64> &&
+                          is_same_v<MeanVarDataType, F64>)
+        {
+            if constexpr(Rank == 4)
+            {
+                add_device_batchnorm_infer_rank_4_f64_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp b/library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp
new file mode 100644
index 0000000000..bd3af891ec
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp
@@ -0,0 +1,194 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef CK_ENABLE_FP32
+// float
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances);
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances);
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances);
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances);
+#endif
+#ifdef CK_ENABLE_FP64
+// double
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           F64_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances);
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           F64_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances);
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           F64_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances);
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           F64_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances);
+#endif
+// Contraction + Bilinear
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContractionMultipleD<
+    NumDimM,
+    NumDimN,
+    NumDimK,
+    ADataType,
+    BDataType,
+    ck::Tuple<DDataType>,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::Bilinear>>
+{
+    using DeviceOp = DeviceContractionMultipleD<NumDimM,
+                                                NumDimN,
+                                                NumDimK,
+                                                ADataType,
+                                                BDataType,
+                                                ck::Tuple<DDataType>,
+                                                EDataType,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::Bilinear>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+#ifdef CK_ENABLE_FP32
+        if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
+                     is_same_v<DDataType, float> && is_same_v<EDataType, float>)
+        {
+            if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
+            {
+                add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance(
+                    op_ptrs);
+                add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance(
+                    op_ptrs);
+                add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance(
+                    op_ptrs);
+                add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance(
+                    op_ptrs);
+            }
+        }
+#endif
+#ifdef CK_ENABLE_FP64
+        if constexpr(is_same_v<ADataType, double> && is_same_v<BDataType, double> &&
+                     is_same_v<DDataType, double> && is_same_v<EDataType, double>)
+        {
+            if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
+            {
+                add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance(
+                    op_ptrs);
+                add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance(
+                    op_ptrs);
+                add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance(
+                    op_ptrs);
+                add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance(
+                    op_ptrs);
+            }
+        }
+#endif
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp b/library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp
new file mode 100644
index 0000000000..4dd735d062
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp
@@ -0,0 +1,193 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef CK_ENABLE_FP32
+// float
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances);
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances);
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances);
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances);
+#endif
+#ifdef CK_ENABLE_FP64
+// double
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           Empty_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances);
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           Empty_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances);
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           Empty_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances);
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           Empty_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances);
+#endif
+// Contraction + Scale
+template <index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceContractionMultipleD<
+    NumDimM,
+    NumDimN,
+    NumDimK,
+    ADataType,
+    BDataType,
+    ck::Tuple<>,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::Scale>>
+{
+    using DeviceOp = DeviceContractionMultipleD<NumDimM,
+                                                NumDimN,
+                                                NumDimK,
+                                                ADataType,
+                                                BDataType,
+                                                ck::Tuple<>,
+                                                EDataType,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::Scale>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+#ifdef CK_ENABLE_FP32
+        if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
+                     is_same_v<EDataType, float>)
+        {
+            if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
+            {
+                add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
+                    op_ptrs);
+                add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
+                    op_ptrs);
+                add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
+                    op_ptrs);
+                add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance(
+                    op_ptrs);
+            }
+        }
+#endif
+#ifdef CK_ENABLE_FP64
+        if constexpr(is_same_v<ADataType, double> && is_same_v<BDataType, double> &&
+                     is_same_v<EDataType, double>)
+        {
+            if constexpr(NumDimM == 2 && NumDimN == 2 && NumDimK == 2)
+            {
+                add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance(
+                    op_ptrs);
+                add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance(
+                    op_ptrs);
+                add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance(
+                    op_ptrs);
+                add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance(
+                    op_ptrs);
+            }
+        }
+#endif
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp b/library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp
new file mode 100644
index 0000000000..571ff0b672
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp
@@ -0,0 +1,347 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef CK_ENABLE_BF16
+// conv1d backward data
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<1,
+                                                  NWC,
+                                                  KXC,
+                                                  NWK,
+                                                  BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+#endif
+#ifdef CK_ENABLE_FP16
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvBwdData<1, NWC, KXC, NWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
+#ifdef CK_ENABLE_FP32
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvBwdData<1, NWC, KXC, NWK, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
+#ifdef CK_ENABLE_INT8
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<1,
+                                                  NWC,
+                                                  KXC,
+                                                  NWK,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+#endif
+#ifdef CK_ENABLE_BF16
+// conv2d backward data
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+#endif
+#ifdef CK_ENABLE_FP16
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  F16,
+                                                  F16,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+#endif
+#ifdef CK_ENABLE_FP32
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  F32,
+                                                  F32,
+                                                  F32,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+#endif
+#ifdef CK_ENABLE_INT8
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+#endif
+#ifdef DL_KERNELS
+#ifdef CK_ENABLE_FP16
+// conv2d dl
+void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  F16,
+                                                  F16,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+#endif
+#ifdef CK_ENABLE_FP32
+void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  F32,
+                                                  F32,
+                                                  F32,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+#endif
+#ifdef CK_ENABLE_INT8
+void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+#endif
+#endif
+#ifdef CK_ENABLE_BF16
+// conv3d backward data
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+#endif
+#ifdef CK_ENABLE_FP16
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  F16,
+                                                  F16,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+#endif
+#ifdef CK_ENABLE_FP32
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  F32,
+                                                  F32,
+                                                  F32,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+#endif
+#ifdef CK_ENABLE_INT8
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+#endif
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceConvBwdData<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceConvBwdData<NumDimSpatial,
+                                       InLayout,
+                                       WeiLayout,
+                                       OutLayout,
+                                       InDataType,
+                                       WeiDataType,
+                                       OutDataType,
+                                       ck::tensor_operation::element_wise::PassThrough,
+                                       ck::tensor_operation::element_wise::PassThrough,
+                                       ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, NWC> && is_same_v<WeiLayout, KXC> &&
+                     is_same_v<OutLayout, NWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(op_ptrs);
+            }
+#ifdef CK_ENABLE_FP16
+            if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                         is_same_v<OutDataType, half_t>)
+            {
+                add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(op_ptrs);
+            }
+#endif
+#ifdef CK_ENABLE_BF16
+            if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                         is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(op_ptrs);
+            }
+#endif
+#ifdef CK_ENABLE_INT8
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<OutDataType, int8_t>)
+            {
+                add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(op_ptrs);
+            }
+#endif
+        }
+        else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
+                          is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
+#ifdef DL_KERNELS
+                add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
+#endif
+            }
+#ifdef CK_ENABLE_FP16
+            if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                         is_same_v<OutDataType, half_t>)
+            {
+                add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
+#ifdef DL_KERNELS
+                add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
+#endif
+            }
+#endif
+#ifdef CK_ENABLE_BF16
+            if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                         is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(op_ptrs);
+            }
+#endif
+#ifdef CK_ENABLE_INT8
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<OutDataType, int8_t>)
+            {
+                add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
+#ifdef DL_KERNELS
+                add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
+#endif
+            }
+#endif
+        }
+        else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWC> &&
+                          is_same_v<WeiLayout, KZYXC> && is_same_v<OutLayout, NDHWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(op_ptrs);
+            }
+#ifdef CK_ENABLE_FP16
+            if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                         is_same_v<OutDataType, half_t>)
+            {
+                add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(op_ptrs);
+            }
+#endif
+#ifdef CK_ENABLE_BF16
+            if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                         is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(op_ptrs);
+            }
+#endif
+#ifdef CK_ENABLE_INT8
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<OutDataType, int8_t>)
+            {
+                add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(op_ptrs);
+            }
+#endif
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp b/library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp
new file mode 100644
index 0000000000..ad2da3364f
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp
@@ -0,0 +1,138 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// conv2d forward
+#ifdef CK_ENABLE_FP16
+void add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvFwd<2, NHWC, KYXC, NHWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvFwd<2, NHWC, KYXC, NHWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
+#ifdef CK_ENABLE_BF16
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvFwd<2,
+                                              NHWC,
+                                              KYXC,
+                                              NHWK,
+                                              BF16,
+                                              BF16,
+                                              BF16,
+                                              PassThrough,
+                                              PassThrough,
+                                              PassThrough>>>& instances);
+#endif
+#ifdef CK_ENABLE_FP32
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvFwd<2, NHWC, KYXC, NHWK, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
+#ifdef CK_ENABLE_INT8
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvFwd<2,
+                                              NHWC,
+                                              KYXC,
+                                              NHWK,
+                                              int8_t,
+                                              int8_t,
+                                              int8_t,
+                                              PassThrough,
+                                              PassThrough,
+                                              PassThrough>>>& instances);
+#endif
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceConvFwd<NumDimSpatial,
+                                                InLayout,
+                                                WeiLayout,
+                                                OutLayout,
+                                                InDataType,
+                                                WeiDataType,
+                                                OutDataType,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceConvFwd<NumDimSpatial,
+                                   InLayout,
+                                   WeiLayout,
+                                   OutLayout,
+                                   InDataType,
+                                   WeiDataType,
+                                   OutDataType,
+                                   ck::tensor_operation::element_wise::PassThrough,
+                                   ck::tensor_operation::element_wise::PassThrough,
+                                   ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWC> &&
+                     is_same_v<WeiLayout, KYXC> && is_same_v<OutLayout, NHWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(op_ptrs);
+            }
+#ifdef CK_ENABLE_FP16
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
+                add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
+            }
+#endif
+#ifdef CK_ENABLE_BF16
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(op_ptrs);
+            }
+#endif
+#ifdef CK_ENABLE_INT8
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                              is_same_v<OutDataType, int8_t>)
+            {
+                add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(op_ptrs);
+            }
+#endif
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp
new file mode 100644
index 0000000000..b03693b00a
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp
@@ -0,0 +1,63 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Normalize                            = ck::tensor_operation::element_wise::Normalize;
+using DeviceNormalizeFromMeanMeanSquarePtr = ck::tensor_operation::device::
+    DeviceElementwisePtr<Tuple<half_t, float, float, half_t, half_t>, Tuple<half_t>, Normalize, 2>;
+
+void add_device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances(
+    std::vector<DeviceNormalizeFromMeanMeanSquarePtr>& instances);
+
+template <typename InputType,
+          typename MeanType,
+          typename MeanSquareType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename OutputType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceElementwise<
+    ck::Tuple<InputType, MeanType, MeanSquareType, GammaDataType, BetaDataType>,
+    ck::Tuple<OutputType>,
+    Normalize,
+    2>>
+{
+    using DeviceOp = DeviceElementwise<
+        ck::Tuple<InputType, MeanType, MeanSquareType, GammaDataType, BetaDataType>,
+        ck::Tuple<OutputType>,
+        Normalize,
+        2>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same<InputType, half_t>::value && is_same<MeanType, float>::value &&
+                     is_same<MeanSquareType, float>::value &&
+                     is_same<GammaDataType, half_t>::value &&
+                     is_same<BetaDataType, half_t>::value && is_same<OutputType, half_t>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances(op_ptrs);
+        }
+
+        return op_ptrs;
+    };
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp
new file mode 100644
index 0000000000..b151395109
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using DeviceGemmAddAddMeanSquareMeanPtr = ck::tensor_operation::device::DeviceGemmReducePtr<1, 2>;
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+    std::vector<DeviceGemmAddAddMeanSquareMeanPtr>&);
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+    std::vector<DeviceGemmAddAddMeanSquareMeanPtr>&);
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
+    std::vector<DeviceGemmAddAddMeanSquareMeanPtr>&);
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
+    std::vector<DeviceGemmAddAddMeanSquareMeanPtr>&);
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+auto get_device_gemm_add_add_mean_squaremean_instances()
+{
+    std::vector<DeviceGemmAddAddMeanSquareMeanPtr> op_ptrs;
+
+    if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
+                 is_same<CDataType, half_t>::value)
+    {
+        if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+                    op_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+                    op_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
+                    op_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
+                    op_ptrs);
+        }
+    }
+
+    return op_ptrs;
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp
new file mode 100644
index 0000000000..931110267a
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#ifdef CK_ENABLE_FP16
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// FP16
+void add_device_elementwise_normalization_rank_2_1_f16_instances(
+    std::vector<std::unique_ptr<DeviceElementwiseNormalization<ck::Tuple<F16, F16>,
+                                                               F16,
+                                                               F16,
+                                                               F32,
+                                                               F16,
+                                                               element_wise::Add,
+                                                               PassThrough,
+                                                               2,
+                                                               1>>>&);
+
+template <typename InDataTypeTuple,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceElementwiseNormalization<
+    InDataTypeTuple,
+    GammaDataType,
+    BetaDataType,
+    F32,
+    YDataType,
+    ck::tensor_operation::element_wise::Add,
+    ck::tensor_operation::element_wise::PassThrough,
+    Rank,
+    NumReduceDim>>
+{
+    using DeviceOp = DeviceElementwiseNormalization<InDataTypeTuple,
+                                                    GammaDataType,
+                                                    BetaDataType,
+                                                    F32,
+                                                    YDataType,
+                                                    ck::tensor_operation::element_wise::Add,
+                                                    ck::tensor_operation::element_wise::PassThrough,
+                                                    Rank,
+                                                    NumReduceDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<GammaDataType, F16> && is_same_v<BetaDataType, F16> &&
+                     is_same_v<YDataType, F16>)
+        {
+            if constexpr(Rank == 2 && NumReduceDim == 1)
+            {
+                add_device_elementwise_normalization_rank_2_1_f16_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
new file mode 100644
index 0000000000..9310fd433b
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
@@ -0,0 +1,492 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+#include <vector>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+#if defined(CK_ENABLE_FP16) && defined(DL_KERNELS)
+void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
+#if defined(CK_ENABLE_FP32) && defined(DL_KERNELS)
+void add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
+#if defined(CK_ENABLE_INT8) && defined(DL_KERNELS)
+void add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
+#ifdef CK_ENABLE_INT8
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
+#ifdef CK_ENABLE_FP16
+void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+#endif
+#ifdef CK_ENABLE_BF16
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
+#ifdef CK_ENABLE_FP32
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
+#ifdef CK_ENABLE_FP64
+void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+
+        DeviceGemm<Col, Row, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f64_f64_f64_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f64_f64_f64_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+#endif
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGemm<ALayout,
+                                             BLayout,
+                                             CLayout,
+                                             ADataType,
+                                             BDataType,
+                                             CDataType,
+                                             ck::tensor_operation::element_wise::PassThrough,
+                                             ck::tensor_operation::element_wise::PassThrough,
+                                             ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGemm<ALayout,
+                                BLayout,
+                                CLayout,
+                                ADataType,
+                                BDataType,
+                                CDataType,
+                                ck::tensor_operation::element_wise::PassThrough,
+                                ck::tensor_operation::element_wise::PassThrough,
+                                ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
+                     is_same_v<CDataType, float>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
+                add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
+#endif
+                add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
+                add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
+#endif
+                add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
+                add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(op_ptrs);
+#endif
+                add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
+                add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(op_ptrs);
+#endif
+                add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(op_ptrs);
+            }
+        }
+#ifdef CK_ENABLE_FP16
+        else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                          is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
+                add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+                add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances(op_ptrs);
+                add_device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+#endif
+                add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
+                add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+                add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(op_ptrs);
+                add_device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+#endif
+                add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+                add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
+                add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+                add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances(op_ptrs);
+                add_device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+#endif
+                add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
+                add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+                add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances(op_ptrs);
+                add_device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+#endif
+                add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+            }
+        }
+#endif
+#ifdef CK_ENABLE_BF16
+        else if constexpr(is_same_v<ADataType, ck::bhalf_t> && is_same_v<BDataType, ck::bhalf_t> &&
+                          is_same_v<CDataType, ck::bhalf_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(op_ptrs);
+            }
+        }
+#endif
+#ifdef CK_ENABLE_INT8
+        else if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
+                          is_same_v<CDataType, int8_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
+                add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(op_ptrs);
+                add_device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instances(op_ptrs);
+#endif
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
+                add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(op_ptrs);
+                add_device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instances(op_ptrs);
+#endif
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
+                add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(op_ptrs);
+                add_device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instances(op_ptrs);
+#endif
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(op_ptrs);
+#ifdef DL_KERNELS
+                add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(op_ptrs);
+                add_device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instances(op_ptrs);
+#endif
+            }
+        }
+#endif
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp
new file mode 100644
index 0000000000..99b2ad1315
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp
@@ -0,0 +1,152 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>&);
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>&);
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>&);
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>&);
+
+// GEMM + Add + Add + FastGelu
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<
+    ALayout,
+    BLayout,
+    ck::Tuple<D0Layout, D1Layout>,
+    ELayout,
+    ADataType,
+    BDataType,
+    ck::Tuple<D0DataType, D1DataType>,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::AddAddFastGelu>>
+{
+    using DeviceOp = DeviceGemmMultipleD<ALayout,
+                                         BLayout,
+                                         ck::Tuple<D0Layout, D1Layout>,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         ck::Tuple<D0DataType, D1DataType>,
+                                         EDataType,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::AddAddFastGelu>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<D0DataType, half_t> && is_same_v<D1DataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp
new file mode 100644
index 0000000000..fd3550c2f0
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp
@@ -0,0 +1,145 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>&);
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>&);
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>&);
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>&);
+
+// GEMM + Add + FastGelu
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
+                                                      BLayout,
+                                                      ck::Tuple<D0Layout>,
+                                                      ELayout,
+                                                      ADataType,
+                                                      BDataType,
+                                                      ck::Tuple<D0DataType>,
+                                                      EDataType,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      AddFastGelu>>
+{
+    using DeviceOp = DeviceGemmMultipleD<ALayout,
+                                         BLayout,
+                                         ck::Tuple<D0Layout>,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         ck::Tuple<D0DataType>,
+                                         EDataType,
+                                         PassThrough,
+                                         PassThrough,
+                                         AddFastGelu>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<D0DataType, half_t> && is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp
new file mode 100644
index 0000000000..481915d00b
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp
@@ -0,0 +1,155 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddMultiply>>>&);
+
+void add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddMultiply>>>&);
+
+void add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddMultiply>>>&);
+
+void add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddMultiply>>>&);
+
+// GEMM + Add + Multiply
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<
+    ALayout,
+    BLayout,
+    ck::Tuple<D0Layout, D1Layout>,
+    ELayout,
+    ADataType,
+    BDataType,
+    ck::Tuple<D0DataType, D1DataType>,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::AddMultiply>>
+{
+    using DeviceOp = DeviceGemmMultipleD<ALayout,
+                                         BLayout,
+                                         ck::Tuple<D0Layout, D1Layout>,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         ck::Tuple<D0DataType, D1DataType>,
+                                         EDataType,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::AddMultiply>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<D0DataType, half_t> && is_same_v<D1DataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp
new file mode 100644
index 0000000000..dd8ecae62c
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp
@@ -0,0 +1,173 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#ifdef CK_ENABLE_FP16
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Row,
+                                                             Row,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Row,
+                                                             Col,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Col,
+                                                             Row,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Col,
+                                                             Col,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+// GEMM + Add + Relu + Add + Layernorm
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename HLayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename HDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleDLayernorm<
+    ALayout,
+    BLayout,
+    ck::Tuple<D0Layout, D1Layout>,
+    HLayout,
+    ADataType,
+    BDataType,
+    ck::Tuple<D0DataType, D1DataType>,
+    GammaDataType,
+    BetaDataType,
+    HDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::AddReluAdd,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGemmMultipleDLayernorm<ALayout,
+                                                  BLayout,
+                                                  ck::Tuple<D0Layout, D1Layout>,
+                                                  HLayout,
+                                                  ADataType,
+                                                  BDataType,
+                                                  ck::Tuple<D0DataType, D1DataType>,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  HDataType,
+                                                  ck::tensor_operation::element_wise::PassThrough,
+                                                  ck::tensor_operation::element_wise::PassThrough,
+                                                  ck::tensor_operation::element_wise::AddReluAdd,
+                                                  ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<D0DataType, half_t> && is_same_v<D1DataType, half_t> &&
+                     is_same_v<GammaDataType, half_t> && is_same_v<BetaDataType, half_t> &&
+                     is_same_v<HDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                         is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp
new file mode 100644
index 0000000000..9d8eaada22
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp
@@ -0,0 +1,147 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#ifdef CK_ENABLE_FP16
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances);
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances);
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances);
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances);
+
+// GEMM + Bilinear
+template <typename ALayout,
+          typename BLayout,
+          typename DLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<
+    ALayout,
+    BLayout,
+    ck::Tuple<DLayout>,
+    ELayout,
+    ADataType,
+    BDataType,
+    ck::Tuple<DDataType>,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::Bilinear>>
+{
+    using DeviceOp = DeviceGemmMultipleD<ALayout,
+                                         BLayout,
+                                         ck::Tuple<DLayout>,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         ck::Tuple<DDataType>,
+                                         EDataType,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::Bilinear>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<DDataType, half_t> && is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<DLayout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<DLayout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<DLayout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<DLayout, Row> && is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp
new file mode 100644
index 0000000000..f6612fc259
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp
@@ -0,0 +1,139 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#ifdef CK_ENABLE_FP16
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>&);
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>&);
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>&);
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>&);
+
+// GEMM + FastGelu
+template <typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
+                                                                                        BLayout,
+                                                                                        Empty_Tuple,
+                                                                                        ELayout,
+                                                                                        ADataType,
+                                                                                        BDataType,
+                                                                                        Empty_Tuple,
+                                                                                        EDataType,
+                                                                                        PassThrough,
+                                                                                        PassThrough,
+                                                                                        FastGelu>>
+{
+    using DeviceOp = DeviceGemmMultipleD<ALayout,
+                                         BLayout,
+                                         Empty_Tuple,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         Empty_Tuple,
+                                         EDataType,
+                                         PassThrough,
+                                         PassThrough,
+                                         FastGelu>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp
new file mode 100644
index 0000000000..badc06dd6f
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp
@@ -0,0 +1,161 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>&);
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>&);
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F8,
+                                                    F32_F32_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>&);
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F8,
+                                                    F32_F32_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>&);
+
+// GEMM + Multiply + Add
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<
+    ALayout,
+    BLayout,
+    ck::Tuple<D0Layout, D1Layout>,
+    ELayout,
+    ADataType,
+    BDataType,
+    ck::Tuple<D0DataType, D1DataType>,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::MultiplyAdd>>
+{
+    using DeviceOp = DeviceGemmMultipleD<ALayout,
+                                         BLayout,
+                                         ck::Tuple<D0Layout, D1Layout>,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         ck::Tuple<D0DataType, D1DataType>,
+                                         EDataType,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::MultiplyAdd>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<D0DataType, half_t> && is_same_v<D1DataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
+                     is_same_v<D0DataType, float> && is_same_v<D1DataType, float> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp
new file mode 100644
index 0000000000..dbd2b8f656
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp
@@ -0,0 +1,235 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f8_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f8_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f8_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f8_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+void add_device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGemmSplitK<ALayout,
+                                                   BLayout,
+                                                   CLayout,
+                                                   ADataType,
+                                                   BDataType,
+                                                   CDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGemmSplitK<ALayout,
+                                      BLayout,
+                                      CLayout,
+                                      ADataType,
+                                      BDataType,
+                                      CDataType,
+                                      ck::tensor_operation::element_wise::PassThrough,
+                                      ck::tensor_operation::element_wise::PassThrough,
+                                      ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
+                     is_same_v<CDataType, float>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                          is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, half_t> &&
+                          is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f8_f16_f16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f8_f16_f16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f8_f16_f16_km_nk_mn_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, f8_t> &&
+                          is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f8_f16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp
new file mode 100644
index 0000000000..2df378b0c6
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_streamk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef CK_ENABLE_FP16
+void add_device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmStreamK<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances);
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmStreamK<
+    ALayout,
+    BLayout,
+    CLayout,
+    ADataType,
+    BDataType,
+    CDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGemmStreamK<ALayout,
+                                       BLayout,
+                                       CLayout,
+                                       ADataType,
+                                       BDataType,
+                                       CDataType,
+                                       ck::tensor_operation::element_wise::PassThrough,
+                                       ck::tensor_operation::element_wise::PassThrough,
+                                       ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+#if 0
+        if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
+                     is_same_v<CDataType, float>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                          is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+            }
+        }
+#endif
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp
new file mode 100644
index 0000000000..26acf4f5f7
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp
@@ -0,0 +1,149 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F16  = ck::half_t;
+using F32  = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using namespace ck::tensor_layout::convolution;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault = ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// f16_f16_f32_f16
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          ConvolutionBackwardDataSpecialization ConvSpec>
+using device_grouped_conv_bwd_data_xdl_f16_instances =
+    std::tuple<
+        // clang-format off
+        // ##############################################|       NDim| ALayout| BLayout|    DsLayout| ELayout| AData| BData| AccData| CShuffle|      DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
+        // ##############################################|    Spatial|        |        |            |        |  Type|  Type|    Type| DataType|        Type|  Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
+        // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
+        // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
+        // generic instance
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,        S<4, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,            1,            1,     S<1, 16, 1, 4>,                1>,
+        // instances for small conv.K and conv.C
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,       S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,        S<4, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,            1,            1,     S<1, 32, 1, 4>,                1>,
+
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,        S<4, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,        S<4, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,        S<4, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,        S<4, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,        S<4, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,        S<4, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>
+        // clang-format on
+        >;
+
+// bf16_bf16_f32_bf16
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          ConvolutionBackwardDataSpecialization ConvSpec>
+using device_grouped_conv_bwd_data_xdl_bf16_instances = std::tuple<
+    // clang-format off
+        // ##############################################|          NDim| ALayout| BLayout|    DsLayout| ELayout| AData| BData| AccData| CShuffle|      DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
+        // ##############################################|       Spatial|        |        |            |        |  Type|  Type|    Type| DataType|        Type|  Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
+        // ##############################################|              |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
+        // ##############################################|              |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
+        // generic instance
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,    S<4, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,            1,            1,     S<1, 16, 1, 4>,                1>,
+        // instances for small conv.K and conv.C
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,   S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,    S<4, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,            1,            1,     S<1, 32, 1, 4>,                1>,
+
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,   S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,   S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,   S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,   S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,   S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,    S<4, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,   S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,    S<4, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,    S<4, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,   S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,    S<4, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,   S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,    S<4, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,    S<4, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>
+    // clang-format on
+    >;
+
+// f32_f32_f32_f32
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          ConvolutionBackwardDataSpecialization ConvSpec>
+using device_grouped_conv_bwd_data_xdl_f32_instances =
+    std::tuple<
+        // clang-format off
+         // ##############################################|       NDim| ALayout| BLayout|    DsLayout| ELayout| AData| BData| AccData| CShuffle|      DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
+         // ##############################################|    Spatial|        |        |            |        |  Type|  Type|    Type| DataType|        Type|  Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
+         // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
+         // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
+        // generic instance
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              4,         1,       S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,         1,            1,            1,     S<1, 16, 1, 4>,                1>,
+        // instances for small conv.K and conv.C
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              4,         1,       S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,        S<4, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,         1,            1,            1,     S<1, 32, 1, 4>,                1>,
+
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 4>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 4>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,        S<4, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 4>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 8>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,        S<4, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 4>,                4>,
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,        S<4, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>
+        // clang-format on
+        >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp
new file mode 100644
index 0000000000..6fc91565b9
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp
@@ -0,0 +1,139 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using namespace ck::tensor_layout::convolution;
+
+using BF16 = ck::bhalf_t;
+using F16  = ck::half_t;
+using F32  = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+static constexpr auto ConvBwdWeightFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
+
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec>
+using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_instances = std::tuple<
+    // clang-format off
+        //#########################################|     Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|              ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                    Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        // generic instance
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              1,              4,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              1,              4,      true,           1,           1,   S<1, 16, 1, 4>,               1>,
+        // instances for small conv.K and conv.C
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,   128,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              1,              1,      true,           1,           1,   S<1, 32, 1, 4>,               1>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    64,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              1,              2,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 32, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 64, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 16, 4>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,   128,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,    32,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  4>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              1,      true,  S<1, 4, 32, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 16, 1>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    64,     4,  4,   32,   32,    1,    2,  S<1, 4, 8,  2>,   S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              4,              2,      true,  S<1, 4, 16, 1>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              4,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+    // clang-format on
+    >;
+
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec>
+using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_instances = std::tuple<
+    // clang-format off
+        //#########################################|     Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|              ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                    Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        // generic instance
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              2,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              2,              4,      true,           1,           1,   S<1, 16, 1, 4>,               2>,   
+        // instance for small conv.K
+        // for fp16 conv.K and conv.C must be divisible by 2
+        // since half_t atomic_add require scalar_per_x_vector % 2 == 0
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              2,              1,      true,           1,           1,   S<1, 32, 1, 4>,               2>,  
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              2,              2,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>,
+
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 32, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              4,      true,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               8>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              2,      true,  S<1, 4, 8,  2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec>
+using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances = std::tuple<
+    // clang-format off
+        //#########################################|     Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|               ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //#########################################|     Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                     Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //#########################################| Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|             Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //#########################################|        |         |          |          |       |        |        |        |            |            |            |                           |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        // generic instance
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              1,              4,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              1,              4,      true,           1,           1,   S<1, 16, 1, 4>,               1>,   
+        // instance for small conv.K
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              1,              1,      true,           1,           1,   S<1, 32, 1, 4>,               1>, 
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              1,              2,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,
+
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 8>,               4>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 16, 4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 32, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 8>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 8,  2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 8,  8>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  8>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              1,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              1,      true,   S<1, 4, 16, 2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 8,  2>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              4,      true,   S<1, 4, 4,  4>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              2,      true,           1,           1,   S<1, 16, 1, 4>,               4>,   
+        DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     F32,    BF16,     F32, PassThrough, PassThrough, PassThrough,                   ConvSpec,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  4>, S<0, 3, 1, 2>,   S<0, 2, 1, 3>,              2,              8,              2,      true,   S<1, 4, 8,  2>,   S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 16, 1, 4>,               4>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp
new file mode 100644
index 0000000000..86ff43e5cf
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp
@@ -0,0 +1,90 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F16  = ck::half_t;
+using F32  = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using namespace ck::tensor_layout::convolution;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto ConvFwdOddC =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+template <typename InLayout,
+          typename WeiLayout,
+          typename DsLayout,
+          typename OutLayout,
+          typename DsDatatype,
+          typename CDEElementOp,
+          ConvolutionForwardSpecialization ConvSpec>
+using device_grouped_conv2d_fwd_dl_f16_instances = std::tuple<
+    // clang-format off
+           // ########################################|        NDim| InData| WeiData|    MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout|          In|          Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+           // ########################################|     Spatial|   Type|    Type|        Type|    Type|    Type|         |          |    Layout|          | Elementwise|  Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+           // ########################################|            |       |        |            |        |        |         |          |          |          |   Operation|    Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+           // ########################################|            |       |        |            |        |        |         |          |          |          |            |             |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        // generic instances
+        // TODO: Change to ScalarPerVector = 1 when inner_product<half_t, half_t, float> will be supported
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2,    F16,     F16,  DsDatatype,     F16,     F32, InLayout, WeiLayout,  DsLayout, OutLayout, PassThrough,  PassThrough,  CDEElementOp,       ConvSpec,    GemmMNKPadding,     8,    16,     4,     2,  2,          1,          2,      1,       S<4, 2>,       S<1, 1>,      S<2, 1, 2, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<1, 1, 1, 2>,        S<2, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 2>,
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2,    F16,     F16,  DsDatatype,     F16,     F32, InLayout, WeiLayout,  DsLayout, OutLayout, PassThrough,  PassThrough,  CDEElementOp,       ConvSpec,    GemmMNKPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 2>,
+
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2,    F16,     F16,  DsDatatype,     F16,     F32, InLayout, WeiLayout,  DsLayout, OutLayout, PassThrough,  PassThrough,  CDEElementOp,       ConvSpec,    GemmMNKPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+template <typename InLayout,
+          typename WeiLayout,
+          typename DsLayout,
+          typename OutLayout,
+          typename DsDatatype,
+          typename CDEElementOp,
+          ConvolutionForwardSpecialization ConvSpec>
+using device_grouped_conv2d_fwd_dl_f32_instances = std::tuple<
+    // clang-format off
+        // clang-format off
+           // ########################################|        NDim| InData| WeiData|    MultpleD| OutData| AccData| InLayout| WeiLayout| MultipleD| OutLayout|          In|          Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+           // ########################################|     Spatial|   Type|    Type|        Type|    Type|    Type|         |          |    Layout|          | Elementwise|  Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+           // ########################################|            |       |        |            |        |        |         |          |          |          |   Operation|    Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+           // ########################################|            |       |        |            |        |        |         |          |          |          |            |             |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        // generic instances
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2,    F32,     F32,  DsDatatype,     F32,     F32, InLayout, WeiLayout,  DsLayout, OutLayout, PassThrough, PassThrough,  CDEElementOp,       ConvSpec,    GemmMNKPadding,     8,    16,      4,    2,  1,          1,          2,      1,       S<4, 2>,       S<1, 1>,      S<2, 1, 2, 1>,      S<1, 1, 8, 1>,    S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 1, 1, 1>,        S<2, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 1>,
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2,    F32,     F32,  DsDatatype,     F32,     F32, InLayout, WeiLayout,  DsLayout, OutLayout, PassThrough, PassThrough,  CDEElementOp,       ConvSpec,    GemmMNKPadding,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 1>,
+
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2,    F32,     F32,  DsDatatype,     F32,     F32, InLayout, WeiLayout,  DsLayout, OutLayout, PassThrough, PassThrough,  CDEElementOp,       ConvSpec,    GemmMNKPadding,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp
new file mode 100644
index 0000000000..23edf35e98
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp
@@ -0,0 +1,180 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F16  = ck::half_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using namespace ck::tensor_layout::convolution;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 = ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 = ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto ConvFwdOddC =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
+
+static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
+
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          ConvolutionForwardSpecialization ConvSpec>
+using device_grouped_conv_fwd_xdl_bf16_instances = std::tuple<
+    // clang-format off
+        //########################################|     NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################|    Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|           |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|           |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        // generic instance
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1>,
+        // instances for small conv.K and conv.C
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          ConvolutionForwardSpecialization ConvSpec>
+using device_grouped_conv_fwd_xdl_f16_instances = std::tuple<
+    // clang-format off
+        //########################################|     NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################|    Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|           |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|           |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        // generic instance
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1>,
+        // instances for small conv.K and conv.C
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          ConvolutionForwardSpecialization ConvSpec>
+using device_grouped_conv_fwd_xdl_f32_instances = std::tuple<
+    // clang-format off
+        //########################################|     NumDim|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|          Ds| EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################|    Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|           |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|           |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        // generic instance
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,         1,           1,           1,              S<1,  8, 1,  8>,              1>,
+        // instances for small conv.K and conv.C
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              1>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F32,   F32,     F32,      F32,    DsLayout,   F32, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>
+    // clang-format on
+    >;
+
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          ConvolutionForwardSpecialization ConvSpec>
+using device_grouped_conv_fwd_xdl_int8_instances = std::tuple<
+    // clang-format off
+        //########################################|     NumDim|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|          Ds|  EData|           A|           B|         CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //########################################|    Spatial| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|    DataType|   Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //########################################|           |       |       |            |       |       |       |        |         |            |       |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //########################################|           |       |       |            |       |       |       |        |         |            |       |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        // generic instance
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1>,
+        // instances for small conv.K and conv.C
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout, int8_t, int8_t, int32_t,   int8_t,    DsLayout, int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp
new file mode 100644
index 0000000000..58c9064535
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp
@@ -0,0 +1,327 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// conv2d backward data
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
+                                                                  GNHWK,
+                                                                  GKYXC,
+                                                                  Empty_Tuple,
+                                                                  GNHWC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
+                                                                  GNHWK,
+                                                                  GKYXC,
+                                                                  Empty_Tuple,
+                                                                  GNHWC,
+                                                                  F32,
+                                                                  F32,
+                                                                  Empty_Tuple,
+                                                                  F32,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
+                                                                  GNHWK,
+                                                                  GKYXC,
+                                                                  Empty_Tuple,
+                                                                  GNHWC,
+                                                                  BF16,
+                                                                  BF16,
+                                                                  Empty_Tuple,
+                                                                  BF16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
+                                                                  NHWGK,
+                                                                  GKYXC,
+                                                                  Empty_Tuple,
+                                                                  NHWGC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
+                                                                  NHWGK,
+                                                                  GKYXC,
+                                                                  Empty_Tuple,
+                                                                  NHWGC,
+                                                                  F32,
+                                                                  F32,
+                                                                  Empty_Tuple,
+                                                                  F32,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
+                                                                  NHWGK,
+                                                                  GKYXC,
+                                                                  Empty_Tuple,
+                                                                  NHWGC,
+                                                                  BF16,
+                                                                  BF16,
+                                                                  Empty_Tuple,
+                                                                  BF16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+// conv3d backward data
+void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  GNDHWK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  GNDHWC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  GNDHWK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  GNDHWC,
+                                                                  F32,
+                                                                  F32,
+                                                                  Empty_Tuple,
+                                                                  F32,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  GNDHWK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  GNDHWC,
+                                                                  BF16,
+                                                                  BF16,
+                                                                  Empty_Tuple,
+                                                                  BF16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  NDHWGK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  NDHWGC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  NDHWGK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  NDHWGC,
+                                                                  F32,
+                                                                  F32,
+                                                                  Empty_Tuple,
+                                                                  F32,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  NDHWGK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  NDHWGC,
+                                                                  BF16,
+                                                                  BF16,
+                                                                  Empty_Tuple,
+                                                                  BF16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+template <ck::index_t NumDimSpatial,
+          typename OutLayout,
+          typename WeiLayout,
+          typename InLayout,
+          typename OutDataType,
+          typename WeiDataType,
+          typename InDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<
+        NumDimSpatial,
+        OutLayout,
+        WeiLayout,
+        Empty_Tuple,
+        InLayout,
+        OutDataType,
+        WeiDataType,
+        Empty_Tuple,
+        InDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp =
+        DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
+                                          OutLayout,
+                                          WeiLayout,
+                                          Empty_Tuple,
+                                          InLayout,
+                                          OutDataType,
+                                          WeiDataType,
+                                          Empty_Tuple,
+                                          InDataType,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+        if constexpr(NumDimSpatial == 2)
+        {
+
+            if constexpr(is_same_v<InLayout, GNHWC> && is_same_v<WeiLayout, GKYXC> &&
+                         is_same_v<OutLayout, GNHWK>)
+            {
+                if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
+                             is_same_v<OutDataType, F16>)
+                {
+                    add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
+                                  is_same_v<OutDataType, F32>)
+                {
+                    add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
+                                  is_same_v<OutDataType, BF16>)
+                {
+                    add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances(
+                        op_ptrs);
+                }
+            }
+            else if constexpr(is_same_v<InLayout, NHWGC> && is_same_v<WeiLayout, GKYXC> &&
+                              is_same_v<OutLayout, NHWGK>)
+            {
+                if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
+                             is_same_v<OutDataType, F16>)
+                {
+                    add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
+                                  is_same_v<OutDataType, F32>)
+                {
+                    add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
+                                  is_same_v<OutDataType, BF16>)
+                {
+                    add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances(
+                        op_ptrs);
+                }
+            }
+        }
+        else if constexpr(NumDimSpatial == 3)
+        {
+
+            if constexpr(is_same_v<InLayout, GNDHWC> && is_same_v<WeiLayout, GKZYXC> &&
+                         is_same_v<OutLayout, GNDHWK>)
+            {
+                if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
+                             is_same_v<OutDataType, F16>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
+                                  is_same_v<OutDataType, F32>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
+                                  is_same_v<OutDataType, BF16>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances(
+                        op_ptrs);
+                }
+            }
+            else if constexpr(is_same_v<InLayout, NDHWGC> && is_same_v<WeiLayout, GKZYXC> &&
+                              is_same_v<OutLayout, NDHWGK>)
+            {
+                if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
+                             is_same_v<OutDataType, F16>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
+                                  is_same_v<OutDataType, F32>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
+                                  is_same_v<OutDataType, BF16>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
+                        op_ptrs);
+                }
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
new file mode 100644
index 0000000000..cad3e1ace8
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
@@ -0,0 +1,367 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// conv1d backward weight
+void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
+                                                           GNWC,
+                                                           GKXC,
+                                                           GNWK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
+                                                           GNWC,
+                                                           GKXC,
+                                                           GNWK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
+                                                           GNWC,
+                                                           GKXC,
+                                                           GNWK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+// conv2d backward weight
+void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           GNHWC,
+                                                           GKYXC,
+                                                           GNHWK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           GNHWC,
+                                                           GKYXC,
+                                                           GNHWK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           GNHWC,
+                                                           GKYXC,
+                                                           GNHWK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+// conv3d backward weight
+void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           GNDHWC,
+                                                           GKZYXC,
+                                                           GNDHWK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           GNDHWC,
+                                                           GKZYXC,
+                                                           GNDHWK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           GNDHWC,
+                                                           GKZYXC,
+                                                           GNDHWK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvBwdWeight<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGroupedConvBwdWeight<NumDimSpatial,
+                                                InLayout,
+                                                WeiLayout,
+                                                OutLayout,
+                                                InDataType,
+                                                WeiDataType,
+                                                OutDataType,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 1)
+        {
+            if constexpr(is_same_v<InLayout, GNWC> && is_same_v<WeiLayout, GKXC> &&
+                         is_same_v<OutLayout, GNWK>)
+            {
+                if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                             is_same_v<OutDataType, float>)
+                {
+                    add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instances(op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                                  is_same_v<OutDataType, half_t>)
+                {
+                    add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances(op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                                  is_same_v<WeiDataType, float> &&
+                                  is_same_v<OutDataType, ck::bhalf_t>)
+                {
+                    add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances(
+                        op_ptrs);
+                }
+            }
+        }
+        else if constexpr(NumDimSpatial == 2)
+        {
+            if constexpr(is_same_v<InLayout, GNHWC> && is_same_v<WeiLayout, GKYXC> &&
+                         is_same_v<OutLayout, GNHWK>)
+            {
+                if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                             is_same_v<OutDataType, float>)
+                {
+                    add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                                  is_same_v<OutDataType, half_t>)
+                {
+                    add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                                  is_same_v<WeiDataType, float> &&
+                                  is_same_v<OutDataType, ck::bhalf_t>)
+                {
+                    add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances(
+                        op_ptrs);
+                }
+            }
+            else if constexpr(is_same_v<InLayout, NHWGC> && is_same_v<WeiLayout, GKYXC> &&
+                              is_same_v<OutLayout, NHWGK>)
+            {
+                if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                             is_same_v<OutDataType, float>)
+                {
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                                  is_same_v<OutDataType, half_t>)
+                {
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                                  is_same_v<WeiDataType, float> &&
+                                  is_same_v<OutDataType, ck::bhalf_t>)
+                {
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances(
+                        op_ptrs);
+                }
+            }
+        }
+        else if constexpr(NumDimSpatial == 3)
+        {
+            if constexpr(is_same_v<InLayout, GNDHWC> && is_same_v<WeiLayout, GKZYXC> &&
+                         is_same_v<OutLayout, GNDHWK>)
+            {
+                if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                             is_same_v<OutDataType, float>)
+                {
+                    add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                                  is_same_v<OutDataType, half_t>)
+                {
+                    add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                                  is_same_v<WeiDataType, float> &&
+                                  is_same_v<OutDataType, ck::bhalf_t>)
+                {
+                    add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
+                        op_ptrs);
+                }
+            }
+            else if constexpr(is_same_v<InLayout, NDHWGC> && is_same_v<WeiLayout, GKZYXC> &&
+                              is_same_v<OutLayout, NDHWGK>)
+            {
+                if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                             is_same_v<OutDataType, float>)
+                {
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                                  is_same_v<OutDataType, half_t>)
+                {
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                                  is_same_v<WeiDataType, float> &&
+                                  is_same_v<OutDataType, ck::bhalf_t>)
+                {
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances(
+                        op_ptrs);
+                }
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
new file mode 100644
index 0000000000..2a0da10c65
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
@@ -0,0 +1,498 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// grouped conv1d forward, GNWC/GKXC/GNWK
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+// grouped conv2d forward, GNHWC/GKYXC/GNHWK
+void add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+// grouped conv2d forward, NHWGC/GKYXC/NHWGK
+void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+// grouped conv3d forward, GNDHWC/GKZYXC/GNDHWK
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    Empty_Tuple,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    Empty_Tuple,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                   InLayout,
+                                                   WeiLayout,
+                                                   Empty_Tuple,
+                                                   OutLayout,
+                                                   InDataType,
+                                                   WeiDataType,
+                                                   Empty_Tuple,
+                                                   OutDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, GNWC> &&
+                     is_same_v<WeiLayout, GKXC> && is_same_v<OutLayout, GNWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                              is_same_v<OutDataType, int8_t>)
+            {
+                add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances(op_ptrs);
+            }
+        }
+        else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
+                          is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
+                add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
+                add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
+                          is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
+                add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
+                add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, GNDHWC> &&
+                          is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, GNDHWK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                              is_same_v<OutDataType, int8_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(op_ptrs);
+            }
+        }
+        else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
+                          is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, NDHWGK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                              is_same_v<OutDataType, int8_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp
new file mode 100644
index 0000000000..070c7e5b17
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp
@@ -0,0 +1,195 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#ifdef CK_ENABLE_FP16
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances);
+
+template <typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
+                                                                                      BLayout,
+                                                                                      Empty_Tuple,
+                                                                                      ELayout,
+                                                                                      ADataType,
+                                                                                      BDataType,
+                                                                                      Empty_Tuple,
+                                                                                      EDataType,
+                                                                                      PassThrough,
+                                                                                      PassThrough,
+                                                                                      PassThrough>>
+{
+    using DeviceOp = DeviceGroupedGemm<ALayout,
+                                       BLayout,
+                                       Empty_Tuple,
+                                       ELayout,
+                                       ADataType,
+                                       BDataType,
+                                       Empty_Tuple,
+                                       EDataType,
+                                       PassThrough,
+                                       PassThrough,
+                                       PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+                add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+                add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+                add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+                add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_bias.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_bias.hpp
new file mode 100644
index 0000000000..9365bf7bc2
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_bias.hpp
@@ -0,0 +1,146 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// fp16_output
+void add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
+                                                         Row,
+                                                         Row_Tuple,
+                                                         Row,
+                                                         F16,
+                                                         F16,
+                                                         F32_Tuple,
+                                                         F16,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         Add>>>& instances);
+
+void add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
+                                                         Col,
+                                                         Row_Tuple,
+                                                         Row,
+                                                         F16,
+                                                         F16,
+                                                         F32_Tuple,
+                                                         F16,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         Add>>>& instances);
+
+// fp32_output
+void add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
+                                                         Row,
+                                                         Row_Tuple,
+                                                         Row,
+                                                         F16,
+                                                         F16,
+                                                         F32_Tuple,
+                                                         F32,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         Add>>>& instances);
+
+void add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
+                                                         Col,
+                                                         Row_Tuple,
+                                                         Row,
+                                                         F16,
+                                                         F16,
+                                                         F32_Tuple,
+                                                         F32,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         Add>>>& instances);
+
+template <typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
+                                                           BLayout,
+                                                           Row_Tuple,
+                                                           ELayout,
+                                                           ADataType,
+                                                           BDataType,
+                                                           F32_Tuple,
+                                                           EDataType,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Add>>
+{
+    using DeviceOp = DeviceGroupedGemmFixedNK<ALayout,
+                                              BLayout,
+                                              Row_Tuple,
+                                              ELayout,
+                                              ADataType,
+                                              BDataType,
+                                              F32_Tuple,
+                                              EDataType,
+                                              PassThrough,
+                                              PassThrough,
+                                              Add>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        // fp16_output
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+            }
+        }
+
+        // fp32_output
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<EDataType, float>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instances(op_ptrs);
+            }
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instances(op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp
new file mode 100644
index 0000000000..eb181ebc7f
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp
@@ -0,0 +1,136 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  FastGelu>>>& instances);
+
+void add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  FastGelu>>>& instances);
+
+void add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  FastGelu>>>& instances);
+
+void add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  FastGelu>>>& instances);
+
+// GroupedGEMM + GELU
+template <typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
+                                                                                      BLayout,
+                                                                                      Empty_Tuple,
+                                                                                      ELayout,
+                                                                                      ADataType,
+                                                                                      BDataType,
+                                                                                      Empty_Tuple,
+                                                                                      EDataType,
+                                                                                      PassThrough,
+                                                                                      PassThrough,
+                                                                                      FastGelu>>
+{
+    using DeviceOp = DeviceGroupedGemm<ALayout,
+                                       BLayout,
+                                       Empty_Tuple,
+                                       ELayout,
+                                       ADataType,
+                                       BDataType,
+                                       Empty_Tuple,
+                                       EDataType,
+                                       PassThrough,
+                                       PassThrough,
+                                       FastGelu>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/image_to_column.hpp b/library/include/ck/library/tensor_operation_instance/gpu/image_to_column.hpp
new file mode 100644
index 0000000000..6c4526ba4e
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/image_to_column.hpp
@@ -0,0 +1,135 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_image_to_column.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// nhwc, 1d
+void add_device_image_to_column_nhwc_1d_bf16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, BF16, BF16>>>& instances);
+
+void add_device_image_to_column_nhwc_1d_f16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, F16, F16>>>& instances);
+
+void add_device_image_to_column_nhwc_1d_f32_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, F32, F32>>>& instances);
+
+void add_device_image_to_column_nhwc_1d_i8_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, int8_t, int8_t>>>& instances);
+// nhwc, 2d
+void add_device_image_to_column_nhwc_2d_bf16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, BF16, BF16>>>& instances);
+
+void add_device_image_to_column_nhwc_2d_f16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, F16, F16>>>& instances);
+
+void add_device_image_to_column_nhwc_2d_f32_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, F32, F32>>>& instances);
+
+void add_device_image_to_column_nhwc_2d_i8_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, int8_t, int8_t>>>& instances);
+// nhwc, 3d
+void add_device_image_to_column_nhwc_3d_bf16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, BF16, BF16>>>& instances);
+
+void add_device_image_to_column_nhwc_3d_f16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, F16, F16>>>& instances);
+
+void add_device_image_to_column_nhwc_3d_f32_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, F32, F32>>>& instances);
+
+void add_device_image_to_column_nhwc_3d_i8_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, int8_t, int8_t>>>& instances);
+
+template <ck::index_t NumDimSpatial, typename InLayout, typename InDataType, typename OutDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::
+        DeviceImageToColumn<NumDimSpatial, InLayout, InDataType, OutDataType>>
+{
+    using DeviceOp = DeviceImageToColumn<NumDimSpatial, InLayout, InDataType, OutDataType>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, GNWC>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
+            {
+                add_device_image_to_column_nhwc_1d_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
+            {
+                add_device_image_to_column_nhwc_1d_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_image_to_column_nhwc_1d_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
+            {
+                add_device_image_to_column_nhwc_1d_i8_instances(op_ptrs);
+            }
+        }
+        else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
+            {
+                add_device_image_to_column_nhwc_2d_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
+            {
+                add_device_image_to_column_nhwc_2d_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_image_to_column_nhwc_2d_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
+            {
+                add_device_image_to_column_nhwc_2d_i8_instances(op_ptrs);
+            }
+        }
+        else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, GNDHWC>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
+            {
+                add_device_image_to_column_nhwc_3d_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
+            {
+                add_device_image_to_column_nhwc_3d_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_image_to_column_nhwc_3d_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
+            {
+                add_device_image_to_column_nhwc_3d_i8_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/image_to_column/device_image_to_column_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/image_to_column/device_image_to_column_instance.hpp
new file mode 100644
index 0000000000..a2603218b2
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/image_to_column/device_image_to_column_instance.hpp
@@ -0,0 +1,121 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using namespace ck::tensor_layout::convolution;
+
+using BF16 = ck::bhalf_t;
+using F16  = ck::half_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+template <ck::index_t NDimSpatial, typename InLayout>
+using device_image_to_column_bf16_instances = std::tuple<
+    // clang-format off
+        //#####################|        Num| InLayout| InDataType| OutDataType| Block|  MPer|  KPer|    Thread| Scalar|
+        //#####################|        Dim|         |           |            |  Size| Block| Block|   Cluster|    Per|
+        //#####################|    Spatial|         |           |            |      |      |      |   Lengths| Vector|
+        //#####################|           |         |           |            |      |      |      |          |       |
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,    64,     8,     8,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,    64,    16,    16,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,    64,    32,    32,   S<8, 8>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,    64,    64,    64,   S<8, 8>,     8>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   128,    16,    16,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   128,    64,    64,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   128,    32,    64,  S<8, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   128,    64,   128,  S<8, 16>,     8>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   256,    16,    16, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   256,    64,    64, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   256,   128,   128, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   256,    64,    64, S<16, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   256,   128,   128, S<16, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   256,   128,   128, S<16, 16>,     8>
+    // clang-format on
+    >;
+
+template <ck::index_t NDimSpatial, typename InLayout>
+using device_image_to_column_f16_instances = std::tuple<
+    // clang-format off
+        //#####################|        Num| InLayout| InDataType| OutDataType| Block|  MPer|  KPer|    Thread| Scalar|
+        //#####################|        Dim|         |           |            |  Size| Block| Block|   Cluster|    Per|
+        //#####################|    Spatial|         |           |            |      |      |      |   Lengths| Vector|
+        //#####################|           |         |           |            |      |      |      |          |       |
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,    64,     8,     8,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,    64,    16,    16,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,    64,    32,    32,   S<8, 8>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,    64,    64,    64,   S<8, 8>,     8>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   128,    16,    16,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   128,    64,    64,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   128,    32,    64,  S<8, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   128,    64,   128,  S<8, 16>,     8>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   256,    16,    16, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   256,    64,    64, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   256,   128,   128, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   256,    64,    64, S<16, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   256,   128,   128, S<16, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   256,   128,   128, S<16, 16>,     8>
+    // clang-format on
+    >;
+
+template <ck::index_t NDimSpatial, typename InLayout>
+using device_image_to_column_f32_instances = std::tuple<
+    // clang-format off
+        //#####################|        Num| InLayout| InDataType| OutDataType| Block|  MPer|  KPer|    Thread| Scalar|
+        //#####################|        Dim|         |           |            |  Size| Block| Block|   Cluster|    Per|
+        //#####################|    Spatial|         |           |            |      |      |      |   Lengths| Vector|
+        //#####################|           |         |           |            |      |      |      |          |       |
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,    64,     8,     8,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,    64,    16,    16,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,    64,    32,    32,   S<8, 8>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   128,    16,    16,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   128,    64,    64,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   128,    32,    64,  S<8, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   256,    16,    16, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   256,    64,    64, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   256,   128,   128, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   256,    64,    64, S<16, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   256,   128,   128, S<16, 16>,     4>
+    // clang-format on
+    >;
+
+template <ck::index_t NDimSpatial, typename InLayout>
+using device_image_to_column_i8_instances = std::tuple<
+    // clang-format off
+        //#####################|        Num| InLayout| InDataType| OutDataType| Block|  MPer|  KPer|    Thread| Scalar|
+        //#####################|        Dim|         |           |            |  Size| Block| Block|   Cluster|    Per|
+        //#####################|    Spatial|         |           |            |      |      |      |   Lengths| Vector|
+        //#####################|           |         |           |            |      |      |      |          |       |
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,    64,     8,     8,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,    64,    16,    16,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,    64,    32,    32,   S<8, 8>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,    64,    64,    64,   S<8, 8>,     8>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   128,    16,    16,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   128,    64,    64,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   128,    32,    64,  S<8, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   128,    64,   128,  S<8, 16>,     8>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   256,    16,    16, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   256,    64,    64, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   256,   128,   128, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   256,    64,    64, S<16, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   256,   128,   128, S<16, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   256,   128,   128, S<16, 16>,     8>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   256,   256,   256, S<16, 16>,     16>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp b/library/include/ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp
new file mode 100644
index 0000000000..63ea4f2891
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+#ifdef CK_ENABLE_FP16
+void add_device_maxpool_bwd_f16_instances(
+    std::vector<std::unique_ptr<DeviceMaxPoolBwd<F16, I32, F16>>>&);
+#endif
+#ifdef CK_ENABLE_BF16
+void add_device_maxpool_bwd_bf16_instances(
+    std::vector<std::unique_ptr<DeviceMaxPoolBwd<BF16, I32, BF16>>>&);
+#endif
+#ifdef CK_ENABLE_FP32
+void add_device_maxpool_bwd_f32_instances(
+    std::vector<std::unique_ptr<DeviceMaxPoolBwd<F32, I32, F32>>>&);
+#endif
+template <typename DOutDataType, typename IndexDataType, typename DInDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceMaxPoolBwd<DOutDataType, IndexDataType, DInDataType>>
+{
+    using DeviceOp = DeviceMaxPoolBwd<DOutDataType, IndexDataType, DInDataType>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+#ifdef CK_ENABLE_FP16
+        if constexpr(is_same_v<DOutDataType, F16> && is_same_v<DInDataType, F16> &&
+                     is_same_v<IndexDataType, I32>)
+            add_device_maxpool_bwd_f16_instances(op_ptrs);
+#endif
+#ifdef CK_ENABLE_BF16
+        else if constexpr(is_same_v<DOutDataType, BF16> && is_same_v<DInDataType, BF16> &&
+                          is_same_v<IndexDataType, I32>)
+            add_device_maxpool_bwd_bf16_instances(op_ptrs);
+#endif
+#ifdef CK_ENABLE_FP32
+        else if constexpr(is_same_v<DOutDataType, F32> && is_same_v<DInDataType, F32> &&
+                          is_same_v<IndexDataType, I32>)
+            add_device_maxpool_bwd_f32_instances(op_ptrs);
+#endif
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/normalization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/normalization.hpp
new file mode 100644
index 0000000000..8e90a7ea98
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/normalization.hpp
@@ -0,0 +1,112 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef CK_ENABLE_FP16
+// FP16
+void add_device_normalization_rank_2_1_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, PassThrough, 2, 1>>>&);
+
+void add_device_normalization_rank_4_3_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, PassThrough, 4, 3>>>&);
+
+void add_device_normalization_rank_5_3_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, PassThrough, 5, 3>>>&);
+#endif
+#ifdef CK_ENABLE_FP32
+// FP32
+void add_device_normalization_rank_2_1_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, PassThrough, 2, 1>>>&);
+
+void add_device_normalization_rank_4_3_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&);
+
+void add_device_normalization_rank_5_3_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, PassThrough, 5, 3>>>&);
+#endif
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceNormalization<
+    XDataType,
+    GammaDataType,
+    BetaDataType,
+    F32,
+    YDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    Rank,
+    NumReduceDim>>
+{
+    using DeviceOp = DeviceNormalization<XDataType,
+                                         GammaDataType,
+                                         BetaDataType,
+                                         F32,
+                                         YDataType,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         Rank,
+                                         NumReduceDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+#ifdef CK_ENABLE_FP16
+        if constexpr(is_same_v<XDataType, F16> && is_same_v<GammaDataType, F16> &&
+                     is_same_v<BetaDataType, F16> && is_same_v<YDataType, F16>)
+        {
+            if constexpr(Rank == 2 && NumReduceDim == 1)
+            {
+                add_device_normalization_rank_2_1_f16_instances(op_ptrs);
+            }
+            else if constexpr(Rank == 4 && NumReduceDim == 3)
+            {
+                add_device_normalization_rank_4_3_f16_instances(op_ptrs);
+            }
+            else if constexpr(Rank == 5 && NumReduceDim == 3)
+            {
+                add_device_normalization_rank_5_3_f16_instances(op_ptrs);
+            }
+        }
+#endif
+#ifdef CK_ENABLE_FP32
+        if constexpr(is_same_v<XDataType, F32> && is_same_v<GammaDataType, F32> &&
+                     is_same_v<BetaDataType, F32> && is_same_v<YDataType, F32>)
+        {
+            if constexpr(Rank == 2 && NumReduceDim == 1)
+            {
+                add_device_normalization_rank_2_1_f32_instances(op_ptrs);
+            }
+            else if constexpr(Rank == 4 && NumReduceDim == 3)
+            {
+                add_device_normalization_rank_4_3_f32_instances(op_ptrs);
+            }
+            else if constexpr(Rank == 5 && NumReduceDim == 3)
+            {
+                add_device_normalization_rank_5_3_f32_instances(op_ptrs);
+            }
+        }
+#endif
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/normalization_swish.hpp b/library/include/ck/library/tensor_operation_instance/gpu/normalization_swish.hpp
new file mode 100644
index 0000000000..2391775299
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/normalization_swish.hpp
@@ -0,0 +1,93 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// FP16
+void add_device_normalization_rank_5_3_swish_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Swish, 5, 3>>>&);
+
+// FP32
+void add_device_normalization_rank_5_3_swish_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Swish, 5, 3>>>&);
+
+// [x, gamma, beta, y] = [f16, f32, f32, f16]
+void add_device_normalization_rank_5_3_swish_f16_f32_f32_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F32, F32, F32, F16, Swish, 5, 3>>>&);
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename YDataType,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceNormalization<XDataType,
+                                                      GammaDataType,
+                                                      BetaDataType,
+                                                      F32,
+                                                      YDataType,
+                                                      ck::tensor_operation::element_wise::Swish,
+                                                      Rank,
+                                                      NumReduceDim>>
+{
+    using DeviceOp = DeviceNormalization<XDataType,
+                                         GammaDataType,
+                                         BetaDataType,
+                                         F32,
+                                         YDataType,
+                                         ck::tensor_operation::element_wise::Swish,
+                                         Rank,
+                                         NumReduceDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<XDataType, F16> && is_same_v<GammaDataType, F16> &&
+                     is_same_v<BetaDataType, F16> && is_same_v<YDataType, F16>)
+        {
+            if constexpr(Rank == 5 && NumReduceDim == 3)
+            {
+                add_device_normalization_rank_5_3_swish_f16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F32> && is_same_v<GammaDataType, F32> &&
+                          is_same_v<BetaDataType, F32> && is_same_v<YDataType, F32>)
+        {
+            if constexpr(Rank == 5 && NumReduceDim == 3)
+            {
+                add_device_normalization_rank_5_3_swish_f32_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F16> && is_same_v<GammaDataType, F32> &&
+                          is_same_v<BetaDataType, F32> && is_same_v<YDataType, F16>)
+        {
+            if constexpr(Rank == 5 && NumReduceDim == 3)
+            {
+                add_device_normalization_rank_5_3_swish_f16_f32_f32_f16_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp b/library/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp
new file mode 100644
index 0000000000..94ee68a409
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp
@@ -0,0 +1,153 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_pool_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto InOutRank  = 5;
+static constexpr auto WindowRank = 3;
+
+static constexpr auto MaxOp = ck::ReduceTensorOp::MAX;
+static constexpr auto AvgOp = ck::ReduceTensorOp::AVG;
+#ifdef CK_ENABLE_FP16
+// FP16
+void add_device_pool3d_fwd_ndhwc_f16_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, NDHWC, NDHWC, MaxOp, false>>>&);
+
+void add_device_pool3d_fwd_ndhwc_f16_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, NDHWC, NDHWC, AvgOp, false>>>&);
+
+// FP16 - return index
+void add_device_pool3d_fwd_ndhwc_index_f16_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, F16, F16, I32, NDHWC, NDHWC, MaxOp, true>>>&);
+#endif
+#ifdef CK_ENABLE_BF16
+// BF16
+void add_device_pool3d_fwd_ndhwc_bf16_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, BF16, BF16, I32, NDHWC, NDHWC, MaxOp, false>>>&);
+
+void add_device_pool3d_fwd_ndhwc_bf16_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, BF16, BF16, I32, NDHWC, NDHWC, AvgOp, false>>>&);
+
+// BF16 - return index
+void add_device_pool3d_fwd_ndhwc_index_bf16_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, BF16, BF16, I32, NDHWC, NDHWC, MaxOp, true>>>&);
+#endif
+#ifdef CK_ENABLE_FP32
+// FP32
+void add_device_pool3d_fwd_ndhwc_f32_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, NDHWC, NDHWC, MaxOp, false>>>&);
+
+void add_device_pool3d_fwd_ndhwc_f32_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, NDHWC, NDHWC, AvgOp, false>>>&);
+
+// FP32 - return index
+void add_device_pool3d_fwd_ndhwc_index_f32_instances(
+    std::vector<std::unique_ptr<
+        DevicePoolFwd<InOutRank, WindowRank, F32, F32, I32, NDHWC, NDHWC, MaxOp, true>>>&);
+#endif
+template <typename InDataType,
+          typename OutDataType,
+          typename IndexDataType,
+          typename InLayout,
+          typename OutLayout,
+          ck::ReduceTensorOp ReduceOpId,
+          bool OutputIndex>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DevicePoolFwd<InOutRank,
+                                                                                  WindowRank,
+                                                                                  InDataType,
+                                                                                  OutDataType,
+                                                                                  IndexDataType,
+                                                                                  InLayout,
+                                                                                  OutLayout,
+                                                                                  ReduceOpId,
+                                                                                  OutputIndex>>
+{
+    using DeviceOp = DevicePoolFwd<InOutRank,
+                                   WindowRank,
+                                   InDataType,
+                                   OutDataType,
+                                   IndexDataType,
+                                   InLayout,
+                                   OutLayout,
+                                   ReduceOpId,
+                                   OutputIndex>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+        if constexpr(is_same_v<InLayout, NDHWC> && is_same_v<OutLayout, NDHWC>)
+        {
+#ifdef CK_ENABLE_FP16
+            if constexpr(is_same_v<InDataType, F16> && is_same_v<OutDataType, F16> &&
+                         is_same_v<IndexDataType, I32>)
+            {
+                if constexpr(OutputIndex && ReduceOpId == MaxOp)
+                {
+                    add_device_pool3d_fwd_ndhwc_index_f16_instances(op_ptrs);
+                }
+                else
+                {
+                    add_device_pool3d_fwd_ndhwc_f16_instances(op_ptrs);
+                }
+            }
+#endif
+#ifdef CK_ENABLE_BF16
+            else if constexpr(is_same_v<InDataType, BF16> && is_same_v<OutDataType, BF16> &&
+                              is_same_v<IndexDataType, I32>)
+            {
+                if constexpr(OutputIndex && ReduceOpId == MaxOp)
+                {
+                    add_device_pool3d_fwd_ndhwc_index_bf16_instances(op_ptrs);
+                }
+                else
+                {
+                    add_device_pool3d_fwd_ndhwc_bf16_instances(op_ptrs);
+                }
+            }
+#endif
+#ifdef CK_ENABLE_FP32
+            else if constexpr(is_same_v<InDataType, F32> && is_same_v<OutDataType, F32> &&
+                              is_same_v<IndexDataType, I32>)
+            {
+                if constexpr(OutputIndex && ReduceOpId == MaxOp)
+                {
+                    add_device_pool3d_fwd_ndhwc_index_f32_instances(op_ptrs);
+                }
+                else
+                {
+                    add_device_pool3d_fwd_ndhwc_f32_instances(op_ptrs);
+                }
+            }
+#endif
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp
new file mode 100644
index 0000000000..19600a90f8
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp
@@ -0,0 +1,233 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef DL_KERNELS
+// Layout(A, B, C) = [Col, Row, Row]
+void add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+
+// Layout(A, B, C) = [Col, Col, Row]
+void add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+
+// Layout(A, B, C) = [Row, Row, Row]
+void add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+
+// Layout(A, B, C) = [Row, Col, Row]
+void add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+#endif
+// Layout(A, B, C) = [Col, Row, Row]
+void add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+
+// Layout(A, B, C) = [Col, Col, Row]
+void add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+
+// Layout(A, B, C) = [Row, Row, Row]
+void add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+
+// Layout(A, B, C) = [Row, Col, Row]
+void add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+
+template <typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename Activation>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<
+    ALayout,
+    BLayout,
+    Empty_Tuple,
+    ELayout,
+    ADataType,
+    BDataType,
+    Empty_Tuple,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    Activation_Mul_Clamp<Activation>>>
+{
+    using DeviceOp = DeviceGemmMultipleD<ALayout,
+                                         BLayout,
+                                         Empty_Tuple,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         Empty_Tuple,
+                                         EDataType,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         Activation_Mul_Clamp<Activation>>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
+                     is_same_v<EDataType, int8_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                if constexpr(is_same_v<Activation, PassThrough>)
+                {
+#ifdef DL_KERNELS
+                    add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(op_ptrs);
+#endif
+                    add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(op_ptrs);
+                }
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                if constexpr(is_same_v<Activation, PassThrough>)
+                {
+#ifdef DL_KERNELS
+                    add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(op_ptrs);
+#endif
+                    add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(op_ptrs);
+                }
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                if constexpr(is_same_v<Activation, PassThrough>)
+                {
+#ifdef DL_KERNELS
+                    add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instances(op_ptrs);
+#endif
+                    add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(op_ptrs);
+                }
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<ELayout, Row>)
+            {
+                if constexpr(is_same_v<Activation, PassThrough>)
+                {
+#ifdef DL_KERNELS
+                    add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instances(op_ptrs);
+#endif
+                    add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(op_ptrs);
+                }
+            }
+
+            return op_ptrs;
+        }
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
\ No newline at end of file
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp
new file mode 100644
index 0000000000..d9f1cdceb9
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp
@@ -0,0 +1,252 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef DL_KERNELS
+// grouped conv2d forward, NHWGC/GKYXC/NHWGK
+void add_device_conv2d_dl_bias_perchannel_quantization_int8_instances(
+    std::vector<
+        std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                      NHWGC,
+                                                      GKYXC,
+                                                      GK_GK_Tuple,
+                                                      NHWGK,
+                                                      int8_t,
+                                                      int8_t,
+                                                      I32_F32_Tuple,
+                                                      int8_t,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      Add_Activation_Mul2_Clamp<PassThrough>>>>&
+        instances);
+
+void add_device_conv2d_dl_bias_relu_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Activation_Mul2_Clamp<Relu>>>>&
+        instances);
+
+void add_device_conv2d_dl_bias_tanh_perchannel_quantization_int8_instances(
+    std::vector<
+        std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                      NHWGC,
+                                                      GKYXC,
+                                                      GK_GK_Tuple,
+                                                      NHWGK,
+                                                      int8_t,
+                                                      int8_t,
+                                                      I32_F32_Tuple,
+                                                      int8_t,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      Add_Mul2_Activation_Mul_Clamp<TanH>>>>&
+        instances);
+#endif
+void add_device_conv2d_xdl_bias_perchannel_quantization_int8_instances(
+    std::vector<
+        std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                      NHWGC,
+                                                      GKYXC,
+                                                      GK_GK_Tuple,
+                                                      NHWGK,
+                                                      int8_t,
+                                                      int8_t,
+                                                      I32_F32_Tuple,
+                                                      int8_t,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      Add_Activation_Mul2_Clamp<PassThrough>>>>&
+        instances);
+
+void add_device_conv2d_xdl_bias_relu_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Activation_Mul2_Clamp<Relu>>>>&
+        instances);
+
+void add_device_conv2d_xdl_bias_tanh_perchannel_quantization_int8_instances(
+    std::vector<
+        std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                      NHWGC,
+                                                      GKYXC,
+                                                      GK_GK_Tuple,
+                                                      NHWGK,
+                                                      int8_t,
+                                                      int8_t,
+                                                      I32_F32_Tuple,
+                                                      int8_t,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      Add_Mul2_Activation_Mul_Clamp<TanH>>>>&
+        instances);
+
+// piecewise activation function
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename DsLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename DsDataType,
+          typename OutDataType,
+          typename Activation>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    DsLayout,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    DsDataType,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    Add_Activation_Mul2_Clamp<Activation>>>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                   InLayout,
+                                                   WeiLayout,
+                                                   DsLayout,
+                                                   OutLayout,
+                                                   InDataType,
+                                                   WeiDataType,
+                                                   DsDataType,
+                                                   OutDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Add_Activation_Mul2_Clamp<Activation>>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
+                     is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_GK_Tuple> &&
+                     is_same_v<OutLayout, NHWGK>)
+        {
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<DsDataType, I32_F32_Tuple> && is_same_v<OutDataType, int8_t>)
+            {
+                if constexpr(is_same_v<Activation, PassThrough>)
+                {
+#ifdef DL_KERNELS
+                    add_device_conv2d_dl_bias_perchannel_quantization_int8_instances(op_ptrs);
+#endif
+                    add_device_conv2d_xdl_bias_perchannel_quantization_int8_instances(op_ptrs);
+                }
+                else if constexpr(is_same_v<Activation, Relu>)
+                {
+#ifdef DL_KERNELS
+                    add_device_conv2d_dl_bias_relu_perchannel_quantization_int8_instances(op_ptrs);
+#endif
+                    add_device_conv2d_xdl_bias_relu_perchannel_quantization_int8_instances(op_ptrs);
+                }
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+// non-piecewise activation function
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename DsLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename DsDataType,
+          typename OutDataType,
+          typename Activation>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    DsLayout,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    DsDataType,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    Add_Mul2_Activation_Mul_Clamp<Activation>>>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                   InLayout,
+                                                   WeiLayout,
+                                                   DsLayout,
+                                                   OutLayout,
+                                                   InDataType,
+                                                   WeiDataType,
+                                                   DsDataType,
+                                                   OutDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Add_Mul2_Activation_Mul_Clamp<Activation>>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
+                     is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_GK_Tuple> &&
+                     is_same_v<OutLayout, NHWGK>)
+        {
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<DsDataType, I32_F32_Tuple> && is_same_v<OutDataType, int8_t>)
+            {
+                if constexpr(is_same_v<Activation, TanH>)
+                {
+#ifdef DL_KERNELS
+                    add_device_conv2d_dl_bias_tanh_perchannel_quantization_int8_instances(op_ptrs);
+#endif
+                    add_device_conv2d_xdl_bias_tanh_perchannel_quantization_int8_instances(op_ptrs);
+                }
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp
new file mode 100644
index 0000000000..9f3919e394
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp
@@ -0,0 +1,250 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef DL_KERNELS
+// grouped conv2d forward, NHWGC/GKYXC/NHWGK
+void add_device_conv2d_dl_bias_perlayer_quantization_int8_instances(
+    std::vector<
+        std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                      NHWGC,
+                                                      GKYXC,
+                                                      GK_Tuple,
+                                                      NHWGK,
+                                                      int8_t,
+                                                      int8_t,
+                                                      I32_Tuple,
+                                                      int8_t,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      Add_Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+
+void add_device_conv2d_dl_bias_relu_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Activation_Mul_Clamp<Relu>>>>&
+        instances);
+
+void add_device_conv2d_dl_bias_tanh_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Mul_Activation_Mul_Clamp<TanH>>>>&
+        instances);
+#endif
+void add_device_conv2d_xdl_bias_perlayer_quantization_int8_instances(
+    std::vector<
+        std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                      NHWGC,
+                                                      GKYXC,
+                                                      GK_Tuple,
+                                                      NHWGK,
+                                                      int8_t,
+                                                      int8_t,
+                                                      I32_Tuple,
+                                                      int8_t,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      Add_Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+
+void add_device_conv2d_xdl_bias_relu_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Activation_Mul_Clamp<Relu>>>>&
+        instances);
+
+void add_device_conv2d_xdl_bias_tanh_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Mul_Activation_Mul_Clamp<TanH>>>>&
+        instances);
+
+// piecewise activation function
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename DsLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename DsDataType,
+          typename OutDataType,
+          typename Activation>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    DsLayout,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    DsDataType,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    Add_Activation_Mul_Clamp<Activation>>>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                   InLayout,
+                                                   WeiLayout,
+                                                   DsLayout,
+                                                   OutLayout,
+                                                   InDataType,
+                                                   WeiDataType,
+                                                   DsDataType,
+                                                   OutDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Add_Activation_Mul_Clamp<Activation>>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
+                     is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_Tuple> &&
+                     is_same_v<OutLayout, NHWGK>)
+        {
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<DsDataType, I32_Tuple> && is_same_v<OutDataType, int8_t>)
+            {
+                if constexpr(is_same_v<Activation, PassThrough>)
+                {
+#ifdef DL_KERNELS
+                    add_device_conv2d_dl_bias_perlayer_quantization_int8_instances(op_ptrs);
+#endif
+                    add_device_conv2d_xdl_bias_perlayer_quantization_int8_instances(op_ptrs);
+                }
+                else if constexpr(is_same_v<Activation, Relu>)
+                {
+#ifdef DL_KERNELS
+                    add_device_conv2d_dl_bias_relu_perlayer_quantization_int8_instances(op_ptrs);
+#endif
+                    add_device_conv2d_xdl_bias_relu_perlayer_quantization_int8_instances(op_ptrs);
+                }
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+// non-piecewise activation function
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename DsLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename DsDataType,
+          typename OutDataType,
+          typename Activation>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    DsLayout,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    DsDataType,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    Add_Mul_Activation_Mul_Clamp<Activation>>>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                   InLayout,
+                                                   WeiLayout,
+                                                   DsLayout,
+                                                   OutLayout,
+                                                   InDataType,
+                                                   WeiDataType,
+                                                   DsDataType,
+                                                   OutDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Add_Mul_Activation_Mul_Clamp<Activation>>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
+                     is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_Tuple> &&
+                     is_same_v<OutLayout, NHWGK>)
+        {
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<DsDataType, I32_Tuple> && is_same_v<OutDataType, int8_t>)
+            {
+                if constexpr(is_same_v<Activation, TanH>)
+                {
+#ifdef DL_KERNELS
+                    add_device_conv2d_dl_bias_tanh_perlayer_quantization_int8_instances(op_ptrs);
+#endif
+                    add_device_conv2d_xdl_bias_tanh_perlayer_quantization_int8_instances(op_ptrs);
+                }
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp
new file mode 100644
index 0000000000..1845cb8a95
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp
@@ -0,0 +1,154 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef DL_KERNELS
+// grouped conv2d forward, NHWGC/GKYXC/NHWGK
+void add_device_conv2d_dl_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Activation_Mul2_Clamp<PassThrough>>>>&
+        instances);
+
+void add_device_conv2d_dl_relu_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Activation_Mul2_Clamp<Relu>>>>&
+        instances);
+#endif
+void add_device_conv2d_xdl_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Activation_Mul2_Clamp<PassThrough>>>>&
+        instances);
+
+void add_device_conv2d_xdl_relu_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Activation_Mul2_Clamp<Relu>>>>&
+        instances);
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename DsLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename DsDataType,
+          typename OutDataType,
+          typename Activation>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    DsLayout,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    DsDataType,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    Activation_Mul2_Clamp<Activation>>>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                   InLayout,
+                                                   WeiLayout,
+                                                   GK_Tuple,
+                                                   OutLayout,
+                                                   InDataType,
+                                                   WeiDataType,
+                                                   F32_Tuple,
+                                                   OutDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Activation_Mul2_Clamp<Activation>>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
+                     is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_Tuple> &&
+                     is_same_v<OutLayout, NHWGK>)
+        {
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<OutDataType, int8_t>)
+            {
+                if constexpr(is_same_v<Activation, PassThrough>)
+                {
+#ifdef DL_KERNELS
+                    add_device_conv2d_dl_perchannel_quantization_int8_instances(op_ptrs);
+#endif
+                    add_device_conv2d_xdl_perchannel_quantization_int8_instances(op_ptrs);
+                }
+                else if constexpr(is_same_v<Activation, Relu>)
+                {
+#ifdef DL_KERNELS
+                    add_device_conv2d_dl_relu_perchannel_quantization_int8_instances(op_ptrs);
+#endif
+                    add_device_conv2d_xdl_relu_perchannel_quantization_int8_instances(op_ptrs);
+                }
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp
new file mode 100644
index 0000000000..c2f41de783
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp
@@ -0,0 +1,151 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef DL_KERNELS
+// grouped conv2d forward, NHWGC/GKYXC/NHWGK
+void add_device_conv2d_dl_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+
+void add_device_conv2d_dl_relu_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Activation_Mul_Clamp<Relu>>>>&
+        instances);
+#endif
+void add_device_conv2d_xdl_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Activation_Mul_Clamp<PassThrough>>>>&
+        instances);
+
+void add_device_conv2d_xdl_relu_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Activation_Mul_Clamp<Relu>>>>&
+        instances);
+
+template <ck::index_t NumDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename Activation>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<
+    NumDimSpatial,
+    InLayout,
+    WeiLayout,
+    Empty_Tuple,
+    OutLayout,
+    InDataType,
+    WeiDataType,
+    Empty_Tuple,
+    OutDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    Activation_Mul_Clamp<Activation>>>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleD<NumDimSpatial,
+                                                   InLayout,
+                                                   WeiLayout,
+                                                   Empty_Tuple,
+                                                   OutLayout,
+                                                   InDataType,
+                                                   WeiDataType,
+                                                   Empty_Tuple,
+                                                   OutDataType,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Activation_Mul_Clamp<Activation>>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
+                     is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
+        {
+            if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                         is_same_v<OutDataType, int8_t>)
+            {
+                if constexpr(is_same_v<Activation, PassThrough>)
+                {
+#ifdef DL_KERNELS
+                    add_device_conv2d_dl_perlayer_quantization_int8_instances(op_ptrs);
+#endif
+                    add_device_conv2d_xdl_perlayer_quantization_int8_instances(op_ptrs);
+                }
+                else if constexpr(is_same_v<Activation, Relu>)
+                {
+#ifdef DL_KERNELS
+                    add_device_conv2d_dl_relu_perlayer_quantization_int8_instances(op_ptrs);
+#endif
+                    add_device_conv2d_xdl_relu_perlayer_quantization_int8_instances(op_ptrs);
+                }
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp
new file mode 100644
index 0000000000..9930b1a6fa
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp"
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp
new file mode 100644
index 0000000000..f6fddc74fb
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp
@@ -0,0 +1,131 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using reduce_configuration_1_instances_blockwise = std::tuple<
+    // clang-format off
+    // BlockSize | MThreadClusterSize | KThreadClusterSize
+    ReductionConfiguration_1<256, 128, 2>,
+    ReductionConfiguration_1<256, 64, 4>,
+    ReductionConfiguration_1<256, 32, 8>,
+    ReductionConfiguration_1<256, 16, 16>,
+    ReductionConfiguration_1<256, 8, 32>,
+    ReductionConfiguration_1<256, 4, 64>,
+    ReductionConfiguration_1<256, 2, 128>,
+    ReductionConfiguration_1<256, 1, 256>
+    // clang-format on
+    >;
+
+#ifdef QUICK_REDUCE_TEST
+using reduce_configuration_2_instances_blockwise = std::tuple<
+    // clang-format off
+    // InSrcVectorDim | InSrcVectorSize | OutDstVectorSize | MThreadSliceSize | KThreadSliceSize
+    ReductionConfiguration_2<0, 2, 2, 2, 1>,
+    ReductionConfiguration_2<0, 1, 1, 2, 1>,
+    ReductionConfiguration_2<1, 2, 1, 1, 2>,
+    ReductionConfiguration_2<0, 1, 1, 3, 1>,
+    ReductionConfiguration_2<1, 1, 1, 1, 3>
+    // clang-format on
+    >;
+#else
+using reduce_configuration_2_instances_blockwise = std::tuple<
+    // clang-format off
+    // InSrcVectorDim | InSrcVectorSize | OutDstVectorSize | MThreadSliceSize | KThreadSliceSize
+    ReductionConfiguration_2<0, 4, 4, 8, 1>,
+    ReductionConfiguration_2<0, 4, 4, 4, 1>,
+    ReductionConfiguration_2<0, 2, 2, 2, 1>,
+
+    ReductionConfiguration_2<1, 4, 1, 1, 8>,
+    ReductionConfiguration_2<1, 4, 1, 1, 4>,
+    ReductionConfiguration_2<1, 2, 1, 1, 2>,
+
+    // special instances
+    ReductionConfiguration_2<0, 1, 1, 3, 1>,
+    ReductionConfiguration_2<0, 1, 1, 5, 1>,
+    ReductionConfiguration_2<0, 1, 1, 7, 1>,
+    ReductionConfiguration_2<0, 1, 1, 11, 1>,
+
+    ReductionConfiguration_2<1, 1, 1, 1, 3>,
+    ReductionConfiguration_2<1, 1, 1, 1, 5>,
+    ReductionConfiguration_2<1, 1, 1, 1, 7>,
+    ReductionConfiguration_2<1, 1, 1, 1, 11>
+    // clang-format on
+    >;
+#endif
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          int Rank,
+          int NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOp,
+          typename AccElementwiseOp,
+          bool PropagateNan,
+          bool OutputIndex>
+void add_device_reduce_instance_blockwise(
+    std::vector<DeviceReducePtr<InDataType,
+                                AccDataType,
+                                OutDataType,
+                                Rank,
+                                NumReduceDim,
+                                ReduceOperation,
+                                InElementwiseOp,
+                                AccElementwiseOp,
+                                PropagateNan,
+                                OutputIndex>>& device_op_instances)
+{
+    static_for<0, std::tuple_size<reduce_configuration_1_instances_blockwise>::value, 1>{}(
+        [&](auto i) {
+            using cfg1 = remove_cvref_t<decltype(std::get<i.value>(
+                reduce_configuration_1_instances_blockwise{}))>;
+
+            static_for<0, std::tuple_size<reduce_configuration_2_instances_blockwise>::value, 1>{}(
+                [&](auto j) {
+                    using cfg2 = remove_cvref_t<decltype(std::get<j.value>(
+                        reduce_configuration_2_instances_blockwise{}))>;
+
+                    using ReduceOpInstance =
+                        DeviceReduceMultiBlock<InDataType,
+                                               AccDataType,
+                                               OutDataType,
+                                               Rank,
+                                               NumReduceDim,
+                                               ReduceOperation,
+                                               InElementwiseOp,
+                                               AccElementwiseOp,
+                                               InMemoryDataOperationEnum::Set,
+                                               PropagateNan,
+                                               OutputIndex,
+                                               false, // HaveIndexInputIfOutputIndex
+                                               cfg1::BlockSize_,
+                                               cfg1::MThreadClusterSize_,
+                                               cfg1::KThreadClusterSize_,
+                                               cfg2::MThreadSliceSize_,
+                                               cfg2::KThreadSliceSize_,
+                                               cfg2::InSrcVectorDim_,
+                                               cfg2::InSrcVectorSize_,
+                                               cfg2::OutDstVectorSize_>;
+
+                    device_op_instances.push_back(
+                        std::make_unique<ReduceOpInstance>(ReduceOpInstance{}));
+                });
+        });
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp
new file mode 100644
index 0000000000..4dd5569cee
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp
new file mode 100644
index 0000000000..d52310a3f3
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp
new file mode 100644
index 0000000000..025500764e
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp
new file mode 100644
index 0000000000..2314e94980
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp
new file mode 100644
index 0000000000..5c2bff16cc
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp
new file mode 100644
index 0000000000..a1279eeccb
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp
new file mode 100644
index 0000000000..aff6a2542b
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp
new file mode 100644
index 0000000000..be8da2243f
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp
new file mode 100644
index 0000000000..652984ae22
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp
new file mode 100644
index 0000000000..be60d1b32d
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp
new file mode 100644
index 0000000000..27e3aa53e3
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp
new file mode 100644
index 0000000000..f7f4870a97
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp
new file mode 100644
index 0000000000..790b5a92ad
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp
new file mode 100644
index 0000000000..ec3bc852e8
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp
new file mode 100644
index 0000000000..8c0c065670
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp
new file mode 100644
index 0000000000..8631495df1
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp
new file mode 100644
index 0000000000..59849c2d45
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp
new file mode 100644
index 0000000000..33cbea85ef
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp
new file mode 100644
index 0000000000..3861829575
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp
new file mode 100644
index 0000000000..f7c05b20a7
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp
new file mode 100644
index 0000000000..2dfecc6b07
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F32, F64, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp
new file mode 100644
index 0000000000..a687938965
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp
new file mode 100644
index 0000000000..cf9c268a64
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp
new file mode 100644
index 0000000000..aeb578f948
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp
new file mode 100644
index 0000000000..73480262c6
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp
new file mode 100644
index 0000000000..74293553e1
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp
new file mode 100644
index 0000000000..8a91f76b34
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp
new file mode 100644
index 0000000000..0ff2c30f34
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp
new file mode 100644
index 0000000000..932299008e
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp
new file mode 100644
index 0000000000..c902e80a2c
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp
new file mode 100644
index 0000000000..e7a1369277
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp
new file mode 100644
index 0000000000..2d1ab69fc7
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp
new file mode 100644
index 0000000000..9f782c11b6
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <int BlockSize, int MThreadClusterSize, int KThreadClusterSize>
+struct ReductionConfiguration_1
+{
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize, "Invalid Configuration!");
+
+    static constexpr int BlockSize_          = BlockSize;
+    static constexpr int MThreadClusterSize_ = MThreadClusterSize;
+    static constexpr int KThreadClusterSize_ = KThreadClusterSize;
+};
+
+template <int InSrcVectorDim,
+          int InSrcVectorSize,
+          int OutDstVectorSize,
+          int MThreadSliceSize,
+          int KThreadSliceSize>
+struct ReductionConfiguration_2
+{
+    static constexpr int InSrcVectorDim_   = InSrcVectorDim;
+    static constexpr int InSrcVectorSize_  = InSrcVectorSize;
+    static constexpr int OutDstVectorSize_ = OutDstVectorSize;
+    static constexpr int MThreadSliceSize_ = MThreadSliceSize;
+    static constexpr int KThreadSliceSize_ = KThreadSliceSize;
+};
+
+using ReduceAdd  = ck::reduce::Add;
+using ReduceMin  = ck::reduce::Min;
+using ReduceMax  = ck::reduce::Max;
+using ReduceAMax = ck::reduce::AMax;
+
+using UnarySquare = ck::tensor_operation::element_wise::UnarySquare;
+using UnarySqrt   = ck::tensor_operation::element_wise::UnarySqrt;
+using UnaryDivide = ck::tensor_operation::element_wise::UnaryDivide;
+using UnaryAbs    = ck::tensor_operation::element_wise::UnaryAbs;
+
+#define QUICK_REDUCE_TEST 1
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp
new file mode 100644
index 0000000000..5a65a2ff88
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp
@@ -0,0 +1,131 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using reduce_configuration_1_instances_multiblock_atomic_add = std::tuple<
+    // clang-format off
+    // BlockSize | MThreadClusterSize | KThreadClusterSize
+    ReductionConfiguration_1<256, 128, 2>,
+    ReductionConfiguration_1<256, 64, 4>,
+    ReductionConfiguration_1<256, 32, 8>,
+    ReductionConfiguration_1<256, 16, 16>,
+    ReductionConfiguration_1<256, 8, 32>,
+    ReductionConfiguration_1<256, 4, 64>,
+    ReductionConfiguration_1<256, 2, 128>,
+    ReductionConfiguration_1<256, 1, 256>
+    // clang-format on
+    >;
+
+#ifdef QUICK_REDUCE_TEST
+using reduce_configuration_2_instances_multiblock_atomic_add = std::tuple<
+    // clang-format off
+    // InSrcVectorDim | InSrcVectorSize | OutDstVectorSize | MThreadSliceSize | KThreadSliceSize
+    ReductionConfiguration_2<0, 2, 2, 2, 1>,
+    ReductionConfiguration_2<0, 1, 1, 2, 1>,
+    ReductionConfiguration_2<1, 2, 1, 1, 2>,
+    ReductionConfiguration_2<0, 1, 1, 3, 1>,
+    ReductionConfiguration_2<1, 1, 1, 1, 3>
+    // clang-format on
+    >;
+#else
+using reduce_configuration_2_instances_multiblock_atomic_add = std::tuple<
+    // clang-format off
+    // InSrcVectorDim | InSrcVectorSize | OutDstVectorSize | MThreadSliceSize | KThreadSliceSize
+    ReductionConfiguration_2<0, 4, 4, 8, 1>,
+    ReductionConfiguration_2<0, 4, 4, 4, 1>,
+    ReductionConfiguration_2<0, 2, 2, 2, 1>,
+
+    ReductionConfiguration_2<1, 4, 1, 1, 8>,
+    ReductionConfiguration_2<1, 4, 1, 1, 4>,
+    ReductionConfiguration_2<1, 2, 1, 1, 2>,
+
+    // special instances
+    ReductionConfiguration_2<0, 1, 1, 3, 1>,
+    ReductionConfiguration_2<0, 1, 1, 5, 1>,
+    ReductionConfiguration_2<0, 1, 1, 7, 1>,
+    ReductionConfiguration_2<0, 1, 1, 11, 1>,
+
+    ReductionConfiguration_2<1, 1, 1, 1, 3>,
+    ReductionConfiguration_2<1, 1, 1, 1, 5>,
+    ReductionConfiguration_2<1, 1, 1, 1, 7>,
+    ReductionConfiguration_2<1, 1, 1, 1, 11>
+    // clang-format on
+    >;
+#endif
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          int Rank,
+          int NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOp,
+          typename AccElementwiseOp,
+          bool PropagateNan,
+          bool OutputIndex>
+void add_device_reduce_instance_multiblock_atomic_add(
+    std::vector<DeviceReducePtr<InDataType,
+                                AccDataType,
+                                OutDataType,
+                                Rank,
+                                NumReduceDim,
+                                ReduceOperation,
+                                InElementwiseOp,
+                                AccElementwiseOp,
+                                PropagateNan,
+                                OutputIndex>>& device_op_instances)
+{
+    static_for<0,
+               std::tuple_size<reduce_configuration_1_instances_multiblock_atomic_add>::value,
+               1>{}([&](auto i) {
+        using cfg1 = remove_cvref_t<decltype(std::get<i.value>(
+            reduce_configuration_1_instances_multiblock_atomic_add{}))>;
+
+        static_for<0,
+                   std::tuple_size<reduce_configuration_2_instances_multiblock_atomic_add>::value,
+                   1>{}([&](auto j) {
+            using cfg2 = remove_cvref_t<decltype(std::get<j.value>(
+                reduce_configuration_2_instances_multiblock_atomic_add{}))>;
+
+            using ReduceOpInstance = DeviceReduceMultiBlock<InDataType,
+                                                            AccDataType,
+                                                            OutDataType,
+                                                            Rank,
+                                                            NumReduceDim,
+                                                            ReduceOperation,
+                                                            InElementwiseOp,
+                                                            AccElementwiseOp,
+                                                            InMemoryDataOperationEnum::AtomicAdd,
+                                                            PropagateNan,
+                                                            OutputIndex,
+                                                            false, // HaveIndexInputIfOutputIndex
+                                                            cfg1::BlockSize_,
+                                                            cfg1::MThreadClusterSize_,
+                                                            cfg1::KThreadClusterSize_,
+                                                            cfg2::MThreadSliceSize_,
+                                                            cfg2::KThreadSliceSize_,
+                                                            cfg2::InSrcVectorDim_,
+                                                            cfg2::InSrcVectorSize_,
+                                                            cfg2::OutDstVectorSize_>;
+
+            device_op_instances.push_back(std::make_unique<ReduceOpInstance>(ReduceOpInstance{}));
+        });
+    });
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp
new file mode 100644
index 0000000000..5814b84bb7
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp
new file mode 100644
index 0000000000..f1a609a6f4
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp
new file mode 100644
index 0000000000..489bef1f56
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp
new file mode 100644
index 0000000000..2507d82308
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp
new file mode 100644
index 0000000000..df4b7b7b49
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp
new file mode 100644
index 0000000000..2748b9b9d0
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp
new file mode 100644
index 0000000000..26230d5456
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp
new file mode 100644
index 0000000000..d5e1499aef
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp
new file mode 100644
index 0000000000..cbc8befb3d
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp
new file mode 100644
index 0000000000..574cc3dd3e
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp
new file mode 100644
index 0000000000..f3267c0d02
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp
@@ -0,0 +1,109 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+#ifdef QUICK_REDUCE_TEST
+using reduce_configuration_2_instances_threadwise = std::tuple<
+    // clang-format off
+    // InSrcVectorDim | InSrcVectorSize | OutDstVectorSize | MThreadSliceSize | KThreadSliceSize
+    ReductionConfiguration_2<0, 2, 2, 2, 1>,
+    ReductionConfiguration_2<0, 1, 1, 2, 1>,
+    ReductionConfiguration_2<1, 2, 1, 1, 2>,
+    ReductionConfiguration_2<0, 1, 1, 3, 1>,
+    ReductionConfiguration_2<1, 1, 1, 1, 3>
+    // clang-format on
+    >;
+#else
+using reduce_configuration_2_instances_threadwise = std::tuple<
+    // clang-format off
+    // InSrcVectorDim | InSrcVectorSize | OutDstVectorSize | MThreadSliceSize | KThreadSliceSize
+    ReductionConfiguration_2<0, 4, 4, 8, 1>,
+    ReductionConfiguration_2<0, 4, 4, 4, 1>,
+    ReductionConfiguration_2<0, 2, 2, 2, 1>,
+
+    ReductionConfiguration_2<1, 4, 1, 1, 8>,
+    ReductionConfiguration_2<1, 4, 1, 1, 4>,
+    ReductionConfiguration_2<1, 2, 1, 1, 2>,
+
+    // special instances
+    ReductionConfiguration_2<0, 1, 1, 3, 1>,
+    ReductionConfiguration_2<0, 1, 1, 5, 1>,
+    ReductionConfiguration_2<0, 1, 1, 7, 1>,
+    ReductionConfiguration_2<0, 1, 1, 11, 1>,
+
+    ReductionConfiguration_2<1, 1, 1, 1, 3>,
+    ReductionConfiguration_2<1, 1, 1, 1, 5>,
+    ReductionConfiguration_2<1, 1, 1, 1, 7>,
+    ReductionConfiguration_2<1, 1, 1, 1, 11>
+    // clang-format on
+    >;
+#endif
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          int Rank,
+          int NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOp,
+          typename AccElementwiseOp,
+          bool PropagateNan,
+          bool OutputIndex>
+void add_device_reduce_instance_threadwise(
+    std::vector<DeviceReducePtr<InDataType,
+                                AccDataType,
+                                OutDataType,
+                                Rank,
+                                NumReduceDim,
+                                ReduceOperation,
+                                InElementwiseOp,
+                                AccElementwiseOp,
+                                PropagateNan,
+                                OutputIndex>>& device_op_instances)
+{
+    using cfg1 = ReductionConfiguration_1<256, 256, 1>;
+
+    static_for<0, std::tuple_size<reduce_configuration_2_instances_threadwise>::value, 1>{}(
+        [&](auto j) {
+            using cfg2 = remove_cvref_t<decltype(std::get<j.value>(
+                reduce_configuration_2_instances_threadwise{}))>;
+
+            using ReduceOpInstance = DeviceReduceThreadWise<InDataType,
+                                                            AccDataType,
+                                                            OutDataType,
+                                                            Rank,
+                                                            NumReduceDim,
+                                                            ReduceOperation,
+                                                            InElementwiseOp,
+                                                            AccElementwiseOp,
+                                                            PropagateNan,
+                                                            OutputIndex,
+                                                            false,
+                                                            false, // HaveIndexInputIfOutputIndex
+                                                            cfg1::BlockSize_,
+                                                            cfg2::MThreadSliceSize_,
+                                                            cfg2::KThreadSliceSize_,
+                                                            cfg2::InSrcVectorDim_,
+                                                            cfg2::InSrcVectorSize_,
+                                                            cfg2::OutDstVectorSize_>;
+
+            device_op_instances.push_back(std::make_unique<ReduceOpInstance>(ReduceOpInstance{}));
+        });
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp
new file mode 100644
index 0000000000..6c50d222ec
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp
new file mode 100644
index 0000000000..5e91f53b0f
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp
new file mode 100644
index 0000000000..1718031560
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp
new file mode 100644
index 0000000000..03bff5912d
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp
new file mode 100644
index 0000000000..676af19336
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp
new file mode 100644
index 0000000000..c67c0a41f8
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp
new file mode 100644
index 0000000000..c9058a9b1d
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp
new file mode 100644
index 0000000000..1fbd3c6e7d
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp
new file mode 100644
index 0000000000..89496a4ed1
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp
new file mode 100644
index 0000000000..fb79ce8adf
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp
new file mode 100644
index 0000000000..7915bd2630
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp
new file mode 100644
index 0000000000..d769fe8154
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp
new file mode 100644
index 0000000000..49938d4434
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp
new file mode 100644
index 0000000000..810249994f
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp
new file mode 100644
index 0000000000..6640ffec70
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp
new file mode 100644
index 0000000000..441bea275c
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp
new file mode 100644
index 0000000000..05912e1e96
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp
new file mode 100644
index 0000000000..b0e16cfeaa
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp
new file mode 100644
index 0000000000..0ee0f7d0fe
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp
new file mode 100644
index 0000000000..feb9d99eba
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp
new file mode 100644
index 0000000000..a9f0d77cb8
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F64, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp
new file mode 100644
index 0000000000..bd1f72250a
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp
new file mode 100644
index 0000000000..60189fd304
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp
new file mode 100644
index 0000000000..b5aa9ce613
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp
new file mode 100644
index 0000000000..0e20a17ff9
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp
new file mode 100644
index 0000000000..c8c1cd7e80
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp
new file mode 100644
index 0000000000..3e71b44673
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp
new file mode 100644
index 0000000000..c8286e56d1
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp
new file mode 100644
index 0000000000..36fbc52f11
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_threadwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp
new file mode 100644
index 0000000000..0900cffa35
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp
new file mode 100644
index 0000000000..726b56066b
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp
new file mode 100644
index 0000000000..0ae9838607
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp
@@ -0,0 +1,31 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp
new file mode 100644
index 0000000000..a3a39b7ccc
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp
@@ -0,0 +1,117 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp"
+#include "ck/utility/reduction_operator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          index_t Rank,
+          index_t NumReduceDim,
+          typename ReduceOperation,
+          typename InElementwiseOp,
+          typename AccElementwiseOp,
+          bool PropagateNan,
+          bool OutputIndex>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceReduce<InDataType,
+                                                                                 AccDataType,
+                                                                                 OutDataType,
+                                                                                 Rank,
+                                                                                 NumReduceDim,
+                                                                                 ReduceOperation,
+                                                                                 InElementwiseOp,
+                                                                                 AccElementwiseOp,
+                                                                                 PropagateNan,
+                                                                                 OutputIndex>>
+{
+    using DeviceOp = DeviceReduce<InDataType,
+                                  AccDataType,
+                                  OutDataType,
+                                  Rank,
+                                  NumReduceDim,
+                                  ReduceOperation,
+                                  InElementwiseOp,
+                                  AccElementwiseOp,
+                                  PropagateNan,
+                                  OutputIndex>;
+
+    using DeviceOpPtr = DeviceReducePtr<InDataType,
+                                        AccDataType,
+                                        OutDataType,
+                                        Rank,
+                                        NumReduceDim,
+                                        ReduceOperation,
+                                        InElementwiseOp,
+                                        AccElementwiseOp,
+                                        PropagateNan,
+                                        OutputIndex>;
+
+    static auto GetInstances()
+    {
+        std::vector<DeviceOpPtr> op_ptrs;
+
+        constexpr bool out_support_atomic_add =
+            ck::reduce::InMemoryDataOperationSupportedOnDataType<
+                InMemoryDataOperationEnum::AtomicAdd,
+                OutDataType>::value;
+        constexpr bool op_support_atomic_add =
+            std::is_same<ReduceOperation, ReduceAdd>::value &&
+            (std::is_same<AccElementwiseOp, PassThrough>::value ||
+             std::is_same<AccElementwiseOp, UnaryDivide>::value);
+        constexpr bool use_atomic_add = (out_support_atomic_add && op_support_atomic_add);
+
+        add_device_reduce_instance_threadwise<InDataType,
+                                              AccDataType,
+                                              OutDataType,
+                                              Rank,
+                                              NumReduceDim,
+                                              ReduceOperation,
+                                              InElementwiseOp,
+                                              AccElementwiseOp,
+                                              PropagateNan,
+                                              OutputIndex>(op_ptrs);
+
+        add_device_reduce_instance_blockwise<InDataType,
+                                             AccDataType,
+                                             OutDataType,
+                                             Rank,
+                                             NumReduceDim,
+                                             ReduceOperation,
+                                             InElementwiseOp,
+                                             AccElementwiseOp,
+                                             PropagateNan,
+                                             OutputIndex>(op_ptrs);
+
+        if constexpr(use_atomic_add)
+        {
+            add_device_reduce_instance_multiblock_atomic_add<InDataType,
+                                                             AccDataType,
+                                                             OutDataType,
+                                                             Rank,
+                                                             NumReduceDim,
+                                                             ReduceOperation,
+                                                             InElementwiseOp,
+                                                             AccElementwiseOp,
+                                                             PropagateNan,
+                                                             OutputIndex>(op_ptrs);
+        };
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp
new file mode 100644
index 0000000000..0a25f8e6ad
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <type_traits>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceSoftmax<InDataType,
+                                                                                  AccDataType,
+                                                                                  OutDataType,
+                                                                                  PassThrough,
+                                                                                  PassThrough,
+                                                                                  Rank,
+                                                                                  NumReduceDim>>
+{
+    using DeviceOp = DeviceSoftmax<InDataType,
+                                   AccDataType,
+                                   OutDataType,
+                                   PassThrough,
+                                   PassThrough,
+                                   Rank,
+                                   NumReduceDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+#ifdef CK_ENABLE_FP16
+        if constexpr(std::is_same_v<InDataType, F16> && std::is_same_v<AccDataType, F32> &&
+                     std::is_same_v<OutDataType, F16>)
+        {
+            if constexpr(Rank == 3)
+            {
+                if constexpr(NumReduceDim == 1)
+                    add_device_softmax_f16_f16_rank3_reduce1_instances(op_ptrs);
+                else if constexpr(NumReduceDim == 2)
+                    add_device_softmax_f16_f16_rank3_reduce2_instances(op_ptrs);
+                else if constexpr(NumReduceDim == 3)
+                    add_device_softmax_f16_f16_rank3_reduce3_instances(op_ptrs);
+            }
+            else if constexpr(Rank == 4)
+            {
+                if constexpr(NumReduceDim == 1)
+                    add_device_softmax_f16_f16_rank4_reduce1_instances(op_ptrs);
+                else if constexpr(NumReduceDim == 2)
+                    add_device_softmax_f16_f16_rank4_reduce2_instances(op_ptrs);
+                else if constexpr(NumReduceDim == 3)
+                    add_device_softmax_f16_f16_rank4_reduce3_instances(op_ptrs);
+                else if constexpr(NumReduceDim == 4)
+                    add_device_softmax_f16_f16_rank4_reduce4_instances(op_ptrs);
+            }
+        }
+#endif
+#ifdef CK_ENABLE_FP32
+        if constexpr(std::is_same_v<InDataType, F32> && std::is_same_v<AccDataType, F32> &&
+                     std::is_same_v<OutDataType, F32>)
+        {
+            if constexpr(Rank == 3)
+            {
+                if constexpr(NumReduceDim == 1)
+                    add_device_softmax_f32_f32_rank3_reduce1_instances(op_ptrs);
+                else if constexpr(NumReduceDim == 2)
+                    add_device_softmax_f32_f32_rank3_reduce2_instances(op_ptrs);
+                else if constexpr(NumReduceDim == 3)
+                    add_device_softmax_f32_f32_rank3_reduce3_instances(op_ptrs);
+            }
+            else if constexpr(Rank == 4)
+            {
+                if constexpr(NumReduceDim == 1)
+                    add_device_softmax_f32_f32_rank4_reduce1_instances(op_ptrs);
+                else if constexpr(NumReduceDim == 2)
+                    add_device_softmax_f32_f32_rank4_reduce2_instances(op_ptrs);
+                else if constexpr(NumReduceDim == 3)
+                    add_device_softmax_f32_f32_rank4_reduce3_instances(op_ptrs);
+                else if constexpr(NumReduceDim == 4)
+                    add_device_softmax_f32_f32_rank4_reduce4_instances(op_ptrs);
+            }
+        }
+#endif
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp
new file mode 100644
index 0000000000..3fd2bd089e
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3, 1>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp
new file mode 100644
index 0000000000..210fdc0a58
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3, 2>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp
new file mode 100644
index 0000000000..894fb034d0
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp
new file mode 100644
index 0000000000..708ef0ce13
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 1>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp
new file mode 100644
index 0000000000..6754e5ceff
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 2>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp
new file mode 100644
index 0000000000..5e111176e1
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp
new file mode 100644
index 0000000000..a3cecb32f8
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce4_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp
new file mode 100644
index 0000000000..8c0782daa5
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <index_t Rank, index_t Reduce>
+using device_softmax_f16_f16_instances = std::tuple<
+    // clang-format off
+    //                InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               1,              1>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  4,                 64,                1,                8,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,                8,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               16,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               32,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,                8,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               16,              1,               8,              8>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               32,              1,               8,              8>,
+    // Reduction on middle dimensions
+    // InSrcVectorDim is 0 since we want to coalesce reads on M dimension
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                8,                4,              0,               1,              1>,
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                8,                4,              0,               8,              4>
+    // clang-format on
+    >;
+
+template <index_t Rank, index_t Reduce>
+using device_softmax_f16_f16_generic_instance = std::tuple<
+    // clang-format off
+    DeviceSoftmaxImpl<       F16,         F32,         F16,     PassThrough,      PassThrough, Rank,       Reduce,        64,                  8,                  8,                1,                1,              1,               1,              1>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp
new file mode 100644
index 0000000000..4cc4690253
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank3_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 1>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp
new file mode 100644
index 0000000000..65724d7888
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank3_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 2>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp
new file mode 100644
index 0000000000..13bd45598e
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank3_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp
new file mode 100644
index 0000000000..d58b424ee9
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank4_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 1>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp
new file mode 100644
index 0000000000..378e45eeb7
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank4_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 2>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp
new file mode 100644
index 0000000000..293df08c7e
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank4_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp
new file mode 100644
index 0000000000..e503a9fec1
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank4_reduce4_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp
new file mode 100644
index 0000000000..90c5ddc8a0
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <index_t Rank, index_t Reduce>
+using device_softmax_f32_f32_instances = std::tuple<
+    // clang-format off
+    //                InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               1,               1>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  4,                 64,                1,                8,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,                8,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               16,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               32,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,                8,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               16,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               32,              1,               4,               4>,
+    // Reduction on middle dimensions
+    // InSrcVectorDim is 0 since we want to coalesce reads on M dimension
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                8,                4,              0,               1,               1>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                8,                4,              0,               4,               4>
+    // clang-format on
+    >;
+
+template <index_t Rank, index_t Reduce>
+using device_softmax_f32_f32_generic_instance = std::tuple<
+    // clang-format off
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,        64,                  8,                  8,                1,                1,              1,               1,               1>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp
new file mode 100644
index 0000000000..10f99acb8d
--- /dev/null
+++ b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp
@@ -0,0 +1,19 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp"
diff --git a/library/include/ck/library/utility/algorithm.hpp b/library/include/ck/library/utility/algorithm.hpp
new file mode 100644
index 0000000000..57136f8a2a
--- /dev/null
+++ b/library/include/ck/library/utility/algorithm.hpp
@@ -0,0 +1,43 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <iterator>
+#include <type_traits>
+#include <utility>
+
+namespace ck {
+namespace ranges {
+template <typename InputRange, typename OutputIterator>
+auto copy(InputRange&& range, OutputIterator iter)
+    -> decltype(std::copy(std::begin(std::forward<InputRange>(range)),
+                          std::end(std::forward<InputRange>(range)),
+                          iter))
+{
+    return std::copy(std::begin(std::forward<InputRange>(range)),
+                     std::end(std::forward<InputRange>(range)),
+                     iter);
+}
+
+template <typename T, typename OutputRange>
+auto fill(OutputRange&& range, const T& init)
+    -> std::void_t<decltype(std::fill(std::begin(std::forward<OutputRange>(range)),
+                                      std::end(std::forward<OutputRange>(range)),
+                                      init))>
+{
+    std::fill(std::begin(std::forward<OutputRange>(range)),
+              std::end(std::forward<OutputRange>(range)),
+              init);
+}
+
+template <typename InputRange, typename OutputIterator, typename UnaryOperation>
+auto transform(InputRange&& range, OutputIterator iter, UnaryOperation unary_op)
+    -> decltype(std::transform(std::begin(range), std::end(range), iter, unary_op))
+{
+    return std::transform(std::begin(range), std::end(range), iter, unary_op);
+}
+
+} // namespace ranges
+} // namespace ck
diff --git a/library/include/ck/library/utility/check_err.hpp b/library/include/ck/library/utility/check_err.hpp
new file mode 100644
index 0000000000..8a72631376
--- /dev/null
+++ b/library/include/ck/library/utility/check_err.hpp
@@ -0,0 +1,234 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <cmath>
+#include <cstdlib>
+#include <iostream>
+#include <iomanip>
+#include <iterator>
+#include <limits>
+#include <type_traits>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/type.hpp"
+#include "ck/host_utility/io.hpp"
+
+#include "ck/library/utility/ranges.hpp"
+
+namespace ck {
+namespace utils {
+
+template <typename Range, typename RefRange>
+typename std::enable_if<
+    std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
+        std::is_floating_point_v<ranges::range_value_t<Range>> &&
+        !std::is_same_v<ranges::range_value_t<Range>, half_t>,
+    bool>::type
+check_err(const Range& out,
+          const RefRange& ref,
+          const std::string& msg = "Error: Incorrect results!",
+          double rtol            = 1e-5,
+          double atol            = 3e-6)
+{
+    if(out.size() != ref.size())
+    {
+        std::cerr << msg << " out.size() != ref.size(), :" << out.size() << " != " << ref.size()
+                  << std::endl;
+        return false;
+    }
+
+    bool res{true};
+    int err_count  = 0;
+    double err     = 0;
+    double max_err = std::numeric_limits<double>::min();
+    for(std::size_t i = 0; i < ref.size(); ++i)
+    {
+        const double o = *std::next(std::begin(out), i);
+        const double r = *std::next(std::begin(ref), i);
+        err            = std::abs(o - r);
+        if(err > atol + rtol * std::abs(r) || !std::isfinite(o) || !std::isfinite(r))
+        {
+            max_err = err > max_err ? err : max_err;
+            err_count++;
+            if(err_count < 5)
+            {
+                std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
+                          << "] != ref[" << i << "]: " << o << " != " << r << std::endl;
+            }
+            res = false;
+        }
+    }
+    if(!res)
+    {
+        const float error_percent =
+            static_cast<float>(err_count) / static_cast<float>(out.size()) * 100.f;
+        std::cerr << "max err: " << max_err;
+        std::cerr << ", number of errors: " << err_count;
+        std::cerr << ", " << error_percent << "% wrong values" << std::endl;
+    }
+    return res;
+}
+
+template <typename Range, typename RefRange>
+typename std::enable_if<
+    std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
+        std::is_same_v<ranges::range_value_t<Range>, bhalf_t>,
+    bool>::type
+check_err(const Range& out,
+          const RefRange& ref,
+          const std::string& msg = "Error: Incorrect results!",
+          double rtol            = 1e-3,
+          double atol            = 1e-3)
+{
+    if(out.size() != ref.size())
+    {
+        std::cerr << msg << " out.size() != ref.size(), :" << out.size() << " != " << ref.size()
+                  << std::endl;
+        return false;
+    }
+
+    bool res{true};
+    int err_count = 0;
+    double err    = 0;
+    // TODO: This is a hack. We should have proper specialization for bhalf_t data type.
+    double max_err = std::numeric_limits<float>::min();
+    for(std::size_t i = 0; i < ref.size(); ++i)
+    {
+        const double o = type_convert<float>(*std::next(std::begin(out), i));
+        const double r = type_convert<float>(*std::next(std::begin(ref), i));
+        err            = std::abs(o - r);
+        if(err > atol + rtol * std::abs(r) || !std::isfinite(o) || !std::isfinite(r))
+        {
+            max_err = err > max_err ? err : max_err;
+            err_count++;
+            if(err_count < 5)
+            {
+                std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
+                          << "] != ref[" << i << "]: " << o << " != " << r << std::endl;
+            }
+            res = false;
+        }
+    }
+    if(!res)
+    {
+        const float error_percent =
+            static_cast<float>(err_count) / static_cast<float>(out.size()) * 100.f;
+        std::cerr << "max err: " << max_err;
+        std::cerr << ", number of errors: " << err_count;
+        std::cerr << ", " << error_percent << "% wrong values" << std::endl;
+    }
+    return res;
+}
+
+template <typename Range, typename RefRange>
+typename std::enable_if<
+    std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
+        std::is_same_v<ranges::range_value_t<Range>, half_t>,
+    bool>::type
+check_err(const Range& out,
+          const RefRange& ref,
+          const std::string& msg = "Error: Incorrect results!",
+          double rtol            = 1e-3,
+          double atol            = 1e-3)
+{
+    if(out.size() != ref.size())
+    {
+        std::cerr << msg << " out.size() != ref.size(), :" << out.size() << " != " << ref.size()
+                  << std::endl;
+        return false;
+    }
+
+    bool res{true};
+    int err_count  = 0;
+    double err     = 0;
+    double max_err = std::numeric_limits<ranges::range_value_t<Range>>::min();
+    for(std::size_t i = 0; i < ref.size(); ++i)
+    {
+        const double o = type_convert<float>(*std::next(std::begin(out), i));
+        const double r = type_convert<float>(*std::next(std::begin(ref), i));
+        err            = std::abs(o - r);
+        if(err > atol + rtol * std::abs(r) || !std::isfinite(o) || !std::isfinite(r))
+        {
+            max_err = err > max_err ? err : max_err;
+            err_count++;
+            if(err_count < 5)
+            {
+                std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
+                          << "] != ref[" << i << "]: " << o << " != " << r << std::endl;
+            }
+            res = false;
+        }
+    }
+    if(!res)
+    {
+        const float error_percent =
+            static_cast<float>(err_count) / static_cast<float>(out.size()) * 100.f;
+        std::cerr << "max err: " << max_err;
+        std::cerr << ", number of errors: " << err_count;
+        std::cerr << ", " << error_percent << "% wrong values" << std::endl;
+    }
+    return res;
+}
+
+template <typename Range, typename RefRange>
+std::enable_if_t<(std::is_same_v<ranges::range_value_t<Range>, ranges::range_value_t<RefRange>> &&
+                  std::is_integral_v<ranges::range_value_t<Range>> &&
+                  !std::is_same_v<ranges::range_value_t<Range>, bhalf_t>)
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+                     || std::is_same_v<ranges::range_value_t<Range>, int4_t>
+#endif
+                 ,
+                 bool>
+check_err(const Range& out,
+          const RefRange& ref,
+          const std::string& msg = "Error: Incorrect results!",
+          double                 = 0,
+          double atol            = 0)
+{
+    if(out.size() != ref.size())
+    {
+        std::cerr << msg << " out.size() != ref.size(), :" << out.size() << " != " << ref.size()
+                  << std::endl;
+        return false;
+    }
+
+    bool res{true};
+    int err_count   = 0;
+    int64_t err     = 0;
+    int64_t max_err = std::numeric_limits<int64_t>::min();
+    for(std::size_t i = 0; i < ref.size(); ++i)
+    {
+        const int64_t o = *std::next(std::begin(out), i);
+        const int64_t r = *std::next(std::begin(ref), i);
+        err             = std::abs(o - r);
+
+        if(err > atol)
+        {
+            max_err = err > max_err ? err : max_err;
+            err_count++;
+            if(err_count < 5)
+            {
+                std::cerr << msg << " out[" << i << "] != ref[" << i << "]: " << o << " != " << r
+                          << std::endl;
+            }
+            res = false;
+        }
+    }
+    if(!res)
+    {
+        const float error_percent =
+            static_cast<float>(err_count) / static_cast<float>(out.size()) * 100.f;
+        std::cerr << "max err: " << max_err;
+        std::cerr << ", number of errors: " << err_count;
+        std::cerr << ", " << error_percent << "% wrong values" << std::endl;
+    }
+    return res;
+}
+
+} // namespace utils
+} // namespace ck
diff --git a/library/include/ck/library/utility/conv_common.hpp b/library/include/ck/library/utility/conv_common.hpp
new file mode 100644
index 0000000000..085454f42d
--- /dev/null
+++ b/library/include/ck/library/utility/conv_common.hpp
@@ -0,0 +1,77 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+
+template <typename... InDesc,
+          typename... WeiDesc,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename LeftPads,
+          typename RightPads>
+constexpr auto get_convolution_output_default_4d_tensor_descriptor(
+    const ck::TensorDescriptor<InDesc...>& in_desc,
+    const ck::TensorDescriptor<WeiDesc...>& wei_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations conv_dilations,
+    const LeftPads& left_pads,
+    const RightPads& right_pads)
+{
+    using namespace ck;
+
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    assert(in_desc.GetNumOfDimension() == 4);
+    assert(wei_desc.GetNumOfDimension() == 4);
+    assert(in_desc.GetLength(I1) == wei_desc.GetLength(I1));
+
+    const auto N  = in_desc.GetLength(I0);
+    const auto Hi = in_desc.GetLength(I2);
+    const auto Wi = in_desc.GetLength(I3);
+
+    const auto K = wei_desc.GetLength(I0);
+    const auto Y = wei_desc.GetLength(I2);
+    const auto X = wei_desc.GetLength(I3);
+
+    const auto LeftPadH = left_pads[I0];
+    const auto LeftPadW = left_pads[I1];
+
+    const auto RightPadH = right_pads[I0];
+    const auto RightPadW = right_pads[I1];
+
+    const auto YEff = (Y - I1) * conv_dilations[I0] + I1;
+    const auto XEff = (X - I1) * conv_dilations[I1] + I1;
+
+    const auto Ho = (Hi + LeftPadH + RightPadH - YEff) / conv_strides[I0] + I1;
+    const auto Wo = (Wi + LeftPadW + RightPadW - XEff) / conv_strides[I1] + I1;
+
+    return make_naive_tensor_descriptor_packed(make_tuple(N, K, Ho, Wo));
+}
+
+template <class InDesc, class WeiDesc, class OutDesc>
+constexpr std::size_t
+calculate_convolution_flops(const InDesc&, const WeiDesc& wei_desc, const OutDesc& out_desc)
+{
+    using namespace ck;
+
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    const index_t N  = out_desc.GetLength(I0);
+    const index_t K  = out_desc.GetLength(I1);
+    const index_t Ho = out_desc.GetLength(I2);
+    const index_t Wo = out_desc.GetLength(I3);
+
+    const index_t C = wei_desc.GetLength(I1);
+    const index_t Y = wei_desc.GetLength(I2);
+    const index_t X = wei_desc.GetLength(I3);
+
+    return std::size_t(2) * N * K * Ho * Wo * C * Y * X;
+}
diff --git a/library/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp b/library/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp
new file mode 100644
index 0000000000..ff697fb71c
--- /dev/null
+++ b/library/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp
@@ -0,0 +1,355 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace utils {
+namespace conv {
+
+namespace detail {
+
+template <typename OldLayout>
+std::vector<std::size_t> get_layout_transpose_gnchw_to_old()
+{
+    // HACK: NHWC/KYXC/NHWK, which is treated as GNHWC/GKYXC/GNHWK by this function,
+    // is used by some legacy kernel. New kernel should use GNHWK/GKYXC/GNHWK
+    // TODO: remove this branch after removing legacy kernel
+    if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NWC> ||
+                 ck::is_same_v<OldLayout, ck::tensor_layout::convolution::KXC> ||
+                 ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NWK>)
+    {
+        return {0, 1, 3, 2};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NHWC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::KYXC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NHWK>)
+    {
+        return {0, 1, 4, 2, 3};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NDHWC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::KZYXC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NDHWK>)
+    {
+        return {0, 1, 5, 2, 3, 4};
+    }
+    // separate from legacy code above
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNCW> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GKCX> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNKW>)
+    {
+        return {0, 1, 2, 3};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNCHW> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GKCYX> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNKHW>)
+    {
+        return {0, 1, 2, 3, 4};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNCDHW> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GKCZYX> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNKDHW>)
+    {
+        return {0, 1, 2, 3, 4, 5};
+    }
+    if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNWC> ||
+                 ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GKXC> ||
+                 ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNWK>)
+    {
+        return {0, 1, 3, 2};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNHWC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GKYXC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNHWK>)
+    {
+        return {0, 1, 4, 2, 3};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNDHWC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GKZYXC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::GNDHWK>)
+    {
+        return {0, 1, 5, 2, 3, 4};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NWGC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::KXGC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NWGK>)
+    {
+        return {2, 0, 3, 1};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NHWGC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::KYXGC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NHWGK>)
+    {
+        return {3, 0, 4, 1, 2};
+    }
+    else if constexpr(ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NDHWGC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::KZYXGC> ||
+                      ck::is_same_v<OldLayout, ck::tensor_layout::convolution::NDHWGK>)
+    {
+        return {4, 0, 5, 1, 2, 3};
+    }
+    else
+    {
+        printf("%s\n", __func__);
+        throw std::runtime_error("wrong! unsupported layout");
+    }
+}
+
+} // namespace detail
+
+// make tensor descriptor for packed input tensor, and order the dimension in the order of GNCHW
+// regardless of physical layout
+template <typename InLayout>
+HostTensorDescriptor
+make_input_host_tensor_descriptor_g_n_c_wis_packed(const ck::utils::conv::ConvParam& param)
+{
+    std::vector<std::size_t> physical_lengths;
+
+    // HACK: NHWC/KYXC/NHWK, which is treated as GNHWC/GKYXC/GNHWK by this function,
+    // is used by some legacy kernel. New kernel should use GNHWK/GKYXC/GNHWK
+    // TODO: remove this branch after removing legacy kernel
+    if constexpr(ck::is_same_v<InLayout, ck::tensor_layout::convolution::NWC> ||
+                 ck::is_same_v<InLayout, ck::tensor_layout::convolution::NHWC> ||
+                 ck::is_same_v<InLayout, ck::tensor_layout::convolution::NDHWC>)
+    {
+        if(param.G_ != 1)
+        {
+            throw std::runtime_error("wrong! G != 1");
+        }
+
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 2,
+                                param.input_spatial_lengths_.begin(),
+                                param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    // separate from legacy code above
+    else if constexpr(ck::is_same_v<InLayout, ck::tensor_layout::convolution::GNCW> ||
+                      ck::is_same_v<InLayout, ck::tensor_layout::convolution::GNCHW> ||
+                      ck::is_same_v<InLayout, ck::tensor_layout::convolution::GNCDHW>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.end(),
+                                param.input_spatial_lengths_.begin(),
+                                param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(ck::is_same_v<InLayout, ck::tensor_layout::convolution::GNWC> ||
+                      ck::is_same_v<InLayout, ck::tensor_layout::convolution::GNHWC> ||
+                      ck::is_same_v<InLayout, ck::tensor_layout::convolution::GNDHWC>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 2,
+                                param.input_spatial_lengths_.begin(),
+                                param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(ck::is_same_v<InLayout, ck::tensor_layout::convolution::NWGC> ||
+                      ck::is_same_v<InLayout, ck::tensor_layout::convolution::NHWGC> ||
+                      ck::is_same_v<InLayout, ck::tensor_layout::convolution::NDHWGC>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 1,
+                                param.input_spatial_lengths_.begin(),
+                                param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else
+    {
+        printf("%s\n", __func__);
+        printf("%s\n", InLayout::name);
+        throw std::runtime_error("wrong! unsupported layout");
+    }
+
+    return transpose_host_tensor_descriptor_given_new2old(
+        HostTensorDescriptor(physical_lengths),
+        detail::get_layout_transpose_gnchw_to_old<InLayout>());
+}
+
+// make tensor descriptor for packed weight tensor, and order the dimension in the order of GKCYX
+// regardless of physical layout
+template <typename WeiLayout>
+HostTensorDescriptor
+make_weight_host_tensor_descriptor_g_k_c_xs_packed(const ck::utils::conv::ConvParam& param)
+{
+    std::vector<std::size_t> physical_lengths;
+
+    // HACK: NHWC/KYXC/NHWK, which is treated as GNHWC/GKYXC/GNHWK by this function,
+    // is used by some legacy kernel. New kernel should use GNHWK/GKYXC/GNHWK
+    // TODO: remove this branch after removing legacy kernel
+    if constexpr(ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KXC> ||
+                 ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KYXC> ||
+                 ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KZYXC>)
+    {
+        if(param.G_ != 1)
+        {
+            throw std::runtime_error("wrong! G != 1");
+        }
+
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.K_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 2,
+                                param.filter_spatial_lengths_.begin(),
+                                param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    // separate from legacy code above
+    else if constexpr(ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KXC> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KYXC> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KZYXC>)
+    {
+        if(param.G_ != 1)
+        {
+            throw std::runtime_error("wrong! G != 1");
+        }
+
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.K_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.end(),
+                                param.filter_spatial_lengths_.begin(),
+                                param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::GKCX> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::GKCYX> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::GKCZYX>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.K_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.end(),
+                                param.filter_spatial_lengths_.begin(),
+                                param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::GKXC> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::GKYXC> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::GKZYXC>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.K_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 2,
+                                param.filter_spatial_lengths_.begin(),
+                                param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KXGC> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KYXGC> ||
+                      ck::is_same_v<WeiLayout, ck::tensor_layout::convolution::KZYXGC>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.K_),
+                                                    static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 1,
+                                param.filter_spatial_lengths_.begin(),
+                                param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else
+    {
+        printf("%s\n", __func__);
+        printf("%s\n", WeiLayout::name);
+        throw std::runtime_error("wrong! unsupported layout");
+    }
+
+    return transpose_host_tensor_descriptor_given_new2old(
+        HostTensorDescriptor(physical_lengths),
+        detail::get_layout_transpose_gnchw_to_old<WeiLayout>());
+}
+
+// make tensor descriptor for packed output tensor, and order the dimension in the order of GNKHW
+// regardless of physical layout
+template <typename OutLayout>
+HostTensorDescriptor
+make_output_host_tensor_descriptor_g_n_k_wos_packed(const ck::utils::conv::ConvParam& param)
+{
+    std::vector<std::size_t> physical_lengths;
+
+    // HACK: NHWC/KYXC/NHWK, which is treated as GNHWC/GKYXC/GNHWK by this function,
+    // is used by some legacy kernel. New kernel should use GNHWK/GKYXC/GNHWK
+    // TODO: remove this branch after removing legacy kernel
+    if constexpr(ck::is_same_v<OutLayout, ck::tensor_layout::convolution::NWK> ||
+                 ck::is_same_v<OutLayout, ck::tensor_layout::convolution::NHWK> ||
+                 ck::is_same_v<OutLayout, ck::tensor_layout::convolution::NDHWK>)
+    {
+        if(param.G_ != 1)
+        {
+            throw std::runtime_error("wrong! G != 1");
+        }
+
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.K_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 2,
+                                param.output_spatial_lengths_.begin(),
+                                param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    // separate from legacy code above
+    else if constexpr(ck::is_same_v<OutLayout, ck::tensor_layout::convolution::GNKW> ||
+                      ck::is_same_v<OutLayout, ck::tensor_layout::convolution::GNKHW> ||
+                      ck::is_same_v<OutLayout, ck::tensor_layout::convolution::GNKDHW>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.K_)};
+
+        physical_lengths.insert(physical_lengths.end(),
+                                param.output_spatial_lengths_.begin(),
+                                param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(ck::is_same_v<OutLayout, ck::tensor_layout::convolution::GNWK> ||
+                      ck::is_same_v<OutLayout, ck::tensor_layout::convolution::GNHWK> ||
+                      ck::is_same_v<OutLayout, ck::tensor_layout::convolution::GNDHWK>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.K_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 2,
+                                param.output_spatial_lengths_.begin(),
+                                param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(ck::is_same_v<OutLayout, ck::tensor_layout::convolution::NWGK> ||
+                      ck::is_same_v<OutLayout, ck::tensor_layout::convolution::NHWGK> ||
+                      ck::is_same_v<OutLayout, ck::tensor_layout::convolution::NDHWGK>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.K_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 1,
+                                param.output_spatial_lengths_.begin(),
+                                param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else
+    {
+        printf("%s\n", __func__);
+        printf("%s\n", OutLayout::name);
+        throw std::runtime_error("wrong! unsupported layout");
+    }
+
+    return transpose_host_tensor_descriptor_given_new2old(
+        HostTensorDescriptor(physical_lengths),
+        detail::get_layout_transpose_gnchw_to_old<OutLayout>());
+}
+
+} // namespace conv
+} // namespace utils
+} // namespace ck
diff --git a/library/include/ck/library/utility/convolution_parameter.hpp b/library/include/ck/library/utility/convolution_parameter.hpp
new file mode 100644
index 0000000000..df6efca108
--- /dev/null
+++ b/library/include/ck/library/utility/convolution_parameter.hpp
@@ -0,0 +1,101 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <numeric>
+#include <iterator>
+#include <vector>
+
+#include "ck/ck.hpp"
+
+#include "ck/library/utility/numeric.hpp"
+
+namespace ck {
+namespace utils {
+namespace conv {
+
+struct ConvParam
+{
+    ConvParam();
+    ConvParam(ck::index_t n_dim,
+              ck::index_t group_count,
+              ck::index_t n_batch,
+              ck::index_t n_out_channels,
+              ck::index_t n_in_channels,
+              const std::vector<ck::index_t>& filters_len,
+              const std::vector<ck::index_t>& input_len,
+              const std::vector<ck::index_t>& strides,
+              const std::vector<ck::index_t>& dilations,
+              const std::vector<ck::index_t>& left_pads,
+              const std::vector<ck::index_t>& right_pads);
+
+    ck::index_t num_dim_spatial_;
+    ck::index_t G_;
+    ck::index_t N_;
+    ck::index_t K_;
+    ck::index_t C_;
+
+    std::vector<ck::index_t> filter_spatial_lengths_;
+    std::vector<ck::index_t> input_spatial_lengths_;
+    std::vector<ck::index_t> output_spatial_lengths_;
+
+    std::vector<ck::index_t> conv_filter_strides_;
+    std::vector<ck::index_t> conv_filter_dilations_;
+
+    std::vector<ck::index_t> input_left_pads_;
+    std::vector<ck::index_t> input_right_pads_;
+
+    std::vector<ck::index_t> GetOutputSpatialLengths() const;
+
+    std::size_t GetFlops() const;
+
+    template <typename InDataType>
+    std::size_t GetInputByte() const
+    {
+        // sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
+        return sizeof(InDataType) *
+               (G_ * N_ * C_ *
+                ck::accumulate_n<std::size_t>(
+                    std::begin(input_spatial_lengths_), num_dim_spatial_, 1, std::multiplies<>()));
+    }
+
+    template <typename WeiDataType>
+    std::size_t GetWeightByte() const
+    {
+        // sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
+        return sizeof(WeiDataType) *
+               (G_ * K_ * C_ *
+                ck::accumulate_n<std::size_t>(
+                    std::begin(filter_spatial_lengths_), num_dim_spatial_, 1, std::multiplies<>()));
+    }
+
+    template <typename OutDataType>
+    std::size_t GetOutputByte() const
+    {
+        // sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
+        return sizeof(OutDataType) * (G_ * N_ * K_ *
+                                      std::accumulate(std::begin(output_spatial_lengths_),
+                                                      std::end(output_spatial_lengths_),
+                                                      static_cast<std::size_t>(1),
+                                                      std::multiplies<std::size_t>()));
+    }
+
+    template <typename InDataType, typename WeiDataType, typename OutDataType>
+    std::size_t GetByte() const
+    {
+        return GetInputByte<InDataType>() + GetWeightByte<WeiDataType>() +
+               GetOutputByte<OutDataType>();
+    }
+};
+
+std::string get_conv_param_parser_helper_msg();
+
+ConvParam parse_conv_param(int num_dim_spatial, int arg_idx, char* const argv[]);
+
+} // namespace conv
+} // namespace utils
+} // namespace ck
+
+std::ostream& operator<<(std::ostream& os, const ck::utils::conv::ConvParam& p);
diff --git a/library/include/ck/library/utility/device_memory.hpp b/library/include/ck/library/utility/device_memory.hpp
new file mode 100644
index 0000000000..d2e611a77c
--- /dev/null
+++ b/library/include/ck/library/utility/device_memory.hpp
@@ -0,0 +1,50 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <hip/hip_runtime.h>
+
+template <typename T>
+__global__ void set_buffer_value(T* p, T x, uint64_t buffer_element_size)
+{
+    for(uint64_t i = threadIdx.x; i < buffer_element_size; i += blockDim.x)
+    {
+        p[i] = x;
+    }
+}
+
+/**
+ * @brief Container for storing data in GPU device memory
+ *
+ */
+struct DeviceMem
+{
+    DeviceMem() : mpDeviceBuf(nullptr), mMemSize(0) {}
+    DeviceMem(std::size_t mem_size);
+    void Realloc(std::size_t mem_size);
+    void* GetDeviceBuffer() const;
+    std::size_t GetBufferSize() const;
+    void ToDevice(const void* p) const;
+    void ToDevice(const void* p, const std::size_t cpySize) const;
+    void FromDevice(void* p) const;
+    void FromDevice(void* p, const std::size_t cpySize) const;
+    void SetZero() const;
+    template <typename T>
+    void SetValue(T x) const;
+    ~DeviceMem();
+
+    void* mpDeviceBuf;
+    std::size_t mMemSize;
+};
+
+template <typename T>
+void DeviceMem::SetValue(T x) const
+{
+    if(mMemSize % sizeof(T) != 0)
+    {
+        throw std::runtime_error("wrong! not entire DeviceMem will be set");
+    }
+
+    set_buffer_value<T><<<1, 1024>>>(static_cast<T*>(mpDeviceBuf), x, mMemSize / sizeof(T));
+}
diff --git a/library/include/ck/library/utility/fill.hpp b/library/include/ck/library/utility/fill.hpp
new file mode 100644
index 0000000000..4e075df43b
--- /dev/null
+++ b/library/include/ck/library/utility/fill.hpp
@@ -0,0 +1,137 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <cmath>
+#include <iterator>
+#include <random>
+#include <type_traits>
+#include <utility>
+
+#include "ck/utility/data_type.hpp"
+
+namespace ck {
+namespace utils {
+
+template <typename T>
+struct FillUniformDistribution
+{
+    float a_{-5.f};
+    float b_{5.f};
+
+    template <typename ForwardIter>
+    void operator()(ForwardIter first, ForwardIter last) const
+    {
+        std::mt19937 gen(11939);
+        std::uniform_real_distribution<float> dis(a_, b_);
+        std::generate(first, last, [&dis, &gen]() { return ck::type_convert<T>(dis(gen)); });
+    }
+
+    template <typename ForwardRange>
+    auto operator()(ForwardRange&& range) const
+        -> std::void_t<decltype(std::declval<const FillUniformDistribution&>()(
+            std::begin(std::forward<ForwardRange>(range)),
+            std::end(std::forward<ForwardRange>(range))))>
+    {
+        (*this)(std::begin(std::forward<ForwardRange>(range)),
+                std::end(std::forward<ForwardRange>(range)));
+    }
+};
+
+// Normally FillUniformDistributionIntegerValue should use std::uniform_int_distribution as below.
+// However this produces segfaults in std::mt19937 which look like inifite loop.
+//      template <typename T>
+//      struct FillUniformDistributionIntegerValue
+//      {
+//          int a_{-5};
+//          int b_{5};
+//
+//          template <typename ForwardIter>
+//          void operator()(ForwardIter first, ForwardIter last) const
+//          {
+//              std::mt19937 gen(11939);
+//              std::uniform_int_distribution<int> dis(a_, b_);
+//              std::generate(
+//                  first, last, [&dis, &gen]() { return ck::type_convert<T>(dis(gen)); });
+//          }
+//      };
+
+// Workaround for uniform_int_distribution not working as expected. See note above.<
+template <typename T>
+struct FillUniformDistributionIntegerValue
+{
+    float a_{-5.f};
+    float b_{5.f};
+
+    template <typename ForwardIter>
+    void operator()(ForwardIter first, ForwardIter last) const
+    {
+        std::mt19937 gen(11939);
+        std::uniform_real_distribution<float> dis(a_, b_);
+        std::generate(
+            first, last, [&dis, &gen]() { return ck::type_convert<T>(std::round(dis(gen))); });
+    }
+
+    template <typename ForwardRange>
+    auto operator()(ForwardRange&& range) const
+        -> std::void_t<decltype(std::declval<const FillUniformDistributionIntegerValue&>()(
+            std::begin(std::forward<ForwardRange>(range)),
+            std::end(std::forward<ForwardRange>(range))))>
+    {
+        (*this)(std::begin(std::forward<ForwardRange>(range)),
+                std::end(std::forward<ForwardRange>(range)));
+    }
+};
+
+template <typename T>
+struct FillMonotonicSeq
+{
+    T init_value_{0};
+    T step_{1};
+
+    template <typename ForwardIter>
+    void operator()(ForwardIter first, ForwardIter last) const
+    {
+        std::generate(first, last, [=, n = init_value_]() mutable {
+            auto tmp = n;
+            n += step_;
+            return tmp;
+        });
+    }
+
+    template <typename ForwardRange>
+    auto operator()(ForwardRange&& range) const
+        -> std::void_t<decltype(std::declval<const FillMonotonicSeq&>()(
+            std::begin(std::forward<ForwardRange>(range)),
+            std::end(std::forward<ForwardRange>(range))))>
+    {
+        (*this)(std::begin(std::forward<ForwardRange>(range)),
+                std::end(std::forward<ForwardRange>(range)));
+    }
+};
+
+template <typename T>
+struct FillConstant
+{
+    T value_{0};
+
+    template <typename ForwardIter>
+    void operator()(ForwardIter first, ForwardIter last) const
+    {
+        std::fill(first, last, value_);
+    }
+
+    template <typename ForwardRange>
+    auto operator()(ForwardRange&& range) const -> std::void_t<
+        decltype(std::declval<const FillConstant&>()(std::begin(std::forward<ForwardRange>(range)),
+                                                     std::end(std::forward<ForwardRange>(range))))>
+    {
+        (*this)(std::begin(std::forward<ForwardRange>(range)),
+                std::end(std::forward<ForwardRange>(range)));
+    }
+};
+
+} // namespace utils
+} // namespace ck
diff --git a/library/include/ck/library/utility/host_common_util.hpp b/library/include/ck/library/utility/host_common_util.hpp
new file mode 100644
index 0000000000..20a8f234db
--- /dev/null
+++ b/library/include/ck/library/utility/host_common_util.hpp
@@ -0,0 +1,136 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <array>
+#include <iostream>
+#include <fstream>
+#include <string>
+#include <algorithm>
+
+#include "ck/ck.hpp"
+
+namespace ck {
+
+namespace host_common {
+
+template <typename T>
+static inline void dumpBufferToFile(const char* fileName, T* data, size_t dataNumItems)
+{
+    std::ofstream outFile(fileName, std::ios::binary);
+    if(outFile)
+    {
+        outFile.write(reinterpret_cast<char*>(data), dataNumItems * sizeof(T));
+        outFile.close();
+        std::cout << "Write output to file " << fileName << std::endl;
+    }
+    else
+    {
+        std::cout << "Could not open file " << fileName << " for writing" << std::endl;
+    }
+};
+
+template <typename T>
+static inline T getSingleValueFromString(const std::string& valueStr)
+{
+    std::istringstream iss(valueStr);
+
+    T val;
+
+    iss >> val;
+
+    return (val);
+};
+
+template <typename T>
+static inline std::vector<T> getTypeValuesFromString(const char* cstr_values)
+{
+    std::string valuesStr(cstr_values);
+
+    std::vector<T> values;
+    std::size_t pos = 0;
+    std::size_t new_pos;
+
+    new_pos = valuesStr.find(',', pos);
+    while(new_pos != std::string::npos)
+    {
+        const std::string sliceStr = valuesStr.substr(pos, new_pos - pos);
+
+        T val = getSingleValueFromString<T>(sliceStr);
+
+        values.push_back(val);
+
+        pos     = new_pos + 1;
+        new_pos = valuesStr.find(',', pos);
+    };
+
+    std::string sliceStr = valuesStr.substr(pos);
+    T val                = getSingleValueFromString<T>(sliceStr);
+
+    values.push_back(val);
+
+    return (values);
+}
+
+template <int NDim>
+static inline std::vector<std::array<index_t, NDim>>
+get_index_set(const std::array<index_t, NDim>& dim_lengths)
+{
+    static_assert(NDim >= 1, "NDim >= 1 is required to use this function!");
+
+    if constexpr(NDim == 1)
+    {
+        std::vector<std::array<index_t, NDim>> index_set;
+
+        for(int i = 0; i < dim_lengths[0]; i++)
+        {
+            std::array<index_t, 1> index{i};
+
+            index_set.push_back(index);
+        };
+
+        return index_set;
+    }
+    else
+    {
+        std::vector<std::array<index_t, NDim>> index_set;
+        std::array<index_t, NDim - 1> partial_dim_lengths;
+
+        std::copy(dim_lengths.begin() + 1, dim_lengths.end(), partial_dim_lengths.begin());
+
+        std::vector<std::array<index_t, NDim - 1>> partial_index_set;
+
+        partial_index_set = get_index_set<NDim - 1>(partial_dim_lengths);
+
+        for(index_t i = 0; i < dim_lengths[0]; i++)
+            for(const auto& partial_index : partial_index_set)
+            {
+                std::array<index_t, NDim> index;
+
+                index[0] = i;
+
+                std::copy(partial_index.begin(), partial_index.end(), index.begin() + 1);
+
+                index_set.push_back(index);
+            };
+
+        return index_set;
+    };
+};
+
+template <int NDim>
+static inline size_t get_offset_from_index(const std::array<index_t, NDim>& strides,
+                                           const std::array<index_t, NDim>& index)
+{
+    size_t offset = 0;
+
+    for(int i = 0; i < NDim; i++)
+        offset += index[i] * strides[i];
+
+    return (offset);
+};
+
+} // namespace host_common
+} // namespace ck
diff --git a/library/include/ck/library/utility/host_gemm.hpp b/library/include/ck/library/utility/host_gemm.hpp
new file mode 100644
index 0000000000..5eb7e3b8c9
--- /dev/null
+++ b/library/include/ck/library/utility/host_gemm.hpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "host_tensor.hpp"
+
+template <typename AType,
+          typename BType,
+          typename CType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+void host_gemm_mk_kn_mn(const Tensor<AType>& a_m_k,
+                        const Tensor<BType>& b_k_n,
+                        Tensor<CType>& c_m_n,
+                        const AElementwiseOperation& a_element_op,
+                        const BElementwiseOperation& b_element_op,
+                        const CElementwiseOperation& c_element_op)
+{
+    auto f_mk_kn_mn = [&](auto m, auto n) {
+        const int K = a_m_k.mDesc.GetLengths()[1];
+
+        float v_acc = 0;
+
+        for(int k = 0; k < K; ++k)
+        {
+            float v_a;
+            float v_b;
+
+            a_element_op(v_a, static_cast<const float>(a_m_k(m, k)));
+            b_element_op(v_b, static_cast<const float>(b_k_n(k, n)));
+
+            v_acc += v_a * v_b;
+        }
+
+        float v_c;
+
+        c_element_op(v_c, v_acc);
+
+        c_m_n(m, n) = v_c;
+    };
+
+    make_ParallelTensorFunctor(f_mk_kn_mn,
+                               c_m_n.mDesc.GetLengths()[0],
+                               c_m_n.mDesc.GetLengths()[1])(std::thread::hardware_concurrency());
+}
diff --git a/library/include/ck/library/utility/host_tensor.hpp b/library/include/ck/library/utility/host_tensor.hpp
new file mode 100644
index 0000000000..09e89157c1
--- /dev/null
+++ b/library/include/ck/library/utility/host_tensor.hpp
@@ -0,0 +1,488 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <cassert>
+#include <iostream>
+#include <iomanip>
+#include <numeric>
+#include <thread>
+#include <utility>
+#include <vector>
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/span.hpp"
+#include "ck/utility/type_convert.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/ranges.hpp"
+
+template <typename Range>
+std::ostream& LogRange(std::ostream& os,
+                       Range&& range,
+                       std::string delim,
+                       int precision = std::cout.precision(),
+                       int width     = 0)
+{
+    bool first = true;
+    for(auto&& v : range)
+    {
+        if(first)
+            first = false;
+        else
+            os << delim;
+        os << std::setw(width) << std::setprecision(precision) << v;
+    }
+    return os;
+}
+
+template <typename T, typename Range>
+std::ostream& LogRangeAsType(std::ostream& os,
+                             Range&& range,
+                             std::string delim,
+                             int precision = std::cout.precision(),
+                             int width     = 0)
+{
+    bool first = true;
+    for(auto&& v : range)
+    {
+        if(first)
+            first = false;
+        else
+            os << delim;
+        os << std::setw(width) << std::setprecision(precision) << static_cast<T>(v);
+    }
+    return os;
+}
+
+template <typename F, typename T, std::size_t... Is>
+auto call_f_unpack_args_impl(F f, T args, std::index_sequence<Is...>)
+{
+    return f(std::get<Is>(args)...);
+}
+
+template <typename F, typename T>
+auto call_f_unpack_args(F f, T args)
+{
+    constexpr std::size_t N = std::tuple_size<T>{};
+
+    return call_f_unpack_args_impl(f, args, std::make_index_sequence<N>{});
+}
+
+template <typename F, typename T, std::size_t... Is>
+auto construct_f_unpack_args_impl(T args, std::index_sequence<Is...>)
+{
+    return F(std::get<Is>(args)...);
+}
+
+template <typename F, typename T>
+auto construct_f_unpack_args(F, T args)
+{
+    constexpr std::size_t N = std::tuple_size<T>{};
+
+    return construct_f_unpack_args_impl<F>(args, std::make_index_sequence<N>{});
+}
+
+struct HostTensorDescriptor
+{
+    HostTensorDescriptor() = default;
+
+    void CalculateStrides();
+
+    template <typename X, typename = std::enable_if_t<std::is_convertible_v<X, std::size_t>>>
+    HostTensorDescriptor(const std::initializer_list<X>& lens) : mLens(lens.begin(), lens.end())
+    {
+        this->CalculateStrides();
+    }
+
+    template <typename Lengths,
+              typename = std::enable_if_t<
+                  std::is_convertible_v<ck::ranges::range_value_t<Lengths>, std::size_t>>>
+    HostTensorDescriptor(const Lengths& lens) : mLens(lens.begin(), lens.end())
+    {
+        this->CalculateStrides();
+    }
+
+    template <typename X,
+              typename Y,
+              typename = std::enable_if_t<std::is_convertible_v<X, std::size_t> &&
+                                          std::is_convertible_v<Y, std::size_t>>>
+    HostTensorDescriptor(const std::initializer_list<X>& lens,
+                         const std::initializer_list<Y>& strides)
+        : mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end())
+    {
+    }
+
+    template <typename Lengths,
+              typename Strides,
+              typename = std::enable_if_t<
+                  std::is_convertible_v<ck::ranges::range_value_t<Lengths>, std::size_t> &&
+                  std::is_convertible_v<ck::ranges::range_value_t<Strides>, std::size_t>>>
+    HostTensorDescriptor(const Lengths& lens, const Strides& strides)
+        : mLens(lens.begin(), lens.end()), mStrides(strides.begin(), strides.end())
+    {
+    }
+
+    std::size_t GetNumOfDimension() const;
+    std::size_t GetElementSize() const;
+    std::size_t GetElementSpaceSize() const;
+
+    const std::vector<std::size_t>& GetLengths() const;
+    const std::vector<std::size_t>& GetStrides() const;
+
+    template <typename... Is>
+    std::size_t GetOffsetFromMultiIndex(Is... is) const
+    {
+        assert(sizeof...(Is) == this->GetNumOfDimension());
+        std::initializer_list<std::size_t> iss{static_cast<std::size_t>(is)...};
+        return std::inner_product(iss.begin(), iss.end(), mStrides.begin(), std::size_t{0});
+    }
+
+    std::size_t GetOffsetFromMultiIndex(std::vector<std::size_t> iss) const
+    {
+        return std::inner_product(iss.begin(), iss.end(), mStrides.begin(), std::size_t{0});
+    }
+
+    friend std::ostream& operator<<(std::ostream& os, const HostTensorDescriptor& desc);
+
+    private:
+    std::vector<std::size_t> mLens;
+    std::vector<std::size_t> mStrides;
+};
+
+template <typename New2Old>
+HostTensorDescriptor transpose_host_tensor_descriptor_given_new2old(const HostTensorDescriptor& a,
+                                                                    const New2Old& new2old)
+{
+    std::vector<std::size_t> new_lengths(a.GetNumOfDimension());
+    std::vector<std::size_t> new_strides(a.GetNumOfDimension());
+
+    for(std::size_t i = 0; i < a.GetNumOfDimension(); i++)
+    {
+        new_lengths[i] = a.GetLengths()[new2old[i]];
+        new_strides[i] = a.GetStrides()[new2old[i]];
+    }
+
+    return HostTensorDescriptor(new_lengths, new_strides);
+}
+
+struct joinable_thread : std::thread
+{
+    template <typename... Xs>
+    joinable_thread(Xs&&... xs) : std::thread(std::forward<Xs>(xs)...)
+    {
+    }
+
+    joinable_thread(joinable_thread&&) = default;
+    joinable_thread& operator=(joinable_thread&&) = default;
+
+    ~joinable_thread()
+    {
+        if(this->joinable())
+            this->join();
+    }
+};
+
+template <typename F, typename... Xs>
+struct ParallelTensorFunctor
+{
+    F mF;
+    static constexpr std::size_t NDIM = sizeof...(Xs);
+    std::array<std::size_t, NDIM> mLens;
+    std::array<std::size_t, NDIM> mStrides;
+    std::size_t mN1d;
+
+    ParallelTensorFunctor(F f, Xs... xs) : mF(f), mLens({static_cast<std::size_t>(xs)...})
+    {
+        mStrides.back() = 1;
+        std::partial_sum(mLens.rbegin(),
+                         mLens.rend() - 1,
+                         mStrides.rbegin() + 1,
+                         std::multiplies<std::size_t>());
+        mN1d = mStrides[0] * mLens[0];
+    }
+
+    std::array<std::size_t, NDIM> GetNdIndices(std::size_t i) const
+    {
+        std::array<std::size_t, NDIM> indices;
+
+        for(std::size_t idim = 0; idim < NDIM; ++idim)
+        {
+            indices[idim] = i / mStrides[idim];
+            i -= indices[idim] * mStrides[idim];
+        }
+
+        return indices;
+    }
+
+    void operator()(std::size_t num_thread = 1) const
+    {
+        std::size_t work_per_thread = (mN1d + num_thread - 1) / num_thread;
+
+        std::vector<joinable_thread> threads(num_thread);
+
+        for(std::size_t it = 0; it < num_thread; ++it)
+        {
+            std::size_t iw_begin = it * work_per_thread;
+            std::size_t iw_end   = std::min((it + 1) * work_per_thread, mN1d);
+
+            auto f = [this, iw_begin, iw_end] {
+                for(std::size_t iw = iw_begin; iw < iw_end; ++iw)
+                {
+                    call_f_unpack_args(this->mF, this->GetNdIndices(iw));
+                }
+            };
+            threads[it] = joinable_thread(f);
+        }
+    }
+};
+
+template <typename F, typename... Xs>
+auto make_ParallelTensorFunctor(F f, Xs... xs)
+{
+    return ParallelTensorFunctor<F, Xs...>(f, xs...);
+}
+
+template <typename T>
+struct Tensor
+{
+    using Descriptor = HostTensorDescriptor;
+    using Data       = std::vector<T>;
+
+    template <typename X>
+    Tensor(std::initializer_list<X> lens) : mDesc(lens), mData(mDesc.GetElementSpaceSize())
+    {
+    }
+
+    template <typename X, typename Y>
+    Tensor(std::initializer_list<X> lens, std::initializer_list<Y> strides)
+        : mDesc(lens, strides), mData(mDesc.GetElementSpaceSize())
+    {
+    }
+
+    template <typename Lengths>
+    Tensor(const Lengths& lens) : mDesc(lens), mData(mDesc.GetElementSpaceSize())
+    {
+    }
+
+    template <typename Lengths, typename Strides>
+    Tensor(const Lengths& lens, const Strides& strides)
+        : mDesc(lens, strides), mData(GetElementSpaceSize())
+    {
+    }
+
+    Tensor(const Descriptor& desc) : mDesc(desc), mData(mDesc.GetElementSpaceSize()) {}
+
+    template <typename OutT>
+    Tensor<OutT> CopyAsType() const
+    {
+        Tensor<OutT> ret(mDesc);
+
+        ck::ranges::transform(
+            mData, ret.mData.begin(), [](auto value) { return ck::type_convert<OutT>(value); });
+
+        return ret;
+    }
+
+    Tensor()              = delete;
+    Tensor(const Tensor&) = default;
+    Tensor(Tensor&&)      = default;
+
+    ~Tensor() = default;
+
+    Tensor& operator=(const Tensor&) = default;
+    Tensor& operator=(Tensor&&) = default;
+
+    template <typename FromT>
+    explicit Tensor(const Tensor<FromT>& other) : Tensor(other.template CopyAsType<T>())
+    {
+    }
+
+    decltype(auto) GetLengths() const { return mDesc.GetLengths(); }
+
+    decltype(auto) GetStrides() const { return mDesc.GetStrides(); }
+
+    std::size_t GetNumOfDimension() const { return mDesc.GetNumOfDimension(); }
+
+    std::size_t GetElementSize() const { return mDesc.GetElementSize(); }
+
+    std::size_t GetElementSpaceSize() const { return mDesc.GetElementSpaceSize(); }
+
+    std::size_t GetElementSpaceSizeInBytes() const { return sizeof(T) * GetElementSpaceSize(); }
+
+    void SetZero() { ck::ranges::fill<T>(mData, 0); }
+
+    template <typename F>
+    void ForEach_impl(F&& f, std::vector<size_t>& idx, size_t rank)
+    {
+        if(rank == mDesc.GetNumOfDimension())
+        {
+            f(*this, idx);
+            return;
+        }
+        // else
+        for(size_t i = 0; i < mDesc.GetLengths()[rank]; i++)
+        {
+            idx[rank] = i;
+            ForEach_impl(std::forward<F>(f), idx, rank + 1);
+        }
+    }
+
+    template <typename F>
+    void ForEach(F&& f)
+    {
+        std::vector<size_t> idx(mDesc.GetNumOfDimension(), 0);
+        ForEach_impl(std::forward<F>(f), idx, size_t(0));
+    }
+
+    template <typename F>
+    void ForEach_impl(const F&& f, std::vector<size_t>& idx, size_t rank) const
+    {
+        if(rank == mDesc.GetNumOfDimension())
+        {
+            f(*this, idx);
+            return;
+        }
+        // else
+        for(size_t i = 0; i < mDesc.GetLengths()[rank]; i++)
+        {
+            idx[rank] = i;
+            ForEach_impl(std::forward<const F>(f), idx, rank + 1);
+        }
+    }
+
+    template <typename F>
+    void ForEach(const F&& f) const
+    {
+        std::vector<size_t> idx(mDesc.GetNumOfDimension(), 0);
+        ForEach_impl(std::forward<const F>(f), idx, size_t(0));
+    }
+
+    template <typename G>
+    void GenerateTensorValue(G g, std::size_t num_thread = 1)
+    {
+        switch(mDesc.GetNumOfDimension())
+        {
+        case 1: {
+            auto f = [&](auto i) { (*this)(i) = g(i); };
+            make_ParallelTensorFunctor(f, mDesc.GetLengths()[0])(num_thread);
+            break;
+        }
+        case 2: {
+            auto f = [&](auto i0, auto i1) { (*this)(i0, i1) = g(i0, i1); };
+            make_ParallelTensorFunctor(f, mDesc.GetLengths()[0], mDesc.GetLengths()[1])(num_thread);
+            break;
+        }
+        case 3: {
+            auto f = [&](auto i0, auto i1, auto i2) { (*this)(i0, i1, i2) = g(i0, i1, i2); };
+            make_ParallelTensorFunctor(
+                f, mDesc.GetLengths()[0], mDesc.GetLengths()[1], mDesc.GetLengths()[2])(num_thread);
+            break;
+        }
+        case 4: {
+            auto f = [&](auto i0, auto i1, auto i2, auto i3) {
+                (*this)(i0, i1, i2, i3) = g(i0, i1, i2, i3);
+            };
+            make_ParallelTensorFunctor(f,
+                                       mDesc.GetLengths()[0],
+                                       mDesc.GetLengths()[1],
+                                       mDesc.GetLengths()[2],
+                                       mDesc.GetLengths()[3])(num_thread);
+            break;
+        }
+        case 5: {
+            auto f = [&](auto i0, auto i1, auto i2, auto i3, auto i4) {
+                (*this)(i0, i1, i2, i3, i4) = g(i0, i1, i2, i3, i4);
+            };
+            make_ParallelTensorFunctor(f,
+                                       mDesc.GetLengths()[0],
+                                       mDesc.GetLengths()[1],
+                                       mDesc.GetLengths()[2],
+                                       mDesc.GetLengths()[3],
+                                       mDesc.GetLengths()[4])(num_thread);
+            break;
+        }
+        case 6: {
+            auto f = [&](auto i0, auto i1, auto i2, auto i3, auto i4, auto i5) {
+                (*this)(i0, i1, i2, i3, i4, i5) = g(i0, i1, i2, i3, i4, i5);
+            };
+            make_ParallelTensorFunctor(f,
+                                       mDesc.GetLengths()[0],
+                                       mDesc.GetLengths()[1],
+                                       mDesc.GetLengths()[2],
+                                       mDesc.GetLengths()[3],
+                                       mDesc.GetLengths()[4],
+                                       mDesc.GetLengths()[5])(num_thread);
+            break;
+        }
+        default: throw std::runtime_error("unspported dimension");
+        }
+    }
+
+    template <typename... Is>
+    std::size_t GetOffsetFromMultiIndex(Is... is) const
+    {
+        return mDesc.GetOffsetFromMultiIndex(is...);
+    }
+
+    template <typename... Is>
+    T& operator()(Is... is)
+    {
+        return mData[mDesc.GetOffsetFromMultiIndex(is...)];
+    }
+
+    template <typename... Is>
+    const T& operator()(Is... is) const
+    {
+        return mData[mDesc.GetOffsetFromMultiIndex(is...)];
+    }
+
+    T& operator()(std::vector<std::size_t> idx)
+    {
+        return mData[mDesc.GetOffsetFromMultiIndex(idx)];
+    }
+
+    const T& operator()(std::vector<std::size_t> idx) const
+    {
+        return mData[mDesc.GetOffsetFromMultiIndex(idx)];
+    }
+
+    typename Data::iterator begin() { return mData.begin(); }
+
+    typename Data::iterator end() { return mData.end(); }
+
+    typename Data::pointer data() { return mData.data(); }
+
+    typename Data::const_iterator begin() const { return mData.begin(); }
+
+    typename Data::const_iterator end() const { return mData.end(); }
+
+    typename Data::const_pointer data() const { return mData.data(); }
+
+    typename Data::size_type size() const { return mData.size(); }
+
+    template <typename U = T>
+    auto AsSpan() const
+    {
+        constexpr std::size_t FromSize = sizeof(T);
+        constexpr std::size_t ToSize   = sizeof(U);
+
+        using Element = std::add_const_t<std::remove_reference_t<U>>;
+        return ck::span<Element>{reinterpret_cast<Element*>(data()), size() * FromSize / ToSize};
+    }
+
+    template <typename U = T>
+    auto AsSpan()
+    {
+        constexpr std::size_t FromSize = sizeof(T);
+        constexpr std::size_t ToSize   = sizeof(U);
+
+        using Element = std::remove_reference_t<U>;
+        return ck::span<Element>{reinterpret_cast<Element*>(data()), size() * FromSize / ToSize};
+    }
+
+    Descriptor mDesc;
+    Data mData;
+};
diff --git a/library/include/ck/library/utility/host_tensor_generator.hpp b/library/include/ck/library/utility/host_tensor_generator.hpp
new file mode 100644
index 0000000000..31ff13aec2
--- /dev/null
+++ b/library/include/ck/library/utility/host_tensor_generator.hpp
@@ -0,0 +1,190 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cmath>
+#include <numeric>
+#include <random>
+
+#include "ck/ck.hpp"
+
+template <typename T>
+struct GeneratorTensor_0
+{
+    template <typename... Is>
+    T operator()(Is...)
+    {
+        return T{0};
+    }
+};
+
+template <typename T>
+struct GeneratorTensor_1
+{
+    T value = 1;
+
+    template <typename... Is>
+    T operator()(Is...)
+    {
+        return value;
+    }
+};
+
+template <>
+struct GeneratorTensor_1<ck::bhalf_t>
+{
+    float value = 1.0;
+
+    template <typename... Is>
+    ck::bhalf_t operator()(Is...)
+    {
+        return ck::type_convert<ck::bhalf_t>(value);
+    }
+};
+
+template <>
+struct GeneratorTensor_1<int8_t>
+{
+    int8_t value = 1;
+
+    template <typename... Is>
+    int8_t operator()(Is...)
+    {
+        return value;
+    }
+};
+
+template <typename T>
+struct GeneratorTensor_2
+{
+    int min_value = 0;
+    int max_value = 1;
+
+    template <typename... Is>
+    T operator()(Is...)
+    {
+        return static_cast<T>((std::rand() % (max_value - min_value)) + min_value);
+    }
+};
+
+template <>
+struct GeneratorTensor_2<ck::bhalf_t>
+{
+    int min_value = 0;
+    int max_value = 1;
+
+    template <typename... Is>
+    ck::bhalf_t operator()(Is...)
+    {
+        float tmp = (std::rand() % (max_value - min_value)) + min_value;
+        return ck::type_convert<ck::bhalf_t>(tmp);
+    }
+};
+
+template <>
+struct GeneratorTensor_2<int8_t>
+{
+    int min_value = 0;
+    int max_value = 1;
+
+    template <typename... Is>
+    int8_t operator()(Is...)
+    {
+        return (std::rand() % (max_value - min_value)) + min_value;
+    }
+};
+
+template <typename T>
+struct GeneratorTensor_3
+{
+    float min_value = 0;
+    float max_value = 1;
+
+    template <typename... Is>
+    T operator()(Is...)
+    {
+        float tmp = float(std::rand()) / float(RAND_MAX);
+
+        return static_cast<T>(min_value + tmp * (max_value - min_value));
+    }
+};
+
+template <>
+struct GeneratorTensor_3<ck::bhalf_t>
+{
+    float min_value = 0;
+    float max_value = 1;
+
+    template <typename... Is>
+    ck::bhalf_t operator()(Is...)
+    {
+        float tmp = float(std::rand()) / float(RAND_MAX);
+
+        float fp32_tmp = min_value + tmp * (max_value - min_value);
+
+        return ck::type_convert<ck::bhalf_t>(fp32_tmp);
+    }
+};
+
+template <typename T>
+struct GeneratorTensor_4
+{
+    std::default_random_engine generator;
+    std::normal_distribution<float> distribution;
+
+    GeneratorTensor_4(float mean, float stddev) : generator(1), distribution(mean, stddev){};
+
+    template <typename... Is>
+    T operator()(Is...)
+    {
+        float tmp = distribution(generator);
+
+        return ck::type_convert<T>(tmp);
+    }
+};
+
+struct GeneratorTensor_Checkboard
+{
+    template <typename... Ts>
+    float operator()(Ts... Xs) const
+    {
+        std::array<ck::index_t, sizeof...(Ts)> dims = {static_cast<ck::index_t>(Xs)...};
+        return std::accumulate(dims.begin(),
+                               dims.end(),
+                               true,
+                               [](bool init, ck::index_t x) -> int { return init != (x % 2); })
+                   ? 1
+                   : -1;
+    }
+};
+
+template <ck::index_t Dim>
+struct GeneratorTensor_Sequential
+{
+    template <typename... Ts>
+    float operator()(Ts... Xs) const
+    {
+        std::array<ck::index_t, sizeof...(Ts)> dims = {{static_cast<ck::index_t>(Xs)...}};
+        return dims[Dim];
+    }
+};
+
+template <typename T, size_t NumEffectiveDim = 2>
+struct GeneratorTensor_Diagonal
+{
+    T value{1};
+
+    template <typename... Ts>
+    T operator()(Ts... Xs) const
+    {
+        std::array<ck::index_t, sizeof...(Ts)> dims = {{static_cast<ck::index_t>(Xs)...}};
+        size_t start_dim                            = dims.size() - NumEffectiveDim;
+        bool pred                                   = true;
+        for(size_t i = start_dim + 1; i < dims.size(); i++)
+        {
+            pred &= (dims[start_dim] == dims[i]);
+        }
+        return pred ? value : T{0};
+    }
+};
diff --git a/library/include/ck/library/utility/iterator.hpp b/library/include/ck/library/utility/iterator.hpp
new file mode 100644
index 0000000000..b44e2d8e3c
--- /dev/null
+++ b/library/include/ck/library/utility/iterator.hpp
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iterator>
+#include <utility>
+
+#include "ck/utility/type.hpp"
+
+namespace ck {
+
+template <typename T>
+using iter_value_t = typename std::iterator_traits<remove_cvref_t<T>>::value_type;
+
+template <typename T>
+using iter_reference_t = decltype(*std::declval<T&>());
+
+template <typename T>
+using iter_difference_t = typename std::iterator_traits<remove_cvref_t<T>>::difference_type;
+
+} // namespace ck
diff --git a/library/include/ck/library/utility/literals.hpp b/library/include/ck/library/utility/literals.hpp
new file mode 100644
index 0000000000..a8bd6303f1
--- /dev/null
+++ b/library/include/ck/library/utility/literals.hpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+namespace literals {
+// [P0330] Literal Suffix for (signed) size_t (C++23)
+// ref: https://wg21.link/p0330r8
+inline constexpr std::size_t operator""_uz(unsigned long long size)
+{
+    return static_cast<std::size_t>(size);
+}
+
+inline constexpr std::size_t operator""_zu(unsigned long long size)
+{
+    return static_cast<std::size_t>(size);
+}
+} // namespace literals
+} // namespace ck
diff --git a/library/include/ck/library/utility/numeric.hpp b/library/include/ck/library/utility/numeric.hpp
new file mode 100644
index 0000000000..9ee118d475
--- /dev/null
+++ b/library/include/ck/library/utility/numeric.hpp
@@ -0,0 +1,16 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iterator>
+#include <numeric>
+
+namespace ck {
+template <typename T, typename ForwardIterator, typename Size, typename BinaryOperation>
+auto accumulate_n(ForwardIterator first, Size count, T init, BinaryOperation op)
+    -> decltype(std::accumulate(first, std::next(first, count), init, op))
+{
+    return std::accumulate(first, std::next(first, count), init, op);
+}
+} // namespace ck
diff --git a/library/include/ck/library/utility/ranges.hpp b/library/include/ck/library/utility/ranges.hpp
new file mode 100644
index 0000000000..f11e4204a0
--- /dev/null
+++ b/library/include/ck/library/utility/ranges.hpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iterator>
+#include <type_traits>
+#include <utility>
+
+#include "ck/library/utility/iterator.hpp"
+
+namespace ck {
+namespace ranges {
+
+template <typename R>
+using iterator_t = decltype(std::begin(std::declval<R&>()));
+
+template <typename R>
+using sentinel_t = decltype(std::end(std::declval<R&>()));
+
+template <typename R>
+using range_size_t = decltype(std::size(std::declval<R&>()));
+
+template <typename R>
+using range_difference_t = ck::iter_difference_t<ranges::iterator_t<R>>;
+
+template <typename R>
+using range_value_t = iter_value_t<ranges::iterator_t<R>>;
+
+template <typename R>
+using range_reference_t = iter_reference_t<ranges::iterator_t<R>>;
+
+template <typename T, typename = void>
+struct is_range : std::false_type
+{
+};
+
+template <typename T>
+struct is_range<
+    T,
+    std::void_t<decltype(std::begin(std::declval<T&>())), decltype(std::end(std::declval<T&>()))>>
+    : std::true_type
+{
+};
+
+template <typename T>
+inline constexpr bool is_range_v = is_range<T>::value;
+
+template <typename T, typename = void>
+struct is_sized_range : std::false_type
+{
+};
+
+template <typename T>
+struct is_sized_range<T, std::void_t<decltype(std::size(std::declval<T&>()))>>
+    : std::bool_constant<is_range_v<T>>
+{
+};
+} // namespace ranges
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/CMakeLists.txt
new file mode 100644
index 0000000000..1d54a141b7
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/CMakeLists.txt
@@ -0,0 +1,104 @@
+function(add_instance_library INSTANCE_NAME)
+    message("adding instance ${INSTANCE_NAME}")
+    add_library(${INSTANCE_NAME} OBJECT ${ARGN})
+    target_compile_features(${INSTANCE_NAME} PUBLIC)
+    set_target_properties(${INSTANCE_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    clang_tidy_check(${INSTANCE_NAME})
+endfunction(add_instance_library INSTANCE_NAME)
+
+
+file(GLOB dir_list LIST_DIRECTORIES true *)
+set(CK_DEVICE_INSTANCES)
+FOREACH(subdir_path ${dir_list})
+set(target_dir)
+IF(IS_DIRECTORY "${subdir_path}")
+    set(cmake_instance)
+    file(READ "${subdir_path}/CMakeLists.txt" cmake_instance)
+    set(add_inst 0)
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp8\" " AND DTYPES MATCHES "fp8")
+            #message("fp8 instance found!")
+            set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp16\"" AND DTYPES MATCHES "fp16")
+            #message("fp16 instance found!")
+            set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp32\"" AND DTYPES MATCHES "fp32")
+            #message("fp32 instance found!")
+            set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp64\"" AND DTYPES MATCHES "fp64")
+            #message("fp64 instance found!")
+            set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"bf16\"" AND DTYPES MATCHES "bf16")
+            #message("bf16 instance found!")
+            set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"int8\"" AND DTYPES MATCHES "int8")
+            #message("int8 instance found!")
+            set(add_inst 1)
+    endif()
+    if(NOT "${cmake_instance}" MATCHES "DTYPES" OR NOT DEFINED DTYPES)
+            #message("instance should be built for all types!")
+            set(add_inst 1)
+    endif()
+    if("${cmake_instance}" MATCHES "ONLY DL_KERNELS" AND NOT DEFINED DL_KERNELS)
+	    message("Found only dl instances, but DL_KERNELS is not set. Skipping.")
+            set(add_inst 0)
+    endif()
+    if(add_inst EQUAL 1)
+      get_filename_component(target_dir ${subdir_path} NAME)
+      add_subdirectory(${target_dir})
+      list(APPEND CK_DEVICE_INSTANCES $<TARGET_OBJECTS:device_${target_dir}_instance>)
+    endif()
+ENDIF()
+ENDFOREACH()
+
+add_library(device_operations STATIC ${CK_DEVICE_INSTANCES})
+add_library(composablekernels::device_operations ALIAS device_operations)
+
+
+set(DEV_OPS_INC_DIRS
+    ${PROJECT_SOURCE_DIR}/include/ck/
+    ${PROJECT_SOURCE_DIR}/library/include/ck/
+)
+
+target_compile_features(device_operations PUBLIC)
+set_target_properties(device_operations PROPERTIES POSITION_INDEPENDENT_CODE ON)
+target_include_directories(device_operations PUBLIC
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/utility>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_description>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/problem_transform>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/device>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/device/impl>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/grid>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/block>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/warp>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/thread>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/tensor_operation/gpu/element>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/utility>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu>
+    $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/tensor_operation_instance/gpu/reduce>
+)
+
+#once new arches are enabled make this an option on the main cmake file
+# and pass down here to be exported
+target_compile_options(device_operations PRIVATE
+    --offload-arch=gfx908
+    --offload-arch=gfx90a
+)
+
+# install(TARGETS device_operations LIBRARY DESTINATION lib)
+rocm_install(TARGETS device_operations
+        EXPORT device_operationsTargets)
+
+rocm_install(DIRECTORY ${DEV_OPS_INC_DIRS} DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/ck)
+rocm_install(EXPORT device_operationsTargets
+        FILE composable_kerneldevice_operationsTargets.cmake
+        NAMESPACE composable_kernel::
+        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
+)
diff --git a/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/CMakeLists.txt
new file mode 100644
index 0000000000..ec079e3ba3
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/CMakeLists.txt
@@ -0,0 +1,11 @@
+set(DEVICE_AVGPOOL_BWD_INSTANCES)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_AVGPOOL_BWD_INSTANCES device_avg_pool3d_bwd_ndhwc_f16_instance.cpp)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_AVGPOOL_BWD_INSTANCES device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_AVGPOOL_BWD_INSTANCES device_avg_pool3d_bwd_ndhwc_f32_instance.cpp)
+endif()
+add_instance_library(device_avg_pool3d_bwd_instance ${DEVICE_AVGPOOL_BWD_INSTANCES})
diff --git a/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/avg_pool3d_bwd_ndhwc_instance_common.hpp b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/avg_pool3d_bwd_ndhwc_instance_common.hpp
new file mode 100644
index 0000000000..c989bbcd3d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/avg_pool3d_bwd_ndhwc_instance_common.hpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using I32   = int32_t;
+using F16   = ck::half_t;
+using BF16  = ck::bhalf_t;
+using F32   = float;
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+using device_avgpool_bwd_ndhwc_f16_instances =
+    // clang-format off
+    std::tuple <
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F16, F16, F32, 256, 256, 1, 1, 1, 1>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F16, F16, F32, 256, 256, 1, 2, 2, 2>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F16, F16, F32, 256, 256, 1, 4, 4, 4>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F16, F16, F32, 256, 256, 1, 8, 8, 8>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F16, F16, F32, 256, 32, 8, 8, 8, 8>
+               // clang-format on
+               >;
+
+using device_avgpool_bwd_ndhwc_bf16_instances =
+    // clang-format off
+    std::tuple <
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<BF16, BF16, F32, 256, 256, 1, 1, 1, 1>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<BF16, BF16, F32, 256, 256, 1, 2, 2, 2>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<BF16, BF16, F32, 256, 256, 1, 4, 4, 4>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<BF16, BF16, F32, 256, 256, 1, 8, 8, 8>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<BF16, BF16, F32, 256, 32, 8, 8, 8, 8>
+               // clang-format on
+               >;
+
+using device_avgpool_bwd_ndhwc_f32_instances =
+    // clang-format off
+    std::tuple <
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F32, F32, F32, 256, 256, 1, 1, 1, 1>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F32, F32, F32, 256, 256, 1, 2, 2, 2>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F32, F32, F32, 256, 256, 1, 4, 4, 4>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F32, F32, F32, 256, 256, 1, 8, 8, 8>,
+        DeviceAvgPool3dBwd_NDHWC_NDHWC<F32, F32, F32, 256, 32, 8, 8, 8, 8>
+               // clang-format on
+               >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp
new file mode 100644
index 0000000000..52a8852f30
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "avg_pool3d_bwd_ndhwc_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_avgpool_bwd_ndhwc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, BF16, BF16, NDHWC, NDHWC>>>& instances)
+{
+    add_device_operation_instances(instances, device_avgpool_bwd_ndhwc_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp
new file mode 100644
index 0000000000..50de10e78e
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "avg_pool3d_bwd_ndhwc_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_avgpool_bwd_ndhwc_f16_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, F16, F16, NDHWC, NDHWC>>>& instances)
+{
+    add_device_operation_instances(instances, device_avgpool_bwd_ndhwc_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp
new file mode 100644
index 0000000000..0d4bb9a67f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "avg_pool3d_bwd_ndhwc_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_avgpool_bwd_ndhwc_f32_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, F32, F32, NDHWC, NDHWC>>>& instances)
+{
+    add_device_operation_instances(instances, device_avgpool_bwd_ndhwc_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..5b342595d8
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt
@@ -0,0 +1,26 @@
+set(BATCHED_GEMM_INSTANCES)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+   list(APPEND BATCHED_GEMM_INSTANCES device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
+                                      device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
+                                      device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
+                                      device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+   list(APPEND BATCHED_GEMM_INSTANCES device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp
+                                      device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp
+                                      device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp
+                                      device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+   list(APPEND BATCHED_GEMM_INSTANCES device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp
+                                      device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp
+                                      device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp
+                                      device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp)
+endif()
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+   list(APPEND BATCHED_GEMM_INSTANCES device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp
+                                      device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp
+                                      device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp
+                                      device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp)
+endif()
+add_instance_library(device_batched_gemm_instance ${BATCHED_GEMM_INSTANCES})
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp
new file mode 100644
index 0000000000..e730e9f589
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //##################| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##################|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##################|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##################|      |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  BF16,  BF16,  BF16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp
new file mode 100644
index 0000000000..f6696ffa9f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|        AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|         Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|             |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp
new file mode 100644
index 0000000000..32d6f258b5
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //#################|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //#################|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //#################|      |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   256,     4,  8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    32,    32,     4,  8,   32,   32,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp
new file mode 100644
index 0000000000..ee246ba56e
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp
@@ -0,0 +1,63 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //#################|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //#################|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //#################|      |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceBatchedGemmXdl< BF16,  BF16,  BF16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
new file mode 100644
index 0000000000..5a9483b309
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|        AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK|          LoopScheduler|                    Pipeline|
+        //##########|         Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar| Prefetch|                       |                            |
+        //##########|             |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector| Stage   |                       |                            |
+        //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+        // pipeline v1, 2 waves
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+        // pipeline v2, 1 wave
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
new file mode 100644
index 0000000000..0fa0719237
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|        AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK|          LoopScheduler|                    Pipeline|
+        //##########|         Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar| Prefetch|                       |                            |
+        //##########|             |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector| Stage   |                       |                            |
+        //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+        // pipeline v1, 2 waves
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+        // pipeline v2, 1 wave
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
new file mode 100644
index 0000000000..2f42f62b0b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
@@ -0,0 +1,110 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK|          LoopScheduler|                    Pipeline|
+        //#################|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar| Prefetch|                       |                            |
+        //#################|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector| Stage   |                       |                            |
+        //#################|      |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   256,     4,  8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    32,    32,     4,  8,   32,   32,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,   256,     4,  8,   16,   16,    1,    8,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,   128,     4,  8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,    64,     4,  8,   16,   16,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,    32,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   256,     4,  8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    32,    32,     4,  8,   32,   32,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,   256,     4,  8,   16,   16,    1,    8,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,   128,     4,  8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,    64,     4,  8,   16,   16,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,    32,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   256,     4,  8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    32,    32,     4,  8,   32,   32,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,   256,     4,  8,   16,   16,    1,    8,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,   128,     4,  8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,    64,     4,  8,   16,   16,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,    32,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
new file mode 100644
index 0000000000..10b4cea7d7
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
@@ -0,0 +1,98 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //#################| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumGemmK|          LoopScheduler|                    Pipeline|
+        //#################|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar| Prefetch|                       |                            |
+        //#################|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector| Stage   |                       |                            |
+        //#################|      |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES         
+        // pipeline v2, 1 wave
+        ,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp
new file mode 100644
index 0000000000..c687eb20f0
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|        AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|         Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|             |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp
new file mode 100644
index 0000000000..b19374ca6d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|        AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|         Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|             |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp
new file mode 100644
index 0000000000..bbd318ba92
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp
@@ -0,0 +1,58 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|        AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|         Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|             |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp
new file mode 100644
index 0000000000..187ccb5ff6
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp
@@ -0,0 +1,63 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|        AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|         Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|             |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp
new file mode 100644
index 0000000000..ec2b2646ff
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using AData   = int8_t;
+using BData   = int8_t;
+using CData   = int8_t;
+using AccData = int32_t;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|          AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer|   ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer|  BBlockTransfer|  BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|           Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|    ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|        DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|       SrcScalar|       DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|               |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave|  Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|     PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |       PerVector|    PerVector_K1|          |                |       PerVector|
+        //##########|               |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                 |               |               |               |               |                 |          |                |               |               |              |                |                |          |                |                |
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,               16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,               16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              1,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             1,              2,              16,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Col,
+                                                  Row,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp
new file mode 100644
index 0000000000..d76cd350c8
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using AData   = int8_t;
+using BData   = int8_t;
+using CData   = int8_t;
+using AccData = int32_t;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|          AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|           Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|               |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|               |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,               16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,               16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,               16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Col,
+                                                  Col,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp
new file mode 100644
index 0000000000..ef65106c2c
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using AData   = int8_t;
+using BData   = int8_t;
+using CData   = int8_t;
+using AccData = int32_t;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|          AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer|  ABlockTransfer|  ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer|  BBlockTransfer|  BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|           Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|       SrcScalar|       DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|       SrcScalar|       DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|               |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |       PerVector|    PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |       PerVector|    PerVector_K1|          |                |       PerVector|
+        //##########|               |      |      |        |        |        |        |            |            |            |      |      |      |      |    |     |     |     |     |                |               |               |               |                |                |          |                |               |               |              |                |                |          |                |                |
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,               16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,               16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,               16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,               16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,               16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,               16,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,               16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,               16,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              16,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Row,
+                                                  Row,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp
new file mode 100644
index 0000000000..078b241f9d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp
@@ -0,0 +1,71 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using AData   = int8_t;
+using BData   = int8_t;
+using CData   = int8_t;
+using AccData = int32_t;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        //##########|          AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer|  ABlockTransfer|  ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer|  BBlockTransfer|  BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|           Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|       SrcScalar|       DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|       SrcScalar|       DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|               |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |       PerVector|    PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |       PerVector|    PerVector_K1|          |                |       PerVector|
+        //##########|               |      |      |        |        |        |        |            |            |            |      |      |      |      |    |     |     |     |     |                |               |               |               |                |                |          |                |               |               |              |                |                |          |                |                |
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>,
+        DeviceBatchedGemmXdl<  AData, BData, CData, AccData,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              16,              16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              16,              16,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemm<Row,
+                                                  Col,
+                                                  Row,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt
new file mode 100644
index 0000000000..6710035ec6
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt
@@ -0,0 +1,6 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+add_instance_library(device_batched_gemm_add_relu_gemm_add_instance
+    device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
+    device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
+)
+endif()
\ No newline at end of file
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
new file mode 100644
index 0000000000..e750f18cc7
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough   = ck::tensor_operation::element_wise::PassThrough;
+using CDE0ElementOp = ck::tensor_operation::element_wise::AddRelu;
+using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =
+    std::tuple<
+        // clang-format off
+        //##################################################| A0Layout| B0Layout|       D0Layout| B1Layout|      D1sLayout| E1Layout| A0Data| B0Data| Acc0DataType|     D0DataType| B1Data| Acc1CData| CShuffle|        D1sData| E1Data|          A0|          B0|          CDE0|          B1|          CDE1| PadGemm0M| PadGemm0N| PadGemm0K| PadGemm1N| PadGemm1K|NumGemm0K| Block|  Gemm0| Gemm0| Gemm0| Gemm1| Gemm1|A0K1|B0K1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| CDE0BlockTransfer| CDE0BlockTransfer|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|   C1Shuffle|   C1Shuffle| CDE1BlockTransferClusterLengths| CDE1BlockTransfer|
+        //##################################################|         |         |               |         |               |         |   Type|   Type|         Type|           Type|   Type|      Type| DataType|           Type|   Type| Elementwise| Elementwise|   Elementwise| Elementwise|   Elementwise|          |          |          |          |          | Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|  ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|      SrcVectorDim|         SrcScalar|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MWaveMPerXdl|   ScalarPerVector|
+        //##################################################|         |         |               |         |               |         |       |       |             |               |       |          |         |               |       |   Operation|   Operation|     Operation|   Operation|     Operation|          |          |          |          |          |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per|Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |                  |         PerVector|  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|            _NBlock_NWaveNPerXdl|     _NWaveNPerXdl|
+        //##################################################|         |         |               |         |               |         |       |       |             |               |       |          |         |               |       |            |            |              |            |              |          |          |          |          |          |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|               |               |               |               |               |               |          |                 |                |                |                |                |                |           |                  |                  |                 |                |                |                |                |                |           |            |            |                                |                  |
+        //generic
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,      true,      true,      true,      true,      true,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 1,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        // no padding
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,                 9,                 4,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    32,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,                 9,                 4,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,    128,   128,    32,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,                  S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,                  S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,                  S<1, 32, 1, 8>,               8>,
+        // Padded fallback kernel
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,      true,      true,      true,      true,      true,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,    S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,                 9,                 4,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Row, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,      true,      true,      true,      true,      true,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,    S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,                  S<1, 32, 1, 8>,               8>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultipleDGemmMultipleD<Row,
+                                                                        Col,
+                                                                        ck::Tuple<Row>,
+                                                                        Row,
+                                                                        ck::Tuple<Row>,
+                                                                        Row,
+                                                                        F16,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        CDE0ElementOp,
+                                                                        PassThrough,
+                                                                        CDE1ElementOp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
new file mode 100644
index 0000000000..37342e5f45
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough   = ck::tensor_operation::element_wise::PassThrough;
+using CDE0ElementOp = ck::tensor_operation::element_wise::AddRelu;
+using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instances =
+    std::tuple<
+        // clang-format off
+        //##################################################| A0Layout| B0Layout|       D0Layout| B1Layout|      D1sLayout| E1Layout| A0Data| B0Data| Acc0DataType|     D0DataType| B1Data| Acc1CData| CShuffle|        D1sData| E1Data|          A0|          B0|          CDE0|          B1|          CDE1| PadGemm0M| PadGemm0N| PadGemm0K| PadGemm1N| PadGemm1K| NumGemm0K| Block|  Gemm0| Gemm0| Gemm0| Gemm1| Gemm1| A0K1| B0K1|B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockTransfer|A0BlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| CDE0BlockTransfer| CDE0BlockTransfer|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|   C1Shuffle|   C1Shuffle| CDE1BlockTransferClusterLengths| CDE1BlockTransfer|
+        //##################################################|         |         |               |         |               |         |   Type|   Type|         Type|           Type|   Type|      Type| DataType|           Type|   Type| Elementwise| Elementwise|   Elementwise| Elementwise|   Elementwise|          |          |          |          |          |  Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|     |     |    |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|      SrcVectorDim|         SrcScalar|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MWaveMPerXdl|   ScalarPerVector|
+        //##################################################|         |         |               |         |               |         |       |       |             |               |       |          |         |               |       |   Operation|   Operation|     Operation|   Operation|     Operation|          |          |          |          |          |     Stage|      |  Block| Block| Block| Block| Block|     |     |    |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |                  |         PerVector|  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|            _NBlock_NWaveNPerXdl|     _NWaveNPerXdl|
+        //##################################################|         |         |               |         |               |         |       |       |             |               |       |          |         |               |       |            |            |              |            |              |          |          |          |          |          |          |      |       |      |      |      |      |     |     |    |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                  |                  |                 |                |                |                |                |                |           |            |            |                                |                  |
+        //generic
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,      true,      true,      true,      true,      true,         1,   256,    128,    64,    32,   128,    32,    8,   8,    4,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 1,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        // no padding
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,    256,   128,    32,   128,    32,    8,   8,    4,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,    128,   128,    64,    64,    32,    8,   8,    4,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,    128,   128,    32,    64,    32,    8,   8,    4,   32,   32,     1,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,    128,   128,    64,   128,    32,    8,   8,    4,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,    128,   128,    32,   128,    32,    8,   8,    4,   32,   32,     1,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,     64,   256,    32,   128,    32,    8,   8,    4,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           8,                  S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,     64,   256,    32,    64,    32,    8,   8,    4,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           4,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,     64,   256,    64,   128,    32,    8,   8,    4,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           8,                  S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,     false,     false,     false,     false,     false,         1,   256,     64,   256,    64,    64,    32,    8,   8,    4,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           4,                  S<1, 32, 1, 8>,               8>,
+        // Padded fallback kernel                                                                                                                                                                                                       
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,      true,      true,      true,      true,      true,         1,   256,    128,   128,    64,   128,    32,    8,   8,    4,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,                 9,                 4,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,           1,           2,                  S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle<      Row,      Col, ck::Tuple<Row>,      Col, ck::Tuple<Row>,      Row,    F16,    F16,          F32, ck::Tuple<F16>,    F16,       F32,      F32, ck::Tuple<F16>,    F16, PassThrough, PassThrough, CDE0ElementOp, PassThrough, CDE1ElementOp,      true,      true,      true,      true,      true,         1,   256,    128,    64,    32,   128,    32,    8,   8,    4,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,                 9,                 4,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,                  S<1, 32, 1, 8>,               8>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultipleDGemmMultipleD<Row,
+                                                                        Col,
+                                                                        ck::Tuple<Row>,
+                                                                        Col,
+                                                                        ck::Tuple<Row>,
+                                                                        Row,
+                                                                        F16,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        ck::Tuple<F16>,
+                                                                        F16,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        CDE0ElementOp,
+                                                                        PassThrough,
+                                                                        CDE1ElementOp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/CMakeLists.txt
new file mode 100644
index 0000000000..b0f37e68fd
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/CMakeLists.txt
@@ -0,0 +1,5 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+add_instance_library(device_batched_gemm_bias_permute_instance
+    device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp
+)
+endif()
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp
new file mode 100644
index 0000000000..a0afaabbc7
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp
@@ -0,0 +1,98 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<F16>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+static constexpr auto ABSpec         = ck::tensor_operation::device::TensorSpecialization::Packed;
+static constexpr auto DESpec         = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// A[g0, m0, m1, k0] * B[g0, n0, n1, n2, k0] + D[g0, m0, m1, n0, n1, n2] = E[g0, n0, m0, n0, n1, m1]
+// m/n/n/n are the fast changing dimension for A/B/D/E
+using device_batched_contraction_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_mnnm_instance =
+    std::tuple<
+        // clang-format off
+        //############################################| NumDimG| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|    DsData| EData|            A|           B|         CDE|           GEMM|              A|              B|             DE| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //############################################|        |        |        |        |  Type|  Type|    Type| DataType|      Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Spacialization| Spacialization| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //############################################|        |        |        |        |      |      |        |         |          |      |    Operation|   Operation|   Operation|               |               |               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //############################################|        |        |        |        |      |      |        |         |          |      |             |            |            |               |               |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               1>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               1>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               1>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               1>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               1>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               1>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               1>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               1>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               1>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               1>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1>,
+        //M1 faster dim
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceBatchedContractionMultipleD_Xdl_CShuffle<       1,       2,       3,       1,   F16,   F16,     F32,      F16, F16_Tuple,   F16,  PassThrough, PassThrough,         Add, GemmMNKPadding,         ABSpec,         ABSpec,         DESpec,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+        // clang-format on
+        >;
+
+void add_device_batched_contraction_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_mnnm_instance(
+    std::vector<std::unique_ptr<DeviceBatchedContractionMultipleD<1,
+                                                                  2,
+                                                                  3,
+                                                                  1,
+                                                                  F16,
+                                                                  F16,
+                                                                  F16_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  Add>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_contraction_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_mnnm_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..cdb1a53901
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt
@@ -0,0 +1,6 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+add_instance_library(device_batched_gemm_gemm_instance
+    device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
+    device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
+)
+endif()
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
new file mode 100644
index 0000000000..67dfc4cd35
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances = std::tuple<
+    // clang-format off
+        //################################| ALayout| B0Layout| B1Layout| CLayout| AData| B0Data| B1Data| CData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //################################|        |         |         |        |  Type|   Type|   Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //################################|        |         |         |        |      |       |       |      |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //################################|        |         |         |        |      |       |       |      |        |         |            |            |            |            |            |               |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |
+        //################################|        |         |         |        |      |       |       |      |        |         |            |            |            |            |            |               |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    2,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   256,    32,    64,    32,   8,   8,    2,   32,   32,     1,     8,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   256,    32,   128,    32,   8,   8,    2,   32,   32,     1,     8,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    32,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    32,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8>,
+        // Padded fallback kernel
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmGemm<Row,
+                                                      Col,
+                                                      Row,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
new file mode 100644
index 0000000000..9001c901cf
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instances = std::tuple<
+    // clang-format off
+        //################################| ALayout| B0Layout| B1Layout| CLayout| AData| B0Data| B1Data| CData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //################################|        |         |         |        |  Type|   Type|   Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //################################|        |         |         |        |      |       |       |      |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //################################|        |         |         |        |      |       |       |      |        |         |            |            |            |            |            |               |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |
+        //################################|        |         |         |        |      |       |       |      |        |         |            |            |            |            |            |               |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    4,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    4,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   256,    32,    64,    32,   8,   8,    4,   32,   32,     1,     8,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   256,    32,   128,    32,   8,   8,    4,   32,   32,     1,     8,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    64,    64,    32,   8,   8,    4,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    32,    64,    32,   8,   8,    4,   32,   32,     1,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    4,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    32,   128,    32,   8,   8,    4,   32,   32,     1,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    32,   128,    32,   8,   8,    4,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           8,               S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    32,    64,    32,   8,   8,    4,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           4,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    64,   128,    32,   8,   8,    4,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           8,               S<1, 16, 1,16>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    64,    64,    32,   8,   8,    4,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           4,               S<1, 32, 1, 8>,               8>,
+        // Padded fallback kernel
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   128,    64,   128,    32,   8,   8,    4,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,           1,           2,               S<1, 32, 1, 8>,               8>,
+        DeviceBatchedGemmGemm_Xdl_CShuffle<     Row,      Col,      Col,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,    64,    32,   128,    32,   8,   8,    4,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,       true,           1,           2,               S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmGemm<Row,
+                                                      Col,
+                                                      Col,
+                                                      Row,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      F16,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough,
+                                                      PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt
new file mode 100644
index 0000000000..444c93b118
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt
@@ -0,0 +1,25 @@
+# ONLY DL_KERNELS
+if(DL_KERNELS)
+  set(BATCHED_GEMM_MULTID_INSTANCES)
+  if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp)
+  endif()
+  if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp)
+    list(APPEND BATCHED_GEMM_MULTID_INSTANCES device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp)
+  endif()
+  add_instance_library(device_batched_gemm_multi_d_instance ${BATCHED_GEMM_MULTID_INSTANCES})
+endif()
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
new file mode 100644
index 0000000000..b20e8a86b1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
@@ -0,0 +1,95 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout| AData| BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |  Type|  Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |      |      |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |      |      |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<8, 2>,       S<4, 2>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 4>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<4, 2>,       S<8, 2>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<2, 4>,       S<2, 8>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<4, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,       S<16, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<4, 1, 4, 2>,        S<4, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<1, 4>,       S<1, 4>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<4, 1, 4, 2>,        S<4, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<8, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<8, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,       S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,       S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp
new file mode 100644
index 0000000000..e98fd24eeb
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout| AData| BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |  Type|  Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |      |      |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |      |      |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        // MPerBlock=128, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    16,    64,    16,  2,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<4, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,       S<16, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<8, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<8, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,       S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,       S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
new file mode 100644
index 0000000000..75042280f0
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
@@ -0,0 +1,95 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout| AData| BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |  Type|  Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |      |      |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |      |      |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128       
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<8, 2>,       S<4, 2>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 4>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<4, 2>,       S<8, 2>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<2, 4>,       S<2, 8>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<1, 4>,       S<1, 4>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp
new file mode 100644
index 0000000000..c8e1acfd4f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout| AData| BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |  Type|  Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |      |      |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |      |      |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128       
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64       
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    16,    64,    16,  2,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
new file mode 100644
index 0000000000..f1ed7a6942
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
@@ -0,0 +1,95 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout| AData| BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |  Type|  Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |      |      |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |      |      |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<8, 2>,       S<4, 2>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 4>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<4, 2>,       S<8, 2>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<2, 4>,       S<2, 8>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<1, 4>,       S<1, 4>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,        S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,        S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp
new file mode 100644
index 0000000000..6e25517ed8
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout| AData| BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |  Type|  Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |      |      |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |      |      |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    16,    64,    16,  2,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,       S<4, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
new file mode 100644
index 0000000000..5a9fd16c83
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
@@ -0,0 +1,95 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout| AData| BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |  Type|  Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |      |      |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |      |      |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // // MPerBlock=128, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<8, 2>,       S<4, 2>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 4>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // // MPerBlock=64, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<4, 2>,       S<8, 2>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<2, 4>,       S<2, 8>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 1, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<1, 4>,       S<1, 4>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp
new file mode 100644
index 0000000000..cf5502926b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout| AData| BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |  Type|  Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |      |      |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |      |      |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    16,    64,    16,  2,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 1, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row,   F16,   F16,     F32, Empty_Tuple, F16, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        F16,
+                                                        F16,
+                                                        Empty_Tuple,
+                                                        F16,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp
new file mode 100644
index 0000000000..43cdcc5f14
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp
@@ -0,0 +1,93 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout|  AData|  BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |   Type|   Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |       |       |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |       |       |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,    64,   128,    16,  4,          2,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  4,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<1, 1, 4, 4>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<4, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  4,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<1, 1, 4, 4>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<4, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  4,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  4,          4,          1,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  4,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 4>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<4, 1, 4, 4>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  4,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<1, 1, 4, 4>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<8, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  4,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 4>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<8, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  4,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  4,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  4,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  4,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  4,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  4,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp
new file mode 100644
index 0000000000..617877499d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp
@@ -0,0 +1,90 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout|  AData|  BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |   Type|   Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |       |       |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |       |       |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<4, 4>,       S<4, 2>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,    32,    32,    32,     8,  4,          2,          4,      1,       S<4, 2>,       S<2, 2>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<4, 1, 8, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<4, 2>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<4, 1, 8, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<4, 1, 8, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,    16,    16,    16,    16,  4,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 4>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 4, 4>,       S<4, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,    16,    16,    16,    16,  4,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 4>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 4, 4>,       S<4, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,     8,    64,    32,  4,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 4>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 4, 4>,      S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,     8,    32,  4,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<1, 1, 4, 4>,      S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,     8,     8,     8,     4,  4,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<1, 1, 4, 4>,       S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,     8,     8,     8,     4,  4,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<1, 1, 4, 4>,       S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,     8,     8,     8,     4,  4,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<1, 1, 4, 4>,       S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough, GemmMNPadding,     8,     8,     8,     4,  4,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<1, 1, 4, 4>,       S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp
new file mode 100644
index 0000000000..1f58aecf50
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp
@@ -0,0 +1,93 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout|  AData|  BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |   Type|   Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |       |       |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |       |       |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,    64,   128,    16,  4,          2,          4,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,      S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp
new file mode 100644
index 0000000000..e0ead0ec49
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp
@@ -0,0 +1,90 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout|  AData|  BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |   Type|   Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |       |       |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |       |       |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<4, 4>,       S<4, 2>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          2,          4,      1,       S<4, 2>,       S<2, 2>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,      S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<4, 2>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,      S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,      S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    16,    16,    16,    16,  2,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    16,    16,    16,    16,  2,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Col,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp
new file mode 100644
index 0000000000..55fa86f37f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp
@@ -0,0 +1,93 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout|  AData|  BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |   Type|   Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |       |       |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |       |       |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,    64,   128,    16,  4,          2,          4,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 8, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,        S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,        S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp
new file mode 100644
index 0000000000..b8af3e61a3
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp
@@ -0,0 +1,90 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout|  AData|  BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |   Type|   Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |       |       |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |       |       |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<4, 4>,       S<4, 2>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          2,          4,      1,       S<4, 2>,       S<2, 2>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 8, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 8, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 8, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    16,    16,    16,    16,  2,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    16,    16,    16,    16,  2,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Row, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Row,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp
new file mode 100644
index 0000000000..29d20e58dd
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp
@@ -0,0 +1,93 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout|  AData|  BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |   Type|   Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |       |       |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |       |       |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // // MPerBlock=128, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // // MPerBlock=64, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,    64,   128,    16,  4,          2,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 1, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 1, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp
new file mode 100644
index 0000000000..dd0f8749c0
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp
@@ -0,0 +1,90 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Empty_Tuple = ck::Tuple<>;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances = std::tuple<
+    // clang-format off
+        // ##########################| ALayout| BLayout|    DsLayout| CLayout|  AData|  BData| AccData|      DsData| CData|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########################|        |        |            |        |   Type|   Type|    Type|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########################|        |        |            |        |       |       |        |            |      |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // ##########################|        |        |            |        |       |       |        |            |      |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<4, 4>,       S<4, 2>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // // MPerBlock=128, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // // MPerBlock=64, NPerBlock=128
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          2,          4,      1,       S<4, 2>,       S<2, 2>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    16,    16,    16,    16,  2,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    16,    16,    16,    16,  2,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceBatchedGemmMultipleD_Dl<     Row,     Col, Empty_Tuple,     Row, int8_t, int8_t, int32_t, Empty_Tuple, int8_t, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Row,
+                                                        Col,
+                                                        Empty_Tuple,
+                                                        Row,
+                                                        int8_t,
+                                                        int8_t,
+                                                        Empty_Tuple,
+                                                        int8_t,
+                                                        PassThrough,
+                                                        PassThrough,
+                                                        PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt
new file mode 100644
index 0000000000..728e35fc3b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt
@@ -0,0 +1,8 @@
+if(DTYPES MATCHES "fp16" OR DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+add_instance_library(device_batched_gemm_reduce_instance
+    device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
+    device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
+    device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
+    device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
+)
+endif()
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
new file mode 100644
index 0000000000..cb89d3cefd
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Identity, Identity>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|      Reduce|      ReduceInEleOp|      ReduceAccEleOp|        Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise|   Operation|                   |                    |    MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|            |                   |                    |     Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |            |            |            |            |                   |                    |              |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   4,   4,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
new file mode 100644
index 0000000000..91eefba0c1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Identity, Identity>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|      Reduce|      ReduceInEleOp|      ReduceAccEleOp|        Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise|   Operation|                   |                    |    MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|            |                   |                    |     Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |            |            |            |            |                   |                    |              |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
new file mode 100644
index 0000000000..c20798f557
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Identity, Identity>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|      Reduce|      ReduceInEleOp|      ReduceAccEleOp|        Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise|   Operation|                   |                    |    MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|            |                   |                    |     Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |            |            |            |            |                   |                    |              |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
new file mode 100644
index 0000000000..3d9ad64b9b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
@@ -0,0 +1,77 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Identity, Identity>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|      Reduce|      ReduceInEleOp|      ReduceAccEleOp|        Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise|   Operation|                   |                    |    MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|            |                   |                    |     Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |         |         |          |                   |            |            |            |            |                   |                    |              |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>,
+        DeviceBatchedGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps,   ReduceMemOp,   GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..5ac55655d4
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt
@@ -0,0 +1,5 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+add_instance_library(device_batched_gemm_softmax_gemm_instance
+    device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
+)
+endif()
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
new file mode 100644
index 0000000000..cf23d01bf2
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
@@ -0,0 +1,131 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+template <bool Masking>
+using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =
+    std::tuple<
+        // clang-format off
+        //#######################################| ALayout| B0Layout| B1Layout| CLayout| AData| B0Data| B1Data| CData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|    MaskOut|
+        //#######################################|        |         |         |        |  Type|   Type|   Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|      Upper|
+        //#######################################|        |         |         |        |      |       |       |      |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|   Triangle|
+        //#######################################|        |         |         |        |      |       |       |      |        |         |            |            |            |            |            |               |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |           |
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    2,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   256,    32,    64,    32,   8,   8,    2,   32,   32,     1,     8,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   256,    32,   128,    32,   8,   8,    2,   32,   32,     1,     8,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    32,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   128,    32,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8,   Masking>,
+        // Padded fallback kernel
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>
+        // clang-format on
+        >;
+
+template <bool Masking>
+using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_irregular_k_instances =
+    std::tuple<
+        // clang-format off
+        //#######################################| ALayout| B0Layout| B1Layout| CLayout| AData| B0Data| B1Data| CData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|    MaskOut|
+        //#######################################|        |         |         |        |  Type|   Type|   Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|      Upper|
+        //#######################################|        |         |         |        |      |       |       |      |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|   Triangle|
+        //#######################################|        |         |         |        |      |       |       |      |        |         |            |            |            |            |            |               |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |           |
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    256,   128,    40,    64,    32,   4,   4,    2,   32,   32,     2,     4,     2,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    256,   128,    40,   128,    32,   4,   4,    2,   32,   32,     2,     4,     4,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   256,    40,    64,    32,   4,   4,    2,   32,   32,     1,     8,     2,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   256,    40,   128,    32,   4,   4,    2,   32,   32,     1,     8,     4,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   128,    40,    64,    32,   4,   4,    2,   32,   32,     1,     4,     2,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
+        DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   128,    40,   128,    32,   4,   4,    2,   32,   32,     1,     4,     4,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmSoftmaxGemm<Row,
+                                                             Col,
+                                                             Row,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             Scale,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             false>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
+            false>{});
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_irregular_k_instances<
+            false>{});
+}
+
+void add_device_batched_gemm_masking_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
+    std::vector<std::unique_ptr<DeviceBatchedGemmSoftmaxGemm<Row,
+                                                             Col,
+                                                             Row,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             Scale,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             true>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
+            true>{});
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_irregular_k_instances<
+            true>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt
new file mode 100644
index 0000000000..d5110e4505
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt
@@ -0,0 +1,11 @@
+set(DEVICE_BATCHED_GEMM_SOFTMAX_GEMM_PERMUTE_INSTANCES)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_BATCHED_GEMM_SOFTMAX_GEMM_PERMUTE_INSTANCES device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp)
+    list(APPEND DEVICE_BATCHED_GEMM_SOFTMAX_GEMM_PERMUTE_INSTANCES device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_BATCHED_GEMM_SOFTMAX_GEMM_PERMUTE_INSTANCES device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp)
+    list(APPEND DEVICE_BATCHED_GEMM_SOFTMAX_GEMM_PERMUTE_INSTANCES device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp)
+endif()
+add_instance_library(device_batched_gemm_softmax_gemm_permute_instance ${DEVICE_BATCHED_GEMM_SOFTMAX_GEMM_PERMUTE_INSTANCES})
+
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
new file mode 100644
index 0000000000..498bf58fb3
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
@@ -0,0 +1,135 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ScaleAdd    = ck::tensor_operation::element_wise::ScaleAdd;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          MaskingSpecialization MaskingSpec>
+using device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances =
+    std::tuple<
+        // clang-format off
+        // #############################################|  NumDimG| NumDimM| NumDimN| NumDimK| NumDimO|  AData|  B0Data|  B1Data|  CData|     Acc0BiasData| Acc1BiasData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM|   ATensorSpec|  B0TensorSpec|  B1TensorSpec|   CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer| MaskingSpec|  D0s Bias|
+        // #############################################|         |        |        |        |        |   Type|    Type|    Type|   Type|             Type|         Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization|              |              |              |              | Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|            | SrcScalar|
+        // #############################################|         |        |        |        |        |       |        |        |       |                 |             |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |              |              |              |              |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            | PerVector|
+        // #############################################|         |        |        |        |        |       |        |        |       |                 |             |        |         |            |            |            |            |            |               |              |              |              |              |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |            |          |
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec,        1>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    2,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,    64,    32,   8,   8,    2,   32,   32,     1,     8,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,   128,    32,   8,   8,    2,   32,   32,     1,     8,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    32,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    32,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        // Padded fallback kernel
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec,        1>,  
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            ck::Tuple<BF16>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            ScaleAdd,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskOutUpperTriangle>{});
+}
+
+void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            ck::Tuple<BF16>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            ScaleAdd,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskDisabled>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
new file mode 100644
index 0000000000..744bd6456d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
@@ -0,0 +1,137 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ScaleAdd    = ck::tensor_operation::element_wise::ScaleAdd;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          MaskingSpecialization MaskingSpec>
+using device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =
+    std::tuple<
+        // clang-format off
+        // #############################################|  NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| AData| B0Data| B1Data| CData|    Acc0BiasData| Acc1BiasData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM|   ATensorSpec|  B0TensorSpec|  B1TensorSpec|   CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer| MaskingSpec|  D0s Bias|
+        // #############################################|         |        |        |        |        |  Type|   Type|   Type|  Type|            Type|         Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization|              |              |              |              | Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|            | SrcScalar|
+        // #############################################|         |        |        |        |        |      |       |       |      |                |             |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |              |              |              |              |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            | PerVector|
+        // #############################################|         |        |        |        |        |      |       |       |      |                |             |        |         |            |            |            |            |            |               |              |              |              |              |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |            |          |
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec,       1>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    2,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+ #if CK_WORKAROUND_SWDEV_388832
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,    64,    32,   8,   8,    2,   32,   32,     1,     8,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+ #endif
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,   128,    32,   8,   8,    2,   32,   32,     1,     8,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    32,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    32,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        // Padded fallback kernel
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec,       1>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            ck::Tuple<F16>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            ScaleAdd,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskOutUpperTriangle>{});
+}
+
+void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            ck::Tuple<F16>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            ScaleAdd,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskDisabled>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
new file mode 100644
index 0000000000..b342612d1c
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
@@ -0,0 +1,133 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          MaskingSpecialization MaskingSpec>
+using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances =
+    std::tuple<
+        // clang-format off
+        // #############################################|  NumDimG| NumDimM| NumDimN| NumDimK| NumDimO|  AData|  B0Data|  B1Data|  CData| Acc0BiasData| Acc1BiasData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM|   ATensorSpec|  B0TensorSpec|  B1TensorSpec|   CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer| MaskingSpec|
+        // #############################################|         |        |        |        |        |   Type|    Type|    Type|   Type|         Type|         Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization|              |              |              |              | Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|            |
+        // #############################################|         |        |        |        |        |       |        |        |       |             |             |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |              |              |              |              |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            |
+        // #############################################|         |        |        |        |        |       |        |        |       |             |             |        |         |            |            |            |            |            |               |              |              |              |              |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |            |
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    2,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,    64,    32,   8,   8,    2,   32,   32,     1,     8,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,   128,    32,   8,   8,    2,   32,   32,     1,     8,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    32,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    32,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        // Padded fallback kernel
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            ck::Tuple<>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            Scale,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskOutUpperTriangle>{});
+}
+
+void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            ck::Tuple<>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            Scale,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskDisabled>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
new file mode 100644
index 0000000000..3fd0c07370
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
@@ -0,0 +1,135 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+
+static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
+
+// c[g, m, n] = a[g, m, k] * b[g, n, k]
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          MaskingSpecialization MaskingSpec>
+using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances =
+    std::tuple<
+        // clang-format off
+        // #############################################|  NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| AData| B0Data| B1Data| CData| Acc0BiasData| Acc1BiasData| AccData| CShuffle|           A|          B0|        Acc0|          B1|           C|           GEMM|   ATensorSpec|  B0TensorSpec|  B1TensorSpec|   CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds|  B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer| MaskingSpec|
+        // #############################################|         |        |        |        |        |  Type|   Type|   Type|  Type|         Type|         Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization|              |              |              |              | Prefetch|  Size|   MPer|  NPer|  KPer|  NPer|  KPer|    |    |     |  XDL|  XDL|  MXdl|  NXdl|  NXdl|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN|    ThreadCluster|   ThreadCluster|  SrcAccessOrder|    SrcVectorDim|       SrcScalar|       DstScalar|  AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|            |
+        // #############################################|         |        |        |        |        |      |       |       |      |             |             |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |              |              |              |              |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            |
+        // #############################################|         |        |        |        |        |      |       |       |      |             |             |        |         |            |            |            |            |            |               |              |              |              |              |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |            |
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    2,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+ #if CK_WORKAROUND_SWDEV_388832
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,    64,    32,   8,   8,    2,   32,   32,     1,     8,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+ #endif
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   256,    32,   128,    32,   8,   8,    2,   32,   32,     1,     8,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    32,    64,    32,   8,   8,    2,   32,   32,     1,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    32,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    32,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,   128,    32,   8,   8,    2,   16,   16,     1,    16,     8,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           8,               S<1, 16, 1,16>,               8, MaskingSpec>,
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
+        // Padded fallback kernel
+        DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>
+        // clang-format on
+        >;
+
+void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            ck::Tuple<>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            Scale,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskOutUpperTriangle>{});
+}
+
+void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            ck::Tuple<>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            Scale,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances<
+            2,
+            1,
+            1,
+            1,
+            1,
+            MaskingSpecialization::MaskDisabled>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt
new file mode 100644
index 0000000000..19a3cc8cd1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt
@@ -0,0 +1,14 @@
+add_instance_library(device_batchnorm_instance
+    device_batchnorm_forward_f16_instance.cpp
+    device_batchnorm_forward_f32_instance.cpp
+    device_batchnorm_forward_bf16_instance.cpp
+    device_batchnorm_forward_f64_instance.cpp
+    device_batchnorm_backward_f16_instance.cpp
+    device_batchnorm_backward_f32_instance.cpp
+    device_batchnorm_backward_bf16_instance.cpp
+    device_batchnorm_backward_f64_instance.cpp
+    device_batchnorm_infer_f16_instance.cpp
+    device_batchnorm_infer_f32_instance.cpp
+    device_batchnorm_infer_bf16_instance.cpp
+    device_batchnorm_infer_f64_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp
new file mode 100644
index 0000000000..bc95d2f1b1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp
@@ -0,0 +1,146 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_bf16_blockwise_instances =
+     std::tuple <
+        // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, DscaleDbiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcVectorSize, DscaleDbiasDstVectorSize, MeanVarSrcVectorSize 
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+     >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_bf16_multiblock_instances =
+     std::tuple <
+        // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcDstVectorSize, BiasDstVectorSize, MeanVarSrcVectorSize 
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<BF16, F32, F32, F32, BF16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_backward_rank_4_3_bf16_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchNormBwd<BF16, F32, F32, F32, BF16, F32, F32, PassThrough, 4, 3>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_bf16_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_bf16_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp
new file mode 100644
index 0000000000..fbc8d0bc60
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp
@@ -0,0 +1,147 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_f16_blockwise_instances =
+     std::tuple <
+        // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, DscaleDbiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcVectorSize, DscaleDbiasDstVectorSize, MeanVarSrcVectorSize 
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,  
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+     >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_f16_multiblock_instances =
+     std::tuple <
+        // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcDstVectorSize, BiasDstVectorSize, MeanVarSrcVectorSize 
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F16, F32, F32, F32, F16, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_backward_rank_4_3_f16_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormBwd<F16, F32, F32, F32, F16, F32, F32, PassThrough, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_f16_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_f16_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp
new file mode 100644
index 0000000000..bed38658a9
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp
@@ -0,0 +1,145 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_f32_blockwise_instances = std::tuple<
+        // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, DscaleDbiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcVectorSize, DscaleDbiasDstVectorSize, MeanVarSrcVectorSize 
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+    >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_f32_multiblock_instances =
+     std::tuple <
+        // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcDstVectorSize, BiasDstVectorSize, MeanVarSrcVectorSize 
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F32, F32, F32, F32, F32, F32, F32, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_backward_rank_4_3_f32_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormBwd<F32, F32, F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_f32_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_f32_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp
new file mode 100644
index 0000000000..fc5ec77e42
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp
@@ -0,0 +1,145 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64 = double;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_f64_blockwise_instances = std::tuple<
+      // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, DscaleDbiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcVectorSize, DscaleDbiasDstVectorSize, MeanVarSrcVectorSize 
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+      DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+    >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename DyElementwiseOp>
+using device_batchnorm_backward_f64_multiblock_instances =
+     std::tuple <
+        // XDataType, DxDataType, DyDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, DyElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XDyDxVectorDim, XSrcVectorSize, DySrcVectorSize, DxDstVectorSize, ScaleSrcDstVectorSize, BiasDstVectorSize, MeanVarSrcVectorSize 
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,  2,  2,    1,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  1,  1,  1,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    0,  2,  2,  2,    1,  1,  1>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    2,  2,  2>,
+        DeviceBatchNormBwdImpl<F64, F64, F64, F64, F64, F64, F64, DyElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,  2,  2,    1,  1,  1,  1,    1,  1,  1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_backward_rank_4_3_f64_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormBwd<F64, F64, F64, F64, F64, F64, F64, PassThrough, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_f64_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_backward_f64_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_bf16_instance.cpp
new file mode 100644
index 0000000000..4e38ee13b2
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_bf16_instance.cpp
@@ -0,0 +1,147 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_bf16_blockwise_instances =
+     std::tuple <
+        // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize 
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+     >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_bf16_multiblock_instances =
+     std::tuple <
+        // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<BF16, BF16, F32, BF16, BF16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_forward_rank_4_3_bf16_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<BF16, BF16, F32, BF16, BF16, F32, PassThrough, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_bf16_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_bf16_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f16_instance.cpp
new file mode 100644
index 0000000000..f087eb7982
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f16_instance.cpp
@@ -0,0 +1,147 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_f16_blockwise_instances =
+     std::tuple <
+        // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize 
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,  
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+     >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_f16_multiblock_instances =
+     std::tuple <
+        // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F16, F16, F32, F16, F16, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_forward_rank_4_3_f16_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<F16, F16, F32, F16, F16, F32, PassThrough, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_f16_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_f16_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f32_instance.cpp
new file mode 100644
index 0000000000..d0f361401a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f32_instance.cpp
@@ -0,0 +1,145 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_f32_blockwise_instances = std::tuple<
+    // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+    >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_f32_multiblock_instances =
+     std::tuple <
+        // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F32, F32, F32, F32, F32, F32, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_forward_rank_4_3_f32_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<F32, F32, F32, F32, F32, F32, PassThrough, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_f32_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_f32_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f64_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f64_instance.cpp
new file mode 100644
index 0000000000..710d07b828
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f64_instance.cpp
@@ -0,0 +1,145 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64 = double;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_f64_blockwise_instances = std::tuple<
+    // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, false, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+    >;
+// clang-format on
+
+// clang-format off
+template <index_t Rank, index_t NumReduceDim, typename YElementwiseOp>
+using device_batchnorm_forward_f64_multiblock_instances =
+     std::tuple <
+        // XDataType, YDataType, AccDataType, ScaleDataType, BiasDataType, MeanVarDataType, YElementwiseOp, Rank, NumReduceDim, UseMultiBlockInK, BLockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XSrcYDstVectorDim, XSrcVectorSize, YDstVectorSize, ScaleSrcVectorSize, BiasSrcVectorSize, MeanVarSrcDstVectorSize
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 128, 2,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 64,  4,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 32,  8,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 16, 16,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 8,  32,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 4,  64,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 2, 128,    2,    2,    1,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    1,    1,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    0,    2,    2,    1,    1,    1>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    2,    2,    2>,
+        DeviceBatchNormFwdImpl<F64, F64, F64, F64, F64, F64, YElementwiseOp, Rank, NumReduceDim, true, 256, 1, 256,    2,    2,    1,    1,    1,    1,    1,    1>
+     >;
+// clang-format on
+
+void add_device_batchnorm_forward_rank_4_3_f64_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchNormFwd<F64, F64, F64, F64, F64, F64, PassThrough, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_f64_blockwise_instances<4, 3, PassThrough>{});
+    add_device_operation_instances(
+        instances, device_batchnorm_forward_f64_multiblock_instances<4, 3, PassThrough>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_bf16_instance.cpp
new file mode 100644
index 0000000000..8801c309f9
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_bf16_instance.cpp
@@ -0,0 +1,55 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Normalize = ck::tensor_operation::element_wise::NormalizeInInfer;
+
+// clang-format off
+template <index_t Rank>
+using device_batchnorm_infer_bf16_instances =
+     std::tuple <
+        // Tuple<XDataType, MeanDataType, VarDataType, ScaleDataType, BiasDataType>, Tuple<YDataType>, NormalizeOp, Rank, MPerThread, Sequence<XVectorSize, MeanDataType, VarDataType, ScaleVectorSize, BiasVectorSize>, Sequence<YVectorSize> 
+        DeviceElementwiseImpl<Tuple<BF16, F32, F32, BF16, BF16>, Tuple<BF16>, Normalize, Rank, 1, Sequence<1, 1, 1, 1, 1>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<BF16, F32, F32, BF16, BF16>, Tuple<BF16>, Normalize, Rank, 2, Sequence<1, 1, 1, 1, 1>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<BF16, F32, F32, BF16, BF16>, Tuple<BF16>, Normalize, Rank, 2, Sequence<2, 1, 1, 1, 1>, Sequence<2> >,
+        DeviceElementwiseImpl<Tuple<BF16, F32, F32, BF16, BF16>, Tuple<BF16>, Normalize, Rank, 2, Sequence<1, 2, 2, 2, 2>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<BF16, F32, F32, BF16, BF16>, Tuple<BF16>, Normalize, Rank, 2, Sequence<2, 2, 2, 2, 2>, Sequence<2> >,
+        DeviceElementwiseImpl<Tuple<BF16, F32, F32, BF16, BF16>, Tuple<BF16>, Normalize, Rank, 4, Sequence<1, 1, 1, 1, 1>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<BF16, F32, F32, BF16, BF16>, Tuple<BF16>, Normalize, Rank, 4, Sequence<2, 1, 1, 1, 1>, Sequence<2> >,
+        DeviceElementwiseImpl<Tuple<BF16, F32, F32, BF16, BF16>, Tuple<BF16>, Normalize, Rank, 4, Sequence<1, 2, 2, 2, 2>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<BF16, F32, F32, BF16, BF16>, Tuple<BF16>, Normalize, Rank, 4, Sequence<2, 2, 2, 2, 2>, Sequence<2> >,
+        DeviceElementwiseImpl<Tuple<BF16, F32, F32, BF16, BF16>, Tuple<BF16>, Normalize, Rank, 4, Sequence<4, 1, 1, 1, 1>, Sequence<4> >,
+        DeviceElementwiseImpl<Tuple<BF16, F32, F32, BF16, BF16>, Tuple<BF16>, Normalize, Rank, 4, Sequence<1, 4, 4, 4, 4>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<BF16, F32, F32, BF16, BF16>, Tuple<BF16>, Normalize, Rank, 4, Sequence<4, 2, 2, 2, 2>, Sequence<4> >,
+        DeviceElementwiseImpl<Tuple<BF16, F32, F32, BF16, BF16>, Tuple<BF16>, Normalize, Rank, 4, Sequence<2, 4, 4, 4, 4>, Sequence<2> >,
+        DeviceElementwiseImpl<Tuple<BF16, F32, F32, BF16, BF16>, Tuple<BF16>, Normalize, Rank, 4, Sequence<4, 4, 4, 4, 4>, Sequence<4> >
+     >;
+// clang-format on
+
+void add_device_batchnorm_infer_rank_4_bf16_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<Tuple<BF16, F32, F32, BF16, BF16>, Tuple<BF16>, Normalize, 4>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_batchnorm_infer_bf16_instances<4>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f16_instance.cpp
new file mode 100644
index 0000000000..b674cfc423
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f16_instance.cpp
@@ -0,0 +1,54 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Normalize = ck::tensor_operation::element_wise::NormalizeInInfer;
+
+// clang-format off
+template <index_t Rank>
+using device_batchnorm_infer_f16_instances =
+     std::tuple <
+        // Tuple<XDataType, MeanDataType, VarDataType, ScaleDataType, BiasDataType>, Tuple<YDataType>, NormalizeOp, Rank, MPerThread, Sequence<XVectorSize, MeanDataType, VarDataType, ScaleVectorSize, BiasVectorSize>, Sequence<YVectorSize> 
+        DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, Rank, 1, Sequence<1, 1, 1, 1, 1>, Sequence<1> >,  
+        DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, Rank, 2, Sequence<1, 1, 1, 1, 1>, Sequence<1> >,  
+        DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, Rank, 2, Sequence<2, 1, 1, 1, 1>, Sequence<2> >,  
+        DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, Rank, 2, Sequence<1, 2, 2, 2, 2>, Sequence<1> >,  
+        DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, Rank, 2, Sequence<2, 2, 2, 2, 2>, Sequence<2> >,  
+        DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, Rank, 4, Sequence<1, 1, 1, 1, 1>, Sequence<1> >,  
+        DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, Rank, 4, Sequence<2, 1, 1, 1, 1>, Sequence<2> >,  
+        DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, Rank, 4, Sequence<1, 2, 2, 2, 2>, Sequence<1> >,  
+        DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, Rank, 4, Sequence<2, 2, 2, 2, 2>, Sequence<2> >,  
+        DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, Rank, 4, Sequence<4, 1, 1, 1, 1>, Sequence<4> >,  
+        DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, Rank, 4, Sequence<1, 4, 4, 4, 4>, Sequence<1> >,  
+        DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, Rank, 4, Sequence<4, 2, 2, 2, 2>, Sequence<4> >,  
+        DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, Rank, 4, Sequence<2, 4, 4, 4, 4>, Sequence<2> >,  
+        DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, Rank, 4, Sequence<4, 4, 4, 4, 4>, Sequence<4> >  
+     >;
+// clang-format on
+
+void add_device_batchnorm_infer_rank_4_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, 4>>>& instances)
+{
+    add_device_operation_instances(instances, device_batchnorm_infer_f16_instances<4>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f32_instance.cpp
new file mode 100644
index 0000000000..05e3650887
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f32_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/utility/tuple.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Normalize = ck::tensor_operation::element_wise::NormalizeInInfer;
+
+// clang-format off
+template <index_t Rank>
+using device_batchnorm_infer_f32_instances =
+     std::tuple <
+        // Tuple<XDataType, MeanDataType, VarDataType, ScaleDataType, BiasDataType>, Tuple<YDataType>, NormalizeOp, Rank, MPerThread, Sequence<XVectorSize, MeanDataType, VarDataType, ScaleVectorSize, BiasVectorSize>, Sequence<YVectorSize> 
+        DeviceElementwiseImpl<Tuple<F32, F32, F32, F32, F32>, Tuple<F32>, Normalize, Rank, 1, Sequence<1, 1, 1, 1, 1>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<F32, F32, F32, F32, F32>, Tuple<F32>, Normalize, Rank, 2, Sequence<1, 1, 1, 1, 1>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<F32, F32, F32, F32, F32>, Tuple<F32>, Normalize, Rank, 2, Sequence<2, 1, 1, 1, 1>, Sequence<2> >,
+        DeviceElementwiseImpl<Tuple<F32, F32, F32, F32, F32>, Tuple<F32>, Normalize, Rank, 2, Sequence<1, 2, 2, 2, 2>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<F32, F32, F32, F32, F32>, Tuple<F32>, Normalize, Rank, 2, Sequence<2, 2, 2, 2, 2>, Sequence<2> >,
+        DeviceElementwiseImpl<Tuple<F32, F32, F32, F32, F32>, Tuple<F32>, Normalize, Rank, 4, Sequence<1, 1, 1, 1, 1>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<F32, F32, F32, F32, F32>, Tuple<F32>, Normalize, Rank, 4, Sequence<2, 1, 1, 1, 1>, Sequence<2> >,
+        DeviceElementwiseImpl<Tuple<F32, F32, F32, F32, F32>, Tuple<F32>, Normalize, Rank, 4, Sequence<1, 2, 2, 2, 2>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<F32, F32, F32, F32, F32>, Tuple<F32>, Normalize, Rank, 4, Sequence<2, 2, 2, 2, 2>, Sequence<2> >,
+        DeviceElementwiseImpl<Tuple<F32, F32, F32, F32, F32>, Tuple<F32>, Normalize, Rank, 4, Sequence<4, 1, 1, 1, 1>, Sequence<4> >,
+        DeviceElementwiseImpl<Tuple<F32, F32, F32, F32, F32>, Tuple<F32>, Normalize, Rank, 4, Sequence<1, 4, 4, 4, 4>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<F32, F32, F32, F32, F32>, Tuple<F32>, Normalize, Rank, 4, Sequence<4, 2, 2, 2, 2>, Sequence<4> >,
+        DeviceElementwiseImpl<Tuple<F32, F32, F32, F32, F32>, Tuple<F32>, Normalize, Rank, 4, Sequence<2, 4, 4, 4, 4>, Sequence<2> >,
+        DeviceElementwiseImpl<Tuple<F32, F32, F32, F32, F32>, Tuple<F32>, Normalize, Rank, 4, Sequence<4, 4, 4, 4, 4>, Sequence<4> >
+     >;
+// clang-format on
+
+void add_device_batchnorm_infer_rank_4_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<Tuple<F32, F32, F32, F32, F32>, Tuple<F32>, Normalize, 4>>>& instances)
+{
+    add_device_operation_instances(instances, device_batchnorm_infer_f32_instances<4>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f64_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f64_instance.cpp
new file mode 100644
index 0000000000..15a02af021
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f64_instance.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/utility/tuple.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64 = double;
+
+using Normalize = ck::tensor_operation::element_wise::NormalizeInInfer;
+
+// clang-format off
+template <index_t Rank>
+using device_batchnorm_infer_f64_instances =
+     std::tuple <
+        // Tuple<XDataType, MeanDataType, VarDataType, ScaleDataType, BiasDataType>, Tuple<YDataType>, NormalizeOp, Rank, MPerThread, Sequence<XVectorSize, MeanDataType, VarDataType, ScaleVectorSize, BiasVectorSize>, Sequence<YVectorSize> 
+        DeviceElementwiseImpl<Tuple<F64, F64, F64, F64, F64>, Tuple<F64>, Normalize, Rank, 1, Sequence<1, 1, 1, 1, 1>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<F64, F64, F64, F64, F64>, Tuple<F64>, Normalize, Rank, 2, Sequence<1, 1, 1, 1, 1>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<F64, F64, F64, F64, F64>, Tuple<F64>, Normalize, Rank, 2, Sequence<2, 1, 1, 1, 1>, Sequence<2> >,
+        DeviceElementwiseImpl<Tuple<F64, F64, F64, F64, F64>, Tuple<F64>, Normalize, Rank, 2, Sequence<1, 2, 2, 2, 2>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<F64, F64, F64, F64, F64>, Tuple<F64>, Normalize, Rank, 2, Sequence<2, 2, 2, 2, 2>, Sequence<2> >,
+        DeviceElementwiseImpl<Tuple<F64, F64, F64, F64, F64>, Tuple<F64>, Normalize, Rank, 4, Sequence<1, 1, 1, 1, 1>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<F64, F64, F64, F64, F64>, Tuple<F64>, Normalize, Rank, 4, Sequence<2, 1, 1, 1, 1>, Sequence<2> >,
+        DeviceElementwiseImpl<Tuple<F64, F64, F64, F64, F64>, Tuple<F64>, Normalize, Rank, 4, Sequence<1, 2, 2, 2, 2>, Sequence<1> >,
+        DeviceElementwiseImpl<Tuple<F64, F64, F64, F64, F64>, Tuple<F64>, Normalize, Rank, 4, Sequence<2, 2, 2, 2, 2>, Sequence<2> >
+     >;
+// clang-format on
+
+void add_device_batchnorm_infer_rank_4_f64_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<Tuple<F64, F64, F64, F64, F64>, Tuple<F64>, Normalize, 4>>>& instances)
+{
+    add_device_operation_instances(instances, device_batchnorm_infer_f64_instances<4>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/contraction_bilinear/CMakeLists.txt
new file mode 100644
index 0000000000..1db6985f61
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_bilinear/CMakeLists.txt
@@ -0,0 +1,17 @@
+set(DEVICE_CONTRACTION_BILINEAR_INSTANCES)
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    #float
+    list(APPEND DEVICE_CONTRACTION_BILINEAR_INSTANCES device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp
+                                                      device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp
+                                                      device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp
+                                                      device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp)
+endif()
+if(DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
+    #double
+    list(APPEND DEVICE_CONTRACTION_BILINEAR_INSTANCES device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp
+                                                      device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp
+                                                      device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp
+                                                      device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp)
+endif()
+add_instance_library(device_contraction_bilinear_instance ${DEVICE_CONTRACTION_BILINEAR_INSTANCES})
+
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp
new file mode 100644
index 0000000000..5587db77e0
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32       = float;
+using F32_Tuple = ck::Tuple<F32>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// k/k/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|    DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|      Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |          |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |          |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp
new file mode 100644
index 0000000000..26262855ea
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32       = float;
+using F32_Tuple = ck::Tuple<F32>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// k/n/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|    DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|      Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |          |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |          |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   1,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   1,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   1,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   1,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp
new file mode 100644
index 0000000000..befc0dcd10
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32       = float;
+using F32_Tuple = ck::Tuple<F32>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// m/k/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|    DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|  CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|      Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|          _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |          |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|          _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |          |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                              |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    16,   1,   4,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   1,   4,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   1,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   1,   4,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   1,   4,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   1,   4,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp
new file mode 100644
index 0000000000..e45b47cf94
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32       = float;
+using F32_Tuple = ck::Tuple<F32>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// m/n/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|    DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|      Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |          |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |          |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    16,   1,   1,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   1,   1,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   1,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   1,   1,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   1,   1,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   1,   1,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, F32_Tuple,   F32,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           F32_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp
new file mode 100644
index 0000000000..f437a227d5
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp
@@ -0,0 +1,76 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64       = double;
+using F64_Tuple = ck::Tuple<F64>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// k/k/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|    DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|      Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |          |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |          |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    64,    16,   2,   2,   16,   16,    4,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1,  8, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   2,   2,   16,   16,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   2,   2,   16,   16,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    32,    16,   2,   2,   16,   16,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    32,   128,    16,   2,   2,   16,   16,    2,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    32,    16,   2,   2,   16,   16,    4,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1,  8, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    32,    64,    16,   2,   2,   16,   16,    2,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1,  8, 1,  8>,               1>
+    // clang-format on
+    >;
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           F64_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp
new file mode 100644
index 0000000000..13fdbeb35c
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp
@@ -0,0 +1,76 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64       = double;
+using F64_Tuple = ck::Tuple<F64>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// k/n/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|    DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|      Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |          |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |          |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   2,   1,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   2,   1,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   2,   1,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   2,   1,   16,   16,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   2,   2,   16,   16,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   2,   1,   16,   16,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   2,   2,   16,   16,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>
+    // clang-format on
+    >;
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           F64_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp
new file mode 100644
index 0000000000..95ef8c4929
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp
@@ -0,0 +1,76 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64       = double;
+using F64_Tuple = ck::Tuple<F64>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// m/k/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|    DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|  CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|      Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|          _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |          |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|          _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |          |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                              |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   1,   2,   16,   16,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   1,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   1,   2,   16,   16,    4,    4,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   1,   2,   16,   16,    4,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   2,   2,   16,   16,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   1,   2,   16,   16,    2,    4,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   2,   2,   16,   16,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 16>,               1>
+    // clang-format on
+    >;
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           F64_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp
new file mode 100644
index 0000000000..290f81d7c9
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp
@@ -0,0 +1,76 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64       = double;
+using F64_Tuple = ck::Tuple<F64>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// m/n/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|    DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|      Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |          |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |          |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   1,   1,   16,   16,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   1,   1,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   1,   1,   16,   16,    4,    4,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   1,   1,   16,   16,    4,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    16,   2,   2,   16,   16,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   1,   1,   16,   16,    2,    4,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, F64_Tuple,   F64,  PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    16,   2,   2,   16,   16,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>
+    // clang-format on
+    >;
+
+void add_device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           F64_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt
new file mode 100644
index 0000000000..aed8bef2a8
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt
@@ -0,0 +1,17 @@
+set(DEVICE_CONTRACTION_SCALE_INSTANCES)
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    #float
+    list(APPEND DEVICE_CONTRACTION_SCALE_INSTANCES device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp
+                                                   device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp
+                                                   device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp
+                                                   device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp)
+endif()
+if(DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
+    #double
+    list(APPEND DEVICE_CONTRACTION_SCALE_INSTANCES device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp
+                                                   device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp
+                                                   device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp
+                                                   device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp)
+endif()
+add_instance_library(device_contraction_scale_instance ${DEVICE_CONTRACTION_SCALE_INSTANCES})
+
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp
new file mode 100644
index 0000000000..16fd1cb407
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp
@@ -0,0 +1,78 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32         = float;
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] = E[m0, m1, n0, n1]
+// k/k/n are the fast changing dimension for A/B/E
+using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp
new file mode 100644
index 0000000000..ff37bf7cce
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp
@@ -0,0 +1,81 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32         = float;
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] = E[m0, m1, n0, n1]
+// k/n/n are the fast changing dimension for A/B/E
+using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   1,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   1,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   1,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   1,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32, PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp
new file mode 100644
index 0000000000..8a1f6f9334
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp
@@ -0,0 +1,81 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32         = float;
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] = E[m0, m1, n0, n1]
+// m/k/n are the fast changing dimension for A/B/E
+using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|      DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|  CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|        Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|          _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |            |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|          _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |            |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                              |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   256,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   256,    16,   1,   4,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   1,   4,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   1,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   1,   4,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   1,   4,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   1,   4,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp
new file mode 100644
index 0000000000..d333f59726
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp
@@ -0,0 +1,81 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32         = float;
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] = E[m0, m1, n0, n1]
+// m/n/n are the fast changing dimension for A/B/E
+using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|      DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|        Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |            |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |            |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   256,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   256,    16,   1,   1,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   1,   1,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   1,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   1,   1,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   1,   1,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   1,   1,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>
+    // clang-format on
+    >;
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F32,
+                                                           F32,
+                                                           Empty_Tuple,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp
new file mode 100644
index 0000000000..4c87b51a9e
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp
@@ -0,0 +1,75 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64         = double;
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// k/k/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|      DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|        Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |            |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |            |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,    64,    64,    64,    16,   2,   2,   16,   16,    4,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1,  8, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   2,   2,   16,   16,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   2,   2,   16,   16,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    32,    16,   2,   2,   16,   16,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    32,   128,    16,   2,   2,   16,   16,    2,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,    64,    64,    32,    16,   2,   2,   16,   16,    4,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1,  8, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,    64,    32,    64,    16,   2,   2,   16,   16,    2,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,              S<1,  8, 1,  8>,               1>
+    // clang-format on
+    >;
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           Empty_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp
new file mode 100644
index 0000000000..fd3f57c6b6
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp
@@ -0,0 +1,75 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64         = double;
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// k/n/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|      DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|        Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |            |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |            |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   2,   1,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   2,   1,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   2,   1,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   2,   1,   16,   16,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   2,   2,   16,   16,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   2,   1,   16,   16,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   2,   2,   16,   16,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>
+    // clang-format on
+    >;
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           Empty_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp
new file mode 100644
index 0000000000..1e53f0b2f5
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp
@@ -0,0 +1,75 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64         = double;
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// m/k/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|      DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|  CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|        Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|          _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |            |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|          _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |            |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                              |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   1,   2,   16,   16,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   1,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   1,   2,   16,   16,    4,    4,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   1,   2,   16,   16,    4,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   2,   2,   16,   16,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   1,   2,   16,   16,    2,    4,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   2,   2,   16,   16,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 16>,               1>
+    // clang-format on
+    >;
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           Empty_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp
new file mode 100644
index 0000000000..d02d146a9b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp
@@ -0,0 +1,75 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+// This (ifndef) is a hack to use customized behavior for buffer load rather than using default
+// setting Don't use this hack unless absolutely necessary!
+// FIXME: make the behavior of buffer load a configurable (template) parameter of each device op
+#define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 1
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64         = double;
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1]
+// m/n/n/n are the fast changing dimension for A/B/D/E
+using device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance = std::tuple<
+    // clang-format off
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|      DsData| EData|            A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|        Type|  Type|  Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |            |      |    Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |            |      |             |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   1,   1,   16,   16,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   1,   1,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,   128,    64,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1,  8>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   1,   1,   16,   16,    4,    4,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   128,    64,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1,  8, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   1,   1,   16,   16,    4,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,   128,    64,    16,   2,   2,   16,   16,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   1,   1,   16,   16,    2,    4,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1>,
+        DeviceContractionMultipleD_Xdl_CShuffle<       2,       2,       2,   F64,   F64,     F64,      F64, Empty_Tuple,   F64,  PassThrough, PassThrough,       Scale, GemmMNKPadding,        1,   256,    64,   128,    16,   2,   2,   16,   16,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              1,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              1,         1,           1,           1,              S<1, 16, 1, 16>,               1>
+    // clang-format on
+    >;
+
+void add_device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance(
+    std::vector<std::unique_ptr<DeviceContractionMultipleD<2,
+                                                           2,
+                                                           2,
+                                                           F64,
+                                                           F64,
+                                                           Empty_Tuple,
+                                                           F64,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           Scale>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000000..75a3670761
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_conv1d_bwd_data_instance
+   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp
+   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp
+   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp
+   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp
new file mode 100644
index 0000000000..e3e90c966d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = bhalf_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NWC = ck::tensor_layout::convolution::NWC;
+using KXC = ck::tensor_layout::convolution::KXC;
+using NWK = ck::tensor_layout::convolution::NWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, wi, c] * wei[k, x, c] = out[n, wo, k]
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_bf16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<1,
+                                                  NWC,
+                                                  KXC,
+                                                  NWK,
+                                                  BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instances{});
+    add_device_operation_instances(
+        instances, device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp
new file mode 100644
index 0000000000..81e9122d95
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp
@@ -0,0 +1,95 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NWC = ck::tensor_layout::convolution::NWC;
+using KXC = ck::tensor_layout::convolution::KXC;
+using NWK = ck::tensor_layout::convolution::NWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, wi, c] * wei[k, x, c] = out[n, wo, k]
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,   8,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,   8,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,   8,  32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,   8,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,   8,  32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,   8,  32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,   8,  32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,       
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,   8,  32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,   8,  32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,   8,  32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,   8,  32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvBwdData<1, NWC, KXC, NWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp
new file mode 100644
index 0000000000..dbc82168f4
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp
@@ -0,0 +1,94 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NWC = ck::tensor_layout::convolution::NWC;
+using KXC = ck::tensor_layout::convolution::KXC;
+using NWK = ck::tensor_layout::convolution::NWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, wi, c] * wei[k, x, c] = out[n, wo, k]
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|        ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|      Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                    |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_f32_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                     ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                   Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                 |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvBwdData<1, NWC, KXC, NWK, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instances{});
+    add_device_operation_instances(
+        instances, device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp
new file mode 100644
index 0000000000..3ac250f3e6
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp
@@ -0,0 +1,99 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NWC = ck::tensor_layout::convolution::NWC;
+using KXC = ck::tensor_layout::convolution::KXC;
+using NWK = ck::tensor_layout::convolution::NWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, wi, c] * wei[k, x, c] = out[n, wo, k]
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_int8_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|    ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |    |     |     | Wave| Wave|  Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |    |     |     |     |     |                 |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  16,   32,   32,    4,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  16,   32,   32,    2,    4,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  16,   32,   32,    4,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  16,   32,   32,    2,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  16,   32,   32,    2,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  16,   32,   32,    2,    1,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  16,   32,   32,    1,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  16,   32,   32,    2,    1,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  16,   32,   32,    1,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  16,   32,   32,    2,    1,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       1, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  16,   32,   32,    1,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<1,
+                                                  NWC,
+                                                  KXC,
+                                                  NWK,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instances{});
+    add_device_operation_instances(
+        instances, device_conv1d_bwd_data_xdl_nwc_kxc_nwk_1x1_s1_p0_int8_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000000..54ef9cc7a6
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt
@@ -0,0 +1,23 @@
+set(CONV2D_BWD_DATA_INSTANCES)
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+   list(APPEND CONV2D_BWD_DATA_INSTANCES device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp)
+   if(DL_KERNELS)
+      list(APPEND CONV2D_BWD_DATA_INSTANCES device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp)
+   endif()
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+   list(APPEND CONV2D_BWD_DATA_INSTANCES device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+   list(APPEND CONV2D_BWD_DATA_INSTANCES device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp)
+   if(DL_KERNELS)
+      list(APPEND CONV2D_BWD_DATA_INSTANCES device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp)
+   endif()
+endif()
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+   list(APPEND CONV2D_BWD_DATA_INSTANCES device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp)
+   if(DL_KERNELS)
+      list(APPEND CONV2D_BWD_DATA_INSTANCES device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp)
+   endif()
+endif()
+add_instance_library(device_conv2d_bwd_data_instance ${CONV2D_BWD_DATA_INSTANCES})
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp
new file mode 100644
index 0000000000..6ca909c35e
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef DL_KERNELS
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances = std::tuple<
+    // clang-format off
+        //#########################|       NDim|     InData|     WeiData|     OutData|     AccData|           In|           Wei|           Out|        Convolution| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //#########################|    Spatial|       Type|        Type|        Type|        Type|  Elementwise|   Elementwise|   Elementwise|            Forward|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //#########################|           |           |            |            |            |    Operation|     Operation|     Operation|     Specialization|      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //#########################|           |           |            |            |            |             |              |              |                   |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+    DeviceConvNdBwdDataNwcKxcNwk_Dl<          2, InDataType, WeiDataType, OutDataType, AccDataType,  InElementOp,  WeiElementOp,  OutElementOp, ConvBwdDataDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<1, 1, 8, 2>,      S<16, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 8, 1>,      S<0, 3, 1, 2>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+using device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //#########################|       NDim|     InData|     WeiData|     OutData|     AccData|           In|           Wei|           Out|                     Convolution| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //#########################|    Spatial|       Type|        Type|        Type|        Type|  Elementwise|   Elementwise|   Elementwise|                         Forward|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //#########################|           |           |            |            |            |    Operation|     Operation|     Operation|                  Specialization|      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //#########################|           |           |            |            |            |             |              |              |                                |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+    DeviceConvNdBwdDataNwcKxcNwk_Dl<          2, InDataType, WeiDataType, OutDataType, AccDataType,  InElementOp,  WeiElementOp,  OutElementOp, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<1, 1, 8, 2>,      S<16, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 8, 1>,      S<0, 3, 1, 2>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  InDataType,
+                                                  WeiDataType,
+                                                  OutDataType,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp
new file mode 100644
index 0000000000..d263e98851
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef DL_KERNELS
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using InDataType  = float;
+using WeiDataType = float;
+using OutDataType = float;
+using AccDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances = std::tuple<
+    // clang-format off
+        //#########################|       NDim|     InData|     WeiData|     OutData|     AccData|           In|           Wei|           Out|        Convolution| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //#########################|    Spatial|       Type|        Type|        Type|        Type|  Elementwise|   Elementwise|   Elementwise|            Forward|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //#########################|           |           |            |            |            |    Operation|     Operation|     Operation|     Specialization|      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //#########################|           |           |            |            |            |             |              |              |                   |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+    DeviceConvNdBwdDataNwcKxcNwk_Dl<          2, InDataType, WeiDataType, OutDataType, AccDataType,  InElementOp,  WeiElementOp,  OutElementOp, ConvBwdDataDefault,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 1, 8, 1>,      S<16, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+using device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances = std::tuple<
+    // clang-format off
+        //#########################|       NDim|     InData|     WeiData|     OutData|     AccData|           In|           Wei|           Out|                     Convolution| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //#########################|    Spatial|       Type|        Type|        Type|        Type|  Elementwise|   Elementwise|   Elementwise|                         Forward|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //#########################|           |           |            |            |            |    Operation|     Operation|     Operation|                  Specialization|      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //#########################|           |           |            |            |            |             |              |              |                                |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+    DeviceConvNdBwdDataNwcKxcNwk_Dl<          2, InDataType, WeiDataType, OutDataType, AccDataType,  InElementOp,  WeiElementOp,  OutElementOp, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 1, 8, 1>,      S<16, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  InDataType,
+                                                  WeiDataType,
+                                                  OutDataType,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp
new file mode 100644
index 0000000000..bc949e757c
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef DL_KERNELS
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using InDataType  = int8_t;
+using WeiDataType = int8_t;
+using OutDataType = int8_t;
+using AccDataType = int32_t;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough  = ck::tensor_operation::element_wise::PassThrough;
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances = std::tuple<
+    // clang-format off
+        //#########################|       NDim|     InData|     WeiData|     OutData|     AccData|           In|           Wei|           Out|        Convolution| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //#########################|    Spatial|       Type|        Type|        Type|        Type|  Elementwise|   Elementwise|   Elementwise|            Forward|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //#########################|           |           |            |            |            |    Operation|     Operation|     Operation|     Specialization|      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //#########################|           |           |            |            |            |             |              |              |                   |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+    DeviceConvNdBwdDataNwcKxcNwk_Dl<          2, InDataType, WeiDataType, OutDataType, AccDataType,  InElementOp,  WeiElementOp,  OutElementOp, ConvBwdDataDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>,      S<1, 1, 8, 4>,      S<16, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 8, 1>,      S<0, 3, 1, 2>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+using device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances = std::tuple<
+    // clang-format off
+        //#########################|       NDim|     InData|     WeiData|     OutData|     AccData|           In|           Wei|           Out|                     Convolution| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //#########################|    Spatial|       Type|        Type|        Type|        Type|  Elementwise|   Elementwise|   Elementwise|                         Forward|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //#########################|           |           |            |            |            |    Operation|     Operation|     Operation|                  Specialization|      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //#########################|           |           |            |            |            |             |              |              |                                |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+    DeviceConvNdBwdDataNwcKxcNwk_Dl<          2, InDataType, WeiDataType, OutDataType, AccDataType,  InElementOp,  WeiElementOp,  OutElementOp, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>,      S<1, 1, 8, 4>,      S<16, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 8, 1>,      S<0, 3, 1, 2>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
+    // clang-format on
+    >;
+
+void add_device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  InDataType,
+                                                  WeiDataType,
+                                                  OutDataType,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
new file mode 100644
index 0000000000..366d1fe160
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
@@ -0,0 +1,158 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+// FIXME: retire dedicated 2D version
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_BF16
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = bhalf_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_bf16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances = std::tuple<
+    // clang-format off
+        //####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|        ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //####################################################################|       |        |        |        |   Operation|   Operation|   Operation|      Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //####################################################################|       |        |        |        |            |            |            |                    |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+using device_conv_dedidecate_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_bf16_instances =
+    std::tuple<
+        // clang-format off
+        //####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                     ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //####################################################################|       |        |        |        |   Operation|   Operation|   Operation|                   Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //####################################################################|       |        |        |        |            |            |            |                                 |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_bf16_instances{});
+    add_device_operation_instances(
+        instances, device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances{});
+    add_device_operation_instances(
+        instances,
+        device_conv_dedidecate_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
new file mode 100644
index 0000000000..422e37e926
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
@@ -0,0 +1,158 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+// FIXME: retire dedicated 2D version
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,   8,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,   8,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,   8,  32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,   8,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,   8,  32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,   8,  32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,   8,  32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,       
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,   8,  32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,   8,  32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,   8,  32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,   8,  32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances = std::tuple<
+    // clang-format off
+        //####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //####################################################################|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //####################################################################|       |        |        |        |            |            |            |                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,   8,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,   8,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,   8,  32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+#if !CK_WORKAROUND_SWDEV_325164
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,   8,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,   8,  32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,   8,  32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+#endif
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,   8,  32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,   8,  32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,   8,  32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,   8,  32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,   8,  32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+using device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances =
+    std::tuple<
+        // clang-format off
+        //####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                     ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //####################################################################|       |        |        |        |   Operation|   Operation|   Operation|                   Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //####################################################################|       |        |        |        |            |            |            |                                 |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  F16,
+                                                  F16,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
new file mode 100644
index 0000000000..5993f6bd7a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
@@ -0,0 +1,155 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+// FIXME: retire dedicated 2D version
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|        ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|      Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                    |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                     ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                   Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                 |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances = std::tuple<
+    // clang-format off
+        //####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|        ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //####################################################################|       |        |        |        |   Operation|   Operation|   Operation|      Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //####################################################################|       |        |        |        |            |            |            |                    |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+using device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances =
+    std::tuple<
+        // clang-format off
+        //####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                     ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //####################################################################|       |        |        |        |   Operation|   Operation|   Operation|                   Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //####################################################################|       |        |        |        |            |            |            |                                 |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  F32,
+                                                  F32,
+                                                  F32,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances{});
+    add_device_operation_instances(
+        instances, device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances{});
+    add_device_operation_instances(
+        instances, device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
new file mode 100644
index 0000000000..36610ae205
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
@@ -0,0 +1,154 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+// FIXME: retire dedicated 2D version
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|    ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |    |     |     | Wave| Wave|  Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |    |     |     |     |     |                 |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  16,   32,   32,    4,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  16,   32,   32,    2,    4,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  16,   32,   32,    4,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  16,   32,   32,    2,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  16,   32,   32,    2,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  16,   32,   32,    2,    1,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  16,   32,   32,    1,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  16,   32,   32,    2,    1,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  16,   32,   32,    1,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  16,   32,   32,    2,    1,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       2, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  16,   32,   32,    1,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances = std::tuple<
+    // clang-format off
+        //####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|        ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //####################################################################|       |        |        |        |   Operation|   Operation|   Operation|      Specialization|      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //####################################################################|       |        |        |        |            |            |            |                    |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+// FIXME: retire dedicated 2D version
+using device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances =
+    std::tuple<
+        // clang-format off
+        //#####################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                     ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //#####################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|    ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //#####################################################################|       |        |        |        |   Operation|   Operation|   Operation|                   Specialization|      |      |      |      |    |     |     | Wave| Wave|  Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //#####################################################################|       |        |        |        |            |            |            |                                 |      |      |      |      |    |     |     |     |     |                 |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  16,   32,   32,    4,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  16,   32,   32,    2,    4,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  16,   32,   32,    4,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  16,   32,   32,    2,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  16,   32,   32,    2,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  16,   32,   32,    2,    1,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  16,   32,   32,    1,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  16,   32,   32,    2,    1,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  16,   32,   32,    1,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  16,   32,   32,    2,    1,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  16,   32,   32,    1,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<2,
+                                                  NHWC,
+                                                  KYXC,
+                                                  NHWK,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances{});
+    add_device_operation_instances(
+        instances, device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances{});
+    add_device_operation_instances(
+        instances, device_conv_dedicated_2d_bwd_data_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv2d_fwd/CMakeLists.txt
new file mode 100644
index 0000000000..96ecc95658
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_fwd/CMakeLists.txt
@@ -0,0 +1,16 @@
+set(DEVICE_CONV2D_FWD_INSTANCES)
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+   list(APPEND DEVICE_CONV2D_FWD_INSTANCES device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+   list(APPEND DEVICE_CONV2D_FWD_INSTANCES device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp)
+   list(APPEND DEVICE_CONV2D_FWD_INSTANCES device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+   list(APPEND DEVICE_CONV2D_FWD_INSTANCES device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp)
+endif()
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+   list(APPEND DEVICE_CONV2D_FWD_INSTANCES device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp)
+endif()
+   
+add_instance_library(device_conv2d_fwd_instance ${DEVICE_CONV2D_FWD_INSTANCES})
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
new file mode 100644
index 0000000000..da96c79a6e
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
@@ -0,0 +1,155 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto ConvFwdOddC =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
+
+// arbitrary conv
+using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// 1x1, pad 0
+using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// 1x1, stride 1, pad 0
+using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_odd_c_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,   128,    64,     2,  4,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   256,   256,    64,     2,  4,   32,   32,    4,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   128,   128,    64,     2,  4,   32,   32,    2,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,    ConvFwdOddC,   128,    64,    64,     2,  4,   32,   32,    1,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvFwd<2, NHWC, KYXC, NHWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_p0_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_odd_c_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
new file mode 100644
index 0000000000..63c612523f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
@@ -0,0 +1,129 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_BF16
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_bf16_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_bf16_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvFwd<2,
+                                              NHWC,
+                                              KYXC,
+                                              NHWK,
+                                              BF16,
+                                              BF16,
+                                              BF16,
+                                              PassThrough,
+                                              PassThrough,
+                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_bf16_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
new file mode 100644
index 0000000000..0f3b9e7939
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
@@ -0,0 +1,121 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_FP16
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvFwd<2, NHWC, KYXC, NHWK, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_f16_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
new file mode 100644
index 0000000000..14f9b5cd6a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
@@ -0,0 +1,120 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_FP32
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_f32_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceConvFwd<2, NHWC, KYXC, NHWK, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_f32_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
new file mode 100644
index 0000000000..3f641cdadc
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
@@ -0,0 +1,126 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWC = ck::tensor_layout::convolution::NHWC;
+using KYXC = ck::tensor_layout::convolution::KYXC;
+using NHWK = ck::tensor_layout::convolution::NHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, hi, wi, c] * wei[k, y, x, c] = out[n, ho, wo, k]
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData|  WeiData| OutData|  AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|     Type|    Type|     Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |         |        |         |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |         |        |         |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwdDefault,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_int8_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData|  WeiData| OutData|  AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|     Type|    Type|     Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |         |        |         |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |         |        |         |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough,   ConvFwd1x1P0,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances = std::tuple<
+    // clang-format off
+        //################################################################| InData|  WeiData| OutData|  AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //################################################################|   Type|     Type|    Type|     Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //################################################################|       |         |        |         |   Operation|   Operation|   Operation|               |      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //################################################################|       |         |        |         |            |            |            |               |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>,
+        DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<  int8_t,  int8_t,  int8_t,  int32_t, PassThrough, PassThrough, PassThrough, ConvFwd1x1S1P0,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvFwd<2,
+                                              NHWC,
+                                              KYXC,
+                                              NHWK,
+                                              int8_t,
+                                              int8_t,
+                                              int8_t,
+                                              PassThrough,
+                                              PassThrough,
+                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0_int8_instances{});
+    add_device_operation_instances(instances,
+                                   device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0_int8_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/CMakeLists.txt
new file mode 100644
index 0000000000..670cd94fc9
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_instance_library(device_conv2d_fwd_bias_relu_instance
+   device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
new file mode 100644
index 0000000000..e34ea06ff4
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
@@ -0,0 +1,154 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddRelu     = ck::tensor_operation::element_wise::AddRelu;
+
+static constexpr auto MemorySet = ck::InMemoryDataOperationEnum::Set;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto ConvFwdOddC =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
+
+// arbitrary conv
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|           Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|  GlobalMemory| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################################|       |        |        |        |   Operation|   Operation|   Operation| DataOperation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################################|       |        |        |        |            |            |            |              |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwdDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// 1x1, pad 0
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|           Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|  GlobalMemory| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################################|       |        |        |        |   Operation|   Operation|   Operation| DataOperation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################################|       |        |        |        |            |            |            |              |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,   ConvFwd1x1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// 1x1, stride 1, pad 0
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|           Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|  GlobalMemory| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################################|       |        |        |        |   Operation|   Operation|   Operation| DataOperation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################################|       |        |        |        |            |            |            |              |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet, ConvFwd1x1S1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// Odd C
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_odd_c_f16_instances = std::tuple<
+    // clang-format off
+        //##########################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|           Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##########################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|  GlobalMemory| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##########################################################################################|       |        |        |        |   Operation|   Operation|   Operation| DataOperation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##########################################################################################|       |        |        |        |            |            |            |              |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,   128,    64,     2,  4,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   256,   256,    64,     2,  4,   32,   32,    4,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   128,   128,    64,     2,  4,   32,   32,    2,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,     AddRelu,     MemorySet,    ConvFwdOddC,   128,    64,    64,     2,  4,   32,   32,    1,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<DeviceConvFwdBiasActivationPtr<PassThrough, PassThrough, AddRelu>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_p0_f16_instances{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_odd_c_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt
new file mode 100644
index 0000000000..68d5f582fd
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt
@@ -0,0 +1,4 @@
+add_instance_library(device_conv2d_fwd_bias_relu_add_instance
+   device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
+)
+
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
new file mode 100644
index 0000000000..3254fcfc26
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
@@ -0,0 +1,154 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+static constexpr auto ConvFwdOddC =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::OddC;
+
+// arbitrary conv
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instances = std::tuple<
+    // clang-format off
+        //##############################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##############################################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwdDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// 1x1, pad 0
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##############################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##############################################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,   ConvFwd1x1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// 1x1, stride 1, pad 0
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##############################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##############################################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd, ConvFwd1x1S1P0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+// Odd C
+using device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_odd_c_f16_instances = std::tuple<
+    // clang-format off
+        //##############################################################################################| InData| WeiData| OutData| AccData|          In|         Wei|         Out|    ConvForward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################################################################################|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##############################################################################################|       |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################################################################################|       |        |        |        |            |            |            |               |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  8, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  4, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<4,  2, 8>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,   128,    64,     2,  4,   32,   32,    2,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   256,   256,    64,     2,  4,   32,   32,    4,    1,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 32, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   128,   128,    64,     2,  4,   32,   32,    2,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 32, 1, 1, 4>,               8>,
+        DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Add_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<    F16,     F16,     F16,     F32, PassThrough, PassThrough,  AddReluAdd,    ConvFwdOddC,   128,    64,    64,     2,  4,   32,   32,    1,    2,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,      true,     S<2, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,      true,           1,           1,             S<1, 1, 16, 1, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<DeviceConvFwdBiasActivationAddPtr<PassThrough, PassThrough, AddReluAdd>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instances{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_p0_f16_instances{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_1x1_s1_p0_f16_instances{});
+    add_device_operation_instances(
+        instances,
+        device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_odd_c_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000000..db92208fd7
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_conv3d_bwd_data_instance
+   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp
+   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp
+   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp
+   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp
new file mode 100644
index 0000000000..94b2a47e50
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = bhalf_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
+using KZYXC = ck::tensor_layout::convolution::KZYXC;
+using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, di, hi, wi, c] * wei[k, z, y, x, c] = out[n, do, ho, wo, k]
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_bf16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instances{});
+    add_device_operation_instances(
+        instances, device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_bf16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp
new file mode 100644
index 0000000000..4244ab7b87
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
+using KZYXC = ck::tensor_layout::convolution::KZYXC;
+using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, di, hi, wi, c] * wei[k, z, y, x, c] = out[n, do, ho, wo, k]
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,   8,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,   8,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,   8,  32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,   8,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,   8,  32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,   8,  32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,   8,  32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,   8,  32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,       
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,   8,  32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,   8,  32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,   8,  32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,   8,  32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_f16_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  F16,
+                                                  F16,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instances{});
+    add_device_operation_instances(
+        instances, device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp
new file mode 100644
index 0000000000..5c7db4ca3b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp
@@ -0,0 +1,101 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
+using KZYXC = ck::tensor_layout::convolution::KZYXC;
+using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, di, hi, wi, c] * wei[k, z, y, x, c] = out[n, do, ho, wo, k]
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|        ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|      Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                    |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_f32_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                     ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                             Data|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                   Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                 |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,  ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  F32,
+                                                  F32,
+                                                  F32,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instances{});
+    add_device_operation_instances(
+        instances, device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp
new file mode 100644
index 0000000000..ebc56487a1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp
@@ -0,0 +1,99 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
+using KZYXC = ck::tensor_layout::convolution::KZYXC;
+using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// Compilation parameters for in[n, wi, c] * wei[k, x, c] = out[n, wo, k]
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|       ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|               Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|     Specialization|      |      |      |      |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                   |      |      |      |      |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   256,     4,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,   128,     4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   256,   128,     4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,   128,     4,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    64,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    64,   128,     4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    64,     4,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,   128,    64,     4,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   256,    64,   128,     4,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,   128,    32,     4,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,   128,    32,   128,     4,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    64,    32,     4,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataDefault,    64,    32,    64,     4,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+using device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_int8_instances = std::tuple<
+    // clang-format off
+        //##############################|     Num| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                    ConvBackward| Block|  MPer|  NPer| K0Per|  K1| MPer| NPer| MXdl| NXdl|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##############################|     Dim|   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Data|  Size| Block| Block| Block|    |  XDL|  XDL|  Per|  Per|    ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##############################| Spatial|       |        |        |        |   Operation|   Operation|   Operation|                  Specialization|      |      |      |      |    |     |     | Wave| Wave|  Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##############################|        |       |        |        |        |            |            |            |                                |      |      |      |      |    |     |     |     |     |                 |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   256,   128,     4,  16,   32,   32,    4,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   256,     4,  16,   32,   32,    2,    4,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,   128,     4,  16,   32,   32,    4,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,   128,     4,  16,   32,   32,    2,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    64,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    64,   128,     4,  16,   32,   32,    2,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    64,     4,  16,   32,   32,    2,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,   128,    64,     4,  16,   32,   32,    2,    1,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   256,    64,   128,     4,  16,   32,   32,    1,    2,      S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 64, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,   128,    32,     4,  16,   32,   32,    2,    1,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              1,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,   128,    32,   128,     4,  16,   32,   32,    1,    2,      S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 32, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    64,    32,     4,  16,   32,   32,    2,    1,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              2,             16,      true,               7,               1>,
+        DeviceConvNdBwdDataNwcKxcNwk_Xdl<       3, int8_t,  int8_t,  int8_t, int32_t, PassThrough, PassThrough, PassThrough, ConvBwdDataFilter1x1Stride1Pad0,    64,    32,    64,     4,  16,   32,   32,    1,    2,      S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,      true,     S<4, 16, 1>,     S<2, 0, 1>,     S<0, 2, 1>,             1,              4,             16,      true,               7,               1>
+    // clang-format on
+    >;
+
+void add_device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceConvBwdData<3,
+                                                  NDHWC,
+                                                  KZYXC,
+                                                  NDHWK,
+                                                  int8_t,
+                                                  int8_t,
+                                                  int8_t,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instances{});
+    add_device_operation_instances(
+        instances, device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_1x1_s1_p0_int8_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/elementwise/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/elementwise/CMakeLists.txt
new file mode 100644
index 0000000000..47516b4162
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/elementwise/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_instance_library(device_elementwise_instance
+    device_normalize_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp b/library/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp
new file mode 100644
index 0000000000..f2a5f0728a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp
@@ -0,0 +1,57 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using inputType      = F16;
+using MeanType       = F32;
+using SquareMeanType = F32;
+using GammaDataType  = F16;
+using BetaDataType   = F16;
+using outputType     = F16;
+
+using Normalize = ck::tensor_operation::element_wise::Normalize;
+using device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances = std::tuple<
+    // clang-format off
+    //###################|<in, mean, square_mean, gamma, beta>| <out>|  functor| NDim| MPerThread| <in, mean, square_mean, gamma, beta ScalarPerVector>| <out ScalarPerVector>|
+    DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>,  Tuple<F16>,  Normalize,  2,   8,       Sequence<8, 1, 1, 8, 8>,      Sequence<8>                >,
+    DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>,  Tuple<F16>,  Normalize,  2,   4,       Sequence<4, 1, 1, 4, 4>,      Sequence<4>                >,
+    DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>,  Tuple<F16>,  Normalize,  2,   2,       Sequence<2, 1, 1, 2, 2>,      Sequence<2>                >
+    // clang-format on
+    >;
+
+using device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_generic_instance = std::tuple<
+    // clang-format off
+    DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>,  Tuple<F16>,  Normalize,  2,   1,       Sequence<1, 1, 1, 1, 1>,      Sequence<1>                >
+    // clang-format on
+    >;
+
+void add_device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances(
+    std::vector<DeviceElementwisePtr<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, 2>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_generic_instance{});
+
+    add_device_operation_instances(
+        instances, device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt
new file mode 100644
index 0000000000..ea53b82be5
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt
@@ -0,0 +1,5 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+add_instance_library(device_elementwise_normalization_instance
+    device_elementwise_normalization_f16_instance.cpp
+)
+endif()
diff --git a/library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp
new file mode 100644
index 0000000000..3e2386ee03
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Add  = ck::tensor_operation::element_wise::Add;
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+template <typename XElementwise, typename YElementwise, index_t Rank, index_t Reduce>
+// clang-format off
+using device_elementwise_normalization_f16_instances =
+    std::tuple <
+        // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 4, 64, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>, // fallback kernel
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 1024, 1, 1024, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 2, 1, 2, 1, 2, 2>
+    >;
+// clang-format on
+
+void add_device_elementwise_normalization_rank_2_1_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwiseNormalization<ck::Tuple<F16, F16>, F16, F16, F32, F16, Add, Pass, 2, 1>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_elementwise_normalization_f16_instances<Add, Pass, 2, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt
new file mode 100644
index 0000000000..5162617b18
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt
@@ -0,0 +1,137 @@
+set(GEMM_INSTANCES)
+if(DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
+    list(APPEND GEMM_INSTANCES device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp)
+    list(APPEND GEMM_INSTANCES device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp)
+    list(APPEND GEMM_INSTANCES device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp)
+    list(APPEND GEMM_INSTANCES device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp)
+   if(DL_KERNELS)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp)
+   endif()
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+   if(DL_KERNELS)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instance.cpp)
+   endif()
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_add_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v1_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v1_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v2_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_add_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v1_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v1_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v2_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_add_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v1_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_add_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v1_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp)
+endif()
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+   if(DL_KERNELS)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp)
+     list(APPEND GEMM_INSTANCES device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp)
+   endif()
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp)
+   list(APPEND GEMM_INSTANCES device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp)
+endif()
+
+add_instance_library(device_gemm_instance ${GEMM_INSTANCES})
+
+
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  set(ENABLE_PIPELINE_V2_OPT OFF)
+
+  if (ENABLE_PIPELINE_V2_OPT)
+    set(MAX_ILP_OPTS
+       -mllvm
+       -amdgpu-enable-max-ilp-scheduling-strategy
+    )
+    set(WAVES_PER_EU_DEFS
+       CK_USE_WAVES_PER_EU=1
+       CK_MIN_WAVES_PER_EU=1
+       CK_MAX_WAVES_PER_EU=1
+    )
+    set(IGLP_OPT_DEFS
+       CK_EXPERIMENTAL_PIPELINE_V2_IGLP_OPT=1
+    )
+
+    # layout=NT
+    set_source_files_properties(device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp PROPERTIES
+       COMPILE_OPTIONS ";;"
+       COMPILE_DEFINITIONS "${WAVES_PER_EU_DEFS};${IGLP_OPT_DEFS}")
+    # layout=NN
+    set_source_files_properties(device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp PROPERTIES
+       COMPILE_OPTIONS "${MAX_ILP_OPTS}"
+       COMPILE_DEFINITIONS "${WAVES_PER_EU_DEFS};${IGLP_OPT_DEFS}")
+    # layout=TT
+    set_source_files_properties(device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp PROPERTIES
+       COMPILE_OPTIONS ";;"
+       COMPILE_DEFINITIONS "${WAVES_PER_EU_DEFS};${IGLP_OPT_DEFS}")
+    # layout=TN
+    set_source_files_properties(device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp PROPERTIES
+       COMPILE_OPTIONS "${MAX_ILP_OPTS}"
+       COMPILE_DEFINITIONS "${WAVES_PER_EU_DEFS};${IGLP_OPT_DEFS}")
+  endif(ENABLE_PIPELINE_V2_OPT)
+endif(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..53fc74a392
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl_dpp8.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        // ##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|      CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|         SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|                Order|                 |                   |
+        // ##########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                     |                 |                   |
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          8,      1,       S<1, 8>,       S<1, 1>,      S<1, 1, 4, 2>,       S<4, 1,  2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>,       S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,     8,    64,    16,  2,          1,          8,      1,       S<1, 8>,       S<4, 1>,      S<1, 1, 4, 2>,      S<16, 1,  2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 2, 2>,      S<16, 1, 2, 2>,       S<1, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 2, 1>,      S<0, 3, 1, 2>,        S<1, 1, 2, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,     8,    64,    16,  2,          1,          8,      1,       S<1, 8>,       S<4, 1>,      S<1, 1, 4, 2>,      S<16, 1,  2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 2, 2>,       S<4, 1, 8, 2>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    16,  2,          1,          8,      1,       S<1, 8>,       S<8, 1>,      S<1, 1, 2, 2>,      S<16, 1,  4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 2, 1>,      S<0, 3, 1, 2>,       S<1, 1, 2, 2>,       S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          8,      1,       S<2, 8>,       S<4, 1>,      S<2, 1, 4, 2>,      S<4, 1,  16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>,       S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    16,  2,          8,          1,      1,       S<1, 1>,       S<8, 8>,      S<1, 1, 2, 2>,      S<16, 1,  4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 2, 1>,      S<0, 3, 1, 2>,       S<1, 1, 2, 2>,       S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,    64,    16,  2,          4,          8,      1,       S<2, 8>,       S<8, 1>,      S<2, 1, 4, 2>,      S<8, 1,  16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>,       S<2, 1, 4, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          1,          8,      8,       S<4, 8>,       S<4, 1>,      S<4, 1, 4, 2>,      S<4, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>,       S<2, 1, 4, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,     8,  2,          1,          8,      8,       S<8, 8>,       S<4, 1>,      S<1, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>,       S<1, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          8,      8,       S<2, 8>,      S<16, 1>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>,       S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..35d46b32f5
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl_dpp8.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        // ##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|      CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|         SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|                Order|                 |                   |
+        // ##########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                     |                 |                   |
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          8,      1,       S<1, 8>,       S<1, 1>,      S<1, 1, 4, 2>,       S<4, 1,  2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>,       S<4, 1, 1, 2>,      S<1, 1,   8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,     8,    64,    16,  2,          1,          8,      1,       S<1, 8>,       S<4, 1>,      S<1, 1, 4, 2>,      S<16, 1,  2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 2, 2>,      S<16, 1, 2, 2>,      S<1, 1,  32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,     8,    64,    16,  2,          1,          8,      1,       S<1, 8>,       S<4, 1>,      S<1, 1, 4, 2>,      S<16, 1,  2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 2, 2>,       S<4, 1, 8, 2>,      S<4, 1,   8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    16,  2,          1,          8,      1,       S<1, 8>,       S<8, 1>,      S<1, 1, 2, 2>,      S<16, 1,  4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 2, 1>,      S<0, 3, 1, 2>,       S<1, 1, 2, 2>,       S<4, 1, 4, 2>,      S<4, 1,  16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          8,      1,       S<2, 8>,       S<4, 1>,      S<2, 1, 4, 2>,      S<4, 1,  16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>,       S<4, 1, 2, 2>,      S<2, 1,  32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    16,  2,          8,          1,      1,       S<1, 1>,       S<8, 8>,      S<1, 1, 2, 2>,      S<16, 1,  4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 2, 1>,      S<0, 3, 1, 2>,       S<1, 1, 2, 2>,       S<4, 1, 4, 2>,      S<4, 1,  16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,    64,    16,  2,          4,          8,      1,       S<2, 8>,       S<8, 1>,      S<2, 1, 4, 2>,      S<8, 1,  16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>,       S<4, 1, 2, 2>,      S<4, 1,  32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          1,          8,      8,       S<4, 8>,       S<4, 1>,      S<4, 1, 4, 2>,      S<4, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>,       S<4, 1, 2, 2>,      S<4, 1,  32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,     8,  2,          1,          8,      8,       S<8, 8>,       S<4, 1>,      S<1, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>,       S<4, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          8,      8,       S<2, 8>,      S<16, 1>,      S<2, 1, 4, 2>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 2>,       S<4, 1, 2, 2>,      S<4, 1,  64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..22e2e5f398
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl_dpp8.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        // ##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|      CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|         SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|                Order|                 |                   |
+        // ##########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                     |                 |                   |
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          8,      1,       S<1, 8>,       S<1, 1>,      S<4, 1, 1, 2>,       S<1, 1,  8, 1>,  S<1, 2, 0 ,3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,     8,    64,    16,  2,          1,          8,      1,       S<1, 8>,       S<4, 1>,      S<4, 1, 1, 2>,        S<4, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<16, 1, 2, 2>,       S<1, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 2, 1>,      S<0, 3, 1, 2>,        S<1, 1, 2, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,     8,    64,    16,  2,          1,          8,      1,       S<1, 8>,       S<4, 1>,      S<4, 1, 1, 2>,        S<4, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<4, 1, 8, 2>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    16,  2,          1,          8,      1,       S<1, 8>,       S<8, 1>,      S<2, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<2, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          8,      1,       S<2, 8>,       S<4, 1>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    16,  2,          8,          1,      1,       S<1, 1>,       S<8, 8>,      S<2, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<2, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,    64,    16,  2,          4,          8,      1,       S<2, 8>,       S<8, 1>,      S<4, 1, 2, 2>,       S<4, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<2, 1, 4, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          1,          8,      8,       S<4, 8>,       S<4, 1>,      S<4, 1, 4, 2>,       S<4, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<2, 1, 4, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,     8,  2,          1,          8,      8,       S<8, 8>,       S<4, 1>,      S<4, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<1, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          8,      8,       S<2, 8>,      S<16, 1>,      S<4, 1, 2, 2>,       S<4, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..49e9416c81
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl_dpp8.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        // ##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|      CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // ##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|         SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // ##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|                Order|                 |                   |
+        // ##########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                     |                 |                   |
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          8,      1,       S<1, 8>,       S<1, 1>,      S<4, 1, 1, 2>,       S<1, 1,  8, 1>,  S<1, 2, 0 ,3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<4, 1, 1, 2>,      S<1, 1,   8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,     8,    64,    16,  2,          1,          8,      1,       S<1, 8>,       S<4, 1>,      S<4, 1, 1, 2>,        S<4, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<16, 1, 2, 2>,      S<1, 1,  32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,     8,    64,    16,  2,          1,          8,      1,       S<1, 8>,       S<4, 1>,      S<4, 1, 1, 2>,        S<4, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<4, 1, 8, 2>,      S<4, 1,   8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    16,  2,          1,          8,      1,       S<1, 8>,       S<8, 1>,      S<2, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<2, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<4, 1, 4, 2>,      S<4, 1,  16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          8,      1,       S<2, 8>,       S<4, 1>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<4, 1, 2, 2>,      S<2, 1,  32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    16,  2,          8,          1,      1,       S<1, 1>,       S<8, 8>,      S<2, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<2, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<4, 1, 4, 2>,      S<4, 1,  16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,    64,    16,  2,          4,          8,      1,       S<2, 8>,       S<8, 1>,      S<4, 1, 2, 2>,       S<4, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<4, 1, 2, 2>,      S<4, 1,  32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          1,          8,      8,       S<4, 8>,       S<4, 1>,      S<4, 1, 4, 2>,       S<4, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<4, 1, 2, 2>,      S<4, 1,  32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,     8,  2,          1,          8,      8,       S<8, 8>,       S<4, 1>,      S<4, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<4, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDlDpp8< F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          8,      8,       S<2, 8>,      S<16, 1>,      S<4, 1, 2, 2>,       S<4, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,       S<4, 1, 2, 2>,      S<4, 1,  64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,  S<0, 1, 2, 3, 4, 5>,                5,                  4>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..ea99a5a30e
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_dl_f16_f16_f16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        // #########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // #########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // #########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // #########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<8, 2>,       S<4, 2>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 4>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<4, 2>,       S<8, 2>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<2, 4>,       S<2, 8>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<4, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,       S<16, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<4, 1, 4, 2>,        S<4, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<1, 4>,       S<1, 4>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<4, 1, 4, 2>,        S<4, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<8, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<8, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,       S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,       S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f16_f16_f16_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp
new file mode 100644
index 0000000000..d444e29aa3
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp
@@ -0,0 +1,73 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances = std::tuple<
+    // clang-format off
+        // #########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // #########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // #########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|               Order|                 |                   |
+        // #########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    16,    64,    16,  2,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<4, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,       S<16, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<8, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<8, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,       S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,       S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,     S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..b83acfa8cb
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_dl_f16_f16_f16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        // #########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // #########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // #########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        // #########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128       
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<8, 2>,       S<4, 2>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 4>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<4, 2>,       S<8, 2>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<2, 4>,       S<2, 8>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<1, 4>,       S<1, 4>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f16_f16_f16_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp
new file mode 100644
index 0000000000..f8f8a0bd33
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp
@@ -0,0 +1,73 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances = std::tuple<
+    // clang-format off
+        // #########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // #########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // #########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        // #########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128       
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64       
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    16,    64,    16,  2,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..d5800e0333
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_dl_f16_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        // #########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // #########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // #########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        // #########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<8, 2>,       S<4, 2>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 4>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 8, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<4, 2>,       S<8, 2>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<2, 4>,       S<2, 8>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 8, 2>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<1, 4>,       S<1, 4>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,        S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,        S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f16_f16_f16_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
new file mode 100644
index 0000000000..840a4fabe1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
@@ -0,0 +1,73 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances = std::tuple<
+    // clang-format off
+        // #########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // #########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // #########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        // #########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<2, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    16,    64,    16,  2,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,       S<4, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..abe52ce1da
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,85 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_dl_f16_f16_f16_mk_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //  ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // // MPerBlock=128, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<8, 2>,       S<4, 2>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,    64,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 4>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // // MPerBlock=64, NPerBlock=128
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<4, 2>,       S<8, 2>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,    64,   128,    16,  2,          4,          4,      1,       S<2, 4>,       S<2, 8>,      S<8, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 2, 2>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 1, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<1, 4>,       S<1, 4>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+        // clang-format on
+        >;
+
+void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f16_f16_f16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
new file mode 100644
index 0000000000..82b1b5dc29
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
@@ -0,0 +1,74 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances =
+    std::tuple<
+        // clang-format off
+        //  ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  2,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 2, 2>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    16,    64,    16,  2,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 1, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+        // clang-format on
+        >;
+
+void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..e696bfdcdc
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_dl_f32_f32_f32_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //  ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                    |                    |                 |                   |
+        DeviceGemmDl<   F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 1>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<2, 1, 4, 1>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 1>, S<0, 1, 2, 3, 4, 5>,                5,                  4>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f32_f32_f32_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..d3ad7c60ec
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_dl_f32_f32_f32_km_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //  ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmDl<   F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 1>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,        S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,                5,                  4>
+        // clang-format on
+        >;
+
+void add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f32_f32_f32_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..a56a36b0ab
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_dl_f32_f32_f32_mk_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //  ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmDl<   F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,        S<1, 1, 1, 1>,      S<2, 1, 4, 1>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 1>, S<0, 1, 2, 3, 4, 5>,                5,                  4>
+        // clang-format on
+        >;
+
+void add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f32_f32_f32_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..63d55e81d6
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_dl_f32_f32_f32_mk_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //  ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        //  ########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        //  ########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        //  ########|      |      |      |        |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        DeviceGemmDl<   F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,        S<1, 1, 1, 1>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,        S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,                5,                  4>
+        // clang-format on
+        >;
+
+void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_f32_f32_f32_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..05f399e218
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_dl_i8_i8_i8_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        // #########|  AData|   BData|   CData|    AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // #########|   Type|    Type|    Type|       Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // #########|       |        |        |           |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        // #########|       |        |        |           |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,    64,   128,    16,  4,          2,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  4,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<1, 1, 4, 4>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<4, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  4,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<1, 1, 4, 4>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<4, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  4,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  4,          4,          1,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  4,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 4>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<4, 1, 4, 4>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  4,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<1, 1, 4, 4>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<8, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  4,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 4>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<8, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  4,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  4,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  4,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  4,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  4,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  4,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,     S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_i8_i8_i8_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp
new file mode 100644
index 0000000000..d1eb8edf97
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp
@@ -0,0 +1,81 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instances = std::tuple<
+    // clang-format off
+        // #########|  AData|   BData|   CData|    AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer|  CThreadTransfer|    CThreadTransfer|
+        // #########|   Type|    Type|    Type|       Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess|  SrcDstVectorDim| DstScalarPerVector|
+        // #########|       |        |        |           |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                 |                   |
+        // #########|       |        |        |           |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                 |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<4, 4>,       S<4, 2>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    32,    32,    32,     8,  4,          2,          4,      1,       S<4, 2>,       S<2, 2>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<4, 1, 8, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<4, 2>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<4, 1, 8, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<2, 1, 4, 4>,       S<4, 1, 8, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    16,    16,    16,    16,  4,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 4>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 4, 4>,       S<4, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    16,    16,    16,    16,  4,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 4>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 4, 4>,       S<4, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,     8,    64,    32,  4,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 4>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 4, 4>,      S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,     8,    32,  4,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<1, 1, 4, 4>,      S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,     8,     8,     8,     4,  4,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<1, 1, 4, 4>,       S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,     8,     8,     8,     4,  4,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<1, 1, 4, 4>,       S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,     8,     8,     8,     4,  4,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<1, 1, 4, 4>,       S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,     8,     8,     8,     4,  4,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 4>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<1, 1, 4, 4>,       S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..8df3bb9641
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_dl_i8_i8_i8_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        // #########|  AData|   BData|   CData|    AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        // #########|   Type|    Type|    Type|       Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        // #########|       |        |        |           |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        // #########|       |        |        |           |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        // MPerBlock=128, NPerBlock=128       
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,    64,   128,    16,  4,          2,          4,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,      S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_i8_i8_i8_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp
new file mode 100644
index 0000000000..3de5458e05
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp
@@ -0,0 +1,81 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instances = std::tuple<
+    // clang-format off
+        // #########|  AData|   BData|   CData|    AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        // #########|   Type|    Type|    Type|       Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        // #########|       |        |        |           |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        // #########|       |        |        |           |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        // MPerBlock=128, NPerBlock=128       
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<4, 4>,       S<4, 2>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<4, 4>,       S<4, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<2, 8>,       S<2, 8>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          2,          4,      1,       S<4, 2>,       S<2, 2>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,      S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<4, 2>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,      S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<2, 1, 4, 4>,        S<4, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,      S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    16,    16,    16,    16,  2,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    16,    16,    16,    16,  2,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t,  int8_t,  int8_t,    int32_t,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<1, 1, 4, 2>,       S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,        S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..8969983356
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_dl_i8_i8_i8_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        // #########|  AData|   BData|   CData|    AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        // #########|   Type|    Type|    Type|       Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        // #########|       |        |        |           |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                |                   |
+        // #########|       |        |        |           |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,    64,   128,    16,  4,          2,          4,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 8, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<16, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,        S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,        S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_i8_i8_i8_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instance.cpp
new file mode 100644
index 0000000000..745a4bb31d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instance.cpp
@@ -0,0 +1,81 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instances = std::tuple<
+    // clang-format off
+        // #########|  AData|   BData|   CData|    AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|       BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        // #########|   Type|    Type|    Type|       Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        // #########|       |        |        |           |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder|  Lengths_K0_N0_N1_K1|               Order|                |                   |
+        // #########|       |        |        |           |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                     |                    |                |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<4, 4>,       S<4, 2>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 8, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=128, NPerBlock=64
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<8, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=128
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<8, 1, 32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          2,          4,      1,       S<4, 2>,       S<2, 2>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 8, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 8, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<2, 1, 4, 4>,       S<4, 1, 8, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    16,    16,    16,    16,  2,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    16,    16,    16,    16,  2,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,        S<4, 1, 4, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,       S<32, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl< int8_t, int8_t, int8_t,      int32_t,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,        S<1, 1, 8, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<4, 1, 1, 2>,      S<0, 3, 1, 2>,        S<1, 1, 1, 2>,      S<1, 1, 4, 2>,        S<4, 1, 2, 1>,   S<0, 3, 1, 2>,  S<0, 3, 1, 2>,      S<1, 1, 4, 1>,      S<0, 3, 1, 2>,        S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..2bda30f82c
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_dl_i8_i8_i8_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        // #########|  AData|   BData|   CData|    AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        // #########|   Type|    Type|    Type|       Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        // #########|       |        |        |           |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        // #########|       |        |        |           |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,   128,    16,  4,          4,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // // MPerBlock=128, NPerBlock=64
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // // MPerBlock=64, NPerBlock=128
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,   256,    64,   128,    16,  4,          2,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=16, NPerBlock=64
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 1, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    16,    64,    16,  2,          1,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 1, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  4>,
+        // MPerBlock=64, NPerBlock=16
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,    16,    16,  2,          4,          1,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    16,    16,    16,    16,  2,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<4, 1>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<4, 1>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  2>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_dl_i8_i8_i8_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instance.cpp
new file mode 100644
index 0000000000..b5e8b8ecd6
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instance.cpp
@@ -0,0 +1,81 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instances = std::tuple<
+    // clang-format off
+        // #########|  AData|   BData|   CData|    AccData| ALayout| BLayout| CLayout|           A|           B|           C|           GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|       ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        // #########|   Type|    Type|    Type|       Type|        |        |        | Elementwise| Elementwise| Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|      DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        // #########|       |        |        |           |        |        |        |   Operation|   Operation|   Operation|               |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder|  Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        // #########|       |        |        |           |        |        |        |            |            |            |               |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                     |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        // MPerBlock=128, NPerBlock=128
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<8, 2>,       S<4, 2>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<4, 4>,       S<4, 2>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   128,   128,   128,    16,  4,          4,          8,      1,       S<2, 8>,       S<2, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 2, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // // MPerBlock=128, NPerBlock=64
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,   128,    64,    16,  4,          4,          2,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // // MPerBlock=64, NPerBlock=128
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<4, 4>,       S<4, 4>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,   256,    64,   128,    16,  4,          2,          4,      1,       S<2, 8>,       S<2, 8>,      S<8, 1, 1, 4>,       S<2, 1, 64, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=64
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<2, 4>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<8, 1>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,    64,     8,  4,          4,          4,      1,       S<4, 2>,       S<8, 1>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 32, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=32, NPerBlock=32
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          2,          4,      1,       S<4, 2>,       S<2, 2>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<4, 2>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    32,    32,    32,     8,  4,          4,          2,      1,       S<2, 2>,       S<2, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>,      S<4, 1, 2, 4>,       S<2, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,        S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=16, NPerBlock=16
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    16,    16,    16,    16,  2,          2,          2,      1,       S<2, 2>,       S<2, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    16,    16,    16,    16,  2,          2,          2,      1,       S<4, 1>,       S<4, 1>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 4, 2>,       S<4, 1, 4, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=64
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,     8,    64,    32,  2,          1,          2,      1,       S<2, 2>,       S<8, 2>,      S<4, 1, 1, 2>,        S<8, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<8, 1, 4, 2>,      S<4, 1, 16, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=64, NPerBlock=8
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,    64,    64,     8,    32,  2,          2,          1,      1,       S<8, 2>,       S<2, 2>,      S<8, 1, 4, 2>,       S<4, 1, 16, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<8, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        // MPerBlock=8, NPerBlock=8
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          1,          2,      1,       S<4, 1>,       S<2, 1>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          1,          2,      1,       S<1, 4>,       S<1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<2, 1>,       S<4, 1>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>,
+        DeviceGemmDl<  int8_t,  int8_t,  int8_t,    int32_t,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,  GemmMNPadding,     8,     8,     8,     4,  2,          2,          1,      1,       S<1, 2>,       S<1, 4>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,        S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>,      S<4, 1, 1, 2>,       S<1, 1, 8, 1>,   S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,        S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,                5,                  1>
+    // clang-format on
+    >;
+
+void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..2e884dfc8a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,100 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                     Version|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..2ca29b1e6f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..706076098d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..5ac458a7b9
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..a644125447
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,  BF16,  BF16,  BF16,     F32,     BF16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..44b6848233
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
@@ -0,0 +1,111 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..23176269c2
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
@@ -0,0 +1,111 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..31a9abe53a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,111 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..201fd93110
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..5d489b207e
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    16,   1,   1,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    16,   1,   1,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    16,   1,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    16,   1,   1,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    16,   1,   1,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    16,   1,   1,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..e09480d57b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|  CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|          _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|          _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                              |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    16,   1,   4,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    16,   1,   4,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    16,   1,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    16,   1,   4,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    16,   1,   4,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    16,   1,   4,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,               S<1, 16, 1, 16>,              4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..34065c334d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    16,   4,   1,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    16,   4,   1,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    16,   4,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    16,   4,   1,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    16,   4,   1,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    16,   4,   1,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..95d7777a79
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp
@@ -0,0 +1,64 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    16,   4,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    16,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    16,   4,   4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    16,   4,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    16,   4,   4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    16,   4,   4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F32,   F32,   F32,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    16,   4,   4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,              4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..d02fb8f70b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
@@ -0,0 +1,69 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //#####################| ALayout| BLayout| CLayout|  AData|  BData|  CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |   Type|   Type|   Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |       |       |       |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |       |       |       |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    64,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,   4,   4,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,  16,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    64,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    64,   4,   4,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    64,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    64,   4,   4,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,   4,   4,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,  16,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,   4,   4,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..abf79262e6
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
@@ -0,0 +1,69 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //#####################| ALayout| BLayout| CLayout|  AData|  BData|  CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |   Type|   Type|   Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |       |       |       |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |       |       |       |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    64,   4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,   4,  16,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,  16,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    64,   4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    64,   4,  16,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    64,   4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    64,   4,  16,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,   4,  16,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,  16,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,   4,  16,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Col,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..5da89c3421
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
@@ -0,0 +1,69 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //#####################| ALayout| BLayout| CLayout|  AData|  BData|  CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |   Type|   Type|   Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |       |       |       |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |       |       |       |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    64,  16,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,  16,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,  16,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    64,  16,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    64,  16,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    64,  16,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    64,  16,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,  16,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,  16,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,  16,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Row,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,             16>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..0ade7a61ce
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
@@ -0,0 +1,66 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#ifdef CK_ENABLE_INT8
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //#####################| ALayout| BLayout| CLayout|  AData|  BData|  CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |   Type|   Type|   Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |       |       |       |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |       |       |       |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    64,  16,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    64,  16,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    64,  16,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    64,  16,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 2>,             16>,
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row, int8_t, int8_t, int8_t, int32_t,  int32_t, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    64,  16,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 2>,             16>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, int8_t, int8_t, int8_t, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/common.hpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/common.hpp
new file mode 100644
index 0000000000..de96a22630
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/common.hpp
@@ -0,0 +1,43 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/utility/data_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault   = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+using InstanceNT = DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>;
+using InstanceNN = DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>;
+using InstanceTT = DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>;
+using InstanceTN = DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>;
+
+template <typename Instance>
+using OwnerList = std::vector<std::unique_ptr<Instance>>;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_add_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_add_instance.cpp
new file mode 100644
index 0000000000..d887b16e60
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_add_instance.cpp
@@ -0,0 +1,38 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Instance  = InstanceNT;
+using Instances = OwnerList<InstanceNT>;
+
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_default_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_interwave_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_default_pipeline_v2_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_default_pipeline_v2_opt_instances(Instances&);
+
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_default_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_interwave_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_default_pipeline_v2_instances(Instances&);
+
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(Instances& instances)
+{
+    add_device_gemm_xdl_f16_f16_f16_km_kn_mn_default_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_km_kn_mn_interwave_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_km_kn_mn_default_pipeline_v2_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_km_kn_mn_default_pipeline_v2_opt_instances(instances);
+
+    add_device_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_default_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_interwave_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_default_pipeline_v2_instances(instances);
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..81fedd50f0
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v1_instance.cpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using Instances =
+    std::tuple<
+        // clang-format off
+        // pipeline v1, 1 wave
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler| Pipeline|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |         |
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |         |
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |         |
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_default_pipeline_v1_instances(
+    OwnerList<InstanceNT>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_instance.cpp
new file mode 100644
index 0000000000..5a0c52c2df
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_instance.cpp
@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using Instances =
+    std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler| Pipeline|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |         |
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |         |
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |         |
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_default_pipeline_v2_instances(
+    OwnerList<InstanceNT>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp
new file mode 100644
index 0000000000..59ffb80bd4
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp
@@ -0,0 +1,35 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using Instances =
+    std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|           LoopScheduler|             Pipeline|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                        |                     |
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                        |                     |
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                        |                     |
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     8,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,  PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_default_pipeline_v2_opt_instances(
+    OwnerList<InstanceNT>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..a64424e8ac
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp
@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using Instances =
+    std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler| Pipeline|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |         |
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |         |
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |         |
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_interwave_pipeline_v1_instances(
+    OwnerList<InstanceNT>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..e313a790e1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v1_instance.cpp
@@ -0,0 +1,32 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// irregular tile size
+using Instances = std::tuple<
+    // clang-format off
+        // pipeline v1, 1 wave
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   64,     16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1, LoopScheduler::Default,          PipelineVersion::v1>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_default_pipeline_v1_instances(
+    OwnerList<InstanceNT>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v2_instance.cpp
new file mode 100644
index 0000000000..050ff6e231
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v2_instance.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// irregular tile size
+using Instances = std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   64,     16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1, LoopScheduler::Default,          PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_default_pipeline_v2_instances(
+    OwnerList<InstanceNT>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..0a0406baec
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// irregular tile size
+using Instances = std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|            LoopScheduler|                   Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                         |                           |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                         |                           |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                         |                           |
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   64,     16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_interwave_pipeline_v1_instances(
+    OwnerList<InstanceNT>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_add_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_add_instance.cpp
new file mode 100644
index 0000000000..7fb1a4845d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_add_instance.cpp
@@ -0,0 +1,38 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Instance  = InstanceNN;
+using Instances = OwnerList<Instance>;
+
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_default_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_interwave_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_default_pipeline_v2_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_default_pipeline_v2_opt_instances(Instances&);
+
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_default_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_interwave_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_default_pipeline_v2_instances(Instances&);
+
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(Instances& instances)
+{
+    add_device_gemm_xdl_f16_f16_f16_km_nk_mn_default_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_km_nk_mn_interwave_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_km_nk_mn_default_pipeline_v2_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_km_nk_mn_default_pipeline_v2_opt_instances(instances);
+
+    add_device_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_default_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_interwave_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_default_pipeline_v2_instances(instances);
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..a0dd60c0f5
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v1_instance.cpp
@@ -0,0 +1,40 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using Instances =
+    std::tuple<
+        // clang-format off
+        // pipeline v1, 1 wave
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_default_pipeline_v1_instances(
+    OwnerList<InstanceNN>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_instance.cpp
new file mode 100644
index 0000000000..122fff4960
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_instance.cpp
@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using Instances =
+    std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_default_pipeline_v2_instances(
+    OwnerList<InstanceNN>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp
new file mode 100644
index 0000000000..9f459aabfc
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp
@@ -0,0 +1,35 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using Instances =
+    std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|              Pipeline|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                      |
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                      |
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                      |
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     8,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,  PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_default_pipeline_v2_opt_instances(
+    OwnerList<InstanceNN>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..3671bea7a3
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp
@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using Instances =
+    std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_interwave_pipeline_v1_instances(
+    OwnerList<InstanceNN>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..d7e7a05505
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v1_instance.cpp
@@ -0,0 +1,32 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// irregular tile size
+using Instances = std::tuple<
+    // clang-format off
+        // pipeline v1, 1 wave
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmMNPadding,   64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_default_pipeline_v1_instances(
+    OwnerList<InstanceNN>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v2_instance.cpp
new file mode 100644
index 0000000000..15e191e5a8
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v2_instance.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// irregular tile size
+using Instances = std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmMNPadding,   64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_default_pipeline_v2_instances(
+    OwnerList<InstanceNN>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..95fc8ecb46
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// irregular tile size
+using Instances = std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmMNPadding,   64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_interwave_pipeline_v1_instances(
+    OwnerList<InstanceNN>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_add_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_add_instance.cpp
new file mode 100644
index 0000000000..e3a7ccefb4
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_add_instance.cpp
@@ -0,0 +1,38 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Instance  = InstanceTT;
+using Instances = OwnerList<Instance>;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_default_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_interwave_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_default_pipeline_v2_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_default_pipeline_v2_opt_instances(Instances&);
+
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_default_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_interwave_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_default_pipeline_v2_instances(Instances&);
+
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(Instances& instances)
+{
+    add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_default_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_interwave_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_default_pipeline_v2_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_default_pipeline_v2_opt_instances(instances);
+
+    add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_default_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_interwave_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_default_pipeline_v2_instances(instances);
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..98db8bad1c
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v1_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using Instances =
+    std::tuple<
+        // clang-format off
+        // pipeline v1, 1 wave
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   256,     4,  8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    32,     4,  8,   32,   32,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,   256,     4,  8,   16,   16,    1,    8,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,   128,     4,  8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,    64,     4,  8,   16,   16,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,    32,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_default_pipeline_v1_instances(
+    OwnerList<InstanceTT>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_instance.cpp
new file mode 100644
index 0000000000..532c348b7e
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_instance.cpp
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using Instances =
+    std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   256,     4,  8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    32,     4,  8,   32,   32,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,   256,     4,  8,   16,   16,    1,    8,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,   128,     4,  8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,    64,     4,  8,   16,   16,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,    32,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_default_pipeline_v2_instances(
+    OwnerList<InstanceTT>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp
new file mode 100644
index 0000000000..b931b8fdfd
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp
@@ -0,0 +1,35 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using Instances =
+    std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|              Pipeline|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                      |
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                      |
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                      |
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     8,  8,   32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Default,  PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_default_pipeline_v2_opt_instances(
+    OwnerList<InstanceTT>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..fa53a3bf0f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using Instances =
+    std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   256,     4,  8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    32,     4,  8,   32,   32,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,   256,     4,  8,   16,   16,    1,    8,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,   128,     4,  8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,    64,     4,  8,   16,   16,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    16,    32,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_interwave_pipeline_v1_instances(
+    OwnerList<InstanceTT>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..bc4b3e801b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp
@@ -0,0 +1,32 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// irregular tile size
+using Instances = std::tuple<
+    // clang-format off
+        // pipeline v1, 1 wave
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_default_pipeline_v1_instances(
+    OwnerList<InstanceTT>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp
new file mode 100644
index 0000000000..0c8e8b2c41
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// irregular tile size
+using Instances = std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_default_pipeline_v2_instances(
+    OwnerList<InstanceTT>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..c9d1913aec
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// irregular tile size
+using Instances = std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_interwave_pipeline_v1_instances(
+    OwnerList<InstanceTT>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_add_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_add_instance.cpp
new file mode 100644
index 0000000000..70bfd0a80c
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_add_instance.cpp
@@ -0,0 +1,38 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Instance  = InstanceTN;
+using Instances = OwnerList<Instance>;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_default_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_interwave_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_default_pipeline_v2_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_default_pipeline_v2_opt_instances(Instances&);
+
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_default_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_interwave_pipeline_v1_instances(Instances&);
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_default_pipeline_v2_instances(Instances&);
+
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(Instances& instances)
+{
+    add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_default_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_interwave_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_default_pipeline_v2_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_default_pipeline_v2_opt_instances(instances);
+
+    add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_default_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_interwave_pipeline_v1_instances(instances);
+    add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_default_pipeline_v2_instances(instances);
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..0f4ccceb50
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v1_instance.cpp
@@ -0,0 +1,44 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using Instances = std::tuple<
+    // clang-format off
+        // pipeline v1, 1 wave
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_default_pipeline_v1_instances(
+    OwnerList<InstanceTN>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_instance.cpp
new file mode 100644
index 0000000000..134755299d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_instance.cpp
@@ -0,0 +1,46 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using Instances = std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_default_pipeline_v2_instances(
+    OwnerList<InstanceTN>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp
new file mode 100644
index 0000000000..76d71ee97a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using Instances = std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|           LoopScheduler|             Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                        |                     |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                        |                     |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                        |                     |
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     8,  8,   32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,  PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_default_pipeline_v2_opt_instances(
+    OwnerList<InstanceTN>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..0410eabb70
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp
@@ -0,0 +1,46 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using Instances = std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_interwave_pipeline_v1_instances(
+    OwnerList<InstanceTN>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..e6e73cdcca
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// irregular tile size
+using Instances = std::tuple<
+    // clang-format off
+        // pipeline v1, 1 wave
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   144,     8,  8,   16,   16,    2,    9,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8,  8, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   144,     4,  8,   16,   16,    2,    9,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v1>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_default_pipeline_v1_instances(
+    OwnerList<InstanceTN>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp
new file mode 100644
index 0000000000..78f18fd582
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp
@@ -0,0 +1,36 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// irregular tile size
+using Instances = std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   144,     8,  8,   16,   16,    2,    9,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8,  8, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   144,     4,  8,   16,   16,    2,    9,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,           1,  LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_default_pipeline_v2_instances(
+    OwnerList<InstanceTN>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
new file mode 100644
index 0000000000..a41919aab7
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
@@ -0,0 +1,36 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// irregular tile size
+using Instances = std::tuple<
+// clang-format off
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        //###########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer| NumPrefetch|          LoopScheduler|                     Pipeline|
+        //###########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|            |                       |                             |
+        //###########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |            |                       |                             |
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   144,     8,  8,   16,   16,    2,    9,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8,  8, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmXdl<   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   144,     4,  8,   16,   16,    2,    9,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,           1,  LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_interwave_pipeline_v1_instances(
+    OwnerList<InstanceTN>& instances)
+{
+    add_device_operation_instances(instances, Instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..a590413acc
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_xdl_f32_f32_f32_km_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f32_f32_f32_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..1d010d1b07
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_xdl_f32_f32_f32_km_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f32_f32_f32_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..f108b75342
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp
@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..b0b4bc012d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp
@@ -0,0 +1,65 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>,
+        DeviceGemmXdl<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..df3bd94fca
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp
@@ -0,0 +1,56 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64 = double;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_xdl_f64_f64_f64_km_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  2,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  2,   16,   16,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  2,   16,   16,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              2,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Row, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f64_f64_f64_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..73b4e77666
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp
@@ -0,0 +1,56 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64 = double;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_xdl_f64_f64_f64_km_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  2,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  2,   16,   16,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Col,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  2,   16,   16,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              2,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f64_f64_f64_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Col, Col, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f64_f64_f64_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..76137a1c3e
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp
@@ -0,0 +1,56 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64 = double;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_f64_f64_f64_mk_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  2,   16,   16,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Row,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  2,   16,   16,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f64_f64_f64_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Row, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f64_f64_f64_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..f0158d8f3d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F64 = double;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
+        //##########|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar|
+        //##########|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|
+        //##########|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  2,   16,   16,    4,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  2,   16,   16,    4,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  2,   16,   16,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  2,   16,   16,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  2,   16,   16,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  2,   16,   16,    2,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  2,   16,   16,    4,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>,
+        DeviceGemmXdl<  F64,   F64,   F64,     F64,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  2,   16,   16,    2,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1>
+        // clang-format on
+        >;
+
+void add_device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemm<Row, Col, Row, F64, F64, F64, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_gemm_xdl_f64_f64_f64_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt
new file mode 100644
index 0000000000..9028829fe9
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt
@@ -0,0 +1,8 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+add_instance_library(device_gemm_add_add_fastgelu_instance
+   device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
+   device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
+   device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
+   device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
+)
+endif()
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..125fbc21a2
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
@@ -0,0 +1,167 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// outout: e[m, n]
+// input: a[k, m], b[k, n], d0[m, n], d1[m, n]
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_generic_instance =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+        // clang-format on
+        >;
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+// irregular tile size
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_generic_instance{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..cc33692d73
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
@@ -0,0 +1,167 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// outout: e[m, n]
+// input: a[k, m], b[n, k], d0[m, n], d1[m, n]
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_generic_instance =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+        // clang-format on
+        >;
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+// irregular tile size
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,      PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_generic_instance{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..704787a080
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
@@ -0,0 +1,167 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// outout: e[m, n]
+// input: a[m, k], b[k, n], d0[m, n], d1[m, n]
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_generic_instance =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+        // clang-format on
+        >;
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+// irregular tile size
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_generic_instance{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..d64c9ec5e1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
@@ -0,0 +1,158 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// outout: e[m, n]
+// input: a[m, k], b[n, k], d0[m, n], d1[m ,n]
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_generic_instance =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+        // clang-format on
+        >;
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES         
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+// irregular tile size
+using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,   F16, PassThrough, PassThrough, AddAddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddAddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_generic_instance{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt
new file mode 100644
index 0000000000..1085966807
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt
@@ -0,0 +1,8 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+add_instance_library(device_gemm_add_fastgelu_instance
+   device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
+   device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
+   device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
+   device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
+)
+endif()
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..e68bd8e7e4
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
@@ -0,0 +1,150 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0)
+// outout: e[m, n]
+// input: a[k, m], b[k, n], d0[m, n]
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_generic_instance =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+        // clang-format on
+        >;
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,      PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_generic_instance{});
+    add_device_operation_instances(
+        instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
new file mode 100644
index 0000000000..5aaa2e8fe5
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
@@ -0,0 +1,150 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// outout: e[m, n]
+// input: a[k, m], b[n, k], d0[m, n], d1[m, n]
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_generic_instance =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+        // clang-format on
+        >;
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,      PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_generic_instance{});
+    add_device_operation_instances(
+        instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..7a2a3dbaf3
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
@@ -0,0 +1,150 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// outout: e[m, n]
+// input: a[m, k], b[k, n], d0[m, n], d1[m, n]
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_generic_instance =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+        // clang-format on
+        >;
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_generic_instance{});
+    add_device_operation_instances(
+        instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
new file mode 100644
index 0000000000..fa33609978
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
@@ -0,0 +1,141 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// outout: e[m, n]
+// input: a[m, k], b[n, k], d0[m, n], d1[m ,n]
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_generic_instance =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+        // clang-format on
+        >;
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES         
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_Tuple,   F16, PassThrough, PassThrough, AddFastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col,     Row_Tuple,    Row,   F16,   F16,     F32,      F32,     F16_Tuple,   F16, PassThrough, PassThrough,    AddFastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddFastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_generic_instance{});
+    add_device_operation_instances(
+        instances, device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_multiply/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/CMakeLists.txt
new file mode 100644
index 0000000000..eb9345cbad
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_gemm_add_multiply_instance
+   device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
+   device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
+   device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
+   device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..d9ae781791
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
@@ -0,0 +1,106 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddMultiply = ck::tensor_operation::element_wise::AddMultiply;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        // no padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|           CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|   Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|     Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |              |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+
+        // M/N/K Padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|           CDE|            GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|   Elementwise|  Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|     Operation|                |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |              |                |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+        // clang-format on
+        >;
+
+void add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..27c6cbe8ab
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
@@ -0,0 +1,106 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddMultiply = ck::tensor_operation::element_wise::AddMultiply;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        // no padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+
+        // M/N/K Padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+        // clang-format on
+        >;
+
+void add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..ff5f0e94a1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
@@ -0,0 +1,106 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddMultiply = ck::tensor_operation::element_wise::AddMultiply;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        // no padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+
+        // M/N/K padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+        // clang-format on
+        >;
+
+void add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..dba625e0ad
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
@@ -0,0 +1,143 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddMultiply = ck::tensor_operation::element_wise::AddMultiply;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        // no padding
+        // N % 8 == 0 && K % 8 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // M/N/K padding
+        // N % 8 == 0 && K % 8 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // M/N/K padding
+        // N % 4 == 0 && K % 4 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+
+        // M/N/K padding
+        // N % 8 == 0 && K % 1 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    AddMultiply, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>
+
+        // clang-format on
+        >;
+
+void add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/CMakeLists.txt
new file mode 100644
index 0000000000..6079f90130
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/CMakeLists.txt
@@ -0,0 +1,8 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+add_instance_library(device_gemm_add_relu_add_layernorm_instance
+   device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp
+   device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp
+   device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp
+   device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp
+)
+endif()
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..28a452c1a1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp
@@ -0,0 +1,130 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// h = layernorm(e, gamma, beta)
+// outout: h[m, n]
+// input: a[k, m], b[k, n], d0[m, n], d1[m, n], gamma[n], beta[n]
+template <LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances = std::tuple<
+    // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|            CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|          PostShuffle|     PostShuffle|            Layernorm|       Layernorm|     LoopScheduler|     Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise|    Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|      Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                 _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |               |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |                |                     |                |                  |             |
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|            CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|           PostShuffle|     PostShuffle|            Layernorm|       Layernorm|          LoopScheduler|              Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise|    Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|  ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                       |                      |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|      Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                  _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                       |                      |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |               |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                      |                |                     |                |                       |                      |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,  PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Interwave, PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,     AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,  PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Col,
+                                                             Row,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v1>{});
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances<
+            LoopScheduler::Interwave,
+            PipelineVersion::v1>{});
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v2>{});
+#endif
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..13366238d6
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp
@@ -0,0 +1,130 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// h = layernorm(e, gamma, beta)
+// outout: h[m, n]
+// input: a[k, m], b[k, n], d0[m, n], d1[m, n], gamma[n], beta[n]
+template <LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances = std::tuple<
+    // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|         CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|          PostShuffle|     PostShuffle|            Layernorm|       Layernorm|     LoopScheduler|     Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                 _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |                |                     |                |                  |             |
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|         CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|          PostShuffle|     PostShuffle|            Layernorm|       Layernorm|          LoopScheduler|               Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                       |                       |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                 _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                       |                       |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |                |                     |                |                       |                       |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,              S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,   PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,              S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Interwave, PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Col,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,              S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,   PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Col,
+                                                             Col,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v1>{});
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances<
+            LoopScheduler::Interwave,
+            PipelineVersion::v1>{});
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v2>{});
+#endif
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..8a4889ee83
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp
@@ -0,0 +1,130 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// h = layernorm(e, gamma, beta)
+// outout: h[m, n]
+// input: a[k, m], b[k, n], d0[m, n], d1[m, n], gamma[n], beta[n]
+template <LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances = std::tuple<
+    // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|         CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|          PostShuffle|     PostShuffle|            Layernorm|       Layernorm|     LoopScheduler|     Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                 _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |                |                     |                |                  |             |
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|         CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|          PostShuffle|     PostShuffle|            Layernorm|       Layernorm|          LoopScheduler|              Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                       |                      |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                 _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                       |                      |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |                |                     |                |                       |                      |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,  PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Interwave, PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Row, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,  PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Row,
+                                                             Row,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v1>{});
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances<
+            LoopScheduler::Interwave,
+            PipelineVersion::v1>{});
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v2>{});
+#endif
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..fc3cbcf905
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp
@@ -0,0 +1,127 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16           = ck::half_t;
+using F32           = float;
+using F16_F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row           = ck::tensor_layout::gemm::RowMajor;
+using Col           = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b), d0, d1)
+// h = layernorm(e, gamma, beta)
+// outout: h[m, n]
+// input: a[k, m], b[k, n], d0[m, n], d1[m, n], gamma[n], beta[n]
+template <LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances = std::tuple<
+    // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|         CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|          PostShuffle|     PostShuffle|            Layernorm|       Layernorm|     LoopScheduler|     Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                 _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                  |             |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |                |                     |                |                  |             |
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 4>,               8,             S<16, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 8>,               8,             S<32, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<32, 4>,               8,             S<32, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 8>,               8,             S<16, 8>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 4>,               8,             S<16, 4>,               1, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,             S<16, 4>,               8,             S<16, 4>,               1, GemmLoopScheduler, GemmPipeline>
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //#######################################|      A|      B|            Ds|      H| AData| BData| AccData| CShuffle|        DsData| EMeanVarData| GammaData|  BetaData| HData|           A|           B|         CDE|           H|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|          PostShuffle|     PostShuffle|            Layernorm|       Layernorm|          LoopScheduler|              Pipeline|
+        //#######################################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|         Type|      Type|      Type|  Type| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ThreadClusterLengths| ScalarPerVector| ThreadClusterLengths| ThreadSliceSize|                       |                      |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |   Operation|   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|                 _M_N|   _NWaveNPerXdl|                 _M_N|              _M|                       |                      |
+        //#######################################|       |       |              |       |      |      |        |         |              |             |          |          |      |            |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                     |                |                     |                |                       |                      |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,  PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Interwave, PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleDLayernorm_Xdl_CShuffle<    Row,    Col, Row_Row_Tuple,    Row,   F16,   F16,     F32,      F32, F16_F16_Tuple,          F16,       F16,       F16,   F16, PassThrough, PassThrough,  AddReluAdd, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<16, 4>,              1,             S<16, 4>,               1, LoopScheduler::Default,  PipelineVersion::v2>
+#endif
+        // clang-format on
+        >;
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Row,
+                                                             Col,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v1>{});
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances<
+            LoopScheduler::Interwave,
+            PipelineVersion::v1>{});
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances<
+            LoopScheduler::Default,
+            PipelineVersion::v2>{});
+#endif
+    add_device_operation_instances(
+        instances,
+        device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt
new file mode 100644
index 0000000000..ccada3a85e
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_gemm_bias_add_reduce_instance
+    device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
+    device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
+    device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
+    device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..bfb95bce85
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[k, m] * b[k, n]
+using device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|C0Data|C1Data|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|          C1|    Reduce|      ReduceInEleOp|      ReduceAccEleOp|      Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise| Elementwise| Operation|                   |                    |  MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |                   |   Operation|   Operation|   Operation|   Operation|          |                   |                    |   Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |                   |            |            |            |            |          |                   |                    |            |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
+    std::vector<DeviceGemmReducePtr<1, ReduceOps::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..d0352339cf
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[k, m] * b[n, k]
+using device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|C0Data|C1Data|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|          C1|    Reduce|      ReduceInEleOp|      ReduceAccEleOp|      Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise| Elementwise| Operation|                   |                    |  MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |                   |   Operation|   Operation|   Operation|   Operation|          |                   |                    |   Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |                   |            |            |            |            |          |                   |                    |            |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
+    std::vector<DeviceGemmReducePtr<1, ReduceOps::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..d5b298ab21
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[m, k] * b[n, k]
+using device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|C0Data|C1Data|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|          C1|    Reduce|      ReduceInEleOp|      ReduceAccEleOp|      Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise| Elementwise| Operation|                   |                    |  MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |                   |   Operation|   Operation|   Operation|   Operation|          |                   |                    |   Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |                   |            |            |            |            |          |                   |                    |            |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+    std::vector<DeviceGemmReducePtr<1, ReduceOps::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..80c8f018f4
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[m, k] * b[n, k]
+using device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances =
+    std::tuple<
+        // clang-format off
+        //##################################| ALayout| BLayout| CLayout|AData| BData| CData|C0Data|C1Data|  GemmAcc| CShuffle| ReduceAcc|         DData|           A|           B|           C|          C1|         Dxs|    DxsInEleOp|    DxsAccEleOp|           D|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //##################################|        |        |        | Type|  Type|  Type|  Type|  Type| DataType| DataType|  DataType|    Type Tuple| Elementwise| Elementwise| Elementwise| Elementwise|      Reduce|              |               |  MemoryData| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |              |   Operation|   Operation|   Operation|   Operation|   Operation|              |               |   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //##################################|        |        |        |     |      |      |      |      |         |         |          |              |            |            |            |            |            |              |               |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>,
+        DeviceGemmBiasAddReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>
+        // clang-format on
+        >;
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+    std::vector<DeviceGemmReducePtr<1, ReduceOps::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bilinear/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_bilinear/CMakeLists.txt
new file mode 100644
index 0000000000..b9f1210b2e
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_bilinear/CMakeLists.txt
@@ -0,0 +1,8 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+add_instance_library(device_gemm_bilinear_instance
+   device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
+   device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
+   device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
+   device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
+)
+endif()
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..74ec9e1f8d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
@@ -0,0 +1,105 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<ck::half_t>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e[m, n] = bilinear(a[k, m] * b[k, n], d[m, n])
+using device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances = std::tuple<
+    // clang-format off
+        // no padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+
+        // M/N/K Padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
new file mode 100644
index 0000000000..eb98b3e7e8
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
@@ -0,0 +1,105 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<ck::half_t>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e[m, n] = bilinear(a[k, m] * b[n, k], d[m, n])
+using device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances = std::tuple<
+    // clang-format off
+        // no padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+
+        // M/N/K Padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..5f4a90125a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
@@ -0,0 +1,105 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<ck::half_t>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e[m, n] = bilinear(a[m, k] * b[k, n], d[m, n])
+using device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances = std::tuple<
+    // clang-format off
+        // no padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+
+        // M/N/K padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
new file mode 100644
index 0000000000..38e3897d6a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
@@ -0,0 +1,142 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<ck::half_t>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e[m, n] = bilinear(a[m, k] * b[n, k], d[m, n])
+using device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances = std::tuple<
+    // clang-format off
+        // no padding
+        // N % 8 == 0 && K % 8 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // M/N/K padding
+        // N % 8 == 0 && K % 8 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+
+        // M/N/K padding
+        // N % 4 == 0 && K % 4 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1,  8>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1, 16>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1,  8, 1,  8>,               4>,
+
+        // M/N/K padding
+        // N % 8 == 0 && K % 1 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    Bilinear, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>
+
+    // clang-format on
+    >;
+
+void add_device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Bilinear>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt
new file mode 100644
index 0000000000..772373dcb0
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt
@@ -0,0 +1,8 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+add_instance_library(device_gemm_fastgelu_instance
+   device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
+   device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
+   device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
+   device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
+)
+endif()
diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..803c44c7f5
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
@@ -0,0 +1,147 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b))
+// outout: e[m, n]
+// input: a[k, m], b[k, n]
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_generic_instance = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+    // clang-format on
+    >;
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,      PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_generic_instance{});
+    add_device_operation_instances(
+        instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..9b9ef3db24
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
@@ -0,0 +1,147 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b))
+// outout: e[m, n]
+// input: a[k, m], b[k, n]
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_generic_instance = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+    // clang-format on
+    >;
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_generic_instance{});
+    add_device_operation_instances(
+        instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..1a0b6c9d1c
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,147 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b))
+// outout: e[m, n]
+// input: a[m, k], b[k, n]
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_generic_instance = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+    // clang-format on
+    >;
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,      PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_generic_instance{});
+    add_device_operation_instances(
+        instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..18b1c0e993
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,138 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/utility/sequence.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmDefault    = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// e = elementwise((a * b))
+// outout: e[m, n]
+// input: a[m, k], b[n, k]
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_generic_instance = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+    // clang-format on
+    >;
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES         
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Interwave,        PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F32, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+// irregular tile size
+using device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|            Ds|      E| AData| BData| AccData| CShuffle|        DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|          LoopScheduler|                    Pipeline|
+        //##############################| Layout| Layout|        Layout| Layout|  Type|  Type|    Type| DataType|          Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                       |                            |
+        //##############################|       |       |              |       |      |      |        |         |              |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                       |                            |
+        // pipeline v1, 1 wave
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v1>
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES        
+        // pipeline v1, 2 waves
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Interwave,      PipelineVersion::v1>
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES        
+        // pipeline v2, 1 wave
+        ,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col,   Empty_Tuple,    Row,   F16,   F16,     F32,      F32,   Empty_Tuple,   F16, PassThrough, PassThrough,       FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,               1, LoopScheduler::Default,        PipelineVersion::v2>
+#endif
+    // clang-format on
+    >;
+
+void add_device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    Empty_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    FastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_generic_instance{});
+    add_device_operation_instances(
+        instances, device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_add/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/CMakeLists.txt
new file mode 100644
index 0000000000..a4f7443455
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/CMakeLists.txt
@@ -0,0 +1,7 @@
+add_instance_library(device_gemm_multiply_add_instance
+   device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
+   device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
+
+   device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp
+   device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..bb38b33405
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        // M/N/K padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+        // clang-format on
+        >;
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..4be9b51aa7
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16       = ck::half_t;
+using F32       = float;
+using F16_Tuple = ck::Tuple<F16, F16>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        // M/N/K padding
+        // N % 8 == 0 && K % 1 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,   F16,     F32,      F16, F16_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>
+
+        // clang-format on
+        >;
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..7946bafc28
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F8        = ck::f8_t;
+using F16       = ck::half_t;
+using F32       = float;
+using F32_Tuple = ck::Tuple<F32, F32>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        // M/N/K padding
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+        // clang-format on
+        >;
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F8,
+                                                    F32_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp
new file mode 100644
index 0000000000..99db9f4270
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F8        = ck::f8_t;
+using F16       = ck::half_t;
+using F32       = float;
+using F32_Tuple = ck::Tuple<F32, F32>;
+
+using Row       = ck::tensor_layout::gemm::RowMajor;
+using Col       = ck::tensor_layout::gemm::ColumnMajor;
+using Row_Tuple = ck::Tuple<Row, Row>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances =
+    std::tuple<
+        // clang-format off
+        // M/N/K padding
+        // N % 8 == 0 && K % 1 == 0
+        //##############################|      A|      B|        Ds|      E| AData| BData| AccData| CShuffle|    DsData| EData|           A|           B|            CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //##############################| Layout| Layout|    Layout| Layout|  Type|  Type|    Type| DataType|      Type|  Type| Elementwise| Elementwise|    Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //##############################|       |       |          |       |      |      |        |         |          |      |   Operation|   Operation|      Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##############################|       |       |          |       |      |      |        |         |          |      |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  4, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 64>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Row_Tuple,    Row,   F16,    F8,     F32,      F32, F32_Tuple,   F16, PassThrough, PassThrough,    MultiplyAdd, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,              S<1,  2, 1, 32>,               1>
+
+        // clang-format on
+        >;
+
+void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F8,
+                                                    F32_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    MultiplyAdd>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_reduce/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_reduce/CMakeLists.txt
new file mode 100644
index 0000000000..2b2cf8c774
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_reduce/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_gemm_reduce_instance
+    device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
+    device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
+    device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
+    device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..f0e7b6ab43
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[k, m] * b[k, n]
+using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //###########################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|    Reduce|      ReduceInEleOp|      ReduceAccEleOp|      Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //###########################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise| Operation|                   |                    |  MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|          |                   |                    |   Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |            |            |            |          |                   |                    |            |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+    // clang-format on
+    >;
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..56815b9ac5
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[k, m] * b[n, k]
+using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //###########################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|    Reduce|      ReduceInEleOp|      ReduceAccEleOp|      Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //###########################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise| Operation|                   |                    |  MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|          |                   |                    |   Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |            |            |            |          |                   |                    |            |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,     false,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Col,      Col,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+    // clang-format on
+    >;
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..e66d46a269
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[m, k] * b[n, k]
+using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //###########################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|    Reduce|      ReduceInEleOp|      ReduceAccEleOp|      Reduce|          GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //###########################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise| Operation|                   |                    |  MemoryData|Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|          |                   |                    |   Operation|              |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |            |            |            |          |                   |                    |            |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,     false,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,      Row,    Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough,   ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,   GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>
+    // clang-format on
+    >;
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..fb1dfac69c
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
@@ -0,0 +1,77 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16              = ck::half_t;
+using F32              = float;
+using ReducePtrsGlobal = ck::Tuple<F32*, F32*>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using ReduceSum   = ck::reduce::Add;
+using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
+
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// c[m, n] = a[m, k] * b[n, k]
+using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //###########################| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         ReduceData|           A|           B|           C|    Reduce|      ReduceInEleOp|      ReduceAccEleOp|      Reduce|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+        //###########################|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|         Type Tuple| Elementwise| Elementwise| Elementwise| Operation|                   |                    |  MemoryData| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |   Operation|   Operation|   Operation|          |                   |                    |   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+        //###########################|        |        |        |     |      |      |         |         |          |                   |            |            |            |          |                   |                    |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 32, 1, 4>,               8,             S<64, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 8>,               8,             S<32, 4>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>,
+        DeviceGemmReduce_Xdl_CShuffle<     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   ReducePtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, ReduceInElementOps, ReduceOutElementOps, ReduceMemOp,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,               S<1, 16, 1, 4>,               8,             S<32, 2>,                         4,                            1>
+    // clang-format on
+    >;
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+    std::vector<DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt
new file mode 100644
index 0000000000..89dfa8f2ed
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt
@@ -0,0 +1,28 @@
+set(GEMM_SPLITK_INSTANCES)
+
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp)
+endif()
+
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp)
+endif()
+
+if(DTYPES MATCHES "fp16" OR DTYPES MATCHES "fp8" OR NOT DEFINED DTYPES)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_instance.cpp)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f8_f16_f16_mk_nk_mn_instance.cpp)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f8_f16_f16_km_kn_mn_instance.cpp)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f8_f16_f16_km_nk_mn_instance.cpp)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instance.cpp)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instance.cpp)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f16_f8_f16_km_kn_mn_instance.cpp)
+  list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instance.cpp)
+endif()
+
+add_instance_library(device_gemm_splitk_instance ${GEMM_SPLITK_INSTANCES})
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..fed2cbbfb9
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              1,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..44ac4c08cd
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              1,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..218d6e0c2a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,104 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+using device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_generic_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              1,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    32,    32,     4,  8,   32,   32,    1,    1,  S<1, 2, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              1,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               2>
+    // clang-format on
+    >;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        //PipelineVersion::v1
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,    64,   192,     4,  8,   32,   32,    1,    3,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 48, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   192,    64,     4,  8,   32,   32,    3,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    32,   192,     4,  8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 24, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   192,    32,     4,  8,   32,   32,    3,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    32,    64,     4,  8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    64,    32,     4,  8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    32,    32,     4,  8,   32,   32,    1,    1,  S<1, 2, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,    64,    16,    32,     4,  8,   16,   16,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               4,  F16, PipelineVersion::v1>,
+
+        //PipelineVersion::v2
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,    64,   192,     4,  8,   32,   32,    1,    3,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 48, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   192,    64,     4,  8,   32,   32,    3,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    32,   192,     4,  8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 24, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   192,    32,     4,  8,   32,   32,    3,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    32,    64,     4,  8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    64,    32,     4,  8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,  F16, PipelineVersion::v2>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_generic_instances{});
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..b87f7ff309
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,95 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_generic_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              1,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              1,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               2,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              1,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              1,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               2,  F16, PipelineVersion::v1>
+    // clang-format on
+    >;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        //PipelineVersion::v1
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,  F16, PipelineVersion::v1>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,  F16, PipelineVersion::v1>,
+
+        //PipelineVersion::v2
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,  F16, PipelineVersion::v2>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,  F16, PipelineVersion::v2>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_generic_instances{});
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..86a86c1035
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_km_kn_mn_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f16_f8_f16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer| Compute|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|    Type|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|        |
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |        |
+        DeviceGemmXdlSplitKCShuffle<  F16,   F8,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,     F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F8,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,     F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F8,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,     F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F8,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,     F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F8,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,     F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F8,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,     F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F8,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,     F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,   F8,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              1,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,     F16>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f16_f8_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f16_f8_f16_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..c423b493eb
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer| Compute|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|    Type|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|        |
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |        |
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              1,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..c0b6ec67bb
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,71 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer| Compute|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|    Type|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|        |
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |        |
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,    64,   192,     4,  8,   32,   32,    1,    3,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 48, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   192,    64,     4,  8,   32,   32,    3,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    32,   192,     4,  8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 24, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   192,    32,     4,  8,   32,   32,    3,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    32,    64,     4,  8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    64,    32,     4,  8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,    F16>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Row, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..fa38640ae2
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer| Compute|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|    Type|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|        |
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |        |
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<  F16,    F8,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,    F16>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..9d15ccd362
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              4,      true,           1,           1,                   S<1, 16, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              1,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..4e9ad58742
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
+using device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              4,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              1,              4,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..330e5aff90
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp
@@ -0,0 +1,61 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              4,      true,           1,           1,                   S<1, 16, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..0db3a15d2a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp
@@ -0,0 +1,66 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
+using device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  4,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  4,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  4,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  4,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  4,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 32, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  4,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 8>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  4,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 4>,               4>,
+        DeviceGemmXdlSplitKCShuffle<  F32,   F32,   F32,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  4,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              4,              4,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              4,              4,      true,           1,           1,                   S<1, 16, 1, 4>,               4>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f8_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f8_f16_f16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..face3f456d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f8_f16_f16_km_kn_mn_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f8_f16_f16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer| Compute|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|    Type|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|        |
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |        |
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,     Row,     Row, PassThrough, PassThrough, PassThrough,    GemmDefault,  256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              1,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f8_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f8_f16_f16_km_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f8_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f8_f16_f16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..727bd221a7
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f8_f16_f16_km_nk_mn_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f8_f16_f16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer| Compute|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|    Type|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|        |
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |        |
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              4,              8,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              2,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,              2,              1,              8,      true,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f8_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Col, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f8_f16_f16_km_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..6f378fd281
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,71 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer| Compute|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|    Type|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|        |
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |        |
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,    64,   192,     4,  8,   32,   32,    1,    3,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 48, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   192,    64,     4,  8,   32,   32,    3,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    32,   192,     4,  8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 24, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   192,    32,     4,  8,   32,   32,    3,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    32,    64,     4,  8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    64,    32,     4,  8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough, GemmMNPadding,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,    F16>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Row, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f8_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f8_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..70f4bdc0ec
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f8_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_splitk_f8_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer| Compute|
+        //#########################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Specialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|    Type|
+        //#########################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|        |
+        //#########################|     |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |        |
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   256,     4,  8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,   128,     4,  8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    64,   128,     4,  8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    64,     4,  8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,    64,     4,  8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,   128,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 32, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 8>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    64,    32,     4,  8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,    F16>,
+        DeviceGemmXdlSplitKCShuffle<   F8,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,    64,    32,    64,     4,  8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,      true,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             3,              8,              8,      true,           1,           1,                   S<1, 16, 1, 4>,               8,    F16>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_splitk_f8_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmSplitK<Row, Col, Row, F8, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_splitk_f8_f16_f16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/gemm_streamk/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_streamk/CMakeLists.txt
new file mode 100644
index 0000000000..d9baf3f062
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_streamk/CMakeLists.txt
@@ -0,0 +1,12 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+add_instance_library(device_gemm_streamk_instance
+   # device_gemm_xdl_streamk_f32_f32_f32_mk_kn_mn_instance.cpp
+   # device_gemm_xdl_streamk_f32_f32_f32_mk_nk_mn_instance.cpp
+   # device_gemm_xdl_streamk_f32_f32_f32_km_kn_mn_instance.cpp
+   # device_gemm_xdl_streamk_f32_f32_f32_km_nk_mn_instance.cpp
+   device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp
+   # device_gemm_xdl_streamk_f16_f16_f16_mk_nk_mn_instance.cpp
+   # device_gemm_xdl_streamk_f16_f16_f16_km_kn_mn_instance.cpp
+   # device_gemm_xdl_streamk_f16_f16_f16_km_nk_mn_instance.cpp
+)
+endif()
diff --git a/library/src/tensor_operation_instance/gpu/gemm_streamk/device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_streamk/device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..60e5d8f5fa
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/gemm_streamk/device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_streamk.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+// static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+// static constexpr auto GemmMNPadding =
+// ck::tensor_operation::device::GemmSpecialization::MNPadding;
+using device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_generic_instances = std::tuple<
+    // clang-format off
+        //##################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##################|     |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,                   S<1, 16, 1, 8>,               2>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,                   S<1, 16, 1, 8>,               2>
+    // clang-format on
+    >;
+
+// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
+using device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //##################|AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
+        //##################| Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
+        //##################|     |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
+        //##################|     |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   192,     4,  8,   32,   32,    1,    3,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 48, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   192,    64,     4,  8,   32,   32,    3,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    64,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,                   S<1, 32, 1, 4>,               8>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,   128,     4,  8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,                   S<1, 16, 1, 4>,               8>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,                   S<1, 32, 1, 8>,               8>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   192,     4,  8,   32,   32,    1,    3,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 24, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,         1,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   192,    32,     4,  8,   32,   32,    3,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,                   S<1, 32, 1, 4>,               8>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,    64,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    64,    32,     4,  8,   32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,                   S<1, 32, 1, 4>,               8>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    32,   128,     4,  8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,                   S<1, 16, 1, 8>,               8>,
+        DeviceGemmXdlStreamK<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,    32,     4,  8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,                   S<1, 32, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmStreamK<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_generic_instances{});
+    add_device_operation_instances(instances,
+                                   device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt
new file mode 100644
index 0000000000..3808e0248f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt
@@ -0,0 +1,5 @@
+add_instance_library(device_grouped_conv1d_bwd_weight_instance
+    device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
+    device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
+    device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
new file mode 100644
index 0000000000..25ea4f48cf
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
+                                                           GNWC,
+                                                           GKXC,
+                                                           GNWK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances<1,
+                                                                    GNWC,
+                                                                    GKXC,
+                                                                    GNWK,
+                                                                    ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances<
+                                       1,
+                                       GNWC,
+                                       GKXC,
+                                       GNWK,
+                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
new file mode 100644
index 0000000000..8065012f12
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
+                                                           GNWC,
+                                                           GKXC,
+                                                           GNWK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_instances<1,
+                                                                   GNWC,
+                                                                   GKXC,
+                                                                   GNWK,
+                                                                   ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_instances<
+                                       1,
+                                       GNWC,
+                                       GKXC,
+                                       GNWK,
+                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
new file mode 100644
index 0000000000..c70a54c2dc
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<1,
+                                                           GNWC,
+                                                           GKXC,
+                                                           GNWK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_instances<1,
+                                                                   GNWC,
+                                                                   GKXC,
+                                                                   GNWK,
+                                                                   ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_instances<
+                                       1,
+                                       GNWC,
+                                       GKXC,
+                                       GNWK,
+                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/CMakeLists.txt
new file mode 100644
index 0000000000..1d90593e37
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_instance_library(device_grouped_conv1d_fwd_instance
+   device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
+   device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
+   device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
+   device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
new file mode 100644
index 0000000000..032ebc1eff
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<1,
+                                                                              GNWC,
+                                                                              GKXC,
+                                                                              Empty_Tuple,
+                                                                              GNWK,
+                                                                              ConvFwdDefault>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<1,
+                                                                              GNWC,
+                                                                              GKXC,
+                                                                              Empty_Tuple,
+                                                                              GNWK,
+                                                                              ConvFwd1x1P0>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<1,
+                                                                              GNWC,
+                                                                              GKXC,
+                                                                              Empty_Tuple,
+                                                                              GNWK,
+                                                                              ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
new file mode 100644
index 0000000000..3a0ddb736f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<1,
+                                                                             GNWC,
+                                                                             GKXC,
+                                                                             Empty_Tuple,
+                                                                             GNWK,
+                                                                             ConvFwdDefault>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<1,
+                                                                             GNWC,
+                                                                             GKXC,
+                                                                             Empty_Tuple,
+                                                                             GNWK,
+                                                                             ConvFwd1x1P0>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<1,
+                                                                             GNWC,
+                                                                             GKXC,
+                                                                             Empty_Tuple,
+                                                                             GNWK,
+                                                                             ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
new file mode 100644
index 0000000000..1db1226b7f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<1,
+                                                                             GNWC,
+                                                                             GKXC,
+                                                                             Empty_Tuple,
+                                                                             GNWK,
+                                                                             ConvFwdDefault>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<1,
+                                                                             GNWC,
+                                                                             GKXC,
+                                                                             Empty_Tuple,
+                                                                             GNWK,
+                                                                             ConvFwd1x1P0>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<1,
+                                                                             GNWC,
+                                                                             GKXC,
+                                                                             Empty_Tuple,
+                                                                             GNWK,
+                                                                             ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp
new file mode 100644
index 0000000000..6fbf631766
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<1,
+                                                              GNWC,
+                                                              GKXC,
+                                                              Empty_Tuple,
+                                                              GNWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_int8_instances<1,
+                                                                              GNWC,
+                                                                              GKXC,
+                                                                              Empty_Tuple,
+                                                                              GNWK,
+                                                                              ConvFwdDefault>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_int8_instances<1,
+                                                                              GNWC,
+                                                                              GKXC,
+                                                                              Empty_Tuple,
+                                                                              GNWK,
+                                                                              ConvFwd1x1P0>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_int8_instances<1,
+                                                                              GNWC,
+                                                                              GKXC,
+                                                                              Empty_Tuple,
+                                                                              GNWK,
+                                                                              ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000000..85ec0f55aa
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt
@@ -0,0 +1,8 @@
+add_instance_library(device_grouped_conv2d_bwd_data_instance
+   device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
+   device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
+   device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
+   device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
+   device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
+   device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
new file mode 100644
index 0000000000..226dca5083
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[g, n, hi, wi, c] * wei[g, k, y, x, c] = in[g, n, ho, wo, k]
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
+                                                                  GNHWK,
+                                                                  GKYXC,
+                                                                  Empty_Tuple,
+                                                                  GNHWC,
+                                                                  BF16,
+                                                                  BF16,
+                                                                  Empty_Tuple,
+                                                                  BF16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_bf16_instances<2,
+                                                        GNHWK,
+                                                        GKYXC,
+                                                        Empty_Tuple,
+                                                        GNHWC,
+                                                        ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_bf16_instances<2,
+                                                        GNHWK,
+                                                        GKYXC,
+                                                        Empty_Tuple,
+                                                        GNHWC,
+                                                        ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
new file mode 100644
index 0000000000..64fbf8bbf2
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[g, n, hi, wi, c] * wei[g, k, y, x, c] = in[g, n, ho, wo, k]
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
+                                                                  GNHWK,
+                                                                  GKYXC,
+                                                                  Empty_Tuple,
+                                                                  GNHWC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f16_instances<2,
+                                                       GNHWK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       GNHWC,
+                                                       ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f16_instances<2,
+                                                       GNHWK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       GNHWC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
new file mode 100644
index 0000000000..f9351d96f2
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[g, n, hi, wi, c] * wei[g, k, y, x, c] = in[g, n, ho, wo, k]
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
+                                                                  GNHWK,
+                                                                  GKYXC,
+                                                                  Empty_Tuple,
+                                                                  GNHWC,
+                                                                  F32,
+                                                                  F32,
+                                                                  Empty_Tuple,
+                                                                  F32,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f32_instances<2,
+                                                       GNHWK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       GNHWC,
+                                                       ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f32_instances<2,
+                                                       GNHWK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       GNHWC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
new file mode 100644
index 0000000000..5d9194798b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[n, hi, wi, g, c] * wei[g, k, y, x, c] = in[n, ho, wo, g, k]
+void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
+                                                                  NHWGK,
+                                                                  GKYXC,
+                                                                  Empty_Tuple,
+                                                                  NHWGC,
+                                                                  BF16,
+                                                                  BF16,
+                                                                  Empty_Tuple,
+                                                                  BF16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_bf16_instances<2,
+                                                        NHWGK,
+                                                        GKYXC,
+                                                        Empty_Tuple,
+                                                        NHWGC,
+                                                        ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_bf16_instances<2,
+                                                        NHWGK,
+                                                        GKYXC,
+                                                        Empty_Tuple,
+                                                        NHWGC,
+                                                        ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
new file mode 100644
index 0000000000..5269bb9652
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[n, hi, wi, g, c] * wei[g, k, y, x, c] = in[n, ho, wo, g, k]
+void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
+                                                                  NHWGK,
+                                                                  GKYXC,
+                                                                  Empty_Tuple,
+                                                                  NHWGC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f16_instances<2,
+                                                       NHWGK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       NHWGC,
+                                                       ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f16_instances<2,
+                                                       NHWGK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       NHWGC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
new file mode 100644
index 0000000000..adfa08c1f7
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[n, hi, wi, g, c] * wei[g, k, y, x, c] = in[n, ho, wo, g, k]
+void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
+                                                                  NHWGK,
+                                                                  GKYXC,
+                                                                  Empty_Tuple,
+                                                                  NHWGC,
+                                                                  F32,
+                                                                  F32,
+                                                                  Empty_Tuple,
+                                                                  F32,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f32_instances<2,
+                                                       NHWGK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       NHWGC,
+                                                       ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f32_instances<2,
+                                                       NHWGK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       NHWGC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt
new file mode 100644
index 0000000000..b7b9fc92d1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt
@@ -0,0 +1,9 @@
+add_instance_library(device_grouped_conv2d_bwd_weight_instance
+    device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
+    device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
+    device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
+    device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
+    device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
+    device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
+)
+
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
new file mode 100644
index 0000000000..cf39c86010
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k]
+void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           GNHWC,
+                                                           GKYXC,
+                                                           GNHWK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances<2,
+                                                                    GNHWC,
+                                                                    GKYXC,
+                                                                    GNHWK,
+                                                                    ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances<
+                                       2,
+                                       GNHWC,
+                                       GKYXC,
+                                       GNHWK,
+                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
new file mode 100644
index 0000000000..d52f0b4d1c
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k]
+void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           GNHWC,
+                                                           GKYXC,
+                                                           GNHWK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_instances<2,
+                                                                   GNHWC,
+                                                                   GKYXC,
+                                                                   GNHWK,
+                                                                   ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_instances<
+                                       2,
+                                       GNHWC,
+                                       GKYXC,
+                                       GNHWK,
+                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
new file mode 100644
index 0000000000..62547a5014
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k]
+void add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           GNHWC,
+                                                           GKYXC,
+                                                           GNHWK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_instances<2,
+                                                                   GNHWC,
+                                                                   GKYXC,
+                                                                   GNHWK,
+                                                                   ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_instances<
+                                       2,
+                                       GNHWC,
+                                       GKYXC,
+                                       GNHWK,
+                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
new file mode 100644
index 0000000000..1cb9991a86
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances<2,
+                                                                    NHWGC,
+                                                                    GKYXC,
+                                                                    NHWGK,
+                                                                    ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances<
+                                       2,
+                                       NHWGC,
+                                       GKYXC,
+                                       NHWGK,
+                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
new file mode 100644
index 0000000000..e64d55c3bb
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_instances<2,
+                                                                   NHWGC,
+                                                                   GKYXC,
+                                                                   NHWGK,
+                                                                   ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_instances<
+                                       2,
+                                       NHWGC,
+                                       GKYXC,
+                                       NHWGK,
+                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
new file mode 100644
index 0000000000..5fa4c9ba3c
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_instances<2,
+                                                                   NHWGC,
+                                                                   GKYXC,
+                                                                   NHWGK,
+                                                                   ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_instances<
+                                       2,
+                                       NHWGC,
+                                       GKYXC,
+                                       NHWGK,
+                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt
new file mode 100644
index 0000000000..708b37c42d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt
@@ -0,0 +1,18 @@
+add_instance_library(device_grouped_conv2d_fwd_instance
+   #xdl
+   # GNHWC, GKYXC, GNHWK
+   device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
+   device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
+   device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
+   # NHWGC, GKYXC, NHWGK
+   device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
+   device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
+   device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
+   #dl
+   # GNHWC, GKYXC, GNHWK
+   device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
+   device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
+   # NHWGC, GKYXC, NHWGK
+   device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
+   device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
new file mode 100644
index 0000000000..1925989838
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
@@ -0,0 +1,57 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f16_instances<GNHWC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              GNHWK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwdDefault>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f16_instances<GNHWC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              GNHWK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwd1x1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f16_instances<GNHWC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              GNHWK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
new file mode 100644
index 0000000000..95ef4bbe37
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
@@ -0,0 +1,57 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f32_instances<GNHWC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              GNHWK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwdDefault>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f32_instances<GNHWC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              GNHWK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwd1x1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f32_instances<GNHWC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              GNHWK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
new file mode 100644
index 0000000000..4a0f7b9b5b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
@@ -0,0 +1,57 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f16_instances<NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwdDefault>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f16_instances<NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwd1x1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f16_instances<NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
new file mode 100644
index 0000000000..94e57bef46
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
@@ -0,0 +1,57 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f32_instances<NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwdDefault>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f32_instances<NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwd1x1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f32_instances<NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
new file mode 100644
index 0000000000..db14ce4306
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k]
+void add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<2,
+                                                                              GNHWC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              GNHWK,
+                                                                              ConvFwdDefault>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<2,
+                                                                              GNHWC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              GNHWK,
+                                                                              ConvFwd1x1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<2,
+                                                                              GNHWC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              GNHWK,
+                                                                              ConvFwd1x1S1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<2,
+                                                                              GNHWC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              GNHWK,
+                                                                              ConvFwdOddC>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
new file mode 100644
index 0000000000..debdb05b4d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for in[g, n, hi ,wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k]
+void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<2,
+                                                                             GNHWC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             GNHWK,
+                                                                             ConvFwdDefault>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<2,
+                                                                             GNHWC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             GNHWK,
+                                                                             ConvFwd1x1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<2,
+                                                                             GNHWC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             GNHWK,
+                                                                             ConvFwd1x1S1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<2,
+                                                                             GNHWC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             GNHWK,
+                                                                             ConvFwdOddC>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
new file mode 100644
index 0000000000..20326f5be1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k]
+void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              GNHWC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              GNHWK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<2,
+                                                                             GNHWC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             GNHWK,
+                                                                             ConvFwdDefault>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<2,
+                                                                             GNHWC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             GNHWK,
+                                                                             ConvFwd1x1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<2,
+                                                                             GNHWC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             GNHWK,
+                                                                             ConvFwd1x1S1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<2,
+                                                                             GNHWC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             GNHWK,
+                                                                             ConvFwdOddC>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
new file mode 100644
index 0000000000..af68c3b07d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<2,
+                                                                              NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              ConvFwdDefault>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<2,
+                                                                              NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              ConvFwd1x1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<2,
+                                                                              NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              ConvFwd1x1S1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<2,
+                                                                              NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              ConvFwdOddC>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
new file mode 100644
index 0000000000..8b1506e0fb
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<2,
+                                                                             NHWGC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             NHWGK,
+                                                                             ConvFwdDefault>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<2,
+                                                                             NHWGC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             NHWGK,
+                                                                             ConvFwd1x1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<2,
+                                                                             NHWGC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             NHWGK,
+                                                                             ConvFwd1x1S1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<2,
+                                                                             NHWGC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             NHWGK,
+                                                                             ConvFwdOddC>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
new file mode 100644
index 0000000000..c8bdfb8c7c
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<2,
+                                                                             NHWGC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             NHWGK,
+                                                                             ConvFwdDefault>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<2,
+                                                                             NHWGC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             NHWGK,
+                                                                             ConvFwd1x1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<2,
+                                                                             NHWGC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             NHWGK,
+                                                                             ConvFwd1x1S1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<2,
+                                                                             NHWGC,
+                                                                             GKYXC,
+                                                                             Empty_Tuple,
+                                                                             NHWGK,
+                                                                             ConvFwdOddC>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000000..8383d622b2
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/CMakeLists.txt
@@ -0,0 +1,8 @@
+add_instance_library(device_grouped_conv3d_bwd_data_instance
+   device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
+   device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
+   device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
+   device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
+   device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
+   device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
new file mode 100644
index 0000000000..8331ea1fda
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
@@ -0,0 +1,50 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = in[g, n, do, ho,
+// wo, k]
+void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  GNDHWK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  GNDHWC,
+                                                                  BF16,
+                                                                  BF16,
+                                                                  Empty_Tuple,
+                                                                  BF16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_bf16_instances<3,
+                                                        GNDHWK,
+                                                        GKZYXC,
+                                                        Empty_Tuple,
+                                                        GNDHWC,
+                                                        ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_bf16_instances<3,
+                                                        GNDHWK,
+                                                        GKZYXC,
+                                                        Empty_Tuple,
+                                                        GNDHWC,
+                                                        ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
new file mode 100644
index 0000000000..1885d49c81
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
@@ -0,0 +1,50 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = in[g, n, do, ho,
+// wo, k]
+void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  GNDHWK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  GNDHWC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f16_instances<3,
+                                                       GNDHWK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       GNDHWC,
+                                                       ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f16_instances<3,
+                                                       GNDHWK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       GNDHWC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
new file mode 100644
index 0000000000..77135fcc05
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
@@ -0,0 +1,50 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = in[g, n, do, ho,
+// wo, k]
+void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  GNDHWK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  GNDHWC,
+                                                                  F32,
+                                                                  F32,
+                                                                  Empty_Tuple,
+                                                                  F32,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f32_instances<3,
+                                                       GNDHWK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       GNDHWC,
+                                                       ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f32_instances<3,
+                                                       GNDHWK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       GNDHWC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
new file mode 100644
index 0000000000..663d41fe0b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
@@ -0,0 +1,50 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[n, di, hi, wi, g, c] * wei[g, k, z, y, x, c] = in[n, do, ho, wo,
+// g, k]
+void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  NDHWGK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  NDHWGC,
+                                                                  BF16,
+                                                                  BF16,
+                                                                  Empty_Tuple,
+                                                                  BF16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_bf16_instances<3,
+                                                        NDHWGK,
+                                                        GKZYXC,
+                                                        Empty_Tuple,
+                                                        NDHWGC,
+                                                        ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_bf16_instances<3,
+                                                        NDHWGK,
+                                                        GKZYXC,
+                                                        Empty_Tuple,
+                                                        NDHWGC,
+                                                        ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
new file mode 100644
index 0000000000..ac0ab44ce3
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
@@ -0,0 +1,50 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[n, di, hi, wi, g, c] * wei[g, k, z, y, x, c] = in[n, do, ho, wo,
+// g, k]
+void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  NDHWGK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  NDHWGC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f16_instances<3,
+                                                       NDHWGK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       NDHWGC,
+                                                       ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f16_instances<3,
+                                                       NDHWGK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       NDHWGC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
new file mode 100644
index 0000000000..50d5cce73d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
@@ -0,0 +1,50 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[n, di, hi, wi, g, c] * wei[g, k, z, y, x, c] = in[n, do, ho, wo,
+// g, k]
+void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  NDHWGK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  NDHWGC,
+                                                                  F32,
+                                                                  F32,
+                                                                  Empty_Tuple,
+                                                                  F32,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f32_instances<3,
+                                                       NDHWGK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       NDHWGC,
+                                                       ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f32_instances<3,
+                                                       NDHWGK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       NDHWGC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt
new file mode 100644
index 0000000000..5118599b4f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt
@@ -0,0 +1,8 @@
+add_instance_library(device_grouped_conv3d_bwd_weight_instance
+    device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
+    device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
+    device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
+    device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
+    device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
+    device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
new file mode 100644
index 0000000000..c8f456db83
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           GNDHWC,
+                                                           GKZYXC,
+                                                           GNDHWK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances<3,
+                                                                    GNDHWC,
+                                                                    GKZYXC,
+                                                                    GNDHWK,
+                                                                    ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances<
+                                       3,
+                                       GNDHWC,
+                                       GKZYXC,
+                                       GNDHWK,
+                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
new file mode 100644
index 0000000000..099123ecbc
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           GNDHWC,
+                                                           GKZYXC,
+                                                           GNDHWK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_instances<3,
+                                                                   GNDHWC,
+                                                                   GKZYXC,
+                                                                   GNDHWK,
+                                                                   ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_instances<
+                                       3,
+                                       GNDHWC,
+                                       GKZYXC,
+                                       GNDHWK,
+                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
new file mode 100644
index 0000000000..0eda980b30
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           GNDHWC,
+                                                           GKZYXC,
+                                                           GNDHWK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_instances<3,
+                                                                   GNDHWC,
+                                                                   GKZYXC,
+                                                                   GNDHWK,
+                                                                   ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_instances<
+                                       3,
+                                       GNDHWC,
+                                       GKZYXC,
+                                       GNDHWK,
+                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
new file mode 100644
index 0000000000..1e5c1946fb
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           BF16,
+                                                           F32,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances<3,
+                                                                    NDHWGC,
+                                                                    GKZYXC,
+                                                                    NDHWGK,
+                                                                    ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances<
+                                       3,
+                                       NDHWGC,
+                                       GKZYXC,
+                                       NDHWGK,
+                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
new file mode 100644
index 0000000000..e0b442bf24
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_instances<3,
+                                                                   NDHWGC,
+                                                                   GKZYXC,
+                                                                   NDHWGK,
+                                                                   ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_instances<
+                                       3,
+                                       NDHWGC,
+                                       GKZYXC,
+                                       NDHWGK,
+                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
new file mode 100644
index 0000000000..4bb7948245
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           F32,
+                                                           F32,
+                                                           F32,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_instances<3,
+                                                                   NDHWGC,
+                                                                   GKZYXC,
+                                                                   NDHWGK,
+                                                                   ConvBwdWeightDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_instances<
+                                       3,
+                                       NDHWGC,
+                                       GKZYXC,
+                                       NDHWGK,
+                                       ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt
new file mode 100644
index 0000000000..cd209dbf9e
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt
@@ -0,0 +1,11 @@
+add_instance_library(device_grouped_conv3d_fwd_instance
+   device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
+   device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
+   device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
+   device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp
+
+   device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
+   device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
+   device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
+   device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
new file mode 100644
index 0000000000..bd8443f5ec
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<3,
+                                                                              GNDHWC,
+                                                                              GKZYXC,
+                                                                              Empty_Tuple,
+                                                                              GNDHWK,
+                                                                              ConvFwdDefault>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<3,
+                                                                              GNDHWC,
+                                                                              GKZYXC,
+                                                                              Empty_Tuple,
+                                                                              GNDHWK,
+                                                                              ConvFwd1x1P0>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<3,
+                                                                              GNDHWC,
+                                                                              GKZYXC,
+                                                                              Empty_Tuple,
+                                                                              GNDHWK,
+                                                                              ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
new file mode 100644
index 0000000000..fdd5c3169d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<3,
+                                                                             GNDHWC,
+                                                                             GKZYXC,
+                                                                             Empty_Tuple,
+                                                                             GNDHWK,
+                                                                             ConvFwdDefault>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<3,
+                                                                             GNDHWC,
+                                                                             GKZYXC,
+                                                                             Empty_Tuple,
+                                                                             GNDHWK,
+                                                                             ConvFwd1x1P0>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<3,
+                                                                             GNDHWC,
+                                                                             GKZYXC,
+                                                                             Empty_Tuple,
+                                                                             GNDHWK,
+                                                                             ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
new file mode 100644
index 0000000000..b486dd80b4
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<3,
+                                                                             GNDHWC,
+                                                                             GKZYXC,
+                                                                             Empty_Tuple,
+                                                                             GNDHWK,
+                                                                             ConvFwdDefault>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<3,
+                                                                             GNDHWC,
+                                                                             GKZYXC,
+                                                                             Empty_Tuple,
+                                                                             GNDHWK,
+                                                                             ConvFwd1x1P0>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<3,
+                                                                             GNDHWC,
+                                                                             GKZYXC,
+                                                                             Empty_Tuple,
+                                                                             GNDHWK,
+                                                                             ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp
new file mode 100644
index 0000000000..de3d866477
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              GNDHWC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              GNDHWK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_int8_instances<3,
+                                                                              GNDHWC,
+                                                                              GKZYXC,
+                                                                              Empty_Tuple,
+                                                                              GNDHWK,
+                                                                              ConvFwdDefault>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_int8_instances<3,
+                                                                              GNDHWC,
+                                                                              GKZYXC,
+                                                                              Empty_Tuple,
+                                                                              GNDHWK,
+                                                                              ConvFwd1x1P0>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_int8_instances<3,
+                                                                              GNDHWC,
+                                                                              GKZYXC,
+                                                                              Empty_Tuple,
+                                                                              GNDHWK,
+                                                                              ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
new file mode 100644
index 0000000000..b3c4c1d9c9
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<3,
+                                                                              NDHWGC,
+                                                                              GKZYXC,
+                                                                              Empty_Tuple,
+                                                                              NDHWGK,
+                                                                              ConvFwdDefault>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<3,
+                                                                              NDHWGC,
+                                                                              GKZYXC,
+                                                                              Empty_Tuple,
+                                                                              NDHWGK,
+                                                                              ConvFwd1x1P0>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_bf16_instances<3,
+                                                                              NDHWGC,
+                                                                              GKZYXC,
+                                                                              Empty_Tuple,
+                                                                              NDHWGK,
+                                                                              ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
new file mode 100644
index 0000000000..d252c7513f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<3,
+                                                                             NDHWGC,
+                                                                             GKZYXC,
+                                                                             Empty_Tuple,
+                                                                             NDHWGK,
+                                                                             ConvFwdDefault>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<3,
+                                                                             NDHWGC,
+                                                                             GKZYXC,
+                                                                             Empty_Tuple,
+                                                                             NDHWGK,
+                                                                             ConvFwd1x1P0>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f16_instances<3,
+                                                                             NDHWGC,
+                                                                             GKZYXC,
+                                                                             Empty_Tuple,
+                                                                             NDHWGK,
+                                                                             ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
new file mode 100644
index 0000000000..0d79c1c08a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
@@ -0,0 +1,52 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<3,
+                                                                             NDHWGC,
+                                                                             GKZYXC,
+                                                                             Empty_Tuple,
+                                                                             NDHWGK,
+                                                                             ConvFwdDefault>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<3,
+                                                                             NDHWGC,
+                                                                             GKZYXC,
+                                                                             Empty_Tuple,
+                                                                             NDHWGK,
+                                                                             ConvFwd1x1P0>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_f32_instances<3,
+                                                                             NDHWGC,
+                                                                             GKZYXC,
+                                                                             Empty_Tuple,
+                                                                             NDHWGK,
+                                                                             ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
new file mode 100644
index 0000000000..881b6df5c5
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              GKZYXC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_int8_instances<3,
+                                                                              NDHWGC,
+                                                                              GKZYXC,
+                                                                              Empty_Tuple,
+                                                                              NDHWGK,
+                                                                              ConvFwdDefault>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_int8_instances<3,
+                                                                              NDHWGC,
+                                                                              GKZYXC,
+                                                                              Empty_Tuple,
+                                                                              NDHWGK,
+                                                                              ConvFwd1x1P0>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv_fwd_xdl_int8_instances<3,
+                                                                              NDHWGC,
+                                                                              GKZYXC,
+                                                                              Empty_Tuple,
+                                                                              NDHWGK,
+                                                                              ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..f1553e3d5b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt
@@ -0,0 +1,12 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+add_instance_library(device_grouped_gemm_instance
+   device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
+   device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
+   device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
+   device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
+   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
+   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
+   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
+   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
+)
+endif()
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..aa161e51c9
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[k, m] * b[k, n] = e[m, n]
+using device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    16,    64,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 4>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              2,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instances{});
+    add_device_operation_instances(
+        instances, device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..c454deac1a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[k, m] * b[n, k] = e[m, n]
+using device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    16,    64,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 4>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              2,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances{});
+    add_device_operation_instances(
+        instances, device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..c829e8863d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[m, k] * b[k, n] = e[m, n]
+using device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances{});
+    add_device_operation_instances(
+        instances, device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..fb30e7a973
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,102 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[m, k] * b[n, k] = e[m, n]
+using device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,    
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   256,    32,   8,   8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances{});
+    add_device_operation_instances(
+        instances, device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..8642562fa3
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,80 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[m, k] * b[k, n] = e[m, n]
+using device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        // Currently AK1 must equal BK1 !
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 16,16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
new file mode 100644
index 0000000000..90223fd9bd
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
@@ -0,0 +1,108 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough                    = ck::tensor_operation::element_wise::PassThrough;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        // Currently AK1 must equal BK1 !
+     // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 16,16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+     // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+     // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+     // DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   192,    64,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 48, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 24, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   192,    32,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,    64,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v1>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v1>,
+
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   192,    64,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 48, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 24, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   192,    32,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,    64,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v2>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v2>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..aa6365cd98
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,75 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough                 = ck::tensor_operation::element_wise::PassThrough;
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[m, k] * b[n, k] = e[m, n]
+using device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
new file mode 100644
index 0000000000..f4460b360b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
@@ -0,0 +1,81 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough                    = ck::tensor_operation::element_wise::PassThrough;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //################################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //################################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //################################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   192,    64,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 48, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,    
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   192,    32,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   192,    32,   8,   8,   32,   32,    1,    3,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   256,    32,   8,   8,   32,   32,    1,    4,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,    64,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,    64,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/CMakeLists.txt
new file mode 100644
index 0000000000..ef8a440c1a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/CMakeLists.txt
@@ -0,0 +1,7 @@
+add_instance_library(device_grouped_gemm_bias_instance
+   device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp
+   device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp
+
+   device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp
+   device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..28be904e65
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using D0DataType = F32;
+using DsDataType = ck::Tuple<D0DataType>;
+
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //############################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //############################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //############################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //############################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 16,16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2, S< 1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2, S< 1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2, S< 1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2, S< 1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>
+        // clang-format on
+        >;
+
+void add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
+                                                         Row,
+                                                         DsLayout,
+                                                         Row,
+                                                         F16,
+                                                         F16,
+                                                         DsDataType,
+                                                         F16,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         Add>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..5f5b86a0e4
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using D0DataType = F32;
+using DsDataType = ck::Tuple<D0DataType>;
+
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //############################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //############################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //############################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //############################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,   128,   256,    64,   8,   8,   32,   32,    2,    4,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,   128,   128,    64,   8,   8,   32,   32,    2,    2,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,   128,    64,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,    64,   128,    64,   8,   8,   32,   32,    1,    2,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,   128,   128,    64,   8,   8,   32,   32,    4,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,   128,    64,    64,   8,   8,   32,   32,    2,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,    64,   128,    64,   8,   8,   32,   32,    2,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,   128,    32,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,    32,   128,    64,   8,   8,   32,   32,    1,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,    32,   256,    64,   8,   8,   32,   32,    1,    4,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,    64,    64,    64,    64,   8,   8,   32,   32,    2,    2,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,    64,    64,    32,    64,   8,   8,   32,   32,    2,    1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F16, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,    64,    32,    64,    64,   8,   8,   32,   32,    1,    2,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>
+        // clang-format on
+        >;
+
+void add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
+                                                         Col,
+                                                         DsLayout,
+                                                         Row,
+                                                         F16,
+                                                         F16,
+                                                         DsDataType,
+                                                         F16,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         Add>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..fa8441431a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using D0DataType = F32;
+using DsDataType = ck::Tuple<D0DataType>;
+
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //############################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //############################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //############################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //############################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 16,16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,    32,   128,    32,   8,   8,   32,   32,    1,    1, S< 1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2, S< 1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2, S< 1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2, S< 1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2, S< 1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2, S< 1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,    32,   256,    32,   8,   8,   32,   32,    1,    2, S< 1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,    16,   128,    32,   8,   8,   16,   16,    1,    4, S< 1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              4>
+        // clang-format on
+        >;
+
+void add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
+                                                         Row,
+                                                         DsLayout,
+                                                         Row,
+                                                         F16,
+                                                         F16,
+                                                         DsDataType,
+                                                         F32,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         Add>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..386a885642
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using D0DataType = F32;
+using DsDataType = ck::Tuple<D0DataType>;
+
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_irregular_tile_instances =
+    std::tuple<
+        // clang-format off
+        //############################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //############################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //############################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //############################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,   128,   256,    64,   8,   8,   32,   32,    2,    4,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,   128,   128,    64,   8,   8,   32,   32,    2,    2,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,   128,    64,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,    64,   128,    64,   8,   8,   32,   32,    1,    2,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   256,    32,   128,    64,   8,   8,   32,   32,    1,    1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,   128,   128,    64,   8,   8,   32,   32,    4,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,   128,    64,    64,   8,   8,   32,   32,    2,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,    64,   128,    64,   8,   8,   32,   32,    2,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,   128,    32,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,    32,   128,    64,   8,   8,   32,   32,    1,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,   128,    32,   256,    64,   8,   8,   32,   32,    1,    4,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,    64,    64,    64,    64,   8,   8,   32,   32,    2,    2,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,    64,    64,    32,    64,   8,   8,   32,   32,    2,    1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              4>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,   F16,   F16,     F32,      F32,  DsDataType,   F32, PassThrough, PassThrough,         Add, GemmMNKPadding,        1,    64,    32,    64,    64,   8,   8,   32,   32,    1,    2,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              4>
+        // clang-format on
+        >;
+
+void add_device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
+                                                         Col,
+                                                         DsLayout,
+                                                         Row,
+                                                         F16,
+                                                         F16,
+                                                         DsDataType,
+                                                         F32,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         Add>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/CMakeLists.txt
new file mode 100644
index 0000000000..a45bf36399
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/CMakeLists.txt
@@ -0,0 +1,8 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+add_instance_library(device_grouped_gemm_fastgelu_instance
+   device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
+   device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
+   device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp
+   device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp
+)
+endif()
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..f4086b6eac
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp
@@ -0,0 +1,104 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[k, m] * b[k, n] = e[m, n]
+using device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   2,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>
+    // clang-format on
+    >;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,    16,    64,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 4>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              2,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,   128,    64,    32,   2,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,    64,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,    64,   128,    32,   2,   2,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  FastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..d68eb76144
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp
@@ -0,0 +1,104 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[k, m] * b[n, k] = e[m, n]
+using device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   2,   8,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,    64,    16,    16,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,    16,    64,    32,   8,   8,   16,   16,    1,    1,     S<4, 16, 4>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              2,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,         1,           1,           1,               S<1, 16, 1, 4>,              1>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,   128,    64,    32,   2,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,    64,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,    64,   128,    32,   2,   8,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Col,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Col,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  FastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..2dfb8caace
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
@@ -0,0 +1,103 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[m, k] * b[k, n] = e[m, n]
+using device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   2,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Row, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Row,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  FastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..598a0b0e29
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
@@ -0,0 +1,104 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// a[m, k] * b[n, k] = e[m, n]
+using device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu,    GemmDefault,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,               8>
+    // clang-format on
+    >;
+
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //###################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //###################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //###################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //###################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,   128,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,    
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,   128,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,    32,   128,    32,   8,   8,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,   128,    32,   256,    32,   8,   8,   32,   32,    1,    4,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl<    Row,    Col, Empty_Tuple,    Row,   F16,   F16,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough,    FastGelu, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
+                                                  Col,
+                                                  Empty_Tuple,
+                                                  Row,
+                                                  F16,
+                                                  F16,
+                                                  Empty_Tuple,
+                                                  F16,
+                                                  PassThrough,
+                                                  PassThrough,
+                                                  FastGelu>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instances{});
+    add_device_operation_instances(
+        instances,
+        device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/image_to_column/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/image_to_column/CMakeLists.txt
new file mode 100644
index 0000000000..de10369374
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/image_to_column/CMakeLists.txt
@@ -0,0 +1,5 @@
+add_instance_library(device_image_to_column_instance
+   device_image_to_column_nhwc_1d_instance.cpp
+   device_image_to_column_nhwc_2d_instance.cpp
+   device_image_to_column_nhwc_3d_instance.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_1d_instance.cpp b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_1d_instance.cpp
new file mode 100644
index 0000000000..c8463623c3
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_1d_instance.cpp
@@ -0,0 +1,39 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/image_to_column/device_image_to_column_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_image_to_column_nhwc_1d_bf16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, BF16, BF16>>>& instances)
+{
+    add_device_operation_instances(instances, device_image_to_column_bf16_instances<1, GNWC>{});
+}
+
+void add_device_image_to_column_nhwc_1d_f16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, F16, F16>>>& instances)
+{
+    add_device_operation_instances(instances, device_image_to_column_f16_instances<1, GNWC>{});
+}
+
+void add_device_image_to_column_nhwc_1d_f32_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, F32, F32>>>& instances)
+{
+    add_device_operation_instances(instances, device_image_to_column_f32_instances<1, GNWC>{});
+}
+
+void add_device_image_to_column_nhwc_1d_i8_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, int8_t, int8_t>>>& instances)
+{
+    add_device_operation_instances(instances, device_image_to_column_i8_instances<1, GNWC>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_2d_instance.cpp b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_2d_instance.cpp
new file mode 100644
index 0000000000..652c7fac2a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_2d_instance.cpp
@@ -0,0 +1,39 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/image_to_column/device_image_to_column_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_image_to_column_nhwc_2d_bf16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, BF16, BF16>>>& instances)
+{
+    add_device_operation_instances(instances, device_image_to_column_bf16_instances<2, GNHWC>{});
+}
+
+void add_device_image_to_column_nhwc_2d_f16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, F16, F16>>>& instances)
+{
+    add_device_operation_instances(instances, device_image_to_column_f16_instances<2, GNHWC>{});
+}
+
+void add_device_image_to_column_nhwc_2d_f32_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, F32, F32>>>& instances)
+{
+    add_device_operation_instances(instances, device_image_to_column_f32_instances<2, GNHWC>{});
+}
+
+void add_device_image_to_column_nhwc_2d_i8_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, int8_t, int8_t>>>& instances)
+{
+    add_device_operation_instances(instances, device_image_to_column_i8_instances<2, GNHWC>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_3d_instance.cpp b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_3d_instance.cpp
new file mode 100644
index 0000000000..07774504d7
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_3d_instance.cpp
@@ -0,0 +1,39 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/image_to_column/device_image_to_column_instance.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_image_to_column_nhwc_3d_bf16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, BF16, BF16>>>& instances)
+{
+    add_device_operation_instances(instances, device_image_to_column_bf16_instances<3, GNDHWC>{});
+}
+
+void add_device_image_to_column_nhwc_3d_f16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, F16, F16>>>& instances)
+{
+    add_device_operation_instances(instances, device_image_to_column_f16_instances<3, GNDHWC>{});
+}
+
+void add_device_image_to_column_nhwc_3d_f32_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, F32, F32>>>& instances)
+{
+    add_device_operation_instances(instances, device_image_to_column_f32_instances<3, GNDHWC>{});
+}
+
+void add_device_image_to_column_nhwc_3d_i8_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, int8_t, int8_t>>>& instances)
+{
+    add_device_operation_instances(instances, device_image_to_column_i8_instances<3, GNDHWC>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/max_pool_bwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/max_pool_bwd/CMakeLists.txt
new file mode 100644
index 0000000000..820a63480e
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/max_pool_bwd/CMakeLists.txt
@@ -0,0 +1,11 @@
+set(DEVICE_MAXPOOL_BWD_INSTANCES)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_MAXPOOL_BWD_INSTANCES device_max_pool_bwd_f16_instance.cpp)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_MAXPOOL_BWD_INSTANCES device_max_pool_bwd_bf16_instance.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_MAXPOOL_BWD_INSTANCES device_max_pool_bwd_f32_instance.cpp)
+endif()
+add_instance_library(device_max_pool_bwd_instance ${DEVICE_MAXPOOL_BWD_INSTANCES})
diff --git a/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_bf16_instance.cpp
new file mode 100644
index 0000000000..40628d58b2
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_bf16_instance.cpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "max_pool_bwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_maxpool_bwd_bf16_instances(
+    std::vector<std::unique_ptr<DeviceMaxPoolBwd<BF16, I32, BF16>>>& instances)
+{
+    add_device_operation_instances(instances, device_maxpool_bwd_instances<BF16, I32, BF16>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f16_instance.cpp
new file mode 100644
index 0000000000..6c5cb27a7a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f16_instance.cpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "max_pool_bwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_maxpool_bwd_f16_instances(
+    std::vector<std::unique_ptr<DeviceMaxPoolBwd<F16, I32, F16>>>& instances)
+{
+    add_device_operation_instances(instances, device_maxpool_bwd_instances<F16, I32, F16>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f32_instance.cpp
new file mode 100644
index 0000000000..0a8d5a7976
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f32_instance.cpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "max_pool_bwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_maxpool_bwd_f32_instances(
+    std::vector<std::unique_ptr<DeviceMaxPoolBwd<F32, I32, F32>>>& instances)
+{
+    add_device_operation_instances(instances, device_maxpool_bwd_instances<F32, I32, F32>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/max_pool_bwd/max_pool_bwd_instance_common.hpp b/library/src/tensor_operation_instance/gpu/max_pool_bwd/max_pool_bwd_instance_common.hpp
new file mode 100644
index 0000000000..0bba106ee2
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/max_pool_bwd/max_pool_bwd_instance_common.hpp
@@ -0,0 +1,35 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using I32  = int32_t;
+using F16  = ck::half_t;
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+template <typename DOutDataType, typename IndexDataType, typename DInDataType>
+using device_maxpool_bwd_instances =
+    // clang-format off
+    std::tuple <
+        DeviceMaxPoolBwdImpl<DOutDataType, IndexDataType, DInDataType, 1>,
+        DeviceMaxPoolBwdImpl<DOutDataType, IndexDataType, DInDataType, 2>,
+        DeviceMaxPoolBwdImpl<DOutDataType, IndexDataType, DInDataType, 4>
+               // clang-format on
+               >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt
new file mode 100644
index 0000000000..4892a9c183
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt
@@ -0,0 +1,15 @@
+set(DEVICE_NORMALIZATION_INSTANCES)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_NORMALIZATION_INSTANCES device_layernorm2d_f16_instance.cpp
+    device_layernorm4d_f16_instance.cpp
+    device_groupnorm_f16_instance.cpp
+    device_groupnorm_swish_f16_instance.cpp
+    device_groupnorm_swish_f16_f32_f32_f16_instance.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_NORMALIZATION_INSTANCES device_layernorm2d_f32_instance.cpp
+    device_layernorm4d_f32_instance.cpp
+    device_groupnorm_f32_instance.cpp
+    device_groupnorm_swish_f32_instance.cpp)
+endif()
+add_instance_library(device_normalization_instance ${DEVICE_NORMALIZATION_INSTANCES})
diff --git a/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f16_instance.cpp
new file mode 100644
index 0000000000..762da1c6ae
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f16_instance.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+void add_device_normalization_rank_5_3_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 5, 3>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_normalization_f16_generic_instance<Pass, 5, 3>{});
+    add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 5, 3>{});
+    add_device_operation_instances(instances,
+                                   device_normalization_splitk_f16_instances<Pass, 5, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f32_instance.cpp
new file mode 100644
index 0000000000..44b553bd16
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f32_instance.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+void add_device_normalization_rank_5_3_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Pass, 5, 3>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_normalization_f32_generic_instance<Pass, 5, 3>{});
+    add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 5, 3>{});
+    add_device_operation_instances(instances,
+                                   device_normalization_splitk_f32_instances<Pass, 5, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_f32_f32_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_f32_f32_f16_instance.cpp
new file mode 100644
index 0000000000..aa662b7dfe
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_f32_f32_f16_instance.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Swish = ck::tensor_operation::element_wise::Swish;
+
+void add_device_normalization_rank_5_3_swish_f16_f32_f32_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F32, F32, F32, F16, Swish, 5, 3>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_normalization_f16_f32_f32_f16_generic_instance<Swish, 5, 3>{});
+    add_device_operation_instances(instances,
+                                   device_normalization_f16_f32_f32_f16_instances<Swish, 5, 3>{});
+    add_device_operation_instances(
+        instances, device_normalization_splitk_f16_f32_f32_f16_instances<Swish, 5, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_instance.cpp
new file mode 100644
index 0000000000..bc5cd801ae
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_instance.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Swish = ck::tensor_operation::element_wise::Swish;
+
+void add_device_normalization_rank_5_3_swish_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Swish, 5, 3>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_normalization_f16_generic_instance<Swish, 5, 3>{});
+    add_device_operation_instances(instances, device_normalization_f16_instances<Swish, 5, 3>{});
+    add_device_operation_instances(instances,
+                                   device_normalization_splitk_f16_instances<Swish, 5, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f32_instance.cpp
new file mode 100644
index 0000000000..4b2ab33570
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f32_instance.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Swish = ck::tensor_operation::element_wise::Swish;
+
+void add_device_normalization_rank_5_3_swish_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Swish, 5, 3>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_normalization_f32_generic_instance<Swish, 5, 3>{});
+    add_device_operation_instances(instances, device_normalization_f32_instances<Swish, 5, 3>{});
+    add_device_operation_instances(instances,
+                                   device_normalization_splitk_f32_instances<Swish, 5, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f16_instance.cpp
new file mode 100644
index 0000000000..0d235f1fa7
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f16_instance.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+void add_device_normalization_rank_2_1_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 2, 1>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_normalization_f16_generic_instance<Pass, 2, 1>{});
+    add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 2, 1>{});
+    add_device_operation_instances(instances,
+                                   device_normalization_splitk_f16_instances<Pass, 2, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f32_instance.cpp
new file mode 100644
index 0000000000..00039531e1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f32_instance.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+void add_device_normalization_rank_2_1_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Pass, 2, 1>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_normalization_f32_generic_instance<Pass, 2, 1>{});
+    add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 2, 1>{});
+    add_device_operation_instances(instances,
+                                   device_normalization_splitk_f32_instances<Pass, 2, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f16_instance.cpp
new file mode 100644
index 0000000000..6bc3950062
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f16_instance.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+void add_device_normalization_rank_4_3_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_normalization_f16_generic_instance<Pass, 4, 3>{});
+    add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 4, 3>{});
+    add_device_operation_instances(instances,
+                                   device_normalization_splitk_f16_instances<Pass, 4, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f32_instance.cpp
new file mode 100644
index 0000000000..b387dc2f3f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f32_instance.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+void add_device_normalization_rank_4_3_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Pass, 4, 3>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_normalization_f32_generic_instance<Pass, 4, 3>{});
+    add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 4, 3>{});
+    add_device_operation_instances(instances,
+                                   device_normalization_splitk_f32_instances<Pass, 4, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/normalization/normalization_instance_common.hpp b/library/src/tensor_operation_instance/gpu/normalization/normalization_instance_common.hpp
new file mode 100644
index 0000000000..7aa3da8eed
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/normalization/normalization_instance_common.hpp
@@ -0,0 +1,201 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_f16_instances =
+    // clang-format off
+    std::tuple <
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8>
+        // clang-format on
+        >;
+
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_splitk_f16_instances =
+    // clang-format off
+    std::tuple <
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8>
+        // clang-format on
+        >;
+
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_f16_generic_instance = std::tuple<
+    // clang-format off
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1>
+    // clang-format on
+    >;
+
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_f32_instances = std::tuple<
+    // clang-format off
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
+    // clang-format on
+    >;
+
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_splitk_f32_instances = std::tuple<
+    // clang-format off
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
+    // clang-format on
+    >;
+
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_f32_generic_instance = std::tuple<
+    // clang-format off
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1>
+    // clang-format on
+    >;
+
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_f16_f32_f32_f16_instances = std::tuple<
+    // clang-format off
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
+    // clang-format on
+    >;
+
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_splitk_f16_f32_f32_f16_instances = std::tuple<
+    // clang-format off
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
+    // clang-format on
+    >;
+
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_f16_f32_f32_f16_generic_instance = std::tuple<
+    // clang-format off
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt
new file mode 100644
index 0000000000..63bbe75465
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt
@@ -0,0 +1,14 @@
+set(DEVICE_POOL3D_FWD_INSTANCES)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
+                                          device_max_pool3d_fwd_ndhwc_f16_instance.cpp)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp
+                                          device_max_pool3d_fwd_ndhwc_bf16_instance.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
+                                          device_max_pool3d_fwd_ndhwc_f32_instance.cpp)
+endif()
+add_instance_library(device_pool3d_fwd_instance ${DEVICE_POOL3D_FWD_INSTANCES})
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp
new file mode 100644
index 0000000000..f102516992
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "pool_fwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+
+void add_device_pool3d_fwd_ndhwc_bf16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, BF16, BF16, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool3d_fwd_ndhwc_instances<BF16, BF16, I32, F32, ReduceOpId, false>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
new file mode 100644
index 0000000000..4ebd50bae6
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "pool_fwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+
+void add_device_pool3d_fwd_ndhwc_f16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F32, ReduceOpId, false>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
new file mode 100644
index 0000000000..dcb19110b2
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "pool_fwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+
+void add_device_pool3d_fwd_ndhwc_f32_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_bf16_instance.cpp
new file mode 100644
index 0000000000..5dc504e178
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_bf16_instance.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "pool_fwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+
+void add_device_pool3d_fwd_ndhwc_bf16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, BF16, BF16, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool3d_fwd_ndhwc_instances<BF16, BF16, I32, BF16, ReduceOpId, false>{});
+}
+
+void add_device_pool3d_fwd_ndhwc_index_bf16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, BF16, BF16, I32, NDHWC, NDHWC, ReduceOpId, true>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool3d_fwd_ndhwc_instances<BF16, BF16, I32, BF16, ReduceOpId, true>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp
new file mode 100644
index 0000000000..46b16bd005
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "pool_fwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+
+void add_device_pool3d_fwd_ndhwc_f16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F16, ReduceOpId, false>{});
+}
+
+void add_device_pool3d_fwd_ndhwc_index_f16_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, true>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F16, ReduceOpId, true>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp
new file mode 100644
index 0000000000..b4b0e74d28
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "pool_fwd_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+
+void add_device_pool3d_fwd_ndhwc_f32_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
+}
+
+void add_device_pool3d_fwd_ndhwc_index_f32_instances(
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, true>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, true>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/pool_fwd_instance_common.hpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/pool_fwd_instance_common.hpp
new file mode 100644
index 0000000000..e8e7813295
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/pool3d_fwd/pool_fwd_instance_common.hpp
@@ -0,0 +1,42 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using I32   = int32_t;
+using F16   = ck::half_t;
+using BF16  = ck::bhalf_t;
+using F32   = float;
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+template <typename InDataType,
+          typename OutDataType,
+          typename IndexDataType,
+          typename ComputeDataType,
+          ReduceTensorOp ReduceOpId,
+          bool OutputIndex>
+using device_pool3d_fwd_ndhwc_instances =
+    // clang-format off
+    std::tuple <
+        DevicePool3dFwd_NDHWC_NDHWC<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1>,
+        DevicePool3dFwd_NDHWC_NDHWC<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2>,
+        DevicePool3dFwd_NDHWC_NDHWC<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4>
+               // clang-format on
+               >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt
new file mode 100644
index 0000000000..00eb6ff1c1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt
@@ -0,0 +1,32 @@
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+
+set(CONV2D_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp)
+set(CONV2D_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp)
+set(CONV2D_BIAS_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp)
+set(CONV2D_BIAS_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp)
+set(GEMM_QUANT_SRC
+    gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
+    gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
+    gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
+    gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
+)
+if(DL_KERNELS)
+    list(APPEND CONV2D_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp)
+    list(APPEND CONV2D_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp)
+    list(APPEND CONV2D_BIAS_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp)
+    list(APPEND CONV2D_BIAS_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp)
+    list(APPEND GEMM_QUANT_SRC
+        gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
+        gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
+        gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
+        gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp)
+endif()
+
+add_instance_library(device_quantization_instance
+    ${CONV2D_PERLAYER_QUANT_SRC}
+    ${CONV2D_PERCHANNEL_QUANT_SRC}
+    ${CONV2D_BIAS_PERLAYER_QUANT_SRC}
+    ${CONV2D_BIAS_PERCHANNEL_QUANT_SRC}
+    ${GEMM_QUANT_SRC}
+)
+endif()
\ No newline at end of file
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/conv2d_quantization_common.hpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/conv2d_quantization_common.hpp
new file mode 100644
index 0000000000..711314985a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/conv2d_quantization_common.hpp
@@ -0,0 +1,68 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Empty_Tuple = ck::Tuple<>;
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using NHWGC       = ck::tensor_layout::convolution::NHWGC;
+using GKYXC       = ck::tensor_layout::convolution::GKYXC;
+using NHWGK       = ck::tensor_layout::convolution::NHWGK;
+using GK          = ck::tensor_layout::convolution::G_K;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Relu        = ck::tensor_operation::element_wise::Relu;
+using TanH        = ck::tensor_operation::element_wise::TanH;
+
+using GK_Tuple      = ck::Tuple<GK>;
+using GK_GK_Tuple   = ck::Tuple<GK, GK>;
+using I32_Tuple     = ck::Tuple<int32_t>;
+using F32_Tuple     = ck::Tuple<float>;
+using I32_F32_Tuple = ck::Tuple<int32_t, float>;
+
+// perlayer
+using Mul_Clamp      = ck::tensor_operation::element_wise::Activation_Mul_Clamp<PassThrough>;
+using Relu_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Relu>;
+
+// bias + perlayer
+using Add_Mul_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<PassThrough>;
+using Add_Relu_Mul_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<Relu>;
+using Add_Mul_TanH_Mul_Clamp =
+    ck::tensor_operation::element_wise::Add_Mul_Activation_Mul_Clamp<TanH>;
+
+// perchannel
+using Mul2_Clamp      = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<PassThrough>;
+using Relu_Mul2_Clamp = ck::tensor_operation::element_wise::Activation_Mul2_Clamp<Relu>;
+
+// bias + perchannel
+using Add_Mul2_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<PassThrough>;
+using Add_Relu_Mul2_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul2_Clamp<Relu>;
+using Add_Mul2_TanH_Mul_Clamp =
+    ck::tensor_operation::element_wise::Add_Mul2_Activation_Mul_Clamp<TanH>;
+
+static constexpr ck::index_t NDimSpatial = 2;
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+static constexpr auto ConvFwdDefault =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
+static constexpr auto ConvFwd1x1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0;
+static constexpr auto ConvFwd1x1S1P0 =
+    ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp
new file mode 100644
index 0000000000..39c4f82fef
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp
@@ -0,0 +1,145 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_conv2d_dl_int8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_conv2d_dl_bias_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Mul2_Clamp>>>& instances)
+{
+    // dl
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_F32_Tuple,
+                                                                           Add_Mul2_Clamp,
+                                                                           ConvFwdDefault,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_F32_Tuple,
+                                                                           Add_Mul2_Clamp,
+                                                                           ConvFwd1x1P0,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_F32_Tuple,
+                                                                           Add_Mul2_Clamp,
+                                                                           ConvFwd1x1S1P0,
+                                                                           4>{});
+}
+
+void add_device_conv2d_dl_bias_relu_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Relu_Mul2_Clamp>>>& instances)
+{
+    // dl
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_F32_Tuple,
+                                                                           Add_Relu_Mul2_Clamp,
+                                                                           ConvFwdDefault,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_F32_Tuple,
+                                                                           Add_Relu_Mul2_Clamp,
+                                                                           ConvFwd1x1P0,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_F32_Tuple,
+                                                                           Add_Relu_Mul2_Clamp,
+                                                                           ConvFwd1x1S1P0,
+                                                                           4>{});
+}
+
+void add_device_conv2d_dl_bias_tanh_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Mul2_TanH_Mul_Clamp>>>& instances)
+{
+    // dl
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_F32_Tuple,
+                                                                           Add_Mul2_TanH_Mul_Clamp,
+                                                                           ConvFwdDefault,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_F32_Tuple,
+                                                                           Add_Mul2_TanH_Mul_Clamp,
+                                                                           ConvFwd1x1P0,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_F32_Tuple,
+                                                                           Add_Mul2_TanH_Mul_Clamp,
+                                                                           ConvFwd1x1S1P0,
+                                                                           4>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp
new file mode 100644
index 0000000000..92e73eb2ee
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp
@@ -0,0 +1,146 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_conv2d_dl_int8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_conv2d_dl_bias_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_Tuple,
+                                                                           Add_Mul_Clamp,
+                                                                           ConvFwdDefault,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_Tuple,
+                                                                           Add_Mul_Clamp,
+                                                                           ConvFwd1x1P0,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_Tuple,
+                                                                           Add_Mul_Clamp,
+                                                                           ConvFwd1x1S1P0,
+                                                                           4>{});
+}
+
+void add_device_conv2d_dl_bias_relu_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Relu_Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_Tuple,
+                                                                           Add_Relu_Mul_Clamp,
+                                                                           ConvFwdDefault,
+                                                                           4>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_Tuple,
+                                                                           Add_Relu_Mul_Clamp,
+                                                                           ConvFwd1x1P0,
+                                                                           4>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_Tuple,
+                                                                           Add_Relu_Mul_Clamp,
+                                                                           ConvFwd1x1S1P0,
+                                                                           4>{});
+}
+
+void add_device_conv2d_dl_bias_tanh_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Mul_TanH_Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_Tuple,
+                                                                           Add_Mul_TanH_Mul_Clamp,
+                                                                           ConvFwdDefault,
+                                                                           4>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_Tuple,
+                                                                           Add_Mul_TanH_Mul_Clamp,
+                                                                           ConvFwd1x1P0,
+                                                                           4>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_Tuple,
+                                                                           NHWGK,
+                                                                           I32_Tuple,
+                                                                           Add_Mul_TanH_Mul_Clamp,
+                                                                           ConvFwd1x1S1P0,
+                                                                           4>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_int8_instance.hpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_int8_instance.hpp
new file mode 100644
index 0000000000..2ec37c8413
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_int8_instance.hpp
@@ -0,0 +1,36 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "conv2d_quantization_common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+template <typename InLayout,
+          typename WeiLayout,
+          typename DsLayout,
+          typename OutLayout,
+          typename DsDatatype,
+          typename OutElementOp,
+          ConvolutionForwardSpecialization ConvSpec,
+          index_t DstScalarPerVector>
+using device_grouped_conv2d_dl_int8_instances =
+    std::tuple<
+        // ###########################################|        NDim| InData| WeiData|   MultpleD|     OutData|     AccData| InLayout| WeiLayout|   MultipleD| OutLayout|           In|           Wei|           Out|    Convolution|              GEMM| Block|  MPer|  NPer| K0Per| K1|      M1Per|      N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        // ###########################################|     Spatial|   Type|    Type|       Type|        Type|        Type|         |          |      Layout|          |  Elementwise|   Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        // ###########################################|            |       |        |           |            |            |         |          |            |          |    Operation|     Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        // ###########################################|            |       |        |           |            |            |         |          |            |          |             |              |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK< NDimSpatial, int8_t,  int8_t, DsDatatype,      int8_t,     int32_t, InLayout, WeiLayout,    DsLayout, OutLayout,  PassThrough,   PassThrough,  OutElementOp,       ConvSpec,          GemmSpec,   256,   128,   128,    16,  4,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5, DstScalarPerVector>
+    >;
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp
new file mode 100644
index 0000000000..1d8b58fd18
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp
@@ -0,0 +1,98 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_conv2d_dl_int8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_conv2d_dl_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Mul2_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_Tuple,
+                                                                           NHWGK,
+                                                                           F32_Tuple,
+                                                                           Mul2_Clamp,
+                                                                           ConvFwdDefault,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_Tuple,
+                                                                           NHWGK,
+                                                                           F32_Tuple,
+                                                                           Mul2_Clamp,
+                                                                           ConvFwd1x1P0,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_Tuple,
+                                                                           NHWGK,
+                                                                           F32_Tuple,
+                                                                           Mul2_Clamp,
+                                                                           ConvFwd1x1S1P0,
+                                                                           4>{});
+}
+
+void add_device_conv2d_dl_relu_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Relu_Mul2_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_Tuple,
+                                                                           NHWGK,
+                                                                           F32_Tuple,
+                                                                           Relu_Mul2_Clamp,
+                                                                           ConvFwdDefault,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_Tuple,
+                                                                           NHWGK,
+                                                                           F32_Tuple,
+                                                                           Relu_Mul2_Clamp,
+                                                                           ConvFwd1x1P0,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           GK_Tuple,
+                                                                           NHWGK,
+                                                                           F32_Tuple,
+                                                                           Relu_Mul2_Clamp,
+                                                                           ConvFwd1x1S1P0,
+                                                                           4>{});
+}
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp
new file mode 100644
index 0000000000..62826d0d2b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp
@@ -0,0 +1,98 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_conv2d_dl_int8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_conv2d_dl_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           Empty_Tuple,
+                                                                           NHWGK,
+                                                                           Empty_Tuple,
+                                                                           Mul_Clamp,
+                                                                           ConvFwdDefault,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           Empty_Tuple,
+                                                                           NHWGK,
+                                                                           Empty_Tuple,
+                                                                           Mul_Clamp,
+                                                                           ConvFwd1x1P0,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           Empty_Tuple,
+                                                                           NHWGK,
+                                                                           Empty_Tuple,
+                                                                           Mul_Clamp,
+                                                                           ConvFwd1x1S1P0,
+                                                                           4>{});
+}
+
+void add_device_conv2d_dl_relu_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Relu_Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           Empty_Tuple,
+                                                                           NHWGK,
+                                                                           Empty_Tuple,
+                                                                           Relu_Mul_Clamp,
+                                                                           ConvFwdDefault,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           Empty_Tuple,
+                                                                           NHWGK,
+                                                                           Empty_Tuple,
+                                                                           Relu_Mul_Clamp,
+                                                                           ConvFwd1x1P0,
+                                                                           4>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_dl_int8_instances<NHWGC,
+                                                                           GKYXC,
+                                                                           Empty_Tuple,
+                                                                           NHWGK,
+                                                                           Empty_Tuple,
+                                                                           Relu_Mul_Clamp,
+                                                                           ConvFwd1x1S1P0,
+                                                                           4>{});
+}
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp
new file mode 100644
index 0000000000..99877c32db
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp
@@ -0,0 +1,142 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_conv2d_xdl_int8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_conv2d_xdl_bias_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Mul2_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_F32_Tuple,
+                                                                            Add_Mul2_Clamp,
+                                                                            ConvFwdDefault,
+                                                                            8>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_F32_Tuple,
+                                                                            Add_Mul2_Clamp,
+                                                                            ConvFwd1x1P0,
+                                                                            8>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_F32_Tuple,
+                                                                            Add_Mul2_Clamp,
+                                                                            ConvFwd1x1S1P0,
+                                                                            8>{});
+}
+
+void add_device_conv2d_xdl_bias_relu_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Relu_Mul2_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_F32_Tuple,
+                                                                            Add_Relu_Mul2_Clamp,
+                                                                            ConvFwdDefault,
+                                                                            8>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_F32_Tuple,
+                                                                            Add_Relu_Mul2_Clamp,
+                                                                            ConvFwd1x1P0,
+                                                                            8>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_F32_Tuple,
+                                                                            Add_Relu_Mul2_Clamp,
+                                                                            ConvFwd1x1S1P0,
+                                                                            8>{});
+}
+
+void add_device_conv2d_xdl_bias_tanh_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Mul2_TanH_Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_F32_Tuple,
+                                                                            Add_Mul2_TanH_Mul_Clamp,
+                                                                            ConvFwdDefault,
+                                                                            8>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_F32_Tuple,
+                                                                            Add_Mul2_TanH_Mul_Clamp,
+                                                                            ConvFwd1x1P0,
+                                                                            8>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_F32_Tuple,
+                                                                            Add_Mul2_TanH_Mul_Clamp,
+                                                                            ConvFwd1x1S1P0,
+                                                                            8>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp
new file mode 100644
index 0000000000..50ccc69f45
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp
@@ -0,0 +1,146 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_conv2d_xdl_int8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_conv2d_xdl_bias_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_Tuple,
+                                                                            Add_Mul_Clamp,
+                                                                            ConvFwdDefault,
+                                                                            8>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_Tuple,
+                                                                            Add_Mul_Clamp,
+                                                                            ConvFwd1x1P0,
+                                                                            8>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_Tuple,
+                                                                            Add_Mul_Clamp,
+                                                                            ConvFwd1x1S1P0,
+                                                                            8>{});
+}
+
+void add_device_conv2d_xdl_bias_relu_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Relu_Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_Tuple,
+                                                                            Add_Relu_Mul_Clamp,
+                                                                            ConvFwdDefault,
+                                                                            8>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_Tuple,
+                                                                            Add_Relu_Mul_Clamp,
+                                                                            ConvFwd1x1P0,
+                                                                            8>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_Tuple,
+                                                                            Add_Relu_Mul_Clamp,
+                                                                            ConvFwd1x1S1P0,
+                                                                            8>{});
+}
+
+void add_device_conv2d_xdl_bias_tanh_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              I32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Add_Mul_TanH_Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_Tuple,
+                                                                            Add_Mul_TanH_Mul_Clamp,
+                                                                            ConvFwdDefault,
+                                                                            8>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_Tuple,
+                                                                            Add_Mul_TanH_Mul_Clamp,
+                                                                            ConvFwd1x1P0,
+                                                                            8>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_Tuple,
+                                                                            NHWGK,
+                                                                            I32_Tuple,
+                                                                            Add_Mul_TanH_Mul_Clamp,
+                                                                            ConvFwd1x1S1P0,
+                                                                            8>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_int8_instance.hpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_int8_instance.hpp
new file mode 100644
index 0000000000..caced6c950
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_int8_instance.hpp
@@ -0,0 +1,48 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "conv2d_quantization_common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename DsDatatype,
+          typename OutElementOp,
+          ConvolutionForwardSpecialization ConvSpec,
+          index_t DstScalarPerVector>
+using device_grouped_conv2d_xdl_int8_instances =
+    std::tuple <
+        //########################################|     NumDim|       A|       B|       Ds|       E|  AData|  BData| AccData| CShuffle|         Ds|  EData|           A|           B|          CDE|    ConvForward|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|     CBlockTransfer|
+        //########################################|    Spatial|  Layout|  Layout|   Layout|  Layout|   Type|   Type|    Type| DataType|   DataType|   Type| Elementwise| Elementwise|  Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl|    ScalarPerVector|
+        //########################################|           |        |        |         |        |       |       |        |         |           |       |   Operation|   Operation|    Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|      _NWaveNPerXdl|
+        //########################################|           |        |        |         |        |       |       |        |         |           |       |            |            |             |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                   |
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>, DstScalarPerVector>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   256,   128,   256,    64,  16,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>, DstScalarPerVector>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   128,   128,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>, DstScalarPerVector>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   256,   128,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>, DstScalarPerVector>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   128,   128,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 2>, DstScalarPerVector>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   128,    64,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>, DstScalarPerVector>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,    64,    64,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 2>, DstScalarPerVector>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   256,   128,    64,    64,  16,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>, DstScalarPerVector>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   256,    64,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>, DstScalarPerVector>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   128,   128,    32,    64,  16,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 2>, DstScalarPerVector>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,   128,    32,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>, DstScalarPerVector>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,    64,    64,    32,    64,  16,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 2>, DstScalarPerVector>,
+        DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, int8_t, int8_t, int32_t,  int32_t, DsDatatype, int8_t, PassThrough, PassThrough, OutElementOp,       ConvSpec,       GemmSpec,        1,    64,    32,    64,    64,  16,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 2>, DstScalarPerVector>
+    >;
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp
new file mode 100644
index 0000000000..526fe73463
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp
@@ -0,0 +1,98 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_conv2d_xdl_int8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_conv2d_xdl_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Mul2_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_Tuple,
+                                                                            NHWGK,
+                                                                            F32_Tuple,
+                                                                            Mul2_Clamp,
+                                                                            ConvFwdDefault,
+                                                                            8>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_Tuple,
+                                                                            NHWGK,
+                                                                            F32_Tuple,
+                                                                            Mul2_Clamp,
+                                                                            ConvFwd1x1P0,
+                                                                            8>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_Tuple,
+                                                                            NHWGK,
+                                                                            F32_Tuple,
+                                                                            Mul2_Clamp,
+                                                                            ConvFwd1x1S1P0,
+                                                                            8>{});
+}
+
+void add_device_conv2d_xdl_relu_perchannel_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              GK_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              F32_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Relu_Mul2_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_Tuple,
+                                                                            NHWGK,
+                                                                            F32_Tuple,
+                                                                            Relu_Mul2_Clamp,
+                                                                            ConvFwdDefault,
+                                                                            8>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_Tuple,
+                                                                            NHWGK,
+                                                                            F32_Tuple,
+                                                                            Relu_Mul2_Clamp,
+                                                                            ConvFwd1x1P0,
+                                                                            8>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            GK_Tuple,
+                                                                            NHWGK,
+                                                                            F32_Tuple,
+                                                                            Relu_Mul2_Clamp,
+                                                                            ConvFwd1x1S1P0,
+                                                                            8>{});
+}
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp
new file mode 100644
index 0000000000..d1d5a66bb9
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp
@@ -0,0 +1,98 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_conv2d_xdl_int8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_conv2d_xdl_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            Empty_Tuple,
+                                                                            NHWGK,
+                                                                            Empty_Tuple,
+                                                                            Mul_Clamp,
+                                                                            ConvFwdDefault,
+                                                                            16>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            Empty_Tuple,
+                                                                            NHWGK,
+                                                                            Empty_Tuple,
+                                                                            Mul_Clamp,
+                                                                            ConvFwd1x1P0,
+                                                                            16>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            Empty_Tuple,
+                                                                            NHWGK,
+                                                                            Empty_Tuple,
+                                                                            Mul_Clamp,
+                                                                            ConvFwd1x1S1P0,
+                                                                            16>{});
+}
+
+void add_device_conv2d_xdl_relu_perlayer_quantization_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              Relu_Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            Empty_Tuple,
+                                                                            NHWGK,
+                                                                            Empty_Tuple,
+                                                                            Relu_Mul_Clamp,
+                                                                            ConvFwdDefault,
+                                                                            16>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            Empty_Tuple,
+                                                                            NHWGK,
+                                                                            Empty_Tuple,
+                                                                            Relu_Mul_Clamp,
+                                                                            ConvFwd1x1P0,
+                                                                            16>{});
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_xdl_int8_instances<NHWGC,
+                                                                            GKYXC,
+                                                                            Empty_Tuple,
+                                                                            NHWGK,
+                                                                            Empty_Tuple,
+                                                                            Relu_Mul_Clamp,
+                                                                            ConvFwd1x1S1P0,
+                                                                            16>{});
+}
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp
new file mode 100644
index 0000000000..0ec498a441
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_quantization_common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <typename OutElementOp>
+using device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //####################|      A|      B|          Ds|      E|  AData|  BData| AccData|      DsData|   EData|           A|           B|           CDE|           GEMM| Block|  MPer|  NPer|  KPer| K1|   M1Per|  N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //####################| Layout| Layout|      Layout| Layout|   Type|   Type|    Type|        Type|    Type| Elementwise| Elementwise|   Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM| Thread| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //####################|       |       |            |       |       |       |        |            |        |   Operation|   Operation|     Operation|               |      |      |      |      |   |        |       |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //####################|       |       |            |       |       |       |        |            |        |            |            |              |               |      |      |      |      |   |        |       |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGemmMultipleD_Dl<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,   256,   128,   128,    16,  4,       4,      4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 4>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 4>,      S<2, 1, 4, 4>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,               5,                  4>
+    // clang-format on
+    >;
+
+template <typename OutElementOp>
+using device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //####################|      A|      B|          Ds|      E|  AData|  BData| AccData|      DsData|   EData|           A|           B|           CDE|           GEMM| Block|  MPer|  NPer|  KPer| K1|   M1Per|  N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //####################| Layout| Layout|      Layout| Layout|   Type|   Type|    Type|        Type|    Type| Elementwise| Elementwise|   Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM| Thread| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //####################|       |       |            |       |       |       |        |            |        |   Operation|   Operation|     Operation|               |      |      |      |      |   |        |       |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //####################|       |       |            |       |       |       |        |            |        |            |            |              |               |      |      |      |      |   |        |       |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGemmMultipleD_Dl<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,   256,   128,   128,    16,  4,      4,       4,      1,       S<8, 2>,       S<8, 2>,      S<2, 1, 4, 4>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5,                  4>
+    // clang-format on
+    >;
+
+template <typename OutElementOp>
+using device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //####################|      A|      B|          Ds|      E|  AData|  BData| AccData|      DsData|   EData|           A|           B|           CDE|           GEMM| Block|  MPer|  NPer|  KPer| K1|   M1Per|  N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //####################| Layout| Layout|      Layout| Layout|   Type|   Type|    Type|        Type|    Type| Elementwise| Elementwise|   Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM| Thread| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //####################|       |       |            |       |       |       |        |            |        |   Operation|   Operation|     Operation|               |      |      |      |      |   |        |       |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //####################|       |       |            |       |       |       |        |            |        |            |            |              |               |      |      |      |      |   |        |       |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGemmMultipleD_Dl<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,   256,   128,   128,    16,  4,       4,      4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>,      S<2, 1, 4, 4>,      S<8, 1,  32, 1>,  S<0, 3, 1, 2>,  S<0, 3, 1, 2>,       S<1, 1, 4, 1>,      S<0, 3, 1, 2>,       S<1, 1, 4, 4>, S<0, 1, 2, 3, 4, 5>,               5,                  4>
+    // clang-format on
+    >;
+
+template <typename OutElementOp>
+using device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //####################|      A|      B|          Ds|      E|  AData|  BData| AccData|      DsData|   EData|           A|           B|           CDE|           GEMM| Block|  MPer|  NPer|  KPer| K1|   M1Per|  N1Per|   KPer|  M11N11Thread|  M11N11Thread|     ABlockTransfer|       ABlockTransfer| ABlockTransfer| ABlockTransfer|      ABlockTransfer|     ABlockTransfer|      ABlockTransfer|     BBlockTransfer|       BBlockTransfer| BBlockTransfer| BBlockTransfer|      BBlockTransfer|     BBlockTransfer|      BBlockTransfer|     CThreadTransfer| CThreadTransfer|    CThreadTransfer|
+        //####################| Layout| Layout|      Layout| Layout|   Type|   Type|    Type|        Type|    Type| Elementwise| Elementwise|   Elementwise| Specialization|  Size| Block| Block| Block|   | ThreadM| Thread| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
+        //####################|       |       |            |       |       |       |        |            |        |   Operation|   Operation|     Operation|               |      |      |      |      |   |        |       |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
+        //####################|       |       |            |       |       |       |        |            |        |            |            |              |               |      |      |      |      |   |        |       |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        DeviceGemmMultipleD_Dl<    Row,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,   256,   128,   128,    16,  4,       4,      4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>,      S<8, 1, 1, 4>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 4>,      S<1, 2, 0, 3>,       S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>,               5,                  4>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..4c02f8179c
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
@@ -0,0 +1,32 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Layout(A, B, C) = [Col, Row, Row]
+void add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instances<Mul_Clamp>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..2c635e96d2
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
@@ -0,0 +1,32 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Layout(A, B, C) = [Col, Col, Row]
+void add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instances<Mul_Clamp>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..c5a52f2cf2
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
@@ -0,0 +1,32 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Layout(A, B, C) = [Row, Row, Row]
+void add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instances<Mul_Clamp>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..fe4a0b4d68
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
@@ -0,0 +1,32 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Layout(A, B, C) = [Row, Col, Row]
+void add_device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(
+        instances, device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instances<Mul_Clamp>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp
new file mode 100644
index 0000000000..d1b18ac492
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp
@@ -0,0 +1,116 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_quantization_common.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <typename OutElementOp, LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
+using device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|      DsData|   EData|           A|           B|           CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|     LoopScheduler|     Pipeline|
+        //##############################| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|        Type|    Type| Elementwise| Elementwise|   Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                  |             |
+        //##############################|       |       |            |       |       |       |        |         |            |        |   Operation|   Operation|     Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                  |             |
+        //##############################|       |       |            |       |       |       |        |         |            |        |            |            |              |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                  |             |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   256,   128,    64,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,   256,    64,   4,   4,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,   256,    64,  16,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,   128,   128,    64,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,   128,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,   128,    64,   4,   4,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,   128,    64,    64,   4,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 2>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,   128,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 2>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,    64,   128,    64,   4,   4,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,    64,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,    64,    64,   4,   4,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,    64,    64,  16,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,    64,   128,    64,   4,   4,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,    64,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>
+    // clang-format on
+    >;
+
+template <typename OutElementOp, LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
+using device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|      DsData|   EData|           A|           B|           CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|     LoopScheduler|     Pipeline|
+        //##############################| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|        Type|    Type| Elementwise| Elementwise|   Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                  |             |
+        //##############################|       |       |            |       |       |       |        |         |            |        |   Operation|   Operation|     Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                  |             |
+        //##############################|       |       |            |       |       |       |        |         |            |        |            |            |              |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                  |             |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   256,   128,    64,   4,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,   256,    64,   4,  16,   32,   32,    2,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,   256,    64,  16,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,   128,   128,    64,   4,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,   128,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,   128,    64,   4,  16,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,   128,    64,    64,   4,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 2>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,   128,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 2>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,    64,   128,    64,   4,  16,   32,   32,    2,    2,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,    64,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,    64,    64,   4,  16,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,    64,    64,  16,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,    64,   128,    64,   4,  16,   32,   32,    1,    2,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Col,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,    64,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,             16,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>
+    // clang-format on
+    >;
+
+template <typename OutElementOp, LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
+using device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|      DsData|   EData|           A|           B|           CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|     LoopScheduler|     Pipeline|
+        //##############################| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|        Type|    Type| Elementwise| Elementwise|   Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                  |             |
+        //##############################|       |       |            |       |       |       |        |         |            |        |   Operation|   Operation|     Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                  |             |
+        //##############################|       |       |            |       |       |       |        |         |            |        |            |            |              |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                  |             |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   256,   128,    64,  16,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,   256,    64,  16,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,   256,    64,  16,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,   128,   128,    64,  16,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,   128,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,   128,    64,  16,   4,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,   128,    64,    64,  16,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 2>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,   128,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 2>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,    64,   128,    64,  16,   4,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,    64,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,             16,         1,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,    64,    64,  16,   4,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,    64,    64,  16,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,    64,   128,    64,  16,   4,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,         0,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Row, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,    64,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,             16,         1,           1,           1,               S<1, 64, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>
+    // clang-format on
+    >;
+
+template <typename OutElementOp, LoopScheduler GemmLoopScheduler, PipelineVersion GemmPipeline>
+using device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances = std::tuple<
+    // clang-format off
+        //##############################|      A|      B|          Ds|      E|  AData|  BData| AccData| CShuffle|      DsData|   EData|           A|           B|           CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|     LoopScheduler|     Pipeline|
+        //##############################| Layout| Layout|      Layout| Layout|   Type|   Type|    Type| DataType|        Type|    Type| Elementwise| Elementwise|   Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|                  |             |
+        //##############################|       |       |            |       |       |       |        |         |            |        |   Operation|   Operation|     Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                  |             |
+        //##############################|       |       |            |       |       |       |        |         |            |        |            |            |              |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                  |             |
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   256,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,   256,    64,  16,  16,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,   128,   128,    64,  16,  16,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,   128,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 2>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,    64,   128,    64,  16,  16,   32,   32,    2,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,    64,    64,    64,    64,  16,  16,   32,   32,    2,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 2>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,   128,    64,    64,  16,  16,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   256,    64,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,   128,    32,    64,  16,  16,   32,   32,    2,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 2>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,   128,    32,   128,    64,  16,  16,   32,   32,    1,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,    64,    64,    32,    64,  16,  16,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 2>,              16, GemmLoopScheduler, GemmPipeline>,
+        DeviceGemmMultipleD_Xdl_CShuffle<    Row,    Col, Empty_Tuple,    Row, int8_t, int8_t, int32_t,  int32_t, Empty_Tuple,  int8_t, PassThrough, PassThrough,  OutElementOp,     MNKPadding,        1,    64,    32,    64,    64,  16,  16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,             16,             16,         1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 16, 1, 2>,              16, GemmLoopScheduler, GemmPipeline>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
new file mode 100644
index 0000000000..e3163c32ed
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Layout(A, B, C) = [Col, Row, Row]
+void add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances<Mul_Clamp,
+                                                                           LoopScheduler::Default,
+                                                                           PipelineVersion::v1>{});
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances<Mul_Clamp,
+                                                                           LoopScheduler::Interwave,
+                                                                           PipelineVersion::v1>{});
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances<Mul_Clamp,
+                                                                           LoopScheduler::Default,
+                                                                           PipelineVersion::v2>{});
+#endif
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
new file mode 100644
index 0000000000..f6a0163aba
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Layout(A, B, C) = [Col, Col, Row]
+void add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances<Mul_Clamp,
+                                                                           LoopScheduler::Default,
+                                                                           PipelineVersion::v1>{});
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances<Mul_Clamp,
+                                                                           LoopScheduler::Interwave,
+                                                                           PipelineVersion::v1>{});
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances<Mul_Clamp,
+                                                                           LoopScheduler::Default,
+                                                                           PipelineVersion::v2>{});
+#endif
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
new file mode 100644
index 0000000000..0cfffbed70
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Layout(A, B, C) = [Row, Row, Row]
+void add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances<Mul_Clamp,
+                                                                           LoopScheduler::Default,
+                                                                           PipelineVersion::v1>{});
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances<Mul_Clamp,
+                                                                           LoopScheduler::Interwave,
+                                                                           PipelineVersion::v1>{});
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances<Mul_Clamp,
+                                                                           LoopScheduler::Default,
+                                                                           PipelineVersion::v2>{});
+#endif
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
new file mode 100644
index 0000000000..289898257b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Layout(A, B, C) = [Row, Col, Row]
+void add_device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Empty_Tuple,
+                                                    Row,
+                                                    int8_t,
+                                                    int8_t,
+                                                    Empty_Tuple,
+                                                    int8_t,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    Mul_Clamp>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances<Mul_Clamp,
+                                                                           LoopScheduler::Default,
+                                                                           PipelineVersion::v1>{});
+#if CK_EXPERIMENTAL_INTER_WAVE_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances<Mul_Clamp,
+                                                                           LoopScheduler::Interwave,
+                                                                           PipelineVersion::v1>{});
+#endif
+#if CK_EXPERIMENTAL_PIPELINE_V2_INSTANCES
+    add_device_operation_instances(
+        instances,
+        device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances<Mul_Clamp,
+                                                                           LoopScheduler::Default,
+                                                                           PipelineVersion::v2>{});
+#endif
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/gemm_quantization_common.hpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/gemm_quantization_common.hpp
new file mode 100644
index 0000000000..e7c2500fef
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/quantization/gemm/gemm_quantization_common.hpp
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using Empty_Tuple   = ck::Tuple<>;
+using Row_Row_Tuple = ck::Tuple<Row, Row>;
+using Col_Col_Tuple = ck::Tuple<Col, Col>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using Relu        = ck::tensor_operation::element_wise::Relu;
+
+using Mul_Clamp      = ck::tensor_operation::element_wise::Activation_Mul_Clamp<PassThrough>;
+using Relu_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Relu>;
+
+using Add_Mul_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<PassThrough>;
+using Add_Relu_Mul_Clamp = ck::tensor_operation::element_wise::Add_Activation_Mul_Clamp<Relu>;
+
+static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt
new file mode 100644
index 0000000000..31ae7226f4
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt
@@ -0,0 +1,76 @@
+add_instance_library(device_reduce_instance
+   device_reduce_instance_blockwise_f16_f16_f16_min.cpp
+   device_reduce_instance_blockwise_f16_f16_f16_max.cpp
+   device_reduce_instance_blockwise_f16_f16_f16_amax.cpp
+   device_reduce_instance_blockwise_f16_f32_f16_add.cpp
+   device_reduce_instance_blockwise_f16_f32_f16_avg.cpp
+   device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp
+   device_reduce_instance_blockwise_f32_f32_f32_add.cpp
+   device_reduce_instance_blockwise_f32_f32_f32_avg.cpp
+   device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp
+   device_reduce_instance_blockwise_f32_f32_f32_min.cpp
+   device_reduce_instance_blockwise_f32_f32_f32_max.cpp
+   device_reduce_instance_blockwise_f32_f32_f32_amax.cpp
+   device_reduce_instance_blockwise_f32_f64_f32_add.cpp
+   device_reduce_instance_blockwise_f32_f64_f32_avg.cpp
+   device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp
+   device_reduce_instance_blockwise_f64_f64_f64_add.cpp
+   device_reduce_instance_blockwise_f64_f64_f64_avg.cpp
+   device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp
+   device_reduce_instance_blockwise_f64_f64_f64_min.cpp
+   device_reduce_instance_blockwise_f64_f64_f64_max.cpp
+   device_reduce_instance_blockwise_f64_f64_f64_amax.cpp
+   device_reduce_instance_blockwise_i8_i32_i8_add.cpp
+   device_reduce_instance_blockwise_i8_i32_i8_avg.cpp
+   device_reduce_instance_blockwise_i8_i8_i8_min.cpp   
+   device_reduce_instance_blockwise_i8_i8_i8_max.cpp   
+   device_reduce_instance_blockwise_i8_i8_i8_amax.cpp   
+   device_reduce_instance_blockwise_b16_f32_b16_add.cpp
+   device_reduce_instance_blockwise_b16_f32_b16_avg.cpp
+   device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp
+   device_reduce_instance_blockwise_b16_f32_b16_min.cpp
+   device_reduce_instance_blockwise_b16_f32_b16_max.cpp
+   device_reduce_instance_blockwise_b16_f32_b16_amax.cpp
+   device_reduce_instance_threadwise_f16_f16_f16_min.cpp
+   device_reduce_instance_threadwise_f16_f16_f16_max.cpp
+   device_reduce_instance_threadwise_f16_f16_f16_amax.cpp
+   device_reduce_instance_threadwise_f16_f32_f16_add.cpp
+   device_reduce_instance_threadwise_f16_f32_f16_avg.cpp
+   device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp
+   device_reduce_instance_threadwise_f32_f32_f32_add.cpp
+   device_reduce_instance_threadwise_f32_f32_f32_avg.cpp
+   device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp
+   device_reduce_instance_threadwise_f32_f32_f32_min.cpp
+   device_reduce_instance_threadwise_f32_f32_f32_max.cpp
+   device_reduce_instance_threadwise_f32_f32_f32_amax.cpp
+   device_reduce_instance_threadwise_f32_f64_f32_add.cpp
+   device_reduce_instance_threadwise_f32_f64_f32_avg.cpp
+   device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp
+   device_reduce_instance_threadwise_f64_f64_f64_add.cpp
+   device_reduce_instance_threadwise_f64_f64_f64_avg.cpp
+   device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp
+   device_reduce_instance_threadwise_f64_f64_f64_min.cpp
+   device_reduce_instance_threadwise_f64_f64_f64_max.cpp
+   device_reduce_instance_threadwise_f64_f64_f64_amax.cpp
+   device_reduce_instance_threadwise_i8_i32_i8_add.cpp
+   device_reduce_instance_threadwise_i8_i32_i8_avg.cpp
+   device_reduce_instance_threadwise_i8_i8_i8_min.cpp
+   device_reduce_instance_threadwise_i8_i8_i8_max.cpp
+   device_reduce_instance_threadwise_i8_i8_i8_amax.cpp
+   device_reduce_instance_threadwise_b16_f32_b16_add.cpp
+   device_reduce_instance_threadwise_b16_f32_b16_avg.cpp
+   device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp
+   device_reduce_instance_threadwise_b16_f32_b16_min.cpp
+   device_reduce_instance_threadwise_b16_f32_b16_max.cpp
+   device_reduce_instance_threadwise_b16_f32_b16_amax.cpp
+   device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp
+   device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp
+   device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp
+   device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp
+   device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp
+   device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp
+   device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp
+   device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp
+   device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp
+   device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp
+)
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.cpp
new file mode 100644
index 0000000000..4771b4aa5d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp
new file mode 100644
index 0000000000..f684dabc1f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp
new file mode 100644
index 0000000000..86e2ca7c47
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.cpp
new file mode 100644
index 0000000000..97d09ff834
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.cpp
new file mode 100644
index 0000000000..87187f493a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp
new file mode 100644
index 0000000000..d3f5cd8656
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp
new file mode 100644
index 0000000000..fb4a50d5e3
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.cpp
new file mode 100644
index 0000000000..f74d22f7ac
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.cpp
new file mode 100644
index 0000000000..0982455add
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.cpp
new file mode 100644
index 0000000000..c6f1c3a633
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp
new file mode 100644
index 0000000000..db9f52a13f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp
new file mode 100644
index 0000000000..1410f51bea
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.cpp
new file mode 100644
index 0000000000..707a655703
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp
new file mode 100644
index 0000000000..17f45c3327
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp
new file mode 100644
index 0000000000..d3753d01c4
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.cpp
new file mode 100644
index 0000000000..83fad7826a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.cpp
new file mode 100644
index 0000000000..c627f68636
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp
new file mode 100644
index 0000000000..09686d98ff
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F32, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.cpp
new file mode 100644
index 0000000000..44519032f3
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp
new file mode 100644
index 0000000000..9e14d80e21
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp
new file mode 100644
index 0000000000..ab217f1160
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F32, F64, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.cpp
new file mode 100644
index 0000000000..3f611da7e0
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp
new file mode 100644
index 0000000000..8695db4004
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp
new file mode 100644
index 0000000000..8d15fda786
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.cpp
new file mode 100644
index 0000000000..c161858b1f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.cpp
new file mode 100644
index 0000000000..e6b87fdead
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp
new file mode 100644
index 0000000000..c2e4acc8d9
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_blockwise<F64, F64, F64, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.cpp
new file mode 100644
index 0000000000..e5a5e07e17
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp
new file mode 100644
index 0000000000..22c6203345
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_blockwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp
new file mode 100644
index 0000000000..1890fb78cd
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.cpp
new file mode 100644
index 0000000000..dc5caf297b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.cpp
new file mode 100644
index 0000000000..02bb0c3b1d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_blockwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp
new file mode 100644
index 0000000000..0b0a403038
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp
new file mode 100644
index 0000000000..fcf95aab7a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<BF16, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp
new file mode 100644
index 0000000000..5151d32a58
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp
new file mode 100644
index 0000000000..5463cc2b4c
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F16, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp
new file mode 100644
index 0000000000..a15e99fe0f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp
new file mode 100644
index 0000000000..ee7a4dd4ce
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp
new file mode 100644
index 0000000000..58f7295a83
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp
new file mode 100644
index 0000000000..43f2370a2c
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp
new file mode 100644
index 0000000000..eaf84a6de1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp
new file mode 100644
index 0000000000..76d63c9a96
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.cpp
new file mode 100644
index 0000000000..eb1d389c39
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp
new file mode 100644
index 0000000000..61225df05d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp
new file mode 100644
index 0000000000..aef231e17a
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.cpp
new file mode 100644
index 0000000000..6d2939e3e0
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.cpp
new file mode 100644
index 0000000000..ff0ddd773b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp
new file mode 100644
index 0000000000..1367022065
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp
new file mode 100644
index 0000000000..f0f9301918
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.cpp
new file mode 100644
index 0000000000..db9f156600
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.cpp
new file mode 100644
index 0000000000..e5644faf29
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.cpp
new file mode 100644
index 0000000000..350259f5f0
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp
new file mode 100644
index 0000000000..89cae55273
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp
new file mode 100644
index 0000000000..b0f82ab453
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F16, F32, F16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.cpp
new file mode 100644
index 0000000000..fedd0f890b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp
new file mode 100644
index 0000000000..64ec1ce157
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp
new file mode 100644
index 0000000000..156f182f55
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.cpp
new file mode 100644
index 0000000000..22b278790d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.cpp
new file mode 100644
index 0000000000..f3070d2975
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp
new file mode 100644
index 0000000000..e17476ba75
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F32, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp
new file mode 100644
index 0000000000..2387725b9f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp
new file mode 100644
index 0000000000..8be94118c4
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp
new file mode 100644
index 0000000000..9ff01615c9
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F32, F64, F32, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp
new file mode 100644
index 0000000000..8a99c7c958
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp
new file mode 100644
index 0000000000..78909c8dbf
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp
new file mode 100644
index 0000000000..7f094c28b7
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp
new file mode 100644
index 0000000000..4dfb477c1e
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp
new file mode 100644
index 0000000000..e492833606
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp
@@ -0,0 +1,27 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp
new file mode 100644
index 0000000000..de8dce357b
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp
@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<F64, F64, F64, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp
new file mode 100644
index 0000000000..0225e75cb0
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
+// clang-format on
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp
new file mode 100644
index 0000000000..6d04d2128f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp
new file mode 100644
index 0000000000..046643abe1
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp
new file mode 100644
index 0000000000..a61302c001
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp
new file mode 100644
index 0000000000..7619e12a8e
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp
@@ -0,0 +1,28 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt
new file mode 100644
index 0000000000..ba0197477f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt
@@ -0,0 +1,20 @@
+set(DEVICE_SOFTMAX_INSTANCES)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_SOFTMAX_INSTANCES device_softmax_f16_f16_instance_rank3_reduce1.cpp
+    device_softmax_f16_f16_instance_rank3_reduce2.cpp
+    device_softmax_f16_f16_instance_rank3_reduce3.cpp
+    device_softmax_f16_f16_instance_rank4_reduce1.cpp
+    device_softmax_f16_f16_instance_rank4_reduce2.cpp
+    device_softmax_f16_f16_instance_rank4_reduce3.cpp
+    device_softmax_f16_f16_instance_rank4_reduce4.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_SOFTMAX_INSTANCES device_softmax_f32_f32_instance_rank3_reduce1.cpp
+    device_softmax_f32_f32_instance_rank3_reduce2.cpp
+    device_softmax_f32_f32_instance_rank3_reduce3.cpp
+    device_softmax_f32_f32_instance_rank4_reduce1.cpp
+    device_softmax_f32_f32_instance_rank4_reduce2.cpp
+    device_softmax_f32_f32_instance_rank4_reduce3.cpp
+    device_softmax_f32_f32_instance_rank4_reduce4.cpp)
+endif()
+add_instance_library(device_softmax_instance ${DEVICE_SOFTMAX_INSTANCES})
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp
new file mode 100644
index 0000000000..36867d993f
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3, 1>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_generic_instance<3, 1>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp
new file mode 100644
index 0000000000..373f33ad59
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3, 2>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_generic_instance<3, 2>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp
new file mode 100644
index 0000000000..d26b92b4f4
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3, 3>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_generic_instance<3, 3>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp
new file mode 100644
index 0000000000..bbb735b6fe
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 1>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_generic_instance<4, 1>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp
new file mode 100644
index 0000000000..92dbe67760
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 2>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_generic_instance<4, 2>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp
new file mode 100644
index 0000000000..354cda85d7
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 3>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_generic_instance<4, 3>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp
new file mode 100644
index 0000000000..edb5e42c10
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank4_reduce4_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 4>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_generic_instance<4, 4>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 4>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp
new file mode 100644
index 0000000000..566be8fc22
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank3_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 1>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<3, 1>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp
new file mode 100644
index 0000000000..f9c76e3116
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank3_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 2>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<3, 2>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp
new file mode 100644
index 0000000000..541e0d71a9
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank3_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 3>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<3, 3>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp
new file mode 100644
index 0000000000..95a38df283
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank4_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 1>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<4, 1>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp
new file mode 100644
index 0000000000..a29b88891d
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank4_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 2>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<4, 2>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp
new file mode 100644
index 0000000000..0da46ea1b4
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank4_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 3>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<4, 3>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp
new file mode 100644
index 0000000000..fa217dc3f5
--- /dev/null
+++ b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp
@@ -0,0 +1,26 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank4_reduce4_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 4>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<4, 4>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 4>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/library/src/utility/CMakeLists.txt b/library/src/utility/CMakeLists.txt
new file mode 100644
index 0000000000..7f6a59eebe
--- /dev/null
+++ b/library/src/utility/CMakeLists.txt
@@ -0,0 +1,28 @@
+## utility
+set(UTILITY_SOURCE
+    device_memory.cpp
+    host_tensor.cpp
+    convolution_parameter.cpp
+)
+
+add_library(utility STATIC ${UTILITY_SOURCE})
+add_library(composable_kernel::utility ALIAS utility)
+
+target_include_directories(utility PUBLIC
+    "$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck>"
+    "$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/utility>"
+)
+
+rocm_install(
+    TARGETS utility
+    EXPORT utilityTargets
+)
+
+rocm_install(
+    EXPORT utilityTargets
+    FILE composable_kernelutilityTargets.cmake
+    NAMESPACE composable_kernel::
+    DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
+)
+
+clang_tidy_check(utility)
diff --git a/library/src/utility/convolution_parameter.cpp b/library/src/utility/convolution_parameter.cpp
new file mode 100644
index 0000000000..57cedd6019
--- /dev/null
+++ b/library/src/utility/convolution_parameter.cpp
@@ -0,0 +1,171 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/host_utility/io.hpp"
+
+#include "ck/library/utility/convolution_parameter.hpp"
+
+namespace ck {
+namespace utils {
+namespace conv {
+
+ConvParam::ConvParam(ck::index_t n_dim,
+                     ck::index_t group_count,
+                     ck::index_t n_batch,
+                     ck::index_t n_out_channels,
+                     ck::index_t n_in_channels,
+                     const std::vector<ck::index_t>& filters_len,
+                     const std::vector<ck::index_t>& input_len,
+                     const std::vector<ck::index_t>& strides,
+                     const std::vector<ck::index_t>& dilations,
+                     const std::vector<ck::index_t>& left_pads,
+                     const std::vector<ck::index_t>& right_pads)
+    : num_dim_spatial_(n_dim),
+      G_(group_count),
+      N_(n_batch),
+      K_(n_out_channels),
+      C_(n_in_channels),
+      filter_spatial_lengths_(filters_len),
+      input_spatial_lengths_(input_len),
+      output_spatial_lengths_(num_dim_spatial_),
+      conv_filter_strides_(strides),
+      conv_filter_dilations_(dilations),
+      input_left_pads_(left_pads),
+      input_right_pads_(right_pads)
+{
+    if(static_cast<ck::index_t>(filter_spatial_lengths_.size()) != num_dim_spatial_ ||
+       static_cast<ck::index_t>(input_spatial_lengths_.size()) != num_dim_spatial_ ||
+       static_cast<ck::index_t>(conv_filter_strides_.size()) != num_dim_spatial_ ||
+       static_cast<ck::index_t>(conv_filter_dilations_.size()) != num_dim_spatial_ ||
+       static_cast<ck::index_t>(input_left_pads_.size()) != num_dim_spatial_ ||
+       static_cast<ck::index_t>(input_right_pads_.size()) != num_dim_spatial_)
+    {
+        throw(
+            std::runtime_error("ConvParam::ConvParam: "
+                               "parameter size is different from number of declared dimensions!"));
+    }
+
+    for(ck::index_t i = 0; i < num_dim_spatial_; ++i)
+    {
+        // XEff = (X - 1) * conv_dilation_w + 1;
+        // Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
+        const ck::index_t x_eff = (filter_spatial_lengths_[i] - 1) * conv_filter_dilations_[i] + 1;
+
+        output_spatial_lengths_[i] =
+            (input_spatial_lengths_[i] + input_left_pads_[i] + input_right_pads_[i] - x_eff) /
+                conv_filter_strides_[i] +
+            1;
+    }
+}
+
+ConvParam::ConvParam()
+    : ConvParam::ConvParam(2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1})
+{
+}
+
+std::vector<ck::index_t> ConvParam::GetOutputSpatialLengths() const
+{
+    return output_spatial_lengths_;
+}
+
+std::size_t ConvParam::GetFlops() const
+{
+    // 2 * G * N * K * C * <output spatial lengths product> * <filter spatial lengths product>
+    return static_cast<std::size_t>(2) * G_ * N_ * K_ * C_ *
+           ck::accumulate_n<std::size_t>(
+               std::begin(output_spatial_lengths_), num_dim_spatial_, 1, std::multiplies<>()) *
+           ck::accumulate_n<std::size_t>(
+               std::begin(filter_spatial_lengths_), num_dim_spatial_, 1, std::multiplies<>());
+}
+
+std::string get_conv_param_parser_helper_msg()
+{
+    std::string msg;
+
+    msg += "Following arguments (depending on number of spatial dims):\n"
+           " Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)\n"
+           " G, N, K, C, \n"
+           " <filter spatial dimensions>, (ie Y, X for 2D)\n"
+           " <input image spatial dimensions>, (ie Hi, Wi for 2D)\n"
+           " <strides>, (ie Sy, Sx for 2D)\n"
+           " <dilations>, (ie Dy, Dx for 2D)\n"
+           " <left padding>, (ie LeftPy, LeftPx for 2D)\n"
+           " <right padding>, (ie RightPy, RightPx for 2D)\n";
+
+    return msg;
+}
+
+ck::utils::conv::ConvParam parse_conv_param(int num_dim_spatial, int arg_idx, char* const argv[])
+{
+    const ck::index_t G = std::stoi(argv[arg_idx++]);
+    const ck::index_t N = std::stoi(argv[arg_idx++]);
+    const ck::index_t K = std::stoi(argv[arg_idx++]);
+    const ck::index_t C = std::stoi(argv[arg_idx++]);
+
+    std::vector<ck::index_t> filter_spatial_lengths(num_dim_spatial);
+    std::vector<ck::index_t> input_spatial_lengths(num_dim_spatial);
+    std::vector<ck::index_t> conv_filter_strides(num_dim_spatial);
+    std::vector<ck::index_t> conv_filter_dilations(num_dim_spatial);
+    std::vector<ck::index_t> input_left_pads(num_dim_spatial);
+    std::vector<ck::index_t> input_right_pads(num_dim_spatial);
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        input_left_pads[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        input_right_pads[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    return ck::utils::conv::ConvParam{num_dim_spatial,
+                                      G,
+                                      N,
+                                      K,
+                                      C,
+                                      filter_spatial_lengths,
+                                      input_spatial_lengths,
+                                      conv_filter_strides,
+                                      conv_filter_dilations,
+                                      input_left_pads,
+                                      input_right_pads};
+}
+} // namespace conv
+} // namespace utils
+} // namespace ck
+
+std::ostream& operator<<(std::ostream& os, const ck::utils::conv::ConvParam& p)
+{
+    os << "ConvParam {"
+       << "\nnum_dim_spatial: " << p.num_dim_spatial_ << "\nG: " << p.G_ << "\nN: " << p.N_
+       << "\nK: " << p.K_ << "\nC: " << p.C_
+       << "\nfilter_spatial_lengths: " << p.filter_spatial_lengths_
+       << "\ninput_spatial_lengths: " << p.input_spatial_lengths_
+       << "\nconv_filter_strides: " << p.conv_filter_strides_
+       << "\nconv_filter_dilations: " << p.conv_filter_dilations_
+       << "\ninput_left_pads: " << p.input_left_pads_
+       << "\ninput_right_pads: " << p.input_right_pads_ << "}\n";
+
+    return os;
+}
diff --git a/library/src/utility/device_memory.cpp b/library/src/utility/device_memory.cpp
new file mode 100644
index 0000000000..61b6326b57
--- /dev/null
+++ b/library/src/utility/device_memory.cpp
@@ -0,0 +1,76 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/host_utility/hip_check_error.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+
+DeviceMem::DeviceMem(std::size_t mem_size) : mMemSize(mem_size)
+{
+    hip_check_error(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
+}
+
+void DeviceMem::Realloc(std::size_t mem_size)
+{
+    if(mpDeviceBuf)
+    {
+        hip_check_error(hipFree(mpDeviceBuf));
+    }
+    mMemSize = mem_size;
+    hip_check_error(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
+}
+
+void* DeviceMem::GetDeviceBuffer() const { return mpDeviceBuf; }
+
+std::size_t DeviceMem::GetBufferSize() const { return mMemSize; }
+
+void DeviceMem::ToDevice(const void* p) const
+{
+    if(mpDeviceBuf)
+    {
+        hip_check_error(
+            hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
+    }
+    else
+    {
+        throw std::runtime_error("ToDevice with an empty pointer");
+    }
+}
+
+void DeviceMem::ToDevice(const void* p, const std::size_t cpySize) const
+{
+    hip_check_error(hipMemcpy(mpDeviceBuf, const_cast<void*>(p), cpySize, hipMemcpyHostToDevice));
+}
+
+void DeviceMem::FromDevice(void* p) const
+{
+    if(mpDeviceBuf)
+    {
+        hip_check_error(hipMemcpy(p, mpDeviceBuf, mMemSize, hipMemcpyDeviceToHost));
+    }
+    else
+    {
+        throw std::runtime_error("FromDevice with an empty pointer");
+    }
+}
+
+void DeviceMem::FromDevice(void* p, const std::size_t cpySize) const
+{
+    hip_check_error(hipMemcpy(p, mpDeviceBuf, cpySize, hipMemcpyDeviceToHost));
+}
+
+void DeviceMem::SetZero() const
+{
+    if(mpDeviceBuf)
+    {
+        hip_check_error(hipMemset(mpDeviceBuf, 0, mMemSize));
+    }
+}
+
+DeviceMem::~DeviceMem()
+{
+    if(mpDeviceBuf)
+    {
+        hip_check_error(hipFree(mpDeviceBuf));
+    }
+}
diff --git a/library/src/utility/host_tensor.cpp b/library/src/utility/host_tensor.cpp
new file mode 100644
index 0000000000..7211552641
--- /dev/null
+++ b/library/src/utility/host_tensor.cpp
@@ -0,0 +1,59 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cassert>
+
+#include "ck/library/utility/host_tensor.hpp"
+
+void HostTensorDescriptor::CalculateStrides()
+{
+    mStrides.clear();
+    mStrides.resize(mLens.size(), 0);
+    if(mStrides.empty())
+        return;
+
+    mStrides.back() = 1;
+    std::partial_sum(
+        mLens.rbegin(), mLens.rend() - 1, mStrides.rbegin() + 1, std::multiplies<std::size_t>());
+}
+
+std::size_t HostTensorDescriptor::GetNumOfDimension() const { return mLens.size(); }
+
+std::size_t HostTensorDescriptor::GetElementSize() const
+{
+    assert(mLens.size() == mStrides.size());
+    return std::accumulate(
+        mLens.begin(), mLens.end(), std::size_t{1}, std::multiplies<std::size_t>());
+}
+
+std::size_t HostTensorDescriptor::GetElementSpaceSize() const
+{
+    std::size_t space = 1;
+    for(std::size_t i = 0; i < mLens.size(); ++i)
+    {
+        if(mLens[i] == 0)
+            continue;
+
+        space += (mLens[i] - 1) * mStrides[i];
+    }
+    return space;
+}
+
+const std::vector<std::size_t>& HostTensorDescriptor::GetLengths() const { return mLens; }
+
+const std::vector<std::size_t>& HostTensorDescriptor::GetStrides() const { return mStrides; }
+
+std::ostream& operator<<(std::ostream& os, const HostTensorDescriptor& desc)
+{
+    os << "dim " << desc.GetNumOfDimension() << ", ";
+
+    os << "lengths {";
+    LogRange(os, desc.GetLengths(), ", ");
+    os << "}, ";
+
+    os << "strides {";
+    LogRange(os, desc.GetStrides(), ", ");
+    os << "}";
+
+    return os;
+}
diff --git a/profiler/CMakeLists.txt b/profiler/CMakeLists.txt
new file mode 100644
index 0000000000..bdd7125ac1
--- /dev/null
+++ b/profiler/CMakeLists.txt
@@ -0,0 +1,5 @@
+include_directories(BEFORE
+    ${CMAKE_CURRENT_LIST_DIR}/include
+)
+
+add_subdirectory(src)
diff --git a/profiler/README.md b/profiler/README.md
new file mode 100644
index 0000000000..d03bfa7fc4
--- /dev/null
+++ b/profiler/README.md
@@ -0,0 +1,224 @@
+## Profile GEMM kernels
+```bash
+#arg1: tensor operation (gemm=GEMM)
+#arg2: data type (0=fp32, 1=fp16)
+#arg3: matrix layout (0=NN, 1=NT, 2=TN, 3=TT)
+#arg4: verification (0=no, 1=yes)
+#arg5: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg6: print matrix value (0=no, 1=yes)
+#arg7: run kernel # of times (>1)
+#arg8 to 13: M, N, K, StrideA, StrideB, StrideC
+
+################        op  datatype  layout  verify  init  log  repeat  M___ N___ K___  StrideA StrideB StrideC
+./bin/ckProfiler      gemm         1       1       1     1    0       5  3840 4096 4096     4096    4096    4096
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```bash
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+....
+Best Perf: 1.1933 ms, 107.977 TFlops, 79.0848 GB/s
+```
+
+## Profile 2d forward convolution kernels
+```bash
+#arg1: tensor operation (conv=Convolution)
+#arg2: data type (0=fp32, 1=fp16)
+#arg3: input tensor layout (0=NCHW, 1=NHWC)
+#arg4: weight tensor layout (0=KCYX, 1=KYXC)
+#arg5: output tensor layout (0=NKHW, 1=NHWK)
+#arg6: verification (0=no, 1=yes)
+#arg7: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg8: print matrix value (0=no, 1=yes)
+#arg9: run kernel # of times (>1)
+#arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, RightPx
+ ################          op datatype  in_layout   wei_layout  out_layout  verify  init  log  repeat  N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
+ ./bin/ckProfiler  conv2d_fwd        1          1            1           1       1     1    0       5  128  256  192 3 3   71   71     2 2       1 1      1 1       1 1
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
+wei_k_c_y_x: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
+out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
+....
+Best Perf: 1.42509 ms, 102.988 TFlops, 234.086 GB/s
+```
+
+## Profile contraction kernels
+```bash
+#arg1: tensor operation (contraction_bilinear=CONTRACTION+Bilinear)
+#arg2: data type (0: fp32; 1: f64)\n"
+#arg3: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
+#                     1: A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
+#                     2: A[k0, k1, m0, m1] * B[k0, k1, n0, n1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1];
+#                     3: A[k0, k1, m0, m1] * B[n0, n1, k0, k1] + D[m0, m1, n0, n1] = E[m0, m1, n0, n1])
+#arg4: verification (0: no; 1: yes)
+#arg5: initialization (0: no init; 1: integer value; 2: decimal value)
+#arg6: print tensor value (0: no; 1: yes)
+#arg7: time kernel (0: no, 1: yes)
+#arg8 and arg9: alpha and beta
+#arg10 to 15: M0, M1, N0, N1, K0, K1
+#arg16 to 31: Strides for A, B, D and E (skip for default)
+
+################                   op  datatype  layout  verify  init  log  time  alpha  beta  M0  M1  N0  N1  K0  K1
+./bin/ckProfiler contraction_bilinear         0       1       0     0    0     1    1.0   1.0 128 128 128 128 128 128
+```
+
+Result (MI100)
+```bash
+a_m_k: dim 4, lengths {128, 128, 128, 128}, strides {2097152, 16384, 128, 1}
+b_k_n: dim 4, lengths {128, 128, 128, 128}, strides {128, 1, 2097152, 16384}
+d_m_n: dim 4, lengths {128, 128, 128, 128}, strides {2097152, 16384, 128, 1}
+e_m_n: dim 4, lengths {128, 128, 128, 128}, strides {2097152, 16384, 128, 1}
+....
+Best Perf: 211.405 ms, 41.6077 TFlops, 15.2372 GB/s
+```
+
+## Profile batched gemm multiple D kernels
+```bash
+#arg1: tensor operation (batched_gemm_multi_d=Batched GEMM multi D);
+#arg2: data type (0: fp16; 1: int8)
+#arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];
+#                     1: A[g, m, k] * B[g, n, k] = C[g, m, n];
+#                     2: A[g, k, m] * B[g, k, n] = C[g, m, n];
+#                     3: A[g, k, m] * B[g, n, k] = C[g, m, n])
+#arg4: verification (0: no; 1: yes)
+#arg5: initialization (0: no init; 1: integer value; 2: decimal value)
+#arg6: print tensor value (0: no; 1: yes)
+#arg7: time kernel (0=n0, 1=yes)
+#arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount
+
+################                   op  datatype  layout  verify  init  log  time    M    N    K StrideA StrideB StrideC BatchStrideA BatchStrideB BatchStrideC BatchCount
+./bin/ckProfiler batched_gemm_multi_d         0       1       0     0    0     1 4096 4096 4096    4096    4096    4096     16777216     16777216     16777216         16
+```
+
+Result (Radeon RX 6800 XT)
+```bash
+arg.a_grid_desc_k0_m0_m1_k1_{2048, 4096, 2}
+arg.b_grid_desc_k0_n0_n1_k1_{2048, 4096, 2}
+arg.e_grid_desc_m_n_{ 4096, 4096}
+....
+Best Perf: 58.0306 ms, 37.8942 TFlops, 27.7545 GB/s
+## Profile grouped convolution backward data kernels
+```bash
+# arg1: tensor operation (grouped_conv_bwd_data: Grouped Convolution Backward Data)
+# arg2: data type (0: Output fp32, Weight fp32, Input fp32
+#                  1: Output fp16, Weight fp16, Input fp16
+#                  2: Output bf16, Weight bf16, Input bf16
+# arg3: tensor layout (0: Output[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Input[G, N, Ho, Wo, K]
+#                      1: Output[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Input[N, Ho, Wo, G, K])
+# arg4: verification (0: no, 1: yes)
+# arg5: initialization (0: no init, 1: integer value, 2: decimal value)
+# arg6: print tensor value (0: no; 1: yes)
+# arg7: time kernel (0: no, 1: yes)
+# Following arguments (depending on number of spatial dims):
+#  Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)
+#  G, N, K, C, 
+#  <filter spatial dimensions>, (ie Y, X for 2D)
+#  <input image spatial dimensions>, (ie Hi, Wi for 2D)
+#  <strides>, (ie Sy, Sx for 2D)
+#  <dilations>, (ie Dy, Dx for 2D)
+#  <left padding>, (ie LeftPy, LeftPx for 2D)
+#  <right padding>, (ie RightPy, RightPx for 2D)
+
+ ################                   op   datatype  layout  verify  init  log  time  Ndims  G  N   K   C  Y  X  Hi  Wi  Sy  Sx  Dy  Dx  LeftPy  LeftPx  RightPy  RightPx
+./bin/ckProfiler grouped_conv_bwd_data          1       0       1     1    0     1      2 32  4 192 192  3  3  28  28   1   1   1   1       1       1        1        1
+
+ ```
+
+Result (MI100, FP16, GNHWC_GKYXC_GNHWK)
+```
+out: dim 5, lengths {32, 4, 192, 28, 28}, strides {602112, 150528, 1, 5376, 192}
+wei: dim 5, lengths {32, 192, 192, 3, 3}, strides {331776, 1728, 1, 576, 192}
+in: dim 5, lengths {32, 4, 192, 28, 28}, strides {602112, 150528, 1, 5376, 192}
+....
+Best configuration parameters:
+name: DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<256, 128, 256, 32, 8, 2, Default, 32, 32, 2, 4, 8, 4, 1, 1>
+avg_time: 0.768321
+tflops: 86.6679
+GB/s: 127.947
+```
+
+## Profile grouped convolution backward weight kernels
+```bash
+# arg1: tensor operation (grouped_conv_bwd_weight: Grouped Convolution Backward Weight)
+# arg2: data type (0: Input fp32, Weight fp32, Output fp32
+#                  1: Input fp16, Weight fp16, Output fp16
+#                  2: Input bf16, Weight fp32, Output bf16)
+# arg3: tensor layout (0: Input[G, N, C, Hi, Wi], Weight[G, K, C, Y, X], Output[G, N, K, Ho, Wo]
+#                      1: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, N, Ho, Wo, K]
+#                      2: Input[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Output[N, Ho, Wo, G, K]
+# arg4: verification (0: no, 1: yes)
+# arg5: initialization (0: no init, 1: integer value, 2: decimal value)
+# arg6: print tensor value (0: no; 1: yes)
+# arg7: time kernel (0: no, 1: yes)
+# Following arguments (depending on number of spatial dims):
+#  Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)
+#  G, N, K, C, 
+#  <filter spatial dimensions>, (ie Y, X for 2D)
+#  <input image spatial dimensions>, (ie Hi, Wi for 2D)
+#  <strides>, (ie Sy, Sx for 2D)
+#  <dilations>, (ie Dy, Dx for 2D)
+#  <left padding>, (ie LeftPy, LeftPx for 2D)
+#  <right padding>, (ie RightPy, RightPx for 2D)
+# SplitK
+
+ ################                   op   datatype  layout  verify  init  log  time  Ndims  G   N   K   C  Y  X  Hi  Wi  Sy  Sx  Dy  Dx  LeftPy  LeftPx  RightPy  RightPx  SplitK
+./bin/ckProfiler grouped_conv_bwd_weight          1       0       1     1    0     1      2 32 256 256 512  3  3  28  28   1   1   1   1       1       0        0        0       1
+
+ ```
+
+Result (MI100, FP16, GNHWC_GKYXC_GNHWK)
+```
+input: dim 5, lengths {32, 512, 1024, 28, 28}, strides {411041792, 802816, 1, 28672, 1024}
+weight: dim 5, lengths {32, 512, 1024, 3, 3}, strides {4718592, 9216, 1, 3072, 1024}
+output: dim 5, lengths {32, 512, 512, 26, 26}, strides {177209344, 346112, 1, 13312, 512}
+....
+Best configuration parameters:
+name: DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 256, 128, 4, Default, 8, 4, 2, 8, 4, 8, 2, 1, 1, 8>
+avg_time: 68.5216
+tflops: 95.337
+GB/s: 69.2301
+```
+Note: This kernel use atomic add, this will cause output buffer to be accumulated multiple times, causing verification failure. To work around it, do not use CK's own timer and do verification at the same time.
+
+## Profile image to column kernels
+```bash
+# arg1: tensor operation (" OP_NAME ": " OP_DESC ")
+# arg2: data type (0: Input fp32, Weight fp32, Output fp32
+#                  1: Input fp16, Weight fp16, Output fp16
+#                  2: Input bf16, Weight bf16, Output bf16
+#                  3: Input int8, Weight int8, Output int8)
+# arg3: tensor layout (0: Input[N, Hi, Wi, C], Output[N * Ho * Wo, Y * X * C])
+# arg4: verification (0: no, 1: yes)
+# arg5: initialization (0: no init, 1: integer value, 2: decimal value)
+# arg6: print tensor value (0: no; 1: yes)
+# arg7: time kernel (0: no, 1: yes)
+# Following arguments (depending on number of spatial dims):
+#  Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)
+#  G, N, K, C, 
+#  <filter spatial dimensions>, (ie Y, X for 2D)
+#  <input image spatial dimensions>, (ie Hi, Wi for 2D)
+#  <strides>, (ie Sy, Sx for 2D)
+#  <dilations>, (ie Dy, Dx for 2D)
+#  <left padding>, (ie LeftPy, LeftPx for 2D)
+#  <right padding>, (ie RightPy, RightPx for 2D)
+
+ ################             op   datatype  layout  verify  init  log  time  Ndims  G   N   K   C  Y  X  Hi  Wi  Sy  Sx  Dy  Dx  LeftPy  LeftPx  RightPy  RightPx
+./bin/ckProfiler image_to_column          0       0       1     1    0     1      2  1 256   1 512  3  3   28  28   1   1   1   1        0       0       0        0
+
+ ```
+
+Result (MI210, FP32, NHWC)
+```
+input: dim 5, lengths {1, 256, 512, 28, 28}, strides {102760448, 401408, 1, 14336, 512}
+output: dim 2, lengths {173056, 4608}, strides {4608, 1}
+....
+Best configuration parameters:
+name: DeviceImageToColumn<128, 32, 64, 4>
+avg_time: 3.12326
+GB/s: 2042.59
+```
diff --git a/profiler/include/profiler/data_type_enum.hpp b/profiler/include/profiler/data_type_enum.hpp
new file mode 100644
index 0000000000..c046c7fabb
--- /dev/null
+++ b/profiler/include/profiler/data_type_enum.hpp
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+namespace ck {
+
+enum struct DataTypeEnum
+{
+    Half     = 0,
+    Float    = 1,
+    Int32    = 2,
+    Int8     = 3,
+    Int8x4   = 4,
+    BFloat16 = 5,
+    Double   = 6,
+    Unknown  = 100,
+};
+
+} // namespace ck
diff --git a/profiler/include/profiler/profile_avg_pool3d_bwd_impl.hpp b/profiler/include/profiler/profile_avg_pool3d_bwd_impl.hpp
new file mode 100644
index 0000000000..e7e8f7213f
--- /dev/null
+++ b/profiler/include/profiler/profile_avg_pool3d_bwd_impl.hpp
@@ -0,0 +1,253 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
+#include "ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename TensorLayout>
+std::vector<ck::index_t> f_tensor_strides_ncdhw(ck::index_t N_,
+                                                ck::index_t C_,
+                                                ck::index_t D,
+                                                ck::index_t H,
+                                                ck::index_t W,
+                                                TensorLayout layout)
+{
+    using namespace ck::literals;
+    (void)N_;
+    if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NDHWC>::value)
+        return {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_};
+    else
+        throw std::runtime_error("not supported yet");
+};
+
+template <typename DOutDataType,
+          typename DInDataType,
+          typename ComputeDataType,
+          typename DOutLayout,
+          typename DInLayout>
+bool profile_avg_pool3d_bwd_impl(int do_verification,
+                                 int init_method,
+                                 bool do_log,
+                                 bool time_kernel,
+                                 std::vector<index_t> in_length, // NCDHW
+                                 std::vector<index_t> window_spatial_lengths,
+                                 std::vector<index_t> window_strides,
+                                 std::vector<index_t> window_dilations,
+                                 std::vector<index_t> input_left_pads,
+                                 std::vector<index_t> input_right_pads)
+{
+    constexpr index_t InOutRank  = 5;
+    constexpr index_t WindowRank = 3;
+
+    if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank ||
+       window_strides.size() != WindowRank || window_dilations.size() != WindowRank ||
+       input_left_pads.size() != WindowRank || input_right_pads.size() != WindowRank)
+    {
+        std::cout << "Parameter is incorrect" << std::endl;
+        return false;
+    }
+
+    std::vector<index_t> out_length(InOutRank);
+
+    int N = in_length[0];
+    int C = in_length[1];
+
+    out_length[0] = N;
+    out_length[1] = C;
+
+    // Calculate Do, Ho, Wo
+    for(int i = 2; i < InOutRank; ++i)
+    {
+        auto pad1             = input_left_pads[i - 2];
+        auto pad2             = input_right_pads[i - 2];
+        auto windows_size     = window_spatial_lengths[i - 2];
+        auto windows_stride   = window_strides[i - 2];
+        auto windows_dilation = window_dilations[i - 2];
+        auto eff              = (windows_size - 1) * windows_dilation + 1;
+        out_length[i]         = (in_length[i] + pad1 + pad2 - eff) / windows_stride + 1;
+    }
+
+    int Di = in_length[2];
+    int Hi = in_length[3];
+    int Wi = in_length[4];
+    int Do = out_length[2];
+    int Ho = out_length[3];
+    int Wo = out_length[4];
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t D, std::size_t H, std::size_t W) {
+            using namespace ck::literals;
+
+            return HostTensorDescriptor({N_, C_, D, H, W},
+                                        {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
+        };
+
+    Tensor<DOutDataType> dout_n_c_do_ho_wo(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+    Tensor<DInDataType> din_n_c_di_hi_wi_device(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
+    Tensor<DInDataType> din_n_c_di_hi_wi_host(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
+
+    switch(init_method)
+    {
+    case 0: dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_1<DOutDataType>{}); break;
+    case 1: dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_2<DOutDataType>{-5, 5}); break;
+    default: dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_3<DOutDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem dout_device_buf(sizeof(DOutDataType) * dout_n_c_do_ho_wo.mDesc.GetElementSpaceSize());
+    DeviceMem din_device_buf(sizeof(DInDataType) *
+                             din_n_c_di_hi_wi_device.mDesc.GetElementSpaceSize());
+
+    dout_device_buf.ToDevice(dout_n_c_do_ho_wo.mData.data());
+
+    using DeviceOp = ck::tensor_operation::device::
+        DeviceAvgPoolBwd<3, DOutDataType, DInDataType, DOutLayout, DInLayout>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using ReferencePoolingBwdInstance =
+            ck::tensor_operation::host::ReferenceAvgPoolBwd<3, DInDataType, DOutDataType>;
+
+        ReferencePoolingBwdInstance ref_pooling_bwd;
+        auto ref_pooling_bwd_argument = ref_pooling_bwd.MakeArgument(din_n_c_di_hi_wi_host,
+                                                                     dout_n_c_do_ho_wo,
+                                                                     window_spatial_lengths,
+                                                                     window_strides,
+                                                                     window_dilations,
+                                                                     input_left_pads,
+                                                                     input_right_pads);
+
+        auto ref_invoker = ref_pooling_bwd.MakeInvoker();
+        ref_invoker.Run(ref_pooling_bwd_argument);
+    }
+
+    int num_kernel = 0;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
+            static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+            {N, C, Do, Ho, Wo},
+            {N, C, Di, Hi, Wi},
+            f_tensor_strides_ncdhw(N, C, Do, Ho, Wo, DOutLayout{}),
+            f_tensor_strides_ncdhw(N, C, Di, Hi, Wi, DInLayout{}),
+            window_spatial_lengths,
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads);
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+        }
+        else
+        {
+            if(time_kernel)
+            {
+                std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
+                LogRange(std::cout << "doutput lengths = ", out_length, ", ") << std::endl;
+            }
+
+            continue;
+        }
+
+        din_device_buf.SetZero();
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+        float avg_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes =
+            dout_n_c_do_ho_wo.mDesc.GetElementSize() * sizeof(DOutDataType) +
+            din_n_c_di_hi_wi_device.mDesc.GetElementSize() * sizeof(DInDataType);
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            din_device_buf.FromDevice(din_n_c_di_hi_wi_device.mData.data());
+            bool pass = ck::utils::check_err(din_n_c_di_hi_wi_device.mData,
+                                             din_n_c_di_hi_wi_host.mData,
+                                             "Error: Incorrect results",
+                                             1e-3,
+                                             1e-3);
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(
+                    std::cout << "din_n_c_di_hi_wi_device: ", din_n_c_di_hi_wi_device.mData, ",")
+                    << std::endl;
+
+                LogRangeAsType<float>(
+                    std::cout << "din_n_c_di_hi_wi_host: ", din_n_c_di_hi_wi_host.mData, ",")
+                    << std::endl;
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "doutput lengths = [", out_length, ", ") << "]." << std::endl;
+                return false;
+            }
+            else
+            {
+                if(time_kernel)
+                    std::cout << "pass" << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", out_length, ",") << std::endl;
+        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp b/profiler/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp
new file mode 100644
index 0000000000..22dab31100
--- /dev/null
+++ b/profiler/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp
@@ -0,0 +1,360 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename A0Layout,
+          typename B0Layout,
+          typename D0sLayout,
+          typename B1Layout,
+          typename D1sLayout,
+          typename E1Layout,
+          typename A0DataType,
+          typename B0DataType,
+          typename D0sDataType,
+          typename B1DataType,
+          typename D1sDataType,
+          typename E1DataType>
+bool profile_batched_gemm_add_relu_gemm_add_impl(bool do_verification,
+                                                 int init_method,
+                                                 bool do_log,
+                                                 bool time_kernel,
+                                                 int M,
+                                                 int N,
+                                                 int K,
+                                                 int O,
+                                                 int BatchCount    = 1,
+                                                 int StrideA0      = -1,
+                                                 int StrideB0      = -1,
+                                                 int StrideD0      = -1,
+                                                 int StrideB1      = -1,
+                                                 int StrideD1      = -1,
+                                                 int StrideE1      = -1,
+                                                 int BatchStrideA0 = -1,
+                                                 int BatchStrideB0 = -1,
+                                                 int BatchStrideD0 = -1,
+                                                 int BatchStrideB1 = -1,
+                                                 int BatchStrideD1 = -1,
+                                                 int BatchStrideE1 = -1)
+
+{
+    using Row = tensor_layout::gemm::RowMajor;
+    using Col = tensor_layout::gemm::ColumnMajor;
+
+    using PassThrough = tensor_operation::element_wise::PassThrough;
+
+    using A0ElementOp   = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using CDE0ElementOp = ck::tensor_operation::element_wise::AddRelu;
+    using B1ElementOp   = PassThrough;
+    using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
+
+    using D0DataType = remove_cvref_t<tuple_element_t<0, D0sDataType>>;
+
+    using D0Layout   = remove_cvref_t<tuple_element_t<0, D0sLayout>>;
+    using D1DataType = remove_cvref_t<tuple_element_t<0, D1sDataType>>;
+    using D1Layout   = remove_cvref_t<tuple_element_t<0, D1sLayout>>;
+
+    // for reference
+    using RefAcc0DataType = float;
+    using RefAcc1DataType = float;
+
+    bool pass = true;
+
+    const int DefaultStrideA0 = ck::is_same_v<A0Layout, Row> ? K : M;
+    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
+    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? O : M;
+    const int DefaultStrideE1 = ck::is_same_v<E1Layout, Row> ? O : M;
+
+    StrideA0 = (StrideA0 < 0) ? DefaultStrideA0 : StrideA0;
+    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideD0 = (StrideD0 < 0) ? DefaultStrideD0 : StrideD0;
+    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideD1 = (StrideD1 < 0) ? DefaultStrideD1 : StrideD1;
+    StrideE1 = (StrideE1 < 0) ? DefaultStrideE1 : StrideE1;
+
+    const int DefaultBatchStrideA0 = (ck::is_same_v<A0Layout, Col> ? K : M) * StrideA0;
+    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideD0 = (ck::is_same_v<D0Layout, Col> ? N : M) * StrideD0;
+    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideD1 = (ck::is_same_v<D1Layout, Col> ? O : M) * StrideD1;
+    const int DefaultBatchStrideE1 = (ck::is_same_v<E1Layout, Col> ? O : M) * StrideE1;
+
+    BatchStrideA0 = BatchStrideA0 < 0 ? DefaultBatchStrideA0 : BatchStrideA0;
+    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideD0 = BatchStrideD0 < 0 ? DefaultBatchStrideD0 : BatchStrideD0;
+    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideD1 = BatchStrideD1 < 0 ? DefaultBatchStrideD1 : BatchStrideD1;
+    BatchStrideE1 = BatchStrideE1 < 0 ? DefaultBatchStrideE1 : BatchStrideE1;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
+        }
+    };
+
+    // E_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<A0DataType> a0_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA0, BatchStrideA0, A0Layout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<D0DataType> d0_g_m_n(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideD0, BatchStrideD0, D0Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<D1DataType> d1_g_m_o(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideD1, BatchStrideD1, D1Layout{}));
+    Tensor<E1DataType> e1_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideE1, BatchStrideE1, E1Layout{}));
+    Tensor<E1DataType> e1_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideE1, BatchStrideE1, E1Layout{}));
+
+    // Host verification: Output of Gemm0 is input A of Gemm1
+    Tensor<RefAcc0DataType> c0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+    Tensor<RefAcc0DataType> e0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+    Tensor<RefAcc1DataType> c1_g_m_o(f_host_tensor_descriptor(BatchCount, M, O, O, M * O, Row{}));
+
+    std::cout << "a0_g_m_k: " << a0_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "d0_g_m_n: " << d0_g_m_n.mDesc << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "d1_g_m_o: " << d1_g_m_o.mDesc << std::endl;
+    std::cout << "e1_g_m_o: " << e1_g_m_o_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_g_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 3});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 3});
+        d0_g_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-2, 3});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 3});
+        d1_g_m_o.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-2, 3});
+        break;
+    default:
+        a0_g_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        d0_g_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        d1_g_m_o.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    DeviceMem a0_g_m_k_device_buf(sizeof(A0DataType) * a0_g_m_k.mDesc.GetElementSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSize());
+    DeviceMem d0_g_m_n_device_buf(sizeof(D0DataType) * d0_g_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSize());
+    DeviceMem d1_g_m_o_device_buf(sizeof(D1DataType) * d1_g_m_o.mDesc.GetElementSpaceSize());
+    DeviceMem e1_g_m_o_device_buf(sizeof(E1DataType) *
+                                  e1_g_m_o_device_result.mDesc.GetElementSize());
+
+    a0_g_m_k_device_buf.ToDevice(a0_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    d0_g_m_n_device_buf.ToDevice(d0_g_m_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+    d1_g_m_o_device_buf.ToDevice(d1_g_m_o.mData.data());
+
+    auto a0_element_op   = A0ElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto cde0_element_op = CDE0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto cde1_element_op = CDE1ElementOp{};
+
+    using DeviceOp =
+        tensor_operation::device::DeviceBatchedGemmMultipleDGemmMultipleD<A0Layout,
+                                                                          B0Layout,
+                                                                          D0sLayout,
+                                                                          B1Layout,
+                                                                          D1sLayout,
+                                                                          E1Layout,
+                                                                          A0DataType,
+                                                                          B0DataType,
+                                                                          D0sDataType,
+                                                                          B1DataType,
+                                                                          D1sDataType,
+                                                                          E1DataType,
+                                                                          A0ElementOp,
+                                                                          B0ElementOp,
+                                                                          CDE0ElementOp,
+                                                                          B1ElementOp,
+                                                                          CDE1ElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    if(do_verification)
+    {
+        // Ref Gemm0
+        using ReferenceGemm0Instance = tensor_operation::host::ReferenceBatchedGemm<A0DataType,
+                                                                                    B0DataType,
+                                                                                    RefAcc0DataType,
+                                                                                    RefAcc0DataType,
+                                                                                    A0ElementOp,
+                                                                                    B0ElementOp,
+                                                                                    PassThrough>;
+
+        // Ref Gemm1
+        using ReferenceGemm1Instance = tensor_operation::host::ReferenceBatchedGemm<RefAcc0DataType,
+                                                                                    B1DataType,
+                                                                                    RefAcc1DataType,
+                                                                                    RefAcc1DataType,
+                                                                                    PassThrough,
+                                                                                    B1ElementOp,
+                                                                                    PassThrough>;
+
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a0_g_m_k, b0_g_k_n, c0_g_m_n, a0_element_op, b0_element_op, PassThrough{});
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        // cde0_elementwise
+        e0_g_m_n.ForEach(
+            [&](auto&, auto idx) { cde0_element_op(e0_g_m_n(idx), c0_g_m_n(idx), d0_g_m_n(idx)); });
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            e0_g_m_n, b1_g_n_o, c1_g_m_o, PassThrough{}, b1_element_op, PassThrough{});
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        // cde1_elementwise
+        e1_g_m_o_host_result.ForEach([&](auto&, auto idx) {
+            cde1_element_op(e1_g_m_o_host_result(idx), c1_g_m_o(idx), d1_g_m_o(idx));
+        });
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<A0DataType*>(a0_g_m_k_device_buf.GetDeviceBuffer()),
+            static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+            std::array<const void*, 1>{d0_g_m_n_device_buf.GetDeviceBuffer()},
+            static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+            std::array<const void*, 1>{d1_g_m_o_device_buf.GetDeviceBuffer()},
+            static_cast<E1DataType*>(e1_g_m_o_device_buf.GetDeviceBuffer()),
+            M,
+            N,
+            K,
+            O,
+            BatchCount,
+            StrideA0,
+            StrideB0,
+            std::array<ck::index_t, 1>{StrideD0},
+            StrideB1,
+            std::array<ck::index_t, 1>{StrideD1},
+            StrideE1,
+            BatchStrideA0,
+            BatchStrideB0,
+            std::array<ck::index_t, 1>{BatchStrideD0},
+            BatchStrideB1,
+            std::array<ck::index_t, 1>{BatchStrideD1},
+            BatchStrideE1,
+            a0_element_op,
+            b0_element_op,
+            cde0_element_op,
+            b1_element_op,
+            cde1_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+            std::size_t num_btype =
+                (sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(D0DataType) * N +
+                 sizeof(B1DataType) * N * O + sizeof(E1DataType) * M * O + sizeof(D1DataType) * O) *
+                BatchCount;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e1_g_m_o_device_buf.FromDevice(e1_g_m_o_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(e1_g_m_o_device_result, e1_g_m_o_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(
+                        std::cout << "e1_g_m_o_host_result : ", e1_g_m_o_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "e1_g_m_o_device_result : ", e1_g_m_o_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp b/profiler/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp
new file mode 100644
index 0000000000..5bee67c1ce
--- /dev/null
+++ b/profiler/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp
@@ -0,0 +1,395 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasesDataType,
+          typename Acc1BiasesDataType,
+          tensor_operation::device::MaskingSpecialization MaskingSpec>
+bool profile_batched_gemm_bias_softmax_gemm_permute_impl(bool do_verification,
+                                                         int init_method,
+                                                         bool do_log,
+                                                         bool time_kernel,
+                                                         int M,
+                                                         int N,
+                                                         int K,
+                                                         int O,
+                                                         int G0,
+                                                         int G1,
+                                                         float alpha = -1.f)
+
+{
+
+    using PassThrough   = tensor_operation::element_wise::PassThrough;
+    using ScaleAdd      = tensor_operation::element_wise::ScaleAdd;
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using C0DEElementOp = ScaleAdd;
+    using Acc0ElementOp = PassThrough;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+    using AccDataType   = float;
+    using D0DataType    = tuple_element_t<0, Acc0BiasesDataType>;
+    using tensor_operation::device::MaskingSpecialization;
+
+    // Ref Gemm0: various type in, fp32 out
+    using ReferenceGemm0Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+    // Ref Softmax: fp32 in, various type out
+    using ReferenceSoftmaxInstance =
+        tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+    // Ref Gemm1: various type in, various type out
+    using ReferenceGemm1Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+    bool pass = true;
+
+    // A layout [G0, M, G1, K]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+    // B0 layout [G0, N, G1, K]
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+    // B1 layout [G0, N, G1, O]
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+    // C layout [G0, M, G1, O]
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+    // D layout [G0, M, G1, N]
+    std::vector<ck::index_t> d0_gs_ms_ns_lengths{G0, G1, M, N};
+    std::vector<ck::index_t> d0_gs_ms_ns_strides{M * G1 * N, N, G1 * N, 1};
+
+    const int BatchCount = G0 * G1;
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+    Tensor<D0DataType> d0_gs_ms_ns(d0_gs_ms_ns_lengths, d0_gs_ms_ns_strides);
+    Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+    Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+    Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
+    std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
+    std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
+
+    std::srand(1); // work around test flakiness
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        // Still unsure whether this kind of deterministic floating point accurary issue is expected
+        // or not. May want to try exact same approach as the GPU kernel in the host reference
+        // GEMM+Softmax+GEMM function to see if the accuracy discrepancy goes away. Until then,
+        // shrink the input value range as it is less likely to produce errors of around ~1e-3.
+        // a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        // b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        // b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-2, 2});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_gs_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) *
+                           c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
+    d0_device_buf.ToDevice(d0_gs_ms_ns.mData.data());
+
+    if(alpha < 0)
+    {
+        alpha = 1.f / std::sqrt(K); // usually 1 / sqrt(head_dim)
+    }
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto c0de_element_op = C0DEElementOp{alpha};
+    auto acc0_element_op = Acc0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    using DeviceOp =
+        tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      ADataType,
+                                                                      B0DataType,
+                                                                      B1DataType,
+                                                                      CDataType,
+                                                                      Acc0BiasesDataType,
+                                                                      ck::Tuple<>,
+                                                                      AElementOp,
+                                                                      B0ElementOp,
+                                                                      C0DEElementOp,
+                                                                      B1ElementOp,
+                                                                      CElementOp,
+                                                                      MaskingSpec>;
+
+    // get device op instances
+    const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+        Tensor<ADataType> a_g_m_k({BatchCount, M, K});
+        Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
+        Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
+        Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N});        // scratch object after gemm0
+        Tensor<ADataType> a1_g_m_n({BatchCount, M, N});            // scratch object after softmax
+        Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
+        Tensor<D0DataType> d0_g_m_n({BatchCount, M, N});
+
+        // permute
+        a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+            a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+        b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+            b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+            b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        d0_gs_ms_ns.ForEach([&](auto& self, auto idx) {
+            d0_g_m_n(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        acc0_g_m_n.ForEach([&](auto&, auto idx) {
+            c0de_element_op(acc0_g_m_n(idx), acc0_g_m_n(idx), d0_g_m_n(idx));
+        });
+        // mask out upper triangle
+        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+            if(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle && idx[1] < idx[2])
+                self(idx) = -ck::NumericLimits<float>::Infinity();
+        });
+
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        // permute
+        c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+            const size_t& g0 = idx[0];
+            const size_t& g1 = idx[1];
+
+            const size_t g = g0 * G1 + g1;
+
+            self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+        });
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+            static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
+            static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
+            static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+            std::array<void*, 1>{
+                d0_device_buf.GetDeviceBuffer()}, // std::array<void*, 1> p_acc0_biases;
+            {},                                   // std::array<void*, 1> p_acc1_biases;
+            a_gs_ms_ks_lengths,
+            a_gs_ms_ks_strides,
+            b0_gs_ns_ks_lengths,
+            b0_gs_ns_ks_strides,
+            b1_gs_os_ns_lengths,
+            b1_gs_os_ns_strides,
+            c_gs_ms_os_lengths,
+            c_gs_ms_os_strides,
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_strides}, // std::array<std::vector<ck::index_t>,
+                                      // 1>{acc0_biases_gs_ms_ns_strides},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_strides},
+            a_element_op,
+            b0_element_op,
+            c0de_element_op,
+            b1_element_op,
+            c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O +
+                                     sizeof(D0DataType) * M * N) *
+                                    BatchCount;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+                // default absolute error and relative error is 0.001
+                double rtol = 1e-3;
+                double atol = 1e-3;
+
+                // when BF16 is taken, set absolute error and relative error to 0.01
+                if(std::is_same_v<ADataType, ck::bhalf_t> &&
+                   std::is_same_v<B0DataType, ck::bhalf_t> &&
+                   std::is_same_v<B1DataType, ck::bhalf_t> &&
+                   std::is_same_v<CDataType, ck::bhalf_t> &&
+                   std::is_same_v<D0DataType, ck::bhalf_t>)
+                {
+                    rtol = 1e-2;
+                    atol = 1e-2;
+                }
+
+                pass = pass & ck::utils::check_err(c_gs_ms_os_device_result,
+                                                   c_gs_ms_os_host_result,
+                                                   "Error: Incorrect results!",
+                                                   rtol,
+                                                   atol);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a_gs_ms_ks: ", a_gs_ms_ks.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b0_gs_ns_ks : ", b0_gs_ns_ks.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b1_gs_os_ns : ", b1_gs_os_ns.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_gs_ms_os_host_result : ", c_gs_ms_os_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_gs_ms_os_device_result : ",
+                                          c_gs_ms_os_device_result.mData,
+                                          ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp b/profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp
new file mode 100644
index 0000000000..f3d2c55617
--- /dev/null
+++ b/profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp
@@ -0,0 +1,319 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename ALayout,
+          typename B0Layout,
+          typename B1Layout,
+          typename CLayout>
+bool profile_batched_gemm_gemm_impl(bool do_verification,
+                                    int init_method,
+                                    bool do_log,
+                                    bool time_kernel,
+                                    int M,
+                                    int N,
+                                    int K,
+                                    int O,
+                                    int BatchCount    = 1,
+                                    int StrideA       = -1,
+                                    int StrideB0      = -1,
+                                    int StrideB1      = -1,
+                                    int StrideC       = -1,
+                                    int BatchStrideA  = -1,
+                                    int BatchStrideB0 = -1,
+                                    int BatchStrideB1 = -1,
+                                    int BatchStrideC  = -1)
+
+{
+
+    using Row           = tensor_layout::gemm::RowMajor;
+    using Col           = tensor_layout::gemm::ColumnMajor;
+    using PassThrough   = tensor_operation::element_wise::PassThrough;
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using B1ElementOp   = PassThrough;
+    using Acc0ElementOp = PassThrough;
+    using CElementOp    = PassThrough;
+    using AccDataType   = float;
+
+    // Ref Gemm0
+    using ReferenceGemm0Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                ADataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                CElementOp>;
+
+    // Ref Gemm
+    using ReferenceGemm1Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+    bool pass = true;
+
+    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideC  = ck::is_same_v<CLayout, Row> ? O : M;
+
+    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
+    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideC  = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
+    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideC  = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
+
+    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
+    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideC  = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
+        }
+    };
+
+    // C_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<CDataType> c_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    Tensor<CDataType> c_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    // Host verification: Output of Gemm0 is input A of Gemm1
+    Tensor<ADataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 3});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 3});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 3});
+        break;
+    case 2:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSize());
+    DeviceMem c_g_m_o_device_buf(sizeof(CDataType) * c_g_m_o_device_result.mDesc.GetElementSize());
+
+    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    using DeviceOp = tensor_operation::device::DeviceBatchedGemmGemm<ALayout,
+                                                                     B0Layout,
+                                                                     B1Layout,
+                                                                     CLayout,
+                                                                     ADataType,
+                                                                     B0DataType,
+                                                                     B1DataType,
+                                                                     CDataType,
+                                                                     AElementOp,
+                                                                     B0ElementOp,
+                                                                     Acc0ElementOp,
+                                                                     B1ElementOp,
+                                                                     CElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // early fail when no instances are found
+    if(op_ptrs.size() == 0)
+    {
+        return false;
+    }
+
+    if(do_verification)
+    {
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, PassThrough{});
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            acc0_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
+            static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+            static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+            static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
+            M,
+            N,
+            K,
+            O,
+            BatchCount,
+            StrideA,
+            StrideB0,
+            StrideB1,
+            StrideC,
+            BatchStrideA,
+            BatchStrideB0,
+            BatchStrideB1,
+            BatchStrideC,
+            a_element_op,
+            b0_element_op,
+            acc0_element_op,
+            b1_element_op,
+            c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                                    BatchCount;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(c_g_m_o_device_result, c_g_m_o_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a_g_m_k: ", a_g_m_k.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b0_g_k_n : ", b0_g_k_n.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b1_g_n_o : ", b1_g_n_o.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_g_m_o_host_result : ", c_g_m_o_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_g_m_o_device_result : ", c_g_m_o_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_batched_gemm_impl.hpp b/profiler/include/profiler/profile_batched_gemm_impl.hpp
new file mode 100644
index 0000000000..936c22f5d8
--- /dev/null
+++ b/profiler/include/profiler/profile_batched_gemm_impl.hpp
@@ -0,0 +1,264 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementOp,
+          typename BElementOp,
+          typename CElementOp,
+          typename DeviceOp>
+bool profile_batched_gemm_impl(int do_verification,
+                               int init_method,
+                               bool do_log,
+                               bool time_kernel,
+                               int M,
+                               int N,
+                               int K,
+                               int BatchStrideA,
+                               int BatchStrideB,
+                               int BatchStrideC,
+                               int StrideA,
+                               int StrideB,
+                               int StrideC,
+                               int BatchCount)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
+        }
+    };
+
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
+    Tensor<BDataType> b_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB, BatchStrideB, BLayout{}));
+    Tensor<CDataType> c_g_m_n_host_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, BatchStrideC, CLayout{}));
+    Tensor<CDataType> c_g_m_n_device_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, BatchStrideC, CLayout{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
+    std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    if(do_verification)
+    {
+        using ReferenceBatchedGemmInstance =
+            ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                             BDataType,
+                                                             CDataType,
+                                                             float,
+                                                             AElementOp,
+                                                             BElementOp,
+                                                             CElementOp>;
+
+        auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
+        auto ref_invoker      = ref_batched_gemm.MakeInvoker();
+
+        auto ref_argument = ref_batched_gemm.MakeArgument(
+            a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_g_m_k.mData.data());
+    b_device_buf.ToDevice(b_g_k_n.mData.data());
+    c_device_buf.ToDevice(c_g_m_n_device_result.mData.data());
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        std::unique_ptr<tensor_operation::device::BaseArgument> argument_ptr;
+        // false branch for multi d dl kernel
+        if constexpr(std::is_same<
+                         DeviceOp,
+                         ck::tensor_operation::device::DeviceBatchedGemm<ALayout,
+                                                                         BLayout,
+                                                                         CLayout,
+                                                                         ADataType,
+                                                                         BDataType,
+                                                                         CDataType,
+                                                                         AElementOp,
+                                                                         BElementOp,
+                                                                         CElementOp>>::value)
+        {
+
+            argument_ptr =
+                op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                            static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                            static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                            M,
+                                            N,
+                                            K,
+                                            StrideA,
+                                            StrideB,
+                                            StrideC,
+                                            BatchStrideA,
+                                            BatchStrideB,
+                                            BatchStrideC,
+                                            BatchCount,
+                                            ck::tensor_operation::element_wise::PassThrough{},
+                                            ck::tensor_operation::element_wise::PassThrough{},
+                                            ck::tensor_operation::element_wise::PassThrough{});
+        }
+        else
+        {
+            argument_ptr =
+                op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                            static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                            {},
+                                            static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                            M,
+                                            N,
+                                            K,
+                                            BatchCount,
+                                            StrideA,
+                                            StrideB,
+                                            {},
+                                            StrideC,
+                                            BatchStrideA,
+                                            BatchStrideB,
+                                            {},
+                                            BatchStrideC,
+                                            ck::tensor_operation::element_wise::PassThrough{},
+                                            ck::tensor_operation::element_wise::PassThrough{},
+                                            ck::tensor_operation::element_wise::PassThrough{});
+        }
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init C to zero before profiling next kernel
+            c_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * BatchCount * M * N * K;
+
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                     sizeof(CDataType) * M * N) *
+                                    BatchCount;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(c_g_m_n_device_result, c_g_m_n_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_g_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_g_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host: ", c_g_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_device: ", c_g_m_n_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_batched_gemm_reduce_impl.hpp b/profiler/include/profiler/profile_batched_gemm_reduce_impl.hpp
new file mode 100644
index 0000000000..901fa338d4
--- /dev/null
+++ b/profiler/include/profiler/profile_batched_gemm_reduce_impl.hpp
@@ -0,0 +1,362 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32                 = float;
+using F16                 = ck::half_t;
+using ReducePtrsGlobal    = ck::Tuple<F32*, F32*>;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Identity, Identity>;
+
+using DeviceGemmReduceNoOpPtr =
+    ck::tensor_operation::device::DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>;
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ReduceDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_batched_gemm_reduce_impl(int do_verification,
+                                      int init_method,
+                                      bool do_log,
+                                      bool time_kernel,
+                                      int M,
+                                      int N,
+                                      int K,
+                                      int StrideA,
+                                      int StrideB,
+                                      int StrideC,
+                                      int BatchCount)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {row * stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {col * stride, 1_uz, stride});
+        }
+    };
+
+    Tensor<ADataType> a_g_m_k(f_host_tensor_descriptor(BatchCount, M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_g_k_n(f_host_tensor_descriptor(BatchCount, K, N, StrideB, BLayout{}));
+
+    Tensor<CDataType> c_g_m_n_host_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> d0_g_m_host_result({BatchCount, M});
+    Tensor<ReduceDataType> d1_g_m_host_result({BatchCount, M});
+
+    Tensor<CDataType> c_g_m_n_device_result(
+        f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> d0_g_m_device_result({BatchCount, M});
+    Tensor<ReduceDataType> d1_g_m_device_result({BatchCount, M});
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
+    std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
+    std::cout << "d0_g_m: " << d0_g_m_host_result.mDesc << std::endl;
+    std::cout << "d1_g_m: " << d1_g_m_host_result.mDesc << std::endl;
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        std::srand(0);
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        break;
+    default:
+        std::srand(0);
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+    }
+
+    using AElementOp            = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp            = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp            = ck::tensor_operation::element_wise::PassThrough;
+    using ReduceOp0             = ck::reduce::Add;
+    using ReduceOp1             = ck::reduce::Add;
+    using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using UnarySquareElementOp  = ck::tensor_operation::element_wise::UnarySquare;
+
+    auto a_element_op                     = AElementOp{};
+    auto b_element_op                     = BElementOp{};
+    auto c_element_op                     = CElementOp{};
+    std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
+
+    const auto reduce0_op = ReduceOp0{};
+    const auto reduce1_op = ReduceOp1{};
+
+    auto passthrough                            = UnaryIdenticElementOp{};
+    auto square                                 = UnarySquareElementOp{};
+    std::array<void*, 2> reduce_in_element_ops  = {&passthrough, &square};
+    std::array<void*, 2> reduce_out_element_ops = {&passthrough, &passthrough};
+
+    if(do_verification)
+    {
+        using ReferenceBatchedGemmInstance =
+            ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                             BDataType,
+                                                             CDataType,
+                                                             float,
+                                                             AElementOp,
+                                                             BElementOp,
+                                                             CElementOp>;
+
+        using ReduceAccDataType = ReduceDataType;
+
+        auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
+        auto ref_invoker      = ref_batched_gemm.MakeInvoker();
+
+        auto ref_argument = ref_batched_gemm.MakeArgument(
+            a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        for(int batch = 0; batch < BatchCount; ++batch)
+        {
+            for(int m = 0; m < M; ++m)
+            {
+                auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+                auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
+
+                for(int n = 0; n < N; ++n)
+                {
+                    ReduceAccDataType d0_val =
+                        ck::type_convert<ReduceAccDataType>(c_g_m_n_host_result(batch, m, n));
+                    ReduceAccDataType d1_val;
+
+                    square(d1_val, d0_val);
+                    reduce0_op(reduce0_acc, d0_val);
+                    reduce1_op(reduce1_acc, d1_val);
+                }
+
+                d0_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce0_acc);
+                d1_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce1_acc);
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
+                                 d0_g_m_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
+                                 d1_g_m_device_result.mDesc.GetElementSpaceSize());
+
+    std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
+                                      reduce1_device_buf.GetDeviceBuffer()};
+
+    a_device_buf.ToDevice(a_g_m_k.mData.data());
+    b_device_buf.ToDevice(b_g_k_n.mData.data());
+
+    // add device GEMM instances
+    std::vector<ck::tensor_operation::device::instance::DeviceGemmReduceNoOpPtr> gemm_ptrs;
+
+    if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
+                 is_same<CDataType, half_t>::value)
+    {
+        if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
+                    gemm_ptrs);
+        }
+    }
+
+    if(gemm_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device GEMM instance found");
+    }
+
+    std::string best_gemm_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device GEMM instances
+    for(auto& gemm_ptr : gemm_ptrs)
+    {
+        auto argument_ptr = gemm_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                          b_device_buf.GetDeviceBuffer(),
+                                                          nullptr,
+                                                          {},
+                                                          c_device_buf.GetDeviceBuffer(),
+                                                          p_reduces,
+                                                          M,
+                                                          N,
+                                                          K,
+                                                          StrideA,
+                                                          StrideB,
+                                                          StrideC,
+                                                          {},
+                                                          gemm_element_ops,
+                                                          {},
+                                                          reduce_in_element_ops,
+                                                          reduce_out_element_ops,
+                                                          BatchCount);
+
+        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
+
+        if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // init DO, D1 to 0
+            reduce0_device_buf.SetZero();
+            reduce1_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::string gemm_name = gemm_ptr->GetTypeString();
+
+            std::size_t flop      = std::size_t(2) * BatchCount * M * N * K;
+            std::size_t num_btype = sizeof(ADataType) * BatchCount * M * K +
+                                    sizeof(BDataType) * BatchCount * K * N +
+                                    sizeof(CDataType) * BatchCount * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << gemm_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_gemm_name  = gemm_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
+                reduce0_device_buf.FromDevice(d0_g_m_device_result.mData.data());
+                reduce1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
+
+                bool c_error  = ck::utils::check_err(c_g_m_n_device_result, c_g_m_n_host_result);
+                bool d0_error = ck::utils::check_err(d0_g_m_device_result, d0_g_m_host_result);
+                bool d1_error = ck::utils::check_err(d1_g_m_device_result, d1_g_m_host_result);
+
+                pass = pass && (c_error == true);
+                pass = pass && (d0_error == true);
+                pass = pass && (d1_error == true);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_g_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_g_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host: ", c_g_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_device: ", c_g_m_n_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "d0_host: ", d0_g_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d0_device: ", d0_g_m_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "d1_host: ", d1_g_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d1_device: ", d1_g_m_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << "does not support this GEMM problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp b/profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp
new file mode 100644
index 0000000000..15a21206c5
--- /dev/null
+++ b/profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp
@@ -0,0 +1,347 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename ALayout,
+          typename B0Layout,
+          typename B1Layout,
+          typename CLayout,
+          bool MaskOutUpperTriangle>
+bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
+                                            int init_method,
+                                            bool do_log,
+                                            bool time_kernel,
+                                            int M,
+                                            int N,
+                                            int K,
+                                            int O,
+                                            int BatchCount    = 1,
+                                            int StrideA       = -1,
+                                            int StrideB0      = -1,
+                                            int StrideB1      = -1,
+                                            int StrideC       = -1,
+                                            int BatchStrideA  = -1,
+                                            int BatchStrideB0 = -1,
+                                            int BatchStrideB1 = -1,
+                                            int BatchStrideC  = -1,
+                                            float alpha       = -1.f)
+
+{
+
+    using Row           = tensor_layout::gemm::RowMajor;
+    using Col           = tensor_layout::gemm::ColumnMajor;
+    using PassThrough   = tensor_operation::element_wise::PassThrough;
+    using Scale         = tensor_operation::element_wise::Scale;
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = Scale;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+    using AccDataType   = float;
+
+    // Ref Gemm0: various type in, fp32 out
+    using ReferenceGemm0Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+    // Ref Softmax: fp32 in, various type out
+    using ReferenceSoftmaxInstance =
+        tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+    // Ref Gemm1: various type in, various type out
+    using ReferenceGemm1Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+    bool pass = true;
+
+    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
+    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
+    const int DefaultStrideC  = ck::is_same_v<CLayout, Row> ? O : M;
+
+    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
+    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
+    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    StrideC  = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
+    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
+    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    const int DefaultBatchStrideC  = (ck::is_same_v<CLayout, Col> ? O : M) * StrideC;
+
+    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
+    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
+    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    BatchStrideC  = BatchStrideC < 0 ? DefaultBatchStrideC : BatchStrideC;
+
+    auto f_host_tensor_descriptor = [](std::size_t batch_count,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        using namespace ck::literals;
+
+        if(std::is_same<decltype(layout), Row>::value)
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
+        }
+        else
+        {
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
+        }
+    };
+
+    // C_m_o = A_m_k * B0_k_n * B1_n_o
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
+    Tensor<B0DataType> b0_g_k_n(
+        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
+    Tensor<B1DataType> b1_g_n_o(
+        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
+    Tensor<CDataType> c_g_m_o_host_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    Tensor<CDataType> c_g_m_o_device_result(
+        f_host_tensor_descriptor(BatchCount, M, O, StrideC, BatchStrideC, CLayout{}));
+    // Host verification: Output of Gemm0 is input A of Gemm1
+    Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+    Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
+    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    std::cout << "c_g_m_o: " << c_g_m_o_host_result.mDesc << std::endl;
+
+    std::srand(1); // work around test flakiness
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        // Still unsure whether this kind of deterministic floating point accurary issue is expected
+        // or not. May want to try exact same approach as the GPU kernel in the host reference
+        // GEMM+Softmax+GEMM function to see if the accuracy discrepancy goes away. Until then,
+        // shrink the input value range as it is less likely to produce errors of around ~1e-3.
+        // a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        // b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        // b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+        break;
+    case 2:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSize());
+    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSize());
+    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSize());
+    DeviceMem c_g_m_o_device_buf(sizeof(CDataType) * c_g_m_o_device_result.mDesc.GetElementSize());
+
+    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
+    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
+    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+
+    if(alpha < 0)
+    {
+        alpha = 1.f / std::sqrt(K); // usually 1 / sqrt(head_dim)
+    }
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    using DeviceOp = tensor_operation::device::DeviceBatchedGemmSoftmaxGemm<ALayout,
+                                                                            B0Layout,
+                                                                            B1Layout,
+                                                                            CLayout,
+                                                                            ADataType,
+                                                                            B0DataType,
+                                                                            B1DataType,
+                                                                            CDataType,
+                                                                            AElementOp,
+                                                                            B0ElementOp,
+                                                                            Acc0ElementOp,
+                                                                            B1ElementOp,
+                                                                            CElementOp,
+                                                                            MaskOutUpperTriangle>;
+
+    // get device op instances
+    const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    if(do_verification)
+    {
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, Scale{alpha});
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        // mask out upper triangle
+        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+            if(MaskOutUpperTriangle && idx[1] < idx[2])
+                self(idx) = -ck::NumericLimits<float>::Infinity();
+        });
+
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
+            static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
+            static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
+            static_cast<CDataType*>(c_g_m_o_device_buf.GetDeviceBuffer()),
+            M,
+            N,
+            K,
+            O,
+            BatchCount,
+            StrideA,
+            StrideB0,
+            StrideB1,
+            StrideC,
+            BatchStrideA,
+            BatchStrideB0,
+            BatchStrideB1,
+            BatchStrideC,
+            a_element_op,
+            b0_element_op,
+            acc0_element_op,
+            b1_element_op,
+            c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                                    BatchCount;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(c_g_m_o_device_result, c_g_m_o_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a_g_m_k: ", a_g_m_k.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b0_g_k_n : ", b0_g_k_n.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b1_g_n_o : ", b1_g_n_o.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_g_m_o_host_result : ", c_g_m_o_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_g_m_o_device_result : ", c_g_m_o_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp b/profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp
new file mode 100644
index 0000000000..f2fcb0b133
--- /dev/null
+++ b/profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp
@@ -0,0 +1,367 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasesDataType,
+          typename Acc1BiasesDataType,
+          tensor_operation::device::MaskingSpecialization MaskingSpec>
+bool profile_batched_gemm_softmax_gemm_permute_impl(bool do_verification,
+                                                    int init_method,
+                                                    bool do_log,
+                                                    bool time_kernel,
+                                                    int M,
+                                                    int N,
+                                                    int K,
+                                                    int O,
+                                                    int G0,
+                                                    int G1,
+                                                    float alpha = -1.f)
+
+{
+
+    using PassThrough   = tensor_operation::element_wise::PassThrough;
+    using Scale         = tensor_operation::element_wise::Scale;
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = Scale;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+    using AccDataType   = float;
+    using tensor_operation::device::MaskingSpecialization;
+
+    // Ref Gemm0: various type in, fp32 out
+    using ReferenceGemm0Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+    // Ref Softmax: fp32 in, various type out
+    using ReferenceSoftmaxInstance =
+        tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+    // Ref Gemm1: various type in, various type out
+    using ReferenceGemm1Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+
+    bool pass = true;
+
+    // A layout [G0, M, G1, K]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+    // B0 layout [G0, N, G1, K]
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+    // B1 layout [G0, N, G1, O]
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+    // C layout [G0, M, G1, O]
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+    const int BatchCount = G0 * G1;
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+    Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+    Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+    Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
+    std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
+    std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
+
+    std::srand(1); // work around test flakiness
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        // Still unsure whether this kind of deterministic floating point accurary issue is expected
+        // or not. May want to try exact same approach as the GPU kernel in the host reference
+        // GEMM+Softmax+GEMM function to see if the accuracy discrepancy goes away. Until then,
+        // shrink the input value range as it is less likely to produce errors of around ~1e-3.
+        // a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        // b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        // b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        break;
+    case 3:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) *
+                           c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
+
+    if(alpha < 0)
+    {
+        alpha = 1.f / std::sqrt(K); // usually 1 / sqrt(head_dim)
+    }
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    using DeviceOp = tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                                   1,
+                                                                                   1,
+                                                                                   1,
+                                                                                   1,
+                                                                                   ADataType,
+                                                                                   B0DataType,
+                                                                                   B1DataType,
+                                                                                   CDataType,
+                                                                                   ck::Tuple<>,
+                                                                                   ck::Tuple<>,
+                                                                                   AElementOp,
+                                                                                   B0ElementOp,
+                                                                                   Acc0ElementOp,
+                                                                                   B1ElementOp,
+                                                                                   CElementOp,
+                                                                                   MaskingSpec>;
+
+    // get device op instances
+    const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+        Tensor<ADataType> a_g_m_k({BatchCount, M, K});
+        Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
+        Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
+        Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N});        // scratch object after gemm0
+        Tensor<ADataType> a1_g_m_n({BatchCount, M, N});            // scratch object after softmax
+        Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
+
+        // permute
+        a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+            a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+        b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+            b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+            b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, Scale{alpha});
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        // mask out upper triangle
+        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+            if(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle && idx[1] < idx[2])
+                self(idx) = -ck::NumericLimits<float>::Infinity();
+        });
+
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        // permute
+        c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+            const size_t& g0 = idx[0];
+            const size_t& g1 = idx[1];
+
+            const size_t g = g0 * G1 + g1;
+
+            self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+        });
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+            static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
+            static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
+            static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+            {}, // std::array<void*, 1> p_acc0_biases;
+            {}, // std::array<void*, 1> p_acc1_biases;
+            a_gs_ms_ks_lengths,
+            a_gs_ms_ks_strides,
+            b0_gs_ns_ks_lengths,
+            b0_gs_ns_ks_strides,
+            b1_gs_os_ns_lengths,
+            b1_gs_os_ns_strides,
+            c_gs_ms_os_lengths,
+            c_gs_ms_os_strides,
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_lengths},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_strides},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_strides},
+            a_element_op,
+            b0_element_op,
+            acc0_element_op,
+            b1_element_op,
+            c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
+                                    BatchCount;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+                // default absolute error and relative error is 0.001
+                double rtol = 1e-3;
+                double atol = 1e-3;
+
+                // when BF16 is taken, set absolute error and relative error to 0.01
+                if(std::is_same_v<ADataType, ck::bhalf_t> &&
+                   std::is_same_v<B0DataType, ck::bhalf_t> &&
+                   std::is_same_v<B1DataType, ck::bhalf_t> &&
+                   std::is_same_v<CDataType, ck::bhalf_t>)
+                {
+                    rtol = 1e-2;
+                    atol = 1e-2;
+                }
+
+                pass = pass & ck::utils::check_err(c_gs_ms_os_device_result,
+                                                   c_gs_ms_os_host_result,
+                                                   "Error: Incorrect results!",
+                                                   rtol,
+                                                   atol);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a_gs_ms_ks: ", a_gs_ms_ks.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b0_gs_ns_ks : ", b0_gs_ns_ks.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "b1_gs_os_ns : ", b1_gs_os_ns.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "c_gs_ms_os_host_result : ", c_gs_ms_os_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_gs_ms_os_device_result : ",
+                                          c_gs_ms_os_device_result.mData,
+                                          ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_batchnorm_backward_impl.hpp b/profiler/include/profiler/profile_batchnorm_backward_impl.hpp
new file mode 100644
index 0000000000..3343b5e66e
--- /dev/null
+++ b/profiler/include/profiler/profile_batchnorm_backward_impl.hpp
@@ -0,0 +1,390 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <stdexcept>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename XDataType,
+          typename DxDataType,
+          typename DyDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename DscaleDbiasDataType,
+          typename MeanVarDataType,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+bool profile_batchnorm_backward_impl(bool do_verification,
+                                     int init_method,
+                                     bool do_dumpout,
+                                     bool time_kernel,
+                                     const std::vector<size_t> inOutLengths,
+                                     const std::vector<int> reduceDims,
+                                     bool haveSavedMeanInvVar,
+                                     double epsilon)
+{
+    if(inOutLengths.size() != Rank || reduceDims.size() != NumBatchNormReduceDim)
+    {
+        throw std::runtime_error("Invalid tensor lengths or number of reduce dimensions!");
+    };
+
+    std::vector<size_t> scaleBiasMeanVarLengths;
+
+    // used for calculating the effective transferred bytes by each operation
+    size_t total_length;
+    size_t invariant_length = 1;
+
+    total_length =
+        std::accumulate(inOutLengths.begin(), inOutLengths.end(), 1, std::multiplies<size_t>{});
+
+    if(std::any_of(reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+        throw std::runtime_error("Invalid reduce dimensions!");
+
+    for(int dim = 0; dim < Rank; dim++)
+    {
+        if(std::none_of(reduceDims.begin(), reduceDims.end(), [&](int d) { return dim == d; }))
+        {
+            scaleBiasMeanVarLengths.push_back(inOutLengths[dim]);
+            invariant_length *= inOutLengths[dim];
+        };
+    }
+
+    // input data of the batchnorm backward algorithm
+    Tensor<XDataType> x(inOutLengths);
+    Tensor<DyDataType> dy(inOutLengths);
+    Tensor<ScaleDataType> bnScale(scaleBiasMeanVarLengths);
+
+    Tensor<MeanVarDataType> savedMean(scaleBiasMeanVarLengths);
+    Tensor<MeanVarDataType> savedInvVar(scaleBiasMeanVarLengths);
+    // savedVariance is only used for initializing savedInvVar
+    Tensor<MeanVarDataType> savedVariance(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm backward algorithm
+    Tensor<DxDataType> dx_ref(inOutLengths);
+    Tensor<DxDataType> dx(inOutLengths);
+
+    Tensor<DscaleDbiasDataType> dscale(scaleBiasMeanVarLengths);
+    Tensor<DscaleDbiasDataType> dbias(scaleBiasMeanVarLengths);
+
+    Tensor<DscaleDbiasDataType> dscale_ref(scaleBiasMeanVarLengths);
+    Tensor<DscaleDbiasDataType> dbias_ref(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = x.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if(haveSavedMeanInvVar)
+    {
+        const float x_mean       = 0.0f;
+        const float x_stddev     = 1.0f;
+        const float noise_stddev = 0.0001f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+
+        // initialize the savedMean to be values with tiny variation to the mean of the x values
+        savedMean.GenerateTensorValue(GeneratorTensor_4<MeanVarDataType>{x_mean, noise_stddev},
+                                      num_thread);
+
+        // initialize the variance to be values with tiny variation to the variance of the x values
+        savedVariance.GenerateTensorValue(
+            GeneratorTensor_4<MeanVarDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+
+        auto it_src       = savedVariance.mData.begin();
+        auto it_dst       = savedInvVar.mData.begin();
+        float tmp_epsilon = std::numeric_limits<float>::epsilon();
+
+        while(it_src != savedVariance.mData.end())
+        {
+            *it_dst = type_convert<AccDataType>(
+                1.0f / std::sqrtf(type_convert<float>(*it_src) + tmp_epsilon));
+
+            it_src++;
+            it_dst++;
+        };
+    }
+    else
+    {
+        const float x_mean   = 0.0f;
+        const float x_stddev = 1.0f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            dy.GenerateTensorValue(GeneratorTensor_0<DyDataType>{}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_0<ScaleDataType>{}, num_thread);
+            break;
+        case 1:
+            dy.GenerateTensorValue(GeneratorTensor_1<DyDataType>{1}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_1<ScaleDataType>{1}, num_thread);
+            break;
+        case 2:
+            dy.GenerateTensorValue(GeneratorTensor_2<DyDataType>{-2, 2}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_2<ScaleDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            dy.GenerateTensorValue(GeneratorTensor_3<DyDataType>{-0.2f, 0.2f}, num_thread);
+            bnScale.GenerateTensorValue(GeneratorTensor_3<ScaleDataType>{-0.5f, 0.5f}, num_thread);
+        }
+    };
+
+    // input data of the batchnorm backward algorithm
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem dy_dev(sizeof(DyDataType) * dy.mDesc.GetElementSpaceSize());
+
+    DeviceMem bnScale_dev(sizeof(ScaleDataType) * bnScale.mDesc.GetElementSpaceSize());
+
+    DeviceMem savedMean_dev(sizeof(MeanVarDataType) * savedMean.mDesc.GetElementSpaceSize());
+    DeviceMem savedInvVar_dev(sizeof(MeanVarDataType) * savedInvVar.mDesc.GetElementSpaceSize());
+
+    // output data of the batchnorm backward algorithm
+    DeviceMem dx_dev(sizeof(DxDataType) * dx.mDesc.GetElementSpaceSize());
+
+    DeviceMem dscale_dev(sizeof(DscaleDbiasDataType) * dscale.mDesc.GetElementSpaceSize());
+    DeviceMem dbias_dev(sizeof(DscaleDbiasDataType) * dbias.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    dy_dev.ToDevice(dy.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+
+    if(haveSavedMeanInvVar)
+    {
+        savedMean_dev.ToDevice(savedMean.mData.data());
+        savedInvVar_dev.ToDevice(savedInvVar.mData.data());
+    };
+
+    std::array<index_t, Rank> arrInOutLengths;
+    std::array<index_t, Rank> arrInOutStrides;
+    std::array<index_t, Rank - NumBatchNormReduceDim> arrScaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumBatchNormReduceDim> arrScaleBiasMeanVarStrides;
+    std::array<int, NumBatchNormReduceDim> arrReduceDims;
+
+    std::copy(inOutLengths.begin(), inOutLengths.end(), arrInOutLengths.begin());
+    std::copy(inOutStrides.begin(), inOutStrides.end(), arrInOutStrides.begin());
+    std::copy(scaleBiasMeanVarLengths.begin(),
+              scaleBiasMeanVarLengths.end(),
+              arrScaleBiasMeanVarLengths.begin());
+    std::copy(scaleBiasMeanVarStrides.begin(),
+              scaleBiasMeanVarStrides.end(),
+              arrScaleBiasMeanVarStrides.begin());
+
+    std::copy(reduceDims.begin(), reduceDims.end(), arrReduceDims.begin());
+
+    using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+    // add device batchnorm-backward instances
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormBwd<XDataType,
+                                                                      DxDataType,
+                                                                      DxDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      DscaleDbiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThroughOp,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using ReferenceBatchNormBwdInstance =
+            ck::tensor_operation::host::ReferenceBatchNormBwd<XDataType,
+                                                              DxDataType,
+                                                              DyDataType,
+                                                              AccDataType,
+                                                              ScaleDataType,
+                                                              DscaleDbiasDataType,
+                                                              MeanVarDataType,
+                                                              PassThroughOp,
+                                                              Rank,
+                                                              NumBatchNormReduceDim>;
+
+        auto batchNormBwd_ref = ReferenceBatchNormBwdInstance{};
+
+        auto argument_ptr_ref = batchNormBwd_ref.MakeArgumentPointer(
+            arrInOutLengths,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrReduceDims,
+            arrScaleBiasMeanVarLengths,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            x.mData.data(),
+            dy.mData.data(),
+            bnScale.mData.data(),
+            haveSavedMeanInvVar ? savedMean.mData.data() : nullptr,
+            haveSavedMeanInvVar ? savedInvVar.mData.data() : nullptr,
+            epsilon,
+            PassThroughOp{},
+            dx_ref.mData.data(),
+            dscale_ref.mData.data(),
+            dbias_ref.mData.data());
+
+        if(!batchNormBwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters not supported by the reference instance, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = batchNormBwd_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+    }
+
+    int num_kernel = 0;
+    bool pass      = true;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            arrInOutLengths,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrReduceDims,
+            arrScaleBiasMeanVarLengths,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            x_dev.GetDeviceBuffer(),
+            dy_dev.GetDeviceBuffer(),
+            bnScale_dev.GetDeviceBuffer(),
+            haveSavedMeanInvVar ? savedMean_dev.GetDeviceBuffer() : nullptr,
+            haveSavedMeanInvVar ? savedInvVar_dev.GetDeviceBuffer() : nullptr,
+            epsilon,
+            PassThroughOp{},
+            dx_dev.GetDeviceBuffer(),
+            dscale_dev.GetDeviceBuffer(),
+            dbias_dev.GetDeviceBuffer());
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            num_kernel++;
+        }
+        else
+        {
+            if(time_kernel)
+            {
+                std::cout << inst_ptr->GetTypeString()
+                          << " skipped due to unsupported argument: " << std::endl;
+            }
+
+            continue;
+        };
+
+        size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+        DeviceMem workspace_dev(workspace_sz);
+
+        inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        size_t num_bytes = 0;
+
+        // inputing of x, dy, scale, outputing of dx, dscale, dbias
+        num_bytes += total_length * (sizeof(XDataType) + sizeof(DyDataType) + sizeof(DxDataType)) +
+                     invariant_length * sizeof(DscaleDbiasDataType) * 2;
+
+        // inputting of savedMean, savedInvVariance
+        if(haveSavedMeanInvVar)
+            num_bytes += invariant_length * sizeof(MeanVarDataType) * 2;
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            using ck::utils::check_err;
+            bool single_pass = true;
+
+            dx_dev.FromDevice(dx.mData.data());
+            dscale_dev.FromDevice(dscale.data());
+            dbias_dev.FromDevice(dbias.data());
+
+            // clang-format off
+            single_pass = single_pass && ck::utils::check_err(dx.mData, dx_ref.mData, "dx result:", 5e-4, 5e-4);
+            single_pass = single_pass && ck::utils::check_err(dscale.mData, dscale_ref.mData, "dScale result:", 3e-3, 3e-3);
+            single_pass = single_pass && ck::utils::check_err(dbias.mData, dbias_ref.mData, "dBias result:", 3e-3, 3e-3);
+            // clang-format on
+
+            pass = pass && single_pass;
+        };
+
+        if(do_dumpout)
+        {
+            using ck::host_common::dumpBufferToFile;
+
+            // clang-format off
+            dumpBufferToFile("dump_x.bin", x.mData.data(), x.mDesc.GetElementSize());
+            dumpBufferToFile("dump_dy.bin", dy.mData.data(), dy.mDesc.GetElementSize());
+            dumpBufferToFile("dump_dx.bin", dx.mData.data(), dx.mDesc.GetElementSize());
+            dumpBufferToFile("dump_dx_ref.bin", dx_ref.mData.data(), dx_ref.mDesc.GetElementSize());
+            dumpBufferToFile("dump_dscale.bin", dscale.mData.data(), dscale.mDesc.GetElementSize());
+            dumpBufferToFile("dump_dscale_ref.bin", dscale_ref.mData.data(), dscale_ref.mDesc.GetElementSize());
+            // clang-format off
+        };
+    }
+
+    if(time_kernel)
+    {
+        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_batchnorm_forward_impl.hpp b/profiler/include/profiler/profile_batchnorm_forward_impl.hpp
new file mode 100644
index 0000000000..2f9538b16c
--- /dev/null
+++ b/profiler/include/profiler/profile_batchnorm_forward_impl.hpp
@@ -0,0 +1,412 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <stdexcept>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+bool profile_batchnorm_forward_impl(int do_verification,
+                                    int init_method,
+                                    bool do_dumpout,
+                                    bool time_kernel,
+                                    const std::vector<size_t> inOutLengths,
+                                    const std::vector<int> reduceDims,
+                                    bool updateMovingAverage,
+                                    bool saveMeanAndInvVariance,
+                                    double averageFactor,
+                                    double epsilon)
+{
+    if(inOutLengths.size() != Rank || reduceDims.size() != NumBatchNormReduceDim)
+    {
+        throw std::runtime_error("Invalid tensor lengths or number of reduce dimensions!");
+    };
+
+    std::vector<size_t> scaleBiasMeanVarLengths;
+
+    // used for calculating the effective transferred bytes by each operation
+    size_t total_length;
+    size_t invariant_length = 1;
+
+    total_length =
+        std::accumulate(inOutLengths.begin(), inOutLengths.end(), 1, std::multiplies<size_t>{});
+
+    if(std::any_of(reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+        throw std::runtime_error("Invalid reduce dimensions!");
+
+    for(int dim = 0; dim < Rank; dim++)
+    {
+        if(std::none_of(reduceDims.begin(), reduceDims.end(), [&](int d) { return dim == d; }))
+        {
+            scaleBiasMeanVarLengths.push_back(inOutLengths[dim]);
+            invariant_length *= inOutLengths[dim];
+        };
+    }
+
+    // input data of the batchnorm forward algorithm
+    Tensor<XDataType> x(inOutLengths);
+    Tensor<ScaleDataType> bnScale(scaleBiasMeanVarLengths);
+    Tensor<BiasDataType> bnBias(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm forward algorithm
+    Tensor<YDataType> y_ref(inOutLengths);
+    Tensor<YDataType> y(inOutLengths);
+
+    Tensor<MeanVarDataType> resultSaveMean_ref(scaleBiasMeanVarLengths);
+    Tensor<MeanVarDataType> resultSaveInvVariance_ref(scaleBiasMeanVarLengths);
+
+    Tensor<MeanVarDataType> resultRunningMean_ref(scaleBiasMeanVarLengths);
+    Tensor<MeanVarDataType> resultRunningVariance_ref(scaleBiasMeanVarLengths);
+
+    auto inOutStrides            = x.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = bnScale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    if(updateMovingAverage)
+    {
+        const float x_mean       = 0.0f;
+        const float x_stddev     = 1.0f;
+        const float noise_stddev = 0.04f;
+
+        // input data in normal distribution
+        x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+
+        // initialize the runningMean to be values with tiny variation to the mean of the x
+        // values
+        resultRunningMean_ref.GenerateTensorValue(
+            GeneratorTensor_4<MeanVarDataType>{x_mean, noise_stddev}, num_thread);
+
+        // initialize the runningVariance to be values with tiny variation to the variance of
+        // the x values
+        resultRunningVariance_ref.GenerateTensorValue(
+            GeneratorTensor_4<MeanVarDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+    }
+    else
+    {
+        if constexpr(ck::is_same_v<XDataType, int8_t>)
+            x.GenerateTensorValue(GeneratorTensor_2<XDataType>{-5, 5}, num_thread);
+        else
+            x.GenerateTensorValue(GeneratorTensor_3<XDataType>{-1.0f, 1.0f}, num_thread);
+    };
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            bnScale.GenerateTensorValue(GeneratorTensor_0<ScaleDataType>{}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_0<BiasDataType>{}, num_thread);
+            break;
+        case 1:
+            bnScale.GenerateTensorValue(GeneratorTensor_1<ScaleDataType>{1}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_1<BiasDataType>{0}, num_thread);
+            break;
+        case 2:
+            bnScale.GenerateTensorValue(GeneratorTensor_2<ScaleDataType>{-5, 5}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            bnScale.GenerateTensorValue(GeneratorTensor_3<ScaleDataType>{-1.0f, 1.0f}, num_thread);
+            bnBias.GenerateTensorValue(GeneratorTensor_3<BiasDataType>{-1.0f, 1.0f}, num_thread);
+        }
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(XDataType) * y.mDesc.GetElementSpaceSize());
+    DeviceMem bnScale_dev(sizeof(ScaleDataType) * bnScale.mDesc.GetElementSpaceSize());
+    DeviceMem bnBias_dev(sizeof(BiasDataType) * bnBias.mDesc.GetElementSpaceSize());
+
+    // mean_dev or resultSaveMean_dev
+    DeviceMem resultSaveMean_dev(sizeof(MeanVarDataType) *
+                                 resultSaveMean_ref.mDesc.GetElementSpaceSize());
+    // meansquare_dev or resultSaveInvVariance_dev
+    DeviceMem resultSaveInvVariance_dev(sizeof(MeanVarDataType) *
+                                        resultSaveInvVariance_ref.mDesc.GetElementSpaceSize());
+    // resultRunningMean_dev
+    DeviceMem resultRunningMean_dev(sizeof(MeanVarDataType) *
+                                    resultRunningMean_ref.mDesc.GetElementSpaceSize());
+    // resultRunningVariance_dev
+    DeviceMem resultRunningVariance_dev(sizeof(MeanVarDataType) *
+                                        resultRunningVariance_ref.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    bnScale_dev.ToDevice(bnScale.mData.data());
+    bnBias_dev.ToDevice(bnBias.mData.data());
+
+    if(updateMovingAverage)
+    {
+        resultRunningMean_dev.ToDevice(resultRunningMean_ref.mData.data());
+        resultRunningVariance_dev.ToDevice(resultRunningVariance_ref.mData.data());
+    };
+
+    // used for storing the device result for verification when updateMovingAverage is enabled
+    Tensor<MeanVarDataType> resultRunningMean(scaleBiasMeanVarLengths);
+    Tensor<MeanVarDataType> resultRunningVariance(scaleBiasMeanVarLengths);
+
+    // used for storing the device result for verification when saveMeanAndInvVariance is enabled
+    Tensor<MeanVarDataType> resultSaveMean(scaleBiasMeanVarLengths);
+    Tensor<MeanVarDataType> resultSaveInvVariance(scaleBiasMeanVarLengths);
+
+    std::array<index_t, Rank> arrInOutLengths;
+    std::array<index_t, Rank> arrInOutStrides;
+    std::array<index_t, Rank - NumBatchNormReduceDim> arrScaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumBatchNormReduceDim> arrScaleBiasMeanVarStrides;
+    std::array<int, NumBatchNormReduceDim> arrReduceDims;
+
+    std::copy(inOutLengths.begin(), inOutLengths.end(), arrInOutLengths.begin());
+    std::copy(inOutStrides.begin(), inOutStrides.end(), arrInOutStrides.begin());
+    std::copy(scaleBiasMeanVarLengths.begin(),
+              scaleBiasMeanVarLengths.end(),
+              arrScaleBiasMeanVarLengths.begin());
+    std::copy(scaleBiasMeanVarStrides.begin(),
+              scaleBiasMeanVarStrides.end(),
+              arrScaleBiasMeanVarStrides.begin());
+
+    std::copy(reduceDims.begin(), reduceDims.end(), arrReduceDims.begin());
+
+    using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+    // add device batchnorm-forward instances
+    using DeviceOp = ck::tensor_operation::device::DeviceBatchNormFwd<XDataType,
+                                                                      YDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      BiasDataType,
+                                                                      MeanVarDataType,
+                                                                      PassThroughOp,
+                                                                      Rank,
+                                                                      NumBatchNormReduceDim>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using ReferenceBatchNormFwdInstance =
+            ck::tensor_operation::host::ReferenceBatchNormFwd<XDataType,
+                                                              YDataType,
+                                                              AccDataType,
+                                                              ScaleDataType,
+                                                              BiasDataType,
+                                                              MeanVarDataType,
+                                                              PassThroughOp,
+                                                              Rank,
+                                                              NumBatchNormReduceDim>;
+
+        auto batchNormFwd_ref = ReferenceBatchNormFwdInstance{};
+
+        auto argument_ptr_ref = batchNormFwd_ref.MakeArgumentPointer(
+            arrInOutLengths,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrReduceDims,
+            arrScaleBiasMeanVarLengths,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            x.mData.data(),
+            bnScale.mData.data(),
+            bnBias.mData.data(),
+            epsilon,
+            PassThroughOp{},
+            y_ref.mData.data(),
+            saveMeanAndInvVariance ? resultSaveMean_ref.mData.data() : nullptr,
+            saveMeanAndInvVariance ? resultSaveInvVariance_ref.mData.data() : nullptr,
+            averageFactor,
+            updateMovingAverage ? resultRunningMean_ref.mData.data() : nullptr,
+            updateMovingAverage ? resultRunningVariance_ref.mData.data() : nullptr);
+
+        if(!batchNormFwd_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters not supported by the reference instance, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = batchNormFwd_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+    }
+
+    int num_kernel = 0;
+    bool pass      = true;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            arrInOutLengths,
+            arrInOutStrides,
+            arrInOutStrides,
+            arrReduceDims,
+            arrScaleBiasMeanVarLengths,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            arrScaleBiasMeanVarStrides,
+            x_dev.GetDeviceBuffer(),
+            bnScale_dev.GetDeviceBuffer(),
+            bnBias_dev.GetDeviceBuffer(),
+            epsilon,
+            PassThroughOp{},
+            y_dev.GetDeviceBuffer(),
+            saveMeanAndInvVariance ? resultSaveMean_dev.GetDeviceBuffer() : nullptr,
+            saveMeanAndInvVariance ? resultSaveInvVariance_dev.GetDeviceBuffer() : nullptr,
+            averageFactor,
+            updateMovingAverage ? resultRunningMean_dev.GetDeviceBuffer() : nullptr,
+            updateMovingAverage ? resultRunningVariance_dev.GetDeviceBuffer() : nullptr);
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            num_kernel++;
+        }
+        else
+        {
+            if(time_kernel)
+            {
+                std::cout << inst_ptr->GetTypeString()
+                          << " skipped due to unsupported argument: " << std::endl;
+            }
+
+            continue;
+        };
+
+        size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
+
+        DeviceMem workspace_dev(workspace_sz);
+
+        inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        size_t num_bytes = 0;
+
+        // inputing of x, scale, bias, outputing of y
+        num_bytes += total_length * (sizeof(XDataType) + sizeof(YDataType)) +
+                     invariant_length * (sizeof(ScaleDataType) + sizeof(BiasDataType));
+
+        // outputing of mean, inv-variance
+        num_bytes += saveMeanAndInvVariance ? invariant_length * sizeof(MeanVarDataType) * 2 : 0;
+
+        // updating of moving mean, variance
+        num_bytes += updateMovingAverage ? invariant_length * sizeof(MeanVarDataType) * 4 : 0;
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            using ck::utils::check_err;
+            bool single_pass;
+
+            y_dev.FromDevice(y.mData.data());
+
+            if constexpr(ck::is_same_v<YDataType, ck::bhalf_t>)
+                single_pass = check_err(y.mData, y_ref.mData, "y results", 1e-2, 1e-2);
+            else
+                single_pass = check_err(y.mData, y_ref.mData, "y results", 4e-3, 4e-3);
+
+            if(updateMovingAverage)
+            {
+                resultRunningMean_dev.FromDevice(resultRunningMean.mData.data());
+                resultRunningVariance_dev.FromDevice(resultRunningVariance.mData.data());
+
+                // clang-format off
+                single_pass = single_pass && check_err(resultRunningMean.mData, resultRunningMean_ref.mData, "average mean results", 1.5e-5, 1.5e-5);
+                single_pass = single_pass && check_err(resultRunningVariance.mData, resultRunningVariance_ref.mData, "average variance results", 1e-5, 1e-5);
+                // clang-format on
+            };
+
+            if(saveMeanAndInvVariance)
+            {
+                resultSaveMean_dev.FromDevice(resultSaveMean.mData.data());
+                resultSaveInvVariance_dev.FromDevice(resultSaveInvVariance.mData.data());
+
+                // clang-format off
+                single_pass = single_pass && check_err(resultSaveMean.mData, resultSaveMean_ref.mData, "mean results", 3e-5, 3e-5);
+                single_pass = single_pass && check_err(resultSaveInvVariance.mData, resultSaveInvVariance_ref.mData, "inv-variance results", 7e-5, 7e-5);
+                // clang-format on
+            };
+
+            pass = pass && single_pass;
+        };
+
+        if(do_dumpout)
+        {
+            using ck::host_common::dumpBufferToFile;
+
+            // clang-format off
+            dumpBufferToFile("dump_x.bin", x.mData.data(), x.mDesc.GetElementSize());
+            dumpBufferToFile("dump_y.bin", y.mData.data(), y.mDesc.GetElementSize());
+            dumpBufferToFile("dump_y_ref.bin", y_ref.mData.data(), y_ref.mDesc.GetElementSize());
+            // clang-format off
+
+            if(saveMeanAndInvVariance)
+            {
+                // clang-format off
+                dumpBufferToFile("dump_mean.bin", resultSaveMean.mData.data(), resultSaveMean.mDesc.GetElementSize());
+                dumpBufferToFile("dump_mean_ref.bin", resultSaveMean_ref.mData.data(), resultSaveMean_ref.mDesc.GetElementSize()); 
+                dumpBufferToFile("dump_invvar.bin", resultSaveInvVariance.mData.data(), resultSaveInvVariance.mDesc.GetElementSize());
+                dumpBufferToFile("dump_invvar_ref.bin", resultSaveInvVariance_ref.mData.data(), resultSaveInvVariance_ref.mDesc.GetElementSize());
+                // clang-format on
+            };
+        };
+    }
+
+    if(time_kernel)
+    {
+        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_batchnorm_infer_impl.hpp b/profiler/include/profiler/profile_batchnorm_infer_impl.hpp
new file mode 100644
index 0000000000..1b31a2aabf
--- /dev/null
+++ b/profiler/include/profiler/profile_batchnorm_infer_impl.hpp
@@ -0,0 +1,335 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <stdexcept>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename XDataType,
+          typename YDataType,
+          typename AccDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          index_t Rank,
+          index_t NumBatchNormReduceDim>
+bool profile_batchnorm_infer_impl(int do_verification,
+                                  int init_method,
+                                  bool do_dumpout,
+                                  bool time_kernel,
+                                  const std::vector<size_t> inOutLengths,
+                                  const std::vector<int> reduceDims,
+                                  double epsilon)
+{
+    if(inOutLengths.size() != Rank || reduceDims.size() != NumBatchNormReduceDim)
+    {
+        throw std::runtime_error("Invalid tensor lengths or number of reduce dimensions!");
+    };
+
+    std::vector<size_t> scaleBiasMeanVarLengths;
+    std::vector<int> invariantDims;
+
+    // used for calculating the effective transferred bytes by each operation
+    size_t total_length;
+    size_t invariant_length = 1;
+
+    total_length =
+        std::accumulate(inOutLengths.begin(), inOutLengths.end(), 1, std::multiplies<size_t>{});
+
+    if(std::any_of(reduceDims.begin(), reduceDims.end(), [](int d) { return d < 0 || d >= Rank; }))
+        throw std::runtime_error("Invalid reduce dimensions!");
+
+    for(int dim = 0; dim < Rank; dim++)
+    {
+        if(std::none_of(reduceDims.begin(), reduceDims.end(), [&](int d) { return dim == d; }))
+        {
+            invariantDims.push_back(dim);
+            scaleBiasMeanVarLengths.push_back(inOutLengths[dim]);
+            invariant_length *= inOutLengths[dim];
+        };
+    }
+
+    // input data of the batchnorm infer algorithm
+    Tensor<XDataType> x(inOutLengths);
+    Tensor<ScaleDataType> scale(scaleBiasMeanVarLengths);
+    Tensor<BiasDataType> bias(scaleBiasMeanVarLengths);
+    Tensor<MeanVarDataType> estimatedMean(scaleBiasMeanVarLengths);
+    Tensor<MeanVarDataType> estimatedVariance(scaleBiasMeanVarLengths);
+
+    // output data of the batchnorm infer algorithm
+    Tensor<YDataType> y_ref(inOutLengths);
+    Tensor<YDataType> y(inOutLengths);
+
+    auto inOutStrides            = x.mDesc.GetStrides();
+    auto scaleBiasMeanVarStrides = scale.mDesc.GetStrides();
+
+    std::size_t num_thread = std::thread::hardware_concurrency();
+
+    const float x_mean       = 0.0f;
+    const float x_stddev     = 1.0f;
+    const float noise_stddev = 0.04f;
+
+    // input data in normal distribution
+    x.GenerateTensorValue(GeneratorTensor_4<XDataType>{x_mean, x_stddev}, num_thread);
+
+    // initialize the estimatedMean to be values with tiny variation to the mean of the x
+    // values
+    estimatedMean.GenerateTensorValue(GeneratorTensor_4<MeanVarDataType>{x_mean, noise_stddev},
+                                      num_thread);
+
+    // initialize the estimatedVariance to be values with tiny variation to the variance of
+    // the x values
+    estimatedVariance.GenerateTensorValue(
+        GeneratorTensor_4<MeanVarDataType>{x_stddev * x_stddev, noise_stddev}, num_thread);
+
+    if(do_verification)
+    {
+        switch(init_method)
+        {
+        case 0:
+            scale.GenerateTensorValue(GeneratorTensor_0<ScaleDataType>{}, num_thread);
+            bias.GenerateTensorValue(GeneratorTensor_0<BiasDataType>{}, num_thread);
+            break;
+        case 1:
+            scale.GenerateTensorValue(GeneratorTensor_1<ScaleDataType>{1}, num_thread);
+            bias.GenerateTensorValue(GeneratorTensor_1<BiasDataType>{0}, num_thread);
+            break;
+        case 2:
+            scale.GenerateTensorValue(GeneratorTensor_2<ScaleDataType>{-5, 5}, num_thread);
+            bias.GenerateTensorValue(GeneratorTensor_2<BiasDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            scale.GenerateTensorValue(GeneratorTensor_3<ScaleDataType>{-1.0f, 1.0f}, num_thread);
+            bias.GenerateTensorValue(GeneratorTensor_3<BiasDataType>{-1.0f, 1.0f}, num_thread);
+        }
+    };
+
+    // these buffers are usually provided by the user application
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(XDataType) * y.mDesc.GetElementSpaceSize());
+    DeviceMem scale_dev(sizeof(ScaleDataType) * scale.mDesc.GetElementSpaceSize());
+    DeviceMem bias_dev(sizeof(BiasDataType) * bias.mDesc.GetElementSpaceSize());
+
+    // estimatedMean_dev
+    DeviceMem estimatedMean_dev(sizeof(MeanVarDataType) *
+                                estimatedMean.mDesc.GetElementSpaceSize());
+    // estimatedVariance_dev
+    DeviceMem estimatedVariance_dev(sizeof(MeanVarDataType) *
+                                    estimatedVariance.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    scale_dev.ToDevice(scale.mData.data());
+    bias_dev.ToDevice(bias.mData.data());
+    estimatedMean_dev.ToDevice(estimatedMean.mData.data());
+    estimatedVariance_dev.ToDevice(estimatedVariance.mData.data());
+
+    std::array<index_t, Rank> arrInOutLengths;
+    std::array<index_t, Rank> arrInOutStrides;
+    std::array<index_t, Rank - NumBatchNormReduceDim> arrScaleBiasMeanVarLengths;
+    std::array<index_t, Rank - NumBatchNormReduceDim> arrScaleBiasMeanVarStrides;
+    std::array<int, NumBatchNormReduceDim> arrReduceDims;
+
+    std::copy(inOutLengths.begin(), inOutLengths.end(), arrInOutLengths.begin());
+    std::copy(inOutStrides.begin(), inOutStrides.end(), arrInOutStrides.begin());
+    std::copy(scaleBiasMeanVarLengths.begin(),
+              scaleBiasMeanVarLengths.end(),
+              arrScaleBiasMeanVarLengths.begin());
+    std::copy(scaleBiasMeanVarStrides.begin(),
+              scaleBiasMeanVarStrides.end(),
+              arrScaleBiasMeanVarStrides.begin());
+
+    std::copy(reduceDims.begin(), reduceDims.end(), arrReduceDims.begin());
+
+    std::array<index_t, Rank> aligned_scaleBiasMeanVarStrides{0};
+
+    int i = 0;
+    for(auto dim : invariantDims)
+    {
+        assert(inOutLengths[dim] == scaleBiasMeanVarLengths[i]);
+
+        aligned_scaleBiasMeanVarStrides[dim] = scaleBiasMeanVarStrides[i];
+        i++;
+    };
+
+    using Normalize = ck::tensor_operation::element_wise::NormalizeInInfer;
+
+    // add device batchnorm-infer instances
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<XDataType, MeanVarDataType, MeanVarDataType, ScaleDataType, BiasDataType>,
+        ck::Tuple<YDataType>,
+        Normalize,
+        Rank>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
+
+        using ReferenceBatchNormInferInstance =
+            ck::tensor_operation::host::ReferenceBatchNormInfer<XDataType,
+                                                                YDataType,
+                                                                AccDataType,
+                                                                ScaleDataType,
+                                                                BiasDataType,
+                                                                MeanVarDataType,
+                                                                PassThroughOp,
+                                                                Rank,
+                                                                NumBatchNormReduceDim>;
+        auto batchNormInfer_ref = ReferenceBatchNormInferInstance{};
+
+        auto argument_ptr_ref =
+            batchNormInfer_ref.MakeArgumentPointer(arrInOutLengths,
+                                                   arrInOutStrides,
+                                                   arrInOutStrides,
+                                                   arrReduceDims,
+                                                   arrScaleBiasMeanVarLengths,
+                                                   arrScaleBiasMeanVarStrides,
+                                                   arrScaleBiasMeanVarStrides,
+                                                   arrScaleBiasMeanVarStrides,
+                                                   x.mData.data(),
+                                                   scale.mData.data(),
+                                                   bias.mData.data(),
+                                                   epsilon,
+                                                   PassThroughOp{},
+                                                   estimatedMean.mData.data(),
+                                                   estimatedVariance.mData.data(),
+                                                   y_ref.mData.data());
+
+        if(!batchNormInfer_ref.IsSupportedArgument(argument_ptr_ref.get()))
+        {
+            std::cout << "The runtime parameters not supported by the reference instance, exiting!"
+                      << std::endl;
+            return (false);
+        };
+
+        auto invoker_ptr_ref = batchNormInfer_ref.MakeInvokerPointer();
+
+        (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+    }
+
+    int num_kernel = 0;
+    bool pass      = true;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(arrInOutLengths,
+                                                          {arrInOutStrides,
+                                                           aligned_scaleBiasMeanVarStrides,
+                                                           aligned_scaleBiasMeanVarStrides,
+                                                           aligned_scaleBiasMeanVarStrides,
+                                                           aligned_scaleBiasMeanVarStrides},
+                                                          {arrInOutStrides},
+                                                          {x_dev.GetDeviceBuffer(),
+                                                           estimatedMean_dev.GetDeviceBuffer(),
+                                                           estimatedVariance_dev.GetDeviceBuffer(),
+                                                           scale_dev.GetDeviceBuffer(),
+                                                           bias_dev.GetDeviceBuffer()},
+                                                          {y_dev.GetDeviceBuffer()},
+                                                          Normalize{epsilon});
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            num_kernel++;
+        }
+        else
+        {
+            if(time_kernel)
+            {
+                std::cout << inst_ptr->GetTypeString()
+                          << " skipped due to unsupported argument: " << std::endl;
+            }
+
+            continue;
+        };
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        size_t num_bytes = 0;
+
+        // inputing of x, scale, bias, outputing of y
+        num_bytes += total_length * (sizeof(XDataType) + sizeof(YDataType)) +
+                     invariant_length *
+                         (sizeof(ScaleDataType) + sizeof(BiasDataType) + sizeof(MeanVarDataType));
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            using ck::utils::check_err;
+            bool single_pass;
+
+            y_dev.FromDevice(y.mData.data());
+
+            if constexpr(ck::is_same_v<YDataType, ck::bhalf_t>)
+                single_pass = check_err(y.mData, y_ref.mData, "y results", 1e-2, 1e-2);
+            else
+                single_pass = check_err(y.mData, y_ref.mData, "y results", 4e-3, 4e-3);
+
+            pass = pass && single_pass;
+        };
+
+        if(do_dumpout)
+        {
+            using ck::host_common::dumpBufferToFile;
+
+            // clang-format off
+            dumpBufferToFile("dump_x.bin", x.mData.data(), x.mDesc.GetElementSize());
+            dumpBufferToFile("dump_y.bin", y.mData.data(), y.mDesc.GetElementSize());
+            dumpBufferToFile("dump_y_ref.bin", y_ref.mData.data(), y_ref.mDesc.GetElementSize());
+            // clang-format off
+        };
+    }
+
+    if(time_kernel)
+    {
+        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_contraction_impl.hpp b/profiler/include/profiler/profile_contraction_impl.hpp
new file mode 100644
index 0000000000..660cc3f9e5
--- /dev/null
+++ b/profiler/include/profiler/profile_contraction_impl.hpp
@@ -0,0 +1,345 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+#include <limits>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp"
+#include "ck/library/tensor_operation_instance/gpu/contraction_scale.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
+
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace profiler {
+
+using Bilinear = ck::tensor_operation::element_wise::Bilinear;
+using Scale    = ck::tensor_operation::element_wise::Scale;
+
+template <typename ALayout,
+          typename BLayout,
+          typename CDELayout,
+          typename DataType,
+          typename DTupleDataType,
+          typename CDElementOp>
+int profile_contraction_impl(ck::index_t do_verification,
+                             ck::index_t init_method,
+                             bool do_log,
+                             bool time_kernel,
+                             CDElementOp cde_element_op,
+                             const std::vector<ck::index_t>& M,
+                             const std::vector<ck::index_t>& N,
+                             const std::vector<ck::index_t>& K,
+                             const std::vector<ck::index_t>& StridesA,
+                             const std::vector<ck::index_t>& StridesB,
+                             const std::vector<ck::index_t>& StridesE,
+                             const std::vector<ck::index_t>& StridesD)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor = [](const std::vector<ck::index_t>& dims01,
+                                       const std::vector<ck::index_t>& dims23,
+                                       const std::vector<ck::index_t>& strides) {
+        std::vector<std::size_t> dims_szt(dims01.begin(), dims01.end());
+        dims_szt.insert(dims_szt.end(), dims23.begin(), dims23.end());
+        std::vector<std::size_t> strides_szt(strides.begin(), strides.end());
+
+        return HostTensorDescriptor(dims_szt, strides);
+    };
+
+    Tensor<DataType> a_m_k(f_host_tensor_descriptor(M, K, StridesA));
+    Tensor<DataType> b_k_n(f_host_tensor_descriptor(K, N, StridesB));
+    Tensor<DataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StridesE));
+    Tensor<DataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StridesE));
+    Tensor<DataType> d_m_n(f_host_tensor_descriptor(M, N, StridesD));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<DataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<DataType>{-5, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<DataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<DataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<DataType>{-0.5, 0.5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<DataType>{-0.5, 0.5});
+    }
+
+    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    DeviceMem a_device_buf(sizeof(DataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(DataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(DataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DataType) * d_m_n.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    e_device_buf.SetZero();
+    d_device_buf.ToDevice(d_m_n.mData.data());
+
+    const std::vector<index_t> a_ms_ks_lengths = {M[0], M[1], K[0], K[1]};
+    const std::vector<index_t> b_ns_ks_lengths = {N[0], N[1], K[0], K[1]};
+    const std::vector<index_t> e_ms_ns_lengths = {M[0], M[1], N[0], N[1]};
+    const std::vector<index_t> d_m_n_lengths   = {M[0], M[1], N[0], N[1]};
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+
+    constexpr ck::index_t NumDim = 2;
+    using DeviceOp               = ck::tensor_operation::device::DeviceContractionMultipleD<NumDim,
+                                                                              NumDim,
+                                                                              NumDim,
+                                                                              DataType,
+                                                                              DataType,
+                                                                              DTupleDataType,
+                                                                              DataType,
+                                                                              AElementOp,
+                                                                              BElementOp,
+                                                                              CDElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // Run reference op
+    if(do_verification)
+    {
+        using ReferenceGemmInstance =
+            ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDim,
+                                                                      NumDim,
+                                                                      NumDim,
+                                                                      DataType,
+                                                                      DataType,
+                                                                      DataType,
+                                                                      DataType,
+                                                                      AElementOp,
+                                                                      BElementOp>;
+
+        auto ref_op      = ReferenceGemmInstance{};
+        auto ref_invoker = ref_op.MakeInvoker();
+
+        Tensor<DataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StridesE));
+
+        auto ref_argument =
+            ref_op.MakeArgument(a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t m0 = 0; m0 < e_m_n_host_result.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < e_m_n_host_result.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t n0 = 0; n0 < e_m_n_host_result.mDesc.GetLengths()[2]; ++n0)
+                {
+                    for(size_t n1 = 0; n1 < e_m_n_host_result.mDesc.GetLengths()[3]; ++n1)
+                    {
+                        if constexpr(is_same<CDElementOp, Bilinear>::value)
+                        {
+                            cde_element_op(e_m_n_host_result(m0, m1, n0, n1),
+                                           c_m_n_host_result(m0, m1, n0, n1),
+                                           d_m_n(m0, m1, n0, n1));
+                        }
+                        else if constexpr(is_same<CDElementOp, Scale>::value)
+                        {
+                            cde_element_op(e_m_n_host_result(m0, m1, n0, n1),
+                                           c_m_n_host_result(m0, m1, n0, n1));
+                        }
+                        else
+                        {
+                            static_assert("Unsupported CDElementOp in contraction profiler.");
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    std::string best_op_name;
+    float best_avg_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        std::unique_ptr<tensor_operation::device::BaseArgument> argument_ptr;
+        if constexpr(is_same<CDElementOp, Bilinear>::value)
+        {
+            argument_ptr = op_ptr->MakeArgumentPointer(
+                static_cast<DataType*>(a_device_buf.GetDeviceBuffer()),
+                static_cast<DataType*>(b_device_buf.GetDeviceBuffer()),
+                std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                static_cast<DataType*>(e_device_buf.GetDeviceBuffer()),
+                a_ms_ks_lengths,
+                StridesA,
+                b_ns_ks_lengths,
+                StridesB,
+                std::array<std::vector<ck::index_t>, 1>{d_m_n_lengths},
+                std::array<std::vector<ck::index_t>, 1>{StridesD},
+                e_ms_ns_lengths,
+                StridesE,
+                a_element_op,
+                b_element_op,
+                cde_element_op);
+        }
+        else if constexpr(is_same<CDElementOp, Scale>::value)
+        {
+            argument_ptr =
+                op_ptr->MakeArgumentPointer(static_cast<DataType*>(a_device_buf.GetDeviceBuffer()),
+                                            static_cast<DataType*>(b_device_buf.GetDeviceBuffer()),
+                                            std::array<const void*, 0>{},
+                                            static_cast<DataType*>(e_device_buf.GetDeviceBuffer()),
+                                            a_ms_ks_lengths,
+                                            StridesA,
+                                            b_ns_ks_lengths,
+                                            StridesB,
+                                            std::array<std::vector<ck::index_t>, 0>{},
+                                            std::array<std::vector<ck::index_t>, 0>{},
+                                            e_ms_ns_lengths,
+                                            StridesE,
+                                            a_element_op,
+                                            b_element_op,
+                                            cde_element_op);
+        }
+        else
+        {
+            static_assert("Unsupported CDElementOp in contraction profiler.");
+        }
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        auto nelems_m = M[0] * M[1];
+        auto nelems_n = N[0] * N[1];
+        auto nelems_k = K[0] * K[1];
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init C to zero before profiling next kernel
+            e_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * nelems_m * nelems_n * nelems_k;
+
+            std::size_t num_btype = sizeof(DataType) * nelems_m * nelems_k +
+                                    sizeof(DataType) * nelems_k * nelems_n +
+                                    sizeof(DataType) * nelems_m * nelems_n;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
+
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                float threshold =
+                    static_cast<DataType>(nelems_k) * std::numeric_limits<DataType>::epsilon();
+                pass = pass & ck::utils::check_err(e_m_n_device_result,
+                                                   e_m_n_host_result,
+                                                   "Error: incorrect results!",
+                                                   threshold,
+                                                   threshold);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host  : ", e_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_device: ", e_m_n_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    if constexpr(is_same<DataType, float>::value)
+    {
+        std::cout << "Best Perf for datatype = f32";
+    }
+    else if constexpr(is_same<DataType, double>::value)
+    {
+        std::cout << "Best Perf for datatype = f64";
+    }
+
+    if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " ALayout =  RowMajor";
+    }
+    else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " ALayout =  ColumnMajor";
+    }
+
+    if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " BLayout =  RowMajor";
+    }
+    else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " BLayout =  ColumnMajor";
+    }
+
+    if constexpr(is_same<CDELayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " CDELayout =  RowMajor";
+    }
+    else if constexpr(is_same<CDELayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " CDELayout =  ColumnMajor";
+    }
+
+    std::cout << " M = " << M << " N = " << N << " K = " << K << " StridesA = " << StridesA
+              << " StridesB = " << StridesB << " StridesE = " << StridesE << " : " << best_avg_time
+              << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_contraction_utils.hpp b/profiler/include/profiler/profile_contraction_utils.hpp
new file mode 100644
index 0000000000..076bbd4559
--- /dev/null
+++ b/profiler/include/profiler/profile_contraction_utils.hpp
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/ck.hpp"
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using Bilinear = ck::tensor_operation::element_wise::Bilinear;
+using Scale    = ck::tensor_operation::element_wise::Scale;
+
+enum struct ContractionMatrixLayout
+{
+    MK_KN_MN_MN, // 0
+    MK_NK_MN_MN, // 1
+    KM_KN_MN_MN, // 2
+    KM_NK_MN_MN, // 3
+};
+
+enum struct ContractionDataType
+{
+    F32_F32_F32_F32, // 0
+    F64_F64_F64_F64, // 1
+};
+
+inline void collect_index_params(char* argv[],
+                                 std::vector<ck::index_t>& params,
+                                 const ck::index_t from,
+                                 const ck::index_t num)
+{
+    for(ck::index_t p = from; p < from + num; p++)
+        params.push_back(std::stoi(argv[p]));
+}
+
+// Defualt strides for row-major: {Dim1 * Dim2 * Dim3, Dim2 * Dim3, Dim3, 1}
+// Defualt strides for column-major: {Dim1, 1, Dim0 * Dim1 * Dim3, Dim0 * Dim1}
+inline void
+assign_default_strides(Row, std::vector<ck::index_t>& strides, std::vector<ck::index_t> dims)
+{
+    strides = {dims[1] * dims[2] * dims[3], dims[2] * dims[3], dims[3], 1};
+}
+
+inline void
+assign_default_strides(Col, std::vector<ck::index_t>& strides, std::vector<ck::index_t> dims)
+{
+    strides = {dims[1], 1, dims[0] * dims[1] * dims[3], dims[0] * dims[1]};
+}
diff --git a/profiler/include/profiler/profile_conv_bwd_data_impl.hpp b/profiler/include/profiler/profile_conv_bwd_data_impl.hpp
new file mode 100644
index 0000000000..52152a90fe
--- /dev/null
+++ b/profiler/include/profiler/profile_conv_bwd_data_impl.hpp
@@ -0,0 +1,248 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename DataType>
+void show_data_nhwc_layout(Tensor<DataType>& nhwc)
+{
+    std::cout << "[";
+    for(int n = 0; n < ck::type_convert<int>(nhwc.mDesc.GetLengths()[0]); n++)
+    {
+        std::cout << "[";
+        for(int hi = 0; hi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[2]); hi++)
+        {
+            std::cout << "[";
+            for(int wi = 0; wi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[3]); wi++)
+            {
+                std::cout << "[";
+                for(int c = 0; c < ck::type_convert<int>(nhwc.mDesc.GetLengths()[1]); c++)
+                {
+                    std::cout << static_cast<float>(nhwc(n, c, hi, wi)) << "  ";
+                }
+                std::cout << "]";
+            }
+            std::cout << "]";
+        }
+        std::cout << "]";
+    }
+    std::cout << "]";
+}
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+bool profile_conv_bwd_data_impl(int do_verification,
+                                int init_method,
+                                bool do_log,
+                                bool time_kernel,
+                                const ck::utils::conv::ConvParam& conv_param)
+{
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    Tensor<InDataType> input_host_result(in_g_n_c_wis_desc);
+    Tensor<InDataType> input_device_result(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> weight(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> output(out_g_n_k_wos_desc);
+
+    std::cout << "input: " << input_host_result.mDesc << std::endl;
+    std::cout << "weight: " << weight.mDesc << std::endl;
+    std::cout << "output: " << output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        output.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        weight.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        output.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+        weight.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * input_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * weight.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * output.mDesc.GetElementSpaceSize());
+
+    out_device_buf.ToDevice(output.mData.data());
+    wei_device_buf.ToDevice(weight.mData.data());
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<NDimSpatial,
+                                                                         InDataType,
+                                                                         WeiDataType,
+                                                                         OutDataType,
+                                                                         InElementOp,
+                                                                         WeiElementOp,
+                                                                         OutElementOp>{};
+
+        auto ref_invoker = ref_conv.MakeInvoker();
+
+        auto ref_argument = ref_conv.MakeArgument(input_host_result,
+                                                  weight,
+                                                  output,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  InElementOp{},
+                                                  WeiElementOp{},
+                                                  OutElementOp{});
+        ref_invoker.Run(ref_argument);
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceConvBwdData<NDimSpatial,
+                                                                     InLayout,
+                                                                     WeiLayout,
+                                                                     OutLayout,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    float best_avg_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device Conv instances
+    bool pass = true;
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                        static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                        conv_param.N_,
+                                        conv_param.K_,
+                                        conv_param.C_,
+                                        conv_param.input_spatial_lengths_,
+                                        conv_param.filter_spatial_lengths_,
+                                        conv_param.output_spatial_lengths_,
+                                        conv_param.conv_filter_strides_,
+                                        conv_param.conv_filter_dilations_,
+                                        conv_param.input_left_pads_,
+                                        conv_param.input_right_pads_,
+                                        in_element_op,
+                                        wei_element_op,
+                                        out_element_op);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // for conv bwd data, some input tensor element are zero, but not written by kernel,
+            // need to set zero
+            in_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = conv_param.GetFlops();
+            std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s" << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                in_device_buf.FromDevice(input_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(input_device_result, input_host_result);
+
+                if(do_log)
+                {
+                    std::cout << "in : ";
+                    show_data_nhwc_layout(output);
+                    std::cout << std::endl;
+
+                    std::cout << "wei: ";
+                    show_data_nhwc_layout(weight);
+                    std::cout << std::endl;
+
+                    std::cout << "out_host  : ";
+                    show_data_nhwc_layout(input_host_result);
+                    std::cout << std::endl;
+
+                    std::cout << "out_device: ";
+                    show_data_nhwc_layout(input_device_result);
+                    std::cout << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best configuration parameters:"
+              << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
+              << "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp b/profiler/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp
new file mode 100644
index 0000000000..436fbdbd75
--- /dev/null
+++ b/profiler/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp
@@ -0,0 +1,278 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using DeviceConvFwdBiasReluAddPtr =
+    DeviceConvFwdBiasActivationAddPtr<ck::tensor_operation::element_wise::PassThrough,
+                                      ck::tensor_operation::element_wise::PassThrough,
+                                      ck::tensor_operation::element_wise::AddReluAdd>;
+
+void add_device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<DeviceConvFwdBiasReluAddPtr>&);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+template <int NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout>
+void profile_conv_fwd_bias_relu_add_impl(int do_verification,
+                                         int init_method,
+                                         bool do_log,
+                                         bool time_kernel,
+                                         ck::index_t N,
+                                         ck::index_t K,
+                                         ck::index_t C,
+                                         std::vector<ck::index_t> input_spatial_lengths,
+                                         std::vector<ck::index_t> filter_spatial_lengths,
+                                         std::vector<ck::index_t> output_spatial_lengths,
+                                         std::vector<ck::index_t> conv_filter_strides,
+                                         std::vector<ck::index_t> conv_filter_dilations,
+                                         std::vector<ck::index_t> input_left_pads,
+                                         std::vector<ck::index_t> input_right_pads)
+{
+    const ck::index_t Y = filter_spatial_lengths[0];
+    const ck::index_t X = filter_spatial_lengths[1];
+
+    const ck::index_t Hi = input_spatial_lengths[0];
+    const ck::index_t Wi = input_spatial_lengths[1];
+
+    const ck::index_t Ho = output_spatial_lengths[0];
+    const ck::index_t Wo = output_spatial_lengths[1];
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
+            using namespace ck::literals;
+
+            if constexpr(is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
+                         is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
+                         is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
+            {
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, H * W, W, 1_uz});
+            }
+            else if constexpr(is_same<decltype(layout), tensor_layout::convolution::NHWC>::value ||
+                              is_same<decltype(layout), tensor_layout::convolution::KYXC>::value ||
+                              is_same<decltype(layout), tensor_layout::convolution::NHWK>::value)
+            {
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, 1_uz, W * C_, C_});
+            }
+        };
+
+    Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
+    Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
+    Tensor<OutDataType> out_n_k_ho_wo_host_result(
+        f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
+    Tensor<OutDataType> out_n_k_ho_wo_device_result(
+        f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
+
+    // bias: assume contiguous 1d vector
+    Tensor<OutDataType> bias_k({K});
+
+    // residual: assume same layout as output tensor
+    Tensor<OutDataType> resi_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
+    std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo_host_result.mDesc << std::endl;
+    std::cout << "bias_k: " << bias_k.mDesc << std::endl;
+    std::cout << "resi_n_k_ho_wo: " << resi_n_k_ho_wo.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        bias_k.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        resi_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        break;
+    default:
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+        bias_k.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+        resi_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+    }
+
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    if(do_verification)
+    {
+        using ReferenceConvFwdInstance =
+            ck::tensor_operation::host::ReferenceConvFwd_Bias_Activation_Add<InDataType,
+                                                                             WeiDataType,
+                                                                             OutDataType,
+                                                                             InElementOp,
+                                                                             WeiElementOp,
+                                                                             OutElementOp>;
+
+        auto ref_conv    = ReferenceConvFwdInstance{};
+        auto ref_invoker = ref_conv.MakeInvoker();
+
+        auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
+                                                  wei_k_c_y_x,
+                                                  out_n_k_ho_wo_host_result,
+                                                  bias_k,
+                                                  resi_n_k_ho_wo,
+                                                  conv_filter_strides,
+                                                  conv_filter_dilations,
+                                                  input_left_pads,
+                                                  input_right_pads,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) *
+                             out_n_k_ho_wo_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(OutDataType) * bias_k.mDesc.GetElementSpaceSize());
+    DeviceMem resi_device_buf(sizeof(OutDataType) * resi_n_k_ho_wo.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+    wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
+    bias_device_buf.ToDevice(bias_k.mData.data());
+    resi_device_buf.ToDevice(resi_n_k_ho_wo.mData.data());
+
+    using DeviceConvFwdBiasReluAddPtr = ck::tensor_operation::device::
+        DeviceConvFwdBiasActivationAddPtr<InElementOp, WeiElementOp, OutElementOp>;
+
+    // add device operator instances
+    std::vector<DeviceConvFwdBiasReluAddPtr> op_ptrs;
+
+    if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::half_t> &&
+                 ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::half_t> &&
+                 ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::half_t>)
+    {
+        ck::tensor_operation::device::instance::
+            add_device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
+    }
+
+    if(op_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device Conv instance found");
+    }
+
+    std::string best_conv_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device Conv instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<const InDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<const WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            static_cast<const OutDataType*>(bias_device_buf.GetDeviceBuffer()),
+            static_cast<const OutDataType*>(resi_device_buf.GetDeviceBuffer()),
+            N,
+            K,
+            C,
+            input_spatial_lengths,
+            filter_spatial_lengths,
+            output_spatial_lengths,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            in_element_op,
+            wei_element_op,
+            out_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string conv_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
+
+            std::size_t num_btype =
+                sizeof(InDataType) * (N * C * Hi * Wi) + sizeof(WeiDataType) * (K * C * Y * X) +
+                sizeof(OutDataType) * (N * K * Ho * Wo) + sizeof(OutDataType) * (K) +
+                sizeof(OutDataType) * (N * K * Ho * Wo);
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << conv_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_conv_name  = conv_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
+
+                ck::utils::check_err(out_n_k_ho_wo_device_result, out_n_k_ho_wo_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "wei: ", wei_k_c_y_x.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "out_host  : ", out_n_k_ho_wo_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "out_device: ", out_n_k_ho_wo_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_conv_fwd_bias_relu_impl.hpp b/profiler/include/profiler/profile_conv_fwd_bias_relu_impl.hpp
new file mode 100644
index 0000000000..808c1a1c90
--- /dev/null
+++ b/profiler/include/profiler/profile_conv_fwd_bias_relu_impl.hpp
@@ -0,0 +1,266 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using DeviceConvFwdBiasReluPtr =
+    DeviceConvFwdBiasActivationPtr<ck::tensor_operation::element_wise::PassThrough,
+                                   ck::tensor_operation::element_wise::PassThrough,
+                                   ck::tensor_operation::element_wise::AddRelu>;
+
+void add_device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances(
+    std::vector<DeviceConvFwdBiasReluPtr>&);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+template <int NDimSpatial,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout>
+void profile_conv_fwd_bias_relu_impl(int do_verification,
+                                     int init_method,
+                                     bool do_log,
+                                     bool time_kernel,
+                                     ck::index_t N,
+                                     ck::index_t K,
+                                     ck::index_t C,
+                                     std::vector<ck::index_t> input_spatial_lengths,
+                                     std::vector<ck::index_t> filter_spatial_lengths,
+                                     std::vector<ck::index_t> output_spatial_lengths,
+                                     std::vector<ck::index_t> conv_filter_strides,
+                                     std::vector<ck::index_t> conv_filter_dilations,
+                                     std::vector<ck::index_t> input_left_pads,
+                                     std::vector<ck::index_t> input_right_pads)
+{
+    const ck::index_t Y = filter_spatial_lengths[0];
+    const ck::index_t X = filter_spatial_lengths[1];
+
+    const ck::index_t Hi = input_spatial_lengths[0];
+    const ck::index_t Wi = input_spatial_lengths[1];
+
+    const ck::index_t Ho = output_spatial_lengths[0];
+    const ck::index_t Wo = output_spatial_lengths[1];
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
+            using namespace ck::literals;
+
+            if constexpr(is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
+                         is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
+                         is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
+            {
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, H * W, W, 1_uz});
+            }
+            else if constexpr(is_same<decltype(layout), tensor_layout::convolution::NHWC>::value ||
+                              is_same<decltype(layout), tensor_layout::convolution::KYXC>::value ||
+                              is_same<decltype(layout), tensor_layout::convolution::NHWK>::value)
+            {
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, 1_uz, W * C_, C_});
+            }
+        };
+
+    Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
+    Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
+    Tensor<OutDataType> out_n_k_ho_wo_host_result(
+        f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
+    Tensor<OutDataType> out_n_k_ho_wo_device_result(
+        f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
+
+    // bias: assume contiguous 1d vector
+    Tensor<OutDataType> bias_k({K});
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
+    std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo_host_result.mDesc << std::endl;
+    std::cout << "bias_k: " << bias_k.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        bias_k.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        break;
+    default:
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+        bias_k.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+    }
+
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    if(do_verification)
+    {
+        using ReferenceConvFwdInstance =
+            ck::tensor_operation::host::ReferenceConvFwd_Bias_Activation<InDataType,
+                                                                         WeiDataType,
+                                                                         OutDataType,
+                                                                         InElementOp,
+                                                                         WeiElementOp,
+                                                                         OutElementOp>;
+
+        auto ref_conv    = ReferenceConvFwdInstance{};
+        auto ref_invoker = ref_conv.MakeInvoker();
+
+        auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
+                                                  wei_k_c_y_x,
+                                                  out_n_k_ho_wo_host_result,
+                                                  bias_k,
+                                                  conv_filter_strides,
+                                                  conv_filter_dilations,
+                                                  input_left_pads,
+                                                  input_right_pads,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+        ref_invoker.Run(ref_argument);
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) *
+                             out_n_k_ho_wo_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(OutDataType) * bias_k.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+    wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
+    bias_device_buf.ToDevice(bias_k.mData.data());
+
+    using DeviceConvFwdBiasReluPtr = ck::tensor_operation::device::
+        DeviceConvFwdBiasActivationPtr<InElementOp, WeiElementOp, OutElementOp>;
+
+    // add device operator instances
+    std::vector<DeviceConvFwdBiasReluPtr> op_ptrs;
+
+    if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::half_t> &&
+                 ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::half_t> &&
+                 ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::half_t>)
+    {
+        ck::tensor_operation::device::instance::
+            add_device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances(op_ptrs);
+    }
+
+    if(op_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device Conv instance found");
+    }
+
+    std::string best_conv_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device Conv instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<const InDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<const WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            static_cast<const OutDataType*>(bias_device_buf.GetDeviceBuffer()),
+            N,
+            K,
+            C,
+            input_spatial_lengths,
+            filter_spatial_lengths,
+            output_spatial_lengths,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads,
+            in_element_op,
+            wei_element_op,
+            out_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string conv_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
+
+            std::size_t num_btype =
+                sizeof(InDataType) * (N * C * Hi * Wi) + sizeof(WeiDataType) * (K * C * Y * X) +
+                sizeof(OutDataType) * (N * K * Ho * Wo) + sizeof(OutDataType) * (K);
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << conv_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_conv_name  = conv_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
+
+                ck::utils::check_err(out_n_k_ho_wo_device_result, out_n_k_ho_wo_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "wei: ", wei_k_c_y_x.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "out_host  : ", out_n_k_ho_wo_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "out_device: ", out_n_k_ho_wo_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_conv_fwd_impl.hpp b/profiler/include/profiler/profile_conv_fwd_impl.hpp
new file mode 100644
index 0000000000..bc2eb25797
--- /dev/null
+++ b/profiler/include/profiler/profile_conv_fwd_impl.hpp
@@ -0,0 +1,221 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/convolution_forward.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+bool profile_conv_fwd_impl(int do_verification,
+                           int init_method,
+                           bool do_log,
+                           bool time_kernel,
+                           const ck::utils::conv::ConvParam& conv_param)
+{
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    Tensor<InDataType> input(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> weight(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> host_output(out_g_n_k_wos_desc);
+    Tensor<OutDataType> device_output(out_g_n_k_wos_desc);
+
+    std::cout << "input: " << input.mDesc << std::endl;
+    std::cout << "weight: " << weight.mDesc << std::endl;
+    std::cout << "output: " << host_output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        weight.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        weight.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * weight.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * device_output.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(input.mData.data());
+    wei_device_buf.ToDevice(weight.mData.data());
+
+    // run reference op
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp>{};
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(input,
+                                                  weight,
+                                                  host_output,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        // init host output to zero
+        host_output.SetZero();
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceConvFwd<NDimSpatial,
+                                                                 InLayout,
+                                                                 WeiLayout,
+                                                                 OutLayout,
+                                                                 InDataType,
+                                                                 WeiDataType,
+                                                                 OutDataType,
+                                                                 InElementOp,
+                                                                 WeiElementOp,
+                                                                 OutElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    float best_avg_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    bool pass = true;
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                        static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                        conv_param.N_,
+                                        conv_param.K_,
+                                        conv_param.C_,
+                                        conv_param.input_spatial_lengths_,
+                                        conv_param.filter_spatial_lengths_,
+                                        conv_param.GetOutputSpatialLengths(),
+                                        conv_param.conv_filter_strides_,
+                                        conv_param.conv_filter_dilations_,
+                                        conv_param.input_left_pads_,
+                                        conv_param.input_right_pads_,
+                                        in_element_op,
+                                        wei_element_op,
+                                        out_element_op);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init output to zero before profiling next kernel
+            out_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = conv_param.GetFlops();
+            std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+            float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
+
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                out_device_buf.FromDevice(device_output.mData.data());
+
+                pass = pass & ck::utils::check_err(device_output, host_output);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "input : ", input.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "weight: ", weight.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "host_output  : ", host_output.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "device_output: ", device_output.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best configuration parameters:"
+              << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
+              << "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_elementwise_layernorm_impl.hpp b/profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
new file mode 100644
index 0000000000..1fd9c81109
--- /dev/null
+++ b/profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
@@ -0,0 +1,266 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
+void host_elementwise2D(HostTensorC& C,
+                        const HostTensorA& A,
+                        const HostTensorB& B,
+                        const std::vector<std::size_t>& shape,
+                        Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0))>;
+
+    for(std::size_t m = 0; m < shape[0]; ++m)
+        for(std::size_t n = 0; n < shape[1]; ++n)
+        {
+            auto a_val  = A(m, n);
+            auto b_val  = B(m, n);
+            ctype c_val = 0;
+            functor(c_val, a_val, b_val);
+            C(m, n) = c_val;
+        }
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType>
+bool profile_elementwise_layernorm_impl(int do_verification,
+                                        int init_method,
+                                        bool do_log,
+                                        bool time_kernel,
+                                        std::vector<index_t> length)
+{
+    using Add         = ck::tensor_operation::element_wise::Add;
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    if(length.size() != 2)
+        return false;
+
+    index_t M      = length[0];
+    index_t N      = length[1];
+    index_t Stride = N;
+
+    constexpr int Rank         = 2;
+    constexpr int NumReduceDim = 1;
+
+    std::vector<index_t> reduce_dim      = {1};
+    std::vector<index_t> gammaBetaLength = {N};
+    std::vector<index_t> gammaBetaStride = {0, 1};
+
+    auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
+        using namespace ck::literals;
+
+        return HostTensorDescriptor({row, col}, {stride, 1_uz});
+    };
+
+    Tensor<ADataType> a(length);
+    Tensor<BDataType> b(length);
+    Tensor<GammaDataType> gamma(gammaBetaLength);
+    Tensor<BetaDataType> beta(gammaBetaLength);
+    Tensor<YDataType> y(length);
+    Tensor<YDataType> host_y(length);
+
+    switch(init_method)
+    {
+    case 0:
+        a.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
+        b.GenerateTensorValue(GeneratorTensor_1<BDataType>{});
+        gamma.GenerateTensorValue(GeneratorTensor_1<GammaDataType>{});
+        beta.GenerateTensorValue(GeneratorTensor_1<BetaDataType>{});
+        break;
+    case 1:
+        a.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        gamma.GenerateTensorValue(GeneratorTensor_2<GammaDataType>{-5, 5});
+        beta.GenerateTensorValue(GeneratorTensor_2<BetaDataType>{-5, 5});
+        break;
+    default:
+        a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0, 1});
+        b.GenerateTensorValue(GeneratorTensor_3<BDataType>{0, 1});
+        gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-0.5, 0.5});
+        beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a_dev(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_dev(sizeof(ADataType) * b.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+
+    a_dev.ToDevice(a.mData.data());
+    b_dev.ToDevice(b.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    beta_dev.ToDevice(beta.mData.data());
+
+    std::array<const void*, 2> input = {a_dev.GetDeviceBuffer(), b_dev.GetDeviceBuffer()};
+
+    // add device normalization instances
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwiseNormalization<
+        ck::Tuple<ADataType, BDataType>,
+        GammaDataType,
+        BetaDataType,
+        AccDataType,
+        YDataType,
+        Add,
+        PassThrough,
+        2,
+        1>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using XDataType             = ADataType;
+        std::vector<std::size_t> mn = {static_cast<unsigned long>(M),
+                                       static_cast<unsigned long>(N)};
+        Tensor<XDataType> x(f_host_tensor_descriptor2d(M, N, Stride));
+        host_elementwise2D<Tensor<ADataType>, Tensor<BDataType>, Tensor<XDataType>, Add>(
+            x, a, b, mn, Add{});
+
+        using ReferenceInstance = ck::tensor_operation::host::ReferenceLayernorm<XDataType,
+                                                                                 GammaDataType,
+                                                                                 BetaDataType,
+                                                                                 YDataType,
+                                                                                 AccDataType,
+                                                                                 PassThrough,
+                                                                                 Rank,
+                                                                                 NumReduceDim>;
+
+        ReferenceInstance ref;
+        auto ref_argument =
+            ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, {M, N}, {1}, 1e-4);
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+    }
+
+    int num_kernel = 0;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            length,
+            {
+                std::vector<ck::index_t>{a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end()},
+                std::vector<ck::index_t>{b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end()},
+            },
+            gammaBetaStride,
+            gammaBetaStride,
+            std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
+            reduce_dim,
+            1e-4,
+            input,
+            gamma_dev.GetDeviceBuffer(),
+            beta_dev.GetDeviceBuffer(),
+            y_dev.GetDeviceBuffer(),
+            Add{},
+            PassThrough{});
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+        }
+        else
+        {
+            continue;
+        }
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes = a.mDesc.GetElementSize() * sizeof(ADataType) +
+                                b.mDesc.GetElementSize() * sizeof(BDataType) +
+                                gamma.mDesc.GetElementSize() * sizeof(GammaDataType) +
+                                beta.mDesc.GetElementSize() * sizeof(BetaDataType) +
+                                y.mDesc.GetElementSize() * sizeof(YDataType);
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            y_dev.FromDevice(y.mData.data());
+
+            bool pass =
+                ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 1e-3, 1e-3);
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(std::cout << "a  : ", a.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "b  : ", b.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "host_y  : ", host_y.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "y  : ", y.mData, ",") << std::endl;
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "lengths = [", length, ", ") << "]." << std::endl;
+                return false;
+            }
+            else
+            {
+                if(time_kernel)
+                    std::cout << "pass" << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", length, ",") << ", ";
+        std::cout << "num_kernel = " << num_kernel << ", best perf = " << best_avg_time << " ms, "
+                  << best_gb_per_sec << " GB/s, " << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is tested" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp b/profiler/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp
new file mode 100644
index 0000000000..81b8d8ddbf
--- /dev/null
+++ b/profiler/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp
@@ -0,0 +1,242 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout>
+bool profile_gemm_add_add_fastgelu_impl(int do_verification,
+                                        int init_method,
+                                        bool /*do_log*/,
+                                        bool time_kernel,
+                                        int M,
+                                        int N,
+                                        int K,
+                                        int StrideA,
+                                        int StrideB,
+                                        int StrideD0,
+                                        int StrideD1,
+                                        int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    using PassThrough    = ck::tensor_operation::element_wise::PassThrough;
+    using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = AddAddFastGelu;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddAddFastGelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp b/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp
new file mode 100644
index 0000000000..6f6d881c1e
--- /dev/null
+++ b/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp
@@ -0,0 +1,232 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename ELayout>
+bool profile_gemm_add_fastgelu_impl(int do_verification,
+                                    int init_method,
+                                    bool /*do_log*/,
+                                    bool time_kernel,
+                                    int M,
+                                    int N,
+                                    int K,
+                                    int StrideA,
+                                    int StrideB,
+                                    int StrideD0,
+                                    int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = AddFastGelu;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddFastGelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 1>{StrideD0},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_gemm_add_multiply_impl.hpp b/profiler/include/profiler/profile_gemm_add_multiply_impl.hpp
new file mode 100644
index 0000000000..25871dfb2e
--- /dev/null
+++ b/profiler/include/profiler/profile_gemm_add_multiply_impl.hpp
@@ -0,0 +1,242 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout>
+bool profile_gemm_add_multiply_impl(int do_verification,
+                                    int init_method,
+                                    bool /*do_log*/,
+                                    bool time_kernel,
+                                    int M,
+                                    int N,
+                                    int K,
+                                    int StrideA,
+                                    int StrideB,
+                                    int StrideD0,
+                                    int StrideD1,
+                                    int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-1, 1});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using AddMultiply = ck::tensor_operation::element_wise::AddMultiply;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = AddMultiply;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
+                                                          BLayout,
+                                                          ck::Tuple<D0Layout, D1Layout>,
+                                                          ELayout,
+                                                          ADataType,
+                                                          BDataType,
+                                                          ck::Tuple<D0DataType, D1DataType>,
+                                                          EDataType,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          CDEElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_gemm_add_relu_add_layernorm_impl.hpp b/profiler/include/profiler/profile_gemm_add_relu_add_layernorm_impl.hpp
new file mode 100644
index 0000000000..4c3d0a0450
--- /dev/null
+++ b/profiler/include/profiler/profile_gemm_add_relu_add_layernorm_impl.hpp
@@ -0,0 +1,346 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EMeanVarDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename HDataType,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp,
+          typename HElementOp>
+void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
+                         const Tensor<ADataType>& a_m_k,
+                         const Tensor<BDataType>& b_k_n,
+                         const Tensor<D0DataType>& d0_m_n,
+                         const Tensor<D1DataType>& d1_m_n,
+                         const Tensor<GammaDataType>& gamma_n,
+                         const Tensor<BetaDataType>& beta_n,
+                         AElementOp a_element_op,
+                         BElementOp b_element_op,
+                         CDEElementOp cde_element_op,
+                         HElementOp h_element_op,
+                         int M,
+                         int N,
+                         AccDataType epsilon = 1e-5)
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    using ReferenceGemm = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                    BDataType,
+                                                                    AccDataType,
+                                                                    AccDataType,
+                                                                    AElementOp,
+                                                                    BElementOp,
+                                                                    PassThrough>;
+
+    using ReferenceLayernorm = ck::tensor_operation::host::ReferenceLayernorm<EMeanVarDataType,
+                                                                              GammaDataType,
+                                                                              BetaDataType,
+                                                                              HDataType,
+                                                                              AccDataType,
+                                                                              HElementOp,
+                                                                              2,
+                                                                              1>;
+
+    Tensor<EMeanVarDataType> e_m_n(HostTensorDescriptor{M, N});
+    Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
+
+    auto ref_gemm         = ReferenceGemm{};
+    auto ref_gemm_invoker = ref_gemm.MakeInvoker();
+
+    auto ref_gemm_argument =
+        ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+    ref_gemm_invoker.Run(ref_gemm_argument);
+
+    for(int n = 0; n < N; ++n)
+    {
+        for(int m = 0; m < M; ++m)
+        {
+            AccDataType e  = static_cast<AccDataType>(e_m_n(m, n));
+            AccDataType d0 = static_cast<AccDataType>(d0_m_n(m, n));
+            AccDataType d1 = static_cast<AccDataType>(d1_m_n(m, n));
+            cde_element_op(e, c_m_n(m, n), d0, d1);
+            e_m_n(m, n) = static_cast<EMeanVarDataType>(e);
+        }
+    }
+
+    ReferenceLayernorm ref_layernorm;
+    auto ref_layernorm_invoker = ref_layernorm.MakeInvoker();
+
+    auto ref_layernorm_argument = ref_layernorm.MakeArgument(
+        e_m_n, gamma_n, beta_n, h_m_n, h_element_op, {M, N}, {1}, epsilon);
+    ref_layernorm_invoker.Run(ref_layernorm_argument);
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EMeanVarDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename HDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename HLayout>
+bool profile_gemm_add_relu_add_layernorm_impl(int do_verification,
+                                              int init_method,
+                                              bool /*do_log*/,
+                                              bool time_kernel,
+                                              int M,
+                                              int N,
+                                              int K,
+                                              int StrideA,
+                                              int StrideB,
+                                              int StrideD0,
+                                              int StrideD1,
+                                              int StrideH,
+                                              AccDataType epsilon = 1e-5)
+{
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    auto f_host_tensor_descriptor2d =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if constexpr(std::is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+    Tensor<D1DataType> d0_m_n(f_host_tensor_descriptor2d(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor2d(M, N, StrideD1, D1Layout{}));
+    Tensor<GammaDataType> gamma_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<BetaDataType> beta_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<HDataType> h_m_n(f_host_tensor_descriptor2d(M, N, StrideH, HLayout{}));
+    Tensor<HDataType> h_m_n_host(f_host_tensor_descriptor2d(M, N, StrideH, HLayout{}));
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-1, 1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-1, 1});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-1, 1});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-1, 1});
+        gamma_n.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-1, 1});
+        beta_n.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-1, 1});
+        break;
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = AddReluAdd;
+    using HElementOp   = PassThrough;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+    const auto h_element_op   = HElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleDLayernorm<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        HLayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        GammaDataType,
+        BetaDataType,
+        HDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddReluAdd,
+        ck::tensor_operation::element_wise::PassThrough>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        host_gemm_layernorm<ADataType,
+                            BDataType,
+                            AccDataType,
+                            D0DataType,
+                            D1DataType,
+                            EMeanVarDataType,
+                            GammaDataType,
+                            BetaDataType,
+                            HDataType>(h_m_n_host,
+                                       a_m_k,
+                                       b_k_n,
+                                       d0_m_n,
+                                       d1_m_n,
+                                       gamma_n,
+                                       beta_n,
+                                       a_element_op,
+                                       b_element_op,
+                                       cde_element_op,
+                                       h_element_op,
+                                       M,
+                                       N,
+                                       epsilon);
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_device_buf(sizeof(GammaDataType) * gamma_n.mDesc.GetElementSpaceSize());
+    DeviceMem beta_device_buf(sizeof(BetaDataType) * beta_n.mDesc.GetElementSpaceSize());
+    DeviceMem h_device_buf(sizeof(HDataType) * h_m_n.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
+    gamma_device_buf.ToDevice(gamma_n.mData.data());
+    beta_device_buf.ToDevice(beta_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    bool pass      = true;
+    int num_kernel = 0;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            {d0_m_n_device_buf.GetDeviceBuffer(), d1_m_n_device_buf.GetDeviceBuffer()},
+            gamma_device_buf.GetDeviceBuffer(),
+            beta_device_buf.GetDeviceBuffer(),
+            h_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            {StrideD0, StrideD1},
+            StrideH,
+            epsilon,
+            a_element_op,
+            b_element_op,
+            cde_element_op,
+            h_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            DeviceMem workspace_dev(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+            // re-init E to zero before profiling a kernel
+            h_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t num_byte =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                (sizeof(D0DataType) + sizeof(D1DataType) + sizeof(HDataType)) * M * N +
+                (sizeof(GammaDataType) + sizeof(BetaDataType)) * N;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            if(time_kernel)
+                std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec
+                          << " GB/s, " << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                h_device_buf.FromDevice(h_m_n.mData.data());
+
+                pass = pass && ck::utils::check_err(
+                                   h_m_n, h_m_n_host, "Error: Incorrect results h_m_n", 1e-2, 1e-2);
+            }
+        }
+        else
+        {
+            if(time_kernel)
+                std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        pass = false;
+    }
+    else
+    {
+        if(time_kernel)
+            std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                      << best_op_name << std::endl;
+    }
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_gemm_bias_add_reduce_impl.hpp b/profiler/include/profiler/profile_gemm_bias_add_reduce_impl.hpp
new file mode 100644
index 0000000000..c0ffea8a32
--- /dev/null
+++ b/profiler/include/profiler/profile_gemm_bias_add_reduce_impl.hpp
@@ -0,0 +1,384 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32                 = float;
+using F16                 = ck::half_t;
+using ReducePtrsGlobal    = ck::Tuple<F32*, F32*>;
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using DeviceGemmBiasAddReduceNoOpPtr =
+    ck::tensor_operation::device::DeviceGemmReducePtr<1, ReducePtrsGlobal::Size()>;
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+    std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+    std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
+    std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
+    std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename BiasDataType,
+          typename D0DataType,
+          typename ReduceDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+void profile_gemm_bias_add_reduce_impl(int do_verification,
+                                       int init_method,
+                                       bool do_log,
+                                       bool time_kernel,
+                                       int M,
+                                       int N,
+                                       int K,
+                                       int StrideA,
+                                       int StrideB,
+                                       int StrideC,
+                                       int StrideD0)
+{
+    auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
+        return HostTensorDescriptor({len}, {stride});
+    };
+
+    auto f_host_tensor_descriptor2d =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor2d(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
+
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+    Tensor<BiasDataType> bias_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> reduce0_m_host_result({M});
+    Tensor<ReduceDataType> reduce1_m_host_result({M});
+
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> reduce0_m_device_result({M});
+    Tensor<ReduceDataType> reduce1_m_device_result({M});
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+    std::cout << "reduce0_m: " << reduce0_m_host_result.mDesc << std::endl;
+    std::cout << "reduce1_m: " << reduce1_m_host_result.mDesc << std::endl;
+
+    std::size_t num_thread = 1;
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        std::srand(0);
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        bias_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        break;
+    default:
+        std::srand(0);
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+        bias_n.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5}, num_thread);
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+    }
+
+    using PassThrough           = ck::tensor_operation::element_wise::PassThrough;
+    using AElementOp            = PassThrough;
+    using BElementOp            = PassThrough;
+    using CElementOp            = PassThrough;
+    using D0ElementOp           = PassThrough;
+    using ReduceOp0             = ck::reduce::Add;
+    using ReduceOp1             = ck::reduce::Add;
+    using UnaryDivElementOp     = ck::tensor_operation::element_wise::UnaryDivide;
+    using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using UnarySquareElementOp  = ck::tensor_operation::element_wise::UnarySquare;
+
+    auto a_element_op                     = AElementOp{};
+    auto b_element_op                     = BElementOp{};
+    auto c_element_op                     = CElementOp{};
+    std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
+
+    auto d0_element_op    = D0ElementOp{};
+    const auto reduce0_op = ReduceOp0{};
+    const auto reduce1_op = ReduceOp1{};
+
+    auto passthrough                            = UnaryIdenticElementOp{};
+    auto square                                 = UnarySquareElementOp{};
+    auto div                                    = UnaryDivElementOp{N};
+    std::array<void*, 2> reduce_in_element_ops  = {&passthrough, &square};
+    std::array<void*, 2> reduce_out_element_ops = {&div, &div};
+
+    if(do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                ReduceDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
+
+        using ReduceAccDataType = ReduceDataType;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+            for(int n = 0; n < N; ++n)
+            {
+                ReduceAccDataType acc = static_cast<ReduceAccDataType>(c_m_n_host_result(m, n)) +
+                                        static_cast<ReduceAccDataType>(bias_n(n));
+
+                ReduceAccDataType d0 = static_cast<ReduceAccDataType>(d0_m_n(m, n));
+                c_element_op(acc, acc);
+                d0_element_op(d0, d0);
+                acc += d0;
+                c_m_n_host_result(m, n) = static_cast<CDataType>(acc);
+            }
+
+        for(int m = 0; m < M; ++m)
+        {
+            auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+            auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
+
+            for(int n = 0; n < N; ++n)
+            {
+                ReduceAccDataType d0_val =
+                    ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
+                ReduceAccDataType d1_val;
+
+                square(d1_val, d0_val);
+                reduce0_op(reduce0_acc, d0_val);
+                reduce1_op(reduce1_acc, d1_val);
+            }
+
+            div(reduce0_acc, reduce0_acc);
+            div(reduce1_acc, reduce1_acc);
+            reduce0_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce0_acc);
+            reduce1_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce1_acc);
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
+                                 reduce0_m_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
+                                 reduce1_m_device_result.mDesc.GetElementSpaceSize());
+
+    std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
+                                      reduce1_device_buf.GetDeviceBuffer()};
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    bias_device_buf.ToDevice(bias_n.mData.data());
+    d0_device_buf.ToDevice(d0_m_n.mData.data());
+
+    // add device GEMM instances
+    std::vector<ck::tensor_operation::device::instance::DeviceGemmBiasAddReduceNoOpPtr> gemm_ptrs;
+
+    if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
+                 is_same<CDataType, half_t>::value)
+    {
+        if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
+                    gemm_ptrs);
+        }
+    }
+
+    if(gemm_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device GEMM instance found");
+    }
+
+    std::string best_gemm_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device GEMM instances
+    for(auto& gemm_ptr : gemm_ptrs)
+    {
+        auto argument_ptr = gemm_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                          b_device_buf.GetDeviceBuffer(),
+                                                          bias_device_buf.GetDeviceBuffer(),
+                                                          {d0_device_buf.GetDeviceBuffer()},
+                                                          c_device_buf.GetDeviceBuffer(),
+                                                          p_reduces,
+                                                          M,
+                                                          N,
+                                                          K,
+                                                          StrideA,
+                                                          StrideB,
+                                                          StrideC,
+                                                          {StrideD0},
+                                                          gemm_element_ops,
+                                                          {&d0_element_op},
+                                                          reduce_in_element_ops,
+                                                          reduce_out_element_ops);
+
+        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
+
+        if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // init DO, D1 to 0
+            reduce0_device_buf.SetZero();
+            reduce1_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::string gemm_name = gemm_ptr->GetTypeString();
+
+            std::size_t flop = std::size_t(2) * M * N * K + std::size_t(2) * M * N;
+
+            std::size_t num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                   sizeof(CDataType) * M * N + sizeof(BiasDataType) * M * N +
+                                   sizeof(D0DataType) * M * N + sizeof(ReduceDataType) * M +
+                                   sizeof(ReduceDataType) * M;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << gemm_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_gemm_name  = gemm_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+                reduce0_device_buf.FromDevice(reduce0_m_device_result.mData.data());
+                reduce1_device_buf.FromDevice(reduce1_m_device_result.mData.data());
+
+                ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+                ck::utils::check_err(reduce0_m_device_result, reduce0_m_host_result);
+                ck::utils::check_err(reduce1_m_device_result, reduce1_m_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host: ", c_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d0_host: ", reduce0_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d0_device: ", reduce0_m_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d1_host: ", reduce1_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d1_device: ", reduce1_m_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << "does not support this GEMM problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_gemm_bilinear_impl.hpp b/profiler/include/profiler/profile_gemm_bilinear_impl.hpp
new file mode 100644
index 0000000000..b540e938b5
--- /dev/null
+++ b/profiler/include/profiler/profile_gemm_bilinear_impl.hpp
@@ -0,0 +1,234 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename DDataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename DLayout,
+          typename ELayout>
+bool profile_gemm_bilinear_impl(int do_verification,
+                                int init_method,
+                                bool /*do_log*/,
+                                bool time_kernel,
+                                int M,
+                                int N,
+                                int K,
+                                int StrideA,
+                                int StrideB,
+                                int StrideD,
+                                int StrideE,
+                                float alpha,
+                                float beta)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DLayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{0.0, 1.0});
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using Bilinear    = ck::tensor_operation::element_wise::Bilinear;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = Bilinear;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{alpha, beta};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<DLayout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<DDataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Bilinear>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_m_n_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d_m_n_device_buf.ToDevice(d_m_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 1>{d_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 1>{StrideD},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_gemm_fastgelu_impl.hpp b/profiler/include/profiler/profile_gemm_fastgelu_impl.hpp
new file mode 100644
index 0000000000..3893f8cdc7
--- /dev/null
+++ b/profiler/include/profiler/profile_gemm_fastgelu_impl.hpp
@@ -0,0 +1,222 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout>
+bool profile_gemm_fastgelu_impl(int do_verification,
+                                int init_method,
+                                bool /*do_log*/,
+                                bool time_kernel,
+                                int M,
+                                int N,
+                                int K,
+                                int StrideA,
+                                int StrideB,
+                                int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = FastGelu;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::FastGelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                        b_device_buf.GetDeviceBuffer(),
+                                                        std::array<const void*, 0>{},
+                                                        e_device_buf.GetDeviceBuffer(),
+                                                        M,
+                                                        N,
+                                                        K,
+                                                        StrideA,
+                                                        StrideB,
+                                                        std::array<ck::index_t, 0>{},
+                                                        StrideE,
+                                                        a_element_op,
+                                                        b_element_op,
+                                                        cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_gemm_impl.hpp b/profiler/include/profiler/profile_gemm_impl.hpp
new file mode 100644
index 0000000000..eaab5dbcc2
--- /dev/null
+++ b/profiler/include/profiler/profile_gemm_impl.hpp
@@ -0,0 +1,254 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType>
+int profile_gemm_impl(int do_verification,
+                      int init_method,
+                      bool do_log,
+                      bool time_kernel,
+                      int M,
+                      int N,
+                      int K,
+                      int StrideA,
+                      int StrideB,
+                      int StrideC)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemm<ALayout,
+                                                              BLayout,
+                                                              CLayout,
+                                                              ADataType,
+                                                              BDataType,
+                                                              CDataType,
+                                                              AElementOp,
+                                                              BElementOp,
+                                                              CElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // Run reference op
+    if(do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
+
+        auto ref_op      = ReferenceGemmInstance{};
+        auto ref_invoker = ref_op.MakeInvoker();
+
+        auto ref_argument = ref_op.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    std::string best_op_name;
+    float best_avg_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                        static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                        M,
+                                        N,
+                                        K,
+                                        StrideA,
+                                        StrideB,
+                                        StrideC,
+                                        a_element_op,
+                                        b_element_op,
+                                        c_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init C to zero before profiling next kernel
+            c_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
+
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host  : ", c_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    if constexpr(is_same<CDataType, float>::value)
+    {
+        std::cout << "Best Perf for datatype = f32";
+    }
+    else if constexpr(is_same<CDataType, half_t>::value)
+    {
+        std::cout << "Best Perf for datatype = f16";
+    }
+    else if constexpr(is_same<CDataType, bhalf_t>::value)
+    {
+        std::cout << "Best Perf for datatype = bf16";
+    }
+    else if constexpr(is_same<CDataType, int8_t>::value)
+    {
+        std::cout << "Best Perf for datatype = int8";
+    }
+
+    if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " ALayout =  RowMajor";
+    }
+    else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " ALayout =  ColumnMajor";
+    }
+
+    if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " BLayout =  RowMajor";
+    }
+    else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " BLayout =  ColumnMajor";
+    }
+
+    std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
+              << " StrideB = " << StrideB << " StrideC = " << StrideC << " : " << best_avg_time
+              << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    return pass ? 0 : 1;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_gemm_multiply_add_impl.hpp b/profiler/include/profiler/profile_gemm_multiply_add_impl.hpp
new file mode 100644
index 0000000000..5fc92d1f8d
--- /dev/null
+++ b/profiler/include/profiler/profile_gemm_multiply_add_impl.hpp
@@ -0,0 +1,242 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout>
+bool profile_gemm_multiply_add_impl(int do_verification,
+                                    int init_method,
+                                    bool /*do_log*/,
+                                    bool time_kernel,
+                                    int M,
+                                    int N,
+                                    int K,
+                                    int StrideA,
+                                    int StrideB,
+                                    int StrideD0,
+                                    int StrideD1,
+                                    int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-1, 1});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = MultiplyAdd;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
+                                                          BLayout,
+                                                          ck::Tuple<D0Layout, D1Layout>,
+                                                          ELayout,
+                                                          ADataType,
+                                                          BDataType,
+                                                          ck::Tuple<D0DataType, D1DataType>,
+                                                          EDataType,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          CDEElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_gemm_reduce_impl.hpp b/profiler/include/profiler/profile_gemm_reduce_impl.hpp
new file mode 100644
index 0000000000..ff801e8afd
--- /dev/null
+++ b/profiler/include/profiler/profile_gemm_reduce_impl.hpp
@@ -0,0 +1,353 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/utility/reduction_operator.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F32                 = float;
+using F16                 = ck::half_t;
+using ReducePtrsGlobal    = ck::Tuple<F32*, F32*>;
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using DeviceGemmReduceNoOpPtr =
+    ck::tensor_operation::device::DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>;
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances(
+    std::vector<DeviceGemmReduceNoOpPtr>&);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ReduceDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_gemm_reduce_impl(int do_verification,
+                              int init_method,
+                              bool do_log,
+                              bool time_kernel,
+                              int M,
+                              int N,
+                              int K,
+                              int StrideA,
+                              int StrideB,
+                              int StrideC)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> reduce0_m_host_result({M});
+    Tensor<ReduceDataType> reduce1_m_host_result({M});
+
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> reduce0_m_device_result({M});
+    Tensor<ReduceDataType> reduce1_m_device_result({M});
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+    std::cout << "reduce0_m: " << reduce0_m_host_result.mDesc << std::endl;
+    std::cout << "reduce1_m: " << reduce1_m_host_result.mDesc << std::endl;
+
+    std::size_t num_thread = 1;
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        std::srand(0);
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        break;
+    default:
+        std::srand(0);
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+    }
+
+    using AElementOp            = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp            = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp            = ck::tensor_operation::element_wise::PassThrough;
+    using ReduceOp0             = ck::reduce::Add;
+    using ReduceOp1             = ck::reduce::Add;
+    using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using UnarySquareElementOp  = ck::tensor_operation::element_wise::UnarySquare;
+    using UnaryDivElementOp     = ck::tensor_operation::element_wise::UnaryDivide;
+
+    auto a_element_op                     = AElementOp{};
+    auto b_element_op                     = BElementOp{};
+    auto c_element_op                     = CElementOp{};
+    std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
+
+    const auto reduce0_op = ReduceOp0{};
+    const auto reduce1_op = ReduceOp1{};
+
+    auto passthrough                            = UnaryIdenticElementOp{};
+    auto square                                 = UnarySquareElementOp{};
+    auto div                                    = UnaryDivElementOp{N};
+    std::array<void*, 2> reduce_in_element_ops  = {&passthrough, &square};
+    std::array<void*, 2> reduce_out_element_ops = {&div, &div};
+
+    if(do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                ReduceDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
+
+        using ReduceAccDataType = ReduceDataType;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+            auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
+
+            for(int n = 0; n < N; ++n)
+            {
+                ReduceAccDataType d0_val =
+                    ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
+                ReduceAccDataType d1_val;
+
+                square(d1_val, d0_val);
+                reduce0_op(reduce0_acc, d0_val);
+                reduce1_op(reduce1_acc, d1_val);
+            }
+
+            div(reduce0_acc, reduce0_acc);
+            div(reduce1_acc, reduce1_acc);
+            reduce0_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce0_acc);
+            reduce1_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce1_acc);
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
+                                 reduce0_m_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
+                                 reduce1_m_device_result.mDesc.GetElementSpaceSize());
+
+    std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
+                                      reduce1_device_buf.GetDeviceBuffer()};
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    // add device GEMM instances
+    std::vector<ck::tensor_operation::device::instance::DeviceGemmReduceNoOpPtr> gemm_ptrs;
+
+    if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
+                 is_same<CDataType, half_t>::value)
+    {
+        if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                     is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances(
+                    gemm_ptrs);
+        }
+        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
+                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
+        {
+            ck::tensor_operation::device::instance::
+                add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances(
+                    gemm_ptrs);
+        }
+    }
+
+    if(gemm_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device GEMM instance found");
+    }
+
+    std::string best_gemm_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device GEMM instances
+    for(auto& gemm_ptr : gemm_ptrs)
+    {
+        auto argument_ptr = gemm_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                          b_device_buf.GetDeviceBuffer(),
+                                                          nullptr,
+                                                          {},
+                                                          c_device_buf.GetDeviceBuffer(),
+                                                          p_reduces,
+                                                          M,
+                                                          N,
+                                                          K,
+                                                          StrideA,
+                                                          StrideB,
+                                                          StrideC,
+                                                          {},
+                                                          gemm_element_ops,
+                                                          {},
+                                                          reduce_in_element_ops,
+                                                          reduce_out_element_ops);
+
+        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
+
+        if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // init DO, D1 to 0
+            reduce0_device_buf.SetZero();
+            reduce1_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::string gemm_name = gemm_ptr->GetTypeString();
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                    sizeof(CDataType) * M * N + sizeof(CDataType) * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                      << " GB/s, " << gemm_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_gemm_name  = gemm_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+                reduce0_device_buf.FromDevice(reduce0_m_device_result.mData.data());
+                reduce1_device_buf.FromDevice(reduce1_m_device_result.mData.data());
+
+                ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+                ck::utils::check_err(reduce0_m_device_result, reduce0_m_host_result);
+                ck::utils::check_err(reduce1_m_device_result, reduce1_m_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host: ", c_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d0_host: ", reduce0_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d0_device: ", reduce0_m_device_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d1_host: ", reduce1_m_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(
+                        std::cout << "d1_device: ", reduce1_m_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << "does not support this GEMM problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_gemm_splitk_impl.hpp b/profiler/include/profiler/profile_gemm_splitk_impl.hpp
new file mode 100644
index 0000000000..fb68bb8811
--- /dev/null
+++ b/profiler/include/profiler/profile_gemm_splitk_impl.hpp
@@ -0,0 +1,293 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_gemm_splitk_impl(int do_verification,
+                              int init_method,
+                              bool do_log,
+                              bool time_kernel,
+                              int M,
+                              int N,
+                              int K,
+                              int StrideA,
+                              int StrideB,
+                              int StrideC,
+                              int KBatch)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-1, 2});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmSplitK<ALayout,
+                                                                    BLayout,
+                                                                    CLayout,
+                                                                    ADataType,
+                                                                    BDataType,
+                                                                    CDataType,
+                                                                    AElementOp,
+                                                                    BElementOp,
+                                                                    CElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // Run reference GEMM
+    if(do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+    float best_kbatch     = 0;
+
+    // profile device GEMM instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        std::vector<int> kbatch_list = {1,  2,  4,  8,  12, 16,  20,  24,  32,  36,  40, 60,
+                                        64, 72, 80, 88, 96, 128, 144, 160, 176, 192, 256};
+
+        if(KBatch > 0)
+        {
+            kbatch_list = {KBatch};
+        }
+
+        for(std::size_t i = 0; i < kbatch_list.size(); i++)
+        {
+            auto kbatch_curr = kbatch_list[i];
+
+            auto argument_ptr =
+                op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                            static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                            static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                            M,
+                                            N,
+                                            K,
+                                            StrideA,
+                                            StrideB,
+                                            StrideC,
+                                            a_element_op,
+                                            b_element_op,
+                                            c_element_op,
+                                            kbatch_curr);
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+
+                // re-init C to zero before profiling next kernel
+                c_device_buf.SetZero();
+
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+
+                if(do_verification)
+                {
+                    c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+                    pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+
+                    if(do_log)
+                    {
+                        LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
+                        LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
+                        LogRangeAsType<float>(
+                            std::cout << "c_host  : ", c_m_n_host_result.mData, ",")
+                            << std::endl;
+                        LogRangeAsType<float>(
+                            std::cout << "c_device: ", c_m_n_device_result.mData, ",")
+                            << std::endl;
+                    }
+                }
+
+                std::string op_name = op_ptr->GetTypeString();
+
+                float ave_time =
+                    invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+                std::size_t flop = std::size_t(2) * M * N * K;
+
+                std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                        sizeof(CDataType) * M * N;
+
+                float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+                float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+                std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
+                          << " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
+                          << kbatch_curr << std::endl;
+
+                // set softer tolerances for fp8
+                if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
+                             is_same_v<CDataType, f8_t>)
+                {
+                    std::string msg = "Error: Incorrect results!";
+                    double rtol     = 1e-1;
+                    double atol     = 1e-1;
+                    pass            = pass & ck::utils::check_err(
+                                      c_m_n_device_result, c_m_n_host_result, msg, rtol, atol);
+                }
+                else
+                {
+                    pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+                }
+
+                if(tflops > best_tflops)
+                {
+                    best_op_name    = op_name;
+                    best_tflops     = tflops;
+                    best_ave_time   = ave_time;
+                    best_gb_per_sec = gb_per_sec;
+                    best_kbatch     = kbatch_curr;
+                }
+            }
+            else
+            {
+                std::cout << op_ptr->GetTypeString() << " does not support this problem"
+                          << std::endl;
+            }
+        }
+    }
+
+    if constexpr(is_same<CDataType, float>::value)
+    {
+        std::cout << "Best Perf for datatype = f32";
+    }
+    else if constexpr(is_same<CDataType, half_t>::value)
+    {
+        std::cout << "Best Perf for datatype = f16";
+    }
+    else if constexpr(is_same<CDataType, bhalf_t>::value)
+    {
+        std::cout << "Best Perf for datatype = bf16";
+    }
+    else if constexpr(is_same<CDataType, int8_t>::value)
+    {
+        std::cout << "Best Perf for datatype = int8";
+    }
+
+    if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " ALayout =  RowMajor";
+    }
+    else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " ALayout =  ColumnMajor";
+    }
+
+    if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " BLayout =  RowMajor";
+    }
+    else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " BLayout =  ColumnMajor";
+    }
+
+    std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
+              << " StrideB = " << StrideB << " StrideC = " << StrideC << " KBatch = " << best_kbatch
+              << " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
+              << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_gemm_streamk_impl.hpp b/profiler/include/profiler/profile_gemm_streamk_impl.hpp
new file mode 100644
index 0000000000..71b54c1f47
--- /dev/null
+++ b/profiler/include/profiler/profile_gemm_streamk_impl.hpp
@@ -0,0 +1,267 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_streamk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_gemm_streamk_impl(int do_verification,
+                               int init_method,
+                               bool do_log,
+                               bool time_kernel,
+                               int M,
+                               int N,
+                               int K,
+                               int StrideA,
+                               int StrideB,
+                               int StrideC,
+                               uint32_t NumSKBlocks = 0xffffffff)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-3, 3});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    c_device_buf.ToDevice(c_m_n_device_result.mData.data());
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmStreamK<ALayout,
+                                                                     BLayout,
+                                                                     CLayout,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     CDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     CElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances, "
+              << (do_verification ? "with verification" : "without verification") << std::endl;
+
+    // Run reference GEMM
+    if(do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device GEMM instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                        static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                        M,
+                                        N,
+                                        K,
+                                        StrideA,
+                                        StrideB,
+                                        StrideC,
+                                        a_element_op,
+                                        b_element_op,
+                                        c_element_op,
+                                        NumSKBlocks);
+        DeviceMem workspace;
+        std::size_t workspace_size = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+        if(workspace_size != 0)
+        {
+            workspace.Realloc(workspace_size);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace.GetDeviceBuffer());
+        }
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init C to zero before profiling next kernel
+            c_device_buf.SetZero();
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+
+            if(do_verification)
+            {
+                c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+                pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host  : ", c_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    if constexpr(is_same<CDataType, float>::value)
+    {
+        std::cout << "Best Perf for datatype = f32";
+    }
+    else if constexpr(is_same<CDataType, half_t>::value)
+    {
+        std::cout << "Best Perf for datatype = f16";
+    }
+    else if constexpr(is_same<CDataType, bhalf_t>::value)
+    {
+        std::cout << "Best Perf for datatype = bf16";
+    }
+    else if constexpr(is_same<CDataType, int8_t>::value)
+    {
+        std::cout << "Best Perf for datatype = int8";
+    }
+
+    if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " ALayout =  RowMajor";
+    }
+    else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " ALayout =  ColumnMajor";
+    }
+
+    if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " BLayout =  RowMajor";
+    }
+    else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " BLayout =  ColumnMajor";
+    }
+
+    std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
+              << " StrideB = " << StrideB << " StrideC = " << StrideC << " : " << best_ave_time
+              << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp b/profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp
new file mode 100644
index 0000000000..93d3430bba
--- /dev/null
+++ b/profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp
@@ -0,0 +1,257 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <ck::index_t NDimSpatial,
+          typename OutLayout,
+          typename WeiLayout,
+          typename InLayout,
+          typename OutDataType,
+          typename WeiDataType,
+          typename InDataType>
+bool profile_grouped_conv_bwd_data_impl(int do_verification,
+                                        int init_method,
+                                        bool do_log,
+                                        bool time_kernel,
+                                        const ck::utils::conv::ConvParam& conv_param)
+{
+    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto out_element_op = OutElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto in_element_op  = InElementOp{};
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    Tensor<OutDataType> out(out_g_n_k_wos_desc);
+    Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
+    Tensor<InDataType> in_host(in_g_n_c_wis_desc);
+    Tensor<InDataType> in_device(in_g_n_c_wis_desc);
+
+    std::cout << "out: " << out.mDesc << std::endl;
+    std::cout << "wei: " << wei.mDesc << std::endl;
+    std::cout << "in: " << in_host.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    case 2:
+        out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
+        wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+        break;
+    default:
+        out.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
+        wei.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
+    }
+
+    DeviceMem out_device_buf(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
+    DeviceMem in_device_buf(sizeof(InDataType) * in_device.mDesc.GetElementSpaceSize());
+
+    out_device_buf.ToDevice(out.mData.data());
+    wei_device_buf.ToDevice(wei.mData.data());
+
+    // reset input to zero
+    in_device_buf.SetZero();
+
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<NDimSpatial,
+                                                                         InDataType,
+                                                                         WeiDataType,
+                                                                         OutDataType,
+                                                                         InElementOp,
+                                                                         WeiElementOp,
+                                                                         OutElementOp>();
+
+        auto ref_invoker = ref_conv.MakeInvoker();
+
+        in_host.SetZero();
+
+        auto ref_argument = ref_conv.MakeArgument(in_host,
+                                                  wei,
+                                                  out,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  out_element_op,
+                                                  wei_element_op,
+                                                  in_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    std::string best_op_name;
+    float best_avg_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    bool pass = true;
+
+    auto run_impl = [&](auto& op_ptr, auto& argument_ptr) {
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init output to zero before profiling next kernel
+            in_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = conv_param.GetFlops();
+            std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+            float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
+
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                in_device_buf.FromDevice(in_device.mData.data());
+
+                pass = pass & ck::utils::check_err(in_device, in_host);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "output : ", out.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "weight: ", wei.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "in_host  : ", in_host.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "in_device: ", in_device.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    };
+
+    // do GEMM
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NDimSpatial,
+                                                                                     OutLayout,
+                                                                                     WeiLayout,
+                                                                                     ck::Tuple<>,
+                                                                                     InLayout,
+                                                                                     OutDataType,
+                                                                                     WeiDataType,
+                                                                                     ck::Tuple<>,
+                                                                                     InDataType,
+                                                                                     OutElementOp,
+                                                                                     WeiElementOp,
+                                                                                     InElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::array<ck::index_t, NDimSpatial + 3> out_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> out_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> wei_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> wei_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> in_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> in_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(out_g_n_k_wos_desc.GetLengths(), out_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), out_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), wei_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), wei_strides);
+    copy(in_g_n_c_wis_desc.GetLengths(), in_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), in_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                        static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                        {},
+                                        static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                        out_lengths,
+                                        out_strides,
+                                        wei_lengths,
+                                        wei_strides,
+                                        {},
+                                        {},
+                                        in_lengths,
+                                        in_strides,
+                                        conv_filter_strides,
+                                        conv_filter_dilations,
+                                        input_left_pads,
+                                        input_right_pads,
+                                        out_element_op,
+                                        wei_element_op,
+                                        in_element_op);
+
+        run_impl(op_ptr, argument_ptr);
+    }
+
+    std::cout << "Best configuration parameters:"
+              << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
+              << "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp b/profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp
new file mode 100644
index 0000000000..48bf639a70
--- /dev/null
+++ b/profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp
@@ -0,0 +1,257 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+bool profile_grouped_conv_bwd_weight_impl(int do_verification,
+                                          int init_method,
+                                          bool do_log,
+                                          bool time_kernel,
+                                          const ck::utils::conv::ConvParam& conv_param,
+                                          ck::index_t split_k)
+{
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    Tensor<InDataType> input(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> weight_host_result(wei_g_k_c_xs_desc);
+    Tensor<WeiDataType> weight_device_result(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> output(out_g_n_k_wos_desc);
+
+    std::cout << "input: " << input.mDesc << std::endl;
+    std::cout << "weight: " << weight_host_result.mDesc << std::endl;
+    std::cout << "output: " << output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        output.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        break;
+    default:
+        input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        output.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) *
+                             weight_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * output.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(input.mData.data());
+    out_device_buf.ToDevice(output.mData.data());
+
+    if(do_verification)
+    {
+        auto ref_conv     = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
+                                                                           InDataType,
+                                                                           WeiDataType,
+                                                                           OutDataType,
+                                                                           InElementOp,
+                                                                           WeiElementOp,
+                                                                           OutElementOp>{};
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(input,
+                                                  weight_host_result,
+                                                  output,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdWeight<NDimSpatial,
+                                                                              InLayout,
+                                                                              WeiLayout,
+                                                                              OutLayout,
+                                                                              InDataType,
+                                                                              WeiDataType,
+                                                                              OutDataType,
+                                                                              InElementOp,
+                                                                              WeiElementOp,
+                                                                              OutElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    float best_avg_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device Conv instances
+    bool all_pass = true;
+
+    std::array<ck::index_t, NDimSpatial + 3> input_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> filter_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> output_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> input_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> weights_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> output_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto range_copy = [](const auto& from, auto to) { std::copy(begin(from), end(from), to); };
+
+    range_copy(in_g_n_c_wis_desc.GetLengths(), begin(input_lengths));
+    range_copy(in_g_n_c_wis_desc.GetStrides(), begin(input_strides));
+    range_copy(wei_g_k_c_xs_desc.GetLengths(), begin(filter_lengths));
+    range_copy(wei_g_k_c_xs_desc.GetStrides(), begin(weights_strides));
+    range_copy(out_g_n_k_wos_desc.GetLengths(), begin(output_lengths));
+    range_copy(out_g_n_k_wos_desc.GetStrides(), begin(output_strides));
+    range_copy(conv_param.conv_filter_strides_, begin(conv_filter_strides));
+    range_copy(conv_param.conv_filter_dilations_, begin(conv_filter_dilations));
+    range_copy(conv_param.input_left_pads_, begin(input_left_pads));
+    range_copy(conv_param.input_right_pads_, begin(input_right_pads));
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                        static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                        static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                        input_lengths,
+                                        input_strides,
+                                        filter_lengths,
+                                        weights_strides,
+                                        output_lengths,
+                                        output_strides,
+                                        conv_filter_strides,
+                                        conv_filter_dilations,
+                                        input_left_pads,
+                                        input_right_pads,
+                                        in_element_op,
+                                        wei_element_op,
+                                        out_element_op,
+                                        split_k);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // using atomic add, so need to reset input
+            wei_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = conv_param.GetFlops();
+            std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / avg_time;
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                wei_device_buf.FromDevice(weight_device_result.mData.data());
+
+                bool pass = ck::utils::check_err(weight_device_result, weight_host_result);
+
+                if(!pass)
+                {
+                    std::cout << "Fail info: " << op_ptr->GetTypeString() << std::endl;
+                }
+
+                all_pass &= pass;
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "output : ", output.mData, ",") << std::endl;
+                    ;
+                    LogRangeAsType<float>(
+                        std::cout << "weight (device): ", weight_device_result.mData, ",")
+                        << std::endl;
+                    ;
+                    LogRangeAsType<float>(
+                        std::cout << "weight (host): ", weight_host_result.mData, ",")
+                        << std::endl;
+                    ;
+                    LogRangeAsType<float>(std::cout << "input: ", input.mData, ",") << std::endl;
+                    ;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best configuration parameters:"
+              << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
+              << "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
+
+    return all_pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp b/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp
new file mode 100644
index 0000000000..9fadfe9699
--- /dev/null
+++ b/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp
@@ -0,0 +1,253 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
+          typename InDataType,
+          typename WeiDataType,
+          typename OutDataType>
+bool profile_grouped_conv_fwd_impl(int do_verification,
+                                   int init_method,
+                                   bool do_log,
+                                   bool time_kernel,
+                                   const ck::utils::conv::ConvParam& conv_param)
+{
+    using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+    using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto in_element_op  = InElementOp{};
+    const auto wei_element_op = WeiElementOp{};
+    const auto out_element_op = OutElementOp{};
+
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> a_g_n_c_wis_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> b_g_k_c_xs_strides{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> e_g_n_k_wos_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+    copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
+    copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
+    copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths);
+    copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides);
+    copy(out_g_n_k_wos_desc.GetLengths(), e_g_n_k_wos_lengths);
+    copy(out_g_n_k_wos_desc.GetStrides(), e_g_n_k_wos_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    Tensor<InDataType> input(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> weight(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> host_output(out_g_n_k_wos_desc);
+    Tensor<OutDataType> device_output(out_g_n_k_wos_desc);
+
+    std::cout << "input: " << input.mDesc << std::endl;
+    std::cout << "weight: " << weight.mDesc << std::endl;
+    std::cout << "output: " << host_output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        weight.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
+        break;
+    default:
+        input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        weight.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpaceSize());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) * weight.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * device_output.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(input.mData.data());
+    wei_device_buf.ToDevice(weight.mData.data());
+
+    // run reference op
+    if(do_verification)
+    {
+        auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                     InDataType,
+                                                                     WeiDataType,
+                                                                     OutDataType,
+                                                                     InElementOp,
+                                                                     WeiElementOp,
+                                                                     OutElementOp>{};
+
+        auto ref_invoker  = ref_conv.MakeInvoker();
+        auto ref_argument = ref_conv.MakeArgument(input,
+                                                  weight,
+                                                  host_output,
+                                                  conv_param.conv_filter_strides_,
+                                                  conv_param.conv_filter_dilations_,
+                                                  conv_param.input_left_pads_,
+                                                  conv_param.input_right_pads_,
+                                                  in_element_op,
+                                                  wei_element_op,
+                                                  out_element_op);
+
+        // init host output to zero
+        host_output.SetZero();
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    std::string best_op_name;
+    float best_avg_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    bool pass = true;
+
+    auto run_impl = [&](auto& op_ptr, auto& argument_ptr) {
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init output to zero before profiling next kernel
+            out_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop      = conv_param.GetFlops();
+            std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
+
+            float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
+
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                out_device_buf.FromDevice(device_output.mData.data());
+
+                pass = pass & ck::utils::check_err(device_output, host_output);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "input : ", input.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "weight: ", weight.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "host_output  : ", host_output.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "device_output: ", device_output.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    };
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                                                 InLayout,
+                                                                                 WeiLayout,
+                                                                                 ck::Tuple<>,
+                                                                                 OutLayout,
+                                                                                 InDataType,
+                                                                                 WeiDataType,
+                                                                                 ck::Tuple<>,
+                                                                                 OutDataType,
+                                                                                 InElementOp,
+                                                                                 WeiElementOp,
+                                                                                 OutElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "xdl found " << op_ptrs.size() << " instances" << std::endl;
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in_device_buf.GetDeviceBuffer(),
+                                                        wei_device_buf.GetDeviceBuffer(),
+                                                        {},
+                                                        out_device_buf.GetDeviceBuffer(),
+                                                        a_g_n_c_wis_lengths,
+                                                        a_g_n_c_wis_strides,
+                                                        b_g_k_c_xs_lengths,
+                                                        b_g_k_c_xs_strides,
+                                                        {},
+                                                        {},
+                                                        e_g_n_k_wos_lengths,
+                                                        e_g_n_k_wos_strides,
+                                                        conv_filter_strides,
+                                                        conv_filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        in_element_op,
+                                                        wei_element_op,
+                                                        out_element_op);
+
+        run_impl(op_ptr, argument_ptr);
+    }
+
+    std::cout << "Best configuration parameters:"
+              << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
+              << "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_grouped_gemm_fastgelu_impl.hpp b/profiler/include/profiler/profile_grouped_gemm_fastgelu_impl.hpp
new file mode 100644
index 0000000000..f05b13b749
--- /dev/null
+++ b/profiler/include/profiler/profile_grouped_gemm_fastgelu_impl.hpp
@@ -0,0 +1,280 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_grouped_gemm_fastgelu_impl(int do_verification,
+                                        int init_method,
+                                        bool do_log,
+                                        bool time_kernel,
+                                        const std::vector<int>& Ms,
+                                        const std::vector<int>& Ns,
+                                        const std::vector<int>& Ks,
+                                        const std::vector<int>& StrideAs,
+                                        const std::vector<int>& StrideBs,
+                                        const std::vector<int>& StrideCs)
+{
+
+    bool pass = true;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    std::size_t group_count = Ms.size();
+
+    if(!(group_count == Ns.size() && group_count == Ks.size() && group_count == StrideAs.size() &&
+         group_count == StrideBs.size() && group_count == StrideCs.size()))
+    {
+        throw std::runtime_error("wrong! inconsistent M/N/Ks, StrideA/B/Cs size\n");
+    }
+
+    std::vector<Tensor<ADataType>> a_m_k;
+    std::vector<Tensor<BDataType>> b_k_n;
+    std::vector<Tensor<CDataType>> c_m_n_device_results;
+
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        a_m_k.push_back(
+            Tensor<ADataType>(f_host_tensor_descriptor(Ms[i], Ks[i], StrideAs[i], ALayout{})));
+        b_k_n.push_back(
+            Tensor<BDataType>(f_host_tensor_descriptor(Ks[i], Ns[i], StrideBs[i], BLayout{})));
+
+        c_m_n_device_results.push_back(
+            Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
+
+        std::cout << "group: " << i << " a_m_k[" << i << "]:" << a_m_k[i].mDesc << ", b_k_n[" << i
+                  << "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
+                  << "]:" << c_m_n_device_results[i].mDesc << std::endl;
+
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            ck::utils::FillUniformDistributionIntegerValue<ADataType>{}(a_m_k[i]);
+            ck::utils::FillUniformDistributionIntegerValue<BDataType>{}(b_k_n[i]);
+            break;
+        default:
+            ck::utils::FillUniformDistribution<ADataType>{0.0, 1.0}(a_m_k[i]);
+            ck::utils::FillUniformDistribution<BDataType>{-0.5, 0.5}(b_k_n[i]);
+        }
+
+        ck::utils::FillConstant<CDataType>{}(c_m_n_device_results[i]);
+    }
+
+    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp = ck::tensor_operation::element_wise::FastGelu;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+    std::vector<DeviceMemPtr> a_device_buf, b_device_buf, c_device_buf;
+
+    a_device_buf.reserve(group_count);
+    b_device_buf.reserve(group_count);
+    c_device_buf.reserve(group_count);
+
+    std::vector<const void*> p_a, p_b;
+    std::vector<void*> p_c;
+
+    p_a.reserve(group_count);
+    p_b.reserve(group_count);
+    p_c.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        a_device_buf.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSpaceSize()));
+        b_device_buf.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(BDataType) * b_k_n[i].mDesc.GetElementSpaceSize()));
+        c_device_buf.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(CDataType) * c_m_n_device_results[i].mDesc.GetElementSpaceSize()));
+
+        a_device_buf[i]->ToDevice(a_m_k[i].mData.data());
+        b_device_buf[i]->ToDevice(b_k_n[i].mData.data());
+        c_device_buf[i]->SetZero();
+
+        gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}});
+
+        p_a.push_back(a_device_buf[i]->GetDeviceBuffer());
+        p_b.push_back(b_device_buf[i]->GetDeviceBuffer());
+        p_c.push_back(c_device_buf[i]->GetDeviceBuffer());
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
+                                                                     BLayout,
+                                                                     ck::Tuple<>,
+                                                                     CLayout,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     ck::Tuple<>,
+                                                                     CDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     CElementOp>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    if(op_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device GEMM instance found");
+    }
+
+    std::string best_gemm_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    auto p_ds = std::vector<std::array<const void*, 0>>{};
+
+    // profile device GEMM instances
+    for(auto& gemm_ptr : op_ptrs)
+    {
+        auto argument_ptr = gemm_ptr->MakeArgumentPointer(
+            p_a, p_b, p_ds, p_c, gemm_descs, a_element_op, b_element_op, c_element_op);
+
+        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
+        DeviceMem gemm_desc_workspace(gemm_ptr->GetWorkSpaceSize(argument_ptr.get()));
+        gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
+
+        if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::string gemm_name = gemm_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = 0, num_btype = 0;
+            for(std::size_t i = 0; i < gemm_descs.size(); i++)
+            {
+                flop += std::size_t(2) * Ms[i] * Ns[i] * Ks[i];
+                num_btype += sizeof(ADataType) * Ms[i] * Ks[i] + sizeof(BDataType) * Ks[i] * Ns[i] +
+                             sizeof(CDataType) * Ms[i] * Ns[i];
+            }
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << gemm_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_gemm_name  = gemm_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                for(std::size_t i = 0; i < gemm_descs.size(); i++)
+                {
+
+                    c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
+                    Tensor<CDataType> c_m_n_host_result(
+                        f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{}));
+
+                    using ReferenceGemmInstance =
+                        ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                  BDataType,
+                                                                  CDataType,
+                                                                  AccDataType,
+                                                                  AElementOp,
+                                                                  BElementOp,
+                                                                  CElementOp>;
+
+                    auto ref_gemm     = ReferenceGemmInstance{};
+                    auto ref_invoker  = ref_gemm.MakeInvoker();
+                    auto ref_argument = ref_gemm.MakeArgument(a_m_k[i],
+                                                              b_k_n[i],
+                                                              c_m_n_host_result,
+                                                              a_element_op,
+                                                              b_element_op,
+                                                              c_element_op);
+
+                    ref_invoker.Run(ref_argument);
+
+                    bool group_pass =
+                        ck::utils::check_err(c_m_n_device_results[i], c_m_n_host_result);
+                    pass = pass && group_pass;
+
+                    std::cout << "group: " << i << " verification result: " << std::boolalpha
+                              << group_pass << std::endl;
+
+                    if(do_log)
+                    {
+                        LogRangeAsType<float>(std::cout << "a : ", a_m_k[i].mData, ",")
+                            << std::endl;
+                        LogRangeAsType<float>(std::cout << "b: ", b_k_n[i].mData, ",") << std::endl;
+                        LogRangeAsType<float>(
+                            std::cout << "c_device: ", c_m_n_device_results[i].mData, ",")
+                            << std::endl;
+                        LogRangeAsType<float>(
+                            std::cout << "c_host  : ", c_m_n_host_result.mData, ",")
+                            << std::endl;
+                    }
+                }
+            }
+        }
+        else
+        {
+            std::cout << "does not support this GEMM problem" << std::endl;
+        }
+    }
+
+    if(do_verification)
+    {
+        std::cout << "Verification: " << (pass ? "SUCCESS" : "FAILURE") << std::endl;
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_grouped_gemm_impl.hpp b/profiler/include/profiler/profile_grouped_gemm_impl.hpp
new file mode 100644
index 0000000000..fe7a397606
--- /dev/null
+++ b/profiler/include/profiler/profile_grouped_gemm_impl.hpp
@@ -0,0 +1,361 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_grouped_gemm_impl(int do_verification,
+                               int init_method,
+                               bool do_log,
+                               bool time_kernel,
+                               const std::vector<int>& Ms,
+                               const std::vector<int>& Ns,
+                               const std::vector<int>& Ks,
+                               const std::vector<int>& StrideAs,
+                               const std::vector<int>& StrideBs,
+                               const std::vector<int>& StrideCs,
+                               int kbatch = 1)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    std::size_t group_count = Ms.size();
+
+    if(!(group_count == Ns.size() && group_count == Ks.size() && group_count == StrideAs.size() &&
+         group_count == StrideBs.size() && group_count == StrideCs.size()))
+    {
+        throw std::runtime_error("wrong! inconsistent M/N/Ks, StrideA/B/Cs size\n");
+    }
+
+    std::vector<Tensor<ADataType>> a_m_k;
+    std::vector<Tensor<BDataType>> b_k_n;
+    std::vector<Tensor<CDataType>> c_m_n_host_results;
+    std::vector<Tensor<CDataType>> c_m_n_device_results;
+
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        a_m_k.push_back(
+            Tensor<ADataType>(f_host_tensor_descriptor(Ms[i], Ks[i], StrideAs[i], ALayout{})));
+        b_k_n.push_back(
+            Tensor<BDataType>(f_host_tensor_descriptor(Ks[i], Ns[i], StrideBs[i], BLayout{})));
+
+        c_m_n_device_results.push_back(
+            Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
+
+        c_m_n_host_results.push_back(
+            Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
+#if DEBUG_LOG
+        std::cout << "group: " << i << " a_m_k[" << i << "]:" << a_m_k[i].mDesc << ", b_k_n[" << i
+                  << "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
+                  << "]:" << c_m_n_device_results[i].mDesc << std::endl;
+#endif // DEBUG_LOG
+        std::size_t num_thread = 1;
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            a_m_k[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
+            b_k_n[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            a_m_k[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
+            b_k_n[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+        }
+    }
+
+    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+    std::vector<DeviceMemPtr> a_device_buf, b_device_buf, c_device_buf;
+
+    a_device_buf.reserve(group_count);
+    b_device_buf.reserve(group_count);
+    c_device_buf.reserve(group_count);
+
+    std::vector<const void*> p_a, p_b;
+    std::vector<void*> p_c;
+
+    p_a.reserve(group_count);
+    p_b.reserve(group_count);
+    p_c.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        a_device_buf.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSpaceSize()));
+        b_device_buf.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(BDataType) * b_k_n[i].mDesc.GetElementSpaceSize()));
+        c_device_buf.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(CDataType) * c_m_n_device_results[i].mDesc.GetElementSpaceSize()));
+
+        a_device_buf[i]->ToDevice(a_m_k[i].mData.data());
+        b_device_buf[i]->ToDevice(b_k_n[i].mData.data());
+
+        gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}});
+
+        p_a.push_back(a_device_buf[i]->GetDeviceBuffer());
+        p_b.push_back(b_device_buf[i]->GetDeviceBuffer());
+        p_c.push_back(c_device_buf[i]->GetDeviceBuffer());
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
+                                                                     BLayout,
+                                                                     ck::Tuple<>,
+                                                                     CLayout,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     ck::Tuple<>,
+                                                                     CDataType,
+                                                                     AElementOp,
+                                                                     BElementOp,
+                                                                     CElementOp>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    if(op_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device GEMM instance found");
+    }
+
+    std::string best_gemm_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+    float best_kbatch     = 0;
+
+    auto p_ds = std::vector<std::array<const void*, 0>>{};
+
+    if(do_verification)
+    {
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                    BDataType,
+                                                                                    CDataType,
+                                                                                    AccDataType,
+                                                                                    AElementOp,
+                                                                                    BElementOp,
+                                                                                    CElementOp>;
+
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_m_k[i],
+                                                      b_k_n[i],
+                                                      c_m_n_host_results[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      c_element_op);
+
+            ref_invoker.Run(ref_argument);
+        }
+    }
+
+    // profile device GEMM instances
+    for(auto& gemm_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            gemm_ptr->MakeArgumentPointer(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_c,
+                                          gemm_descs,
+                                          ck::tensor_operation::element_wise::PassThrough{},
+                                          ck::tensor_operation::element_wise::PassThrough{},
+                                          ck::tensor_operation::element_wise::PassThrough{});
+
+        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
+
+        DeviceMem gemm_desc_workspace(gemm_ptr->GetWorkSpaceSize(argument_ptr.get()));
+
+        gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
+        std::string gemm_name = gemm_ptr->GetTypeString();
+
+        using DeviceOpSplitK = ck::tensor_operation::device::DeviceGroupedGemmSplitK<ALayout,
+                                                                                     BLayout,
+                                                                                     ck::Tuple<>,
+                                                                                     CLayout,
+                                                                                     ADataType,
+                                                                                     BDataType,
+                                                                                     ck::Tuple<>,
+                                                                                     CDataType,
+                                                                                     AElementOp,
+                                                                                     BElementOp,
+                                                                                     CElementOp>;
+
+        // skip non-splitk grouped_gemm
+        if(dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get()) == nullptr)
+        {
+            continue;
+        }
+
+        std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 24, 32, 48, 64};
+
+        if(kbatch > 0)
+        {
+            kbatch_list = {kbatch};
+        }
+
+        for(std::size_t j = 0; j < kbatch_list.size(); j++)
+        {
+
+            auto kbatch_curr = kbatch_list[j];
+
+            dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get())
+                ->SetKBatchSize(argument_ptr.get(), kbatch_curr);
+
+            if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+                for(std::size_t i = 0; i < gemm_descs.size(); i++)
+                    c_device_buf[i]->SetZero();
+
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+
+                if(do_verification)
+                {
+                    bool instance_pass = true;
+                    for(std::size_t i = 0; i < gemm_descs.size(); i++)
+                    {
+
+                        c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
+
+                        if(std::is_same_v<CDataType, ck::half_t> && kbatch_curr > 1)
+                        {
+                            instance_pass =
+                                instance_pass && ck::utils::check_err(c_m_n_device_results[i],
+                                                                      c_m_n_host_results[i],
+                                                                      "Error: Incorrect results!",
+                                                                      0.06);
+                        }
+                        else
+                        {
+                            instance_pass =
+                                instance_pass && ck::utils::check_err(c_m_n_device_results[i],
+                                                                      c_m_n_host_results[i]);
+                        }
+
+                        if(do_log)
+                        {
+                            LogRangeAsType<float>(std::cout << "a : ", a_m_k[i].mData, ",")
+                                << std::endl;
+                            LogRangeAsType<float>(std::cout << "b: ", b_k_n[i].mData, ",")
+                                << std::endl;
+                            LogRangeAsType<float>(
+                                std::cout << "c_device: ", c_m_n_device_results[i].mData, ",")
+                                << std::endl;
+                            LogRangeAsType<float>(
+                                std::cout << "c_host  : ", c_m_n_host_results[i].mData, ",")
+                                << std::endl;
+                        }
+                    }
+
+                    std::cout << "Instance: " << gemm_name << " verification "
+                              << (instance_pass ? "SUCCEED" : "FAILED") << std::endl;
+
+                    pass = pass && instance_pass;
+                }
+
+                float ave_time =
+                    invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+                if(time_kernel)
+                {
+                    std::size_t flop = 0, num_btype = 0;
+                    for(std::size_t i = 0; i < gemm_descs.size(); i++)
+                    {
+                        flop += std::size_t(2) * Ms[i] * Ns[i] * Ks[i];
+
+                        num_btype += sizeof(ADataType) * Ms[i] * Ks[i] +
+                                     sizeof(BDataType) * Ks[i] * Ns[i] +
+                                     sizeof(CDataType) * Ms[i] * Ns[i];
+                    }
+
+                    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+                    float gb_per_sec = num_btype / 1.E6 / ave_time;
+                    std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
+                              << " TFlops, " << gb_per_sec << " GB/s, " << gemm_name << ", KBatch "
+                              << kbatch_curr << std::endl;
+
+                    if(tflops > best_tflops)
+                    {
+                        best_gemm_name  = gemm_name;
+                        best_tflops     = tflops;
+                        best_ave_time   = ave_time;
+                        best_gb_per_sec = gb_per_sec;
+                        best_kbatch     = kbatch_curr;
+                    }
+                }
+            }
+            else
+            {
+                std::cout << "Instance: " << gemm_name << ", does not support this GEMM problem"
+                          << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+                  << best_gb_per_sec << " GB/s, " << best_gemm_name << ", KBatch = " << best_kbatch
+                  << std::endl;
+    }
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_groupnorm_impl.hpp b/profiler/include/profiler/profile_groupnorm_impl.hpp
new file mode 100644
index 0000000000..f88ba8453c
--- /dev/null
+++ b/profiler/include/profiler/profile_groupnorm_impl.hpp
@@ -0,0 +1,212 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/normalization.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType>
+bool profile_groupnorm_impl(int do_verification,
+                            int init_method,
+                            bool do_log,
+                            bool time_kernel,
+                            std::vector<index_t> length)
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    if(length.size() != 5)
+        return false;
+
+    index_t G = length[3];
+    index_t C = length[4];
+
+    std::vector<index_t> reduce_dim      = {1, 2, 4};
+    std::vector<index_t> gammaBetaLength = {G, C};
+    std::vector<index_t> gammaBetaStride = {0, 0, 0, C, 1};
+
+    Tensor<XDataType> x(length);
+    Tensor<GammaDataType> gamma(gammaBetaLength);
+    Tensor<BetaDataType> beta(gammaBetaLength);
+    Tensor<YDataType> y(length);
+    Tensor<YDataType> host_y(length);
+
+    switch(init_method)
+    {
+    case 0:
+        x.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        gamma.GenerateTensorValue(GeneratorTensor_1<GammaDataType>{});
+        beta.GenerateTensorValue(GeneratorTensor_1<BetaDataType>{});
+        break;
+    case 1:
+        x.GenerateTensorValue(GeneratorTensor_2<XDataType>{-5, 5});
+        gamma.GenerateTensorValue(GeneratorTensor_2<GammaDataType>{-5, 5});
+        beta.GenerateTensorValue(GeneratorTensor_2<BetaDataType>{-5, 5});
+        break;
+    default:
+        x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1});
+        gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-0.5, 0.5});
+        beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    beta_dev.ToDevice(beta.mData.data());
+
+    // add device normalization instances
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
+                                                                       GammaDataType,
+                                                                       BetaDataType,
+                                                                       AccDataType,
+                                                                       YDataType,
+                                                                       PassThrough,
+                                                                       5,
+                                                                       3>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using ReferenceInstance = ck::tensor_operation::host::ReferenceGroupnorm<XDataType,
+                                                                                 GammaDataType,
+                                                                                 BetaDataType,
+                                                                                 YDataType,
+                                                                                 AccDataType,
+                                                                                 PassThrough>;
+
+        ReferenceInstance ref;
+        auto ref_argument = ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, length, 1e-6);
+        auto ref_invoker  = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+    }
+
+    int num_kernel = 0;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            length,
+            std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
+            gammaBetaStride,
+            gammaBetaStride,
+            std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
+            reduce_dim,
+            1e-6,
+            x_dev.GetDeviceBuffer(),
+            gamma_dev.GetDeviceBuffer(),
+            beta_dev.GetDeviceBuffer(),
+            y_dev.GetDeviceBuffer(),
+            nullptr,
+            nullptr,
+            PassThrough{});
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+        }
+        else
+        {
+            continue;
+        }
+
+        size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
+        DeviceMem workspace_dev(workspace_sz);
+        inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes = x.mDesc.GetElementSize() * sizeof(XDataType) +
+                                gamma.mDesc.GetElementSize() * sizeof(GammaDataType) +
+                                beta.mDesc.GetElementSize() * sizeof(BetaDataType) +
+                                y.mDesc.GetElementSize() * sizeof(YDataType);
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            y_dev.FromDevice(y.mData.data());
+
+            bool pass = ck::utils::check_err(y, host_y, "Error: Incorrect results", 1e-3, 1e-3);
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(std::cout << "x  : ", x.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "host_y  : ", host_y.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "y  : ", y.mData, ",") << std::endl;
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "lengths = [", length, ", ") << "]." << std::endl;
+                return false;
+            }
+            else
+            {
+                if(time_kernel)
+                    std::cout << "pass" << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", length, ",") << std::endl;
+        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_image_to_column_impl.hpp b/profiler/include/profiler/profile_image_to_column_impl.hpp
new file mode 100644
index 0000000000..cc929e9220
--- /dev/null
+++ b/profiler/include/profiler/profile_image_to_column_impl.hpp
@@ -0,0 +1,200 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+#include <limits>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_image_to_column.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp"
+#include "ck/library/tensor_operation_instance/gpu/image_to_column.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+template <index_t NDimSpatial,
+          typename InputLayout,
+          typename InputDataType,
+          typename OutputDataType>
+bool profile_image_to_column_impl(int do_verification,
+                                  int init_method,
+                                  bool do_log,
+                                  bool time_kernel,
+                                  const ck::utils::conv::ConvParam& conv_param)
+{
+    const ck::index_t NDoHoWo =
+        conv_param.N_ *
+        ck::accumulate_n<ck::index_t>(
+            conv_param.output_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+    const ck::index_t CZYX =
+        conv_param.C_ *
+        ck::accumulate_n<ck::index_t>(
+            conv_param.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+
+    const auto in_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InputLayout>(
+            conv_param);
+    const auto out_desc = HostTensorDescriptor({NDoHoWo, CZYX});
+
+    std::array<ck::index_t, NDimSpatial> input_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial + 3> input_g_n_c_wis_strides{};
+    std::array<ck::index_t, 2> output_m_k_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
+
+    copy(conv_param.input_spatial_lengths_, input_spatial_lengths);
+    copy(conv_param.filter_spatial_lengths_, filter_spatial_lengths);
+    copy(conv_param.output_spatial_lengths_, output_spatial_lengths);
+    copy(in_desc.GetStrides(), input_g_n_c_wis_strides);
+    copy(out_desc.GetStrides(), output_m_k_strides);
+    copy(conv_param.conv_filter_strides_, conv_filter_strides);
+    copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_param.input_left_pads_, input_left_pads);
+    copy(conv_param.input_right_pads_, input_right_pads);
+
+    Tensor<InputDataType> input(in_desc);
+    Tensor<OutputDataType> host_output(out_desc);
+    Tensor<OutputDataType> device_output(out_desc);
+
+    std::cout << "input: " << input.mDesc << std::endl;
+    std::cout << "output: " << host_output.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1: input.GenerateTensorValue(GeneratorTensor_2<InputDataType>{-5, 5}); break;
+    default: input.GenerateTensorValue(GeneratorTensor_3<InputDataType>{0.0, 1.0});
+    }
+
+    DeviceMem in_device_buf(sizeof(InputDataType) * input.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutputDataType) * device_output.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(input.mData.data());
+
+    // run reference op
+    if(do_verification)
+    {
+        auto ref_image_to_column = ck::tensor_operation::host::
+            ReferenceImageToColumn<NDimSpatial, InputLayout, InputDataType, OutputDataType>{};
+
+        auto ref_invoker  = ref_image_to_column.MakeInvoker();
+        auto ref_argument = ref_image_to_column.MakeArgument(input,
+                                                             host_output,
+                                                             conv_param.filter_spatial_lengths_,
+                                                             conv_param.conv_filter_strides_,
+                                                             conv_param.conv_filter_dilations_,
+                                                             conv_param.input_left_pads_,
+                                                             conv_param.input_right_pads_);
+
+        // init host output to zero
+        host_output.SetZero();
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    using DeviceOp = ck::tensor_operation::device::
+        DeviceImageToColumn<NDimSpatial, InputLayout, InputDataType, OutputDataType>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    bool pass                   = true;
+    bool is_supporting_instance = false;
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            static_cast<InputDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<OutputDataType*>(out_device_buf.GetDeviceBuffer()),
+            conv_param.N_,
+            conv_param.C_,
+            input_spatial_lengths,
+            filter_spatial_lengths,
+            output_spatial_lengths,
+            input_g_n_c_wis_strides,
+            output_m_k_strides,
+            conv_filter_strides,
+            conv_filter_dilations,
+            input_left_pads,
+            input_right_pads);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            is_supporting_instance = true;
+            // re-init output to zero before profiling next kernel
+            out_device_buf.SetZero();
+            std::string op_name = op_ptr->GetTypeString();
+            auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+            std::size_t num_btype =
+                NDoHoWo * CZYX * (sizeof(OutputDataType) + sizeof(InputDataType));
+            float gb_per_sec = num_btype / 1.E6 / avg_time;
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(avg_time < best_avg_time)
+            {
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                out_device_buf.FromDevice(device_output.mData.data());
+                pass = pass & ck::utils::check_err(device_output, host_output);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "input : ", input.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "host_output  : ", host_output.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "device_output: ", device_output.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best configuration parameters:"
+              << "\nname: " << best_op_name << "\navg_time: " << best_avg_time
+              << "\nGB/s: " << best_gb_per_sec << std::endl;
+
+    return is_supporting_instance && pass;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_layernorm_impl.hpp b/profiler/include/profiler/profile_layernorm_impl.hpp
new file mode 100644
index 0000000000..f969646c2f
--- /dev/null
+++ b/profiler/include/profiler/profile_layernorm_impl.hpp
@@ -0,0 +1,231 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/normalization.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename XDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename ComputeDataType,
+          typename YDataType,
+          index_t Rank>
+bool profile_layernorm_impl(int do_verification,
+                            int init_method,
+                            bool do_log,
+                            bool time_kernel,
+                            std::vector<index_t> length)
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    if(length.size() < 2)
+        return false;
+
+    // Assume normalize dimension except for batch (first) dimension
+    std::vector<index_t> reduce_length{length.begin() + 1, length.end()};
+    std::vector<index_t> reduce_dim;
+    for(int i = 1; i < Rank; ++i)
+        reduce_dim.push_back(i);
+
+    Tensor<XDataType> x(length);
+    Tensor<GammaDataType> gamma(reduce_length);
+    Tensor<BetaDataType> beta(reduce_length);
+    Tensor<YDataType> y(length);
+    Tensor<YDataType> host_y(length);
+
+    std::vector<index_t> strideXY =
+        std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()};
+    std::vector<index_t> strideGammaBeta = strideXY;
+    strideGammaBeta[0]                   = 0;
+
+    switch(init_method)
+    {
+    case 0:
+        x.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        gamma.GenerateTensorValue(GeneratorTensor_1<GammaDataType>{});
+        beta.GenerateTensorValue(GeneratorTensor_1<BetaDataType>{});
+        y.GenerateTensorValue(GeneratorTensor_1<YDataType>{});
+        break;
+    case 1:
+        x.GenerateTensorValue(GeneratorTensor_2<XDataType>{-5, 5});
+        gamma.GenerateTensorValue(GeneratorTensor_2<GammaDataType>{-5, 5});
+        beta.GenerateTensorValue(GeneratorTensor_2<BetaDataType>{-5, 5});
+        y.GenerateTensorValue(GeneratorTensor_2<YDataType>{-5, 5});
+        break;
+    default:
+        x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1});
+        gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-0.5, 0.5});
+        beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-0.5, 0.5});
+        y.GenerateTensorValue(GeneratorTensor_3<YDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+
+    x_dev.ToDevice(x.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    beta_dev.ToDevice(beta.mData.data());
+
+    constexpr int NumReduceDim = Rank - 1;
+
+    // add device normalization instances
+    using DeviceOp = ck::tensor_operation::device::DeviceNormalization<XDataType,
+                                                                       GammaDataType,
+                                                                       BetaDataType,
+                                                                       ComputeDataType,
+                                                                       YDataType,
+                                                                       PassThrough,
+                                                                       Rank,
+                                                                       NumReduceDim>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using ReferenceInstance = ck::tensor_operation::host::ReferenceLayernorm<XDataType,
+                                                                                 GammaDataType,
+                                                                                 BetaDataType,
+                                                                                 YDataType,
+                                                                                 ComputeDataType,
+                                                                                 PassThrough,
+                                                                                 Rank,
+                                                                                 NumReduceDim>;
+
+        ReferenceInstance ref;
+        auto ref_argument =
+            ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, length, reduce_dim, 1e-4);
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+    }
+
+    int num_kernel = 0;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(length,
+                                                          strideXY,
+                                                          strideGammaBeta,
+                                                          strideGammaBeta,
+                                                          strideXY,
+                                                          reduce_dim,
+                                                          1e-4,
+                                                          x_dev.GetDeviceBuffer(),
+                                                          gamma_dev.GetDeviceBuffer(),
+                                                          beta_dev.GetDeviceBuffer(),
+                                                          y_dev.GetDeviceBuffer(),
+                                                          nullptr,
+                                                          nullptr,
+                                                          PassThrough{});
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+        }
+        else
+        {
+            if(time_kernel)
+            {
+                std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
+                LogRange(std::cout << "input lengths = ", length, ", ") << std::endl;
+            }
+
+            continue;
+        }
+
+        size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
+        DeviceMem workspace_dev(workspace_sz);
+        inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes = x.mDesc.GetElementSize() * sizeof(XDataType) +
+                                gamma.mDesc.GetElementSize() * sizeof(GammaDataType) +
+                                beta.mDesc.GetElementSize() * sizeof(BetaDataType) +
+                                y.mDesc.GetElementSize() * sizeof(YDataType);
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            y_dev.FromDevice(y.mData.data());
+
+            bool pass =
+                ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 1e-3, 1e-3);
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(std::cout << "x  : ", x.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "host_y  : ", host_y.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "y  : ", y.mData, ",") << std::endl;
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "lengths = [", length, ", ") << "]." << std::endl;
+                return false;
+            }
+            else
+            {
+                if(time_kernel)
+                    std::cout << "pass" << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", length, ",") << ", ";
+        LogRange(std::cout << "stride = ", strideXY, ",") << ", ";
+        LogRange(std::cout << "reduce dims ", reduce_dim, ",") << std::endl;
+        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_max_pool3d_bwd_impl.hpp b/profiler/include/profiler/profile_max_pool3d_bwd_impl.hpp
new file mode 100644
index 0000000000..15fb4e9034
--- /dev/null
+++ b/profiler/include/profiler/profile_max_pool3d_bwd_impl.hpp
@@ -0,0 +1,288 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
+#include "ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename InDataType,
+          typename OutDataType,
+          typename IndexDataType,
+          typename DOutDataType,
+          typename DInDataType,
+          bool PropagateNan>
+bool profile_max_pool3d_bwd_impl(int do_verification,
+                                 int init_method,
+                                 bool do_log,
+                                 bool time_kernel,
+                                 std::vector<index_t> in_length, // NCDHW
+                                 std::vector<index_t> window_spatial_lengths,
+                                 std::vector<index_t> window_strides,
+                                 std::vector<index_t> window_dilations,
+                                 std::vector<index_t> input_left_pads,
+                                 std::vector<index_t> input_right_pads)
+{
+    // AtomicAdd only support f32 for now. ComputeDataType must be float32
+    using ComputeDataType = float;
+
+    constexpr index_t InOutRank  = 5;
+    constexpr index_t WindowRank = 3;
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank ||
+       window_strides.size() != WindowRank || window_dilations.size() != WindowRank ||
+       input_left_pads.size() != WindowRank || input_right_pads.size() != WindowRank)
+    {
+        std::cout << "Parameter is incorrect" << std::endl;
+        return false;
+    }
+
+    std::vector<index_t> out_length(InOutRank);
+
+    int N = in_length[0];
+    int C = in_length[1];
+
+    out_length[0] = N;
+    out_length[1] = C;
+
+    // Calculate Do, Ho, Wo
+    for(int i = 2; i < InOutRank; ++i)
+    {
+        auto pad1             = input_left_pads[i - 2];
+        auto pad2             = input_right_pads[i - 2];
+        auto windows_size     = window_spatial_lengths[i - 2];
+        auto windows_stride   = window_strides[i - 2];
+        auto windows_dilation = window_dilations[i - 2];
+        auto eff              = (windows_size - 1) * windows_dilation + 1;
+        out_length[i]         = (in_length[i] + pad1 + pad2 - eff) / windows_stride + 1;
+    }
+
+    int Di = in_length[2];
+    int Hi = in_length[3];
+    int Wi = in_length[4];
+    int Do = out_length[2];
+    int Ho = out_length[3];
+    int Wo = out_length[4];
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t D, std::size_t H, std::size_t W) {
+            using namespace ck::literals;
+
+            return HostTensorDescriptor({N_, C_, D, H, W},
+                                        {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
+        };
+
+    Tensor<InDataType> in_n_c_di_hi_wi(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
+    Tensor<OutDataType> out_n_c_do_ho_wo(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+    Tensor<IndexDataType> out_indices_n_c_do_ho_wo(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+    Tensor<DOutDataType> dout_n_c_do_ho_wo(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+    Tensor<DInDataType> din_n_c_di_hi_wi_host(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
+
+    Tensor<DInDataType> din_n_c_di_hi_wi_device(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
+
+    switch(init_method)
+    {
+    case 0:
+        in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{});
+        dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_1<DOutDataType>{});
+        break;
+    case 1:
+        in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_2<DOutDataType>{-5, 5});
+        break;
+    default:
+        in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
+        dout_n_c_do_ho_wo.GenerateTensorValue(GeneratorTensor_3<DOutDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem indices_device_buf(sizeof(IndexDataType) *
+                                 out_indices_n_c_do_ho_wo.mDesc.GetElementSpaceSize());
+    DeviceMem dout_device_buf(sizeof(DOutDataType) * dout_n_c_do_ho_wo.mDesc.GetElementSpaceSize());
+    DeviceMem din_device_buf(sizeof(DInDataType) *
+                             din_n_c_di_hi_wi_device.mDesc.GetElementSpaceSize());
+
+    // Generate index data from forwarding
+    {
+        using ReferencePoolingFwdInstance =
+            ck::tensor_operation::host::ReferencePoolingFwd<InOutRank,
+                                                            WindowRank,
+                                                            InDataType,
+                                                            OutDataType,
+                                                            ComputeDataType,
+                                                            IndexDataType,
+                                                            ck::ReduceTensorOp::MAX,
+                                                            false,
+                                                            true>;
+
+        ReferencePoolingFwdInstance ref_pooling_fwd;
+        auto ref_pooling_fwd_argument = ref_pooling_fwd.MakeArgument(in_n_c_di_hi_wi,
+                                                                     out_n_c_do_ho_wo,
+                                                                     out_indices_n_c_do_ho_wo,
+                                                                     window_spatial_lengths,
+                                                                     window_strides,
+                                                                     window_dilations,
+                                                                     input_left_pads,
+                                                                     input_right_pads);
+        auto ref_pooling_fwd_invoker  = ref_pooling_fwd.MakeInvoker();
+        ref_pooling_fwd_invoker.Run(ref_pooling_fwd_argument);
+    }
+
+    indices_device_buf.ToDevice(out_indices_n_c_do_ho_wo.mData.data());
+    dout_device_buf.ToDevice(dout_n_c_do_ho_wo.mData.data());
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceMaxPoolBwd<DOutDataType, IndexDataType, DInDataType>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using ReferencePoolingBwdInstance =
+            ck::tensor_operation::host::ReferenceMaxPoolBwd<DOutDataType,
+                                                            IndexDataType,
+                                                            ComputeDataType,
+                                                            DInDataType,
+                                                            PassThrough>;
+
+        ReferencePoolingBwdInstance ref_pooling_bwd;
+        auto ref_pooling_bwd_argument = ref_pooling_bwd.MakeArgument(
+            dout_n_c_do_ho_wo, out_indices_n_c_do_ho_wo, din_n_c_di_hi_wi_host, PassThrough{});
+        auto ref_invoker = ref_pooling_bwd.MakeInvoker();
+        ref_invoker.Run(ref_pooling_bwd_argument);
+    }
+
+    int num_kernel = 0;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
+            static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+            dout_n_c_do_ho_wo.mDesc.GetElementSpaceSize(),
+            din_n_c_di_hi_wi_device.mDesc.GetElementSpaceSize(),
+            window_spatial_lengths,
+            window_strides,
+            window_dilations);
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+        }
+        else
+        {
+            if(time_kernel)
+            {
+                std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
+                LogRange(std::cout << "doutput lengths = ", out_length, ", ") << std::endl;
+            }
+
+            continue;
+        }
+
+        size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
+        DeviceMem workspace_device_buf(workspace_sz);
+        inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_device_buf.GetDeviceBuffer());
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+        float avg_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes =
+            dout_n_c_do_ho_wo.mDesc.GetElementSize() * sizeof(DOutDataType) +
+            out_indices_n_c_do_ho_wo.mDesc.GetElementSize() * sizeof(IndexDataType) +
+            din_n_c_di_hi_wi_device.mDesc.GetElementSize() * sizeof(DInDataType);
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            din_device_buf.FromDevice(din_n_c_di_hi_wi_device.mData.data());
+
+            bool pass = ck::utils::check_err(din_n_c_di_hi_wi_device.mData,
+                                             din_n_c_di_hi_wi_host.mData,
+                                             "Error: Incorrect results",
+                                             1e-3,
+                                             1e-3);
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(
+                    std::cout << "out_indices_n_c_do_ho_wo: ", out_indices_n_c_do_ho_wo.mData, ",")
+                    << std::endl;
+
+                LogRangeAsType<float>(
+                    std::cout << "din_n_c_di_hi_wi_device: ", din_n_c_di_hi_wi_device.mData, ",")
+                    << std::endl;
+
+                LogRangeAsType<float>(
+                    std::cout << "din_n_c_di_hi_wi_host: ", din_n_c_di_hi_wi_host.mData, ",")
+                    << std::endl;
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "doutput lengths = [", out_length, ", ") << "]." << std::endl;
+                return false;
+            }
+            else
+            {
+                if(time_kernel)
+                    std::cout << "pass" << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", out_length, ",") << std::endl;
+        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_pool3d_fwd_impl.hpp b/profiler/include/profiler/profile_pool3d_fwd_impl.hpp
new file mode 100644
index 0000000000..02fde48d6e
--- /dev/null
+++ b/profiler/include/profiler/profile_pool3d_fwd_impl.hpp
@@ -0,0 +1,280 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename InDataType,
+          typename OutDataType,
+          typename ComputeDataType,
+          typename IndexDataType,
+          typename InLayout,
+          typename OutLayout,
+          ck::ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool OutputIndex>
+bool profile_pool3d_fwd_impl(int do_verification,
+                             int init_method,
+                             bool do_log,
+                             bool time_kernel,
+                             std::vector<index_t> in_length, // NCDHW
+                             std::vector<index_t> window_spatial_lengths,
+                             std::vector<index_t> window_strides,
+                             std::vector<index_t> window_dilations,
+                             std::vector<index_t> input_left_pads,
+                             std::vector<index_t> input_right_pads)
+{
+    constexpr index_t InOutRank  = 5;
+    constexpr index_t WindowRank = 3;
+
+    if(in_length.size() != InOutRank || window_spatial_lengths.size() != WindowRank ||
+       window_strides.size() != WindowRank || window_dilations.size() != WindowRank ||
+       input_left_pads.size() != WindowRank || input_right_pads.size() != WindowRank)
+        return false;
+
+    std::vector<index_t> out_length(InOutRank);
+
+    int N = in_length[0];
+    int C = in_length[1];
+
+    out_length[0] = N;
+    out_length[1] = C;
+
+    // Calculate Do, Ho, Wo
+    for(int i = 2; i < InOutRank; ++i)
+    {
+        auto pad1             = input_left_pads[i - 2];
+        auto pad2             = input_right_pads[i - 2];
+        auto windows_size     = window_spatial_lengths[i - 2];
+        auto windows_stride   = window_strides[i - 2];
+        auto windows_dilation = window_dilations[i - 2];
+        auto eff              = (windows_size - 1) * windows_dilation + 1;
+        out_length[i]         = (in_length[i] + pad1 + pad2 - eff) / windows_stride + 1;
+    }
+
+    int Di = in_length[2];
+    int Hi = in_length[3];
+    int Wi = in_length[4];
+    int Do = out_length[2];
+    int Ho = out_length[3];
+    int Wo = out_length[4];
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t N_, std::size_t C_, std::size_t D, std::size_t H, std::size_t W) {
+            using namespace ck::literals;
+
+            return HostTensorDescriptor({N_, C_, D, H, W},
+                                        {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
+        };
+
+    Tensor<InDataType> in_n_c_di_hi_wi(f_host_tensor_descriptor(N, C, Di, Hi, Wi));
+    Tensor<OutDataType> out_n_c_do_ho_wo_host(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+    Tensor<IndexDataType> out_indices_n_c_do_ho_wo_host(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+
+    Tensor<OutDataType> out_n_c_do_ho_wo_device(f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+    Tensor<IndexDataType> out_indices_n_c_do_ho_wo_device(
+        f_host_tensor_descriptor(N, C, Do, Ho, Wo));
+
+    switch(init_method)
+    {
+    case 0: in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{}); break;
+    case 1: in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
+    default: in_n_c_di_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_di_hi_wi.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) *
+                             out_n_c_do_ho_wo_device.mDesc.GetElementSpaceSize());
+    DeviceMem out_indices_device_buf(sizeof(IndexDataType) *
+                                     out_indices_n_c_do_ho_wo_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in_n_c_di_hi_wi.mData.data());
+
+    // add device normalization instances
+    using DeviceOp = ck::tensor_operation::device::DevicePoolFwd<InOutRank,
+                                                                 WindowRank,
+                                                                 InDataType,
+                                                                 OutDataType,
+                                                                 IndexDataType,
+                                                                 InLayout,
+                                                                 OutLayout,
+                                                                 ReduceOpId,
+                                                                 OutputIndex>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using ReferenceInstance = ck::tensor_operation::host::ReferencePoolingFwd<InOutRank,
+                                                                                  WindowRank,
+                                                                                  InDataType,
+                                                                                  OutDataType,
+                                                                                  ComputeDataType,
+                                                                                  IndexDataType,
+                                                                                  ReduceOpId,
+                                                                                  PropagateNan,
+                                                                                  OutputIndex>;
+
+        ReferenceInstance ref;
+        auto ref_argument = ref.MakeArgument(in_n_c_di_hi_wi,
+                                             out_n_c_do_ho_wo_host,
+                                             out_indices_n_c_do_ho_wo_host,
+                                             window_spatial_lengths,
+                                             window_strides,
+                                             window_dilations,
+                                             input_left_pads,
+                                             input_right_pads);
+        auto ref_invoker  = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+    }
+
+    int num_kernel = 0;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+            static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+            static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
+            in_length,
+            window_spatial_lengths,
+            out_length,
+            {Di * C * Hi * Wi, 1, C * Hi * Wi, Wi * C, C},
+            {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C},
+            {Do * C * Ho * Wo, 1, C * Ho * Wo, Wo * C, C},
+            window_strides,
+            window_dilations,
+            input_left_pads,
+            input_right_pads,
+            {2, 3, 4});
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+        }
+        else
+        {
+            if(time_kernel)
+            {
+                std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
+                LogRange(std::cout << "input lengths = ", in_length, ", ") << std::endl;
+            }
+
+            continue;
+        }
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes = in_n_c_di_hi_wi.mDesc.GetElementSize() * sizeof(InDataType) +
+                                out_n_c_do_ho_wo_host.mDesc.GetElementSize() * sizeof(OutDataType);
+
+        if constexpr(OutputIndex)
+            num_bytes +=
+                out_indices_n_c_do_ho_wo_host.mDesc.GetElementSize() * sizeof(IndexDataType);
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            out_device_buf.FromDevice(out_n_c_do_ho_wo_device.mData.data());
+
+            bool pass = ck::utils::check_err(out_n_c_do_ho_wo_device.mData,
+                                             out_n_c_do_ho_wo_host.mData,
+                                             "Error: Incorrect results",
+                                             1e-3,
+                                             1e-3);
+
+            if constexpr(OutputIndex)
+            {
+                out_indices_device_buf.FromDevice(out_indices_n_c_do_ho_wo_device.mData.data());
+
+                pass = pass && ck::utils::check_err(out_indices_n_c_do_ho_wo_device,
+                                                    out_indices_n_c_do_ho_wo_host);
+            }
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(
+                    std::cout << "in_n_c_di_hi_wi  : ", in_n_c_di_hi_wi.mData, ",")
+                    << std::endl;
+                LogRangeAsType<float>(
+                    std::cout << "out_n_c_do_ho_wo_host  : ", out_n_c_do_ho_wo_host.mData, ",")
+                    << std::endl;
+                LogRangeAsType<float>(
+                    std::cout << "out_n_c_do_ho_wo_device  : ", out_n_c_do_ho_wo_device.mData, ",")
+                    << std::endl;
+
+                if constexpr(OutputIndex)
+                    LogRangeAsType<float>(std::cout << "out_indices_n_c_do_ho_wo_device  : ",
+                                          out_indices_n_c_do_ho_wo_device.mData,
+                                          ",")
+                        << std::endl;
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "lengths = [", in_length, ", ") << "]." << std::endl;
+                return false;
+            }
+            else
+            {
+                if(time_kernel)
+                    std::cout << "pass" << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", in_length, ",") << std::endl;
+        std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_reduce_impl.hpp b/profiler/include/profiler/profile_reduce_impl.hpp
new file mode 100644
index 0000000000..b54aa65aef
--- /dev/null
+++ b/profiler/include/profiler/profile_reduce_impl.hpp
@@ -0,0 +1,517 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/reduction_enums.hpp"
+#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_reduce.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <index_t Rank,
+          index_t NumReduceDim,
+          ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool UseIndex>
+struct ReduceDescription
+{
+    static constexpr index_t Rank_              = Rank;
+    static constexpr index_t NumReduceDim_      = NumReduceDim;
+    static constexpr ReduceTensorOp ReduceOpId_ = ReduceOpId;
+    static constexpr bool PropagateNan_         = PropagateNan;
+    static constexpr bool UseIndex_             = UseIndex;
+};
+
+using reduce_description_instances =
+    std::tuple<ReduceDescription<4, 3, ReduceTensorOp::ADD, false, false>, // for ADD
+               ReduceDescription<4, 4, ReduceTensorOp::ADD, false, false>,
+               ReduceDescription<4, 1, ReduceTensorOp::ADD, false, false>,
+               ReduceDescription<2, 1, ReduceTensorOp::ADD, false, false>,
+
+               ReduceDescription<4, 3, ReduceTensorOp::AVG, false, false>, // for AVG
+               ReduceDescription<4, 4, ReduceTensorOp::AVG, false, false>,
+               ReduceDescription<4, 1, ReduceTensorOp::AVG, false, false>,
+               ReduceDescription<2, 1, ReduceTensorOp::AVG, false, false>,
+
+               ReduceDescription<4, 3, ReduceTensorOp::NORM2, false, false>, // for NORM2
+               ReduceDescription<4, 4, ReduceTensorOp::NORM2, false, false>,
+               ReduceDescription<4, 1, ReduceTensorOp::NORM2, false, false>,
+               ReduceDescription<2, 1, ReduceTensorOp::NORM2, false, false>,
+
+               ReduceDescription<4, 3, ReduceTensorOp::MIN, false, false>, // for MIN
+               ReduceDescription<4, 4, ReduceTensorOp::MIN, false, false>,
+               ReduceDescription<4, 1, ReduceTensorOp::MIN, false, false>,
+               ReduceDescription<2, 1, ReduceTensorOp::MIN, false, false>,
+               ReduceDescription<4, 3, ReduceTensorOp::MAX, false, false>, // for MAX
+               ReduceDescription<4, 4, ReduceTensorOp::MAX, false, false>,
+               ReduceDescription<4, 1, ReduceTensorOp::MAX, false, false>,
+               ReduceDescription<2, 1, ReduceTensorOp::MAX, false, false>,
+               ReduceDescription<4, 3, ReduceTensorOp::AMAX, false, false>, // for AMAX
+               ReduceDescription<4, 4, ReduceTensorOp::AMAX, false, false>,
+               ReduceDescription<4, 1, ReduceTensorOp::AMAX, false, false>,
+               ReduceDescription<2, 1, ReduceTensorOp::AMAX, false, false>,
+
+               ReduceDescription<4, 3, ReduceTensorOp::MIN, false, true>, // for MIN
+               ReduceDescription<4, 4, ReduceTensorOp::MIN, false, true>,
+               ReduceDescription<4, 1, ReduceTensorOp::MIN, false, true>,
+               ReduceDescription<2, 1, ReduceTensorOp::MIN, false, true>,
+               ReduceDescription<4, 3, ReduceTensorOp::MAX, false, true>, // for MAX
+               ReduceDescription<4, 4, ReduceTensorOp::MAX, false, true>,
+               ReduceDescription<4, 1, ReduceTensorOp::MAX, false, true>,
+               ReduceDescription<2, 1, ReduceTensorOp::MAX, false, true>,
+               ReduceDescription<4, 3, ReduceTensorOp::AMAX, false, true>, // for AMAX
+               ReduceDescription<4, 4, ReduceTensorOp::AMAX, false, true>,
+               ReduceDescription<4, 1, ReduceTensorOp::AMAX, false, true>,
+               ReduceDescription<2, 1, ReduceTensorOp::AMAX, false, true>>;
+
+template <typename DescriptionType>
+bool description_match(const DescriptionType& description,
+                       int Rank,
+                       const std::vector<int>& reduceDims,
+                       ReduceTensorOp ReduceOpId,
+                       bool PropagateNan,
+                       bool UseIndex)
+{
+    if(description.Rank_ != Rank || description.ReduceOpId_ != ReduceOpId ||
+       description.PropagateNan_ != PropagateNan || description.UseIndex_ != UseIndex)
+        return (false);
+
+    if(DescriptionType::NumReduceDim_ != reduceDims.size())
+        return (false);
+
+    bool result = true;
+
+    return (result);
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+template <int Rank, int NumReduceDim>
+static inline std::array<int, Rank - NumReduceDim>
+get_invariant_dims(const std::array<int, NumReduceDim>& reduceDims)
+{
+    int reduceFlag = 0;
+
+    // flag the bits for the reduceDims
+    for(int i = 0; i < NumReduceDim; i++)
+    {
+        reduceFlag |= 1 << reduceDims[i];
+    };
+
+    std::array<int, Rank - NumReduceDim> invariantDims;
+
+    // collect invariant dimensions
+    int dim = 0;
+    for(int i = 0; i < Rank; i++)
+        if((reduceFlag & (1 << i)) == 0)
+        {
+            invariantDims[dim] = i;
+            dim++;
+        };
+
+    return invariantDims;
+};
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          int Rank,
+          int NumReduceDim,
+          ReduceTensorOp ReduceOpId,
+          bool PropagateNan,
+          bool UseIndex>
+bool profile_reduce_impl_impl(bool do_verification,
+                              int init_method,
+                              bool do_dumpout,
+                              bool time_kernel,
+                              const std::vector<size_t>& inLengths,
+                              const std::array<int, NumReduceDim>& reduceDims,
+                              float alpha,
+                              float beta)
+{
+    using namespace ck::tensor_operation::device;
+    using namespace ck::tensor_operation::device::instance;
+    using ck::host_common::dumpBufferToFile;
+
+    constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
+
+    constexpr bool op_support_indices =
+        (ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
+         ReduceOpId == ReduceTensorOp::AMAX);
+
+    constexpr bool OutputIndex = (op_support_indices && UseIndex);
+
+    // 1) If InDataType is half_t, must use half_t as AccDataType for indexable reduction operations
+    // 2) If InDataType is half_t, must use float as AccDataType for non-indexable reduction
+    // operations
+    constexpr bool invalid_reduce_1 =
+        std::is_same<InDataType, half_t>::value &&
+        ((!op_support_indices && !std::is_same<AccDataType, float>::value) ||
+         (op_support_indices && !std::is_same<AccDataType, half_t>::value));
+
+    // 1) If InDataType is float, must use float as AccDataType for indexable reduction operations
+    constexpr bool invalid_reduce_2 =
+        std::is_same<InDataType, float>::value &&
+        (op_support_indices && !std::is_same<AccDataType, float>::value);
+
+    // 1) The indices can only be used when the reduction operation is indexable
+    constexpr bool invalid_reduce_3 = (!op_support_indices && UseIndex);
+
+    // 1) If InDataType is int8_t, must use int8_t as AccDataType for indexable reduction operations
+    // 2) If InDataType is int8_t, must use int32_t as AccDataType for non-indexable reduction
+    // operations
+    constexpr bool invalid_reduce_4 =
+        std::is_same<InDataType, int8_t>::value &&
+        ((!op_support_indices && !std::is_same<AccDataType, int32_t>::value) ||
+         (op_support_indices && !std::is_same<AccDataType, int8_t>::value));
+
+    // 1) If InDataType is int8_t, the supported operation must be either indexable operations or
+    // ADD/AVG
+    constexpr bool invalid_reduce_5 = std::is_same<InDataType, int8_t>::value &&
+                                      (!op_support_indices && ReduceOpId != ReduceTensorOp::ADD &&
+                                       ReduceOpId != ReduceTensorOp::AVG);
+
+    // 1) If InDataType is bhalf_t, must use float as AccDataType for all reduction operations
+    constexpr bool invalid_reduce_6 =
+        std::is_same<InDataType, bhalf_t>::value && !std::is_same<AccDataType, float>::value;
+
+    constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 ||
+                                     invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6);
+
+    int num_kernel = 0;
+    bool pass      = true;
+
+    if constexpr(!invalid_reduce)
+    {
+        Tensor<InDataType> in(inLengths);
+
+        std::vector<size_t> outLengths;
+
+        const auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
+
+        if(reduceDims.size() == Rank)
+            outLengths.push_back(1);
+        else
+            for(auto dim : invariantDims)
+                outLengths.push_back(inLengths[dim]);
+
+        Tensor<OutDataType> out_ref(outLengths);
+        Tensor<OutDataType> out(outLengths);
+        Tensor<int32_t> out_indices_ref(outLengths);
+        Tensor<int32_t> out_indices(outLengths);
+
+        auto inStrides  = in.mDesc.GetStrides();
+        auto outStrides = out.mDesc.GetStrides();
+
+        size_t invariant_total_length = out.mDesc.GetElementSize();
+        size_t reduce_total_length    = in.mDesc.GetElementSize() / invariant_total_length;
+
+        std::size_t num_thread = 1;
+
+        if(do_verification)
+        {
+            switch(init_method)
+            {
+            case 0: break;
+            case 1:
+                in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread);
+                if(beta != 0.0f)
+                    out_ref.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread);
+                break;
+            case 2:
+                in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread);
+                if(beta != 0.0f)
+                    out_ref.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread);
+                break;
+            default:
+                in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0}, num_thread);
+                if(beta != 0.0f)
+                    out_ref.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0},
+                                                num_thread);
+            }
+
+            if(beta != 0.0f)
+                for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
+                    out.mData[i] = out_ref.mData[i];
+        };
+
+        // these buffers are usually provided by the user application
+        DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+        DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
+
+        in_dev.ToDevice(in.mData.data());
+
+        if(beta != 0.0f)
+            out_dev.ToDevice(out.mData.data());
+
+        size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int) : 0;
+
+        DeviceMem out_indices_dev(indicesSizeInBytes);
+
+        float best_avg_time   = 0;
+        float best_gb_per_sec = 0;
+
+        using InElementwiseOperation =
+            typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
+        using AccElementwiseOperation =
+            typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
+
+        using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
+
+        InElementwiseOperation in_elementwise_op;
+        AccElementwiseOperation acc_elementwise_op;
+
+        std::tie(in_elementwise_op, acc_elementwise_op) =
+            reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
+                static_cast<int32_t>(reduce_total_length));
+
+        using ReduceOp = ck::tensor_operation::device::DeviceReduce<InDataType,
+                                                                    AccDataType,
+                                                                    OutDataType,
+                                                                    Rank,
+                                                                    NumReduceDim,
+                                                                    ReduceOperation,
+                                                                    InElementwiseOperation,
+                                                                    AccElementwiseOperation,
+                                                                    PropagateNan,
+                                                                    OutputIndex>;
+        const auto reduce_ptrs =
+            ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+                ReduceOp>::GetInstances();
+
+        if(reduce_ptrs.empty())
+        {
+            throw std::runtime_error("Wrong! No device REDUCE instance found");
+        };
+
+        std::array<index_t, Rank> arrInLengths;
+        std::array<index_t, Rank> arrInStrides;
+        std::array<index_t, NumOutDim> arrOutLengths;
+        std::array<index_t, NumOutDim> arrOutStrides;
+
+        ck::ranges::copy(inLengths, arrInLengths.begin());
+        ck::ranges::copy(inStrides, arrInStrides.begin());
+        ck::ranges::copy(outLengths, arrOutLengths.begin());
+        ck::ranges::copy(outStrides, arrOutStrides.begin());
+
+        if(do_verification)
+        {
+            using ReferenceReduceInstance =
+                ck::tensor_operation::host::ReferenceReduce<InDataType,
+                                                            AccDataType,
+                                                            OutDataType,
+                                                            Rank,
+                                                            NumReduceDim,
+                                                            ReduceOperation,
+                                                            InElementwiseOperation,
+                                                            AccElementwiseOperation,
+                                                            PropagateNan,
+                                                            OutputIndex>;
+
+            auto reduce_ref = ReferenceReduceInstance{};
+
+            auto argument_ptr_ref = reduce_ref.MakeArgumentPointer(arrInLengths,
+                                                                   arrInStrides,
+                                                                   arrOutLengths,
+                                                                   arrOutStrides,
+                                                                   reduceDims,
+                                                                   static_cast<double>(alpha),
+                                                                   static_cast<double>(beta),
+                                                                   in.mData.data(),
+                                                                   nullptr,
+                                                                   out_ref.mData.data(),
+                                                                   out_indices_ref.mData.data(),
+                                                                   in_elementwise_op,
+                                                                   acc_elementwise_op);
+
+            if(!reduce_ref.IsSupportedArgument(argument_ptr_ref.get()))
+            {
+                std::cout
+                    << "The runtime parameters not supported by the reduce reference, exiting!"
+                    << std::endl;
+                return (false);
+            };
+
+            auto invoker_ptr_ref = reduce_ref.MakeInvokerPointer();
+
+            (void)invoker_ptr_ref->Run(argument_ptr_ref.get());
+        };
+
+        for(auto& reduce_ptr : reduce_ptrs)
+        {
+            auto argument_ptr = reduce_ptr->MakeArgumentPointer(arrInLengths,
+                                                                arrInStrides,
+                                                                arrOutLengths,
+                                                                arrOutStrides,
+                                                                reduceDims,
+                                                                static_cast<double>(alpha),
+                                                                static_cast<double>(beta),
+                                                                in_dev.GetDeviceBuffer(),
+                                                                nullptr,
+                                                                out_dev.GetDeviceBuffer(),
+                                                                out_indices_dev.GetDeviceBuffer(),
+                                                                in_elementwise_op,
+                                                                acc_elementwise_op);
+
+            if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
+                continue;
+            else
+                num_kernel++;
+
+            std::string reduce_name = reduce_ptr->GetTypeString();
+
+            auto invoker_ptr = reduce_ptr->MakeInvokerPointer();
+
+            float avg_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t num_bytes =
+                invariant_total_length * reduce_total_length * sizeof(InDataType) +
+                invariant_total_length * sizeof(OutDataType);
+
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            if(time_kernel)
+                std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                          << reduce_name << std::endl;
+
+            if(gb_per_sec > best_gb_per_sec)
+            {
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                bool single_pass;
+
+                out_dev.FromDevice(out.mData.data());
+                single_pass = ck::utils::check_err(out, out_ref);
+
+                if(OutputIndex)
+                {
+                    out_indices_dev.FromDevice(out_indices.mData.data());
+                    single_pass = single_pass && ck::utils::check_err(out_indices, out_indices_ref);
+                };
+
+                if(!single_pass)
+                {
+                    std::cout << "Fail Info: " << reduce_ptr->GetTypeString() << std::endl;
+                }
+
+                pass = pass && single_pass;
+            };
+
+            if(do_dumpout)
+            {
+                dumpBufferToFile("dump_in.bin", in.mData.data(), in.mDesc.GetElementSize());
+                dumpBufferToFile("dump_out.bin", out.mData.data(), out.mDesc.GetElementSize());
+                dumpBufferToFile(
+                    "dump_out_host.bin", out_ref.mData.data(), out_ref.mDesc.GetElementSize());
+                if(OutputIndex)
+                {
+                    dumpBufferToFile("dump_indices.bin",
+                                     out_indices.mData.data(),
+                                     out_indices.mDesc.GetElementSize());
+                    dumpBufferToFile("dump_indices_host.bin",
+                                     out_indices_ref.mData.data(),
+                                     out_indices_ref.mDesc.GetElementSize());
+                };
+            };
+        };
+
+        if(time_kernel && num_kernel > 0)
+            std::cout << "Best Perf: " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s"
+                      << std::endl;
+    }
+    else
+    {
+        throw std::runtime_error(
+            "The requested reduction operation is not supported, please check!");
+    };
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is applicable" << std::endl;
+        return false;
+    };
+
+    return pass;
+};
+
+template <typename InDataType, typename AccDataType, typename OutDataType>
+bool profile_reduce_impl(bool do_verification,
+                         int init_method,
+                         bool do_dumpout,
+                         bool time_kernel,
+                         const std::vector<size_t>& inLengths,
+                         const std::vector<int>& reduceDims,
+                         ReduceTensorOp ReduceOpId,
+                         bool PropagateNan,
+                         bool UseIndex,
+                         float alpha,
+                         float beta)
+{
+    bool matched = false;
+    bool pass    = true;
+
+    using tuple_of_description_instances =
+        tensor_operation::device::instance::reduce_description_instances;
+
+    const auto tuple_object = tuple_of_description_instances{};
+
+    static_for<0, std::tuple_size<tuple_of_description_instances>::value, 1>{}([&](auto i) {
+        if(matched)
+            return;
+
+        using descType = remove_cvref_t<decltype(std::get<i>(tuple_object))>;
+
+        if(!description_match(
+               descType{}, inLengths.size(), reduceDims, ReduceOpId, PropagateNan, UseIndex))
+            return;
+
+        std::array<ck::index_t, descType::NumReduceDim_> arrReduceDims;
+
+        ck::ranges::copy(reduceDims, arrReduceDims.begin());
+
+        pass = pass && profile_reduce_impl_impl<InDataType,
+                                                AccDataType,
+                                                OutDataType,
+                                                descType::Rank_,
+                                                descType::NumReduceDim_,
+                                                static_cast<ReduceTensorOp>(descType::ReduceOpId_),
+                                                descType::PropagateNan_,
+                                                descType::UseIndex_>(do_verification,
+                                                                     init_method,
+                                                                     do_dumpout,
+                                                                     time_kernel,
+                                                                     inLengths,
+                                                                     arrReduceDims,
+                                                                     alpha,
+                                                                     beta);
+
+        matched = true;
+    });
+
+    return pass;
+};
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/include/profiler/profile_softmax_impl.hpp b/profiler/include/profiler/profile_softmax_impl.hpp
new file mode 100644
index 0000000000..daaf565149
--- /dev/null
+++ b/profiler/include/profiler/profile_softmax_impl.hpp
@@ -0,0 +1,227 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <iomanip>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/utility/data_type.hpp"
+
+namespace ck {
+namespace profiler {
+
+enum struct SoftmaxDataType
+{
+    F32_F32, // in, out
+    F16_F16,
+    BF16_BF16,
+    INT8_INT8,
+};
+
+// clang-format off
+template <typename SoftmaxDataType> std::string type_to_string();
+template <> std::string type_to_string<float>()   { return "f32"; }
+template <> std::string type_to_string<half_t>()  { return "f16"; }
+template <> std::string type_to_string<bhalf_t>() { return "bf16"; }
+template <> std::string type_to_string<int8_t>()  { return "int8"; }
+template <> std::string type_to_string<int32_t>() { return "int32"; }
+// clang-format on
+
+template <typename InDataType,
+          typename AccDataType,
+          typename OutDataType,
+          index_t Rank,
+          index_t NumReduceDim>
+bool profile_softmax_impl(int do_verification,
+                          int init_method,
+                          bool do_log,
+                          bool time_kernel,
+                          std::vector<index_t> in_length,
+                          std::vector<index_t> in_strides,
+                          std::vector<index_t> reduce_dims,
+                          double alpha,
+                          double beta)
+{
+    if(Rank != in_length.size())
+    {
+        throw std::runtime_error("Input tensor rank is different from template argument Rank!");
+    };
+
+    if(NumReduceDim != reduce_dims.size())
+    {
+        throw std::runtime_error(
+            "Input reduce_dims rank is different from template argument NumReduceDim!");
+    };
+
+    Tensor<InDataType> in = in_strides.empty() ? Tensor<InDataType>(in_length)
+                                               : Tensor<InDataType>(in_length, in_strides);
+    Tensor<OutDataType> out(in.mDesc);
+    Tensor<OutDataType> prior_out(in.mDesc);
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        ck::utils::FillUniformDistributionIntegerValue<InDataType>{-5.f, 5.f}(in.begin(), in.end());
+        ck::utils::FillUniformDistributionIntegerValue<OutDataType>{-5.f, 5.f}(prior_out.begin(),
+                                                                               prior_out.end());
+        break;
+    default:
+        ck::utils::FillUniformDistribution<InDataType>{0.0f, 1.0f}(in);
+        ck::utils::FillUniformDistribution<OutDataType>{-0.5f, 0.5f}(prior_out);
+    }
+
+    Tensor<OutDataType> out_ref(prior_out);
+
+    if(do_verification)
+    {
+        using ReferenceSoftmax =
+            tensor_operation::host::ReferenceSoftmax<InDataType, OutDataType, AccDataType>;
+        ReferenceSoftmax{}.MakeInvoker().Run({in, out_ref, alpha, beta, reduce_dims});
+    }
+
+    DeviceMem in_dev(in.GetElementSpaceSizeInBytes());
+    DeviceMem out_dev(out.GetElementSpaceSizeInBytes());
+    in_dev.ToDevice(in.data());
+
+    std::vector<index_t> in_tensor_lengths(in.GetLengths().begin(), in.GetLengths().end());
+    std::vector<index_t> in_tensor_strides(in.GetStrides().begin(), in.GetStrides().end());
+
+    // add device softmax instances
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using DeviceOp    = tensor_operation::device::DeviceSoftmax<InDataType,
+                                                             AccDataType,
+                                                             OutDataType,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             Rank,
+                                                             NumReduceDim>;
+
+    // get device op instances
+    const auto instances = tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+    std::cout << "found " << instances.size() << " instances" << std::endl;
+
+    if(instances.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device normalization instance found");
+    }
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    std::vector<bool> instance_pass;
+
+    for(auto& inst_ptr : instances)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(in_tensor_lengths,
+                                                          in_tensor_strides,
+                                                          reduce_dims,
+                                                          alpha,
+                                                          beta,
+                                                          in_dev.GetDeviceBuffer(),
+                                                          out_dev.GetDeviceBuffer(),
+                                                          PassThrough{},
+                                                          PassThrough{});
+
+        if(!inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
+            LogRange(std::cout << "input lengths = [", in_length, ", ")
+                << "], "
+                << "scaler = [" << alpha << ", " << beta << "]";
+            LogRange(std::cout << ", reduce dims = [", reduce_dims, ", ") << "]." << std::endl;
+            instance_pass.push_back(true);
+            continue;
+        }
+
+        out_dev.ToDevice(prior_out.data());
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+        float avg_time   = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        if(time_kernel)
+        {
+            std::size_t num_bytes =
+                in.GetElementSize() * sizeof(InDataType) +
+                (beta == 0.0f ? 1 : 2) * out.GetElementSize() * sizeof(OutDataType);
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << inst_ptr->GetTypeString() << std::endl;
+
+            if(avg_time < best_avg_time)
+            {
+                best_instance_name = inst_ptr->GetTypeString();
+                best_avg_time      = avg_time;
+                best_gb_per_sec    = gb_per_sec;
+            }
+        }
+
+        if(do_verification)
+        {
+            out_dev.FromDevice(out.data());
+            bool pass = true;
+            if(std::is_same<InDataType, int8_t>::value)
+            {
+                pass = pass && ck::utils::check_err(
+                                   out.mData, out_ref.mData, "Error: Incorrect results!", 0, 1);
+                if(do_log)
+                {
+                    LogRangeAsType<int>(std::cout << "in  : ", in.mData, ",") << std::endl;
+                    LogRangeAsType<int>(std::cout << "out_ref  : ", out_ref.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<int>(std::cout << "out  : ", out.mData, ",") << std::endl;
+                }
+            }
+            else
+            {
+                pass = pass && ck::utils::check_err(out.mData, out_ref.mData);
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "in  : ", in.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "out_ref  : ", out_ref.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "out  : ", out.mData, ",") << std::endl;
+                }
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "input lengths = [", in_length, ", ")
+                    << "], "
+                    << "scaler = [" << alpha << ", " << beta << "]." << std::endl;
+            }
+            instance_pass.push_back(pass);
+        }
+    }
+    if(time_kernel)
+    {
+        std::cout << "Best Perf for datatype = " << type_to_string<InDataType>() << "_"
+                  << type_to_string<OutDataType>() << ", ";
+        LogRange(std::cout << "length = ", in_tensor_lengths, ",") << ", ";
+        LogRange(std::cout << "stride = ", in_tensor_strides, ",") << ", ";
+        LogRange(std::cout << "reduce dims ", reduce_dims, ",") << ", ";
+        std::cout << "alpha = " << alpha << ", "
+                  << "beta = " << beta << ", " << best_avg_time << " ms, " << best_gb_per_sec
+                  << " GB/s, " << best_instance_name << std::endl;
+    }
+    return std::all_of(
+        std::begin(instance_pass), std::end(instance_pass), [](bool p) { return p; });
+}
+
+} // namespace profiler
+} // namespace ck
diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt
new file mode 100644
index 0000000000..7da7613f26
--- /dev/null
+++ b/profiler/src/CMakeLists.txt
@@ -0,0 +1,102 @@
+# ckProfiler
+set(PROFILER_SOURCES
+    profiler.cpp
+    profile_gemm.cpp
+    profile_gemm_splitk.cpp
+    profile_gemm_bias_add_reduce.cpp
+    profile_gemm_add_multiply.cpp
+    profile_gemm_multiply_add.cpp
+    profile_gemm_reduce.cpp
+    profile_batched_gemm.cpp
+    profile_batched_gemm_reduce.cpp
+    profile_conv_fwd.cpp
+    profile_conv_fwd_bias_relu.cpp
+    profile_conv_fwd_bias_relu_add.cpp
+    profile_conv_bwd_data.cpp
+    profile_grouped_conv_fwd.cpp
+    profile_grouped_conv_bwd_weight.cpp
+    profile_reduce.cpp
+    profile_groupnorm.cpp
+    profile_layernorm.cpp
+    profile_max_pool3d_fwd.cpp
+    profile_avg_pool3d_bwd.cpp
+    profile_max_pool3d_bwd.cpp
+    profile_softmax.cpp
+    profile_batchnorm_fwd.cpp
+    profile_batchnorm_bwd.cpp
+    profile_batchnorm_infer.cpp
+    profile_contraction_bilinear.cpp
+    profile_contraction_scale.cpp
+    profile_grouped_conv_bwd_data.cpp
+    profile_image_to_column.cpp
+)
+if(DL_KERNELS)
+  list(APPEND PROFILER_SOURCES profile_batched_gemm_multi_d.cpp)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  list(APPEND PROFILER_SOURCES profile_batched_gemm_gemm.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
+  list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
+  list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
+  list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
+endif()
+
+set(PROFILER_EXECUTABLE ckProfiler)
+
+add_executable(${PROFILER_EXECUTABLE} ${PROFILER_SOURCES})
+target_compile_options(${PROFILER_EXECUTABLE} PRIVATE -Wno-global-constructors)
+
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool3d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_avg_pool3d_bwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_max_pool_bwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance)
+if(DL_KERNELS)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_streamk_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
+endif()
+rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)
diff --git a/profiler/src/profile_avg_pool3d_bwd.cpp b/profiler/src/profile_avg_pool3d_bwd.cpp
new file mode 100644
index 0000000000..0ff50a5292
--- /dev/null
+++ b/profiler/src/profile_avg_pool3d_bwd.cpp
@@ -0,0 +1,175 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <unordered_map>
+
+#include "profiler/data_type_enum.hpp"
+#include "profiler/profile_avg_pool3d_bwd_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+struct maxPoolbwdArgParser
+{
+    std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}},
+                                                                   {"wsize", {}},
+                                                                   {"wstride", {}},
+                                                                   {"wdilation", {}},
+                                                                   {"pad1", {}},
+                                                                   {"pad2", {}}};
+
+    bool parse_opt(int argc, char* argv[], const std::string& key, int i)
+    {
+        if(std::string("--") + key == argv[i])
+        {
+            int pos = i;
+            while(++i < argc && argv[i][0] != '-') {}
+            int end = i;
+            for(int j = pos + 1; j < end; j++)
+            {
+                long_opts[key].push_back(std::stoi(argv[j]));
+            }
+            return true;
+        }
+        return false;
+    }
+
+    void operator()(int argc, char* argv[])
+    {
+        for(auto& kv : long_opts)
+        {
+            for(int i = 1; i < argc; i++)
+            {
+                if(parse_opt(argc, argv, kv.first, i))
+                    break;
+            }
+        }
+    }
+};
+
+void print_help_avg_pool3d_bwd()
+{
+    std::cout << "arg1: data type (0: fp16; 1: fp32; 5: bf16)\n"
+              << "arg2: verification (0: no; 1: yes)\n"
+              << "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg4: print tensor value (0: no; 1: yes)\n"
+              << "arg5: time kernel (0=no, 1=yes)\n"
+              << "--length: input tensor length for NCDHW(e.g, --length 2 32 30 30 30) \n"
+              << "--wsize: window size for ZYX (e.g, --wsize 2 2 2) \n"
+              << "--wstride: window stride for DHW (e.g, --wstride 2 2 2) \n"
+              << "--wdilation: window dilation for DHW (e.g, --wdilation 1 1 1) \n"
+              << "--pad1: left side of padding in DHW (e.g, --pad1 1 1 1) \n"
+              << "--pad2: right side of padding in DHW (e.g, --pad2 1 1 1) \n"
+              << "eg: ckProfiler avg_pool3d_bwd 0 1 2 0 1 --length 2 32 30 30 30 --wsize 2 2 2 "
+                 "--wstride 2 2 2 --wdilation 1 1 1 --pad1 1 1 1 --pad2 1 1 1"
+              << std::endl;
+}
+
+int profile_avg_pool3d_bwd(int argc, char* argv[])
+{
+    ck::DataTypeEnum data_type = ck::DataTypeEnum::Half;
+    bool do_verification       = true;
+    int init_method            = 0;
+    bool do_log                = false;
+    bool time_kernel           = true;
+
+    std::vector<index_t> in_length = {2, 32, 30, 30, 30};
+    std::vector<index_t> wsize     = {2, 2, 2};
+    std::vector<index_t> wstride   = {2, 2, 2};
+    std::vector<index_t> wdilation = {1, 1, 1};
+    std::vector<index_t> pad1      = {1, 1, 1};
+    std::vector<index_t> pad2      = {1, 1, 1};
+
+    if(argc != 2 && argc != 33)
+    {
+        print_help_avg_pool3d_bwd();
+        return 0;
+    }
+    else if(argc == 33)
+    {
+        data_type       = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
+        do_verification = std::stoi(argv[3]);
+        init_method     = std::stoi(argv[4]);
+        do_log          = std::stoi(argv[5]);
+        time_kernel     = std::stoi(argv[6]);
+
+        // parse the long options
+        maxPoolbwdArgParser arg_parser;
+        arg_parser(argc, argv);
+        in_length = arg_parser.long_opts["length"];
+        wsize     = arg_parser.long_opts["wsize"];
+        wstride   = arg_parser.long_opts["wstride"];
+        wdilation = arg_parser.long_opts["wdilation"];
+        pad1      = arg_parser.long_opts["pad1"];
+        pad2      = arg_parser.long_opts["pad2"];
+    }
+
+#ifdef CK_ENABLE_FP16
+    using F16 = ck::half_t;
+#endif
+#ifdef CK_ENABLE_BF16
+    using BF16 = ck::bhalf_t;
+#endif
+#ifdef CK_ENABLE_FP32
+    using F32 = float;
+#endif
+    using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+    if(false)
+        ;
+#ifdef CK_ENABLE_FP16
+    else if(data_type == ck::DataTypeEnum::Half)
+    {
+        ck::profiler::profile_avg_pool3d_bwd_impl<F16, F16, F16, NDHWC, NDHWC>(do_verification,
+                                                                               init_method,
+                                                                               do_log,
+                                                                               time_kernel,
+                                                                               in_length,
+                                                                               wsize,
+                                                                               wstride,
+                                                                               wdilation,
+                                                                               pad1,
+                                                                               pad2);
+    }
+#endif
+#ifdef CK_ENABLE_BF16
+    else if(data_type == ck::DataTypeEnum::BFloat16)
+    {
+        ck::profiler::profile_avg_pool3d_bwd_impl<BF16, BF16, BF16, NDHWC, NDHWC>(do_verification,
+                                                                                  init_method,
+                                                                                  do_log,
+                                                                                  time_kernel,
+                                                                                  in_length,
+                                                                                  wsize,
+                                                                                  wstride,
+                                                                                  wdilation,
+                                                                                  pad1,
+                                                                                  pad2);
+    }
+#endif
+#ifdef CK_ENABLE_FP32
+    else if(data_type == ck::DataTypeEnum::Float)
+    {
+        ck::profiler::profile_avg_pool3d_bwd_impl<F32, F32, F32, NDHWC, NDHWC>(do_verification,
+                                                                               init_method,
+                                                                               do_log,
+                                                                               time_kernel,
+                                                                               in_length,
+                                                                               wsize,
+                                                                               wstride,
+                                                                               wdilation,
+                                                                               pad1,
+                                                                               pad2);
+    }
+#endif
+    else
+    {
+        throw std::runtime_error("not implemented yet");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("avg_pool3d_bwd", "max_pool bwd", profile_avg_pool3d_bwd);
diff --git a/profiler/src/profile_batched_gemm.cpp b/profiler/src/profile_batched_gemm.cpp
new file mode 100644
index 0000000000..222532b7bb
--- /dev/null
+++ b/profiler/src/profile_batched_gemm.cpp
@@ -0,0 +1,222 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdint>
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_batched_gemm_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+};
+
+enum struct GemmDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+#define OP_NAME "batched_gemm"
+#define OP_DESC "Batched GEMM"
+
+int profile_batched_gemm(int argc, char* argv[])
+{
+    if(argc != 18)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16, 2: bf16, 3: int8)\n");
+        printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n");
+        printf("                     1: A[g, m, k] * B[g, n, k] = C[g, m, n];\n");
+        printf("                     2: A[g, k, m] * B[g, k, n] = C[g, m, n];\n");
+        printf("                     3: A[g, k, m] * B[g, n, k] = C[g, m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
+        printf("arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideC = std::stoi(argv[13]);
+
+    const int BatchStrideA = std::stoi(argv[14]);
+    const int BatchStrideB = std::stoi(argv[15]);
+    const int BatchStrideC = std::stoi(argv[16]);
+
+    const int BatchCount = std::stoi(argv[17]);
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+    using INT8 = int8_t;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile =
+        [&](auto a_type, auto b_type, auto c_type, auto a_layout, auto b_layout, auto c_layout) {
+            using ADataType = decltype(a_type);
+            using BDataType = decltype(b_type);
+            using CDataType = decltype(c_type);
+
+            using ALayout = decltype(a_layout);
+            using BLayout = decltype(b_layout);
+            using CLayout = decltype(c_layout);
+
+            const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+            const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+            const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
+
+            const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA;
+            const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB;
+            const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+            const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Row> ? M : K) * StrideA_;
+            const int DefaultBatchStrideB = (ck::is_same_v<BLayout, Row> ? K : N) * StrideB_;
+            const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Row> ? M : N) * StrideC_;
+
+            const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA;
+            const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB;
+            const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC;
+
+            using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+            using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+            using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+            using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
+            bool pass = ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                                BDataType,
+                                                                CDataType,
+                                                                ALayout,
+                                                                BLayout,
+                                                                CLayout,
+                                                                AElementOp,
+                                                                BElementOp,
+                                                                CElementOp,
+                                                                DeviceOp>(do_verification,
+                                                                          init_method,
+                                                                          do_log,
+                                                                          time_kernel,
+                                                                          M,
+                                                                          N,
+                                                                          K,
+                                                                          BatchStrideA_,
+                                                                          BatchStrideB_,
+                                                                          BatchStrideC_,
+                                                                          StrideA_,
+                                                                          StrideB_,
+                                                                          StrideC_,
+                                                                          BatchCount);
+
+            return pass ? 0 : 1;
+        };
+
+    if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, Col{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F16{}, F16{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F16{}, F16{}, F16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(F16{}, F16{}, F16{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(F16{}, F16{}, F16{}, Col{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(BF16{}, BF16{}, BF16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(BF16{}, BF16{}, BF16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(BF16{}, BF16{}, BF16{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(BF16{}, BF16{}, BF16{}, Col{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(INT8{}, INT8{}, INT8{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(INT8{}, INT8{}, INT8{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(INT8{}, INT8{}, INT8{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(INT8{}, INT8{}, INT8{}, Col{}, Col{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm);
diff --git a/profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp b/profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp
new file mode 100644
index 0000000000..3d29c4b84a
--- /dev/null
+++ b/profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp
@@ -0,0 +1,214 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+#define OP_NAME "batched_gemm_add_relu_gemm_add"
+#define OP_DESC "Batched GEMM+Add+Relu+GEMM+Add"
+
+int profile_batched_gemm_add_relu_gemm_add(int argc, char* argv[])
+{
+    enum struct GemmMatrixLayout
+    {
+        MK_NK_MN_NO_MO_MO, // 0
+        MK_NK_MN_ON_MO_MO, // 1
+    };
+
+    enum struct GemmDataType
+    {
+        F32_F32_F32_F32_F32_F32, // 0
+        F16_F16_F16_F16_F16_F16, // 1
+    };
+
+    GemmDataType data_type  = GemmDataType::F16_F16_F16_F16_F16_F16;
+    GemmMatrixLayout layout = GemmMatrixLayout::MK_NK_MN_NO_MO_MO;
+    bool do_verification    = true;
+    int init_method         = 1;
+    bool do_log             = 0;
+    bool time_kernel        = false;
+
+    // GEMM shape
+    ck::index_t M             = 1024;
+    ck::index_t N             = 1024;
+    ck::index_t K             = 64;
+    ck::index_t O             = 128;
+    ck::index_t BatchCount    = 4;
+    ck::index_t StrideA0      = -1;
+    ck::index_t StrideB0      = -1;
+    ck::index_t StrideD0      = -1;
+    ck::index_t StrideB1      = -1;
+    ck::index_t StrideD1      = -1;
+    ck::index_t StrideE1      = -1;
+    ck::index_t BatchStrideA0 = -1;
+    ck::index_t BatchStrideB0 = -1;
+    ck::index_t BatchStrideD0 = -1;
+    ck::index_t BatchStrideB1 = -1;
+    ck::index_t BatchStrideD1 = -1;
+    ck::index_t BatchStrideE1 = -1;
+
+    if(argc == 8)
+    {
+        data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+        layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+        do_verification = std::stoi(argv[4]);
+        init_method     = std::stoi(argv[5]);
+        do_log          = std::stoi(argv[6]);
+        time_kernel     = std::stoi(argv[7]);
+    }
+    else if(argc == 13)
+    {
+        data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+        layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+        do_verification = std::stoi(argv[4]);
+        init_method     = std::stoi(argv[5]);
+        do_log          = std::stoi(argv[6]);
+        time_kernel     = std::stoi(argv[7]);
+
+        M          = std::stoi(argv[8]);
+        N          = std::stoi(argv[9]);
+        K          = std::stoi(argv[10]);
+        O          = std::stoi(argv[11]);
+        BatchCount = std::stoi(argv[12]);
+    }
+    else if(argc == 25)
+    {
+        data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+        layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+        do_verification = std::stoi(argv[4]);
+        init_method     = std::stoi(argv[5]);
+        do_log          = std::stoi(argv[6]);
+        time_kernel     = std::stoi(argv[7]);
+
+        M          = std::stoi(argv[8]);
+        N          = std::stoi(argv[9]);
+        K          = std::stoi(argv[10]);
+        O          = std::stoi(argv[11]);
+        BatchCount = std::stoi(argv[12]);
+
+        StrideA0 = std::stoi(argv[13]);
+        StrideB0 = std::stoi(argv[14]);
+        StrideD0 = std::stoi(argv[15]);
+        StrideB1 = std::stoi(argv[16]);
+        StrideD1 = std::stoi(argv[17]);
+        StrideE1 = std::stoi(argv[18]);
+
+        BatchStrideA0 = std::stoi(argv[19]);
+        BatchStrideB0 = std::stoi(argv[20]);
+        BatchStrideD0 = std::stoi(argv[21]);
+        BatchStrideB1 = std::stoi(argv[22]);
+        BatchStrideD1 = std::stoi(argv[23]);
+        BatchStrideE1 = std::stoi(argv[24]);
+    }
+    else
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (1: fp16)\n");
+        printf("arg3: matrix layout (0: Relu(A0[m, k] * B0[n, k] + D0[m, n]) * B1[n, o] + D1[m, o] "
+               "= E1[m, o]; 1: Relu(A0[m, k] * B0[n, k] + D0[m, n]) * B1[o, n] + D1[m, o] = "
+               "E1[m, o];)\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 12: M, N, K, O, Batch\n");
+        printf("arg13 to 18: StrideA0, StrideB0, StrideD0, StrideB1, StrideD1, StrideE1\n");
+        printf("arg19 to 24: BatchStrideA0, BatchStrideB0, BatchStrideD0, BatchStrideB1, "
+               "BatchStrideD1, BatchStrideE1 \n");
+        exit(1);
+    }
+
+    if(data_type == GemmDataType::F16_F16_F16_F16_F16_F16 &&
+       layout == GemmMatrixLayout::MK_NK_MN_NO_MO_MO)
+    {
+        ck::profiler::profile_batched_gemm_add_relu_gemm_add_impl<Row,            // A0Layout,
+                                                                  Col,            // B0Layout,
+                                                                  ck::Tuple<Row>, // D0sLayout,
+                                                                  Row,            // B1Layout,
+                                                                  ck::Tuple<Row>, // D1sLayout,
+                                                                  Row,            // E1Layout,
+                                                                  F16,            // A0DataType,
+                                                                  F16,            // B0DataType,
+                                                                  ck::Tuple<F16>, // D0DataType,
+                                                                  F16,            // B1DataType,
+                                                                  ck::Tuple<F16>, // D1sDataType
+                                                                  F16>            // E1DataType,
+            (do_verification,
+             init_method,
+             do_log,
+             time_kernel,
+             M,
+             N,
+             K,
+             O,
+             BatchCount,
+             StrideA0,
+             StrideB0,
+             StrideD0,
+             StrideB1,
+             StrideD1,
+             StrideE1,
+             BatchStrideA0,
+             BatchStrideB0,
+             BatchStrideD0,
+             BatchStrideB1,
+             BatchStrideD1,
+             BatchStrideE1);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16_F16_F16_F16 &&
+            layout == GemmMatrixLayout::MK_NK_MN_ON_MO_MO)
+    {
+        ck::profiler::profile_batched_gemm_add_relu_gemm_add_impl<Row,            // A0Layout,
+                                                                  Col,            // B0Layout,
+                                                                  ck::Tuple<Row>, // D0sLayout,
+                                                                  Col,            // B1Layout,
+                                                                  ck::Tuple<Row>, // D1sLayout,
+                                                                  Row,            // E1Layout,
+                                                                  F16,            // A0DataType,
+                                                                  F16,            // B0DataType,
+                                                                  ck::Tuple<F16>, // D0DataType,
+                                                                  F16,            // B1DataType,
+                                                                  ck::Tuple<F16>, // D1sDataType
+                                                                  F16>            // E1DataType,
+            (do_verification,
+             init_method,
+             do_log,
+             time_kernel,
+             M,
+             N,
+             K,
+             O,
+             BatchCount,
+             StrideA0,
+             StrideB0,
+             StrideD0,
+             StrideB1,
+             StrideD1,
+             StrideE1,
+             BatchStrideA0,
+             BatchStrideB0,
+             BatchStrideD0,
+             BatchStrideB1,
+             BatchStrideD1,
+             BatchStrideE1);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this data_type & layout is not implemented");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_add_relu_gemm_add);
diff --git a/profiler/src/profile_batched_gemm_gemm.cpp b/profiler/src/profile_batched_gemm_gemm.cpp
new file mode 100644
index 0000000000..9a99874d1c
--- /dev/null
+++ b/profiler/src/profile_batched_gemm_gemm.cpp
@@ -0,0 +1,187 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_batched_gemm_gemm_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+#define OP_NAME "batched_gemm_gemm"
+#define OP_DESC "Batched GEMM+GEMM"
+
+int profile_batched_gemm_gemm(int argc, char* argv[])
+{
+    enum struct GemmMatrixLayout
+    {
+        MK_NK_NO_MO, // 0
+        MK_NK_ON_MO, // 0
+    };
+
+    enum struct GemmDataType
+    {
+        F32_F32_F32_F32, // 0
+        F16_F16_F16_F16, // 1
+    };
+
+    GemmDataType data_type  = GemmDataType::F16_F16_F16_F16;
+    GemmMatrixLayout layout = GemmMatrixLayout::MK_NK_NO_MO;
+    bool do_verification    = true;
+    int init_method         = 1;
+    bool do_log             = 0;
+    bool time_kernel        = false;
+
+    // GEMM shape
+    ck::index_t M             = 1024;
+    ck::index_t N             = 1024;
+    ck::index_t K             = 64;
+    ck::index_t O             = 128;
+    ck::index_t BatchCount    = 4;
+    ck::index_t StrideA0      = -1;
+    ck::index_t StrideB0      = -1;
+    ck::index_t StrideB1      = -1;
+    ck::index_t StrideE1      = -1;
+    ck::index_t BatchStrideA0 = -1;
+    ck::index_t BatchStrideB0 = -1;
+    ck::index_t BatchStrideB1 = -1;
+    ck::index_t BatchStrideE1 = -1;
+
+    if(argc == 8)
+    {
+        data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+        layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+        do_verification = std::stoi(argv[4]);
+        init_method     = std::stoi(argv[5]);
+        do_log          = std::stoi(argv[6]);
+        time_kernel     = std::stoi(argv[7]);
+    }
+    else if(argc == 13)
+    {
+        data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+        layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+        do_verification = std::stoi(argv[4]);
+        init_method     = std::stoi(argv[5]);
+        do_log          = std::stoi(argv[6]);
+        time_kernel     = std::stoi(argv[7]);
+
+        M          = std::stoi(argv[8]);
+        N          = std::stoi(argv[9]);
+        K          = std::stoi(argv[10]);
+        O          = std::stoi(argv[11]);
+        BatchCount = std::stoi(argv[12]);
+    }
+    else if(argc == 21)
+    {
+        data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+        layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+        do_verification = std::stoi(argv[4]);
+        init_method     = std::stoi(argv[5]);
+        do_log          = std::stoi(argv[6]);
+        time_kernel     = std::stoi(argv[7]);
+
+        M          = std::stoi(argv[8]);
+        N          = std::stoi(argv[9]);
+        K          = std::stoi(argv[10]);
+        O          = std::stoi(argv[11]);
+        BatchCount = std::stoi(argv[12]);
+
+        StrideA0 = std::stoi(argv[13]);
+        StrideB0 = std::stoi(argv[14]);
+        StrideB1 = std::stoi(argv[15]);
+        StrideE1 = std::stoi(argv[16]);
+
+        BatchStrideA0 = std::stoi(argv[17]);
+        BatchStrideB0 = std::stoi(argv[18]);
+        BatchStrideB1 = std::stoi(argv[19]);
+        BatchStrideE1 = std::stoi(argv[20]);
+    }
+    else
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (1: fp16)\n");
+        printf("arg3: matrix layout (0: Relu(A0[m, k] * B0[n, k] + D0[m, n]) * B1[n, o] + D1[m, o] "
+               "= E1[m, o];  1: Relu(A0[m, k] * B0[n, k] + D0[m, n]) * B1[o, n] + D1[m, o] = E1[m, "
+               "o];)\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 12: M, N, K, O, Batch\n");
+        printf("arg13 to 16: StrideA0, StrideB0, StrideB1, StrideE1\n");
+        printf("arg17 to 20: BatchStrideA0, BatchStrideB0, BatchStrideB1, BatchStrideE1 \n");
+        exit(1);
+    }
+
+    if(data_type == GemmDataType::F16_F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_NO_MO)
+    {
+        ck::profiler::profile_batched_gemm_gemm_impl<F16, // A0DataType,
+                                                     F16, // B0DataType,
+                                                     F16, // B1DataType,
+                                                     F16, // E1DataType,
+                                                     Row, // A0Layout,
+                                                     Col, // B0Layout,
+                                                     Row, // B1Layout,
+                                                     Row> // E1Layout,
+            (do_verification,
+             init_method,
+             do_log,
+             time_kernel,
+             M,
+             N,
+             K,
+             O,
+             BatchCount,
+             StrideA0,
+             StrideB0,
+             StrideB1,
+             StrideE1,
+             BatchStrideA0,
+             BatchStrideB0,
+             BatchStrideB1,
+             BatchStrideE1);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_ON_MO)
+    {
+        ck::profiler::profile_batched_gemm_gemm_impl<F16, // A0DataType,
+                                                     F16, // B0DataType,
+                                                     F16, // B1DataType,
+                                                     F16, // E1DataType,
+                                                     Row, // A0Layout,
+                                                     Col, // B0Layout,
+                                                     Col, // B1Layout,
+                                                     Row> // E1Layout,
+            (do_verification,
+             init_method,
+             do_log,
+             time_kernel,
+             M,
+             N,
+             K,
+             O,
+             BatchCount,
+             StrideA0,
+             StrideB0,
+             StrideB1,
+             StrideE1,
+             BatchStrideA0,
+             BatchStrideB0,
+             BatchStrideB1,
+             BatchStrideE1);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this data_type & layout is not implemented");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_gemm);
diff --git a/profiler/src/profile_batched_gemm_multi_d.cpp b/profiler/src/profile_batched_gemm_multi_d.cpp
new file mode 100644
index 0000000000..7cd4636d98
--- /dev/null
+++ b/profiler/src/profile_batched_gemm_multi_d.cpp
@@ -0,0 +1,194 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdint>
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_batched_gemm_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+};
+
+enum struct GemmDataType
+{
+    F16_F16_F16,    // 0
+    INT8_INT8_INT8, // 1
+};
+
+#define OP_NAME "batched_gemm_multi_d"
+#define OP_DESC "Batched GEMM multi D"
+
+int profile_batched_gemm_multi_d(int argc, char* argv[])
+{
+    if(argc != 18)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp16; 1: int8)\n");
+        printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n");
+        printf("                     1: A[g, m, k] * B[g, n, k] = C[g, m, n];\n");
+        printf("                     2: A[g, k, m] * B[g, k, n] = C[g, m, n];\n");
+        printf("                     3: A[g, k, m] * B[g, n, k] = C[g, m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
+        printf("arg8 to 17: M, N, K, StrideA, StrideB, StrideC, BatchStrideA, BatchStrideB, BatchStrideC, BatchCount\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideC = std::stoi(argv[13]);
+
+    const int BatchStrideA = std::stoi(argv[14]);
+    const int BatchStrideB = std::stoi(argv[15]);
+    const int BatchStrideC = std::stoi(argv[16]);
+
+    const int BatchCount = std::stoi(argv[17]);
+
+    using F16 = ck::half_t;
+#ifdef CK_ENABLE_INT8
+    using INT8 = int8_t;
+#endif
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile =
+        [&](auto a_type, auto b_type, auto c_type, auto a_layout, auto b_layout, auto c_layout) {
+            using ADataType  = decltype(a_type);
+            using BDataType  = decltype(b_type);
+            using CDataType  = decltype(c_type);
+            using DsDataType = ck::Tuple<>;
+
+            using ALayout  = decltype(a_layout);
+            using BLayout  = decltype(b_layout);
+            using CLayout  = decltype(c_layout);
+            using DsLayout = ck::Tuple<>;
+
+            const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+            const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+            const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
+
+            const int StrideA_ = (StrideA < 0) ? DefaultStrideA : StrideA;
+            const int StrideB_ = (StrideB < 0) ? DefaultStrideB : StrideB;
+            const int StrideC_ = (StrideC < 0) ? DefaultStrideC : StrideC;
+
+            const int DefaultBatchStrideA = (ck::is_same_v<ALayout, Row> ? M : K) * StrideA_;
+            const int DefaultBatchStrideB = (ck::is_same_v<BLayout, Row> ? K : N) * StrideB_;
+            const int DefaultBatchStrideC = (ck::is_same_v<CLayout, Row> ? M : N) * StrideC_;
+
+            const int BatchStrideA_ = (BatchStrideA < 0) ? DefaultBatchStrideA : BatchStrideA;
+            const int BatchStrideB_ = (BatchStrideB < 0) ? DefaultBatchStrideB : BatchStrideB;
+            const int BatchStrideC_ = (BatchStrideC < 0) ? DefaultBatchStrideC : BatchStrideC;
+
+            using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+            using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+            using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+            using DeviceOp = ck::tensor_operation::device::DeviceBatchedGemmMultiD<ALayout,
+                                                                                   BLayout,
+                                                                                   DsLayout,
+                                                                                   CLayout,
+                                                                                   ADataType,
+                                                                                   BDataType,
+                                                                                   DsDataType,
+                                                                                   CDataType,
+                                                                                   AElementOp,
+                                                                                   BElementOp,
+                                                                                   CElementOp>;
+
+            bool pass = ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                                BDataType,
+                                                                CDataType,
+                                                                ALayout,
+                                                                BLayout,
+                                                                CLayout,
+                                                                AElementOp,
+                                                                BElementOp,
+                                                                CElementOp,
+                                                                DeviceOp>(do_verification,
+                                                                          init_method,
+                                                                          do_log,
+                                                                          time_kernel,
+                                                                          M,
+                                                                          N,
+                                                                          K,
+                                                                          BatchStrideA_,
+                                                                          BatchStrideB_,
+                                                                          BatchStrideC_,
+                                                                          StrideA_,
+                                                                          StrideB_,
+                                                                          StrideC_,
+                                                                          BatchCount);
+
+            return pass ? 0 : 1;
+        };
+
+    if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F16{}, F16{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F16{}, F16{}, F16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(F16{}, F16{}, F16{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(F16{}, F16{}, F16{}, Col{}, Col{}, Row{});
+    }
+#ifdef CK_ENABLE_INT8
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(INT8{}, INT8{}, INT8{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(INT8{}, INT8{}, INT8{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(INT8{}, INT8{}, INT8{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(INT8{}, INT8{}, INT8{}, Col{}, Col{}, Row{});
+    }
+#endif
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_multi_d);
diff --git a/profiler/src/profile_batched_gemm_reduce.cpp b/profiler/src/profile_batched_gemm_reduce.cpp
new file mode 100644
index 0000000000..9620d63caf
--- /dev/null
+++ b/profiler/src/profile_batched_gemm_reduce.cpp
@@ -0,0 +1,159 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_batched_gemm_reduce_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "batched_gemm_reduce"
+#define OP_DESC "Batched GEMM+Reduce"
+
+int profile_batched_gemm_reduce(int argc, char* argv[])
+{
+    enum struct GemmMatrixLayout
+    {
+        MK_KN_MN, // 0
+        MK_NK_MN, // 1
+        KM_KN_MN, // 2
+        KM_NK_MN, // 3
+    };
+
+    enum struct GemmReduceDataType
+    {
+        F32_F32_F32_F32_F32, // 0
+        F16_F16_F16_F32_F32, // 1
+    };
+
+    if(argc != 15)
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
+        printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, BatchCount\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmReduceDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideC = std::stoi(argv[13]);
+
+    const int BatchCount = std::stoi(argv[14]);
+
+    if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                       ck::half_t,
+                                                       ck::half_t,
+                                                       float,
+                                                       ck::tensor_layout::gemm::RowMajor,
+                                                       ck::tensor_layout::gemm::RowMajor,
+                                                       ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? K : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            BatchCount);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                       ck::half_t,
+                                                       ck::half_t,
+                                                       float,
+                                                       ck::tensor_layout::gemm::RowMajor,
+                                                       ck::tensor_layout::gemm::ColumnMajor,
+                                                       ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? K : StrideA,
+            (StrideB < 0) ? K : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            BatchCount);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                       ck::half_t,
+                                                       ck::half_t,
+                                                       float,
+                                                       ck::tensor_layout::gemm::ColumnMajor,
+                                                       ck::tensor_layout::gemm::RowMajor,
+                                                       ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            BatchCount);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                       ck::half_t,
+                                                       ck::half_t,
+                                                       float,
+                                                       ck::tensor_layout::gemm::ColumnMajor,
+                                                       ck::tensor_layout::gemm::ColumnMajor,
+                                                       ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? K : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            BatchCount);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this data_type & layout is not implemented");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_reduce);
diff --git a/profiler/src/profile_batchnorm_bwd.cpp b/profiler/src/profile_batchnorm_bwd.cpp
new file mode 100644
index 0000000000..1738d53dbe
--- /dev/null
+++ b/profiler/src/profile_batchnorm_bwd.cpp
@@ -0,0 +1,207 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <getopt.h>
+
+#include "ck/library/utility/host_common_util.hpp"
+#include "profiler/profile_batchnorm_backward_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+using namespace std;
+
+static const struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                             {"reduceDims", required_argument, nullptr, 'R'},
+                                             {"dumpout", required_argument, nullptr, 'o'},
+                                             {"verify", required_argument, nullptr, 'v'},
+                                             {"help", no_argument, nullptr, '?'},
+                                             {nullptr, 0, nullptr, 0}};
+
+class BatchnormBwdArgParser
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths;
+    std::vector<int> reduceDims;
+
+    bool do_verification = false;
+    bool do_dumpout      = false;
+
+    bool haveSavedMeanInvVar;
+
+    int data_type    = 0;
+    int init_method  = 2;
+    bool time_kernel = false;
+
+    BatchnormBwdArgParser()  = default;
+    ~BatchnormBwdArgParser() = default;
+
+    void show_usage(const char* cmd)
+    {
+        // clang-format off
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension lengths, must have 4 integers for nhwc" << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of dimensions to reduce on" << std::endl;  
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the result by comparing with the host-based batch-normalization" << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2 -- 1/0 to indicate whether to use saved mean and invVariance" << std::endl;
+        std::cout << "Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)" << std::endl;
+        std::cout << "Arg4 -- time kernel (0=no, 1=yes)" << std::endl;
+        // clang-format on
+    };
+
+    int operator()(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        optind++; // to skip the module name
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:v:o:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case 'o':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_dumpout = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return -1;
+                };
+                break;
+
+            default:
+                show_usage(argv[0]);
+                std::cerr << "Invalid cmd-line options!" << std::endl;
+                return -1;
+            };
+        };
+
+        if(optind + 4 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type           = std::atoi(argv[optind++]);
+        haveSavedMeanInvVar = std::atoi(argv[optind++]);
+        init_method         = std::atoi(argv[optind++]);
+        time_kernel         = static_cast<bool>(std::atoi(argv[optind++]));
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return -1;
+
+        return 0;
+    };
+}; // end of class AppArgs
+
+static const double epsilon = std::numeric_limits<float>::epsilon();
+
+int profile_batchnorm_backward(int argc, char* argv[])
+{
+    using ck::profiler::profile_batchnorm_backward_impl;
+
+    BatchnormBwdArgParser arg_parser;
+
+    if(arg_parser(argc, argv) != 0)
+        return -1;
+
+    using F16  = ck::half_t;
+    using F32  = float;
+    using BF16 = ck::bhalf_t;
+    using F64  = double;
+
+    if(arg_parser.data_type == 0)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<F16, F32, F32, F32, F16, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+    else if(arg_parser.data_type == 1)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<F32, F32, F32, F32, F32, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+    else if(arg_parser.data_type == 5)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<BF16, F32, F32, F32, BF16, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+    else if(arg_parser.data_type == 6)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<F64, F64, F64, F64, F64, F64, F64, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("bnorm_bwd", "Batchnorm backward", profile_batchnorm_backward);
diff --git a/profiler/src/profile_batchnorm_fwd.cpp b/profiler/src/profile_batchnorm_fwd.cpp
new file mode 100644
index 0000000000..507fb4b450
--- /dev/null
+++ b/profiler/src/profile_batchnorm_fwd.cpp
@@ -0,0 +1,219 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <getopt.h>
+
+#include "ck/library/utility/host_common_util.hpp"
+#include "profiler/profile_batchnorm_forward_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+using namespace std;
+
+static const struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                             {"reduceDims", required_argument, nullptr, 'R'},
+                                             {"dumpout", required_argument, nullptr, 'o'},
+                                             {"verify", required_argument, nullptr, 'v'},
+                                             {"help", no_argument, nullptr, '?'},
+                                             {nullptr, 0, nullptr, 0}};
+
+class BatchnormFwdArgParser
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths;
+    std::vector<int> reduceDims;
+
+    bool do_verification = false;
+    bool do_dumpout      = false;
+
+    bool updateMovingAverage;
+    bool saveMeanAndInvVariance;
+
+    int data_type    = 0;
+    int init_method  = 2;
+    bool time_kernel = false;
+
+    BatchnormFwdArgParser()  = default;
+    ~BatchnormFwdArgParser() = default;
+
+    void show_usage(const char* cmd)
+    {
+        // clang-format off
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension lengths, must have 4 integers for nhwc" << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of dimensions to reduce on" << std::endl;  
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the result by comparing with the host-based batch-normalization" << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2: 1/0 to indicate whether to update the moving average and variance (0=no, 1=yes)" << std::endl;
+        std::cout << "Arg3: 1/0 to indicate whether to save the calculated mean and invVariance (0=no, 1=yes)" << std::endl;
+        std::cout << "Arg4: init method used for bnScale and bnBias (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)" << std::endl;
+        std::cout << "Arg5: time kernel (0=no, 1=yes)" << std::endl;
+        // clang-format on
+    };
+
+    int operator()(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        optind++; // to skip the module name
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:v:o:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case 'o':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_dumpout = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return -1;
+                };
+                break;
+
+            default:
+                show_usage(argv[0]);
+                std::cerr << "Invalid cmd-line options!" << std::endl;
+                return -1;
+            };
+        };
+
+        if(optind + 5 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type              = std::atoi(argv[optind++]);
+        updateMovingAverage    = std::atoi(argv[optind++]);
+        saveMeanAndInvVariance = std::atoi(argv[optind++]);
+        init_method            = std::atoi(argv[optind++]);
+        time_kernel            = static_cast<bool>(std::atoi(argv[optind++]));
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return -1;
+
+        return 0;
+    };
+}; // end of class AppArgs
+
+static const double epsilon       = std::numeric_limits<float>::epsilon();
+static const double averageFactor = 0.1;
+
+int profile_batchnorm_forward(int argc, char* argv[])
+{
+    using ck::profiler::profile_batchnorm_forward_impl;
+
+    BatchnormFwdArgParser arg_parser;
+
+    if(arg_parser(argc, argv) != 0)
+        return -1;
+
+    using F16  = ck::half_t;
+    using F32  = float;
+    using BF16 = ck::bhalf_t;
+    using F64  = double;
+
+    if(arg_parser.data_type == 0)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_forward_impl<F16, F16, F32, F16, F16, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.updateMovingAverage,
+                arg_parser.saveMeanAndInvVariance,
+                epsilon,
+                averageFactor);
+        };
+    }
+    else if(arg_parser.data_type == 1)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_forward_impl<F32, F32, F32, F32, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.updateMovingAverage,
+                arg_parser.saveMeanAndInvVariance,
+                epsilon,
+                averageFactor);
+        };
+    }
+    else if(arg_parser.data_type == 5)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_forward_impl<BF16, BF16, F32, BF16, BF16, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.updateMovingAverage,
+                arg_parser.saveMeanAndInvVariance,
+                epsilon,
+                averageFactor);
+        };
+    }
+    else if(arg_parser.data_type == 6)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_forward_impl<F64, F64, F64, F64, F64, F64, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.updateMovingAverage,
+                arg_parser.saveMeanAndInvVariance,
+                epsilon,
+                averageFactor);
+        };
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("bnorm_fwd", "Batchnorm forward", profile_batchnorm_forward);
diff --git a/profiler/src/profile_batchnorm_infer.cpp b/profiler/src/profile_batchnorm_infer.cpp
new file mode 100644
index 0000000000..f1c19bc36e
--- /dev/null
+++ b/profiler/src/profile_batchnorm_infer.cpp
@@ -0,0 +1,202 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <getopt.h>
+
+#include "ck/library/utility/host_common_util.hpp"
+#include "profiler/profile_batchnorm_infer_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+using namespace std;
+
+static const struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                             {"reduceDims", required_argument, nullptr, 'R'},
+                                             {"dumpout", required_argument, nullptr, 'o'},
+                                             {"verify", required_argument, nullptr, 'v'},
+                                             {"help", no_argument, nullptr, '?'},
+                                             {nullptr, 0, nullptr, 0}};
+
+class BatchnormInferArgParser
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths;
+    std::vector<int> reduceDims;
+
+    bool do_verification = false;
+    bool do_dumpout      = false;
+
+    bool updateMovingAverage;
+    bool saveMeanAndInvVariance;
+
+    int data_type    = 0;
+    int init_method  = 2;
+    bool time_kernel = false;
+
+    BatchnormInferArgParser()  = default;
+    ~BatchnormInferArgParser() = default;
+
+    void show_usage(const char* cmd)
+    {
+        // clang-format off
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension lengths, must have 4 integers for nhwc" << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of dimensions to reduce on" << std::endl;  
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the result by comparing with the host-based batch-normalization" << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2: init method used for bnScale and bnBias (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)" << std::endl;
+        std::cout << "Arg3: time kernel (0=no, 1=yes)" << std::endl;
+        // clang-format on
+    };
+
+    int operator()(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        optind++; // to skip the module name
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:v:o:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case 'o':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_dumpout = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return -1;
+                };
+                break;
+
+            default:
+                show_usage(argv[0]);
+                std::cerr << "Invalid cmd-line options!" << std::endl;
+                return -1;
+            };
+        };
+
+        if(optind + 3 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type   = std::atoi(argv[optind++]);
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind++]));
+
+        if(data_type != 0 && data_type != 1 && data_type != 5 && data_type != 6)
+            return -1;
+
+        return 0;
+    };
+}; // end of class AppArgs
+
+static const double epsilon = std::numeric_limits<float>::epsilon();
+
+int profile_batchnorm_infer(int argc, char* argv[])
+{
+    using ck::profiler::profile_batchnorm_infer_impl;
+
+    BatchnormInferArgParser arg_parser;
+
+    if(arg_parser(argc, argv) != 0)
+        return -1;
+
+    using F16  = ck::half_t;
+    using F32  = float;
+    using BF16 = ck::bhalf_t;
+    using F64  = double;
+
+    if(arg_parser.data_type == 0)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_infer_impl<F16, F16, F32, F16, F16, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                epsilon);
+        };
+    }
+    else if(arg_parser.data_type == 1)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_infer_impl<F32, F32, F32, F32, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                epsilon);
+        };
+    }
+    else if(arg_parser.data_type == 5)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_infer_impl<BF16, BF16, F32, BF16, BF16, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                epsilon);
+        };
+    }
+    else if(arg_parser.data_type == 6)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_infer_impl<F64, F64, F64, F64, F64, F64, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                epsilon);
+        };
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("bnorm_infer", "Batchnorm inference", profile_batchnorm_infer);
diff --git a/profiler/src/profile_contraction_bilinear.cpp b/profiler/src/profile_contraction_bilinear.cpp
new file mode 100644
index 0000000000..6ed1841204
--- /dev/null
+++ b/profiler/src/profile_contraction_bilinear.cpp
@@ -0,0 +1,165 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <vector>
+
+#include "profiler/profile_contraction_impl.hpp"
+#include "profiler/profile_contraction_utils.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "contraction_bilinear"
+#define OP_DESC "CONTRACTION+Bilinear"
+
+static void print_helper_msg()
+{
+    std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+              << "arg2: data type (0: fp32; 1: f64)\n"
+              << "arg3: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + "
+                 "D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
+              << "                     1: A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + "
+                 "D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
+              << "                     2: A[k0, k1, m0, m1] * B[k0, k1, n0, n1] + "
+                 "D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
+              << "                     3: A[k0, k1, m0, m1] * B[n0, n1, k0, k1] + "
+                 "D[m0, m1, n0, n1] = E[m0, m1, n0, n1])\n"
+              << "arg4: verification (0: no; 1: yes)\n"
+              << "arg5: initialization (0: no init; 1: integer value; 2: decimal "
+              << "value)\n"
+              << "arg6: print tensor value (0: no; 1: yes)\n"
+              << "arg7: time kernel (0: no, 1: yes)\n"
+              << "arg8 and arg9: alpha and beta\n"
+              << "arg10 to 15: M0, M1, N0, N1, K0, K1\n"
+              << "arg16 to 31: Strides for A, B, D and E (skip for default)\n"
+              << std::endl;
+}
+
+int profile_contraction_bilinear(int argc, char* argv[])
+{
+    const bool default_strides = argc == 16;
+
+    if(argc != 32 && argc != 16)
+    {
+        print_helper_msg();
+        exit(1);
+    }
+
+    const auto data_type          = static_cast<ContractionDataType>(std::stoi(argv[2]));
+    const auto layout             = static_cast<ContractionMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification    = std::stoi(argv[4]);
+    const ck::index_t init_method = std::stoi(argv[5]);
+    const bool do_log             = std::stoi(argv[6]);
+    const bool time_kernel        = std::stoi(argv[7]);
+    const float alpha             = std::stof(argv[8]);
+    const float beta              = std::stof(argv[9]);
+
+    std::vector<ck::index_t> M;
+    std::vector<ck::index_t> N;
+    std::vector<ck::index_t> K;
+    const ck::index_t dims_arg_num = 10;
+    collect_index_params(argv, M, dims_arg_num, 2);
+    collect_index_params(argv, N, dims_arg_num + 2, 2);
+    collect_index_params(argv, K, dims_arg_num + 4, 2);
+
+    std::vector<ck::index_t> StridesA;
+    std::vector<ck::index_t> StridesB;
+    std::vector<ck::index_t> StridesE;
+    std::vector<ck::index_t> StridesD;
+    if(!default_strides)
+    {
+        collect_index_params(argv, StridesA, dims_arg_num + 6, 4);
+        collect_index_params(argv, StridesB, dims_arg_num + 10, 4);
+        collect_index_params(argv, StridesE, dims_arg_num + 14, 4);
+        collect_index_params(argv, StridesD, dims_arg_num + 18, 4);
+    }
+
+    using F32 = float;
+    using F64 = double;
+
+    auto profile = [&](auto a_layout, auto b_layout, auto cde_layout, auto type) {
+        using ALayout   = decltype(a_layout);
+        using BLayout   = decltype(b_layout);
+        using CDELayout = decltype(cde_layout);
+
+        using DataType = decltype(type);
+
+        if(default_strides)
+        {
+            assign_default_strides(a_layout, StridesA, {M[0], M[1], K[0], K[1]});
+            assign_default_strides(b_layout, StridesB, {K[0], K[1], N[0], N[1]});
+            assign_default_strides(cde_layout, StridesE, {M[0], M[1], N[0], N[1]});
+            assign_default_strides(cde_layout, StridesD, {M[0], M[1], N[0], N[1]});
+        }
+        bool pass = ck::profiler::profile_contraction_impl<ALayout,
+                                                           BLayout,
+                                                           CDELayout,
+                                                           DataType,
+                                                           ck::Tuple<DataType>,
+                                                           Bilinear>(do_verification,
+                                                                     init_method,
+                                                                     do_log,
+                                                                     time_kernel,
+                                                                     Bilinear{alpha, beta},
+                                                                     M,
+                                                                     N,
+                                                                     K,
+                                                                     StridesA,
+                                                                     StridesB,
+                                                                     StridesE,
+                                                                     StridesD);
+
+        return pass;
+    };
+
+    if(data_type == ContractionDataType::F32_F32_F32_F32 &&
+       layout == ContractionMatrixLayout::MK_KN_MN_MN)
+    {
+        return profile(Row{}, Row{}, Row{}, F32{});
+    }
+    else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
+            layout == ContractionMatrixLayout::MK_NK_MN_MN)
+    {
+        return profile(Row{}, Col{}, Row{}, F32{});
+    }
+    else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
+            layout == ContractionMatrixLayout::KM_KN_MN_MN)
+    {
+        return profile(Col{}, Row{}, Row{}, F32{});
+    }
+    else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
+            layout == ContractionMatrixLayout::KM_NK_MN_MN)
+    {
+        return profile(Col{}, Col{}, Row{}, F32{});
+    }
+    else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
+            layout == ContractionMatrixLayout::MK_KN_MN_MN)
+    {
+        return profile(Row{}, Row{}, Row{}, F64{});
+    }
+    else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
+            layout == ContractionMatrixLayout::MK_NK_MN_MN)
+    {
+        return profile(Row{}, Col{}, Row{}, F64{});
+    }
+    else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
+            layout == ContractionMatrixLayout::KM_KN_MN_MN)
+    {
+        return profile(Col{}, Row{}, Row{}, F64{});
+    }
+    else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
+            layout == ContractionMatrixLayout::KM_NK_MN_MN)
+    {
+        return profile(Col{}, Col{}, Row{}, F64{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_contraction_bilinear);
diff --git a/profiler/src/profile_contraction_scale.cpp b/profiler/src/profile_contraction_scale.cpp
new file mode 100644
index 0000000000..6784b916f6
--- /dev/null
+++ b/profiler/src/profile_contraction_scale.cpp
@@ -0,0 +1,162 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <vector>
+
+#include "profiler/profile_contraction_impl.hpp"
+#include "profiler/profile_contraction_utils.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "contraction_scale"
+#define OP_DESC "CONTRACTION+Scale"
+
+static void print_helper_msg()
+{
+    std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+              << "arg2: data type (0: fp32; 1: f64)\n"
+              << "arg3: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + "
+                 "D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
+              << "                     1: A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + "
+                 "D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
+              << "                     2: A[k0, k1, m0, m1] * B[k0, k1, n0, n1] + "
+                 "D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
+              << "                     3: A[k0, k1, m0, m1] * B[n0, n1, k0, k1] + "
+                 "D[m0, m1, n0, n1] = E[m0, m1, n0, n1])\n"
+              << "arg4: verification (0: no; 1: yes)\n"
+              << "arg5: initialization (0: no init; 1: integer value; 2: decimal "
+              << "value)\n"
+              << "arg6: print tensor value (0: no; 1: yes)\n"
+              << "arg7: time kernel (0: no, 1: yes)\n"
+              << "arg8: alpha\n"
+              << "arg9 to 14: M0, M1, N0, N1, K0, K1\n"
+              << "arg15 to 30: Strides for A, B, D and E (skip for default)\n"
+              << std::endl;
+}
+
+int profile_contraction_scale(int argc, char* argv[])
+{
+    const bool default_strides = argc == 15;
+
+    if(argc != 31 && argc != 15)
+    {
+        print_helper_msg();
+        exit(1);
+    }
+
+    const auto data_type          = static_cast<ContractionDataType>(std::stoi(argv[2]));
+    const auto layout             = static_cast<ContractionMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification    = std::stoi(argv[4]);
+    const ck::index_t init_method = std::stoi(argv[5]);
+    const bool do_log             = std::stoi(argv[6]);
+    const bool time_kernel        = std::stoi(argv[7]);
+    const float alpha             = std::stof(argv[8]);
+
+    std::vector<ck::index_t> M;
+    std::vector<ck::index_t> N;
+    std::vector<ck::index_t> K;
+    const ck::index_t dims_arg_num = 9;
+    collect_index_params(argv, M, dims_arg_num, 2);
+    collect_index_params(argv, N, dims_arg_num + 2, 2);
+    collect_index_params(argv, K, dims_arg_num + 4, 2);
+
+    std::vector<ck::index_t> StridesA;
+    std::vector<ck::index_t> StridesB;
+    std::vector<ck::index_t> StridesE;
+    std::vector<ck::index_t> StridesD;
+    if(!default_strides)
+    {
+        collect_index_params(argv, StridesA, dims_arg_num + 6, 4);
+        collect_index_params(argv, StridesB, dims_arg_num + 10, 4);
+        collect_index_params(argv, StridesE, dims_arg_num + 14, 4);
+        collect_index_params(argv, StridesD, dims_arg_num + 18, 4);
+    }
+
+    using F32 = float;
+    using F64 = double;
+
+    auto profile = [&](auto a_layout, auto b_layout, auto cde_layout, auto type) {
+        using ALayout   = decltype(a_layout);
+        using BLayout   = decltype(b_layout);
+        using CDELayout = decltype(cde_layout);
+
+        using DataType = decltype(type);
+
+        if(default_strides)
+        {
+            assign_default_strides(a_layout, StridesA, {M[0], M[1], K[0], K[1]});
+            assign_default_strides(b_layout, StridesB, {K[0], K[1], N[0], N[1]});
+            assign_default_strides(cde_layout, StridesE, {M[0], M[1], N[0], N[1]});
+            assign_default_strides(cde_layout, StridesD, {M[0], M[1], N[0], N[1]});
+        }
+
+        bool pass = ck::profiler::
+            profile_contraction_impl<ALayout, BLayout, CDELayout, DataType, ck::Tuple<>, Scale>(
+                do_verification,
+                init_method,
+                do_log,
+                time_kernel,
+                Scale{alpha},
+                M,
+                N,
+                K,
+                StridesA,
+                StridesB,
+                StridesE,
+                StridesD);
+
+        return pass;
+    };
+
+    if(data_type == ContractionDataType::F32_F32_F32_F32 &&
+       layout == ContractionMatrixLayout::MK_KN_MN_MN)
+    {
+        return profile(Row{}, Row{}, Row{}, F32{});
+    }
+    else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
+            layout == ContractionMatrixLayout::MK_NK_MN_MN)
+    {
+        return profile(Row{}, Col{}, Row{}, F32{});
+    }
+    else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
+            layout == ContractionMatrixLayout::KM_KN_MN_MN)
+    {
+        return profile(Col{}, Row{}, Row{}, F32{});
+    }
+    else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
+            layout == ContractionMatrixLayout::KM_NK_MN_MN)
+    {
+        return profile(Col{}, Col{}, Row{}, F32{});
+    }
+    else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
+            layout == ContractionMatrixLayout::MK_KN_MN_MN)
+    {
+        return profile(Row{}, Row{}, Row{}, F64{});
+    }
+    else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
+            layout == ContractionMatrixLayout::MK_NK_MN_MN)
+    {
+        return profile(Row{}, Col{}, Row{}, F64{});
+    }
+    else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
+            layout == ContractionMatrixLayout::KM_KN_MN_MN)
+    {
+        return profile(Col{}, Row{}, Row{}, F64{});
+    }
+    else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
+            layout == ContractionMatrixLayout::KM_NK_MN_MN)
+    {
+        return profile(Col{}, Col{}, Row{}, F64{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_contraction_scale);
diff --git a/profiler/src/profile_conv_bwd_data.cpp b/profiler/src/profile_conv_bwd_data.cpp
new file mode 100644
index 0000000000..e08a39aeb0
--- /dev/null
+++ b/profiler/src/profile_conv_bwd_data.cpp
@@ -0,0 +1,198 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_conv_bwd_data_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+namespace {
+
+enum struct ConvLayout
+{
+    NCHW_KCYX_NKHW, // 0
+    NHWC_KYXC_NHWK, // 1
+};
+
+enum struct ConvDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+#define OP_NAME "conv_bwd_data"
+#define OP_DESC "Convolution Backward Data"
+
+static void print_helper_msg()
+{
+    std::cout
+        << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+        << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
+        << "                 1: Input fp16, Weight fp16, Output fp16\n"
+        << "                 2: Input bf16, Weight bf16, Output bf16\n"
+        << "                 3: Input int8, Weight int8, Output int8)\n"
+        << "arg3: tensor layout (0: Input[N, C, Hi, Wi], Weight[K, C, Y, X], Output[N, K, Ho, Wo]\n"
+        << "                     1: Input[N, Hi, Wi, C], Weight[K, Y, X, C], Output[N, Ho, Wo, "
+           "K])\n"
+        << "arg4: verification (0: no, 1: yes)\n"
+        << "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
+        << "arg6: print tensor value (0: no; 1: yes)\n"
+        << "arg7: time kernel (0: no, 1: yes)\n"
+        << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+} // namespace
+
+int profile_conv_bwd_data(int argc, char* argv[])
+{
+    // 8 for control, 1 for num_dim_spatial
+    if(argc < 9)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<ConvLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+    const int num_dim_spatial  = std::stoi(argv[8]);
+
+    // 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
+    if(argc != 8 + 1 + 4 + 6 * num_dim_spatial)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+#ifdef CK_ENABLE_INT8
+    using INT8 = int8_t;
+#endif
+
+    using NWC   = ck::tensor_layout::convolution::NWC;
+    using NHWC  = ck::tensor_layout::convolution::NHWC;
+    using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+    using KXC   = ck::tensor_layout::convolution::KXC;
+    using KYXC  = ck::tensor_layout::convolution::KYXC;
+    using KZYXC = ck::tensor_layout::convolution::KZYXC;
+
+    using NWK   = ck::tensor_layout::convolution::NWK;
+    using NHWK  = ck::tensor_layout::convolution::NHWK;
+    using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+    constexpr auto I1 = ck::Number<1>{};
+    constexpr auto I2 = ck::Number<2>{};
+    constexpr auto I3 = ck::Number<3>{};
+
+    auto profile = [&](auto num_dim_spatial_tmp,
+                       auto in_layout,
+                       auto wei_layout,
+                       auto out_layout,
+                       auto in_type,
+                       auto wei_type,
+                       auto out_type) {
+        constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
+
+        using InLayout  = decltype(in_layout);
+        using WeiLayout = decltype(wei_layout);
+        using OutLayout = decltype(out_layout);
+
+        using InDataType  = decltype(in_type);
+        using WeiDataType = decltype(wei_type);
+        using OutDataType = decltype(out_type);
+
+        bool pass = ck::profiler::profile_conv_bwd_data_impl<NDimSpatial,
+                                                             InLayout,
+                                                             WeiLayout,
+                                                             OutLayout,
+                                                             InDataType,
+                                                             WeiDataType,
+                                                             OutDataType>(
+            do_verification, init_method, do_log, time_kernel, params);
+
+        return pass ? 0 : 1;
+    };
+
+    if(num_dim_spatial == 1 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, BF16{}, BF16{}, BF16{});
+        }
+#ifdef CK_ENABLE_INT8
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, INT8{}, INT8{}, INT8{});
+        }
+#endif
+    }
+    else if(num_dim_spatial == 2 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, BF16{}, BF16{}, BF16{});
+        }
+#ifdef CK_ENABLE_INT8
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, INT8{}, INT8{}, INT8{});
+        }
+#endif
+    }
+    else if(num_dim_spatial == 3 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, BF16{}, BF16{}, BF16{});
+        }
+#ifdef CK_ENABLE_INT8
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, INT8{}, INT8{}, INT8{});
+        }
+#endif
+    }
+
+    std::cout << "this data_type & layout is not implemented" << std::endl;
+
+    return 1;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_bwd_data);
diff --git a/profiler/src/profile_conv_fwd.cpp b/profiler/src/profile_conv_fwd.cpp
new file mode 100644
index 0000000000..701999d8a9
--- /dev/null
+++ b/profiler/src/profile_conv_fwd.cpp
@@ -0,0 +1,192 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_conv_fwd_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+namespace {
+
+enum struct ConvLayout
+{
+    NCHW_KCYX_NKHW, // 0
+    NHWC_KYXC_NHWK, // 1
+};
+
+enum struct ConvDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+#define OP_NAME "conv_fwd"
+#define OP_DESC "Convolution Forward"
+
+static void print_helper_msg()
+{
+    std::cout
+        // clang-format-off
+        << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+        << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
+        << "                 1: Input fp16, Weight fp16, Output fp16\n"
+        << "                 2: Input bf16, Weight bf16, Output bf16\n"
+        << "                 3: Input int8, Weight int8, Output int8)\n"
+        << "arg3: tensor layout (0: Input[N, C, Hi, Wi], Weight[K, C, Y, X], Output[N, K, Ho, Wo]\n"
+        << "                     1: Input[N, Hi, Wi, C], Weight[K, Y, X, C], Output[N, Ho, Wo, "
+           "K])\n"
+        << "arg4: verification (0: no, 1: yes)\n"
+        << "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
+        << "arg6: print tensor value (0: no; 1: yes)\n"
+        << "arg7: time kernel (0: no, 1: yes)\n"
+        << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+    // clang-format-on
+}
+
+} // namespace
+
+int profile_conv_fwd(int argc, char* argv[])
+{
+    // 8 for control, 1 for num_dim_spatial
+    if(argc < 9)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<ConvLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+    const int num_dim_spatial  = std::stoi(argv[8]);
+
+    // 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
+    if(argc != 8 + 1 + 4 + 6 * num_dim_spatial)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+    using INT8 = int8_t;
+
+    using NWC   = ck::tensor_layout::convolution::NWC;
+    using NHWC  = ck::tensor_layout::convolution::NHWC;
+    using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+    using KXC   = ck::tensor_layout::convolution::KXC;
+    using KYXC  = ck::tensor_layout::convolution::KYXC;
+    using KZYXC = ck::tensor_layout::convolution::KZYXC;
+
+    using NWK   = ck::tensor_layout::convolution::NWK;
+    using NHWK  = ck::tensor_layout::convolution::NHWK;
+    using NDHWK = ck::tensor_layout::convolution::NDHWK;
+
+    constexpr auto I1 = ck::Number<1>{};
+    constexpr auto I2 = ck::Number<2>{};
+    constexpr auto I3 = ck::Number<3>{};
+
+    auto profile = [&](auto num_dim_spatial_tmp,
+                       auto in_layout,
+                       auto wei_layout,
+                       auto out_layout,
+                       auto in_type,
+                       auto wei_type,
+                       auto out_type) {
+        constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
+
+        using InLayout  = decltype(in_layout);
+        using WeiLayout = decltype(wei_layout);
+        using OutLayout = decltype(out_layout);
+
+        using InDataType  = decltype(in_type);
+        using WeiDataType = decltype(wei_type);
+        using OutDataType = decltype(out_type);
+
+        bool pass = ck::profiler::profile_conv_fwd_impl<NDimSpatial,
+                                                        InLayout,
+                                                        WeiLayout,
+                                                        OutLayout,
+                                                        InDataType,
+                                                        WeiDataType,
+                                                        OutDataType>(
+            do_verification, init_method, do_log, time_kernel, params);
+
+        return pass ? 0 : 1;
+    };
+
+    if(num_dim_spatial == 1 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I1, NWC{}, KXC{}, NWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 2 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I2, NHWC{}, KYXC{}, NHWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 3 && layout == ConvLayout::NHWC_KYXC_NHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I3, NDHWC{}, KZYXC{}, NDHWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+
+    std::cout << "this data_type & layout is not implemented" << std::endl;
+
+    return 1;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_fwd);
diff --git a/profiler/src/profile_conv_fwd_bias_relu.cpp b/profiler/src/profile_conv_fwd_bias_relu.cpp
new file mode 100644
index 0000000000..31055ec1d1
--- /dev/null
+++ b/profiler/src/profile_conv_fwd_bias_relu.cpp
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_conv_fwd_bias_relu_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct ConvDataType
+{
+    F32_F32_F32, // 0
+    F16_F16_F16, // 1
+};
+
+enum struct ConvInputLayout
+{
+    NCHW, // 0
+    NHWC, // 1
+};
+
+enum struct ConvWeightLayout
+{
+    KCYX, // 0
+    KYXC, // 1
+};
+
+enum struct ConvOutputLayout
+{
+    NKHW, // 0
+    NHWK, // 1
+};
+
+#define OP_NAME "conv_fwd_bias_relu"
+#define OP_DESC "Convolution Forward+Bias+ReLU"
+
+int profile_conv_fwd_bias_relu(int argc, char* argv[])
+{
+    if(argc != 25)
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16)\n");
+        printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
+        printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
+        printf("arg5: output tensor layout (0: NKHW; 1: NHWK)\n");
+        printf("arg6: verification (0: no; 1: yes)\n");
+        printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg8: print tensor value (0: no; 1: yes)\n");
+        printf("arg9: time kernel (0=n0, 1=yes)\n");
+        printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
+               "RightPx\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto in_layout       = static_cast<ConvInputLayout>(std::stoi(argv[3]));
+    const auto wei_layout      = static_cast<ConvWeightLayout>(std::stoi(argv[4]));
+    const auto out_layout      = static_cast<ConvOutputLayout>(std::stoi(argv[5]));
+    const bool do_verification = std::stoi(argv[6]);
+    const int init_method      = std::stoi(argv[7]);
+    const bool do_log          = std::stoi(argv[8]);
+    const bool time_kernel     = std::stoi(argv[9]);
+
+    const ck::index_t N  = std::stoi(argv[10]);
+    const ck::index_t K  = std::stoi(argv[11]);
+    const ck::index_t C  = std::stoi(argv[12]);
+    const ck::index_t Y  = std::stoi(argv[13]);
+    const ck::index_t X  = std::stoi(argv[14]);
+    const ck::index_t Hi = std::stoi(argv[15]);
+    const ck::index_t Wi = std::stoi(argv[16]);
+
+    const ck::index_t conv_stride_h   = std::stoi(argv[17]);
+    const ck::index_t conv_stride_w   = std::stoi(argv[18]);
+    const ck::index_t conv_dilation_h = std::stoi(argv[19]);
+    const ck::index_t conv_dilation_w = std::stoi(argv[20]);
+    const ck::index_t in_left_pad_h   = std::stoi(argv[21]);
+    const ck::index_t in_left_pad_w   = std::stoi(argv[22]);
+    const ck::index_t in_right_pad_h  = std::stoi(argv[23]);
+    const ck::index_t in_right_pad_w  = std::stoi(argv[24]);
+
+    const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
+    const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
+
+    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
+    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
+
+    if(data_type == ConvDataType::F16_F16_F16 && in_layout == ConvInputLayout::NHWC &&
+       wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
+    {
+        ck::profiler::profile_conv_fwd_bias_relu_impl<2,
+                                                      ck::half_t,
+                                                      ck::half_t,
+                                                      ck::half_t,
+                                                      ck::tensor_layout::convolution::NHWC,
+                                                      ck::tensor_layout::convolution::KYXC,
+                                                      ck::tensor_layout::convolution::NHWK>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            N,
+            K,
+            C,
+            std::vector<ck::index_t>{Hi, Wi},
+            std::vector<ck::index_t>{Y, X},
+            std::vector<ck::index_t>{Ho, Wo},
+            std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
+            std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
+            std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
+            std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
+    }
+    else
+    {
+        throw std::runtime_error("wrong! data_type & layout for this operator is not implemented");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_fwd_bias_relu);
diff --git a/profiler/src/profile_conv_fwd_bias_relu_add.cpp b/profiler/src/profile_conv_fwd_bias_relu_add.cpp
new file mode 100644
index 0000000000..8c2439a0c7
--- /dev/null
+++ b/profiler/src/profile_conv_fwd_bias_relu_add.cpp
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_conv_fwd_bias_relu_add_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct ConvDataType
+{
+    F32_F32_F32, // 0
+    F16_F16_F16, // 1
+};
+
+enum struct ConvInputLayout
+{
+    NCHW, // 0
+    NHWC, // 1
+};
+
+enum struct ConvWeightLayout
+{
+    KCYX, // 0
+    KYXC, // 1
+};
+
+enum struct ConvOutputLayout
+{
+    NKHW, // 0
+    NHWK, // 1
+};
+
+#define OP_NAME "conv_fwd_bias_relu_add"
+#define OP_DESC "Convolution Forward+Bias+ReLU+Add"
+
+int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
+{
+    if(argc != 25)
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16)\n");
+        printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
+        printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
+        printf("arg5: output tensor layout (0: NKHW; 1: NHWK)\n");
+        printf("arg6: verification (0: no; 1: yes)\n");
+        printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg8: print tensor value (0: no; 1: yes)\n");
+        printf("arg9: time kernel (0=n0, 1=yes)\n");
+        printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
+               "RightPx\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto in_layout       = static_cast<ConvInputLayout>(std::stoi(argv[3]));
+    const auto wei_layout      = static_cast<ConvWeightLayout>(std::stoi(argv[4]));
+    const auto out_layout      = static_cast<ConvOutputLayout>(std::stoi(argv[5]));
+    const bool do_verification = std::stoi(argv[6]);
+    const int init_method      = std::stoi(argv[7]);
+    const bool do_log          = std::stoi(argv[8]);
+    const bool time_kernel     = std::stoi(argv[9]);
+
+    const ck::index_t N  = std::stoi(argv[10]);
+    const ck::index_t K  = std::stoi(argv[11]);
+    const ck::index_t C  = std::stoi(argv[12]);
+    const ck::index_t Y  = std::stoi(argv[13]);
+    const ck::index_t X  = std::stoi(argv[14]);
+    const ck::index_t Hi = std::stoi(argv[15]);
+    const ck::index_t Wi = std::stoi(argv[16]);
+
+    const ck::index_t conv_stride_h   = std::stoi(argv[17]);
+    const ck::index_t conv_stride_w   = std::stoi(argv[18]);
+    const ck::index_t conv_dilation_h = std::stoi(argv[19]);
+    const ck::index_t conv_dilation_w = std::stoi(argv[20]);
+    const ck::index_t in_left_pad_h   = std::stoi(argv[21]);
+    const ck::index_t in_left_pad_w   = std::stoi(argv[22]);
+    const ck::index_t in_right_pad_h  = std::stoi(argv[23]);
+    const ck::index_t in_right_pad_w  = std::stoi(argv[24]);
+
+    const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
+    const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
+
+    const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
+    const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
+
+    if(data_type == ConvDataType::F16_F16_F16 && in_layout == ConvInputLayout::NHWC &&
+       wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
+    {
+        ck::profiler::profile_conv_fwd_bias_relu_add_impl<2,
+                                                          ck::half_t,
+                                                          ck::half_t,
+                                                          ck::half_t,
+                                                          ck::tensor_layout::convolution::NHWC,
+                                                          ck::tensor_layout::convolution::KYXC,
+                                                          ck::tensor_layout::convolution::NHWK>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            N,
+            K,
+            C,
+            std::vector<ck::index_t>{Hi, Wi},
+            std::vector<ck::index_t>{Y, X},
+            std::vector<ck::index_t>{Ho, Wo},
+            std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
+            std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
+            std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
+            std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
+    }
+    else
+    {
+        throw std::runtime_error("wrong! data_type & layout for this operator is not implemented");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_fwd_bias_relu_add);
diff --git a/profiler/src/profile_gemm.cpp b/profiler/src/profile_gemm.cpp
new file mode 100644
index 0000000000..9ca7fc4c88
--- /dev/null
+++ b/profiler/src/profile_gemm.cpp
@@ -0,0 +1,206 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+};
+
+enum struct GemmDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+#define OP_NAME "gemm"
+#define OP_DESC "GEMM"
+
+static void print_helper_msg()
+{
+    std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+              << "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
+              << "arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n"
+              << "                     1: A[m, k] * B[n, k] = C[m, n];\n"
+              << "                     2: A[k, m] * B[k, n] = C[m, n];\n"
+              << "                     3: A[k, m] * B[n, k] = C[m, n])\n"
+              << "arg4: verification (0: no; 1: yes)\n"
+              << "arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg6: print tensor value (0: no; 1: yes)\n"
+              << "arg7: time kernel (0: no, 1: yes)\n"
+              << "arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n"
+              << std::endl;
+}
+
+int profile_gemm(int argc, char* argv[])
+{
+    if(argc != 14)
+    {
+        print_helper_msg();
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideC = std::stoi(argv[13]);
+
+    using F32 = float;
+    using F16 = ck::half_t;
+#ifdef CK_ENABLE_BF16
+    using BF16 = ck::bhalf_t;
+#endif
+#ifdef CK_ENABLE_INT8
+    using INT8  = int8_t;
+    using INT32 = int32_t;
+#endif
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_layout,
+                       auto b_layout,
+                       auto c_layout,
+                       auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto c_type) {
+        using ALayout = decltype(a_layout);
+        using BLayout = decltype(b_layout);
+        using CLayout = decltype(c_layout);
+
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using CDataType   = decltype(c_type);
+
+        const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
+
+        bool pass =
+            ck::profiler::profile_gemm_impl<ALayout,
+                                            BLayout,
+                                            CLayout,
+                                            ADataType,
+                                            BDataType,
+                                            AccDataType,
+                                            CDataType>(do_verification,
+                                                       init_method,
+                                                       do_log,
+                                                       time_kernel,
+                                                       M,
+                                                       N,
+                                                       K,
+                                                       (StrideA < 0) ? DefaultStrideA : StrideA,
+                                                       (StrideB < 0) ? DefaultStrideB : StrideB,
+                                                       (StrideC < 0) ? DefaultStrideC : StrideC);
+
+        return pass ? 0 : 1;
+    };
+
+    if(false)
+        ;
+#ifdef CK_ENABLE_FP32
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(Row{}, Row{}, Row{}, F32{}, F32{}, F32{}, F32{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(Row{}, Col{}, Row{}, F32{}, F32{}, F32{}, F32{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(Col{}, Row{}, Row{}, F32{}, F32{}, F32{}, F32{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(Col{}, Col{}, Row{}, F32{}, F32{}, F32{}, F32{});
+    }
+#endif
+#ifdef CK_ENABLE_FP16
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(Row{}, Row{}, Row{}, F16{}, F16{}, F32{}, F16{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(Row{}, Col{}, Row{}, F16{}, F16{}, F32{}, F16{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(Col{}, Row{}, Row{}, F16{}, F16{}, F32{}, F16{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(Col{}, Col{}, Row{}, F16{}, F16{}, F32{}, F16{});
+    }
+#endif
+#ifdef CK_ENABLE_BF16
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(Row{}, Row{}, Row{}, BF16{}, BF16{}, F32{}, BF16{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(Row{}, Col{}, Row{}, BF16{}, BF16{}, F32{}, BF16{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(Col{}, Row{}, Row{}, BF16{}, BF16{}, F32{}, BF16{});
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(Col{}, Col{}, Row{}, BF16{}, BF16{}, F32{}, BF16{});
+    }
+#endif
+#ifdef CK_ENABLE_INT8
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(Row{}, Row{}, Row{}, INT8{}, INT8{}, INT32{}, INT8{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(Row{}, Col{}, Row{}, INT8{}, INT8{}, INT32{}, INT8{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(Col{}, Row{}, Row{}, INT8{}, INT8{}, INT32{}, INT8{});
+    }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(Col{}, Col{}, Row{}, INT8{}, INT8{}, INT32{}, INT8{});
+    }
+#endif
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm);
diff --git a/profiler/src/profile_gemm_add_add_fastgelu.cpp b/profiler/src/profile_gemm_add_add_fastgelu.cpp
new file mode 100644
index 0000000000..8af3768a48
--- /dev/null
+++ b/profiler/src/profile_gemm_add_add_fastgelu.cpp
@@ -0,0 +1,158 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_add_add_fastgelu_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_add_add_fastgelu"
+#define OP_DESC "GEMM+Add+Add+FastGeLU"
+
+int profile_gemm_add_add_fastgelu(int argc, char* argv[])
+{
+    enum struct MatrixLayout
+    {
+        MK_KN_MN_MN_MN, // 0
+        MK_NK_MN_MN_MN, // 1
+        KM_KN_MN_MN_MN, // 2
+        KM_NK_MN_MN_MN, // 3
+    };
+
+    enum struct MatrixDataType
+    {
+        F32_F32_F32_F32_F32,      // 0
+        F16_F16_F16_F16_F16,      // 1
+        BF16_BF16_BF16_BF16_BF16, // 2
+        INT8_INT8_INT8_INT8_INT8, // 3
+    };
+
+    if(argc != 16)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n] + D1[m, n]);\n");
+        printf("                     1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n] + D1[m, n]);\n");
+        printf("                     2: E[m, n] = FastGeLU(A[k, m] * B[k, n] + D0[m, n] + D1[m, n]);\n");
+        printf("                     3: E[m, n] = FastGeLU(A[k, m] * B[n, k] + D0[m, n] + D1[m, n]))\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 15: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<MatrixDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<MatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA  = std::stoi(argv[11]);
+    const int StrideB  = std::stoi(argv[12]);
+    const int StrideD0 = std::stoi(argv[13]);
+    const int StrideD1 = std::stoi(argv[14]);
+    const int StrideE  = std::stoi(argv[15]);
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto d0_type,
+                       auto d1_type,
+                       auto e_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto d0_layout,
+                       auto d1_layout,
+                       auto e_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using D0DataType  = decltype(d0_type);
+        using D1DataType  = decltype(d1_type);
+        using EDataType   = decltype(e_type);
+
+        using ALayout  = decltype(a_layout);
+        using BLayout  = decltype(b_layout);
+        using D0Layout = decltype(d0_layout);
+        using D1Layout = decltype(d1_layout);
+        using ELayout  = decltype(e_layout);
+
+        const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB  = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
+        const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
+        const int DefaultStrideE  = ck::is_same_v<ELayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_add_add_fastgelu_impl<ADataType,
+                                                                     BDataType,
+                                                                     AccDataType,
+                                                                     D0DataType,
+                                                                     D1DataType,
+                                                                     EDataType,
+                                                                     ALayout,
+                                                                     BLayout,
+                                                                     D0Layout,
+                                                                     D1Layout,
+                                                                     ELayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
+            (StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
+            (StrideE < 0) ? DefaultStrideE : StrideE);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == MatrixDataType::F16_F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
+            layout == MatrixLayout::MK_NK_MN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
+            layout == MatrixLayout::KM_KN_MN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Row{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
+            layout == MatrixLayout::KM_NK_MN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add_add_fastgelu);
diff --git a/profiler/src/profile_gemm_add_fastgelu.cpp b/profiler/src/profile_gemm_add_fastgelu.cpp
new file mode 100644
index 0000000000..a09bb8340d
--- /dev/null
+++ b/profiler/src/profile_gemm_add_fastgelu.cpp
@@ -0,0 +1,146 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_add_fastgelu_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_add_fastgelu"
+#define OP_DESC "GEMM+Add+FastGeLU"
+
+int profile_gemm_add_fastgelu(int argc, char* argv[])
+{
+    enum struct MatrixLayout
+    {
+        MK_KN_MN_MN, // 0
+        MK_NK_MN_MN, // 1
+        KM_KN_MN_MN, // 2
+        KM_NK_MN_MN, // 3
+    };
+
+    enum struct MatrixDataType
+    {
+        F32_F32_F32_F32,     // 0
+        F16_F16_F16_F16,     // 1
+        BF16_BF16_BF16_BF16, // 2
+        INT8_INT8_INT8_INT8, // 3
+    };
+
+    if(argc != 15)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n]);\n");
+        printf("                     1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n]);\n");
+        printf("                     2: E[m, n] = FastGeLU(A[k, m] * B[k, n] + D0[m, n]);\n");
+        printf("                     3: E[m, n] = FastGeLU(A[k, m] * B[n, k] + D0[m, n]))\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideD0, StrideE\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<MatrixDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<MatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA  = std::stoi(argv[11]);
+    const int StrideB  = std::stoi(argv[12]);
+    const int StrideD0 = std::stoi(argv[13]);
+    const int StrideE  = std::stoi(argv[14]);
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto d0_type,
+                       auto e_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto d0_layout,
+                       auto e_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using D0DataType  = decltype(d0_type);
+        using EDataType   = decltype(e_type);
+
+        using ALayout  = decltype(a_layout);
+        using BLayout  = decltype(b_layout);
+        using D0Layout = decltype(d0_layout);
+        using ELayout  = decltype(e_layout);
+
+        const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB  = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
+        const int DefaultStrideE  = ck::is_same_v<ELayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_add_fastgelu_impl<ADataType,
+                                                                 BDataType,
+                                                                 AccDataType,
+                                                                 D0DataType,
+                                                                 EDataType,
+                                                                 ALayout,
+                                                                 BLayout,
+                                                                 D0Layout,
+                                                                 ELayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
+            (StrideE < 0) ? DefaultStrideE : StrideE);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::MK_NK_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::KM_KN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::KM_NK_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add_fastgelu);
diff --git a/profiler/src/profile_gemm_add_multiply.cpp b/profiler/src/profile_gemm_add_multiply.cpp
new file mode 100644
index 0000000000..560467c264
--- /dev/null
+++ b/profiler/src/profile_gemm_add_multiply.cpp
@@ -0,0 +1,158 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_add_multiply_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_add_multiply"
+#define OP_DESC "GEMM+Add+MULTIPLY"
+
+int profile_gemm_add_multiply(int argc, char* argv[])
+{
+    enum struct MatrixLayout
+    {
+        MK_KN_MN_MN_MN, // 0
+        MK_NK_MN_MN_MN, // 1
+        KM_KN_MN_MN_MN, // 2
+        KM_NK_MN_MN_MN, // 3
+    };
+
+    enum struct MatrixDataType
+    {
+        F32_F32_F32_F32_F32,      // 0
+        F16_F16_F16_F16_F16,      // 1
+        BF16_BF16_BF16_BF16_BF16, // 2
+        INT8_INT8_INT8_INT8_INT8, // 3
+    };
+
+    if(argc != 16)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg3: matrix layout (0: E[m, n] = AddMultiply((A[m, k] * B[k, n] + D0[m, n]) x D1[m, n]);\n");
+        printf("                     1: E[m, n] = AddMultiply((A[m, k] * B[k, n] + D0[m, n]) x D1[m, n]);\n");
+        printf("                     2: E[m, n] = AddMultiply((A[m, k] * B[k, n] + D0[m, n]) x D1[m, n]);\n");
+        printf("                     3: E[m, n] = AddMultiply((A[m, k] * B[k, n] + D0[m, n]) x D1[m, n]))\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 15: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<MatrixDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<MatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA  = std::stoi(argv[11]);
+    const int StrideB  = std::stoi(argv[12]);
+    const int StrideD0 = std::stoi(argv[13]);
+    const int StrideD1 = std::stoi(argv[14]);
+    const int StrideE  = std::stoi(argv[15]);
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto d0_type,
+                       auto d1_type,
+                       auto e_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto d0_layout,
+                       auto d1_layout,
+                       auto e_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using D0DataType  = decltype(d0_type);
+        using D1DataType  = decltype(d1_type);
+        using EDataType   = decltype(e_type);
+
+        using ALayout  = decltype(a_layout);
+        using BLayout  = decltype(b_layout);
+        using D0Layout = decltype(d0_layout);
+        using D1Layout = decltype(d1_layout);
+        using ELayout  = decltype(e_layout);
+
+        const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB  = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
+        const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
+        const int DefaultStrideE  = ck::is_same_v<ELayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_add_multiply_impl<ADataType,
+                                                                 BDataType,
+                                                                 AccDataType,
+                                                                 D0DataType,
+                                                                 D1DataType,
+                                                                 EDataType,
+                                                                 ALayout,
+                                                                 BLayout,
+                                                                 D0Layout,
+                                                                 D1Layout,
+                                                                 ELayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
+            (StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
+            (StrideE < 0) ? DefaultStrideE : StrideE);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == MatrixDataType::F16_F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
+            layout == MatrixLayout::MK_NK_MN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
+            layout == MatrixLayout::KM_KN_MN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Row{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
+            layout == MatrixLayout::KM_NK_MN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add_multiply);
diff --git a/profiler/src/profile_gemm_add_relu_add_layernorm.cpp b/profiler/src/profile_gemm_add_relu_add_layernorm.cpp
new file mode 100644
index 0000000000..558d255ce1
--- /dev/null
+++ b/profiler/src/profile_gemm_add_relu_add_layernorm.cpp
@@ -0,0 +1,215 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_add_relu_add_layernorm_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_add_relu_add_layernorm"
+#define OP_DESC "GEMM+Add+Relu+Add+Layernorm"
+
+int profile_gemm_add_relu_add_layernorm(int argc, char* argv[])
+{
+    enum struct MatrixLayout
+    {
+        MK_KN_MN_MN_MN, // 0
+        MK_NK_MN_MN_MN, // 1
+        KM_KN_MN_MN_MN, // 2
+        KM_NK_MN_MN_MN, // 3
+    };
+
+    enum struct MatrixDataType
+    {
+        F32,  // 0
+        F16,  // 1
+        BF16, // 2
+    };
+
+    if(argc != 16)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16)\n");
+        printf("arg3: matrix layout (0: H[m, n] = Layernorm(Relu(A[m, k] * B[k, n] + D0[m, n]) + D1[m, n]);\n");
+        printf("                     1: H[m, n] = Layernorm(Relu(A[m, k] * B[n, k] + D0[m, n]) + D1[m, n]);\n");
+        printf("                     2: H[m, n] = Layernorm(Relu(A[k, m] * B[k, n] + D0[m, n]) + D1[m, n]);\n");
+        printf("                     3: H[m, n] = Layernorm(Relu(A[k, m] * B[n, k] + D0[m, n]) + D1[m, n]))\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 15: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideH\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<MatrixDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<MatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA  = std::stoi(argv[11]);
+    const int StrideB  = std::stoi(argv[12]);
+    const int StrideD0 = std::stoi(argv[13]);
+    const int StrideD1 = std::stoi(argv[14]);
+    const int StrideH  = std::stoi(argv[15]);
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto d0_type,
+                       auto d1_type,
+                       auto e_mean_var_type,
+                       auto gamma_type,
+                       auto beta_type,
+                       auto h_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto d0_layout,
+                       auto d1_layout,
+                       auto h_layout) {
+        using ADataType        = decltype(a_type);
+        using BDataType        = decltype(b_type);
+        using AccDataType      = decltype(acc_type);
+        using D0DataType       = decltype(d0_type);
+        using D1DataType       = decltype(d1_type);
+        using EMeanVarDataType = decltype(e_mean_var_type);
+        using GammaDataType    = decltype(gamma_type);
+        using BetaDataType     = decltype(beta_type);
+        using HDataType        = decltype(h_type);
+
+        using ALayout  = decltype(a_layout);
+        using BLayout  = decltype(b_layout);
+        using D0Layout = decltype(d0_layout);
+        using D1Layout = decltype(d1_layout);
+        using HLayout  = decltype(h_layout);
+
+        const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB  = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
+        const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
+        const int DefaultStrideH  = ck::is_same_v<HLayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_add_relu_add_layernorm_impl<ADataType,
+                                                                           BDataType,
+                                                                           AccDataType,
+                                                                           D0DataType,
+                                                                           D1DataType,
+                                                                           EMeanVarDataType,
+                                                                           GammaDataType,
+                                                                           BetaDataType,
+                                                                           HDataType,
+                                                                           ALayout,
+                                                                           BLayout,
+                                                                           D0Layout,
+                                                                           D1Layout,
+                                                                           HLayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
+            (StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
+            (StrideH < 0) ? DefaultStrideH : StrideH);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == MatrixDataType::F16 && layout == MatrixLayout::MK_KN_MN_MN_MN)
+    {
+        return profile(F16{},
+                       F16{},
+                       F32{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       Row{},
+                       Row{},
+                       Row{},
+                       Row{},
+                       Row{});
+    }
+    else if(data_type == MatrixDataType::F16 && layout == MatrixLayout::MK_NK_MN_MN_MN)
+    {
+        return profile(F16{},
+                       F16{},
+                       F32{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       Row{},
+                       Col{},
+                       Row{},
+                       Row{},
+                       Row{});
+    }
+    else if(data_type == MatrixDataType::F16 && layout == MatrixLayout::KM_KN_MN_MN_MN)
+    {
+        return profile(F16{},
+                       F16{},
+                       F32{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       Col{},
+                       Row{},
+                       Row{},
+                       Row{},
+                       Row{});
+    }
+    else if(data_type == MatrixDataType::F16 && layout == MatrixLayout::KM_NK_MN_MN_MN)
+    {
+        return profile(F16{},
+                       F16{},
+                       F32{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       F16{},
+                       Col{},
+                       Col{},
+                       Row{},
+                       Row{},
+                       Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add_relu_add_layernorm);
diff --git a/profiler/src/profile_gemm_bias_add_reduce.cpp b/profiler/src/profile_gemm_bias_add_reduce.cpp
new file mode 100644
index 0000000000..76daffbc67
--- /dev/null
+++ b/profiler/src/profile_gemm_bias_add_reduce.cpp
@@ -0,0 +1,167 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_bias_add_reduce_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_bias_add_reduce"
+#define OP_DESC "GEMM+Bias+Add+Reduce"
+
+int profile_gemm_bias_add_reduce(int argc, char* argv[])
+{
+    enum struct GemmMatrixLayout
+    {
+        MK_KN_MN, // 0
+        MK_NK_MN, // 1
+        KM_KN_MN, // 2
+        KM_NK_MN, // 3
+    };
+
+    enum struct GemmReduceDataType
+    {
+        F32_F32_F32_F32_F32_F32_F32, // 0
+        F16_F16_F16_F16_F16_F32_F32, // 1
+    };
+
+    if(!(argc == 14 || argc == 15))
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
+        printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideC, StrideC1\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmReduceDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA  = std::stoi(argv[11]);
+    const int StrideB  = std::stoi(argv[12]);
+    const int StrideC  = std::stoi(argv[13]);
+    const int StrideC1 = std::stoi(argv[14]);
+
+    if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
+       layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        float,
+                                                        ck::tensor_layout::gemm::RowMajor,
+                                                        ck::tensor_layout::gemm::RowMajor,
+                                                        ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            (StrideC1 < 0) ? N : StrideC1);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        float,
+                                                        ck::tensor_layout::gemm::RowMajor,
+                                                        ck::tensor_layout::gemm::ColumnMajor,
+                                                        ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            (StrideC1 < 0) ? N : StrideC1);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        float,
+                                                        ck::tensor_layout::gemm::ColumnMajor,
+                                                        ck::tensor_layout::gemm::RowMajor,
+                                                        ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            (StrideC1 < 0) ? N : StrideC1);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        ck::profiler::profile_gemm_bias_add_reduce_impl<ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        ck::half_t,
+                                                        float,
+                                                        ck::tensor_layout::gemm::ColumnMajor,
+                                                        ck::tensor_layout::gemm::ColumnMajor,
+                                                        ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            (StrideC1 < 0) ? N : StrideC1);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this data_type & layout is not implemented");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_bias_add_reduce);
diff --git a/profiler/src/profile_gemm_bilinear.cpp b/profiler/src/profile_gemm_bilinear.cpp
new file mode 100644
index 0000000000..a1a48616b4
--- /dev/null
+++ b/profiler/src/profile_gemm_bilinear.cpp
@@ -0,0 +1,152 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_bilinear_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_bilinear"
+#define OP_DESC "GEMM+Bilinear"
+
+int profile_gemm_bilinear(int argc, char* argv[])
+{
+    enum struct MatrixLayout
+    {
+        MK_KN_MN_MN, // 0
+        MK_NK_MN_MN, // 1
+        KM_KN_MN_MN, // 2
+        KM_NK_MN_MN, // 3
+    };
+
+    enum struct MatrixDataType
+    {
+        F32_F32_F32_F32,     // 0
+        F16_F16_F16_F16,     // 1
+        BF16_BF16_BF16_BF16, // 2
+        INT8_INT8_INT8_INT8, // 3
+    };
+
+    if(argc != 17)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg3: matrix layout (0: E[m, n] = alpha * A[m, k] * B[k, n] + beta * D[m, n];\n");
+        printf("                     1: E[m, n] = alpha * A[m, k] * B[n, k] + beta * D[m, n];\n");
+        printf("                     2: E[m, n] = alpha * A[k, m] * B[k, n] + beta * D[m, n];\n");
+        printf("                     3: E[m, n] = alpha * A[k, m] * B[n, k] + beta * D[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideD, StrideE\n");
+        printf("arg15 to 16: alhpa, beta\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<MatrixDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<MatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideD = std::stoi(argv[13]);
+    const int StrideE = std::stoi(argv[14]);
+
+    const float alpha = std::stof(argv[15]);
+    const float beta  = std::stof(argv[16]);
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto d_type,
+                       auto e_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto d_layout,
+                       auto e_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using DDataType   = decltype(d_type);
+        using EDataType   = decltype(e_type);
+
+        using ALayout = decltype(a_layout);
+        using BLayout = decltype(b_layout);
+        using DLayout = decltype(d_layout);
+        using ELayout = decltype(e_layout);
+
+        const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideD = ck::is_same_v<DLayout, Row> ? N : M;
+        const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_bilinear_impl<ADataType,
+                                                             BDataType,
+                                                             AccDataType,
+                                                             DDataType,
+                                                             EDataType,
+                                                             ALayout,
+                                                             BLayout,
+                                                             DLayout,
+                                                             ELayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideD < 0) ? DefaultStrideD : StrideD,
+            (StrideE < 0) ? DefaultStrideE : StrideE,
+            alpha,
+            beta);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::MK_NK_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::KM_KN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16 && layout == MatrixLayout::KM_NK_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_bilinear);
diff --git a/profiler/src/profile_gemm_fastgelu.cpp b/profiler/src/profile_gemm_fastgelu.cpp
new file mode 100644
index 0000000000..93573002ef
--- /dev/null
+++ b/profiler/src/profile_gemm_fastgelu.cpp
@@ -0,0 +1,137 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_fastgelu_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_fastgelu"
+#define OP_DESC "GEMM+FastGeLU"
+
+int profile_gemm_fastgelu(int argc, char* argv[])
+{
+    enum struct MatrixLayout
+    {
+        MK_KN_MN, // 0
+        MK_NK_MN, // 1
+        KM_KN_MN, // 2
+        KM_NK_MN, // 3
+    };
+
+    enum struct MatrixDataType
+    {
+        F32_F32_F32,    // 0
+        F16_F16_F16,    // 1
+        BF16_BF16_BF16, // 2
+        INT8_INT8_INT8, // 3
+    };
+
+    if(argc != 14)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n]);\n");
+        printf("                     1: E[m, n] = FastGeLU(A[m, k] * B[n, k]);\n");
+        printf("                     2: E[m, n] = FastGeLU(A[k, m] * B[k, n]);\n");
+        printf("                     3: E[m, n] = FastGeLU(A[k, m] * B[n, k]))\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideE\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<MatrixDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<MatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideE = std::stoi(argv[13]);
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto e_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto e_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using EDataType   = decltype(e_type);
+
+        using ALayout = decltype(a_layout);
+        using BLayout = decltype(b_layout);
+        using ELayout = decltype(e_layout);
+
+        const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_fastgelu_impl<ADataType,
+                                                             BDataType,
+                                                             AccDataType,
+                                                             EDataType,
+                                                             ALayout,
+                                                             BLayout,
+                                                             ELayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideE < 0) ? DefaultStrideE : StrideE);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == MatrixDataType::F16_F16_F16 && layout == MatrixLayout::MK_KN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16 && layout == MatrixLayout::MK_NK_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16 && layout == MatrixLayout::KM_KN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16 && layout == MatrixLayout::KM_NK_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_fastgelu);
diff --git a/profiler/src/profile_gemm_multiply_add.cpp b/profiler/src/profile_gemm_multiply_add.cpp
new file mode 100644
index 0000000000..fd1f5c65c1
--- /dev/null
+++ b/profiler/src/profile_gemm_multiply_add.cpp
@@ -0,0 +1,153 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_multiply_add_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_multiply_add"
+#define OP_DESC "GEMM+MULTIPLY+ADD"
+
+int profile_gemm_multiply_add(int argc, char* argv[])
+{
+    enum struct MatrixLayout
+    {
+        MK_KN_MN_MN_MN, // 0
+        MK_NK_MN_MN_MN, // 1
+    };
+
+    enum struct MatrixDataType
+    {
+        F16_F16_F16_F16_F16, // 0
+        F16_F8_F32_F32_F16,  // 1
+    };
+
+    if(argc != 16)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp16; 1: fp16Afp8B)\n");
+        printf("arg3: matrix layout (0: E[m, n] = Multiply_Add((A[m, k] * B[k, n]) x D1[m, n] + D0[m, n]);\n");
+        printf("                     1: E[m, n] = Multiply_Add((A[m, k] * B[n, k]) x D1[m, n] + D0[m, n]);\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 15: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<MatrixDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<MatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA  = std::stoi(argv[11]);
+    const int StrideB  = std::stoi(argv[12]);
+    const int StrideD0 = std::stoi(argv[13]);
+    const int StrideD1 = std::stoi(argv[14]);
+    const int StrideE  = std::stoi(argv[15]);
+
+    using F8  = ck::f8_t;
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto d0_type,
+                       auto d1_type,
+                       auto e_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto d0_layout,
+                       auto d1_layout,
+                       auto e_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using D0DataType  = decltype(d0_type);
+        using D1DataType  = decltype(d1_type);
+        using EDataType   = decltype(e_type);
+
+        using ALayout  = decltype(a_layout);
+        using BLayout  = decltype(b_layout);
+        using D0Layout = decltype(d0_layout);
+        using D1Layout = decltype(d1_layout);
+        using ELayout  = decltype(e_layout);
+
+        const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB  = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
+        const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
+        const int DefaultStrideE  = ck::is_same_v<ELayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_multiply_add_impl<ADataType,
+                                                                 BDataType,
+                                                                 AccDataType,
+                                                                 D0DataType,
+                                                                 D1DataType,
+                                                                 EDataType,
+                                                                 ALayout,
+                                                                 BLayout,
+                                                                 D0Layout,
+                                                                 D1Layout,
+                                                                 ELayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
+            (StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
+            (StrideE < 0) ? DefaultStrideE : StrideE);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == MatrixDataType::F16_F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
+            layout == MatrixLayout::MK_NK_MN_MN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F8_F32_F32_F16 &&
+            layout == MatrixLayout::MK_KN_MN_MN_MN)
+    {
+        return profile(F16{}, F8{}, F32{}, F32{}, F32{}, F16{}, Row{}, Row{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::F16_F8_F32_F32_F16 &&
+            layout == MatrixLayout::MK_NK_MN_MN_MN)
+    {
+        return profile(F16{}, F8{}, F32{}, F32{}, F32{}, F16{}, Row{}, Col{}, Row{}, Row{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_multiply_add);
diff --git a/profiler/src/profile_gemm_reduce.cpp b/profiler/src/profile_gemm_reduce.cpp
new file mode 100644
index 0000000000..48f6f5eb49
--- /dev/null
+++ b/profiler/src/profile_gemm_reduce.cpp
@@ -0,0 +1,154 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_reduce_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_reduce"
+#define OP_DESC "GEMM+Reduce"
+
+int profile_gemm_reduce(int argc, char* argv[])
+{
+    enum struct GemmMatrixLayout
+    {
+        MK_KN_MN, // 0
+        MK_NK_MN, // 1
+        KM_KN_MN, // 2
+        KM_NK_MN, // 3
+    };
+
+    enum struct GemmReduceDataType
+    {
+        F32_F32_F32_F32_F32, // 0
+        F16_F16_F16_F32_F32, // 1
+    };
+
+    if(!(argc == 14 || argc == 15))
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
+        printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
+        printf("arg14: split k into  mulitiple batch\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmReduceDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideC = std::stoi(argv[13]);
+
+    if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        ck::profiler::profile_gemm_reduce_impl<ck::half_t,
+                                               ck::half_t,
+                                               ck::half_t,
+                                               float,
+                                               ck::tensor_layout::gemm::RowMajor,
+                                               ck::tensor_layout::gemm::RowMajor,
+                                               ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? K : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_gemm_reduce_impl<ck::half_t,
+                                               ck::half_t,
+                                               ck::half_t,
+                                               float,
+                                               ck::tensor_layout::gemm::RowMajor,
+                                               ck::tensor_layout::gemm::ColumnMajor,
+                                               ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? K : StrideA,
+            (StrideB < 0) ? K : StrideB,
+            (StrideC < 0) ? N : StrideC);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        ck::profiler::profile_gemm_reduce_impl<ck::half_t,
+                                               ck::half_t,
+                                               ck::half_t,
+                                               float,
+                                               ck::tensor_layout::gemm::ColumnMajor,
+                                               ck::tensor_layout::gemm::RowMajor,
+                                               ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? N : StrideB,
+            (StrideC < 0) ? N : StrideC);
+    }
+    else if(data_type == GemmReduceDataType::F16_F16_F16_F32_F32 &&
+            layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        ck::profiler::profile_gemm_reduce_impl<ck::half_t,
+                                               ck::half_t,
+                                               ck::half_t,
+                                               float,
+                                               ck::tensor_layout::gemm::ColumnMajor,
+                                               ck::tensor_layout::gemm::ColumnMajor,
+                                               ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? K : StrideB,
+            (StrideC < 0) ? N : StrideC);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this data_type & layout is not implemented");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_reduce);
diff --git a/profiler/src/profile_gemm_splitk.cpp b/profiler/src/profile_gemm_splitk.cpp
new file mode 100644
index 0000000000..617e0b9cd4
--- /dev/null
+++ b/profiler/src/profile_gemm_splitk.cpp
@@ -0,0 +1,189 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_splitk_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+};
+
+enum struct GemmDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+    F8_F16_F16,     // 4
+    F16_F8_F16,     // 5
+};
+
+#define OP_NAME "gemm_splitk"
+#define OP_DESC "Split-K GEMM"
+
+int profile_gemm_splitk(int argc, char* argv[])
+{
+    if(argc != 15)
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
+        printf("arg14: split k into  mulitiple batch\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideC = std::stoi(argv[13]);
+    const int KBatch  = std::stoi(argv[14]);
+
+    using F32 = float;
+    using F16 = ck::half_t;
+    using F8  = ck::f8_t;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto c_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto c_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using CDataType   = decltype(c_type);
+
+        using ALayout = decltype(a_layout);
+        using BLayout = decltype(b_layout);
+        using CLayout = decltype(c_layout);
+
+        const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_splitk_impl<ADataType,
+                                                           BDataType,
+                                                           AccDataType,
+                                                           CDataType,
+                                                           ALayout,
+                                                           BLayout,
+                                                           CLayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideC < 0) ? DefaultStrideC : StrideC,
+            KBatch);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(F8{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(F8{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(F16{}, F8{}, F32{}, F16{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(F16{}, F8{}, F32{}, F16{}, Col{}, Col{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_splitk);
diff --git a/profiler/src/profile_gemm_streamk.cpp b/profiler/src/profile_gemm_streamk.cpp
new file mode 100644
index 0000000000..a0a49eb36d
--- /dev/null
+++ b/profiler/src/profile_gemm_streamk.cpp
@@ -0,0 +1,155 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_streamk_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+};
+
+enum struct GemmDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+#define OP_NAME "gemm_streamk"
+#define OP_DESC "StreamK GEMM"
+
+int profile_gemm_streamk(int argc, char* argv[])
+{
+    if(argc < 14)
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideC\n");
+        printf("arg14: num_sk_blocks (optional)\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA = std::stoi(argv[11]);
+    const int StrideB = std::stoi(argv[12]);
+    const int StrideC = std::stoi(argv[13]);
+    const uint32_t NumSKBlocks =
+        argc >= 15 ? static_cast<uint32_t>(std::stoul(std::string(argv[14]))) : 0xffffffff;
+
+    using F32 = float;
+    using F16 = ck::half_t;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto c_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto c_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using CDataType   = decltype(c_type);
+
+        using ALayout = decltype(a_layout);
+        using BLayout = decltype(b_layout);
+        using CLayout = decltype(c_layout);
+
+        const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_streamk_impl<ADataType,
+                                                            BDataType,
+                                                            AccDataType,
+                                                            CDataType,
+                                                            ALayout,
+                                                            BLayout,
+                                                            CLayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA <= 0) ? DefaultStrideA : StrideA,
+            (StrideB <= 0) ? DefaultStrideB : StrideB,
+            (StrideC <= 0) ? DefaultStrideC : StrideC,
+            NumSKBlocks);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{});
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_streamk);
diff --git a/profiler/src/profile_grouped_conv_bwd_data.cpp b/profiler/src/profile_grouped_conv_bwd_data.cpp
new file mode 100644
index 0000000000..55d199317a
--- /dev/null
+++ b/profiler/src/profile_grouped_conv_bwd_data.cpp
@@ -0,0 +1,186 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_grouped_conv_bwd_data_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+namespace {
+
+enum struct ConvLayout
+{
+    GNHWC_GKYXC_GNHWK, // 0
+    NHWGC_GKYXC_NHWGK, // 1
+};
+
+enum struct ConvDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+};
+
+#define OP_NAME "grouped_conv_bwd_data"
+#define OP_DESC "Grouped Convolution Backward Data"
+
+static void print_helper_msg()
+{
+    std::cout
+        // clang-format off
+        << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+        << "arg2: data type (0: Output fp32, Weight fp32, Input fp32\n"
+        << "                 1: Output fp16, Weight fp16, Input fp16\n"
+        << "                 2: Output bf16, Weight bf16, Input bf16\n"
+        << "arg3: tensor layout (0: Output[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Input[G, N, Ho, Wo, K]\n"
+        << "                     1: Output[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Input[N, Ho, Wo, G, K])\n"
+        << "arg4: verification (0: no, 1: yes)\n"
+        << "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
+        << "arg6: print tensor value (0: no; 1: yes)\n"
+        << "arg7: time kernel (0: no, 1: yes)\n"
+        << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+    // clang-format on
+}
+
+} // namespace
+
+int profile_grouped_conv_bwd_data(int argc, char* argv[])
+{
+    // 8 for control, 1 for num_dim_spatial
+    if(argc < 9)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<ConvLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+    const int num_dim_spatial  = std::stoi(argv[8]);
+
+    // 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
+    if(argc != 8 + 1 + 4 + 6 * num_dim_spatial)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+
+    using namespace ck::tensor_layout::convolution;
+
+    constexpr auto I2 = ck::Number<2>{};
+    constexpr auto I3 = ck::Number<3>{};
+
+    auto profile = [&](auto num_dim_spatial_tmp,
+                       auto out_layout,
+                       auto wei_layout,
+                       auto in_layout,
+                       auto wei_type,
+                       auto out_type,
+                       auto in_type) {
+        constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
+
+        using OutLayout = decltype(out_layout);
+        using WeiLayout = decltype(wei_layout);
+        using InLayout  = decltype(in_layout);
+
+        using OutDataType = decltype(out_type);
+        using WeiDataType = decltype(wei_type);
+        using InDataType  = decltype(in_type);
+
+        bool pass = ck::profiler::profile_grouped_conv_bwd_data_impl<NDimSpatial,
+                                                                     OutLayout,
+                                                                     WeiLayout,
+                                                                     InLayout,
+                                                                     OutDataType,
+                                                                     WeiDataType,
+                                                                     InDataType>(
+            do_verification, init_method, do_log, time_kernel, params);
+
+        return pass ? 0 : 1;
+    };
+
+    if(num_dim_spatial == 2)
+    {
+        if(layout == ConvLayout::GNHWC_GKYXC_GNHWK)
+        {
+            if(data_type == ConvDataType::F32_F32_F32)
+            {
+                return profile(I2, GNHWK{}, GKYXC{}, GNHWC{}, F32{}, F32{}, F32{});
+            }
+            else if(data_type == ConvDataType::F16_F16_F16)
+            {
+                return profile(I2, GNHWK{}, GKYXC{}, GNHWC{}, F16{}, F16{}, F16{});
+            }
+            else if(data_type == ConvDataType::BF16_BF16_BF16)
+            {
+                return profile(I2, GNHWK{}, GKYXC{}, GNHWC{}, BF16{}, BF16{}, BF16{});
+            }
+        }
+        else if(layout == ConvLayout::NHWGC_GKYXC_NHWGK)
+        {
+            if(data_type == ConvDataType::F32_F32_F32)
+            {
+                return profile(I2, NHWGK{}, GKYXC{}, NHWGC{}, F32{}, F32{}, F32{});
+            }
+            else if(data_type == ConvDataType::F16_F16_F16)
+            {
+                return profile(I2, NHWGK{}, GKYXC{}, NHWGC{}, F16{}, F16{}, F16{});
+            }
+            else if(data_type == ConvDataType::BF16_BF16_BF16)
+            {
+                return profile(I2, NHWGK{}, GKYXC{}, NHWGC{}, BF16{}, BF16{}, BF16{});
+            }
+        }
+    }
+    else if(num_dim_spatial == 3)
+    {
+        if(layout == ConvLayout::GNHWC_GKYXC_GNHWK)
+        {
+            if(data_type == ConvDataType::F32_F32_F32)
+            {
+                return profile(I3, GNDHWK{}, GKZYXC{}, GNDHWC{}, F32{}, F32{}, F32{});
+            }
+            else if(data_type == ConvDataType::F16_F16_F16)
+            {
+                return profile(I3, GNDHWK{}, GKZYXC{}, GNDHWC{}, F16{}, F16{}, F16{});
+            }
+            else if(data_type == ConvDataType::BF16_BF16_BF16)
+            {
+                return profile(I3, GNDHWK{}, GKZYXC{}, GNDHWC{}, BF16{}, BF16{}, BF16{});
+            }
+        }
+        else if(layout == ConvLayout::NHWGC_GKYXC_NHWGK)
+        {
+            if(data_type == ConvDataType::F32_F32_F32)
+            {
+                return profile(I3, NDHWGK{}, GKZYXC{}, NDHWGC{}, F32{}, F32{}, F32{});
+            }
+            else if(data_type == ConvDataType::F16_F16_F16)
+            {
+                return profile(I3, NDHWGK{}, GKZYXC{}, NDHWGC{}, F16{}, F16{}, F16{});
+            }
+            else if(data_type == ConvDataType::BF16_BF16_BF16)
+            {
+                return profile(I3, NDHWGK{}, GKZYXC{}, NDHWGC{}, BF16{}, BF16{}, BF16{});
+            }
+        }
+    }
+
+    std::cout << "this data_type & layout is not implemented" << std::endl;
+
+    return 1;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_conv_bwd_data);
diff --git a/profiler/src/profile_grouped_conv_bwd_weight.cpp b/profiler/src/profile_grouped_conv_bwd_weight.cpp
new file mode 100644
index 0000000000..be8a3230f7
--- /dev/null
+++ b/profiler/src/profile_grouped_conv_bwd_weight.cpp
@@ -0,0 +1,207 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <initializer_list>
+#include <iostream>
+#include <numeric>
+
+#include "profiler/profile_grouped_conv_bwd_weight_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+namespace {
+
+enum struct ConvLayout
+{
+    GNCHW_GKCYX_GNKHW, // 0
+    GNHWC_GKYXC_GNHWK, // 1
+    NHWGC_GKYXC_NHWGK, // 2
+};
+
+enum struct ConvDataType
+{
+    F32_F32_F32,   // 0
+    F16_F16_F16,   // 1
+    BF16_F32_BF16, // 2
+};
+
+#define OP_NAME "grouped_conv_bwd_weight"
+#define OP_DESC "Grouped Convolution Backward Weight"
+
+static void print_helper_msg()
+{
+    std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+              << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
+              << "                 1: Input fp16, Weight fp16, Output fp16\n"
+              << "                 2: Input bf16, Weight fp32, Output bf16)\n"
+              << "arg3: tensor layout (0: Input[G, N, C, Hi, Wi], Weight[G, K, C, Y, X], Output[G, "
+                 "N, K, Ho, Wo]\n"
+              << "                     1: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, "
+                 "N, Ho, Wo, K]\n"
+              << "                     2: Input[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Output[N, "
+                 "Ho, Wo, G, K]\n"
+              << "arg4: verification (0: no, 1: yes)\n"
+              << "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
+              << "arg6: print tensor value (0: no; 1: yes)\n"
+              << "arg7: time kernel (0: no, 1: yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << " SplitK\n"
+              << std::endl;
+}
+
+} // namespace
+
+int profile_grouped_conv_bwd_weight(int argc, char* argv[])
+{
+    // 8 for control, 1 for num_dim_spatial
+    if(argc < 9)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<ConvLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+    const int num_dim_spatial  = std::stoi(argv[8]);
+
+    // 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial, 1 for split-K
+    if(argc != 8 + 1 + 4 + 6 * num_dim_spatial + 1)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
+
+    ck::index_t split_k = std::stoi(argv[8 + 1 + 4 + 6 * num_dim_spatial]);
+    split_k             = std::max(1, split_k);
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+
+    using namespace ck::tensor_layout::convolution;
+
+    constexpr auto I1 = ck::Number<1>{};
+    constexpr auto I2 = ck::Number<2>{};
+    constexpr auto I3 = ck::Number<3>{};
+
+    auto profile = [&](auto num_dim_spatial_tmp,
+                       auto in_layout,
+                       auto wei_layout,
+                       auto out_layout,
+                       auto in_type,
+                       auto wei_type,
+                       auto out_type) {
+        constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
+
+        using InLayout  = decltype(in_layout);
+        using WeiLayout = decltype(wei_layout);
+        using OutLayout = decltype(out_layout);
+
+        using InDataType  = decltype(in_type);
+        using WeiDataType = decltype(wei_type);
+        using OutDataType = decltype(out_type);
+
+        bool pass = ck::profiler::profile_grouped_conv_bwd_weight_impl<NDimSpatial,
+                                                                       InLayout,
+                                                                       WeiLayout,
+                                                                       OutLayout,
+                                                                       InDataType,
+                                                                       WeiDataType,
+                                                                       OutDataType>(
+            do_verification, init_method, do_log, time_kernel, params, split_k);
+
+        return pass ? 0 : 1;
+    };
+
+    if(num_dim_spatial == 1 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I1, GNWC{}, GKXC{}, GNWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I1, GNWC{}, GKXC{}, GNWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_F32_BF16)
+        {
+            // fp32 atomic add is used for weight tensor in bf16 kernel
+            return profile(I1, GNWC{}, GKXC{}, GNWK{}, BF16{}, F32{}, BF16{});
+        }
+    }
+    else if(num_dim_spatial == 2 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_F32_BF16)
+        {
+            // fp32 atomic add is used for weight tensor in bf16 kernel
+            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, BF16{}, F32{}, BF16{});
+        }
+    }
+    else if(num_dim_spatial == 2 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_F32_BF16)
+        {
+            // fp32 atomic add is used for weight tensor in bf16 kernel
+            return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, BF16{}, F32{}, BF16{});
+        }
+    }
+    else if(num_dim_spatial == 3 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_F32_BF16)
+        {
+            // fp32 atomic add is used for weight tensor in bf16 kernel
+            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, BF16{}, F32{}, BF16{});
+        }
+    }
+    else if(num_dim_spatial == 3 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_F32_BF16)
+        {
+            // fp32 atomic add is used for weight tensor in bf16 kernel
+            return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, BF16{}, F32{}, BF16{});
+        }
+    }
+
+    std::cout << "this data_type & layout is not implemented" << std::endl;
+
+    return 1;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_conv_bwd_weight);
diff --git a/profiler/src/profile_grouped_conv_fwd.cpp b/profiler/src/profile_grouped_conv_fwd.cpp
new file mode 100644
index 0000000000..d0b424cde6
--- /dev/null
+++ b/profiler/src/profile_grouped_conv_fwd.cpp
@@ -0,0 +1,260 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_grouped_conv_fwd_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+namespace {
+
+enum struct ConvLayout
+{
+    GNHWC_GKYXC_GNHWK, // 0
+    NHWGC_GKYXC_NHWGK, // 1
+};
+
+enum struct ConvDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+#define OP_NAME "grouped_conv_fwd"
+#define OP_DESC "Grouped Convolution Forward"
+
+static void print_helper_msg()
+{
+    std::cout
+        // clang-format off
+        << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+        << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
+        << "                 1: Input fp16, Weight fp16, Output fp16\n"
+        << "                 2: Input bf16, Weight bf16, Output bf16\n"
+        << "                 3: Input int8, Weight int8, Output int8)\n"
+        << "arg3: tensor layout (0: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, N, Ho, Wo, K]\n"
+        << "                     1: Input[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Output[N, Ho, Wo, G, K])\n"
+        << "arg4: verification (0: no, 1: yes)\n"
+        << "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
+        << "arg6: print tensor value (0: no; 1: yes)\n"
+        << "arg7: time kernel (0: no, 1: yes)\n"
+        << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+    // clang-format on
+}
+
+} // namespace
+
+int profile_grouped_conv_fwd(int argc, char* argv[])
+{
+    // 8 for control, 1 for num_dim_spatial
+    if(argc < 9)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto data_type       = static_cast<ConvDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<ConvLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+    const int num_dim_spatial  = std::stoi(argv[8]);
+
+    // 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
+    if(argc != 8 + 1 + 4 + 6 * num_dim_spatial)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+    using INT8 = int8_t;
+
+    //
+    using GNWC   = ck::tensor_layout::convolution::GNWC;
+    using GNHWC  = ck::tensor_layout::convolution::GNHWC;
+    using GNDHWC = ck::tensor_layout::convolution::GNDHWC;
+
+    using GKXC   = ck::tensor_layout::convolution::GKXC;
+    using GKYXC  = ck::tensor_layout::convolution::GKYXC;
+    using GKZYXC = ck::tensor_layout::convolution::GKZYXC;
+
+    using GNWK   = ck::tensor_layout::convolution::GNWK;
+    using GNHWK  = ck::tensor_layout::convolution::GNHWK;
+    using GNDHWK = ck::tensor_layout::convolution::GNDHWK;
+
+    //
+    using NWGC   = ck::tensor_layout::convolution::NWGC;
+    using NHWGC  = ck::tensor_layout::convolution::NHWGC;
+    using NDHWGC = ck::tensor_layout::convolution::NDHWGC;
+
+    using NWGK   = ck::tensor_layout::convolution::NWGK;
+    using NHWGK  = ck::tensor_layout::convolution::NHWGK;
+    using NDHWGK = ck::tensor_layout::convolution::NDHWGK;
+
+    constexpr auto I1 = ck::Number<1>{};
+    constexpr auto I2 = ck::Number<2>{};
+    constexpr auto I3 = ck::Number<3>{};
+
+    auto profile = [&](auto num_dim_spatial_tmp,
+                       auto in_layout,
+                       auto wei_layout,
+                       auto out_layout,
+                       auto in_type,
+                       auto wei_type,
+                       auto out_type) {
+        constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
+
+        using InLayout  = decltype(in_layout);
+        using WeiLayout = decltype(wei_layout);
+        using OutLayout = decltype(out_layout);
+
+        using InDataType  = decltype(in_type);
+        using WeiDataType = decltype(wei_type);
+        using OutDataType = decltype(out_type);
+
+        bool pass = ck::profiler::profile_grouped_conv_fwd_impl<NDimSpatial,
+                                                                InLayout,
+                                                                WeiLayout,
+                                                                OutLayout,
+                                                                InDataType,
+                                                                WeiDataType,
+                                                                OutDataType>(
+            do_verification, init_method, do_log, time_kernel, params);
+
+        return pass ? 0 : 1;
+    };
+
+    // GNHWC_GKYXC_GNHWK
+    if(num_dim_spatial == 1 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I1, GNWC{}, GKXC{}, GNWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I1, GNWC{}, GKXC{}, GNWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I1, GNWC{}, GKXC{}, GNWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I1, GNWC{}, GKXC{}, GNWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 2 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 3 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    // NHWGC_GKYXC_NHWGK
+    else if(num_dim_spatial == 1 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I1, NWGC{}, GKXC{}, NWGK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I1, NWGC{}, GKXC{}, NWGK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I1, NWGC{}, GKXC{}, NWGK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I1, NWGC{}, GKXC{}, NWGK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 2 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+    else if(num_dim_spatial == 3 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
+    {
+        if(data_type == ConvDataType::F32_F32_F32)
+        {
+            return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F32{}, F32{}, F32{});
+        }
+        else if(data_type == ConvDataType::F16_F16_F16)
+        {
+            return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F16{}, F16{}, F16{});
+        }
+        else if(data_type == ConvDataType::BF16_BF16_BF16)
+        {
+            return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, BF16{}, BF16{}, BF16{});
+        }
+        else if(data_type == ConvDataType::INT8_INT8_INT8)
+        {
+            return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, INT8{}, INT8{}, INT8{});
+        }
+    }
+
+    std::cout << "this data_type & layout is not implemented" << std::endl;
+
+    return 1;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_conv_fwd);
diff --git a/profiler/src/profile_grouped_gemm.cpp b/profiler/src/profile_grouped_gemm.cpp
new file mode 100644
index 0000000000..5636656ba3
--- /dev/null
+++ b/profiler/src/profile_grouped_gemm.cpp
@@ -0,0 +1,182 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_grouped_gemm_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+    MK_KN_NM, // 4
+    MK_NK_NM, // 5
+    KM_KN_NM, // 6
+    KM_NK_NM, // 7
+};
+
+enum struct GemmDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+#define OP_NAME "grouped_gemm"
+#define OP_DESC "Grouped GEMM"
+
+namespace {
+
+std::vector<int> argToIntArray(char* input)
+{
+    std::vector<int> out;
+
+    std::istringstream in(input);
+
+    std::string item;
+
+    while(std::getline(in, item, ','))
+    {
+        out.push_back(std::stoi(item));
+    }
+
+    return out;
+}
+
+int profile_grouped_gemm(int argc, char* argv[])
+{
+    if(argc < 14)
+    {
+        std::cout
+            << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+            << "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
+            << "arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n"
+            << "                     1: A[m, k] * B[n, k] = C[m, n];\n"
+            << "                     2: A[k, m] * B[k, n] = C[m, n];\n"
+            << "                     3: A[k, m] * B[n, k] = C[m, n])\n"
+            << "arg4: verification (0: no; 1: yes)\n"
+            << "arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+            << "arg6: print tensor value (0: no; 1: yes)\n"
+            << "arg7: time kernel (0=n0, 1=yes)\n"
+            << "arg8 to 13: Ms, Ns, Ks, StrideAs, StrideBs, StrideCs (e.g., 256,256 128,128 64,64 "
+               "64,64 64,64 128,128)\n"
+            << "arg15: kbatch value (default 4)\n"
+            << std::endl;
+
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const auto Ms = argToIntArray(argv[8]);
+    const auto Ns = argToIntArray(argv[9]);
+    const auto Ks = argToIntArray(argv[10]);
+
+    const auto StrideAs = argToIntArray(argv[11]);
+    const auto StrideBs = argToIntArray(argv[12]);
+    const auto StrideCs = argToIntArray(argv[13]);
+    const int kbatch    = argc == 15 ? std::stoi(argv[14]) : 1;
+#ifdef CK_ENABLE_FP16
+    if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        ck::profiler::profile_grouped_gemm_impl<ck::half_t,
+                                                ck::half_t,
+                                                ck::half_t,
+                                                float,
+                                                ck::tensor_layout::gemm::RowMajor,
+                                                ck::tensor_layout::gemm::RowMajor,
+                                                ck::tensor_layout::gemm::RowMajor>(do_verification,
+                                                                                   init_method,
+                                                                                   do_log,
+                                                                                   time_kernel,
+                                                                                   Ms,
+                                                                                   Ns,
+                                                                                   Ks,
+                                                                                   StrideAs,
+                                                                                   StrideBs,
+                                                                                   StrideCs,
+                                                                                   kbatch);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_grouped_gemm_impl<ck::half_t,
+                                                ck::half_t,
+                                                ck::half_t,
+                                                float,
+                                                ck::tensor_layout::gemm::RowMajor,
+                                                ck::tensor_layout::gemm::ColumnMajor,
+                                                ck::tensor_layout::gemm::RowMajor>(do_verification,
+                                                                                   init_method,
+                                                                                   do_log,
+                                                                                   time_kernel,
+                                                                                   Ms,
+                                                                                   Ns,
+                                                                                   Ks,
+                                                                                   StrideAs,
+                                                                                   StrideBs,
+                                                                                   StrideCs,
+                                                                                   kbatch);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        ck::profiler::profile_grouped_gemm_impl<ck::half_t,
+                                                ck::half_t,
+                                                ck::half_t,
+                                                float,
+                                                ck::tensor_layout::gemm::ColumnMajor,
+                                                ck::tensor_layout::gemm::RowMajor,
+                                                ck::tensor_layout::gemm::RowMajor>(do_verification,
+                                                                                   init_method,
+                                                                                   do_log,
+                                                                                   time_kernel,
+                                                                                   Ms,
+                                                                                   Ns,
+                                                                                   Ks,
+                                                                                   StrideAs,
+                                                                                   StrideBs,
+                                                                                   StrideCs,
+                                                                                   kbatch);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        ck::profiler::profile_grouped_gemm_impl<ck::half_t,
+                                                ck::half_t,
+                                                ck::half_t,
+                                                float,
+                                                ck::tensor_layout::gemm::ColumnMajor,
+                                                ck::tensor_layout::gemm::ColumnMajor,
+                                                ck::tensor_layout::gemm::RowMajor>(do_verification,
+                                                                                   init_method,
+                                                                                   do_log,
+                                                                                   time_kernel,
+                                                                                   Ms,
+                                                                                   Ns,
+                                                                                   Ks,
+                                                                                   StrideAs,
+                                                                                   StrideBs,
+                                                                                   StrideCs,
+                                                                                   kbatch);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
+    }
+#endif
+    return 0;
+}
+
+} // anonymous namespace
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_gemm);
diff --git a/profiler/src/profile_grouped_gemm_fastgelu.cpp b/profiler/src/profile_grouped_gemm_fastgelu.cpp
new file mode 100644
index 0000000000..50ecf25cae
--- /dev/null
+++ b/profiler/src/profile_grouped_gemm_fastgelu.cpp
@@ -0,0 +1,177 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_grouped_gemm_fastgelu_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+    MK_KN_NM, // 4
+    MK_NK_NM, // 5
+    KM_KN_NM, // 6
+    KM_NK_NM, // 7
+};
+
+enum struct GemmDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+};
+
+#define OP_NAME "grouped_gemm_fastgelu"
+#define OP_DESC "Grouped GEMM+FastGelu"
+
+namespace {
+
+std::vector<int> argToIntArray(char* input)
+{
+    std::vector<int> out;
+
+    std::istringstream in(input);
+
+    std::string item;
+
+    while(std::getline(in, item, ','))
+    {
+        out.push_back(std::stoi(item));
+    }
+
+    return out;
+}
+
+int profile_grouped_gemm_fastgelu(int argc, char* argv[])
+{
+    if(!(argc == 14))
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=n0, 1=yes)\n");
+        printf("arg8 to 13: Ms, Ns, Ks, StrideAs, StrideBs, StrideCs (e.g., 256,256 128,128 64,64 "
+               "64,64 64,64 128,128)\n");
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const auto Ms = argToIntArray(argv[8]);
+    const auto Ns = argToIntArray(argv[9]);
+    const auto Ks = argToIntArray(argv[10]);
+
+    const auto StrideAs = argToIntArray(argv[11]);
+    const auto StrideBs = argToIntArray(argv[12]);
+    const auto StrideCs = argToIntArray(argv[13]);
+
+    if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    {
+        ck::profiler::profile_grouped_gemm_fastgelu_impl<ck::half_t,
+                                                         ck::half_t,
+                                                         ck::half_t,
+                                                         float,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_grouped_gemm_fastgelu_impl<ck::half_t,
+                                                         ck::half_t,
+                                                         ck::half_t,
+                                                         float,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::ColumnMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    {
+        ck::profiler::profile_grouped_gemm_fastgelu_impl<ck::half_t,
+                                                         ck::half_t,
+                                                         ck::half_t,
+                                                         float,
+                                                         ck::tensor_layout::gemm::ColumnMajor,
+                                                         ck::tensor_layout::gemm::RowMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs);
+    }
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    {
+        ck::profiler::profile_grouped_gemm_fastgelu_impl<ck::half_t,
+                                                         ck::half_t,
+                                                         ck::half_t,
+                                                         float,
+                                                         ck::tensor_layout::gemm::ColumnMajor,
+                                                         ck::tensor_layout::gemm::ColumnMajor,
+                                                         ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            Ms,
+            Ns,
+            Ks,
+            StrideAs,
+            StrideBs,
+            StrideCs);
+    }
+    else
+    {
+        throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
+    }
+
+    return 0;
+}
+
+} // anonymous namespace
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_grouped_gemm_fastgelu);
diff --git a/profiler/src/profile_groupnorm.cpp b/profiler/src/profile_groupnorm.cpp
new file mode 100644
index 0000000000..d55784ff0a
--- /dev/null
+++ b/profiler/src/profile_groupnorm.cpp
@@ -0,0 +1,112 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <unordered_map>
+
+#include "profiler/data_type_enum.hpp"
+#include "profiler/profile_groupnorm_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+struct GroupnormArgParser
+{
+    std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}}};
+
+    bool parse_opt(int argc, char* argv[], const std::string& key, int i)
+    {
+        if(std::string("--") + key == argv[i])
+        {
+            int pos = i;
+            while(++i < argc && argv[i][0] != '-') {}
+            int end = i;
+            for(int j = pos + 1; j < end; j++)
+            {
+                long_opts[key].push_back(std::stoi(argv[j]));
+            }
+            return true;
+        }
+        return false;
+    }
+
+    void operator()(int argc, char* argv[])
+    {
+        for(auto& kv : long_opts)
+        {
+            for(int i = 1; i < argc; i++)
+            {
+                if(parse_opt(argc, argv, kv.first, i))
+                    break;
+            }
+        }
+    }
+};
+
+#define OP_NAME "groupnorm"
+#define OP_DESC "Group Normalization"
+
+void print_help_groupnorm()
+{
+    std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+              << "arg2: data type (0: fp16; 1: fp32)\n"
+              << "arg3: verification (0: no; 1: yes)\n"
+              << "arg4: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg5: print tensor value (0: no; 1: yes)\n"
+              << "arg6: time kernel (0=no, 1=yes)\n"
+              << "--length: tensor extents (e.g, --length 1 16 16 32 40) \n"
+              << std::endl;
+}
+
+int profile_groupnorm(int argc, char* argv[])
+{
+    ck::DataTypeEnum data_type  = ck::DataTypeEnum::Half;
+    bool do_verification        = false;
+    int init_method             = 0;
+    bool do_log                 = false;
+    bool time_kernel            = 1;
+    std::vector<index_t> length = {64, 16, 16, 32, 40};
+
+    if(argc != 1 && argc != 13)
+    {
+        print_help_groupnorm();
+        return 0;
+    }
+
+    if(argc == 13)
+    {
+        data_type       = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
+        do_verification = std::stoi(argv[3]);
+        init_method     = std::stoi(argv[4]);
+        do_log          = std::stoi(argv[5]);
+        time_kernel     = std::stoi(argv[6]);
+
+        // parse the long options
+        GroupnormArgParser arg_parser;
+        arg_parser(argc, argv);
+        length = arg_parser.long_opts["length"];
+    }
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    if(data_type == ck::DataTypeEnum::Float)
+    {
+        ck::profiler::profile_groupnorm_impl<F32, F32, F32, F32, F32>(
+            do_verification, init_method, do_log, time_kernel, length);
+    }
+    else if(data_type == ck::DataTypeEnum::Half)
+    {
+        ck::profiler::profile_groupnorm_impl<F16, F16, F16, F32, F16>(
+            do_verification, init_method, do_log, time_kernel, length);
+    }
+    else
+    {
+        throw std::runtime_error("not implemented yet");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_groupnorm);
diff --git a/profiler/src/profile_image_to_column.cpp b/profiler/src/profile_image_to_column.cpp
new file mode 100644
index 0000000000..bf4312a6cf
--- /dev/null
+++ b/profiler/src/profile_image_to_column.cpp
@@ -0,0 +1,169 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_image_to_column_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+namespace {
+
+enum struct ConvLayout
+{
+    NHWC, // 0
+};
+
+enum struct DataType
+{
+    F32_F32,   // 0
+    F16_F16,   // 1
+    BF16_BF16, // 2
+    INT8_INT8, // 3
+};
+
+#define OP_NAME "image_to_column"
+#define OP_DESC "Image To Column"
+
+static void print_helper_msg()
+{
+    std::cout
+        // clang-format off
+        << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
+        << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
+        << "                 1: Input fp16, Weight fp16, Output fp16\n"
+        << "                 2: Input bf16, Weight bf16, Output bf16\n"
+        << "                 3: Input int8, Weight int8, Output int8)\n"
+        << "arg3: tensor layout (0: Input[N, Hi, Wi, C], Output[N * Ho * Wo, Y * X * C])\n"
+        << "arg4: verification (0: no, 1: yes)\n"
+        << "arg5: initialization (0: no init, 1: integer value, 2: decimal value)\n"
+        << "arg6: print tensor value (0: no; 1: yes)\n"
+        << "arg7: time kernel (0: no, 1: yes)\n"
+        << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+    // clang-format on
+}
+
+} // namespace
+
+int profile_image_to_column(int argc, char* argv[])
+{
+    // 8 for control, 1 for num_dim_spatial
+    if(argc < 9)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto data_type       = static_cast<DataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<ConvLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+    const int num_dim_spatial  = std::stoi(argv[8]);
+
+    // 8 for control, 1 for num_dim_spatial, 4 for G/N/K/C, and 6 * num_dim_spatial
+    if(argc != 8 + 1 + 4 + 6 * num_dim_spatial)
+    {
+        print_helper_msg();
+        return 1;
+    }
+
+    const auto params = ck::utils::conv::parse_conv_param(num_dim_spatial, 9, argv);
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+    using INT8 = int8_t;
+
+    using namespace ck::tensor_layout::convolution;
+
+    constexpr auto I1 = ck::Number<1>{};
+    constexpr auto I2 = ck::Number<2>{};
+    constexpr auto I3 = ck::Number<3>{};
+
+    auto profile = [&](auto num_dim_spatial_tmp, auto in_layout, auto in_type, auto out_type) {
+        constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
+
+        using InLayout = decltype(in_layout);
+
+        using InDataType  = decltype(in_type);
+        using OutDataType = decltype(out_type);
+
+        bool pass = ck::profiler::
+            profile_image_to_column_impl<NDimSpatial, InLayout, InDataType, OutDataType>(
+                do_verification, init_method, do_log, time_kernel, params);
+
+        return pass ? 0 : 1;
+    };
+
+    // NHWC
+    if(layout == ConvLayout::NHWC)
+    {
+        if(num_dim_spatial == 1)
+        {
+            if(data_type == DataType::F32_F32)
+            {
+                return profile(I1, GNWC{}, F32{}, F32{});
+            }
+            else if(data_type == DataType::F16_F16)
+            {
+                return profile(I1, GNWC{}, F16{}, F16{});
+            }
+            else if(data_type == DataType::BF16_BF16)
+            {
+                return profile(I1, GNWC{}, BF16{}, BF16{});
+            }
+            else if(data_type == DataType::INT8_INT8)
+            {
+                return profile(I1, GNWC{}, INT8{}, INT8{});
+            }
+        }
+        else if(num_dim_spatial == 2)
+        {
+            if(data_type == DataType::F32_F32)
+            {
+                return profile(I2, GNHWC{}, F32{}, F32{});
+            }
+            else if(data_type == DataType::F16_F16)
+            {
+                return profile(I2, GNHWC{}, F16{}, F16{});
+            }
+            else if(data_type == DataType::BF16_BF16)
+            {
+                return profile(I2, GNHWC{}, BF16{}, BF16{});
+            }
+            else if(data_type == DataType::INT8_INT8)
+            {
+                return profile(I2, GNHWC{}, INT8{}, INT8{});
+            }
+        }
+        else if(num_dim_spatial == 3)
+        {
+            if(data_type == DataType::F32_F32)
+            {
+                return profile(I3, GNDHWC{}, F32{}, F32{});
+            }
+            else if(data_type == DataType::F16_F16)
+            {
+                return profile(I3, GNDHWC{}, F16{}, F16{});
+            }
+            else if(data_type == DataType::BF16_BF16)
+            {
+                return profile(I3, GNDHWC{}, BF16{}, BF16{});
+            }
+            else if(data_type == DataType::INT8_INT8)
+            {
+                return profile(I3, GNDHWC{}, INT8{}, INT8{});
+            }
+        }
+    }
+
+    std::cout << "this data_type & layout is not implemented" << std::endl;
+
+    return 1;
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_image_to_column);
diff --git a/profiler/src/profile_layernorm.cpp b/profiler/src/profile_layernorm.cpp
new file mode 100644
index 0000000000..7bf210e678
--- /dev/null
+++ b/profiler/src/profile_layernorm.cpp
@@ -0,0 +1,101 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <unordered_map>
+
+#include "profiler/data_type_enum.hpp"
+#include "profiler/profile_layernorm_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+struct LayernormArgParser
+{
+    std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}}};
+
+    bool parse_opt(int argc, char* argv[], const std::string& key, int i)
+    {
+        if(std::string("--") + key == argv[i])
+        {
+            int pos = i;
+            while(++i < argc && argv[i][0] != '-') {}
+            int end = i;
+            for(int j = pos + 1; j < end; j++)
+            {
+                long_opts[key].push_back(std::stoi(argv[j]));
+            }
+            return true;
+        }
+        return false;
+    }
+
+    void operator()(int argc, char* argv[])
+    {
+        for(auto& kv : long_opts)
+        {
+            for(int i = 1; i < argc; i++)
+            {
+                if(parse_opt(argc, argv, kv.first, i))
+                    break;
+            }
+        }
+    }
+};
+
+void print_help_layernorm()
+{
+    std::cout << "arg1: data type (0: fp16; 1: fp32)\n"
+              << "arg2: verification (0: no; 1: yes)\n"
+              << "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg4: print tensor value (0: no; 1: yes)\n"
+              << "arg5: time kernel (0=no, 1=yes)\n"
+              << "--length: tensor extents (e.g, --length 1024 1024) \n"
+              << std::endl;
+}
+
+int profile_layernorm(int argc, char* argv[])
+{
+    if(argc <= 2)
+    {
+        print_help_layernorm();
+        return 0;
+    }
+
+    LayernormArgParser arg_parser;
+
+    // short unnamed options
+    const ck::DataTypeEnum data_type = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
+    const bool do_verification       = std::stoi(argv[3]);
+    const int init_method            = std::stoi(argv[4]);
+    const bool do_log                = std::stoi(argv[5]);
+    const bool time_kernel           = std::stoi(argv[6]);
+
+    // parse the long options
+    arg_parser(argc, argv);
+    const std::vector<index_t> length = arg_parser.long_opts["length"];
+
+    using F16          = ck::half_t;
+    using F32          = float;
+    constexpr int rank = 2;
+
+    if(data_type == ck::DataTypeEnum::Half)
+    {
+        ck::profiler::profile_layernorm_impl<F16, F16, F16, F32, F16, rank>(
+            do_verification, init_method, do_log, time_kernel, length);
+    }
+    else if(data_type == ck::DataTypeEnum::Float)
+    {
+        ck::profiler::profile_layernorm_impl<F32, F32, F32, F32, F32, rank>(
+            do_verification, init_method, do_log, time_kernel, length);
+    }
+    else
+    {
+        throw std::runtime_error("not implemented yet");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("layernorm", "Layer Normalization", profile_layernorm);
diff --git a/profiler/src/profile_max_pool3d_bwd.cpp b/profiler/src/profile_max_pool3d_bwd.cpp
new file mode 100644
index 0000000000..45a64df423
--- /dev/null
+++ b/profiler/src/profile_max_pool3d_bwd.cpp
@@ -0,0 +1,177 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <unordered_map>
+
+#include "profiler/data_type_enum.hpp"
+#include "profiler/profile_max_pool3d_bwd_impl.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+struct maxPoolbwdArgParser
+{
+    std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}},
+                                                                   {"wsize", {}},
+                                                                   {"wstride", {}},
+                                                                   {"wdilation", {}},
+                                                                   {"pad1", {}},
+                                                                   {"pad2", {}}};
+
+    bool parse_opt(int argc, char* argv[], const std::string& key, int i)
+    {
+        if(std::string("--") + key == argv[i])
+        {
+            int pos = i;
+            while(++i < argc && argv[i][0] != '-') {}
+            int end = i;
+            for(int j = pos + 1; j < end; j++)
+            {
+                long_opts[key].push_back(std::stoi(argv[j]));
+            }
+            return true;
+        }
+        return false;
+    }
+
+    void operator()(int argc, char* argv[])
+    {
+        for(auto& kv : long_opts)
+        {
+            for(int i = 1; i < argc; i++)
+            {
+                if(parse_opt(argc, argv, kv.first, i))
+                    break;
+            }
+        }
+    }
+};
+
+void print_help_max_pool3d_bwd()
+{
+    std::cout << "arg1: data type (0: fp16; 1: fp32; 5: bf16)\n"
+              << "arg2: verification (0: no; 1: yes)\n"
+              << "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg4: print tensor value (0: no; 1: yes)\n"
+              << "arg5: time kernel (0=no, 1=yes)\n"
+              << "--length: input tensor length for NCDHW(e.g, --length 2 32 30 30 30) \n"
+              << "--wsize: window size for ZYX (e.g, --wsize 2 2 2) \n"
+              << "--wstride: window stride for DHW (e.g, --wstride 2 2 2) \n"
+              << "--wdilation: window dilation for DHW (e.g, --wdilation 1 1 1) \n"
+              << "--pad1: left side of padding in DHW (e.g, --pad1 1 1 1) \n"
+              << "--pad2: right side of padding in DHW (e.g, --pad2 1 1 1) \n"
+              << "eg: ckProfiler max_pool3d_bwd 0 1 2 0 1 --length 2 32 30 30 30 --wsize 2 2 2 "
+                 "--wstride 2 2 2 --wdilation 1 1 1 --pad1 1 1 1 --pad2 1 1 1"
+              << std::endl;
+}
+
+int profile_max_pool3d_bwd(int argc, char* argv[])
+{
+    ck::DataTypeEnum data_type = ck::DataTypeEnum::Half;
+    bool do_verification       = true;
+    int init_method            = 0;
+    bool do_log                = false;
+    bool time_kernel           = true;
+
+    std::vector<index_t> in_length = {2, 32, 30, 30, 30};
+    std::vector<index_t> wsize     = {2, 2, 2};
+    std::vector<index_t> wstride   = {2, 2, 2};
+    std::vector<index_t> wdilation = {1, 1, 1};
+    std::vector<index_t> pad1      = {1, 1, 1};
+    std::vector<index_t> pad2      = {1, 1, 1};
+
+    if(argc != 2 && argc != 33)
+    {
+        print_help_max_pool3d_bwd();
+        return 0;
+    }
+    else if(argc == 33)
+    {
+        data_type       = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
+        do_verification = std::stoi(argv[3]);
+        init_method     = std::stoi(argv[4]);
+        do_log          = std::stoi(argv[5]);
+        time_kernel     = std::stoi(argv[6]);
+
+        // parse the long options
+        maxPoolbwdArgParser arg_parser;
+        arg_parser(argc, argv);
+        in_length = arg_parser.long_opts["length"];
+        wsize     = arg_parser.long_opts["wsize"];
+        wstride   = arg_parser.long_opts["wstride"];
+        wdilation = arg_parser.long_opts["wdilation"];
+        pad1      = arg_parser.long_opts["pad1"];
+        pad2      = arg_parser.long_opts["pad2"];
+    }
+
+#ifdef CK_ENABLE_FP16
+    using F16 = ck::half_t;
+#endif
+#ifdef CK_ENABLE_BF16
+    using BF16 = ck::bhalf_t;
+#endif
+#ifdef CK_ENABLE_FP32
+    using F32 = float;
+#endif
+    using I32 = int32_t;
+
+    if(false)
+        ;
+#ifdef CK_ENABLE_FP16
+    else if(data_type == ck::DataTypeEnum::Half)
+    {
+        ck::profiler::profile_max_pool3d_bwd_impl<F16, F16, I32, F16, F16, false>(do_verification,
+                                                                                  init_method,
+                                                                                  do_log,
+                                                                                  time_kernel,
+                                                                                  in_length,
+                                                                                  wsize,
+                                                                                  wstride,
+                                                                                  wdilation,
+                                                                                  pad1,
+                                                                                  pad2);
+    }
+#endif
+#ifdef CK_ENABLE_BF16
+    else if(data_type == ck::DataTypeEnum::BFloat16)
+    {
+        ck::profiler::profile_max_pool3d_bwd_impl<BF16, BF16, I32, BF16, BF16, false>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            in_length,
+            wsize,
+            wstride,
+            wdilation,
+            pad1,
+            pad2);
+    }
+#endif
+#ifdef CK_ENABLE_FP32
+    else if(data_type == ck::DataTypeEnum::Float)
+    {
+        ck::profiler::profile_max_pool3d_bwd_impl<F32, F32, I32, F32, F32, false>(do_verification,
+                                                                                  init_method,
+                                                                                  do_log,
+                                                                                  time_kernel,
+                                                                                  in_length,
+                                                                                  wsize,
+                                                                                  wstride,
+                                                                                  wdilation,
+                                                                                  pad1,
+                                                                                  pad2);
+    }
+#endif
+    else
+    {
+        throw std::runtime_error("not implemented yet");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("max_pool3d_bwd", "max_pool3d bwd", profile_max_pool3d_bwd);
diff --git a/profiler/src/profile_max_pool3d_fwd.cpp b/profiler/src/profile_max_pool3d_fwd.cpp
new file mode 100644
index 0000000000..52fdf29fe4
--- /dev/null
+++ b/profiler/src/profile_max_pool3d_fwd.cpp
@@ -0,0 +1,245 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <unordered_map>
+
+#include "profiler/data_type_enum.hpp"
+#include "profiler/profile_pool3d_fwd_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+struct maxPoolFwdArgParser
+{
+    std::unordered_map<std::string, std::vector<int>> long_opts = {{"length", {}},
+                                                                   {"wsize", {}},
+                                                                   {"wstride", {}},
+                                                                   {"wdilation", {}},
+                                                                   {"pad1", {}},
+                                                                   {"pad2", {}}};
+
+    bool parse_opt(int argc, char* argv[], const std::string& key, int i)
+    {
+        if(std::string("--") + key == argv[i])
+        {
+            int pos = i;
+            while(++i < argc && argv[i][0] != '-') {}
+            int end = i;
+            for(int j = pos + 1; j < end; j++)
+            {
+                long_opts[key].push_back(std::stoi(argv[j]));
+            }
+            return true;
+        }
+        return false;
+    }
+
+    void operator()(int argc, char* argv[])
+    {
+        for(auto& kv : long_opts)
+        {
+            for(int i = 1; i < argc; i++)
+            {
+                if(parse_opt(argc, argv, kv.first, i))
+                    break;
+            }
+        }
+    }
+};
+
+void print_help_max_pool3d_fwd()
+{
+    std::cout << "arg1: data type (0: fp16; 1: fp32; 5: bf16)\n"
+              << "arg2: verification (0: no; 1: yes)\n"
+              << "arg3: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg4: print tensor value (0: no; 1: yes)\n"
+              << "arg5: time kernel (0=no, 1=yes)\n"
+              << "arg6: return index (0=no, 1=yes)\n"
+              << "--length: input tensor length for NCDHW(e.g, --length 2 32 30 30 30) \n"
+              << "--wsize: window size for ZYX (e.g, --wsize 2 2 2) \n"
+              << "--wstride: window stride for DHW (e.g, --wstride 2 2 2) \n"
+              << "--wdilation: window dilation for DHW (e.g, --wdilation 1 1 1) \n"
+              << "--pad1: left side of padding in DHW (e.g, --pad1 1 1 1) \n"
+              << "--pad2: right side of padding in DHW (e.g, --pad2 1 1 1) \n"
+              << "eg: ckProfiler max_pool3d_fwd 0 1 2 0 1 0 --length 2 32 30 30 30 --wsize 2 2 2 "
+                 "--wstride 2 2 2 --wdilation 1 1 1 --pad1 1 1 1 --pad2 1 1 1"
+              << std::endl;
+}
+
+int profile_max_pool3d_fwd(int argc, char* argv[])
+{
+    ck::DataTypeEnum data_type = ck::DataTypeEnum::Half;
+    bool do_verification       = true;
+    int init_method            = 0;
+    bool do_log                = false;
+    bool time_kernel           = true;
+    bool return_index          = false;
+
+    std::vector<index_t> in_length = {2, 32, 30, 30, 30};
+    std::vector<index_t> wsize     = {2, 2, 2};
+    std::vector<index_t> wstride   = {2, 2, 2};
+    std::vector<index_t> wdilation = {1, 1, 1};
+    std::vector<index_t> pad1      = {1, 1, 1};
+    std::vector<index_t> pad2      = {1, 1, 1};
+
+    if(argc != 2 && argc != 34)
+    {
+        print_help_max_pool3d_fwd();
+        return 0;
+    }
+    else if(argc == 34)
+    {
+        data_type       = static_cast<ck::DataTypeEnum>(std::stoi(argv[2]));
+        do_verification = std::stoi(argv[3]);
+        init_method     = std::stoi(argv[4]);
+        do_log          = std::stoi(argv[5]);
+        time_kernel     = std::stoi(argv[6]);
+        return_index    = std::stoi(argv[7]);
+
+        // parse the long options
+        maxPoolFwdArgParser arg_parser;
+        arg_parser(argc, argv);
+        in_length = arg_parser.long_opts["length"];
+        wsize     = arg_parser.long_opts["wsize"];
+        wstride   = arg_parser.long_opts["wstride"];
+        wdilation = arg_parser.long_opts["wdilation"];
+        pad1      = arg_parser.long_opts["pad1"];
+        pad2      = arg_parser.long_opts["pad2"];
+    }
+
+#ifdef CK_ENABLE_FP16
+    using F16 = ck::half_t;
+#endif
+#ifdef CK_ENABLE_BF16
+    using BF16 = ck::bhalf_t;
+#endif
+#ifdef CK_ENABLE_FP32
+    using F32 = float;
+#endif
+    using I32   = int32_t;
+    using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+#if 1
+    constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
+#else
+    constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
+#endif
+
+    if(false)
+        ;
+#ifdef CK_ENABLE_FP16
+    else if(data_type == ck::DataTypeEnum::Half)
+    {
+        if(return_index)
+            ck::profiler::
+                profile_pool3d_fwd_impl<F16, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, false, true>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    in_length,
+                    wsize,
+                    wstride,
+                    wdilation,
+                    pad1,
+                    pad2);
+        else
+            ck::profiler::
+                profile_pool3d_fwd_impl<F16, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, false, false>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    in_length,
+                    wsize,
+                    wstride,
+                    wdilation,
+                    pad1,
+                    pad2);
+    }
+#endif
+#ifdef CK_ENABLE_BF16
+    else if(data_type == ck::DataTypeEnum::BFloat16)
+    {
+        if(return_index)
+            ck::profiler::profile_pool3d_fwd_impl<BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  I32,
+                                                  NDHWC,
+                                                  NDHWC,
+                                                  ReduceOpId,
+                                                  false,
+                                                  true>(do_verification,
+                                                        init_method,
+                                                        do_log,
+                                                        time_kernel,
+                                                        in_length,
+                                                        wsize,
+                                                        wstride,
+                                                        wdilation,
+                                                        pad1,
+                                                        pad2);
+        else
+            ck::profiler::profile_pool3d_fwd_impl<BF16,
+                                                  BF16,
+                                                  BF16,
+                                                  I32,
+                                                  NDHWC,
+                                                  NDHWC,
+                                                  ReduceOpId,
+                                                  false,
+                                                  false>(do_verification,
+                                                         init_method,
+                                                         do_log,
+                                                         time_kernel,
+                                                         in_length,
+                                                         wsize,
+                                                         wstride,
+                                                         wdilation,
+                                                         pad1,
+                                                         pad2);
+    }
+#endif
+#ifdef CK_ENABLE_FP32
+    else if(data_type == ck::DataTypeEnum::Float)
+    {
+        if(return_index)
+            ck::profiler::
+                profile_pool3d_fwd_impl<F32, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, false, true>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    in_length,
+                    wsize,
+                    wstride,
+                    wdilation,
+                    pad1,
+                    pad2);
+        else
+            ck::profiler::
+                profile_pool3d_fwd_impl<F32, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, false, false>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    in_length,
+                    wsize,
+                    wstride,
+                    wdilation,
+                    pad1,
+                    pad2);
+    }
+#endif
+    else
+    {
+        throw std::runtime_error("not implemented yet");
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("max_pool3d_fwd", "max_pool3d fwd", profile_max_pool3d_fwd);
diff --git a/profiler/src/profile_reduce.cpp b/profiler/src/profile_reduce.cpp
new file mode 100644
index 0000000000..e4af5680a5
--- /dev/null
+++ b/profiler/src/profile_reduce.cpp
@@ -0,0 +1,434 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <fstream>
+#include <cstdlib>
+#include <vector>
+#include <stdexcept>
+#include <sstream>
+#include <getopt.h>
+
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/utility/host_common_util.hpp"
+
+#include "profiler/profile_reduce_impl.hpp"
+#include "profiler/data_type_enum.hpp"
+#include "profiler_operation_registry.hpp"
+
+using namespace std;
+
+using ck::ReduceTensorOp;
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"reduceDims", required_argument, nullptr, 'R'},
+                                       {"reduceOp", required_argument, nullptr, 'O'},
+                                       {"compType", required_argument, nullptr, 'C'},
+                                       {"outType", required_argument, nullptr, 'W'},
+                                       {"nanOpt", required_argument, nullptr, 'N'},
+                                       {"indicesOpt", required_argument, nullptr, 'I'},
+                                       {"scales", required_argument, nullptr, 'S'},
+                                       {"half", no_argument, nullptr, '?'},
+                                       {"double", no_argument, nullptr, '?'},
+                                       {"int8", no_argument, nullptr, '?'},
+                                       {"bf16", no_argument, nullptr, '?'},
+                                       {"dumpout", required_argument, nullptr, 'o'},
+                                       {"verify", required_argument, nullptr, 'v'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+static void check_reduce_dims(const int rank, const std::vector<int>& reduceDims)
+{
+    for(auto dim : reduceDims)
+    {
+        if(dim < 0 || dim >= rank)
+            throw std::runtime_error("Invalid dimension index specified for Reducing");
+    };
+
+    unsigned int flag = 0;
+
+    for(auto dim : reduceDims)
+    {
+        if(flag & (0x1 << dim))
+            throw std::runtime_error("All toReduce dimensions should be different!");
+        flag = flag | (0x1 << dim);
+    };
+};
+
+class ReduceProfilerArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    bool use_half   = false;
+    bool use_double = false;
+    bool use_int8   = false;
+    bool use_bf16   = false;
+
+    std::vector<size_t> inLengths;
+    std::vector<size_t> outLengths;
+    std::vector<int> reduceDims;
+
+    std::vector<float> scales;
+
+    ReduceTensorOp reduceOp     = ReduceTensorOp::ADD;
+    ck::DataTypeEnum compTypeId = ck::DataTypeEnum::Float;
+    ck::DataTypeEnum outTypeId  = ck::DataTypeEnum::Float;
+
+    bool compType_assigned = false;
+    bool outType_assigned  = false;
+
+    int nanOpt           = 0;
+    int indicesOpt       = 0;
+    bool do_verification = false;
+    bool do_dumpout      = false;
+
+    int init_method;
+    bool time_kernel;
+
+    ReduceProfilerArgs()  = default;
+    ~ReduceProfilerArgs() = default;
+
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
+                  << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of to-reduce dimensions"
+                  << std::endl;
+        std::cout << "--reduceOp or -O, enum value indicating the reduction operations"
+                  << std::endl;
+        std::cout << "--compType or -C, enum value indicating the type of accumulated values used "
+                     "during the reduction"
+                  << std::endl;
+        std::cout << "--outType or -W, optional enum value indicating the type of the reduced "
+                     "output, which could be float when the input data is half"
+                  << std::endl;
+        std::cout
+            << "--nanOpt or -N, 1/0 value indicates the selection to use or not use Nan-Propagation"
+            << std::endl;
+        std::cout << "--indicesOpt or -I, 1/0 value indicates the selection to use or not use "
+                     "index in reduction"
+                  << std::endl;
+        std::cout << "--scales or -S, comma separated two float values for alpha and beta"
+                  << std::endl;
+        std::cout << "--half, use fp16 for the input and output tensor data types" << std::endl;
+        std::cout << "--double, use fp64 for the input and output tensor data types" << std::endl;
+        std::cout << "--int8, use int8 for the input and output tensor data types" << std::endl;
+        std::cout << "--bf16, use bfloat16 for the input and output tensor data types" << std::endl;
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
+                     "comparing with the host-based reduction"
+                  << std::endl;
+        std::cout << "--dumpout or -o, 1/0 to indicate where to save the reduction result to files "
+                     "for further analysis"
+                  << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        optind++; // to skip the "reduce" module name
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:O:C:W:N:I:S:v:o:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'O':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceOp = static_cast<ReduceTensorOp>(std::atoi(optarg));
+                break;
+            case 'C':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                compTypeId        = static_cast<ck::DataTypeEnum>(std::atoi(optarg));
+                compType_assigned = true;
+                break;
+            case 'W':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                outTypeId        = static_cast<ck::DataTypeEnum>(std::atoi(optarg));
+                outType_assigned = true;
+                break;
+            case 'N':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                nanOpt = std::atoi(optarg);
+                break;
+            case 'I':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                indicesOpt = std::atoi(optarg);
+                break;
+            case 'S':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                scales = getTypeValuesFromString<float>(optarg);
+
+                if(scales.size() != 2)
+                    throw std::runtime_error("Invalid option format!");
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case 'o':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_dumpout = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "half")
+                    use_half = true;
+                else if(std::string(long_options[option_index].name) == "double")
+                    use_double = true;
+                else if(std::string(long_options[option_index].name) == "int8")
+                    use_int8 = true;
+                else if(std::string(long_options[option_index].name) == "bf16")
+                    use_bf16 = true;
+                else if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+
+            default:
+                show_usage(argv[0]);
+                std::cerr << "Invalid cmd-line options!" << std::endl;
+                return (-1);
+            };
+        };
+
+        if(optind + 2 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        init_method = std::atoi(argv[optind++]);
+        time_kernel = static_cast<bool>(std::atoi(argv[optind]));
+
+        if(scales.empty())
+        {
+            scales.push_back(1.0f);
+            scales.push_back(0.0f);
+        };
+
+        if(reduceOp == ReduceTensorOp::MIN || reduceOp == ReduceTensorOp::MAX ||
+           reduceOp == ReduceTensorOp::AMAX)
+        {
+            // for indexable operations, no need to assign compType and outType, just let them be
+            // same as inType
+            compType_assigned = false;
+            outType_assigned  = false;
+        };
+
+        return (0);
+    };
+
+}; // end of class AppArgs
+
+int profile_reduce(int argc, char* argv[])
+{
+    using ck::DataTypeEnum;
+    using ck::profiler::profile_reduce_impl;
+
+    ReduceProfilerArgs args;
+
+    if(args.processArgs(argc, argv) < 0)
+        return (-1);
+
+    int rank = args.inLengths.size();
+
+    check_reduce_dims(rank, args.reduceDims);
+
+    if(args.reduceOp == ReduceTensorOp::MUL || args.reduceOp == ReduceTensorOp::NORM1)
+        throw std::runtime_error("MUL and NORM1 are not supported by composable kernel!");
+
+    if(args.use_half)
+    {
+        if(!args.compType_assigned)
+            args.compTypeId = DataTypeEnum::Half;
+
+        if(args.outType_assigned &&
+           (args.outTypeId != DataTypeEnum::Half && args.outTypeId != DataTypeEnum::Float))
+            args.outTypeId = DataTypeEnum::Float;
+
+        if(!args.outType_assigned)
+            args.outTypeId = DataTypeEnum::Half;
+
+        if(args.compTypeId == DataTypeEnum::Half)
+        {
+            profile_reduce_impl<ck::half_t, ck::half_t, ck::half_t>(
+                args.do_verification,
+                args.init_method,
+                args.do_dumpout,
+                args.time_kernel,
+                args.inLengths,
+                args.reduceDims,
+                args.reduceOp,
+                static_cast<bool>(args.nanOpt),
+                static_cast<bool>(args.indicesOpt),
+                args.scales[0],
+                args.scales[1]);
+        }
+        else if(args.compTypeId == DataTypeEnum::Float)
+        {
+            profile_reduce_impl<ck::half_t, float, ck::half_t>(args.do_verification,
+                                                               args.init_method,
+                                                               args.do_dumpout,
+                                                               args.time_kernel,
+                                                               args.inLengths,
+                                                               args.reduceDims,
+                                                               args.reduceOp,
+                                                               static_cast<bool>(args.nanOpt),
+                                                               static_cast<bool>(args.indicesOpt),
+                                                               args.scales[0],
+                                                               args.scales[1]);
+        }
+        else
+            throw std::runtime_error("Invalid compType assignment!");
+    }
+    else if(args.use_double)
+    {
+        profile_reduce_impl<double, double, double>(args.do_verification,
+                                                    args.init_method,
+                                                    args.do_dumpout,
+                                                    args.time_kernel,
+                                                    args.inLengths,
+                                                    args.reduceDims,
+                                                    args.reduceOp,
+                                                    static_cast<bool>(args.nanOpt),
+                                                    static_cast<bool>(args.indicesOpt),
+                                                    args.scales[0],
+                                                    args.scales[1]);
+    }
+    else if(args.use_int8)
+    {
+        if(!args.compType_assigned)
+            args.compTypeId = DataTypeEnum::Int8;
+
+        if(args.outType_assigned &&
+           (args.outTypeId != DataTypeEnum::Int8 && args.outTypeId != DataTypeEnum::Int32))
+            args.outTypeId = DataTypeEnum::Int32;
+
+        if(!args.outType_assigned)
+            args.outTypeId = DataTypeEnum::Int8;
+
+        if(args.compTypeId == DataTypeEnum::Int8)
+        {
+            profile_reduce_impl<int8_t, int8_t, int8_t>(args.do_verification,
+                                                        args.init_method,
+                                                        args.do_dumpout,
+                                                        args.time_kernel,
+                                                        args.inLengths,
+                                                        args.reduceDims,
+                                                        args.reduceOp,
+                                                        static_cast<bool>(args.nanOpt),
+                                                        static_cast<bool>(args.indicesOpt),
+                                                        args.scales[0],
+                                                        args.scales[1]);
+        }
+        else if(args.compTypeId == DataTypeEnum::Int32)
+        {
+            profile_reduce_impl<int8_t, int32_t, int8_t>(args.do_verification,
+                                                         args.init_method,
+                                                         args.do_dumpout,
+                                                         args.time_kernel,
+                                                         args.inLengths,
+                                                         args.reduceDims,
+                                                         args.reduceOp,
+                                                         static_cast<bool>(args.nanOpt),
+                                                         static_cast<bool>(args.indicesOpt),
+                                                         args.scales[0],
+                                                         args.scales[1]);
+        }
+        else
+            throw std::runtime_error("Invalid compType assignment!");
+    }
+    else if(args.use_bf16)
+    {
+        if(args.outType_assigned &&
+           (args.outTypeId != DataTypeEnum::BFloat16 && args.outTypeId != DataTypeEnum::Float))
+            args.outTypeId = DataTypeEnum::Float;
+
+        if(!args.outType_assigned)
+            args.outTypeId = DataTypeEnum::BFloat16;
+
+        profile_reduce_impl<ck::bhalf_t, float, ck::bhalf_t>(args.do_verification,
+                                                             args.init_method,
+                                                             args.do_dumpout,
+                                                             args.time_kernel,
+                                                             args.inLengths,
+                                                             args.reduceDims,
+                                                             args.reduceOp,
+                                                             static_cast<bool>(args.nanOpt),
+                                                             static_cast<bool>(args.indicesOpt),
+                                                             args.scales[0],
+                                                             args.scales[1]);
+    }
+    else
+    {
+        if(args.compTypeId == DataTypeEnum::Float)
+        {
+            profile_reduce_impl<float, float, float>(args.do_verification,
+                                                     args.init_method,
+                                                     args.do_dumpout,
+                                                     args.time_kernel,
+                                                     args.inLengths,
+                                                     args.reduceDims,
+                                                     args.reduceOp,
+                                                     static_cast<bool>(args.nanOpt),
+                                                     static_cast<bool>(args.indicesOpt),
+                                                     args.scales[0],
+                                                     args.scales[1]);
+        }
+        else if(args.compTypeId == DataTypeEnum::Double)
+        {
+            profile_reduce_impl<float, double, float>(args.do_verification,
+                                                      args.init_method,
+                                                      args.do_dumpout,
+                                                      args.time_kernel,
+                                                      args.inLengths,
+                                                      args.reduceDims,
+                                                      args.reduceOp,
+                                                      static_cast<bool>(args.nanOpt),
+                                                      static_cast<bool>(args.indicesOpt),
+                                                      args.scales[0],
+                                                      args.scales[1]);
+        }
+        else
+            throw std::runtime_error("Invalid compType assignment!");
+    };
+
+    return (0);
+};
+
+REGISTER_PROFILER_OPERATION("reduce", "Reduce", profile_reduce);
diff --git a/profiler/src/profile_softmax.cpp b/profiler/src/profile_softmax.cpp
new file mode 100644
index 0000000000..dfe8d95c90
--- /dev/null
+++ b/profiler/src/profile_softmax.cpp
@@ -0,0 +1,288 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <unordered_map>
+
+#include "profiler/profile_softmax_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+using ck::profiler::SoftmaxDataType;
+
+struct ArgParser
+{
+    std::unordered_map<std::string, std::vector<int>> long_opts = {
+        {"length", {}}, {"stride", {}}, {"reduce", {}}, {"alpha", {}}, {"beta", {}}};
+
+    bool parse_opt(int argc, char* argv[], const std::string& key, int i)
+    {
+        if(std::string("--") + key == argv[i])
+        {
+            int pos = i;
+            while(++i < argc && argv[i][0] != '-') {}
+            int end = i;
+            for(int j = pos + 1; j < end; j++)
+            {
+                long_opts[key].push_back(std::stoi(argv[j]));
+            }
+            return true;
+        }
+        return false;
+    }
+
+    void operator()(int argc, char* argv[])
+    {
+        for(auto& kv : long_opts)
+        {
+            for(int i = 1; i < argc; i++)
+            {
+                if(parse_opt(argc, argv, kv.first, i))
+                    break;
+            }
+        }
+    }
+};
+
+void print_help()
+{
+    std::cout << "arg1: tensor operation (softmax)\n"
+              << "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
+              << "arg3: verification (0: no; 1: yes)\n"
+              << "arg4: initialization (0: no init; 1: integer value; 2: decimal value)\n"
+              << "arg5: print tensor value (0: no; 1: yes)\n"
+              << "arg6: time kernel (0=n0, 1=yes)\n"
+              << "--length: tensor extents (e.g, --length 8 4 256) \n"
+              << "--stride: tensor strides (e.g, --stride 1024 256 1)\n"
+              << "--reduce: to-reduce dimensions (e.g, --reduce 2)\n"
+              << "--alpha: alpha scaling value\n"
+              << "--beta: beta scaling value\n"
+              << std::endl;
+}
+
+int profile_softmax(int argc, char* argv[])
+{
+    if(argc <= 2)
+    {
+        print_help();
+        return 0;
+    }
+
+    ArgParser arg_parser;
+
+    // short unnamed options
+    const SoftmaxDataType data_type = static_cast<SoftmaxDataType>(std::stoi(argv[2]));
+    const bool do_verification      = std::stoi(argv[3]);
+    const int init_method           = std::stoi(argv[4]);
+    const bool do_log               = std::stoi(argv[5]);
+    const bool time_kernel          = std::stoi(argv[6]);
+
+    // parse the long options
+    arg_parser(argc, argv);
+    const std::vector<index_t> length = arg_parser.long_opts["length"];
+    const std::vector<index_t> stride = arg_parser.long_opts["stride"];
+    const std::vector<index_t> reduce = arg_parser.long_opts["reduce"];
+    const index_t alpha =
+        arg_parser.long_opts["alpha"].empty() ? 1 : arg_parser.long_opts["alpha"][0];
+    const index_t beta = arg_parser.long_opts["beta"].empty() ? 0 : arg_parser.long_opts["beta"][0];
+
+    // Rank 3
+    if(length.size() == 3)
+    {
+        if(data_type == SoftmaxDataType::F16_F16)
+        {
+            if(reduce.size() == 1)
+                ck::profiler::profile_softmax_impl<ck::half_t, float, ck::half_t, 3, 1>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    length,
+                    stride,
+                    reduce,
+                    double(alpha),
+                    double(beta));
+            else if(reduce.size() == 2)
+                ck::profiler::profile_softmax_impl<ck::half_t, float, ck::half_t, 3, 2>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    length,
+                    stride,
+                    reduce,
+                    double(alpha),
+                    double(beta));
+            else if(reduce.size() == 3)
+                ck::profiler::profile_softmax_impl<ck::half_t, float, ck::half_t, 3, 3>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    length,
+                    stride,
+                    reduce,
+                    double(alpha),
+                    double(beta));
+            else
+                throw std::runtime_error("invalid number of dimensions to reduce");
+        }
+        else if(data_type == SoftmaxDataType::F32_F32)
+        {
+            if(reduce.size() == 1)
+                ck::profiler::profile_softmax_impl<float, float, float, 3, 1>(do_verification,
+                                                                              init_method,
+                                                                              do_log,
+                                                                              time_kernel,
+                                                                              length,
+                                                                              stride,
+                                                                              reduce,
+                                                                              double(alpha),
+                                                                              double(beta));
+            else if(reduce.size() == 2)
+                ck::profiler::profile_softmax_impl<float, float, float, 3, 2>(do_verification,
+                                                                              init_method,
+                                                                              do_log,
+                                                                              time_kernel,
+                                                                              length,
+                                                                              stride,
+                                                                              reduce,
+                                                                              double(alpha),
+                                                                              double(beta));
+            else if(reduce.size() == 3)
+                ck::profiler::profile_softmax_impl<float, float, float, 3, 3>(do_verification,
+                                                                              init_method,
+                                                                              do_log,
+                                                                              time_kernel,
+                                                                              length,
+                                                                              stride,
+                                                                              reduce,
+                                                                              double(alpha),
+                                                                              double(beta));
+            else
+                throw std::runtime_error("invalid number of dimensions to reduce");
+        }
+        else
+        {
+            throw std::runtime_error("not implemented yet");
+        }
+    }
+    // Rank 4
+    else if(length.size() == 4)
+    {
+        if(data_type == SoftmaxDataType::F16_F16)
+        {
+            if(reduce.size() == 1)
+                ck::profiler::profile_softmax_impl<ck::half_t, float, ck::half_t, 4, 1>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    length,
+                    stride,
+                    reduce,
+                    double(alpha),
+                    double(beta));
+            else if(reduce.size() == 2)
+                ck::profiler::profile_softmax_impl<ck::half_t, float, ck::half_t, 4, 2>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    length,
+                    stride,
+                    reduce,
+                    double(alpha),
+                    double(beta));
+            else if(reduce.size() == 3)
+                ck::profiler::profile_softmax_impl<ck::half_t, float, ck::half_t, 4, 3>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    length,
+                    stride,
+                    reduce,
+                    double(alpha),
+                    double(beta));
+            else if(reduce.size() == 4)
+                ck::profiler::profile_softmax_impl<ck::half_t, float, ck::half_t, 4, 4>(
+                    do_verification,
+                    init_method,
+                    do_log,
+                    time_kernel,
+                    length,
+                    stride,
+                    reduce,
+                    double(alpha),
+                    double(beta));
+            else
+                throw std::runtime_error("invalid number of dimensions to reduce");
+        }
+        else if(data_type == SoftmaxDataType::F32_F32)
+        {
+            if(reduce.size() == 1)
+                ck::profiler::profile_softmax_impl<float, float, float, 4, 1>(do_verification,
+                                                                              init_method,
+                                                                              do_log,
+                                                                              time_kernel,
+                                                                              length,
+                                                                              stride,
+                                                                              reduce,
+                                                                              double(alpha),
+                                                                              double(beta));
+            else if(reduce.size() == 2)
+                ck::profiler::profile_softmax_impl<float, float, float, 4, 2>(do_verification,
+                                                                              init_method,
+                                                                              do_log,
+                                                                              time_kernel,
+                                                                              length,
+                                                                              stride,
+                                                                              reduce,
+                                                                              double(alpha),
+                                                                              double(beta));
+            else if(reduce.size() == 3)
+                ck::profiler::profile_softmax_impl<float, float, float, 4, 3>(do_verification,
+                                                                              init_method,
+                                                                              do_log,
+                                                                              time_kernel,
+                                                                              length,
+                                                                              stride,
+                                                                              reduce,
+                                                                              double(alpha),
+                                                                              double(beta));
+            else if(reduce.size() == 4)
+                ck::profiler::profile_softmax_impl<float, float, float, 4, 4>(do_verification,
+                                                                              init_method,
+                                                                              do_log,
+                                                                              time_kernel,
+                                                                              length,
+                                                                              stride,
+                                                                              reduce,
+                                                                              double(alpha),
+                                                                              double(beta));
+            else
+                throw std::runtime_error("invalid number of dimensions to reduce");
+        }
+        else
+        {
+            throw std::runtime_error("not implemented yet");
+        }
+    }
+    else
+    {
+        throw std::runtime_error("not implemented yet");
+    }
+
+    return 0;
+}
+
+// hijack main() for quick debugging
+// int main(int argc, char* argv[])
+// {
+//     profile_normalization(argc, argv);
+//     return 0;
+// }
+
+REGISTER_PROFILER_OPERATION("softmax", "Softmax", profile_softmax);
diff --git a/profiler/src/profiler.cpp b/profiler/src/profiler.cpp
new file mode 100644
index 0000000000..0f528c008f
--- /dev/null
+++ b/profiler/src/profiler.cpp
@@ -0,0 +1,30 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+
+#include "profiler_operation_registry.hpp"
+
+static void print_helper_message()
+{
+    std::cout << "arg1: tensor operation " << ProfilerOperationRegistry::GetInstance() << std::endl;
+}
+
+int main(int argc, char* argv[])
+{
+    if(argc == 1)
+    {
+        print_helper_message();
+    }
+    else if(const auto operation = ProfilerOperationRegistry::GetInstance().Get(argv[1]);
+            operation.has_value())
+    {
+        return (*operation)(argc, argv);
+    }
+    else
+    {
+        std::cerr << "cannot find operation: " << argv[1] << std::endl;
+        return EXIT_FAILURE;
+    }
+}
diff --git a/profiler/src/profiler_operation_registry.hpp b/profiler/src/profiler_operation_registry.hpp
new file mode 100644
index 0000000000..276b7b38dc
--- /dev/null
+++ b/profiler/src/profiler_operation_registry.hpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <map>
+#include <optional>
+#include <string_view>
+#include <utility>
+
+class ProfilerOperationRegistry final
+{
+    ProfilerOperationRegistry()  = default;
+    ~ProfilerOperationRegistry() = default;
+
+    public:
+    using Operation = std::function<int(int, char*[])>;
+
+    private:
+    struct Entry final
+    {
+        explicit Entry(std::string_view description, Operation operation) noexcept
+            : description_(description), operation_(std::move(operation))
+        {
+        }
+
+        std::string_view description_;
+        Operation operation_;
+    };
+
+    std::map<std::string_view, Entry> entries_;
+
+    friend std::ostream& operator<<(std::ostream& stream, const ProfilerOperationRegistry& registry)
+    {
+        stream << "{\n";
+        for(auto& [name, entry] : registry.entries_)
+        {
+            stream << "\t" << name << ": " << entry.description_ << "\n";
+        }
+        stream << "}";
+
+        return stream;
+    }
+
+    public:
+    static ProfilerOperationRegistry& GetInstance()
+    {
+        static ProfilerOperationRegistry registry;
+        return registry;
+    }
+
+    std::optional<Operation> Get(std::string_view name) const
+    {
+        const auto found = entries_.find(name);
+        if(found == end(entries_))
+        {
+            return std::nullopt;
+        }
+
+        return (found->second).operation_;
+    }
+
+    bool Add(std::string_view name, std::string_view description, Operation operation)
+    {
+        return entries_
+            .emplace(std::piecewise_construct,
+                     std::forward_as_tuple(name),
+                     std::forward_as_tuple(description, std::move(operation)))
+            .second;
+    }
+};
+
+#define PP_CONCAT(x, y) PP_CONCAT_IMPL(x, y)
+#define PP_CONCAT_IMPL(x, y) x##y
+
+#define REGISTER_PROFILER_OPERATION(name, description, operation)              \
+    static const bool PP_CONCAT(operation_registration_result_, __COUNTER__) = \
+        ::ProfilerOperationRegistry::GetInstance().Add(name, description, operation)
diff --git a/rbuild.ini b/rbuild.ini
new file mode 100644
index 0000000000..3649cedf0a
--- /dev/null
+++ b/rbuild.ini
@@ -0,0 +1,8 @@
+[develop]
+cxx = ${rocm_path}/bin/hipcc
+cc = ${rocm_path}/llvm/bin/clang
+ignore = pcre
+deps =
+    -f dev-requirements.txt
+define =
+    BUILD_DEV=On
diff --git a/requirements.txt b/requirements.txt
new file mode 100644
index 0000000000..8b13789179
--- /dev/null
+++ b/requirements.txt
@@ -0,0 +1 @@
+
diff --git a/script/check_copyright_year.sh b/script/check_copyright_year.sh
new file mode 100755
index 0000000000..f7709472ef
--- /dev/null
+++ b/script/check_copyright_year.sh
@@ -0,0 +1,17 @@
+#!/bin/bash
+
+current_year=$(date +%Y)
+exit_code=0
+
+for file in $@; do
+    if grep -q "Copyright (c)" $file
+    then
+        if ! grep -q "Copyright (c).*$current_year" $file
+        then
+            echo "ERROR: File $file has a copyright notice without the current year ($current_year)."
+            exit_code=1
+        fi
+    fi
+done
+
+exit $exit_code
diff --git a/script/clang-format-overwrite.sh b/script/clang-format-overwrite.sh
new file mode 100755
index 0000000000..da83254f00
--- /dev/null
+++ b/script/clang-format-overwrite.sh
@@ -0,0 +1,2 @@
+#find . -name deps -prune -o -name build -prune -o -iname '*.h' -o -iname '*.hpp' -o -iname '*.cpp' -o -iname '*.h.in' -o -iname '*.hpp.in' -o -iname '*.cpp.in' -o -iname '*.cl' -o -iname '*.cuh' -o -iname '*.cu' -o -iname '*.inc' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-12 -i -style=file {}'
+git status --porcelain | awk '$1 != "D" && (match($2, "\\.cpp|hpp|inc")) {print $2}' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-12 -i -style=file {}'
diff --git a/script/cmake-ck-dev.sh b/script/cmake-ck-dev.sh
new file mode 100755
index 0000000000..da9e92b862
--- /dev/null
+++ b/script/cmake-ck-dev.sh
@@ -0,0 +1,18 @@
+#!/bin/bash
+rm -f CMakeCache.txt
+rm -f *.cmake
+rm -rf CMakeFiles
+
+MY_PROJECT_SOURCE=$1
+
+cmake                                                                                             \
+-D CMAKE_PREFIX_PATH=/opt/rocm                                                                    \
+-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                         \
+-D CMAKE_CXX_FLAGS="-std=c++17 -O3 -ftemplate-backtrace-limit=0  -fPIE  -Wno-gnu-line-marker      \
+-save-temps=$PWD"                                                                                 \
+-D CMAKE_BUILD_TYPE=Release                                                                       \
+-D BUILD_DEV=ON                                                                                   \
+-D GPU_TARGETS="gfx908;gfx90a;gfx940"                                                             \
+-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON                                                                 \
+-D USE_BITINT_EXTENSION_INT4=OFF                                                                  \
+${MY_PROJECT_SOURCE}
diff --git a/script/cmake-ck-release.sh b/script/cmake-ck-release.sh
new file mode 100755
index 0000000000..787eabbf96
--- /dev/null
+++ b/script/cmake-ck-release.sh
@@ -0,0 +1,18 @@
+#!/bin/bash
+rm -f CMakeCache.txt
+rm -f *.cmake
+rm -rf CMakeFiles
+
+MY_PROJECT_SOURCE=$1
+
+cmake                                                                                             \
+-D CMAKE_PREFIX_PATH=/opt/rocm                                                                    \
+-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                                                         \
+-D CMAKE_CXX_FLAGS="-O3"                                                                          \
+-D CMAKE_BUILD_TYPE=Release                                                                       \
+-D BUILD_DEV=OFF                                                                                  \
+-D GPU_TARGETS="gfx908;gfx90a;gfx940"                                                             \
+-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON                                                                 \
+-D USE_BITINT_EXTENSION_INT4=OFF                                                                  \
+${MY_PROJECT_SOURCE}
+
diff --git a/script/count_vgpr.sh b/script/count_vgpr.sh
new file mode 100755
index 0000000000..07debc53a8
--- /dev/null
+++ b/script/count_vgpr.sh
@@ -0,0 +1,20 @@
+#!/bin/bash
+FILE=$1
+
+for num in {0..255}
+do
+    base_pattern="(\[?${num}\b|\[\d*:${num}\])"
+    spattern="s${base_pattern}"
+    vpattern="v${base_pattern}"
+    apattern="a${base_pattern}"
+    scount=$(grep -P $spattern $FILE | wc -l)
+    vcount=$(grep -P $vpattern $FILE | wc -l)
+    acount=$(grep -P $apattern $FILE | wc -l)
+    bash -c "echo -n v${num}   $vcount && \
+             echo -n , s${num} $scount && \
+             echo -n , a${num} $acount"
+    if [[ $scount -ne 0 || $vcount -ne 0 || $acount -ne 0 ]]; then
+        echo -n "  *"
+    fi
+    echo ""
+done
diff --git a/script/hipclang_opt.sh b/script/hipclang_opt.sh
new file mode 100755
index 0000000000..c51bd51d97
--- /dev/null
+++ b/script/hipclang_opt.sh
@@ -0,0 +1,25 @@
+rm *.ll *.s
+
+BC_FILE=$1
+
+/opt/rocm/llvm/bin/llvm-dis $BC_FILE -o original.ll
+/opt/rocm/llvm/bin/opt -S -inline -inline-threshold=104857 original.ll > inline.ll
+/opt/rocm/llvm/bin/opt -S -sroa inline.ll > sroa.ll
+/opt/rocm/llvm/bin/opt -S -O3 sroa.ll > o3.ll
+
+/opt/rocm/llvm/bin/llc -mcpu=gfx906 original.ll
+/opt/rocm/llvm/bin/llc -mcpu=gfx906 inline.ll
+/opt/rocm/llvm/bin/llc -mcpu=gfx906 sroa.ll
+/opt/rocm/llvm/bin/llc -mcpu=gfx906 o3.ll
+
+#/opt/rocm/llvm/bin/opt -S -O3 -sroa inline.ll > o3.ll
+#/opt/rocm/llvm/bin/opt -S -O3 -sroa o3.ll > o3_2.ll
+#/opt/rocm/llvm/bin/opt -S -O3 -sroa o3_2.ll > o3_3.ll
+#/opt/rocm/llvm/bin/opt -S -O3 -sroa o3_3.ll > o3_4.ll
+
+#/opt/rocm/llvm/bin/llc -mcpu=gfx908 opt.ll
+#/opt/rocm/llvm/bin/llc -mcpu=gfx908 inline.ll
+#/opt/rocm/llvm/bin/llc -mcpu=gfx908 o3.ll
+#/opt/rocm/llvm/bin/llc -mcpu=gfx908 o3_2.ll
+#/opt/rocm/llvm/bin/llc -mcpu=gfx908 o3_3.ll
+#/opt/rocm/llvm/bin/llc -mcpu=gfx908 o3_4.ll
diff --git a/script/install_precommit.sh b/script/install_precommit.sh
new file mode 100755
index 0000000000..296280bb03
--- /dev/null
+++ b/script/install_precommit.sh
@@ -0,0 +1,20 @@
+#!/bin/bash
+
+run_and_check() {
+    "$@"
+    status=$?
+    if [ $status -ne 0 ]; then
+        echo "Error with \"$@\": Exited with status $status"
+        exit $status
+    fi
+    return $status
+}
+
+echo "I: Installing tools required for pre-commit checks..."
+run_and_check apt install clang-format-12
+
+echo "I: Installing pre-commit itself..."
+run_and_check pip3 install pre-commit
+run_and_check pre-commit install
+
+echo "I: Installation successful."
diff --git a/script/parse_perf_data.py b/script/parse_perf_data.py
new file mode 100644
index 0000000000..4cb13e6243
--- /dev/null
+++ b/script/parse_perf_data.py
@@ -0,0 +1,290 @@
+#!/usr/bin/env python3
+import os, io, argparse, datetime, re
+import numpy as np
+import sqlalchemy
+from sqlalchemy.types import NVARCHAR, Float, Integer
+import pymysql
+import pandas as pd
+from sshtunnel import SSHTunnelForwarder
+
+def print_to_string(*args, **kwargs):
+    output = io.StringIO()
+    print(*args, file=output, **kwargs)
+    contents = output.getvalue()
+    output.close()
+    return contents
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Parse results from tf benchmark runs')
+    parser.add_argument('filename', type=str, help='Log file to prase or directory containing log files')
+    args = parser.parse_args()
+    files = []
+    if os.path.isdir(args.filename):
+        all_files = os.listdir(args.filename)
+        for name in all_files:
+            if not 'log' in name:
+                continue
+            files.append(os.path.join(args.filename, name))
+    else:
+        files = [args.filename]
+    args.files = files
+    return args
+
+def main():
+    args = parse_args()
+    tests = []
+    kernels=[]
+    tflops=[]
+    dtype=[]
+    alayout=[]
+    blayout=[]
+    M=[]
+    N=[]
+    K=[]
+    StrideA=[]
+    StrideB=[]
+    StrideC=[]
+    #parse results, get the Tflops value for "Best Perf" kernels
+
+    glue=""
+    for filename in args.files:
+        for line in open(filename):
+            if 'Branch name' in line:
+                lst=line.split()
+                branch_name=lst[2]
+            if 'On branch' in line:
+                lst=line.split()
+                branch_name=lst[2]
+            if 'Node name' in line:
+                lst=line.split()
+                node_id=lst[2]
+            if 'GPU_arch' in line:
+                lst=line.split()
+                gpu_arch=lst[2]
+            if 'HIP version' in line:
+                lst=line.split()
+                hip_vers=lst[2]
+            if 'Compute Unit' in line:
+                lst=line.split()
+                compute_units=lst[2]
+            if 'InstalledDir' in line:
+                lst=line.split()
+                rocm_vers=lst[1][lst[1].find('/opt/rocm-')+len('/opt/rocm-'):lst[1].rfind('/llvm/bin')]
+    print("Branch name:",branch_name)
+    print("Node name:",node_id)
+    print("GPU_arch:",gpu_arch)
+    print("Compute units:",compute_units)
+    print("ROCM_version:",rocm_vers)
+    print("HIP_version:",hip_vers)
+
+
+    #parse gemm performance tests:
+    if 'gemm' in filename:
+        for filename in args.files:
+            for line in open(filename):
+                if 'Best Perf' in line:
+                    lst=line.split()
+                    if len(lst)>=37: #the line is complete
+                        tests.append(glue.join(lst[5:30]))
+                        kernels.append(glue.join(lst[37:]))
+                        tflops.append(lst[33])
+                        dtype.append(lst[5])
+                        alayout.append(lst[8])
+                        blayout.append(lst[11])
+                        M.append(lst[14])
+                        N.append(lst[17])
+                        K.append(lst[20])
+                        StrideA.append(lst[23])
+                        StrideB.append(lst[26])
+                        StrideC.append(lst[29])
+                    elif len(lst)<37 and len(lst)>=33: #the tflops are available
+                        tests.append(glue.join(lst[5:30]))
+                        kernels.append("N/A")
+                        tflops.append(lst[33])
+                        dtype.append(lst[5])
+                        alayout.append(lst[8])
+                        blayout.append(lst[11])
+                        M.append(lst[14])
+                        N.append(lst[17])
+                        K.append(lst[20])
+                        StrideA.append(lst[23])
+                        StrideB.append(lst[26])
+                        StrideC.append(lst[29])
+                        print("warning: incomplete line:",lst)
+                    elif len(lst)<33: #even the tflops are not available
+                        print("Error in ckProfiler output!")
+                        print("warning: incomplete line=",lst)
+        #sort results
+        #sorted_tests = sorted(tests)
+        #print("sorted tests:",sorted_tests)
+        sorted_tflops = [x for _,x in sorted(zip(tests,tflops))]
+        #sorted_kernels = [x for _,x in sorted(zip(tests,kernels))]
+        test_list=list(range(1,len(tests)+1))
+
+    #parse resnet50 performance tests:
+    if 'resnet50' in filename:
+        for filename in args.files:
+            for line in open(filename):
+                if 'Best Perf' in line:
+                    lst=line.split()
+                    tflops.append(lst[4])
+
+    print("Number of tests:",len(tflops))
+    sql_hostname = '127.0.0.1'
+    sql_username = os.environ["dbuser"]
+    sql_password = os.environ["dbpassword"]
+    sql_main_database = 'miopen_perf'
+    sql_port = 3306
+    ssh_host = os.environ["dbsship"]
+    ssh_user = os.environ["dbsshuser"]
+    ssh_port = int(os.environ["dbsshport"])
+    ssh_pass = os.environ["dbsshpassword"]
+
+    with SSHTunnelForwarder(
+            (ssh_host, ssh_port),
+            ssh_username=ssh_user,
+            ssh_password=ssh_pass,
+            remote_bind_address=(sql_hostname, sql_port)) as tunnel:
+
+        sqlEngine = sqlalchemy.create_engine('mysql+pymysql://{0}:{1}@{2}:{3}/{4}'.
+            format(sql_username, sql_password, sql_hostname, tunnel.local_bind_port, sql_main_database))
+        conn = sqlEngine.connect()
+
+        #save gemm performance tests:
+        if 'gemm' in filename:
+
+            #write the ck_gemm_test_params table
+            #only needed once the test set changes
+            '''
+            sorted_dtypes = [x for _,x in sorted(zip(tests,dtype))]
+            sorted_alayout = [x for _,x in sorted(zip(tests,alayout))]
+            sorted_blayout = [x for _,x in sorted(zip(tests,blayout))]
+            sorted_M = [x for _,x in sorted(zip(tests,M))]
+            sorted_N = [x for _,x in sorted(zip(tests,N))]
+            sorted_K = [x for _,x in sorted(zip(tests,K))]
+            sorted_StrideA = [x for _,x in sorted(zip(tests,StrideA))]
+            sorted_StrideB = [x for _,x in sorted(zip(tests,StrideB))]
+            sorted_StrideC = [x for _,x in sorted(zip(tests,StrideC))]
+            ck_gemm_params=[test_list,sorted_dtypes,sorted_alayout,sorted_blayout,
+                        sorted_M,sorted_N,sorted_K,sorted_StrideA,sorted_StrideB,
+                        sorted_StrideC]
+            df=pd.DataFrame(np.transpose(ck_gemm_params),columns=['Test_number','Data_type',
+                'Alayout','BLayout','M','N','K', 'StrideA','StrideB','StrideC'])
+            print(df)
+
+            dtypes = {
+                'Test_number': Integer(),
+                'Data_type': NVARCHAR(length=5),
+                'Alayout': NVARCHAR(length=12),
+                'Blayout': NVARCHAR(length=12),
+                'M': Integer(),
+                'N': Integer(),
+                'K': Integer(),
+                'StrideA': Integer(),
+                'StrideB': Integer(),
+                'StrideC': Integer()
+                }
+            df.to_sql("ck_gemm_test_params",conn,if_exists='replace',index=False, dtype=dtypes)
+            '''
+
+            #read baseline results for the latest develop branch
+            query = '''SELECT * from ck_gemm_tflops WHERE Datetime = (SELECT MAX(Datetime) FROM ck_gemm_tflops where Branch_ID='develop' );'''
+            tflops_base = pd.read_sql_query(query, conn)
+
+            #write new results to the db
+            testlist=[]
+            for i in range(1,len(tests)+1):
+                testlist.append("Test%i"%i)
+            ck_gemm_tflops=[str(branch_name),str(node_id),str(gpu_arch),compute_units,str(rocm_vers),str(hip_vers),str(datetime.datetime.now())]
+            flops=pd.DataFrame(data=[ck_gemm_tflops],columns=['Branch_ID','Node_ID','GPU_arch','Compute Units','ROCM_version','HIP_version','Datetime'])
+            df_add=pd.DataFrame(data=[sorted_tflops],columns=testlist)
+            flops=pd.concat([flops,df_add],axis=1)
+            print("new tflops for gemm tests:",flops)
+            flops.to_sql("ck_gemm_tflops",conn,if_exists='append',index=False)
+
+        #save resnet50 performance tests:
+        if 'resnet50' in filename:
+            #read baseline results for the latest develop branch
+            query = '''SELECT * from ck_resnet50_N256_tflops WHERE Datetime = (SELECT MAX(Datetime) FROM ck_resnet50_N256_tflops where Branch_ID='develop' );'''
+            tflops_base_N256 = pd.read_sql_query(query, conn)
+            query = '''SELECT * from ck_resnet50_N4_tflops WHERE Datetime = (SELECT MAX(Datetime) FROM ck_resnet50_N4_tflops where Branch_ID='develop' );'''
+            tflops_base_N4 = pd.read_sql_query(query, conn)
+
+            #write new results to the db
+            testlist=[]
+            for i in range(1,50):
+                testlist.append("Layer%i"%i)
+            ck_resnet_tflops=[str(branch_name),str(node_id),str(gpu_arch),compute_units,str(rocm_vers),str(hip_vers),str(datetime.datetime.now())]
+            flops0=pd.DataFrame(data=[ck_resnet_tflops],columns=['Branch_ID','Node_ID','GPU_arch','Compute Units','ROCM_version','HIP_version','Datetime'])
+            df_add=pd.DataFrame(data=[tflops[0:49]],columns=testlist)
+            flops=pd.concat([flops0,df_add],axis=1)
+            print("new tflops for N=256 resnet50 test:",flops)
+            flops.to_sql("ck_resnet50_N256_tflops",conn,if_exists='append',index=False)
+            df_add=pd.DataFrame(data=[tflops[49:98]],columns=testlist)
+            flops=pd.concat([flops0,df_add],axis=1)
+            print("new tflops for N=4 resnet50 test:",flops)
+            flops.to_sql("ck_resnet50_N4_tflops",conn,if_exists='append',index=False)
+
+        conn.close()
+
+    #compare the results to the baseline if baseline exists
+    regression=0
+    if 'gemm' in filename:
+        if not tflops_base.empty:
+            base=tflops_base[testlist].to_numpy(dtype='float')
+            base_list=base[0]
+            ave_perf=0
+            for i in range(len(base_list)):
+                # success criterion:
+                if base_list[i]>1.01*float(sorted_tflops[i]):
+                    print("test # ",i,"shows regression by {:.3f}%".format(
+                        (float(sorted_tflops[i])-base_list[i])/base_list[i]*100))
+                    regression=1
+                ave_perf=ave_perf+float(sorted_tflops[i])/base_list[i]
+            if regression==0:
+                print("no regressions found")
+            ave_perf=ave_perf/len(base_list)
+            print("average performance relative to baseline:",ave_perf)
+        else:
+            print("could not find a baseline")
+    if 'resnet50' in filename:
+        if not tflops_base_N256.empty:
+            base=tflops_base_N256[testlist].to_numpy(dtype='float')
+            base_list=base[0]
+            ave_perf=0
+            for i in range(len(base_list)):
+                # success criterion:
+                if base_list[i]>1.01*float(tflops[i]):
+                    print("layer # ",i,"shows regression by {:.3f}%".format(
+                        (float(tflops[i])-base_list[i])/base_list[i]*100))
+                    regression=1
+                ave_perf=ave_perf+float(tflops[i])/base_list[i]
+            if regression==0:
+                print("no regressions found")
+            ave_perf=ave_perf/len(base_list)
+            print("average performance relative to baseline:",ave_perf)
+        else:
+            print("could not find a baseline for N=256")
+        if not tflops_base_N4.empty:
+            base=tflops_base_N4[testlist].to_numpy(dtype='float')
+            base_list=base[0]
+            ave_perf=0
+            for i in range(len(base_list)):
+                # success criterion:
+                if base_list[i]>1.01*float(tflops[i+49]):
+                    print("layer # ",i,"shows regression by {:.3f}%".format(
+                        (float(tflops[i+49])-base_list[i])/base_list[i]*100))
+                    regression=1
+                ave_perf=ave_perf+float(tflops[i+49])/base_list[i]
+            if regression==0:
+                print("no regressions found")
+            ave_perf=ave_perf/len(base_list)
+            print("average performance relative to baseline:",ave_perf)
+        else:
+            print("could not find a baseline for N=4")
+
+    #return 0 if performance criteria met, otherwise return 1
+    return regression
+
+if __name__ == '__main__':
+    main()
\ No newline at end of file
diff --git a/script/process_perf_data.py b/script/process_perf_data.py
new file mode 100644
index 0000000000..e8b8e1458c
--- /dev/null
+++ b/script/process_perf_data.py
@@ -0,0 +1,309 @@
+#!/usr/bin/env python3
+import os, io, argparse, datetime
+#import numpy as np
+import sqlalchemy
+from sqlalchemy.types import NVARCHAR, Float, Integer
+from sqlalchemy import text
+import pymysql
+import pandas as pd
+from sshtunnel import SSHTunnelForwarder
+
+def print_to_string(*args, **kwargs):
+    output = io.StringIO()
+    print(*args, file=output, **kwargs)
+    contents = output.getvalue()
+    output.close()
+    return contents
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Parse results from tf benchmark runs')
+    parser.add_argument('filename', type=str, help='Log file to prase or directory containing log files')
+    args = parser.parse_args()
+    files = []
+    if os.path.isdir(args.filename):
+        all_files = os.listdir(args.filename)
+        for name in all_files:
+            if not 'log' in name:
+                continue
+            files.append(os.path.join(args.filename, name))
+    else:
+        files = [args.filename]
+    args.files = files
+    return args
+
+def get_log_params(logfile):
+    print("logfile=",logfile)
+    branch_name=' '
+    node_id=' '
+    gpu_arch=' '
+    hip_vers=' '
+    compute_units=0
+    environment=' '
+    rocm_vers=' '
+    for line in open(logfile):
+        if 'Branch name' in line:
+            lst=line.split()
+            branch_name=lst[2]
+        if 'On branch' in line:
+            lst=line.split()
+            branch_name=lst[2]
+        if 'Node name' in line:
+            lst=line.split()
+            node_id=lst[2]
+        if 'GPU_arch' in line:
+            lst=line.split()
+            gpu_arch=lst[2]
+        if 'HIP version' in line:
+            lst=line.split()
+            hip_vers=lst[2]
+        if 'Compute Unit' in line:
+            lst=line.split()
+            compute_units=lst[2]
+        if 'Environment type' in line:
+            lst=line.split()
+            environment=lst[2]
+        if 'InstalledDir' in line:
+            lst=line.split()
+            rocm_vers=lst[1][lst[1].find('/opt/rocm-')+len('/opt/rocm-'):lst[1].rfind('/llvm/bin')]
+    return branch_name, node_id, gpu_arch, compute_units, rocm_vers, hip_vers, environment
+
+def parse_logfile(logfile):
+    glue=''
+    res=[]
+    tests=[]
+    kernels=[]
+    tflops=[]
+    dtype=[]
+    alayout=[]
+    blayout=[]
+    M=[]
+    N=[]
+    K=[]
+    StrideA=[]
+    StrideB=[]
+    StrideC=[]
+    if 'perf_gemm.log' in logfile:
+        for line in open(logfile):
+            if 'Best Perf' in line:
+                lst=line.split()
+                if len(lst)>=37: #the line is complete
+                    tests.append(glue.join(lst[5:30]))
+                    kernels.append(glue.join(lst[37:]))
+                    tflops.append(lst[33])
+                    dtype.append(lst[5])
+                    alayout.append(lst[8])
+                    blayout.append(lst[11])
+                    M.append(lst[14])
+                    N.append(lst[17])
+                    K.append(lst[20])
+                    StrideA.append(lst[23])
+                    StrideB.append(lst[26])
+                    StrideC.append(lst[29])
+                elif len(lst)<37 and len(lst)>=33: #the tflops are available
+                    tests.append(glue.join(lst[5:30]))
+                    kernels.append("N/A")
+                    tflops.append(lst[33])
+                    dtype.append(lst[5])
+                    alayout.append(lst[8])
+                    blayout.append(lst[11])
+                    M.append(lst[14])
+                    N.append(lst[17])
+                    K.append(lst[20])
+                    StrideA.append(lst[23])
+                    StrideB.append(lst[26])
+                    StrideC.append(lst[29])
+                    print("warning: incomplete line:",lst)
+                elif len(lst)<33: #even the tflops are not available
+                    print("Error in ckProfiler output!")
+                    print("warning: incomplete line=",lst)
+        #sort results
+        #sorted_tests = sorted(tests)
+        res = [x for _,x in sorted(zip(tests,tflops))]
+        #sorted_kernels = [x for _,x in sorted(zip(tests,kernels))]
+        test_list=list(range(1,len(tests)+1))
+    #parse conv_fwd and conv_bwd performance tests:
+    elif 'conv_fwd' in logfile or 'conv_bwd_data' in logfile:
+        for line in open(logfile):
+            if 'tflops:' in line:
+                lst=line.split()
+                res.append(lst[1])
+    #parse all other performance tests:
+    elif 'resnet50' in logfile or 'batched_gemm' in logfile or 'grouped_gemm' in logfile  or 'gemm_bilinear' in logfile or 'reduction' in logfile:
+        for line in open(logfile):
+            if 'Best Perf' in line:
+                lst=line.split()
+                res.append(lst[4])
+    elif 'onnx_gemm' in logfile or 'splitK_gemm' in logfile:
+        for line in open(logfile):
+            if 'Best Perf' in line:
+                lst=line.split()
+                res.append(lst[33])
+    return res
+
+
+def get_baseline(table, connection):
+    query = text('''SELECT * from '''+table+''' WHERE Datetime = (SELECT MAX(Datetime) FROM '''+table+''' where Branch_ID='develop' );''')
+    return pd.read_sql(query, connection)
+
+def store_new_test_result(table_name, test_results, testlist, branch_name, node_id, gpu_arch, compute_units, rocm_vers, hip_vers, environment, connection):
+    params=[str(branch_name),str(node_id),str(gpu_arch),compute_units,str(rocm_vers),str(hip_vers),str(environment),str(datetime.datetime.now())]
+    df=pd.DataFrame(data=[params],columns=['Branch_ID','Node_ID','GPU_arch','Compute Units','ROCM_version','HIP_version','Environment','Datetime'])
+    df_add=pd.DataFrame(data=[test_results],columns=testlist)
+    df=pd.concat([df,df_add],axis=1)
+    #print("new test results dataframe:",df)
+    df.to_sql(table_name,connection,if_exists='append',index=False)
+    return 0
+
+def compare_test_to_baseline(baseline,test,testlist):
+    regression=0
+    if not baseline.empty:
+        base=baseline[testlist].to_numpy(dtype='float')
+        base_list=base[0]
+        ave_perf=0
+        for i in range(len(base_list)):
+            # success criterion:
+            if base_list[i]>1.01*float(test[i]):
+                print("test # ",i,"shows regression by {:.3f}%".format(
+                    (float(test[i])-base_list[i])/base_list[i]*100))
+                regression=1
+            if base_list[i]>0: ave_perf=ave_perf+float(test[i])/base_list[i]
+        if regression==0:
+            print("no regressions found")
+        ave_perf=ave_perf/len(base_list)
+        print("average performance relative to baseline:",ave_perf)
+    else:
+        print("could not find a baseline")
+    return regression
+
+'''
+def post_test_params(tlist,connection):
+    sorted_dtypes = [x for _,x in sorted(zip(tests,dtype))]
+    sorted_alayout = [x for _,x in sorted(zip(tests,alayout))]
+    sorted_blayout = [x for _,x in sorted(zip(tests,blayout))]
+    sorted_M = [x for _,x in sorted(zip(tests,M))]
+    sorted_N = [x for _,x in sorted(zip(tests,N))]
+    sorted_K = [x for _,x in sorted(zip(tests,K))]
+    sorted_StrideA = [x for _,x in sorted(zip(tests,StrideA))]
+    sorted_StrideB = [x for _,x in sorted(zip(tests,StrideB))]
+    sorted_StrideC = [x for _,x in sorted(zip(tests,StrideC))]
+    ck_gemm_params=[tlist,sorted_dtypes,sorted_alayout,sorted_blayout,
+                sorted_M,sorted_N,sorted_K,sorted_StrideA,sorted_StrideB,
+                sorted_StrideC]
+    df=pd.DataFrame(np.transpose(ck_gemm_params),columns=['Test_number','Data_type',
+        'Alayout','BLayout','M','N','K', 'StrideA','StrideB','StrideC'])
+    print(df)
+
+    dtypes = {
+        'Test_number': Integer(),
+        'Data_type': NVARCHAR(length=5),
+        'Alayout': NVARCHAR(length=12),
+        'Blayout': NVARCHAR(length=12),
+        'M': Integer(),
+        'N': Integer(),
+        'K': Integer(),
+        'StrideA': Integer(),
+        'StrideB': Integer(),
+        'StrideC': Integer()
+        }
+    df.to_sql("ck_gemm_test_params",connection,if_exists='replace',index=False, dtype=dtypes)
+'''
+
+def main():
+    args = parse_args()
+    results=[]
+    tflops_base=[]
+    testlist=[]
+    #parse the test parameters from the logfile
+    for filename in args.files:
+        branch_name, node_id, gpu_arch, compute_units, rocm_vers, hip_vers, environment = get_log_params(filename)
+
+    print("Branch name:",branch_name)
+    print("Node name:",node_id)
+    print("GPU_arch:",gpu_arch)
+    print("Compute units:",compute_units)
+    print("ROCM_version:",rocm_vers)
+    print("HIP_version:",hip_vers)
+    print("Environment:",environment)
+    #parse results, get the Tflops value for "Best Perf" kernels
+    results=parse_logfile(filename)
+
+    print("Number of tests:",len(results))
+    sql_hostname = '127.0.0.1'
+    sql_username = os.environ["dbuser"]
+    sql_password = os.environ["dbpassword"]
+    sql_main_database = 'miopen_perf'
+    sql_port = 3306
+    ssh_host = os.environ["dbsship"]
+    ssh_user = os.environ["dbsshuser"]
+    ssh_port = int(os.environ["dbsshport"])
+    ssh_pass = os.environ["dbsshpassword"]
+
+    with SSHTunnelForwarder(
+            (ssh_host, ssh_port),
+            ssh_username=ssh_user,
+            ssh_password=ssh_pass,
+            remote_bind_address=(sql_hostname, sql_port)) as tunnel:
+
+        sqlEngine = sqlalchemy.create_engine('mysql+pymysql://{0}:{1}@{2}:{3}/{4}'.
+            format(sql_username, sql_password, sql_hostname, tunnel.local_bind_port, sql_main_database))
+        conn = sqlEngine.connect()
+
+        #save gemm performance tests:
+        if 'perf_gemm.log' in filename:
+            #write the ck_gemm_test_params table only needed once the test set changes
+            #post_test_params(test_list,conn)
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_gemm_tflops"
+        if 'batched_gemm' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_batched_gemm_tflops"
+        if 'grouped_gemm' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_grouped_gemm_tflops"
+        if 'conv_fwd' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_conv_fwd_tflops"
+        if 'conv_bwd_data' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_conv_bwd_data_tflops"
+        if 'gemm_bilinear' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_gemm_bilinear_tflops"
+        if 'reduction' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_reduction_GBps"
+        if 'resnet50_N4' in filename:
+            for i in range(1,50):
+                testlist.append("Layer%i"%i)
+            table_name="ck_resnet50_N4_tflops"
+        if 'resnet50_N256' in filename:
+            for i in range(1,50):
+                testlist.append("Layer%i"%i)
+            table_name="ck_resnet50_N256_tflops"
+        if 'onnx_gemm' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_onnx_gemm_tflops"
+        if 'splitK_gemm' in filename:
+            for i in range(1,len(results)+1):
+                testlist.append("Test%i"%i)
+            table_name="ck_splitK_gemm_tflops"
+
+        tflops_base = get_baseline(table_name,conn)
+        store_new_test_result(table_name, results, testlist, branch_name, node_id, gpu_arch, compute_units, rocm_vers, hip_vers, environment, conn)
+        conn.close()
+
+    #compare the results to the baseline if baseline exists
+    regression=0
+    regression=compare_test_to_baseline(tflops_base,results,testlist)
+    return regression
+
+if __name__ == '__main__':
+    main()
diff --git a/script/process_perf_data.sh b/script/process_perf_data.sh
new file mode 100755
index 0000000000..15fc5cb15f
--- /dev/null
+++ b/script/process_perf_data.sh
@@ -0,0 +1,15 @@
+#!/bin/bash 
+#
+# in order to run this script you'd need the following python packages:
+
+#pip3 install --upgrade pip
+#pip3 install sqlalchemy pymysql pandas sshtunnel
+
+# you would also need to set up some environment variables in order to 
+# post your new test results to the database and compare them to the baseline
+# please contact Illia.Silin@amd.com for more details
+
+#process results
+python3 process_perf_data.py perf_gemm.log
+python3 process_perf_data.py perf_resnet50_N256.log
+python3 process_perf_data.py perf_resnet50_N4.log
diff --git a/script/process_qa_data.sh b/script/process_qa_data.sh
new file mode 100755
index 0000000000..abf1e6234e
--- /dev/null
+++ b/script/process_qa_data.sh
@@ -0,0 +1,23 @@
+#!/bin/bash 
+#
+# in order to run this script you'd need the following python packages:
+
+#pip3 install --upgrade pip
+#pip3 install sqlalchemy pymysql pandas sshtunnel
+
+# you would also need to set up some environment variables in order to 
+# post your new test results to the database and compare them to the baseline
+# please contact Illia.Silin@amd.com for more details
+
+#process results
+python3 process_perf_data.py perf_gemm.log
+python3 process_perf_data.py perf_resnet50_N256.log
+python3 process_perf_data.py perf_resnet50_N4.log
+python3 process_perf_data.py perf_batched_gemm.log
+python3 process_perf_data.py perf_grouped_gemm.log
+python3 process_perf_data.py perf_conv_fwd.log
+python3 process_perf_data.py perf_conv_bwd_data.log
+python3 process_perf_data.py perf_gemm_bilinear.log
+python3 process_perf_data.py perf_reduction.log
+python3 process_perf_data.py perf_splitK_gemm.log
+python3 process_perf_data.py perf_onnx_gemm.log
diff --git a/script/profile_batched_gemm.sh b/script/profile_batched_gemm.sh
new file mode 100755
index 0000000000..f90baaed68
--- /dev/null
+++ b/script/profile_batched_gemm.sh
@@ -0,0 +1,37 @@
+#!/bin/bash
+ 
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+ 
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+ 
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC BatchStrideA BatchStrideB BatchStrideC BatchCount
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  960  1024 1024       -1     -1      -1            -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1920  2048 2048       -1     -1      -1            -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 3840  4096 4096       -1     -1      -1            -1           -1           -1          4
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 7680  8192 8192       -1     -1      -1            -1           -1           -1          2
+ 
+ #######  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC BatchStrideA BatchStrideB BatchStrideC BatchCount
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024     1024    1024    1024           -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048     2048    2048    2048           -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096     4096    4096    4096           -1           -1           -1          4
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192     8192    8192    8192           -1           -1           -1          2
+ 
+ #######  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC BatchStrideA BatchStrideB BatchStrideC BatchCount
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024     1056    1056    1056           -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048     2080    2080    2080           -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096     4128    4128    4128           -1           -1           -1          4
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192     8224    8224    8224           -1           -1           -1          2
+ 
+ #######  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC BatchStrideA BatchStrideB BatchStrideC BatchCount
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024     1088    1088    1088           -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048     2112    2112    2112           -1           -1           -1          8
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096     4160    4160    4160           -1           -1           -1          4
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192     8256    8256    8256           -1           -1           -1          2
diff --git a/script/profile_conv_bwd_data.sh b/script/profile_conv_bwd_data.sh
new file mode 100755
index 0000000000..a1d2f450c9
--- /dev/null
+++ b/script/profile_conv_bwd_data.sh
@@ -0,0 +1,38 @@
+#!/bin/bash
+ 
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+
+ N=$8
+
+# Resnet50
+########  op  datatype  layout  verify  init  log  time conv_dim G__ N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256 1024 1 1   14   14     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512 1024 1 1   14   14     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  128 3 3   28   28     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512  128 1 1   28   28     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  128 3 3   56   56     2 2       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512 2048 1 1    7    7     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N 1024  256 1 1   14   14     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256  256 3 3   14   14     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256  256 3 3   28   28     2 2       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  256 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64  256 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512  512 3 3   14   14     2 2       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  512 1 1   28   28     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256  512 1 1   28   28     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N 2048  512 1 1    7    7     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512  512 3 3    7    7     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256   64 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64   64 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64   64 3 3   56   56     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64    3 7 7  224  224     2 2       1 1      3 3       3 3
diff --git a/script/profile_conv_fwd.sh b/script/profile_conv_fwd.sh
new file mode 100755
index 0000000000..a1d2f450c9
--- /dev/null
+++ b/script/profile_conv_fwd.sh
@@ -0,0 +1,38 @@
+#!/bin/bash
+ 
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+
+ N=$8
+
+# Resnet50
+########  op  datatype  layout  verify  init  log  time conv_dim G__ N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256 1024 1 1   14   14     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512 1024 1 1   14   14     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  128 3 3   28   28     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512  128 1 1   28   28     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  128 3 3   56   56     2 2       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512 2048 1 1    7    7     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N 1024  256 1 1   14   14     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256  256 3 3   14   14     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256  256 3 3   28   28     2 2       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  256 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64  256 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512  512 3 3   14   14     2 2       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  128  512 1 1   28   28     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256  512 1 1   28   28     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N 2048  512 1 1    7    7     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  512  512 3 3    7    7     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N  256   64 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64   64 1 1   56   56     1 1       1 1      0 0       0 0
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64   64 3 3   56   56     1 1       1 1      1 1       1 1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME        2   1  $N   64    3 7 7  224  224     2 2       1 1      3 3       3 3
diff --git a/script/profile_gemm.sh b/script/profile_gemm.sh
new file mode 100755
index 0000000000..b88159e74d
--- /dev/null
+++ b/script/profile_gemm.sh
@@ -0,0 +1,58 @@
+#!/bin/bash
+
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+echo $DRIVER
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+
+
+# 120 CU
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  960  1024 1024       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  960  2048 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1920  1024 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1920  2048 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 3840  4096 4096       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 7680  8192 8192       -1     -1      -1
+ 
+# 104 CU
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  832  1024 1024       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  832  2048 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1664  1024 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1664  2048 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 3328  4096 4096       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 6656  8192 8192       -1     -1      -1
+ 
+# 110 CU
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1280  1408 1024       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1280  2816 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2560  1408 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2560  2816 2048       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 5120  5632 4096       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 7040  8192 8192       -1     -1      -1
+
+# testing different strides
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024	    1024   1024    1024
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048	    2048   2048    2048
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096	    4096   4096    4096
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192	    8192   8192    8192
+ 
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024	    1056   1056    1056
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048	    2080   2080    2080
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096	    4128   4128    4128
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192	    8224   8224    8224
+ 
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024	    1088   1088    1088
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048	    2112   2112    2112
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096	    4160   4160    4160
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192	    8256   8256    8256
diff --git a/script/profile_gemm_bilinear.sh b/script/profile_gemm_bilinear.sh
new file mode 100755
index 0000000000..e6edefae85
--- /dev/null
+++ b/script/profile_gemm_bilinear.sh
@@ -0,0 +1,41 @@
+#!/bin/bash
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+ 
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideD StrideE Alpha Beta
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  960  1024 1024       -1      -1      -1      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1920  2048 2048       -1      -1      -1      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 3840  4096 4096       -1      -1      -1      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 7680  8192 8192       -1      -1      -1      -1     1    1
+ 
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideD StrideE Alpha Beta
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  960  1024 1024       -1      -1       0      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1920  2048 2048       -1      -1       0      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 3840  4096 4096       -1      -1       0      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 7680  8192 8192       -1      -1       0      -1     1    1
+ 
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideD StrideE Alpha Beta
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1000  1000 1000       -1      -1       0      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2000  2000 2000       -1      -1       0      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4000  4000 4000       -1      -1       0      -1     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8000  8000 8000       -1      -1       0      -1     1    1
+ 
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideD StrideE Alpha Beta
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024     1056    1056    1056    1056     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048     2080    2080    2080    2080     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096     4128    4128    4128    4128     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192     8224    8224    8224    8224     1    1
+ 
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideD StrideE Alpha Beta
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024     1088    1088    1088    1088     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048     2112    2112    2112    2112     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 4096  4096 4096     4160    4160    4160    4160     1    1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 8192  8192 8192     8256    8256    8256    8256     1    1
\ No newline at end of file
diff --git a/script/profile_grouped_gemm.sh b/script/profile_grouped_gemm.sh
new file mode 100755
index 0000000000..8adb7c81ac
--- /dev/null
+++ b/script/profile_grouped_gemm.sh
@@ -0,0 +1,18 @@
+#!/bin/bash
+ 
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+ 
+########  op  datatype  layout  verify  init  log  time  Ms______________ Ns______________ Ks_____________ StrideAs___________ StrideBs__________  StrideCs___________
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  256,512,1024,768 128,256,384,1024 128,192,256,512 1024,1025,1044,1026 1024,1024,1024,1024 1025,1024,1028,1024
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  512,768,2048,128 128,256,384,1024 128,192,256,512 1024,1025,2053,1026 1024,1024,1024,1024 1025,1024,2054,1024
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  256,512,1024,768 512,256,768,1024 128,192,256,512 1024,1045,1034,1026 1024,1024,1024,1024 1025,1063,1028,1024
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  512,768,4096,768 128,768,512,2048 128,192,256,512 1024,1027,4096,2050 1024,1024,1024,2048 1025,1024,4099,2049
diff --git a/script/profile_onnx_gemm.sh b/script/profile_onnx_gemm.sh
new file mode 100755
index 0000000000..c2721e7f59
--- /dev/null
+++ b/script/profile_onnx_gemm.sh
@@ -0,0 +1,31 @@
+#!/bin/bash
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+echo $DRIVER
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+# GEMM kernel benchmarks used by ONNX 
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  768  768        -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  768  2304       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  768  3072       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  3072 768        -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  1024 1024       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  1024 3072       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  1024 4096       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  384  4096 1024       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 768 768        -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 768 2304       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 768 3072       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 3072 768       -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 1024 1024      -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 1024 3072      -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 1024 4096      -1     -1      -1
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  24576 4096 1024      -1     -1      -1
+ 
diff --git a/script/profile_reduce_no_index.sh b/script/profile_reduce_no_index.sh
new file mode 100755
index 0000000000..66bfe1dcd3
--- /dev/null
+++ b/script/profile_reduce_no_index.sh
@@ -0,0 +1,78 @@
+#!/bin/bash
+DRIVER="../build/bin/ckProfiler"
+VERIFY="-v $1"
+INIT=$2
+NREPEAT=$3
+PRECISION=$4
+##PRECISION=--half
+##PRECISION=--double
+##PRECISION=--int8
+##PRECISION=--bf16
+
+if [ -n $PRECISION ] && [ "$PRECISION" = "--half" -o "$PRECISION" = "--bf16" ]; then
+   ACCTYPE="-C 1"
+elif [ -n $PRECISION ] && [ "$PRECISION" = "--int8" ]; then
+   ACCTYPE="-C 2"
+fi
+
+#### 0 - ADD,  5 - AVG,  7 - NORM2
+Operations="0 5"
+
+#### 0 - ADD,  5 - AVG,    for int8, no NORM2 supported
+if [ -n $PRECISION ] && [ "$PRECISION" = "--int8" -o "$PRECISION" = "--half" ]; then
+   Operations=5
+fi
+
+## for generic validation
+for op in $Operations; do
+    set -x
+    #######        datatype   layout          reduce dims  op     acctype   verify  init  repeats
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,1,2,3   -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 1         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 2         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 3         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 1,2,3     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,2,3     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,1,3     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,22960    -R 0         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,22960    -R 1         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 4,1469440    -R 0         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 4,1469440    -R 1         -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    set +x
+done
+
+#### 0 - ADD,  5 - AVG,  7 - NORM2
+Operations=5
+
+## for performance evaluation (resnet50 NHWC => C)
+for op in $Operations; do
+    set -x
+    #######        datatype   layout             reduce dims  op     acctype   verify  init  repeats
+    $DRIVER reduce $PRECISION -D 256,14,14,1024  -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT 
+    $DRIVER reduce $PRECISION -D 256,28,28,128   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,58,58,128   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,7,7,2048    -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,14,14,256   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,30,30,256   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,56,56,256   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,16,16,512   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,28,28,512   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,7,7,512     -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,56,56,64    -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 256,230,230,3   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,14,14,1024  -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,28,28,128   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,58,58,128   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,7,7,2048    -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,14,14,256   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,30,30,256   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,56,56,256   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,16,16,512   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,28,28,512   -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,7,7,512     -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    $DRIVER reduce $PRECISION -D 128,56,56,64    -R 0,1,2     -O $op $ACCTYPE  $VERIFY $INIT $NREPEAT
+    set +x
+done 
+
diff --git a/script/profile_reduce_with_index.sh b/script/profile_reduce_with_index.sh
new file mode 100755
index 0000000000..43543f4430
--- /dev/null
+++ b/script/profile_reduce_with_index.sh
@@ -0,0 +1,70 @@
+#!/bin/bash
+DRIVER="../build/bin/ckProfiler"
+VERIFY="-v $1"
+INIT=$2
+NREPEAT=$3
+PRECISION=$4
+##PRECISION=--half
+##PRECISION=--double
+##PRECISION=--int8
+##PRECISION=--bf16
+
+#### 2 - MIN,  3 - MAX,  4 - AMAX
+Operations="2 4"
+
+## for generic validation
+for op in $Operations; do
+    for use_idx in 0 1; do
+        set -x
+        #######        datatype   layout          reduce dims  op     use index    verify  init  repeats
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,1,2,3   -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 1         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 2         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 3         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 1,2,3     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,2,3     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 64,4,280,82  -R 0,1,3     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,22960    -R 0         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,22960    -R 1         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 4,1469440    -R 0         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 4,1469440    -R 1         -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        set +x
+    done
+done
+
+Operations=2
+
+## for performance evaluation (resnet50 NHWC => C)
+for op in $Operations; do
+    for use_idx in 0 1; do
+        set -x
+        #######        datatype   layout             reduce dims  op     use index    verify  init  repeats
+        $DRIVER reduce $PRECISION -D 256,14,14,1024  -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,28,28,128   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,58,58,128   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,7,7,2048    -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,14,14,256   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,30,30,256   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,56,56,256   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,16,16,512   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,28,28,512   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,7,7,512     -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,56,56,64    -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 256,230,230,3   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,14,14,1024  -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,28,28,128   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,58,58,128   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,7,7,2048    -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,14,14,256   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,30,30,256   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,56,56,256   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,16,16,512   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,28,28,512   -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,7,7,512     -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        $DRIVER reduce $PRECISION -D 128,56,56,64    -R 0,1,2     -O $op -I $use_idx  $VERIFY $INIT $NREPEAT
+        set +x
+    done
+done 
+
diff --git a/script/profile_resnet50.sh b/script/profile_resnet50.sh
new file mode 100755
index 0000000000..b55cb2ccef
--- /dev/null
+++ b/script/profile_resnet50.sh
@@ -0,0 +1,69 @@
+#!/bin/bash
+
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+
+OP=$1
+DATATYPE=$2
+IN_LAYOUT=$3
+WEI_LAYOUT=$4
+OUT_LAYOUT=$5
+VERIFY=$6
+INIT=$7
+LOG=$8
+TIME=$9
+
+ N=${10}
+
+# Resnet50
+######## op____________________  datatype  in_layout   wei_layout  out_layout  verify  init  log  time  N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N   64    3 7 7  224 224    2   2     1   1    3   3     3   3
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N   64   64 1 1   56  56    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N   64   64 3 3   56  56    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256   64 1 1   56  56    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N   64  256 1 1   56  56    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N   64   64 3 3   56  56    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256   64 1 1   56  56    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N   64  256 1 1   56  56    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N   64   64 3 3   56  56    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256   64 1 1   56  56    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  256 1 1   56  56    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  128 3 3   56  56    2   2     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512  128 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  512 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  128 3 3   28  28    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512  128 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  512 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  128 3 3   28  28    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512  128 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  512 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  128  128 3 3   28  28    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512  128 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256  512 1 1   28  28    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256  256 3 3   28  28    2   2     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 1024  256 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256 1024 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256  256 3 3   14  14    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 1024  256 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256 1024 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256  256 3 3   14  14    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 1024  256 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256 1024 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256  256 3 3   14  14    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 1024  256 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256 1024 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256  256 3 3   14  14    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 1024  256 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256 1024 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  256  256 3 3   14  14    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 1024  256 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512 1024 1 1   14  14    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512  512 3 3   14  14    2   2     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 2048  512 1 1    7   7    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512 2048 1 1    7   7    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512  512 3 3    7   7    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 2048  512 1 1    7   7    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512 2048 1 1    7   7    1   1     1   1    0   0     0   0
+ $DRIVER conv_fwd_bias_relu     $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N  512  512 3 3    7   7    1   1     1   1    1   1     1   1
+ $DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT  $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $TIME   $N 2048  512 1 1    7   7    1   1     1   1    0   0     0   0
diff --git a/script/profile_splitK_gemm.sh b/script/profile_splitK_gemm.sh
new file mode 100755
index 0000000000..d62f0e4753
--- /dev/null
+++ b/script/profile_splitK_gemm.sh
@@ -0,0 +1,41 @@
+#!/bin/bash
+
+## GPU visibility
+export HIP_VISIBLE_DEVICES=0
+DRIVER="../build/bin/ckProfiler"
+echo $DRIVER
+OP=$1
+DATATYPE=$2
+LAYOUT=$3
+VERIFY=$4
+INIT=$5
+LOG=$6
+TIME=$7
+KBatch=$8
+
+
+# 120 CU
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC  KBatch_
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  960  1024 1024       -1     -1      -1   $KBatch
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  960  2048 2048       -1     -1      -1   $KBatch
+ 
+# 104 CU
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC KBatch_
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  832  1024 1024       -1     -1      -1  $KBatch
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME  832  2048 2048       -1     -1      -1  $KBatch
+ 
+# 110 CU
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC KBatch_
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1280  1408 1024       -1     -1      -1  $KBatch
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1280  2816 2048       -1     -1      -1  $KBatch
+
+# testing different strides
+########  op  datatype  layout  verify  init  log  time  M___ N___ K___  StrideA StrideB StrideC KBatch_
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024	    1024   1024    1024  $KBatch
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048	    2048   2048    2048  $KBatch
+ 
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024	    1056   1056    1056  $KBatch
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048	    2080   2080    2080  $KBatch
+ 
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 1024  1024 1024	    1088   1088    1088  $KBatch
+ $DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $TIME 2048  2048 2048	    2112   2112    2112  $KBatch
diff --git a/script/run_full_performance_tests.sh b/script/run_full_performance_tests.sh
new file mode 100755
index 0000000000..eae334ae2d
--- /dev/null
+++ b/script/run_full_performance_tests.sh
@@ -0,0 +1,149 @@
+#!/bin/bash 
+#
+# in order to run this script you'd first need to build the ckProfiler executable in ../build/bin/
+# you would also need to set up some environment variables in order to 
+# post your new test results to the database and compare them to the baseline
+# please contact Illia.Silin@amd.com for more details
+#
+# run the script as "./run_full_performance_tests.sh <verification> <tag for your test environment> <branch name> < node name>
+# input arguments: 
+# verification = 0 : do not verify result correctness on CPU
+#              = 1 : verifuy correctness on CPU (may take a long time)
+# environment tag  : a string describing the specifics of your test environment
+# branch name      : name of the branch in git repo (git status | grep -e 'On branch')
+# node name        : $hostname
+
+#get the command line arguments:
+export verify=$1
+echo 'Verification: ' $verify
+export env_type=$2
+echo 'Environment type: ' $env_type
+export branch=$3
+echo 'Branch name: ' $branch
+export host_name=$4
+echo 'Host name: ' $host_name
+function print_log_header(){
+	rm -f $1;
+	echo 'On branch ' $3 &> $1;
+	echo 'Node name: ' $4 >> $1;
+	#get GPU_arch and number of compute units from rocminfo
+	echo -n "GPU_arch: " >> $1; rocminfo | grep "Name:" | grep "gfx" >> $1;
+	rocminfo | grep "Compute Unit:" >> $1;
+	hipcc --version | grep -e 'HIP version'  >> $1;
+	echo 'Environment type: ' $2 >> $1;
+	/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> $1;
+}
+
+#run gemm tests
+export gemm_log="perf_gemm.log"
+print_log_header $gemm_log $env_type $branch $host_name
+./profile_gemm.sh gemm 0 0 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 0 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 0 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 0 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 0 1 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 1 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 1 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 1 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 0 2 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 2 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 2 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 2 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 0 3 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 3 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 3 $verify 1 0 1 2>&1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 3 $verify 1 0 1 2>&1 | tee -a $gemm_log
+
+#run batched_gemm tests
+export batched_gemm_log="perf_batched_gemm.log"
+print_log_header $batched_gemm_log $env_type $branch $host_name
+./profile_batched_gemm.sh batched_gemm 0 0 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 0 1 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 0 2 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 0 3 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 1 0 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 1 1 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 1 2 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 1 3 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 2 0 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 2 1 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 2 2 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 2 3 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 3 0 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 3 1 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 3 2 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+./profile_batched_gemm.sh batched_gemm 3 3 $verify 1 0 1 2>&1 | tee -a $batched_gemm_log
+
+#run grouped_gemm tests
+export grouped_gemm_log="perf_grouped_gemm.log"
+print_log_header $grouped_gemm_log $env_type $branch $host_name
+./profile_grouped_gemm.sh grouped_gemm 1 0 $verify 1 0 1 2>&1 | tee -a $grouped_gemm_log
+./profile_grouped_gemm.sh grouped_gemm 1 1 $verify 1 0 1 2>&1 | tee -a $grouped_gemm_log
+./profile_grouped_gemm.sh grouped_gemm 1 2 $verify 1 0 1 2>&1 | tee -a $grouped_gemm_log
+./profile_grouped_gemm.sh grouped_gemm 1 3 $verify 1 0 1 2>&1 | tee -a $grouped_gemm_log
+
+#run GEMM+Bilinear tests
+export gemm_bilinear_log="perf_gemm_bilinear.log"
+print_log_header $gemm_bilinear_log $env_type $branch $host_name
+./profile_gemm_bilinear.sh gemm_bilinear 1 0 $verify 1 0 1 2>&1 | tee -a $gemm_bilinear_log
+./profile_gemm_bilinear.sh gemm_bilinear 1 1 $verify 1 0 1 2>&1 | tee -a $gemm_bilinear_log
+./profile_gemm_bilinear.sh gemm_bilinear 1 2 $verify 1 0 1 2>&1 | tee -a $gemm_bilinear_log
+./profile_gemm_bilinear.sh gemm_bilinear 1 3 $verify 1 0 1 2>&1 | tee -a $gemm_bilinear_log
+
+#run conv_fwd tests
+export conv_fwd_log="perf_conv_fwd.log"
+print_log_header $conv_fwd_log $env_type $branch $host_name
+./profile_conv_fwd.sh conv_fwd 0 1 $verify 1 0 1 256 2>&1 | tee -a $conv_fwd_log
+./profile_conv_fwd.sh conv_fwd 1 1 $verify 1 0 1 256 2>&1 | tee -a $conv_fwd_log
+./profile_conv_fwd.sh conv_fwd 2 1 $verify 1 0 1 256 2>&1 | tee -a $conv_fwd_log
+./profile_conv_fwd.sh conv_fwd 3 1 $verify 1 0 1 256 2>&1 | tee -a $conv_fwd_log
+
+#run conv_bwd_data tests
+export conv_bwd_data_log="perf_conv_bwd_data.log"
+print_log_header $conv_bwd_data_log $env_type $branch $host_name
+./profile_conv_bwd_data.sh conv_bwd_data 0 1 $verify 1 0 1 256 2>&1 | tee -a $conv_bwd_data_log
+./profile_conv_bwd_data.sh conv_bwd_data 1 1 $verify 1 0 1 256 2>&1 | tee -a $conv_bwd_data_log
+./profile_conv_bwd_data.sh conv_bwd_data 2 1 $verify 1 0 1 256 2>&1 | tee -a $conv_bwd_data_log
+./profile_conv_bwd_data.sh conv_bwd_data 3 1 $verify 1 0 1 256 2>&1 | tee -a $conv_bwd_data_log
+
+#run resnet50 tests
+export resnet256_log="perf_resnet50_N256.log"
+print_log_header $resnet256_log $env_type $branch $host_name
+./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 256 2>&1 | tee -a $resnet256_log
+export resnet4_log="perf_resnet50_N4.log"
+print_log_header $resnet4_log $env_type $branch $host_name
+./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 4 2>&1 | tee -a $resnet4_log
+
+#run reduction tests
+export reduction_log="perf_reduction.log"
+print_log_header $reduction_log $env_type $branch $host_name
+./profile_reduce_with_index.sh $verify 2 10 --half 2>&1 | tee -a $reduction_log
+./profile_reduce_no_index.sh $verify 2 10 --half 2>&1 | tee -a $reduction_log
+
+#run splitK_gemm tests, first correctness verification, then performance
+export splitK_gemm_ver_log="perf_splitK_gemm_verify.log"
+print_log_header $splitK_gemm_ver_log $env_type $branch $host_name
+./profile_splitK_gemm.sh gemm_splitk 0 0 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+./profile_splitK_gemm.sh gemm_splitk 0 1 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+./profile_splitK_gemm.sh gemm_splitk 0 2 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+./profile_splitK_gemm.sh gemm_splitk 0 3 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+./profile_splitK_gemm.sh gemm_splitk 1 0 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+./profile_splitK_gemm.sh gemm_splitk 1 1 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+./profile_splitK_gemm.sh gemm_splitk 1 2 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+./profile_splitK_gemm.sh gemm_splitk 1 3 $verify 1 0 0 4 2>&1 | tee -a $splitK_gemm_ver_log
+export splitK_gemm_log="perf_splitK_gemm.log"
+print_log_header $splitK_gemm_log $env_type $branch $host_name
+./profile_splitK_gemm.sh gemm_splitk 0 0 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+./profile_splitK_gemm.sh gemm_splitk 0 1 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+./profile_splitK_gemm.sh gemm_splitk 0 2 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+./profile_splitK_gemm.sh gemm_splitk 0 3 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+./profile_splitK_gemm.sh gemm_splitk 1 0 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+./profile_splitK_gemm.sh gemm_splitk 1 1 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+./profile_splitK_gemm.sh gemm_splitk 1 2 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+./profile_splitK_gemm.sh gemm_splitk 1 3 0 1 0 1 4 2>&1 | tee -a $splitK_gemm_log
+
+#run ONNX gemm tests
+export onnx_log="perf_onnx_gemm.log"
+print_log_header $onnx_log $env_type $branch $host_name
+./profile_onnx_gemm.sh gemm 0 0 $verify 1 0 1 2>&1 | tee -a $onnx_log
+./profile_onnx_gemm.sh gemm 1 0 $verify 1 0 1 2>&1 | tee -a $onnx_log
diff --git a/script/run_performance_tests.sh b/script/run_performance_tests.sh
new file mode 100755
index 0000000000..4e3a6fc8eb
--- /dev/null
+++ b/script/run_performance_tests.sh
@@ -0,0 +1,60 @@
+#!/bin/bash 
+#
+# in order to run this script you'd first need to build the ckProfiler executable in ../build/bin/
+# run the script as "./run_performance_tests.sh <verification> <tag for your test environment> <branch name> < node name>
+# input arguments: 
+# verification = 0 : do not verify result correctness on CPU
+#              = 1 : verify correctness on CPU (may take a long time)
+# environment tag  : a string describing the specifics of your test environment
+# branch name      : name of the branch in git repo (git status | grep -e 'On branch')
+# node name        : $hostname
+
+#get the command line arguments:
+export verify=$1
+echo 'Verification: ' $verify
+export env_type=$2
+echo 'Environment type: ' $env_type
+export branch=$3
+echo 'Branch name: ' $branch
+export host_name=$4
+echo 'Host name: ' $host_name
+
+function print_log_header(){
+	rm -f $1;
+	echo 'On branch ' $3 &> $1;
+	echo 'Node name: ' $4 >> $1;
+	#get GPU_arch and number of compute units from rocminfo
+	echo -n "GPU_arch: " >> $1; rocminfo | grep "Name:" | grep "gfx" >> $1;
+	rocminfo | grep "Compute Unit:" >> $1;
+	hipcc --version | grep -e 'HIP version'  >> $1;
+	echo 'Environment type: ' $2 >> $1;
+	/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> $1;
+}
+
+#run gemm tests
+export gemm_log="perf_gemm.log"
+print_log_header $gemm_log $env_type $branch $host_name
+./profile_gemm.sh gemm 0 0 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 0 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 0 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 0 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 0 1 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 1 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 1 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 1 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 0 2 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 2 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 2 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 2 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 0 3 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 1 3 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 2 3 $verify 1 0 1 | tee -a $gemm_log
+./profile_gemm.sh gemm 3 3 $verify 1 0 1 | tee -a $gemm_log
+
+#run resnet50 tests
+export resnet256_log="perf_resnet50_N256.log"
+print_log_header $resnet256_log $env_type $branch $host_name
+./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 256 | tee -a $resnet256_log
+export resnet4_log="perf_resnet50_N4.log"
+print_log_header $resnet4_log $env_type $branch $host_name
+./profile_resnet50.sh conv_fwd_bias_relu 1 1 1 1 $verify 1 0 1 4 | tee -a $resnet4_log
diff --git a/script/test_convnd_fwd.sh b/script/test_convnd_fwd.sh
new file mode 100644
index 0000000000..1bd7a6b5d7
--- /dev/null
+++ b/script/test_convnd_fwd.sh
@@ -0,0 +1,110 @@
+#!/usr/bin/env bash
+
+# set -e
+
+DIM1=False
+DIM2=True
+DIM3=False
+DATE=220317
+GIT_HASH=4e6dfda
+LOG_DIR=${DATE}_${GIT_HASH}
+SUFFIX=${GIT_HASH}
+
+
+#--------------------------------------------------------------------------
+#   Commandline arguments parsing
+#   like: cmd -key[--key] value
+#--------------------------------------------------------------------------
+
+POSITIONAL=()
+while [[ $# -gt 0 ]]
+do
+key="$1"
+
+case $key in
+    -d1|--d1)
+    DIM1=True
+    echo DIM1: "${DIM1}"
+    shift # past argument
+    ;;
+    -d2|--d2)
+    DIM2=True
+    echo DIM2: "${DIM2}"
+    shift # past argument
+    ;;
+    -d3|--d3)
+    DIM3=True
+    echo DIM3: "${DIM3}"
+    shift # past argument
+    ;;
+    -all|--all)
+    DIM1=True
+    DIM2=True
+    DIM3=True
+    echo DIM1: "${DIM1}"
+    echo DIM2: "${DIM2}"
+    echo DIM3: "${DIM3}"
+    shift # past argument
+    ;;
+    -s|--suffix)
+    SUFFIX=${SUFFIX}_"$2"
+    echo SUFFIX: "${SUFFIX}"
+    shift # past argument
+    shift # past value
+    ;;
+    *)    # unknown option
+    POSITIONAL+=("$1") # save it in an array for later
+    shift # past argument
+    ;;
+esac
+done
+set -- "${POSITIONAL[@]}" # restore positional parameters
+
+#--------------------------------------------------------------------------
+
+# NUMACTL="numactl --cpunodebind=1 --membind=1"
+NUMACTL=
+# ENV_CONF=
+GPU=mi100
+PROF_ITER_COUNT=10000
+LOG_DIR_PATH=../log/${LOG_DIR}
+set -x
+
+#-------------------------------------------------------------------------------
+#               1D
+#-------------------------------------------------------------------------------
+
+if [[ "${DIM1}" == "True" ]]; then
+    mkdir -p ${LOG_DIR_PATH}
+    echo ">>>>>>>> RUN test conv1d nwc <<<<<<<<<<"
+    CMD="./../build/bin/test_conv1d_fwd"
+    ${NUMACTL} ${CMD} 2>&1 \
+        | tee ${LOG_DIR_PATH}/test_conv1d_fwd_nwc_${SUFFIX}_${GPU}.log
+
+fi
+
+#-------------------------------------------------------------------------------
+#               2D
+#-------------------------------------------------------------------------------
+
+if [[ "${DIM2}" == "True" ]]; then
+    mkdir -p ${LOG_DIR_PATH}
+    echo ">>>>>>>> RUN test conv2d nhwc <<<<<<<<<<"
+    CMD="./../build/bin/test_conv2d_fwd"
+    ${NUMACTL} ${CMD} 2>&1 \
+        | tee ${LOG_DIR_PATH}/test_conv2d_fwd_nhwc_${SUFFIX}_${GPU}.log
+
+fi
+
+#-------------------------------------------------------------------------------
+#               3D
+#-------------------------------------------------------------------------------
+
+if [[ "${DIM3}" == "True" ]]; then
+    mkdir -p ${LOG_DIR_PATH}
+    echo ">>>>>>>> RUN test conv3d ndhwc <<<<<<<<<<"
+    CMD="./../build/bin/test_conv3d_fwd"
+    ${NUMACTL} ${CMD} 2>&1 \
+        | tee ${LOG_DIR_PATH}/test_conv3d_fwd_ndhwc_${SUFFIX}_${GPU}.log
+
+fi
diff --git a/script/test_reduce_no_index.sh b/script/test_reduce_no_index.sh
new file mode 100755
index 0000000000..b956303837
--- /dev/null
+++ b/script/test_reduce_no_index.sh
@@ -0,0 +1,63 @@
+#!/bin/bash
+
+## The following will be used for CI
+
+set -x
+
+## for float
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2,3  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,3  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,2,3  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1,2,3  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 2  0 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 3  0 2
+
+## for float64
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2,3  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,3  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,2,3  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1,2,3  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 2  6 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 3  6 2
+
+## for float16
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2,3  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,3  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,2,3  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1,2,3  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 2  1 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 3  1 2
+
+## for int8_t
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2,3  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,3  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,2,3  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1,2,3  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 2  3 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 3  3 2
+
+## for bfloat16
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2,3  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,2  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,1,3  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0,2,3  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1,2,3  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 0  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 1  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 2  5 2
+bin/test_reduce_no_index -D 64,4,280,82  -R 3  5 2
+
+set +x
+
diff --git a/script/test_reduce_with_index.sh b/script/test_reduce_with_index.sh
new file mode 100755
index 0000000000..b0843ba6c1
--- /dev/null
+++ b/script/test_reduce_with_index.sh
@@ -0,0 +1,63 @@
+#!/bin/bash
+
+## The following will be used for CI
+
+set -x
+
+## for float
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2,3  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,3  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,2,3  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1,2,3  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 2  0 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 3  0 2
+
+## for float64
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2,3  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,3  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,2,3  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1,2,3  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 2  6 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 3  6 2
+
+## for float16
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2,3  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,3  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,2,3  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1,2,3  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 2  1 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 3  1 2
+
+## for int8_t
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2,3  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,3  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,2,3  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1,2,3  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 2  3 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 3  3 2
+
+## for bfloat16
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2,3  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,2  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,1,3  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0,2,3  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1,2,3  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 0  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 1  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 2  5 2
+bin/test_reduce_with_index -D 64,4,280,82  -R 3  5 2
+
+set +x
+
diff --git a/script/uninstall_precommit.sh b/script/uninstall_precommit.sh
new file mode 100755
index 0000000000..b0d4d15166
--- /dev/null
+++ b/script/uninstall_precommit.sh
@@ -0,0 +1 @@
+pre-commit uninstall
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
new file mode 100644
index 0000000000..8fddd60858
--- /dev/null
+++ b/test/CMakeLists.txt
@@ -0,0 +1,66 @@
+include_directories(BEFORE
+    ${PROJECT_SOURCE_DIR}/
+    ${PROJECT_SOURCE_DIR}/profiler/include
+)
+
+include(googletest)
+
+add_custom_target(tests)
+
+function(add_test_executable TEST_NAME)
+    message("adding test ${TEST_NAME}")
+    add_executable(${TEST_NAME} ${ARGN})
+    add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}>)
+    add_dependencies(tests ${TEST_NAME})
+    add_dependencies(check ${TEST_NAME})
+    rocm_install(TARGETS ${TEST_NAME} COMPONENT tests)
+endfunction(add_test_executable TEST_NAME)
+
+include(GoogleTest)
+
+function(add_gtest_executable TEST_NAME)
+    message("adding gtest ${TEST_NAME}")
+    add_executable(${TEST_NAME} ${ARGN})
+    add_dependencies(tests ${TEST_NAME})
+    add_dependencies(check ${TEST_NAME})
+
+    # suppress gtest warnings
+    target_compile_options(${TEST_NAME} PRIVATE -Wno-global-constructors -Wno-undef)
+    target_link_libraries(${TEST_NAME} PRIVATE gtest_main)
+    add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}>)
+    rocm_install(TARGETS ${TEST_NAME} COMPONENT tests)
+endfunction(add_gtest_executable TEST_NAME)
+
+add_subdirectory(magic_number_division)
+add_subdirectory(space_filling_curve)
+add_subdirectory(conv_util)
+add_subdirectory(reference_conv_fwd)
+add_subdirectory(gemm)
+add_subdirectory(gemm_layernorm)
+add_subdirectory(gemm_split_k)
+add_subdirectory(gemm_reduce)
+add_subdirectory(batched_gemm)
+add_subdirectory(batched_gemm_reduce)
+add_subdirectory(batched_gemm_gemm)
+add_subdirectory(batched_gemm_softmax_gemm)
+add_subdirectory(batched_gemm_softmax_gemm_permute)
+add_subdirectory(grouped_gemm)
+add_subdirectory(reduce)
+add_subdirectory(convnd_fwd)
+add_subdirectory(convnd_bwd_data)
+add_subdirectory(grouped_convnd_fwd)
+add_subdirectory(grouped_convnd_bwd_weight)
+add_subdirectory(block_to_ctile_map)
+add_subdirectory(softmax)
+add_subdirectory(normalization)
+add_subdirectory(data_type)
+add_subdirectory(elementwise_normalization)
+add_subdirectory(batchnorm)
+add_subdirectory(contraction)
+add_subdirectory(pool)
+add_subdirectory(batched_gemm_multi_d)
+add_subdirectory(grouped_convnd_bwd_data)
+add_subdirectory(image_to_column)
+if(GPU_TARGETS MATCHES "gfx11")
+    add_subdirectory(wmma_op)
+endif()
diff --git a/test/batched_gemm/CMakeLists.txt b/test/batched_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..8a3269e90f
--- /dev/null
+++ b/test/batched_gemm/CMakeLists.txt
@@ -0,0 +1,27 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_test_executable(test_batched_gemm_fp16 batched_gemm_fp16.cpp)
+      target_link_libraries(test_batched_gemm_fp16 PRIVATE utility)
+      target_link_libraries(test_batched_gemm_fp16 PRIVATE device_batched_gemm_instance)
+   endif()
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+      add_test_executable(test_batched_gemm_fp32 batched_gemm_fp32.cpp)
+      target_link_libraries(test_batched_gemm_fp32 PRIVATE utility)
+      target_link_libraries(test_batched_gemm_fp32 PRIVATE device_batched_gemm_instance)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+      add_test_executable(test_batched_gemm_bf16 batched_gemm_bf16.cpp)
+      target_link_libraries(test_batched_gemm_bf16 PRIVATE utility)
+      target_link_libraries(test_batched_gemm_bf16 PRIVATE device_batched_gemm_instance)
+   endif()
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_test_executable(test_batched_gemm_int8 batched_gemm_int8.cpp)
+      target_link_libraries(test_batched_gemm_int8 PRIVATE utility)
+      target_link_libraries(test_batched_gemm_int8 PRIVATE device_batched_gemm_instance)
+   endif()
+   set(target 1)
+ endif()
+endforeach()
\ No newline at end of file
diff --git a/test/batched_gemm/batched_gemm_bf16.cpp b/test/batched_gemm/batched_gemm_bf16.cpp
new file mode 100644
index 0000000000..5d12a1e956
--- /dev/null
+++ b/test/batched_gemm/batched_gemm_bf16.cpp
@@ -0,0 +1,114 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "profiler/profile_batched_gemm_impl.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
+
+namespace {
+using ADataType = ck::bhalf_t;
+using BDataType = ck::bhalf_t;
+using CDataType = ck::bhalf_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+} // namespace
+
+int main()
+{
+    int M          = 256;
+    int N          = 256;
+    int K          = 128;
+    int BatchCount = 3;
+
+    bool pass = true;
+
+    using namespace ck::tensor_operation::device;
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Row,
+                                                           Row,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Row,
+                                                                             Row,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Row,
+                                                           Col,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Row,
+                                                                             Col,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Col,
+                                                           Row,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Col,
+                                                                             Row,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Col,
+                                                           Col,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Col,
+                                                                             Col,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount);
+
+    std::cout << "test BatchedGEMM bf16: " << (pass ? "Pass" : "Fail") << std::endl;
+    return pass ? 0 : 1;
+}
diff --git a/test/batched_gemm/batched_gemm_fp16.cpp b/test/batched_gemm/batched_gemm_fp16.cpp
new file mode 100644
index 0000000000..a2b61d951a
--- /dev/null
+++ b/test/batched_gemm/batched_gemm_fp16.cpp
@@ -0,0 +1,114 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "profiler/profile_batched_gemm_impl.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
+
+namespace {
+using ADataType = ck::half_t;
+using BDataType = ck::half_t;
+using CDataType = ck::half_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+} // namespace
+
+int main()
+{
+    int M          = 512;
+    int N          = 256;
+    int K          = 128;
+    int BatchCount = 3;
+
+    bool pass = true;
+
+    using namespace ck::tensor_operation::device;
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Row,
+                                                           Row,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Row,
+                                                                             Row,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Row,
+                                                           Col,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Row,
+                                                                             Col,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Col,
+                                                           Row,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Col,
+                                                                             Row,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Col,
+                                                           Col,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Col,
+                                                                             Col,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount);
+
+    std::cout << "test BatchedGEMM fp16: " << (pass ? "Pass" : "Fail") << std::endl;
+    return pass ? 0 : 1;
+}
diff --git a/test/batched_gemm/batched_gemm_fp32.cpp b/test/batched_gemm/batched_gemm_fp32.cpp
new file mode 100644
index 0000000000..2b18d166e6
--- /dev/null
+++ b/test/batched_gemm/batched_gemm_fp32.cpp
@@ -0,0 +1,114 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "profiler/profile_batched_gemm_impl.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
+
+namespace {
+using ADataType = float;
+using BDataType = float;
+using CDataType = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+} // namespace
+
+int main()
+{
+    int M          = 256;
+    int N          = 256;
+    int K          = 128;
+    int BatchCount = 3;
+
+    bool pass = true;
+
+    using namespace ck::tensor_operation::device;
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Row,
+                                                           Row,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Row,
+                                                                             Row,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Row,
+                                                           Col,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Row,
+                                                                             Col,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Col,
+                                                           Row,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Col,
+                                                                             Row,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Col,
+                                                           Col,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Col,
+                                                                             Col,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount);
+
+    std::cout << "test BatchedGEMM fp32: " << (pass ? "Pass" : "Fail") << std::endl;
+    return pass ? 0 : 1;
+}
diff --git a/test/batched_gemm/batched_gemm_int8.cpp b/test/batched_gemm/batched_gemm_int8.cpp
new file mode 100644
index 0000000000..f607eaa84b
--- /dev/null
+++ b/test/batched_gemm/batched_gemm_int8.cpp
@@ -0,0 +1,114 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "profiler/profile_batched_gemm_impl.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
+
+namespace {
+using ADataType = int8_t;
+using BDataType = int8_t;
+using CDataType = int8_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+} // namespace
+
+int main()
+{
+    int M          = 256;
+    int N          = 256;
+    int K          = 128;
+    int BatchCount = 3;
+
+    bool pass = true;
+
+    using namespace ck::tensor_operation::device;
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Row,
+                                                           Row,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Row,
+                                                                             Row,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Row,
+                                                           Col,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Row,
+                                                                             Col,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Col,
+                                                           Row,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Col,
+                                                                             Row,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
+                                                           BDataType,
+                                                           CDataType,
+                                                           Col,
+                                                           Col,
+                                                           Row,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           DeviceBatchedGemm<Col,
+                                                                             Col,
+                                                                             Row,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             PassThrough,
+                                                                             PassThrough,
+                                                                             PassThrough>>(
+                       true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount);
+
+    std::cout << "test BatchedGEMM int8: " << (pass ? "Pass" : "Fail") << std::endl;
+    return pass ? 0 : 1;
+}
diff --git a/test/batched_gemm_gemm/CMakeLists.txt b/test/batched_gemm_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..404e74f3bc
--- /dev/null
+++ b/test/batched_gemm_gemm/CMakeLists.txt
@@ -0,0 +1,13 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_custom_target(test_batched_gemm_gemm)
+      add_gtest_executable(test_batched_gemm_gemm_fp16 test_batched_gemm_gemm_fp16.cpp)
+      target_link_libraries(test_batched_gemm_gemm_fp16 PRIVATE utility device_batched_gemm_gemm_instance)
+      add_dependencies(test_batched_gemm_gemm test_batched_gemm_gemm_fp16)
+      set(target 1)
+   endif()
+ endif()
+endforeach()
\ No newline at end of file
diff --git a/test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp b/test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp
new file mode 100644
index 0000000000..1a8d5c2e55
--- /dev/null
+++ b/test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp
@@ -0,0 +1,149 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "test_batched_gemm_gemm_util.hpp"
+
+template <typename Tuple>
+class TestBatchedGemmGemmFP16 : public TestBatchedGemmGemm<Tuple>
+{
+};
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    std::tuple<F16, F16, F16, F16, Row, Col, Row, Row>,
+    std::tuple<F16, F16, F16, F16, Row, Col, Col, Row>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestBatchedGemmGemmFP16, KernelTypes);
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16) { this->Run(); }
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_PadM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {136, 128, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_PadN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 136, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_PadK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 40, 128, 1},
+        {128, 128, 136, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_PadO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 136, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_OddM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {129, 128, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_OddN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 129, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_OddK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 33, 128, 1},
+        {128, 128, 129, 128, 1},
+    };
+    this->Run();
+}
+
+// If kernel B1Layout is RowMajor, expect not to support odd O size
+TYPED_TEST(TestBatchedGemmGemmFP16, Test_FP16_OddO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 129, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmGemmFP16, DISABLED_Bench_FP16)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {256, 256, 64, 64, 768},
+        {256, 256, 128, 128, 768},
+        {512, 512, 64, 64, 768},
+        {512, 512, 128, 128, 768},
+        {1024, 1024, 64, 64, 768},
+        {1024, 1024, 128, 128, 768},
+        {2048, 2048, 64, 64, 768},
+        {2048, 2048, 128, 128, 768},
+        {4096, 4096, 64, 64, 768},
+        {4096, 4096, 128, 128, 768},
+    };
+    this->bench_  = true;
+    this->verify_ = false;
+    this->Run();
+}
+
+using ck::tensor_operation::device::GemmSpecialization;
+
+TEST(TestBatchedGemmGemmInterface, GemmSpecializationSizeMatch)
+{
+    int P = 120; // requires padding
+    int Q = 128; // do not require padding
+
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
+    // clang-format on
+}
+
+TEST(TestBatchedGemmGemmInterface, GemmSpecializationSizeMismatch)
+{
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
+    // Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
+    // Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
+    // clang-format on
+}
diff --git a/test/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp b/test/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp
new file mode 100644
index 0000000000..b0fffc466e
--- /dev/null
+++ b/test/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp
@@ -0,0 +1,189 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include <vector>
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp"
+#include "profiler/profile_batched_gemm_gemm_impl.hpp"
+
+using ck::tensor_operation::device::GemmSpecialization;
+
+template <ck::index_t N>
+using I = ck::Number<N>;
+
+using F16 = ck::half_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <typename Tuple>
+struct TestBatchedGemmGemm : public ::testing::Test
+{
+    using ADataType  = std::tuple_element_t<0, Tuple>;
+    using B0DataType = std::tuple_element_t<1, Tuple>;
+    using B1DataType = std::tuple_element_t<2, Tuple>;
+    using CDataType  = std::tuple_element_t<3, Tuple>;
+    using ALayout    = std::tuple_element_t<4, Tuple>;
+    using B0Layout   = std::tuple_element_t<5, Tuple>;
+    using B1Layout   = std::tuple_element_t<6, Tuple>;
+    using CLayout    = std::tuple_element_t<7, Tuple>;
+
+    std::vector<std::vector<int>> lengths_ = {
+        {256, 256, 64, 64, 4},
+        {256, 256, 128, 128, 4},
+        {512, 512, 64, 64, 2},
+        {512, 512, 128, 128, 2},
+        {1024, 1024, 64, 64, 1},
+        {1024, 1024, 128, 128, 1},
+    };
+    bool bench_  = false;
+    bool verify_ = true;
+
+    void RunSingle(int M, int N, int K, int O, int BatchCount)
+    {
+        bool pass = ck::profiler::profile_batched_gemm_gemm_impl<ADataType,
+                                                                 B0DataType,
+                                                                 B1DataType,
+                                                                 CDataType,
+                                                                 ALayout,
+                                                                 B0Layout,
+                                                                 B1Layout,
+                                                                 CLayout>(
+            verify_, 1, false, bench_, M, N, K, O, BatchCount);
+
+        EXPECT_TRUE(pass);
+    }
+
+    void Run()
+    {
+        for(auto lengths : this->lengths_)
+        {
+            int M          = lengths[0];
+            int N          = lengths[1];
+            int K          = lengths[2];
+            int O          = lengths[3];
+            int BatchCount = lengths[4];
+
+            this->RunSingle(M, N, K, O, BatchCount);
+        }
+    }
+};
+
+template <GemmSpecialization GemmSpec>
+struct DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    using ALayout  = Row;
+    using B0Layout = Col;
+    using B1Layout = Row;
+    using CLayout  = Row;
+
+    using ADataType        = F16;
+    using B0DataType       = F16;
+    using B1DataType       = F16;
+    using AccDataType      = float;
+    using CShuffleDataType = float;
+    using CDataType        = F16;
+
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = PassThrough;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+
+    template <ck::index_t... Is>
+    using S = ck::Sequence<Is...>;
+
+    // static constexpr auto GemmSpec = std::tuple_element_t<0, Tuple>::value;
+
+    using DeviceGemmGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmGemm_Xdl_CShuffle<
+        ALayout,
+        B0Layout,
+        B1Layout,
+        CLayout,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        Acc0ElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        128,         // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        4,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<8, 32, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8>;             // CShuffleBlockTransferScalarPerVector_NPerBlock
+
+    bool IsSupported(int M, int N, int K, int O)
+    {
+        auto gemm     = DeviceGemmGemmInstance{};
+        auto invoker  = gemm.MakeInvoker();
+        auto argument = gemm.MakeArgument(static_cast<ADataType*>(nullptr),
+                                          static_cast<B0DataType*>(nullptr),
+                                          static_cast<B1DataType*>(nullptr),
+                                          static_cast<CDataType*>(nullptr),
+                                          M,
+                                          N,
+                                          K,
+                                          O,
+                                          0,              // BatchCount
+                                          0,              // StrideA
+                                          0,              // StrideB0
+                                          0,              // StrideB1
+                                          0,              // StrideC
+                                          0,              // BatchStrideA
+                                          0,              // BatchStrideB0
+                                          0,              // BatchStrideB1
+                                          0,              // BatchStrideC
+                                          PassThrough{},  // a_element_op
+                                          PassThrough{},  // b0_element_op
+                                          PassThrough{},  // acc0_element_op
+                                          PassThrough{},  // b1_element_op
+                                          PassThrough{}); // c_element_op
+
+        return gemm.IsSupportedArgument(argument);
+    }
+};
diff --git a/test/batched_gemm_multi_d/CMakeLists.txt b/test/batched_gemm_multi_d/CMakeLists.txt
new file mode 100644
index 0000000000..825f0dd235
--- /dev/null
+++ b/test/batched_gemm_multi_d/CMakeLists.txt
@@ -0,0 +1,4 @@
+if(DL_KERNELS)
+    add_gtest_executable(test_batched_gemm_multi_d test_batched_gemm_multi_d.cpp)
+    target_link_libraries(test_batched_gemm_multi_d PRIVATE utility device_batched_gemm_multi_d_instance)
+endif()
diff --git a/test/batched_gemm_multi_d/test_batched_gemm_multi_d.cpp b/test/batched_gemm_multi_d/test_batched_gemm_multi_d.cpp
new file mode 100644
index 0000000000..6c04086e0e
--- /dev/null
+++ b/test/batched_gemm_multi_d/test_batched_gemm_multi_d.cpp
@@ -0,0 +1,76 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <gtest/gtest.h>
+
+#include "profiler/profile_batched_gemm_impl.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp"
+
+namespace {
+using F16 = ck::half_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using Empty_Tuple = ck::Tuple<>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+template <typename Tuple>
+class TestBatchedGemmMultiD : public ::testing::Test
+{
+    protected:
+    using ALayout = std::tuple_element_t<0, Tuple>;
+    using BLayout = std::tuple_element_t<1, Tuple>;
+    using CLayout = std::tuple_element_t<2, Tuple>;
+
+    static constexpr int M          = 512;
+    static constexpr int N          = 256;
+    static constexpr int K          = 128;
+    static constexpr int BatchCount = 3;
+
+    template <typename DataType>
+    void Run()
+    {
+        using namespace ck::tensor_operation::device;
+
+        const bool pass =
+            ck::profiler::profile_batched_gemm_impl<DataType,
+                                                    DataType,
+                                                    DataType,
+                                                    ALayout,
+                                                    BLayout,
+                                                    CLayout,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    DeviceBatchedGemmMultiD<ALayout,
+                                                                            BLayout,
+                                                                            Empty_Tuple,
+                                                                            CLayout,
+                                                                            DataType,
+                                                                            DataType,
+                                                                            Empty_Tuple,
+                                                                            DataType,
+                                                                            PassThrough,
+                                                                            PassThrough,
+                                                                            PassThrough>>(
+                true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount);
+        EXPECT_TRUE(pass);
+    }
+};
+
+using KernelTypes = ::testing::Types<std::tuple<Row, Row, Row>,
+                                     std::tuple<Row, Col, Row>,
+                                     std::tuple<Col, Row, Row>,
+                                     std::tuple<Col, Col, Row>>;
+} // namespace
+
+TYPED_TEST_SUITE(TestBatchedGemmMultiD, KernelTypes);
+#ifdef __fp16
+TYPED_TEST(TestBatchedGemmMultiD, f16) { this->template Run<F16>(); }
+#endif
+#ifdef CK_ENABLE_INT8
+TYPED_TEST(TestBatchedGemmMultiD, int8) { this->template Run<int8_t>(); }
+#endif
diff --git a/test/batched_gemm_reduce/CMakeLists.txt b/test/batched_gemm_reduce/CMakeLists.txt
new file mode 100644
index 0000000000..af95a50eab
--- /dev/null
+++ b/test/batched_gemm_reduce/CMakeLists.txt
@@ -0,0 +1,12 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+     add_test_executable(test_batched_gemm_reduce_fp16 batched_gemm_reduce_fp16.cpp)
+     target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE utility)
+     target_link_libraries(test_batched_gemm_reduce_fp16 PRIVATE device_batched_gemm_reduce_instance)
+     set(target 1)
+   endif()
+ endif()
+endforeach()
diff --git a/test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp b/test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp
new file mode 100644
index 0000000000..dd2638ce89
--- /dev/null
+++ b/test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "profiler/profile_batched_gemm_reduce_impl.hpp"
+
+int main()
+{
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    int M = 512;
+    int N = 256;
+    int K = 128;
+
+    int BatchCount = 3;
+
+    bool pass = true;
+
+    pass = pass && ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                                  ck::half_t,
+                                                                  ck::half_t,
+                                                                  float,
+                                                                  Row,
+                                                                  Row,
+                                                                  Row>(
+                       true, 1, false, false, M, N, K, K, N, N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                                  ck::half_t,
+                                                                  ck::half_t,
+                                                                  float,
+                                                                  Row,
+                                                                  Col,
+                                                                  Row>(
+                       true, 1, false, false, M, N, K, K, K, N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                                  ck::half_t,
+                                                                  ck::half_t,
+                                                                  float,
+                                                                  Col,
+                                                                  Row,
+                                                                  Row>(
+                       true, 1, false, false, M, N, K, M, N, N, BatchCount);
+
+    pass = pass && ck::profiler::profile_batched_gemm_reduce_impl<ck::half_t,
+                                                                  ck::half_t,
+                                                                  ck::half_t,
+                                                                  float,
+                                                                  Col,
+                                                                  Col,
+                                                                  Row>(
+                       true, 1, false, false, M, N, K, M, K, N, BatchCount);
+
+    if(pass)
+    {
+        std::cout << "test BatchedGEMM+Reduce fp16: Pass" << std::endl;
+        return 0;
+    }
+    else
+    {
+        std::cout << "test BatchedGEMM+Reduce fp16: Fail" << std::endl;
+        return -1;
+    }
+}
diff --git a/test/batched_gemm_softmax_gemm/CMakeLists.txt b/test/batched_gemm_softmax_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..c49175a2eb
--- /dev/null
+++ b/test/batched_gemm_softmax_gemm/CMakeLists.txt
@@ -0,0 +1,13 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+     add_custom_target(test_batched_gemm_softmax_gemm)
+     add_gtest_executable(test_batched_gemm_softmax_gemm_fp16 test_batched_gemm_softmax_gemm_fp16.cpp)
+     target_link_libraries(test_batched_gemm_softmax_gemm_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_instance)
+     add_dependencies(test_batched_gemm_softmax_gemm test_batched_gemm_softmax_gemm_fp16)
+     set(target 1)
+   endif()
+ endif()
+endforeach()
\ No newline at end of file
diff --git a/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp b/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp
new file mode 100644
index 0000000000..cb46a995c6
--- /dev/null
+++ b/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp
@@ -0,0 +1,176 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "test_batched_gemm_softmax_gemm_util.hpp"
+
+template <typename Tuple>
+class TestBatchedGemmSoftmaxGemmFP16 : public TestBatchedGemmSoftmaxGemm<Tuple>
+{
+};
+
+using Masked = std::true_type;
+using NoMask = std::false_type;
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    std::tuple<F16, F16, F16, F16, Row, Col, Row, Row, NoMask>,
+    std::tuple<F16, F16, F16, F16, Row, Col, Row, Row, Masked>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestBatchedGemmSoftmaxGemmFP16, KernelTypes);
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16) { this->Run(); }
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_PadM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {136, 128, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_PadN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 136, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_PadK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 40, 128, 1},
+        {128, 128, 136, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_PadO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 136, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_OddM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {129, 128, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_OddN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 129, 32, 128, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_OddK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 33, 128, 1},
+        {128, 128, 129, 128, 1},
+    };
+    this->Run();
+}
+
+// If kernel B1Layout is RowMajor, expect not to support odd O size
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, Test_FP16_OddO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 129, 1},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, DISABLED_Bench_FP16)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {256, 256, 64, 64, 768},
+        {256, 256, 128, 128, 768},
+        {512, 512, 64, 64, 768},
+        {512, 512, 128, 128, 768},
+        {1024, 1024, 64, 64, 768},
+        {1024, 1024, 128, 128, 768},
+        {2048, 2048, 64, 64, 768},
+        {2048, 2048, 128, 128, 768},
+        {4096, 4096, 64, 64, 768},
+        {4096, 4096, 128, 128, 768},
+    };
+    this->bench_  = true;
+    this->verify_ = false;
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, DISABLED_Bench_FP16_IrregularK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{{256, 256, 160, 160, 16},
+                                                   {256, 64, 160, 64, 16},
+                                                   {1024, 1024, 80, 80, 16},
+                                                   {1024, 64, 80, 64, 16},
+                                                   {4096, 4096, 40, 40, 16},
+                                                   {4096, 64, 40, 64, 16}};
+    this->bench_   = true;
+    this->verify_  = false;
+    this->Run();
+}
+
+using ck::tensor_operation::device::GemmSpecialization;
+
+TEST(TestBatchedGemmSoftmaxGemmInterface, GemmSpecializationSizeMatch)
+{
+    int P = 120; // requires padding
+    int Q = 128; // do not require padding
+
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
+    // clang-format on
+}
+
+TEST(TestBatchedGemmSoftmaxGemmInterface, GemmSpecializationSizeMismatch)
+{
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
+    // Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
+    // Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
+    // clang-format on
+}
+
+TYPED_TEST(TestBatchedGemmSoftmaxGemmFP16, AdhocTest)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {49, 49, 64, 64, 24},
+        {64, 49, 64, 64, 24},
+        {1020, 1020, 64, 128, 24},
+        {576, 576, 64, 64, 24},
+    };
+    this->Run();
+}
diff --git a/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp b/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp
new file mode 100644
index 0000000000..d8ee744c60
--- /dev/null
+++ b/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp
@@ -0,0 +1,197 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include <vector>
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp"
+#include "profiler/profile_batched_gemm_softmax_gemm_impl.hpp"
+using ck::tensor_operation::device::GemmSpecialization;
+
+template <ck::index_t N>
+using I = ck::Number<N>;
+
+using F16 = ck::half_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <typename Tuple>
+struct TestBatchedGemmSoftmaxGemm : public ::testing::Test
+{
+    using ADataType   = std::tuple_element_t<0, Tuple>;
+    using B0DataType  = std::tuple_element_t<1, Tuple>;
+    using B1DataType  = std::tuple_element_t<2, Tuple>;
+    using CDataType   = std::tuple_element_t<3, Tuple>;
+    using ALayout     = std::tuple_element_t<4, Tuple>;
+    using B0Layout    = std::tuple_element_t<5, Tuple>;
+    using B1Layout    = std::tuple_element_t<6, Tuple>;
+    using CLayout     = std::tuple_element_t<7, Tuple>;
+    using MaskingType = std::tuple_element_t<8, Tuple>;
+
+    std::vector<std::vector<int>> lengths_ = {{256, 256, 64, 64, 4},
+                                              {256, 256, 128, 128, 4},
+                                              {512, 512, 64, 64, 2},
+                                              {512, 512, 128, 128, 2},
+                                              {1024, 1024, 64, 64, 1},
+                                              {1024, 1024, 128, 128, 1},
+                                              {256, 256, 160, 160, 4},
+                                              {256, 64, 160, 64, 4},
+                                              {1024, 1024, 80, 80, 2},
+                                              {1024, 64, 80, 64, 2},
+                                              {4096, 4096, 40, 40, 1},
+                                              {4096, 64, 40, 64, 1}};
+
+    bool bench_  = false;
+    bool verify_ = true;
+
+    void RunSingle(int M, int N, int K, int O, int BatchCount)
+    {
+        bool pass = ck::profiler::profile_batched_gemm_softmax_gemm_impl<ADataType,
+                                                                         B0DataType,
+                                                                         B1DataType,
+                                                                         CDataType,
+                                                                         ALayout,
+                                                                         B0Layout,
+                                                                         B1Layout,
+                                                                         CLayout,
+                                                                         MaskingType::value>(
+            verify_, 1, false, bench_, M, N, K, O, BatchCount);
+
+        EXPECT_TRUE(pass);
+    }
+
+    void Run()
+    {
+        for(auto lengths : this->lengths_)
+        {
+            int M          = lengths[0];
+            int N          = lengths[1];
+            int K          = lengths[2];
+            int O          = lengths[3];
+            int BatchCount = lengths[4];
+
+            this->RunSingle(M, N, K, O, BatchCount);
+        }
+    }
+};
+
+template <GemmSpecialization GemmSpec>
+struct DeviceInstanceWrapper_TNTT_FP16_M128_N128_K32_O128
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    using ALayout  = Row;
+    using B0Layout = Col;
+    using B1Layout = Row;
+    using CLayout  = Row;
+
+    using ADataType        = F16;
+    using B0DataType       = F16;
+    using B1DataType       = F16;
+    using AccDataType      = float;
+    using CShuffleDataType = float;
+    using CDataType        = F16;
+
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = PassThrough;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+
+    template <ck::index_t... Is>
+    using S = ck::Sequence<Is...>;
+
+    // static constexpr auto GemmSpec = std::tuple_element_t<0, Tuple>::value;
+
+    using DeviceGemmGemmInstance =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<
+            ALayout,
+            B0Layout,
+            B1Layout,
+            CLayout,
+            ADataType,
+            B0DataType,
+            B1DataType,
+            CDataType,
+            AccDataType,
+            CShuffleDataType,
+            AElementOp,
+            B0ElementOp,
+            Acc0ElementOp,
+            B1ElementOp,
+            CElementOp,
+            GemmSpec,
+            1,
+            256,
+            128,         // MPerBlock
+            128,         // NPerBlock
+            32,          // KPerBlock
+            128,         // Gemm1NPerBlock
+            32,          // Gemm1KPerBlock
+            8,           // AK1
+            8,           // BK1
+            2,           // B1K1
+            32,          // MPerXDL
+            32,          // NPerXDL
+            1,           // MXdlPerWave
+            4,           // NXdlPerWave
+            4,           // Gemm1NXdlPerWave
+            S<4, 64, 1>, // ABlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<4, 64, 1>, // BBlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<8, 32, 1>, // B1BlockTransfer
+            S<0, 2, 1>,
+            S<0, 2, 1>,
+            1,
+            4,
+            2,
+            false,
+            1,              // CShuffleMXdlPerWavePerShuffle
+            2,              // CShuffleNXdlPerWavePerShuffle
+            S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+            8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+            false>;
+
+    bool IsSupported(int M, int N, int K, int O)
+    {
+        auto gemm     = DeviceGemmGemmInstance{};
+        auto invoker  = gemm.MakeInvoker();
+        auto argument = gemm.MakeArgument(static_cast<ADataType*>(nullptr),
+                                          static_cast<B0DataType*>(nullptr),
+                                          static_cast<B1DataType*>(nullptr),
+                                          static_cast<CDataType*>(nullptr),
+                                          M,
+                                          N,
+                                          K,
+                                          O,
+                                          0,              // BatchCount
+                                          0,              // StrideA
+                                          0,              // StrideB0
+                                          0,              // StrideB1
+                                          0,              // StrideC
+                                          0,              // BatchStrideA
+                                          0,              // BatchStrideB0
+                                          0,              // BatchStrideB1
+                                          0,              // BatchStrideC
+                                          PassThrough{},  // a_element_op
+                                          PassThrough{},  // b0_element_op
+                                          PassThrough{},  // acc0_element_op
+                                          PassThrough{},  // b1_element_op
+                                          PassThrough{}); // c_element_op
+
+        return gemm.IsSupportedArgument(argument);
+    }
+};
diff --git a/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt b/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt
new file mode 100644
index 0000000000..ae28daa80e
--- /dev/null
+++ b/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt
@@ -0,0 +1,26 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   if(DTYPES MATCHES "fp16" OR DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+     add_custom_target(test_batched_gemm_softmax_gemm_permute)
+   endif()
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+     add_gtest_executable(test_batched_gemm_softmax_gemm_permute_fp16 test_batched_gemm_softmax_gemm_permute_fp16.cpp)
+     add_gtest_executable(test_batched_gemm_bias_softmax_gemm_permute_fp16 test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp)
+     target_link_libraries(test_batched_gemm_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
+     target_link_libraries(test_batched_gemm_bias_softmax_gemm_permute_fp16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
+     add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_fp16)
+     add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_bias_softmax_gemm_permute_fp16)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+     add_gtest_executable(test_batched_gemm_softmax_gemm_permute_bf16 test_batched_gemm_softmax_gemm_permute_bf16.cpp)
+     add_gtest_executable(test_batched_gemm_bias_softmax_gemm_permute_bf16 test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp)
+     target_link_libraries(test_batched_gemm_softmax_gemm_permute_bf16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
+     target_link_libraries(test_batched_gemm_bias_softmax_gemm_permute_bf16 PRIVATE utility device_batched_gemm_softmax_gemm_permute_instance)
+     add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_softmax_gemm_permute_bf16)
+     add_dependencies(test_batched_gemm_softmax_gemm_permute test_batched_gemm_bias_softmax_gemm_permute_bf16)
+   endif()
+   set(target 1)
+ endif()
+endforeach()
\ No newline at end of file
diff --git a/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp
new file mode 100644
index 0000000000..ef88ce6d81
--- /dev/null
+++ b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp
@@ -0,0 +1,182 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "test_batched_gemm_bias_softmax_gemm_permute_util.hpp"
+
+template <typename Tuple>
+class TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16
+    : public TestBatchedGemmMaskingScaleSoftmaxGemmPermute<Tuple>
+{
+};
+
+using I1_t = ck::Number<1>;
+using I2_t = ck::Number<2>;
+
+using MaskDisabled_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskDisabled>;
+using MaskOutUpperTriangle_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskOutUpperTriangle>;
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, BF16, BF16, BF16, BF16, ck::Tuple<BF16>, ck::Tuple<>, MaskDisabled_t>,
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, BF16, BF16, BF16, BF16, ck::Tuple<BF16>, ck::Tuple<>, MaskOutUpperTriangle_t>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, KernelTypes);
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, DISABLED_Test_BF16) { this->Run(); }
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {136, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 136, 32, 128, 3, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 40, 128, 2, 4},
+        {128, 128, 136, 128, 4, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 136, 1, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {129, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 129, 32, 128, 4, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 33, 128, 2, 3},
+        {128, 128, 129, 128, 2, 3},
+    };
+    this->Run();
+}
+
+// If kernel B1Layout is RowMajor, expect not to support odd O size
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 129, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, DISABLED_Bench_BF16_IrregularK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{{256, 256, 160, 160, 1, 16},
+                                                   {256, 64, 160, 64, 1, 16},
+                                                   {1024, 1024, 80, 80, 1, 16},
+                                                   {1024, 64, 80, 64, 1, 16},
+                                                   {4096, 4096, 40, 40, 1, 16},
+                                                   {4096, 64, 40, 64, 1, 16}};
+    this->bench_   = true;
+    this->verify_  = false;
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, DISABLED_Bench_BF16)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {256, 256, 64, 64, 48, 16},
+        {256, 256, 128, 128, 48, 16},
+        {512, 512, 64, 64, 48, 16},
+        {512, 512, 128, 128, 48, 16},
+        {1024, 1024, 64, 64, 48, 16},
+        {1024, 1024, 128, 128, 48, 16},
+        {2048, 2048, 64, 64, 48, 16},
+        {2048, 2048, 128, 128, 48, 16},
+        {4096, 4096, 64, 64, 48, 16},
+        {4096, 4096, 128, 128, 48, 16},
+    };
+    this->bench_  = true;
+    this->verify_ = false;
+    this->Run();
+}
+
+using ck::tensor_operation::device::GemmSpecialization;
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMatch)
+{
+    int P = 120; // requires padding
+    int Q = 128; // do not require padding
+
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
+    // clang-format on
+}
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMismatch)
+{
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
+    // Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
+    // Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
+    // clang-format on
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, AdhocTest)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {49, 49, 64, 64, 4, 6},
+        {64, 49, 64, 64, 4, 6},
+        {1020, 1020, 64, 128, 4, 6},
+        {576, 576, 64, 64, 4, 6},
+    };
+    this->Run();
+}
diff --git a/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp
new file mode 100644
index 0000000000..b38b10d195
--- /dev/null
+++ b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp
@@ -0,0 +1,182 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "test_batched_gemm_softmax_gemm_permute_util.hpp"
+
+template <typename Tuple>
+class TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16
+    : public TestBatchedGemmMaskingScaleSoftmaxGemmPermute<Tuple>
+{
+};
+
+using I1_t = ck::Number<1>;
+using I2_t = ck::Number<2>;
+
+using MaskDisabled_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskDisabled>;
+using MaskOutUpperTriangle_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskOutUpperTriangle>;
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, F16, F16, F16, F16, ck::Tuple<F16>, ck::Tuple<>, MaskDisabled_t>,
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, F16, F16, F16, F16, ck::Tuple<F16>, ck::Tuple<>, MaskOutUpperTriangle_t>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, KernelTypes);
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16) { this->Run(); }
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {136, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 136, 32, 128, 3, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 40, 128, 2, 4},
+        {128, 128, 136, 128, 4, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 136, 1, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {129, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 129, 32, 128, 4, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 33, 128, 2, 3},
+        {128, 128, 129, 128, 2, 3},
+    };
+    this->Run();
+}
+
+// If kernel B1Layout is RowMajor, expect not to support odd O size
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 129, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, DISABLED_Bench_FP16_IrregularK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{{256, 256, 160, 160, 1, 16},
+                                                   {256, 64, 160, 64, 1, 16},
+                                                   {1024, 1024, 80, 80, 1, 16},
+                                                   {1024, 64, 80, 64, 1, 16},
+                                                   {4096, 4096, 40, 40, 1, 16},
+                                                   {4096, 64, 40, 64, 1, 16}};
+    this->bench_   = true;
+    this->verify_  = false;
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, DISABLED_Bench_FP16)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {256, 256, 64, 64, 48, 16},
+        {256, 256, 128, 128, 48, 16},
+        {512, 512, 64, 64, 48, 16},
+        {512, 512, 128, 128, 48, 16},
+        {1024, 1024, 64, 64, 48, 16},
+        {1024, 1024, 128, 128, 48, 16},
+        {2048, 2048, 64, 64, 48, 16},
+        {2048, 2048, 128, 128, 48, 16},
+        {4096, 4096, 64, 64, 48, 16},
+        {4096, 4096, 128, 128, 48, 16},
+    };
+    this->bench_  = true;
+    this->verify_ = false;
+    this->Run();
+}
+
+using ck::tensor_operation::device::GemmSpecialization;
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMatch)
+{
+    int P = 120; // requires padding
+    int Q = 128; // do not require padding
+
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
+    // clang-format on
+}
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMismatch)
+{
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
+    // Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
+    // Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
+    // clang-format on
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, AdhocTest)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {49, 49, 64, 64, 4, 6},
+        {64, 49, 64, 64, 4, 6},
+        {1020, 1020, 64, 128, 4, 6},
+        {576, 576, 64, 64, 4, 6},
+    };
+    this->Run();
+}
diff --git a/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_util.hpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_util.hpp
new file mode 100644
index 0000000000..d7c39367c8
--- /dev/null
+++ b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_util.hpp
@@ -0,0 +1,380 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include <vector>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp"
+
+using ck::tensor_operation::device::GemmSpecialization;
+using ck::tensor_operation::device::MaskingSpecialization;
+using ck::tensor_operation::device::TensorSpecialization;
+
+template <ck::index_t N>
+using I = ck::Number<N>;
+
+using F16  = ck::half_t;
+using BF16 = ck::bhalf_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <typename Tuple>
+struct TestBatchedGemmMaskingScaleSoftmaxGemmPermute : public ::testing::Test
+{
+    using NumDimGType      = std::tuple_element_t<0, Tuple>;
+    using NumDimMType      = std::tuple_element_t<1, Tuple>;
+    using NumDimNType      = std::tuple_element_t<2, Tuple>;
+    using NumDimKType      = std::tuple_element_t<3, Tuple>;
+    using NumDimOType      = std::tuple_element_t<4, Tuple>;
+    using ADataType        = std::tuple_element_t<5, Tuple>;
+    using B0DataType       = std::tuple_element_t<6, Tuple>;
+    using B1DataType       = std::tuple_element_t<7, Tuple>;
+    using CDataType        = std::tuple_element_t<8, Tuple>;
+    using Acc0BiasDataType = std::tuple_element_t<9, Tuple>;
+    using Acc1BiasDataType = std::tuple_element_t<10, Tuple>;
+    using MaskingType      = std::tuple_element_t<11, Tuple>;
+
+    std::vector<std::vector<int>> lengths_ = {
+        {256, 256, 64, 64, 6, 4},
+        {256, 256, 128, 128, 4, 6},
+        {512, 512, 64, 64, 3, 2},
+        {512, 512, 128, 128, 2, 3},
+        {1024, 1024, 64, 64, 3, 1},
+        {1024, 1024, 128, 128, 1, 1},
+    };
+    bool bench_  = false;
+    bool verify_ = true;
+
+    void RunSingle(int M, int N, int K, int O, int G0, int G1)
+    {
+        bool pass =
+            ck::profiler::profile_batched_gemm_bias_softmax_gemm_permute_impl<NumDimGType::value,
+                                                                              NumDimMType::value,
+                                                                              NumDimNType::value,
+                                                                              NumDimKType::value,
+                                                                              NumDimOType::value,
+                                                                              ADataType,
+                                                                              B0DataType,
+                                                                              B1DataType,
+                                                                              CDataType,
+                                                                              Acc0BiasDataType,
+                                                                              Acc1BiasDataType,
+                                                                              MaskingType::value>(
+                verify_, 2, false, bench_, M, N, K, O, G0, G1);
+
+        EXPECT_TRUE(pass);
+    }
+
+    void Run()
+    {
+        for(auto lengths : this->lengths_)
+        {
+            int M  = lengths[0];
+            int N  = lengths[1];
+            int K  = lengths[2];
+            int O  = lengths[3];
+            int G0 = lengths[4];
+            int G1 = lengths[5];
+
+            this->RunSingle(M, N, K, O, G0, G1);
+        }
+    }
+};
+
+template <GemmSpecialization GemmSpec>
+struct DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using ScaleAdd    = ck::tensor_operation::element_wise::ScaleAdd;
+
+    template <ck::index_t... Is>
+    using S = ck::Sequence<Is...>;
+
+    using ADataType        = F16;
+    using B0DataType       = F16;
+    using B1DataType       = F16;
+    using AccDataType      = float;
+    using CShuffleDataType = F16;
+    using CDataType        = F16;
+
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = ScaleAdd;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+
+    // static constexpr auto GemmSpec = std::tuple_element_t<0, Tuple>::value;
+
+    using DeviceGemmGemmInstance =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+            2,
+            1,
+            1,
+            1,
+            1,
+            ADataType,
+            B0DataType,
+            B1DataType,
+            CDataType,
+            ck::Tuple<F16>,
+            ck::Tuple<>,
+            AccDataType,
+            CShuffleDataType,
+            AElementOp,
+            B0ElementOp,
+            Acc0ElementOp,
+            B1ElementOp,
+            CElementOp,
+            GemmSpec,
+            TensorSpecialization::Default, // ATensorSpec
+            TensorSpecialization::Default, // B0TensorSpec
+            TensorSpecialization::Default, // B1TensorSpec
+            TensorSpecialization::Default, // CTensorSpec
+            1,
+            256,
+            128,         // MPerBlock
+            128,         // NPerBlock
+            32,          // KPerBlock
+            128,         // Gemm1NPerBlock
+            32,          // Gemm1KPerBlock
+            8,           // AK1
+            8,           // BK1
+            2,           // B1K1
+            32,          // MPerXDL
+            32,          // NPerXDL
+            1,           // MXdlPerWave
+            4,           // NXdlPerWave
+            4,           // Gemm1NXdlPerWave
+            S<4, 64, 1>, // ABlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<4, 64, 1>, // BBlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<8, 32, 1>, // B1BlockTransfer
+            S<0, 2, 1>,
+            S<0, 2, 1>,
+            1,
+            4,
+            2,
+            false,
+            1,              // CShuffleMXdlPerWavePerShuffle
+            2,              // CShuffleNXdlPerWavePerShuffle
+            S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+            8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+            MaskingSpecialization::MaskOutUpperTriangle>; // MaskOutUpperTriangle
+
+    bool IsSupported(int M, int N, int K, int O)
+    {
+        const int G0 = 1, G1 = 1;
+
+        // A layout [G0, M, G1, K]
+        std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+        std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+        // B0 layout [G0, N, G1, K]
+        std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+        std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+        // B1 layout [G0, N, G1, O]
+        std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+        std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+        // C layout [G0, M, G1, O]
+        std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+        std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+        // D layout [G0, M, G1, N]
+        std::vector<ck::index_t> d0_gs_ms_ns_lengths{G0, G1, M, N};
+        std::vector<ck::index_t> d0_gs_ms_ns_strides{M * G1 * N, N, G1 * N, 1};
+
+        auto gemm     = DeviceGemmGemmInstance{};
+        auto invoker  = gemm.MakeInvoker();
+        auto argument = gemm.MakeArgument(static_cast<ADataType*>(nullptr),
+                                          static_cast<B0DataType*>(nullptr),
+                                          static_cast<B1DataType*>(nullptr),
+                                          static_cast<CDataType*>(nullptr),
+                                          std::array<void*, 1>{nullptr}, // p_acc0_biases
+                                          {},                            // p_acc1_biases
+                                          a_gs_ms_ks_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b0_gs_ns_ks_lengths,
+                                          b0_gs_ns_ks_strides,
+                                          b1_gs_os_ns_lengths,
+                                          b1_gs_os_ns_strides,
+                                          c_gs_ms_os_lengths,
+                                          c_gs_ms_os_strides,
+                                          std::array<std::vector<ck::index_t>, 1>{
+                                              d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+                                          std::array<std::vector<ck::index_t>, 1>{
+                                              d0_gs_ms_ns_strides}, // acc0_biases_gs_ms_ns_strides
+                                          {},                       // acc1_biases_gs_ms_os_lengths
+                                          {},                       // acc1_biases_gs_ms_os_strides
+                                          PassThrough{},            // a_element_op
+                                          PassThrough{},            // b0_element_op
+                                          Acc0ElementOp{1.f},       // acc0_element_op
+                                          PassThrough{},            // b1_element_op
+                                          PassThrough{});           // c_element_op
+
+        return gemm.IsSupportedArgument(argument);
+    }
+};
+
+template <GemmSpecialization GemmSpec>
+struct DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using ScaleAdd    = ck::tensor_operation::element_wise::ScaleAdd;
+
+    template <ck::index_t... Is>
+    using S = ck::Sequence<Is...>;
+
+    using ADataType        = BF16;
+    using B0DataType       = BF16;
+    using B1DataType       = BF16;
+    using AccDataType      = float;
+    using CShuffleDataType = BF16;
+    using CDataType        = BF16;
+
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = ScaleAdd;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+
+    // static constexpr auto GemmSpec = std::tuple_element_t<0, Tuple>::value;
+
+    using DeviceGemmGemmInstance =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+            2,
+            1,
+            1,
+            1,
+            1,
+            ADataType,
+            B0DataType,
+            B1DataType,
+            CDataType,
+            ck::Tuple<BF16>,
+            ck::Tuple<>,
+            AccDataType,
+            CShuffleDataType,
+            AElementOp,
+            B0ElementOp,
+            Acc0ElementOp,
+            B1ElementOp,
+            CElementOp,
+            GemmSpec,
+            TensorSpecialization::Default, // ATensorSpec
+            TensorSpecialization::Default, // B0TensorSpec
+            TensorSpecialization::Default, // B1TensorSpec
+            TensorSpecialization::Default, // CTensorSpec
+            1,
+            256,
+            128,         // MPerBlock
+            128,         // NPerBlock
+            32,          // KPerBlock
+            128,         // Gemm1NPerBlock
+            32,          // Gemm1KPerBlock
+            8,           // AK1
+            8,           // BK1
+            2,           // B1K1
+            32,          // MPerXDL
+            32,          // NPerXDL
+            1,           // MXdlPerWave
+            4,           // NXdlPerWave
+            4,           // Gemm1NXdlPerWave
+            S<4, 64, 1>, // ABlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<4, 64, 1>, // BBlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<8, 32, 1>, // B1BlockTransfer
+            S<0, 2, 1>,
+            S<0, 2, 1>,
+            1,
+            4,
+            2,
+            false,
+            1,              // CShuffleMXdlPerWavePerShuffle
+            2,              // CShuffleNXdlPerWavePerShuffle
+            S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+            8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+            MaskingSpecialization::MaskOutUpperTriangle>; // MaskOutUpperTriangle
+
+    bool IsSupported(int M, int N, int K, int O)
+    {
+        const int G0 = 1, G1 = 1;
+
+        // A layout [G0, M, G1, K]
+        std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+        std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+        // B0 layout [G0, N, G1, K]
+        std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+        std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+        // B1 layout [G0, N, G1, O]
+        std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+        std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+        // C layout [G0, M, G1, O]
+        std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+        std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+        // D layout [G0, M, G1, N]
+        std::vector<ck::index_t> d0_gs_ms_ns_lengths{G0, G1, M, N};
+        std::vector<ck::index_t> d0_gs_ms_ns_strides{M * G1 * N, N, G1 * N, 1};
+
+        auto gemm     = DeviceGemmGemmInstance{};
+        auto invoker  = gemm.MakeInvoker();
+        auto argument = gemm.MakeArgument(static_cast<ADataType*>(nullptr),
+                                          static_cast<B0DataType*>(nullptr),
+                                          static_cast<B1DataType*>(nullptr),
+                                          static_cast<CDataType*>(nullptr),
+                                          std::array<void*, 1>{nullptr}, // p_acc0_biases
+                                          {},                            // p_acc1_biases
+                                          a_gs_ms_ks_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b0_gs_ns_ks_lengths,
+                                          b0_gs_ns_ks_strides,
+                                          b1_gs_os_ns_lengths,
+                                          b1_gs_os_ns_strides,
+                                          c_gs_ms_os_lengths,
+                                          c_gs_ms_os_strides,
+                                          std::array<std::vector<ck::index_t>, 1>{
+                                              d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+                                          std::array<std::vector<ck::index_t>, 1>{
+                                              d0_gs_ms_ns_strides}, // acc0_biases_gs_ms_ns_strides
+                                          {},                       // acc1_biases_gs_ms_os_lengths
+                                          {},                       // acc1_biases_gs_ms_os_strides
+                                          PassThrough{},            // a_element_op
+                                          PassThrough{},            // b0_element_op
+                                          Acc0ElementOp{1.f},       // acc0_element_op
+                                          PassThrough{},            // b1_element_op
+                                          PassThrough{});           // c_element_op
+
+        return gemm.IsSupportedArgument(argument);
+    }
+};
diff --git a/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp
new file mode 100644
index 0000000000..8e0baede11
--- /dev/null
+++ b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp
@@ -0,0 +1,182 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "test_batched_gemm_softmax_gemm_permute_util.hpp"
+
+template <typename Tuple>
+class TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16
+    : public TestBatchedGemmMaskingScaleSoftmaxGemmPermute<Tuple>
+{
+};
+
+using I1_t = ck::Number<1>;
+using I2_t = ck::Number<2>;
+
+using MaskDisabled_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskDisabled>;
+using MaskOutUpperTriangle_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskOutUpperTriangle>;
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, MaskDisabled_t>,
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, BF16, BF16, BF16, BF16, ck::Tuple<>, ck::Tuple<>, MaskOutUpperTriangle_t>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, KernelTypes);
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16) { this->Run(); }
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {136, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 136, 32, 128, 3, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 40, 128, 2, 4},
+        {128, 128, 136, 128, 4, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_PadO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 136, 1, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {129, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 129, 32, 128, 4, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 33, 128, 2, 3},
+        {128, 128, 129, 128, 2, 3},
+    };
+    this->Run();
+}
+
+// If kernel B1Layout is RowMajor, expect not to support odd O size
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Test_BF16_OddO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 129, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Bench_BF16_IrregularK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{{256, 256, 160, 160, 1, 16},
+                                                   {256, 64, 160, 64, 1, 16},
+                                                   {1024, 1024, 80, 80, 1, 16},
+                                                   {1024, 64, 80, 64, 1, 16},
+                                                   {4096, 4096, 40, 40, 1, 16},
+                                                   {4096, 64, 40, 64, 1, 16}};
+    this->bench_   = true;
+    this->verify_  = false;
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, Bench_BF16)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {256, 256, 64, 64, 48, 16},
+        {256, 256, 128, 128, 48, 16},
+        {512, 512, 64, 64, 48, 16},
+        {512, 512, 128, 128, 48, 16},
+        {1024, 1024, 64, 64, 48, 16},
+        {1024, 1024, 128, 128, 48, 16},
+        {2048, 2048, 64, 64, 48, 16},
+        {2048, 2048, 128, 128, 48, 16},
+        {4096, 4096, 64, 64, 48, 16},
+        {4096, 4096, 128, 128, 48, 16},
+    };
+    this->bench_  = true;
+    this->verify_ = false;
+    this->Run();
+}
+
+using ck::tensor_operation::device::GemmSpecialization;
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMatch)
+{
+    int P = 120; // requires padding
+    int Q = 128; // do not require padding
+
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
+    // clang-format on
+}
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMismatch)
+{
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
+    // Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
+    // Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
+    // clang-format on
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteBF16, AdhocTest)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {49, 49, 64, 64, 4, 6},
+        {64, 49, 64, 64, 4, 6},
+        {1020, 1020, 64, 128, 4, 6},
+        {576, 576, 64, 64, 4, 6},
+    };
+    this->Run();
+}
diff --git a/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp
new file mode 100644
index 0000000000..81d404109f
--- /dev/null
+++ b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp
@@ -0,0 +1,182 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "test_batched_gemm_softmax_gemm_permute_util.hpp"
+
+template <typename Tuple>
+class TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16
+    : public TestBatchedGemmMaskingScaleSoftmaxGemmPermute<Tuple>
+{
+};
+
+using I1_t = ck::Number<1>;
+using I2_t = ck::Number<2>;
+
+using MaskDisabled_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskDisabled>;
+using MaskOutUpperTriangle_t =
+    ck::integral_constant<MaskingSpecialization, MaskingSpecialization::MaskOutUpperTriangle>;
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, MaskDisabled_t>,
+    std::tuple<I2_t, I1_t, I1_t, I1_t, I1_t, F16, F16, F16, F16, ck::Tuple<>, ck::Tuple<>, MaskOutUpperTriangle_t>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, KernelTypes);
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16) { this->Run(); }
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {136, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 136, 32, 128, 3, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 40, 128, 2, 4},
+        {128, 128, 136, 128, 4, 2},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_PadO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 136, 1, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddM)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {129, 128, 32, 128, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddN)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 129, 32, 128, 4, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 33, 128, 2, 3},
+        {128, 128, 129, 128, 2, 3},
+    };
+    this->Run();
+}
+
+// If kernel B1Layout is RowMajor, expect not to support odd O size
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, Test_FP16_OddO)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {128, 128, 32, 129, 2, 3},
+    };
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, DISABLED_Bench_FP16_IrregularK)
+{
+    this->lengths_ = std::vector<std::vector<int>>{{256, 256, 160, 160, 1, 16},
+                                                   {256, 64, 160, 64, 1, 16},
+                                                   {1024, 1024, 80, 80, 1, 16},
+                                                   {1024, 64, 80, 64, 1, 16},
+                                                   {4096, 4096, 40, 40, 1, 16},
+                                                   {4096, 64, 40, 64, 1, 16}};
+    this->bench_   = true;
+    this->verify_  = false;
+    this->Run();
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, DISABLED_Bench_FP16)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {256, 256, 64, 64, 48, 16},
+        {256, 256, 128, 128, 48, 16},
+        {512, 512, 64, 64, 48, 16},
+        {512, 512, 128, 128, 48, 16},
+        {1024, 1024, 64, 64, 48, 16},
+        {1024, 1024, 128, 128, 48, 16},
+        {2048, 2048, 64, 64, 48, 16},
+        {2048, 2048, 128, 128, 48, 16},
+        {4096, 4096, 64, 64, 48, 16},
+        {4096, 4096, 128, 128, 48, 16},
+    };
+    this->bench_  = true;
+    this->verify_ = false;
+    this->Run();
+}
+
+using ck::tensor_operation::device::GemmSpecialization;
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMatch)
+{
+    int P = 120; // requires padding
+    int Q = 128; // do not require padding
+
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(Q, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MPadding>{}.IsSupported(P, Q, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NPadding>{}.IsSupported(Q, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KPadding>{}.IsSupported(Q, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNPadding>{}.IsSupported(P, P, Q, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKPadding>{}.IsSupported(P, Q, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKPadding>{}.IsSupported(Q, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(P, P, P, Q));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::OPadding>{}.IsSupported(Q, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MOPadding>{}.IsSupported(P, Q, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NOPadding>{}.IsSupported(Q, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::KOPadding>{}.IsSupported(Q, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNOPadding>{}.IsSupported(P, P, Q, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MKOPadding>{}.IsSupported(P, Q, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::NKOPadding>{}.IsSupported(Q, P, P, P));
+    EXPECT_TRUE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(P, P, P, P));
+    // clang-format on
+}
+
+TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteInterface, GemmSpecializationSizeMismatch)
+{
+    // IsSupported(M, N, K, O)
+    // clang-format off
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::Default>{}.IsSupported(128, 128, 120, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKPadding>{}.IsSupported(128, 128, 128, 120));
+    // Kernel can't support odd K size because SrcVectorDim == KDim and must satisfy SizeKRaw % ABSrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 129, 128));
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 130, 128));
+    // Kernel can't support odd O size because SrcVectorDim == ODim and must satisfy SizeORaw % B1SrcScalarPerVector == 0
+    EXPECT_FALSE(DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128<GemmSpecialization::MNKOPadding>{}.IsSupported(128, 128, 128, 129));
+    // clang-format on
+}
+
+TYPED_TEST(TestBatchedGemmMaskingScaleSoftmaxGemmPermuteFP16, AdhocTest)
+{
+    this->lengths_ = std::vector<std::vector<int>>{
+        {49, 49, 64, 64, 4, 6},
+        {64, 49, 64, 64, 4, 6},
+        {1020, 1020, 64, 128, 4, 6},
+        {576, 576, 64, 64, 4, 6},
+    };
+    this->Run();
+}
diff --git a/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp
new file mode 100644
index 0000000000..9df03ffd2a
--- /dev/null
+++ b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp
@@ -0,0 +1,368 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include <vector>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp"
+
+using ck::tensor_operation::device::GemmSpecialization;
+using ck::tensor_operation::device::MaskingSpecialization;
+using ck::tensor_operation::device::TensorSpecialization;
+
+template <ck::index_t N>
+using I = ck::Number<N>;
+
+using F16  = ck::half_t;
+using BF16 = ck::bhalf_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <typename Tuple>
+struct TestBatchedGemmMaskingScaleSoftmaxGemmPermute : public ::testing::Test
+{
+    using NumDimGType      = std::tuple_element_t<0, Tuple>;
+    using NumDimMType      = std::tuple_element_t<1, Tuple>;
+    using NumDimNType      = std::tuple_element_t<2, Tuple>;
+    using NumDimKType      = std::tuple_element_t<3, Tuple>;
+    using NumDimOType      = std::tuple_element_t<4, Tuple>;
+    using ADataType        = std::tuple_element_t<5, Tuple>;
+    using B0DataType       = std::tuple_element_t<6, Tuple>;
+    using B1DataType       = std::tuple_element_t<7, Tuple>;
+    using CDataType        = std::tuple_element_t<8, Tuple>;
+    using Acc0BiasDataType = std::tuple_element_t<9, Tuple>;
+    using Acc1BiasDataType = std::tuple_element_t<10, Tuple>;
+    using MaskingType      = std::tuple_element_t<11, Tuple>;
+
+    std::vector<std::vector<int>> lengths_ = {
+        {256, 256, 64, 64, 6, 4},
+        {256, 256, 128, 128, 4, 6},
+        {512, 512, 64, 64, 3, 2},
+        {512, 512, 128, 128, 2, 3},
+        {1024, 1024, 64, 64, 3, 1},
+        {1024, 1024, 128, 128, 1, 1},
+    };
+    bool bench_  = false;
+    bool verify_ = true;
+
+    void RunSingle(int M, int N, int K, int O, int G0, int G1)
+    {
+        bool pass =
+            ck::profiler::profile_batched_gemm_softmax_gemm_permute_impl<NumDimGType::value,
+                                                                         NumDimMType::value,
+                                                                         NumDimNType::value,
+                                                                         NumDimKType::value,
+                                                                         NumDimOType::value,
+                                                                         ADataType,
+                                                                         B0DataType,
+                                                                         B1DataType,
+                                                                         CDataType,
+                                                                         ck::Tuple<>,
+                                                                         ck::Tuple<>,
+                                                                         MaskingType::value>(
+                verify_, 2, false, bench_, M, N, K, O, G0, G1);
+
+        EXPECT_TRUE(pass);
+    }
+
+    void Run()
+    {
+        for(auto lengths : this->lengths_)
+        {
+            int M  = lengths[0];
+            int N  = lengths[1];
+            int K  = lengths[2];
+            int O  = lengths[3];
+            int G0 = lengths[4];
+            int G1 = lengths[5];
+
+            this->RunSingle(M, N, K, O, G0, G1);
+        }
+    }
+};
+
+template <GemmSpecialization GemmSpec>
+struct DeviceInstanceWrapper_G2M1N1K1O1_TNTT_FP16_M128_N128_K32_O128
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using Scale       = ck::tensor_operation::element_wise::Scale;
+
+    template <ck::index_t... Is>
+    using S = ck::Sequence<Is...>;
+
+    using ADataType        = F16;
+    using B0DataType       = F16;
+    using B1DataType       = F16;
+    using AccDataType      = float;
+    using CShuffleDataType = F16;
+    using CDataType        = F16;
+
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = Scale;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+
+    // static constexpr auto GemmSpec = std::tuple_element_t<0, Tuple>::value;
+
+    using DeviceGemmGemmInstance =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+            2,
+            1,
+            1,
+            1,
+            1,
+            ADataType,
+            B0DataType,
+            B1DataType,
+            CDataType,
+            ck::Tuple<>,
+            ck::Tuple<>,
+            AccDataType,
+            CShuffleDataType,
+            AElementOp,
+            B0ElementOp,
+            Acc0ElementOp,
+            B1ElementOp,
+            CElementOp,
+            GemmSpec,
+            TensorSpecialization::Default, // ATensorSpec
+            TensorSpecialization::Default, // B0TensorSpec
+            TensorSpecialization::Default, // B1TensorSpec
+            TensorSpecialization::Default, // CTensorSpec
+            1,
+            256,
+            128,         // MPerBlock
+            128,         // NPerBlock
+            32,          // KPerBlock
+            128,         // Gemm1NPerBlock
+            32,          // Gemm1KPerBlock
+            8,           // AK1
+            8,           // BK1
+            2,           // B1K1
+            32,          // MPerXDL
+            32,          // NPerXDL
+            1,           // MXdlPerWave
+            4,           // NXdlPerWave
+            4,           // Gemm1NXdlPerWave
+            S<4, 64, 1>, // ABlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<4, 64, 1>, // BBlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<8, 32, 1>, // B1BlockTransfer
+            S<0, 2, 1>,
+            S<0, 2, 1>,
+            1,
+            4,
+            2,
+            false,
+            1,              // CShuffleMXdlPerWavePerShuffle
+            2,              // CShuffleNXdlPerWavePerShuffle
+            S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+            8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+            MaskingSpecialization::MaskOutUpperTriangle>; // MaskOutUpperTriangle
+
+    bool IsSupported(int M, int N, int K, int O)
+    {
+        const int G0 = 1, G1 = 1;
+
+        // A layout [G0, M, G1, K]
+        std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+        std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+        // B0 layout [G0, N, G1, K]
+        std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+        std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+        // B1 layout [G0, N, G1, O]
+        std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+        std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+        // C layout [G0, M, G1, O]
+        std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+        std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+        auto gemm     = DeviceGemmGemmInstance{};
+        auto invoker  = gemm.MakeInvoker();
+        auto argument = gemm.MakeArgument(static_cast<ADataType*>(nullptr),
+                                          static_cast<B0DataType*>(nullptr),
+                                          static_cast<B1DataType*>(nullptr),
+                                          static_cast<CDataType*>(nullptr),
+                                          {}, // p_acc0_biases
+                                          {}, // p_acc1_biases
+                                          a_gs_ms_ks_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b0_gs_ns_ks_lengths,
+                                          b0_gs_ns_ks_strides,
+                                          b1_gs_os_ns_lengths,
+                                          b1_gs_os_ns_strides,
+                                          c_gs_ms_os_lengths,
+                                          c_gs_ms_os_strides,
+                                          {},             // acc0_biases_gs_ms_ns_lengths
+                                          {},             // acc0_biases_gs_ms_ns_strides
+                                          {},             // acc1_biases_gs_ms_os_lengths
+                                          {},             // acc1_biases_gs_ms_os_strides
+                                          PassThrough{},  // a_element_op
+                                          PassThrough{},  // b0_element_op
+                                          Scale{1.f},     // acc0_element_op
+                                          PassThrough{},  // b1_element_op
+                                          PassThrough{}); // c_element_op
+
+        return gemm.IsSupportedArgument(argument);
+    }
+};
+
+template <GemmSpecialization GemmSpec>
+struct DeviceInstanceWrapper_G2M1N1K1O1_TNTT_BF16_M128_N128_K32_O128
+{
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using Scale       = ck::tensor_operation::element_wise::Scale;
+
+    template <ck::index_t... Is>
+    using S = ck::Sequence<Is...>;
+
+    using ADataType        = BF16;
+    using B0DataType       = BF16;
+    using B1DataType       = BF16;
+    using AccDataType      = float;
+    using CShuffleDataType = BF16;
+    using CDataType        = BF16;
+
+    using AElementOp    = PassThrough;
+    using B0ElementOp   = PassThrough;
+    using Acc0ElementOp = Scale;
+    using B1ElementOp   = PassThrough;
+    using CElementOp    = PassThrough;
+
+    // static constexpr auto GemmSpec = std::tuple_element_t<0, Tuple>::value;
+
+    using DeviceGemmGemmInstance =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+            2,
+            1,
+            1,
+            1,
+            1,
+            ADataType,
+            B0DataType,
+            B1DataType,
+            CDataType,
+            ck::Tuple<>,
+            ck::Tuple<>,
+            AccDataType,
+            CShuffleDataType,
+            AElementOp,
+            B0ElementOp,
+            Acc0ElementOp,
+            B1ElementOp,
+            CElementOp,
+            GemmSpec,
+            TensorSpecialization::Default, // ATensorSpec
+            TensorSpecialization::Default, // B0TensorSpec
+            TensorSpecialization::Default, // B1TensorSpec
+            TensorSpecialization::Default, // CTensorSpec
+            1,
+            256,
+            128,         // MPerBlock
+            128,         // NPerBlock
+            32,          // KPerBlock
+            128,         // Gemm1NPerBlock
+            32,          // Gemm1KPerBlock
+            8,           // AK1
+            8,           // BK1
+            2,           // B1K1
+            32,          // MPerXDL
+            32,          // NPerXDL
+            1,           // MXdlPerWave
+            4,           // NXdlPerWave
+            4,           // Gemm1NXdlPerWave
+            S<4, 64, 1>, // ABlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<4, 64, 1>, // BBlockTransfer
+            S<1, 0, 2>,
+            S<1, 0, 2>,
+            2,
+            8,
+            8,
+            true,
+            S<8, 32, 1>, // B1BlockTransfer
+            S<0, 2, 1>,
+            S<0, 2, 1>,
+            1,
+            4,
+            2,
+            false,
+            1,              // CShuffleMXdlPerWavePerShuffle
+            2,              // CShuffleNXdlPerWavePerShuffle
+            S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+            8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+            MaskingSpecialization::MaskOutUpperTriangle>; // MaskOutUpperTriangle
+
+    bool IsSupported(int M, int N, int K, int O)
+    {
+        const int G0 = 1, G1 = 1;
+
+        // A layout [G0, M, G1, K]
+        std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+        std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+        // B0 layout [G0, N, G1, K]
+        std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+        std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+        // B1 layout [G0, N, G1, O]
+        std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+        std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+        // C layout [G0, M, G1, O]
+        std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+        std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+        auto gemm     = DeviceGemmGemmInstance{};
+        auto invoker  = gemm.MakeInvoker();
+        auto argument = gemm.MakeArgument(static_cast<ADataType*>(nullptr),
+                                          static_cast<B0DataType*>(nullptr),
+                                          static_cast<B1DataType*>(nullptr),
+                                          static_cast<CDataType*>(nullptr),
+                                          {}, // p_acc0_biases
+                                          {}, // p_acc1_biases
+                                          a_gs_ms_ks_lengths,
+                                          a_gs_ms_ks_strides,
+                                          b0_gs_ns_ks_lengths,
+                                          b0_gs_ns_ks_strides,
+                                          b1_gs_os_ns_lengths,
+                                          b1_gs_os_ns_strides,
+                                          c_gs_ms_os_lengths,
+                                          c_gs_ms_os_strides,
+                                          {},             // acc0_biases_gs_ms_ns_lengths
+                                          {},             // acc0_biases_gs_ms_ns_strides
+                                          {},             // acc1_biases_gs_ms_os_lengths
+                                          {},             // acc1_biases_gs_ms_os_strides
+                                          PassThrough{},  // a_element_op
+                                          PassThrough{},  // b0_element_op
+                                          Scale{1.f},     // acc0_element_op
+                                          PassThrough{},  // b1_element_op
+                                          PassThrough{}); // c_element_op
+
+        return gemm.IsSupportedArgument(argument);
+    }
+};
diff --git a/test/batchnorm/CMakeLists.txt b/test/batchnorm/CMakeLists.txt
new file mode 100644
index 0000000000..2a528f9c37
--- /dev/null
+++ b/test/batchnorm/CMakeLists.txt
@@ -0,0 +1,6 @@
+add_gtest_executable(test_batchnorm_fwd_rank_4 batchnorm_fwd_rank_4.cpp)
+add_gtest_executable(test_batchnorm_bwd_rank_4 batchnorm_bwd_rank_4.cpp)
+add_gtest_executable(test_batchnorm_infer_rank_4 batchnorm_infer_rank_4.cpp)
+target_link_libraries(test_batchnorm_fwd_rank_4 PRIVATE utility device_batchnorm_instance)
+target_link_libraries(test_batchnorm_bwd_rank_4 PRIVATE utility device_batchnorm_instance)
+target_link_libraries(test_batchnorm_infer_rank_4 PRIVATE utility device_batchnorm_instance)
diff --git a/test/batchnorm/batchnorm_bwd_rank_4.cpp b/test/batchnorm/batchnorm_bwd_rank_4.cpp
new file mode 100644
index 0000000000..a4696cf2a3
--- /dev/null
+++ b/test/batchnorm/batchnorm_bwd_rank_4.cpp
@@ -0,0 +1,92 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <tuple>
+#include <gtest/gtest.h>
+
+#include "profiler/profile_batchnorm_backward_impl.hpp"
+
+using F16  = ck::half_t;
+using F32  = float;
+using BF16 = ck::bhalf_t;
+using F64  = double;
+
+template <typename Tuple>
+class TestBatchNormBwdRank4 : public ::testing::Test
+{
+    private:
+    const double epsilon = std::numeric_limits<float>::epsilon();
+
+    protected:
+    using XDataType       = std::tuple_element_t<0, Tuple>;
+    using DxDataType      = std::tuple_element_t<1, Tuple>;
+    using DyDataType      = std::tuple_element_t<2, Tuple>;
+    using AccDataType     = std::tuple_element_t<3, Tuple>;
+    using ScaleDataType   = std::tuple_element_t<4, Tuple>;
+    using BiasDataType    = std::tuple_element_t<5, Tuple>;
+    using MeanVarDataType = std::tuple_element_t<6, Tuple>;
+
+    std::vector<std::vector<size_t>> list_of_lengths = {
+        {128, 16, 3, 1024}, {128, 16, 6, 512}, {1, 1, 1, 1}, {4, 4, 4, 4}, {32, 32, 32, 32}};
+    std::vector<int> reduceDims;
+
+    template <int NumReduceDim>
+    void Run()
+    {
+        for(auto& inOutLengths : list_of_lengths)
+        {
+            bool pass = true;
+
+            EXPECT_FALSE(reduceDims.size() != NumReduceDim);
+
+            pass = pass && ck::profiler::profile_batchnorm_backward_impl<XDataType,
+                                                                         DxDataType,
+                                                                         DyDataType,
+                                                                         AccDataType,
+                                                                         ScaleDataType,
+                                                                         BiasDataType,
+                                                                         MeanVarDataType,
+                                                                         4,
+                                                                         NumReduceDim>(
+                               true, 3, false, false, inOutLengths, reduceDims, true, epsilon);
+
+            pass = pass && ck::profiler::profile_batchnorm_backward_impl<XDataType,
+                                                                         DxDataType,
+                                                                         DyDataType,
+                                                                         AccDataType,
+                                                                         ScaleDataType,
+                                                                         BiasDataType,
+                                                                         MeanVarDataType,
+                                                                         4,
+                                                                         NumReduceDim>(
+                               true, 3, false, false, inOutLengths, reduceDims, false, epsilon);
+
+            EXPECT_TRUE(pass);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<std::tuple<F16, F32, F32, F32, F16, F32, F32>,
+                                     std::tuple<F32, F32, F32, F32, F32, F32, F32>,
+                                     std::tuple<BF16, F32, F32, F32, BF16, F32, F32>,
+                                     std::tuple<F64, F64, F64, F64, F64, F64, F64>>;
+
+TYPED_TEST_SUITE(TestBatchNormBwdRank4, KernelTypes);
+
+// nhwc
+TYPED_TEST(TestBatchNormBwdRank4, nhwc)
+{
+    this->reduceDims = {0, 1, 2};
+    this->template Run<3>();
+}
+
+// nchw
+TYPED_TEST(TestBatchNormBwdRank4, nchw)
+{
+    this->reduceDims = {0, 2, 3};
+    this->template Run<3>();
+}
diff --git a/test/batchnorm/batchnorm_fwd_rank_4.cpp b/test/batchnorm/batchnorm_fwd_rank_4.cpp
new file mode 100644
index 0000000000..9b6fbd0f66
--- /dev/null
+++ b/test/batchnorm/batchnorm_fwd_rank_4.cpp
@@ -0,0 +1,109 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <tuple>
+#include <gtest/gtest.h>
+
+#include "profiler/profile_batchnorm_forward_impl.hpp"
+
+using F16  = ck::half_t;
+using F32  = float;
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F64  = double;
+
+template <typename Tuple>
+class TestBatchNormFwdRank4 : public ::testing::Test
+{
+    private:
+    const double epsilon       = std::numeric_limits<float>::epsilon();
+    const double averageFactor = 0.1;
+
+    protected:
+    using XDataType       = std::tuple_element_t<0, Tuple>;
+    using YDataType       = std::tuple_element_t<1, Tuple>;
+    using AccDataType     = std::tuple_element_t<2, Tuple>;
+    using ScaleDataType   = std::tuple_element_t<3, Tuple>;
+    using BiasDataType    = std::tuple_element_t<4, Tuple>;
+    using MeanVarDataType = std::tuple_element_t<5, Tuple>;
+
+    std::vector<std::vector<size_t>> list_of_lengths = {
+        {128, 16, 3, 1024}, {128, 16, 6, 512}, {1, 1, 1, 1}, {4, 4, 4, 4}, {32, 32, 32, 32}};
+    std::vector<int> reduceDims;
+
+    template <int NumReduceDim>
+    void Run()
+    {
+        for(auto& inOutLengths : list_of_lengths)
+        {
+            bool pass = true;
+
+            EXPECT_FALSE(reduceDims.size() != NumReduceDim);
+
+            pass =
+                pass && ck::profiler::profile_batchnorm_forward_impl<XDataType,
+                                                                     YDataType,
+                                                                     AccDataType,
+                                                                     ScaleDataType,
+                                                                     BiasDataType,
+                                                                     MeanVarDataType,
+                                                                     4,
+                                                                     NumReduceDim>(true,
+                                                                                   3,
+                                                                                   false,
+                                                                                   false,
+                                                                                   inOutLengths,
+                                                                                   reduceDims,
+                                                                                   true,
+                                                                                   true,
+                                                                                   epsilon,
+                                                                                   averageFactor);
+
+            pass =
+                pass && ck::profiler::profile_batchnorm_forward_impl<XDataType,
+                                                                     YDataType,
+                                                                     AccDataType,
+                                                                     ScaleDataType,
+                                                                     BiasDataType,
+                                                                     MeanVarDataType,
+                                                                     4,
+                                                                     NumReduceDim>(true,
+                                                                                   3,
+                                                                                   false,
+                                                                                   false,
+                                                                                   inOutLengths,
+                                                                                   reduceDims,
+                                                                                   false,
+                                                                                   false,
+                                                                                   epsilon,
+                                                                                   averageFactor);
+
+            EXPECT_TRUE(pass);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<std::tuple<F16, F16, F32, F16, F16, F32>,
+                                     std::tuple<F32, F32, F32, F32, F32, F32>,
+                                     std::tuple<BF16, BF16, F32, BF16, BF16, F32>,
+                                     std::tuple<F64, F64, F64, F64, F64, F64>>;
+
+TYPED_TEST_SUITE(TestBatchNormFwdRank4, KernelTypes);
+
+// nhwc
+TYPED_TEST(TestBatchNormFwdRank4, nhwc)
+{
+    this->reduceDims = {0, 1, 2};
+    this->template Run<3>();
+}
+
+// nchw
+TYPED_TEST(TestBatchNormFwdRank4, nchw)
+{
+    this->reduceDims = {0, 2, 3};
+    this->template Run<3>();
+}
diff --git a/test/batchnorm/batchnorm_infer_rank_4.cpp b/test/batchnorm/batchnorm_infer_rank_4.cpp
new file mode 100644
index 0000000000..ecb4043b36
--- /dev/null
+++ b/test/batchnorm/batchnorm_infer_rank_4.cpp
@@ -0,0 +1,89 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <tuple>
+#include <gtest/gtest.h>
+
+#include "profiler/profile_batchnorm_infer_impl.hpp"
+
+using F16  = ck::half_t;
+using F32  = float;
+using BF16 = ck::bhalf_t;
+using F64  = double;
+
+template <typename Tuple>
+class TestBatchNormInferRank4 : public ::testing::Test
+{
+    private:
+    const double epsilon = std::numeric_limits<float>::epsilon();
+
+    protected:
+    using XDataType       = std::tuple_element_t<0, Tuple>;
+    using YDataType       = std::tuple_element_t<1, Tuple>;
+    using AccDataType     = std::tuple_element_t<2, Tuple>;
+    using ScaleDataType   = std::tuple_element_t<3, Tuple>;
+    using BiasDataType    = std::tuple_element_t<4, Tuple>;
+    using MeanVarDataType = std::tuple_element_t<5, Tuple>;
+
+    std::vector<std::vector<size_t>> list_of_lengths = {
+        {128, 16, 3, 1024}, {128, 16, 6, 512}, {4, 4, 4, 4}, {32, 32, 32, 32}};
+    std::vector<int> reduceDims;
+
+    template <int NumReduceDim>
+    void Run()
+    {
+        for(auto& inOutLengths : list_of_lengths)
+        {
+            bool pass = true;
+
+            EXPECT_FALSE(reduceDims.size() != NumReduceDim);
+
+            pass = pass && ck::profiler::profile_batchnorm_infer_impl<XDataType,
+                                                                      YDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      BiasDataType,
+                                                                      MeanVarDataType,
+                                                                      4,
+                                                                      NumReduceDim>(
+                               true, 3, false, false, inOutLengths, reduceDims, epsilon);
+
+            pass = pass && ck::profiler::profile_batchnorm_infer_impl<XDataType,
+                                                                      YDataType,
+                                                                      AccDataType,
+                                                                      ScaleDataType,
+                                                                      BiasDataType,
+                                                                      MeanVarDataType,
+                                                                      4,
+                                                                      NumReduceDim>(
+                               true, 3, false, false, inOutLengths, reduceDims, epsilon);
+
+            EXPECT_TRUE(pass);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<std::tuple<F16, F16, F32, F16, F16, F32>,
+                                     std::tuple<F32, F32, F32, F32, F32, F32>,
+                                     std::tuple<BF16, BF16, F32, BF16, BF16, F32>,
+                                     std::tuple<F64, F64, F64, F64, F64, F64>>;
+
+TYPED_TEST_SUITE(TestBatchNormInferRank4, KernelTypes);
+
+// nhwc
+TYPED_TEST(TestBatchNormInferRank4, nhwc)
+{
+    this->reduceDims = {0, 1, 2};
+    this->template Run<3>();
+}
+
+// nchw
+TYPED_TEST(TestBatchNormInferRank4, nchw)
+{
+    this->reduceDims = {0, 2, 3};
+    this->template Run<3>();
+}
diff --git a/test/block_swizzle_test/block_swizzle_test.cpp b/test/block_swizzle_test/block_swizzle_test.cpp
new file mode 100644
index 0000000000..29e118c2ad
--- /dev/null
+++ b/test/block_swizzle_test/block_swizzle_test.cpp
@@ -0,0 +1,406 @@
+#include <stdio.h>
+#include <string>
+#include <algorithm>
+#include <vector>
+#include <limits>
+#include "simple_args.h"
+
+simple_args_t create_arg(int argc, char** argv)
+{
+    simple_args_t args;
+    args.insert("m", "1024", "matrix m")
+        .insert("n", "1024", "matrix n")
+        .insert("k", "1024", "matrix k")
+        .insert("m_per_block", "128", "m_per_block")
+        .insert("n_per_block", "128", "n_per_block")
+        .insert("k_per_block", "32", "k_per_block")
+        .insert("num_cu", "104", "num cu")
+        .insert("occupancy", "2", "occupancy")
+        .parse(argc, argv);
+    return args;
+}
+
+namespace impl {
+template <typename T>
+T integer_divide_ceil(T n, T d)
+{
+    return (n + d - 1) / d;
+}
+
+template <typename T>
+T min(T a, T b)
+{
+    return a > b ? b : a;
+}
+
+template <typename T>
+T max(T a, T b)
+{
+    return a > b ? a : b;
+}
+
+} // namespace impl
+
+struct block_dispatcher_t
+{
+    public:
+    uint32_t m_per_block;
+    uint32_t n_per_block;
+    uint32_t k_per_block;
+    uint32_t num_cu;
+    uint32_t occupancy;
+    uint32_t m;
+    uint32_t n;
+    uint32_t k;
+
+    //--------------------------------------
+
+    uint32_t sk_num_blocks;
+    uint32_t sk_num_big_blocks;
+    uint32_t sk_total_iters;
+
+    // uint32_t sk_num_blocks_per_tile;    // how many
+
+    uint32_t dp_start_block_idx;
+    uint32_t dp_iters_per_block;
+    uint32_t dp_num_blocks;
+
+    uint32_t k_iters_per_tile;
+    uint32_t k_iters_per_big_block;
+    //--------------------------------------
+
+    static constexpr uint32_t min_k_iters_per_sk_block = 1;
+
+    void dump()
+    {
+        printf("%dx%dx%d(%dx%dx%d), cu:%d, occ:%d, grids:%d, sk_num_big_blocks:%d, "
+               "sk_num_blocks:%d, sk_total_iters:%d, dp_start_block_idx:%d, dp_iters_per_block:%d, "
+               "dp_num_blocks:%d, k_iters_per_tile:%d, k_iters_per_big_block:%d\n",
+               m,
+               n,
+               k,
+               m_per_block,
+               n_per_block,
+               k_per_block,
+               num_cu,
+               occupancy,
+               get_grid_dims_x(),
+               sk_num_big_blocks,
+               sk_num_blocks,
+               sk_total_iters,
+               dp_start_block_idx,
+               dp_iters_per_block,
+               dp_num_blocks,
+               k_iters_per_tile,
+               k_iters_per_big_block);
+    }
+
+    block_dispatcher_t(uint32_t m_per_block_,
+                       uint32_t n_per_block_,
+                       uint32_t k_per_block_,
+                       uint32_t num_cu_,
+                       uint32_t occupancy_,
+                       uint32_t m_,
+                       uint32_t n_,
+                       uint32_t k_)
+        : m_per_block(m_per_block_),
+          n_per_block(n_per_block_),
+          k_per_block(k_per_block_),
+          num_cu(num_cu_),
+          occupancy(occupancy_),
+          m(m_),
+          n(n_),
+          k(k_)
+    {
+        init();
+    }
+
+    uint32_t get_grid_dims_x() { return dp_start_block_idx + dp_num_blocks; }
+
+    uint32_t get_block_idx(uint32_t bid)
+    {
+        // block id is linearily allocated along sk blocks (dp blocks are fine)
+        // this function will compute blockIdx.x and the linear sk block mapping
+        // uint32_t block_idx = 0;
+        // if(bid < sk_num_big_blocks) {
+        //     uint32_t current_k_iter = bid * k_iters_per_big_block;
+        //     tile_idx = current_k_iter / k_iters_per_tile;
+        // }
+        return bid;
+    }
+
+    uint32_t get_current_itr(uint32_t block_idx)
+    {
+        uint32_t current_itr = 0;
+        if(block_idx < sk_num_big_blocks)
+        {
+            current_itr = block_idx * k_iters_per_big_block;
+        }
+        else if(block_idx < sk_num_blocks)
+        {
+            current_itr = (sk_num_big_blocks * k_iters_per_big_block) +
+                          (block_idx - sk_num_big_blocks) * (k_iters_per_big_block - 1);
+        }
+        else if(block_idx >= dp_start_block_idx)
+        {
+            current_itr = sk_total_iters + (block_idx - dp_start_block_idx) * dp_iters_per_block;
+        }
+        return current_itr;
+    }
+
+    void get_block_itr(uint32_t block_idx, uint32_t& iter_start, uint32_t& iter_end)
+    {
+        if(block_idx < sk_num_big_blocks)
+        {
+            iter_start = block_idx * k_iters_per_big_block;
+            iter_end   = iter_start + k_iters_per_big_block;
+        }
+        else if(block_idx < sk_num_blocks)
+        {
+            iter_start = (sk_num_big_blocks * k_iters_per_big_block) +
+                         (block_idx - sk_num_big_blocks) * (k_iters_per_big_block - 1);
+            iter_end = iter_start + (k_iters_per_big_block - 1);
+        }
+        else if(block_idx >= dp_start_block_idx)
+        {
+            iter_start = sk_total_iters + (block_idx - dp_start_block_idx) * dp_iters_per_block;
+            iter_end   = iter_start + dp_iters_per_block;
+        }
+    }
+
+    private:
+    void init()
+    {
+        uint32_t num_tiles =
+            impl::integer_divide_ceil(m, m_per_block) * impl::integer_divide_ceil(n, n_per_block);
+        k_iters_per_tile = impl::integer_divide_ceil(k, k_per_block);
+
+        // one cu can hold one wg at one time, from the whole chip's point of view
+        // if number of wg is same as num_cu, we call it 1 dispatch
+        // if number of wg is 2x num_cu, we call it 2 dispatches.
+        // one dispatch can deliever wg same as num_cu (full dispatch), or less than num_cu (partial
+        // dispatch)
+        //
+        uint32_t full_dispatches         = num_tiles / num_cu;
+        uint32_t full_dispatch_tiles     = full_dispatches * num_cu;
+        uint32_t partial_dispatche_tiles = num_tiles - full_dispatch_tiles;
+
+        uint32_t sk_occupancy = occupancy;
+        uint32_t dp_tiles     = full_dispatch_tiles;
+        uint32_t sk_tiles     = partial_dispatche_tiles;
+
+        if(full_dispatches < occupancy)
+        {
+            // in this case, we allocate all blocks as sk blocks
+            // sk_occupancy = occupancy - full_dispatches;
+            sk_occupancy = 1; // TODO: single occ seems better
+            dp_tiles     = full_dispatch_tiles;
+            sk_tiles     = partial_dispatche_tiles;
+        }
+        else if((occupancy > 1) && (full_dispatches % occupancy == occupancy - 1))
+        {
+            // e.g. occupancy = 2, full_dispatches = 3, 5, 7 ...
+            //      occupancy = 3, full_dispatches = 5, 8, 11 ...
+            //      occupancy = 4, full_dispatches = 7, 11 ...
+            sk_occupancy = 1; // left 1 slot for sk occupancy
+            dp_tiles     = full_dispatch_tiles;
+            sk_tiles     = partial_dispatche_tiles;
+        }
+        else
+        {
+            // others, we reduce 1 dispatch from dp, together with partial dispatch,
+            // to construct sk dispatch
+            sk_occupancy = occupancy - ((full_dispatches - 1) % occupancy);
+            dp_tiles     = full_dispatch_tiles - num_cu;
+            sk_tiles     = partial_dispatche_tiles + num_cu;
+        }
+
+        // dp_num_blocks = dp_tiles;
+        // dp_start_block_idx = num_cu * sk_occupancy;
+        dp_iters_per_block = k_iters_per_tile;
+
+        sk_total_iters = k_iters_per_tile * sk_tiles;
+
+        // printf("num_tiles:%d, full_dispatches:%d, full_dispatch_tiles:%d,
+        // partial_dispatche_tiles:%d\n",
+        //         num_tiles, full_dispatches, full_dispatch_tiles, partial_dispatche_tiles);
+
+        {
+            uint32_t min_sk_tiles = (sk_tiles >= num_cu) ? num_cu : (sk_tiles + 1);
+            uint32_t max_sk_tiles =
+                (sk_tiles >= num_cu) ? num_cu * sk_occupancy
+                                     : impl::min(num_cu, sk_total_iters / min_k_iters_per_sk_block);
+
+            // if use dp for sk-block, how many iters do we need
+            uint32_t dp_for_sk_iters = k_iters_per_tile;
+
+            uint32_t best_sk_score =
+                std::numeric_limits<int>::max(); // we need to find the smallest sk iters
+            for(uint32_t tentative_sk_blocks = min_sk_tiles; tentative_sk_blocks < max_sk_tiles;
+                tentative_sk_blocks++)
+            {
+                uint32_t tentative_sk_iters_per_block =
+                    (sk_total_iters + tentative_sk_blocks - 1) / tentative_sk_blocks;
+                uint32_t tentative_sk_iters = tentative_sk_iters_per_block;
+                uint32_t sk_blocks_per_tile = (tentative_sk_blocks + sk_tiles - 1) / sk_tiles;
+
+                // TODO: carefully adjust this parameter
+                //       the more sk_blocks_per_tile, the worse the overhead
+                uint32_t cross_sk_blocks_overhead = sk_blocks_per_tile;
+                if(tentative_sk_blocks % sk_tiles != 0)
+                {
+                    // penalty for uneven divide
+                    cross_sk_blocks_overhead +=
+                        sk_blocks_per_tile * tentative_sk_iters_per_block / 50;
+                }
+
+                uint32_t tentative_sk_score = tentative_sk_iters + cross_sk_blocks_overhead;
+
+                if(tentative_sk_score < best_sk_score)
+                {
+                    best_sk_score = tentative_sk_score;
+                    sk_num_blocks = tentative_sk_blocks;
+                }
+            }
+
+            if(best_sk_score >= dp_for_sk_iters)
+            {
+                sk_num_blocks = 0;
+            }
+
+            if(sk_num_blocks == 0)
+            {
+                sk_num_big_blocks     = 0;
+                k_iters_per_big_block = 0;
+
+                dp_num_blocks      = num_tiles; // all tile to be dp block
+                dp_start_block_idx = 0;
+                sk_total_iters     = 0; // clear this tiles
+            }
+            else
+            {
+                uint32_t k_iters_per_sk_block = sk_total_iters / sk_num_blocks;
+                sk_num_big_blocks     = sk_total_iters - k_iters_per_sk_block * sk_num_blocks;
+                k_iters_per_big_block = k_iters_per_sk_block + 1;
+
+                dp_num_blocks      = dp_tiles;
+                dp_start_block_idx = (sk_num_blocks + num_cu - 1) / num_cu * num_cu;
+            }
+        }
+    }
+};
+
+struct tile_work_t
+{
+    uint32_t tile_idx;
+    uint32_t iter_begin;
+    uint32_t k_begin;
+    uint32_t k_end;
+    uint32_t k_iters_remaining;
+};
+
+int main(int argc, char** argv)
+{
+    simple_args_t arg = create_arg(argc, argv);
+    block_dispatcher_t block_dispatcher{arg.get_uint32("m_per_block"),
+                                        arg.get_uint32("n_per_block"),
+                                        arg.get_uint32("k_per_block"),
+                                        arg.get_uint32("num_cu"),
+                                        arg.get_uint32("occupancy"),
+                                        arg.get_uint32("m"),
+                                        arg.get_uint32("n"),
+                                        arg.get_uint32("k")};
+    block_dispatcher.dump();
+    // simulate actual kernel launch
+    uint32_t dim_x = block_dispatcher.get_grid_dims_x();
+    uint32_t total_k_iters =
+        impl::integer_divide_ceil(arg.get_uint32("k"), arg.get_uint32("k_per_block"));
+    uint32_t num_tiles =
+        impl::integer_divide_ceil(arg.get_uint32("m"), arg.get_uint32("m_per_block")) *
+        impl::integer_divide_ceil(arg.get_uint32("n"), arg.get_uint32("n_per_block"));
+
+    std::vector<int> valid_tile_record(num_tiles * total_k_iters);
+
+    for(uint32_t bid = 0; bid < dim_x; bid++)
+    {
+        uint32_t block_idx = block_dispatcher.get_block_idx(bid);
+        bool is_sk_block   = block_idx < (block_dispatcher.sk_num_blocks);
+        bool is_dp_block   = block_idx >= block_dispatcher.dp_start_block_idx;
+        uint32_t iter_start, iter_end;
+        block_dispatcher.get_block_itr(block_idx, iter_start, iter_end);
+        uint32_t total_iter_length = iter_end - iter_start;
+
+        while(true)
+        {
+            uint32_t iter_length_mod = iter_end % block_dispatcher.k_iters_per_tile;
+            uint32_t current_iter_length =
+                impl::min(iter_length_mod == 0 ? (iter_end - iter_start) : iter_length_mod,
+                          total_iter_length);
+            uint32_t tile_idx = (iter_end - 1) / block_dispatcher.k_iters_per_tile;
+            uint32_t tile_iter_start =
+                ((iter_end - 1) % block_dispatcher.k_iters_per_tile) - current_iter_length + 1;
+
+            if(is_sk_block)
+            {
+                printf("[sk_block] bid:%3d, block_idx:%3d, tile_idx:%3d, iter_start:%d(%d | %d), "
+                       "iter_end:%d (len:%d)\n",
+                       bid,
+                       block_idx,
+                       tile_idx,
+                       iter_end - current_iter_length,
+                       tile_iter_start,
+                       iter_start,
+                       iter_end,
+                       current_iter_length);
+            }
+            else if(is_dp_block)
+            {
+                printf("[dp_block] bid:%3d, block_idx:%3d, tile_idx:%3d, iter_start:%d(%d | %d), "
+                       "iter_end:%d (len:%d)\n",
+                       bid,
+                       block_idx,
+                       tile_idx,
+                       iter_end - current_iter_length,
+                       tile_iter_start,
+                       iter_start,
+                       iter_end,
+                       current_iter_length);
+            }
+            else
+            {
+                printf("[other   ] bid:%3d, block_idx:%3d\n", bid, block_idx);
+            }
+
+            // some validation check
+            for(auto i = iter_end - current_iter_length; i < iter_end; i++)
+            {
+                if(i >= valid_tile_record.size())
+                {
+                    printf("unexpected, current iter:%d larger than max:%d\n",
+                           i,
+                           valid_tile_record.size());
+                    return -1;
+                }
+                valid_tile_record[i] = 1;
+            }
+
+            iter_end -= current_iter_length;
+            if(iter_end <= iter_start)
+                break;
+        }
+    }
+
+    int untouched = 0;
+    for(auto i = 0; i < valid_tile_record.size(); i++)
+    {
+        if(valid_tile_record[i] != 1)
+        {
+            printf("untouched at %d (%d)\n", i, valid_tile_record.size());
+            untouched++;
+        }
+    }
+    printf("untouched %d/%d, %s\n",
+           untouched,
+           valid_tile_record.size(),
+           untouched == 0 ? "valid" : "fail");
+}
diff --git a/test/block_swizzle_test/rebuild.sh b/test/block_swizzle_test/rebuild.sh
new file mode 100644
index 0000000000..b07eb55048
--- /dev/null
+++ b/test/block_swizzle_test/rebuild.sh
@@ -0,0 +1,3 @@
+CC=g++
+
+$CC -Wall -std=c++17 -Iinclude -O3 block_swizzle_test.cpp -o block_swizzle_test.exe
\ No newline at end of file
diff --git a/test/block_swizzle_test/simple_args.h b/test/block_swizzle_test/simple_args.h
new file mode 100644
index 0000000000..7b10456ce9
--- /dev/null
+++ b/test/block_swizzle_test/simple_args.h
@@ -0,0 +1,159 @@
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <stdlib.h>
+#include <string>
+#include <unordered_map>
+#include <vector>
+#include <assert.h>
+
+struct arg_content_t
+{
+    std::string name; // key
+    std::string value;
+    std::string help_text;
+};
+
+class simple_args_t
+{
+    public:
+    simple_args_t() {}
+    simple_args_t& insert(const std::string& name_,
+                          const std::string& default_value_,
+                          const std::string& help_text_)
+    {
+        arg_content_t arg{name_, default_value_, help_text_};
+
+        if(arg_map.count(arg.name) != 0)
+        {
+            std::cout << "arg:" << arg.name << "already exist" << std::endl;
+        }
+        else
+        {
+            arg_map[arg.name] = arg;
+        }
+        return *this;
+    }
+    void usage()
+    {
+        for(auto& content : arg_map)
+        {
+            std::vector<std::string> help_text_lines;
+            size_t pos = 0;
+            for(size_t next_pos = content.second.help_text.find('\n', pos);
+                next_pos != std::string::npos;)
+            {
+                help_text_lines.push_back(
+                    std::string(content.second.help_text.begin() + pos,
+                                content.second.help_text.begin() + next_pos++));
+                pos      = next_pos;
+                next_pos = content.second.help_text.find('\n', pos);
+            }
+            help_text_lines.push_back(std::string(content.second.help_text.begin() + pos,
+                                                  content.second.help_text.end()));
+
+            int arg_name_width = 16 - content.second.name.length();
+            arg_name_width     = arg_name_width > 0 ? arg_name_width : 2;
+            std::cout << std::setw(4) << "-" << content.second.name << std::setw(arg_name_width)
+                      << " " << help_text_lines[0] << std::endl;
+
+            for(auto help_next_line = std::next(help_text_lines.begin());
+                help_next_line != help_text_lines.end();
+                ++help_next_line)
+            {
+                std::cout << std::setw(28) << " " << *help_next_line << std::endl;
+            }
+        }
+    }
+    bool parse(int argc, char* argv[], int start_index = 1)
+    {
+        if(argc <= start_index)
+        {
+            // std::cout << "not enough args (" << argc << ") with starting index " << start_index
+            // << std::endl;
+            return true;
+        }
+        for(int i = start_index; i < argc; i++)
+        {
+            std::string cur_arg = std::string(argv[i]);
+            if(cur_arg[0] != '-')
+            {
+                std::cout << "illegal input" << std::endl;
+                usage();
+                return false;
+            }
+            else if(cur_arg[0] == '-' && cur_arg[1] == '?')
+            {
+                usage();
+                return false;
+            }
+            else
+            {
+                size_t found_equal = cur_arg.find('=');
+                if(found_equal == std::string::npos || found_equal == (cur_arg.length() - 1))
+                {
+                    std::cout << "failed while parsing \"" << cur_arg << "\", "
+                              << "arg must be in the form \"-name=value\"" << std::endl;
+                    return false;
+                }
+                std::string arg_name  = cur_arg.substr(1, found_equal - 1);
+                std::string arg_value = cur_arg.substr(found_equal + 1);
+                if(arg_map.count(arg_name) == 0)
+                {
+                    std::cout << "no such arg \"" << arg_name << "\" registered" << std::endl;
+                    return false;
+                }
+                arg_map[arg_name].value = arg_value;
+            }
+        }
+        return true;
+    }
+
+    std::string get(const std::string& name) const { return get_str(name); }
+
+    std::string get_str(const std::string& name) const
+    {
+        assert(arg_map.count(name) != 0);
+        std::string value = arg_map.at(name).value;
+        return value;
+    }
+
+    int get_int(const std::string& name) const
+    {
+        assert(arg_map.count(name) != 0);
+        int value = atoi(arg_map.at(name).value.c_str());
+        return value;
+    }
+
+    uint32_t get_uint32(const std::string& name) const
+    {
+        assert(arg_map.count(name) != 0);
+        uint32_t value = strtoul(arg_map.at(name).value.c_str(), nullptr, 10);
+        return value;
+    }
+
+    uint64_t get_uint64(const std::string& name) const
+    {
+        assert(arg_map.count(name) != 0);
+        uint64_t value = strtoull(arg_map.at(name).value.c_str(), nullptr, 10);
+        return value;
+    }
+
+    double get_double(const std::string& name) const
+    {
+        assert(arg_map.count(name) != 0);
+        double value = atof(arg_map.at(name).value.c_str());
+        return value;
+    }
+
+    float get_float(const std::string& name) const
+    {
+        assert(arg_map.count(name) != 0);
+        float value = atof(arg_map.at(name).value.c_str());
+        return value;
+    }
+
+    private:
+    std::unordered_map<std::string, arg_content_t> arg_map;
+};
diff --git a/test/block_to_ctile_map/CMakeLists.txt b/test/block_to_ctile_map/CMakeLists.txt
new file mode 100644
index 0000000000..97dfbb2b55
--- /dev/null
+++ b/test/block_to_ctile_map/CMakeLists.txt
@@ -0,0 +1 @@
+add_gtest_executable(test_block_to_ctile_map test_block_to_ctile_map.cpp)
\ No newline at end of file
diff --git a/test/block_to_ctile_map/test_block_to_ctile_map.cpp b/test/block_to_ctile_map/test_block_to_ctile_map.cpp
new file mode 100644
index 0000000000..b8e349eda1
--- /dev/null
+++ b/test/block_to_ctile_map/test_block_to_ctile_map.cpp
@@ -0,0 +1,322 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+
+using namespace ck;
+
+static auto I0 = Number<0>{};
+static auto I1 = Number<1>{};
+static auto I2 = Number<2>{};
+
+TEST(BlockToCTileMap, TestBlockToCTileMap_M00_N00_M01_N01_DeviceCTileIndexCheck1)
+{
+    const index_t M         = 384;
+    const index_t N         = 384;
+    const index_t MPerBlock = 128;
+    const index_t NPerBlock = 128;
+    const index_t MBlock    = M / MPerBlock;
+    const index_t NBlock    = N / NPerBlock;
+    const index_t M01       = 4;
+    const index_t N01       = 4;
+
+    auto c_grid_desc_m_n = make_naive_tensor_descriptor_packed(make_tuple(M, N));
+
+    printf("(M, N, MPerBlock, NPerBlock, M01, N01) = (%d, %d, %d, %d, %d, %d)\n",
+           M,
+           N,
+           MPerBlock,
+           NPerBlock,
+           M01,
+           N01);
+
+    BlockToCTileMap_M00_N00_M01_N01<MPerBlock, NPerBlock, decltype(c_grid_desc_m_n), true> tile_map(
+        c_grid_desc_m_n, M01, N01);
+
+    EXPECT_TRUE(tile_map.CheckValidity(c_grid_desc_m_n) == true);
+    EXPECT_TRUE(tile_map.CalculateGridSize(c_grid_desc_m_n) == 16);
+
+    // clang-format off
+    std::vector<std::vector<int>> expected_m0idx_n0idx_valid = {
+        {0, 0, 1},
+        {0, 1, 1},
+        {0, 2, 1},
+        {0, 3, 0},
+        {1, 0, 1},
+        {1, 1, 1},
+        {1, 2, 1},
+        {1, 3, 0},
+        {2, 0, 1},
+        {2, 1, 1},
+        {2, 2, 1},
+        {2, 3, 0},
+        {3, 0, 0},
+        {3, 1, 0},
+        {3, 2, 0},
+        {3, 3, 0}
+    };
+    // clang-format on
+
+    for(index_t i = 0; i < tile_map.CalculateGridSize(c_grid_desc_m_n); i++)
+    {
+        auto m0n0_idx = tile_map.CalculateBottomIndex(make_multi_index(i));
+        std::cout << "block_1d_id = " << i << ", m0, n0 = " << m0n0_idx[I0] << ", " << m0n0_idx[I1];
+        std::cout << ", valid = " << tile_map.ValidCTileIndex(m0n0_idx, make_tuple(MBlock, NBlock))
+                  << std::endl;
+        bool equal =
+            expected_m0idx_n0idx_valid[i] ==
+            std::vector<int>{m0n0_idx[I0],
+                             m0n0_idx[I1],
+                             tile_map.ValidCTileIndex(m0n0_idx, make_tuple(MBlock, NBlock))};
+        EXPECT_TRUE(equal);
+    }
+}
+
+TEST(BlockToCTileMap, TestBlockToCTileMap_M00_N00_M01_N01_DeviceCTileIndexCheck0)
+{
+    const index_t M         = 384;
+    const index_t N         = 384;
+    const index_t MPerBlock = 128;
+    const index_t NPerBlock = 128;
+
+    const index_t M01 = 4;
+    const index_t N01 = 4;
+
+    auto c_grid_desc_m_n = make_naive_tensor_descriptor_packed(make_tuple(M, N));
+
+    printf("(M, N, MPerBlock, NPerBlock, M01, N01) = (%d, %d, %d, %d, %d, %d)\n",
+           M,
+           N,
+           MPerBlock,
+           NPerBlock,
+           M01,
+           N01);
+
+    BlockToCTileMap_M00_N00_M01_N01<MPerBlock, NPerBlock, decltype(c_grid_desc_m_n), false>
+        tile_map(c_grid_desc_m_n, M01, N01);
+
+    EXPECT_TRUE(tile_map.CheckValidity(c_grid_desc_m_n) == false);
+}
+
+TEST(BlockToCTileMap, TestBlockToCTileMap_M00_N0_M01_DeviceCTileIndexCheck1)
+{
+    const index_t M         = 384;
+    const index_t N         = 512;
+    const index_t MPerBlock = 128;
+    const index_t NPerBlock = 128;
+    const index_t MBlock    = M / MPerBlock;
+    const index_t NBlock    = N / NPerBlock;
+    const index_t M01       = 4;
+
+    auto c_grid_desc_m_n = make_naive_tensor_descriptor_packed(make_tuple(M, N));
+
+    printf("(M, N, MPerBlock, NPerBlock, M01) = (%d, %d, %d, %d, %d)\n",
+           M,
+           N,
+           MPerBlock,
+           NPerBlock,
+           M01);
+
+    BlockToCTileMap_M00_N0_M01<MPerBlock, NPerBlock, decltype(c_grid_desc_m_n), true> tile_map(
+        c_grid_desc_m_n, M01);
+
+    EXPECT_TRUE(tile_map.CheckValidity(c_grid_desc_m_n) == true);
+    EXPECT_TRUE(tile_map.CalculateGridSize(c_grid_desc_m_n) == 16);
+
+    // clang-format off
+    std::vector<std::vector<int>> expected_m0idx_n0idx_valid = {
+        {0, 0, 1},
+        {1, 0, 1},
+        {2, 0, 1},
+        {3, 0, 0},
+        {0, 1, 1},
+        {1, 1, 1},
+        {2, 1, 1},
+        {3, 1, 0},
+        {0, 2, 1},
+        {1, 2, 1},
+        {2, 2, 1},
+        {3, 2, 0},
+        {0, 3, 1},
+        {1, 3, 1},
+        {2, 3, 1},
+        {3, 3, 0}
+    };
+    // clang-format on
+
+    for(index_t i = 0; i < tile_map.CalculateGridSize(c_grid_desc_m_n); i++)
+    {
+        auto m0n0_idx = tile_map.CalculateBottomIndex(make_multi_index(i));
+        std::cout << "block_1d_id = " << i << ", m0, n0 = " << m0n0_idx[I0] << ", " << m0n0_idx[I1];
+        std::cout << ", valid = " << tile_map.ValidCTileIndex(m0n0_idx, make_tuple(MBlock, NBlock))
+                  << std::endl;
+        bool equal =
+            expected_m0idx_n0idx_valid[i] ==
+            std::vector<int>{m0n0_idx[I0],
+                             m0n0_idx[I1],
+                             tile_map.ValidCTileIndex(m0n0_idx, make_tuple(MBlock, NBlock))};
+        EXPECT_TRUE(equal);
+    }
+}
+
+TEST(BlockToCTileMap, TestBlockToCTileMap_M00_N0_M01_DeviceCTileIndexCheck0)
+{
+    const index_t M         = 512;
+    const index_t N         = 384;
+    const index_t MPerBlock = 128;
+    const index_t NPerBlock = 128;
+
+    auto c_grid_desc_m_n = make_naive_tensor_descriptor_packed(make_tuple(M, N));
+
+    // clang-format off
+    std::vector<std::tuple<int, int, bool>> expected_m0_gridsize_validity = {
+        {5, 15, false},
+        {4, 12, true},
+        {3, 18, false},
+        {2, 12, true},
+        {1, 12, true}
+    };
+    // clang-format on
+
+    for(auto e : expected_m0_gridsize_validity)
+    {
+        const index_t M01 = std::get<0>(e);
+
+        printf("(M, N, MPerBlock, NPerBlock, M01) = (%d, %d, %d, %d, %d)\n",
+               M,
+               N,
+               MPerBlock,
+               NPerBlock,
+               M01);
+
+        BlockToCTileMap_M00_N0_M01<MPerBlock, NPerBlock, decltype(c_grid_desc_m_n), false> tile_map(
+            c_grid_desc_m_n, M01);
+
+        EXPECT_EQ(tile_map.CalculateGridSize(c_grid_desc_m_n), std::get<1>(e));
+        EXPECT_EQ(tile_map.CheckValidity(c_grid_desc_m_n), std::get<2>(e));
+    }
+}
+
+TEST(BlockToCTileMap, TestBlockToCTileMap_M00_N0_M01Adapt)
+{
+    const index_t M         = 768;
+    const index_t N         = 384;
+    const index_t MPerBlock = 128;
+    const index_t NPerBlock = 128;
+    const index_t MBlock    = M / MPerBlock;
+    const index_t NBlock    = N / NPerBlock;
+    constexpr index_t M01   = 4;
+
+    auto c_grid_desc_m_n = make_naive_tensor_descriptor_packed(make_tuple(M, N));
+
+    printf("(M, N, MPerBlock, NPerBlock, M01) = (%d, %d, %d, %d, %d)\n",
+           M,
+           N,
+           MPerBlock,
+           NPerBlock,
+           M01);
+
+    BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, decltype(c_grid_desc_m_n)> tile_map(
+        c_grid_desc_m_n, M01);
+
+    EXPECT_TRUE(tile_map.CheckValidity(c_grid_desc_m_n) == true);
+    EXPECT_TRUE(tile_map.CalculateGridSize(c_grid_desc_m_n) == 18);
+
+    // clang-format off
+    std::vector<std::vector<int>> expected_m0idx_n0idx_valid = {
+        {0, 0, 1},
+        {1, 0, 1},
+        {2, 0, 1},
+        {3, 0, 1},
+        {0, 1, 1},
+        {1, 1, 1},
+        {2, 1, 1},
+        {3, 1, 1},
+        {0, 2, 1},
+        {1, 2, 1},
+        {2, 2, 1},
+        {3, 2, 1},
+        {4, 0, 1},
+        {5, 0, 1},
+        {4, 1, 1},
+        {5, 1, 1},
+        {4, 2, 1},
+        {5, 2, 1},
+    };
+    // clang-format on
+
+    for(index_t i = 0; i < tile_map.CalculateGridSize(c_grid_desc_m_n); i++)
+    {
+        auto m0n0_idx = tile_map.CalculateBottomIndex(make_multi_index(i));
+        std::cout << "block_1d_id = " << i << ", m0, n0 = " << m0n0_idx[I0] << ", " << m0n0_idx[I1];
+        std::cout << ", valid = " << tile_map.ValidCTileIndex(m0n0_idx, make_tuple(MBlock, NBlock))
+                  << std::endl;
+        bool equal =
+            expected_m0idx_n0idx_valid[i] ==
+            std::vector<int>{m0n0_idx[I0],
+                             m0n0_idx[I1],
+                             tile_map.ValidCTileIndex(m0n0_idx, make_tuple(MBlock, NBlock))};
+        EXPECT_TRUE(equal);
+    }
+}
+
+TEST(BlockToCTileMap, TestBlockToCTileMap_KSplit_M00_N0_M01Adapt)
+{
+    const index_t M         = 768;
+    const index_t N         = 384;
+    const index_t MPerBlock = 128;
+    const index_t NPerBlock = 128;
+    const index_t MBlock    = M / MPerBlock;
+    const index_t NBlock    = N / NPerBlock;
+    constexpr index_t M01   = 4;
+    const index_t KSplit    = 3;
+
+    auto c_grid_desc_m_n = make_naive_tensor_descriptor_packed(make_tuple(M, N));
+
+    printf("(M, N, MPerBlock, NPerBlock, M01) = (%d, %d, %d, %d, %d)\n",
+           M,
+           N,
+           MPerBlock,
+           NPerBlock,
+           M01);
+
+    BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, decltype(c_grid_desc_m_n)>
+        tile_map(c_grid_desc_m_n, M01, KSplit);
+
+    EXPECT_TRUE(tile_map.CheckValidity(c_grid_desc_m_n) == true);
+    EXPECT_TRUE(tile_map.CalculateGridSize(c_grid_desc_m_n) == 18 * KSplit);
+
+    std::vector<std::vector<int>> expected_ksplitidx_m0idx_n0idx_valid = {
+        {0, 0, 0, 1}, {0, 1, 0, 1}, {0, 2, 0, 1}, {0, 3, 0, 1}, {0, 0, 1, 1}, {0, 1, 1, 1},
+        {0, 2, 1, 1}, {0, 3, 1, 1}, {0, 0, 2, 1}, {0, 1, 2, 1}, {0, 2, 2, 1}, {0, 3, 2, 1},
+        {0, 4, 0, 1}, {0, 5, 0, 1}, {0, 4, 1, 1}, {0, 5, 1, 1}, {0, 4, 2, 1}, {0, 5, 2, 1},
+        {1, 0, 0, 1}, {1, 1, 0, 1}, {1, 2, 0, 1}, {1, 3, 0, 1}, {1, 0, 1, 1}, {1, 1, 1, 1},
+        {1, 2, 1, 1}, {1, 3, 1, 1}, {1, 0, 2, 1}, {1, 1, 2, 1}, {1, 2, 2, 1}, {1, 3, 2, 1},
+        {1, 4, 0, 1}, {1, 5, 0, 1}, {1, 4, 1, 1}, {1, 5, 1, 1}, {1, 4, 2, 1}, {1, 5, 2, 1},
+        {2, 0, 0, 1}, {2, 1, 0, 1}, {2, 2, 0, 1}, {2, 3, 0, 1}, {2, 0, 1, 1}, {2, 1, 1, 1},
+        {2, 2, 1, 1}, {2, 3, 1, 1}, {2, 0, 2, 1}, {2, 1, 2, 1}, {2, 2, 2, 1}, {2, 3, 2, 1},
+        {2, 4, 0, 1}, {2, 5, 0, 1}, {2, 4, 1, 1}, {2, 5, 1, 1}, {2, 4, 2, 1}, {2, 5, 2, 1},
+    };
+
+    for(index_t i = 0; i < tile_map.CalculateGridSize(c_grid_desc_m_n); i++)
+    {
+        auto ksplitm0n0_idx = tile_map.CalculateBottomIndex(make_multi_index(i));
+        std::cout << "block_1d_id = " << i << ", ksplit, m0, n0 = " << ksplitm0n0_idx[I0] << ", "
+                  << ksplitm0n0_idx[I1] << ", " << ksplitm0n0_idx[I2];
+        std::cout << ", valid = "
+                  << tile_map.ValidCTileIndex(ksplitm0n0_idx, make_tuple(MBlock, NBlock))
+                  << std::endl;
+        bool equal =
+            expected_ksplitidx_m0idx_n0idx_valid[i] ==
+            std::vector<int>{ksplitm0n0_idx[I0],
+                             ksplitm0n0_idx[I1],
+                             ksplitm0n0_idx[I2],
+                             tile_map.ValidCTileIndex(ksplitm0n0_idx, make_tuple(MBlock, NBlock))};
+        EXPECT_TRUE(equal);
+    }
+}
diff --git a/test/contraction/CMakeLists.txt b/test/contraction/CMakeLists.txt
new file mode 100644
index 0000000000..1f6e0ed341
--- /dev/null
+++ b/test/contraction/CMakeLists.txt
@@ -0,0 +1,11 @@
+add_gtest_executable(test_contraction test_contraction.cpp)
+target_link_libraries(test_contraction PRIVATE utility device_contraction_bilinear_instance device_contraction_scale_instance)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+    add_gtest_executable(test_contraction_interface test_contraction_interface.cpp)
+    target_link_libraries(test_contraction_interface PRIVATE utility device_contraction_bilinear_instance device_contraction_scale_instance)
+   set(target 1)
+ endif()
+endforeach()
diff --git a/test/contraction/test_contraction.cpp b/test/contraction/test_contraction.cpp
new file mode 100644
index 0000000000..c86b849235
--- /dev/null
+++ b/test/contraction/test_contraction.cpp
@@ -0,0 +1,138 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <memory>
+#include <initializer_list>
+#include <vector>
+#include <tuple>
+#include <gtest/gtest.h>
+
+#include "profiler/profile_contraction_impl.hpp"
+
+using F32 = float;
+using F64 = double;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using Bilinear = ck::tensor_operation::element_wise::Bilinear;
+using Scale    = ck::tensor_operation::element_wise::Scale;
+
+struct MemoryParams
+{
+    std::vector<ck::index_t> M;
+    std::vector<ck::index_t> N;
+    std::vector<ck::index_t> K;
+    std::vector<ck::index_t> StridesA;
+    std::vector<ck::index_t> StridesB;
+    std::vector<ck::index_t> StridesC;
+    std::vector<ck::index_t> StridesD;
+};
+
+template <typename Tuple>
+class TestContraction : public ::testing::Test
+{
+    protected:
+    using ALayout        = std::tuple_element_t<0, Tuple>;
+    using BLayout        = std::tuple_element_t<1, Tuple>;
+    using CDLayout       = std::tuple_element_t<2, Tuple>;
+    using DataType       = std::tuple_element_t<3, Tuple>;
+    using DTupleDataType = std::tuple_element_t<4, Tuple>;
+    using CDElementOp    = std::tuple_element_t<5, Tuple>;
+
+    std::vector<MemoryParams> list_of_memory_params = {{{32, 32},
+                                                        {32, 32},
+                                                        {32, 32},
+                                                        {32768, 1024, 32, 1},
+                                                        {32768, 1024, 32, 1},
+                                                        {32768, 1024, 32, 1},
+                                                        {32768, 1024, 32, 1}},
+                                                       {{16, 16},
+                                                        {32, 32},
+                                                        {16, 16},
+                                                        {4096, 256, 16, 1},
+                                                        {16, 1, 8192, 256},
+                                                        {16384, 1024, 32, 1},
+                                                        {16384, 1024, 32, 1}}};
+
+    std::vector<ck::index_t> init_methods = {0, 1, 2};
+    std::unique_ptr<CDElementOp> p_cd_element_op;
+    void Run()
+    {
+        for(auto& memory_params : list_of_memory_params)
+        {
+            for(const ck::index_t init_method : init_methods)
+            {
+                bool pass =
+                    ck::profiler::profile_contraction_impl<ALayout,
+                                                           BLayout,
+                                                           CDLayout,
+                                                           DataType,
+                                                           DTupleDataType,
+                                                           CDElementOp>(true /*do_verification*/,
+                                                                        init_method,
+                                                                        false /*do_logs*/,
+                                                                        false /*time_kernel*/,
+                                                                        *p_cd_element_op,
+                                                                        memory_params.M,
+                                                                        memory_params.N,
+                                                                        memory_params.K,
+                                                                        memory_params.StridesA,
+                                                                        memory_params.StridesB,
+                                                                        memory_params.StridesC,
+                                                                        memory_params.StridesD);
+                EXPECT_TRUE(pass);
+            }
+        }
+    }
+};
+
+template <typename Tuple>
+class TestContractionScale : public TestContraction<Tuple>
+{
+};
+
+template <typename Tuple>
+class TestContractionBilinear : public TestContraction<Tuple>
+{
+};
+
+using BilinearKernelTypes =
+    ::testing::Types<std::tuple<Row, Row, Row, F32, ck::Tuple<F32>, Bilinear>,
+                     std::tuple<Row, Col, Row, F32, ck::Tuple<F32>, Bilinear>,
+                     std::tuple<Col, Row, Row, F32, ck::Tuple<F32>, Bilinear>,
+                     std::tuple<Col, Col, Row, F32, ck::Tuple<F32>, Bilinear>,
+                     std::tuple<Row, Row, Row, F64, ck::Tuple<F32>, Bilinear>,
+                     std::tuple<Row, Col, Row, F64, ck::Tuple<F32>, Bilinear>,
+                     std::tuple<Col, Row, Row, F64, ck::Tuple<F32>, Bilinear>,
+                     std::tuple<Col, Col, Row, F64, ck::Tuple<F32>, Bilinear>>;
+
+using ScaleKernelTypes = ::testing::Types<std::tuple<Row, Row, Row, F32, ck::Tuple<>, Scale>,
+                                          std::tuple<Row, Col, Row, F32, ck::Tuple<>, Scale>,
+                                          std::tuple<Col, Row, Row, F32, ck::Tuple<>, Scale>,
+                                          std::tuple<Col, Col, Row, F32, ck::Tuple<>, Scale>,
+                                          std::tuple<Row, Row, Row, F64, ck::Tuple<>, Scale>,
+                                          std::tuple<Row, Col, Row, F64, ck::Tuple<>, Scale>,
+                                          std::tuple<Col, Row, Row, F64, ck::Tuple<>, Scale>,
+                                          std::tuple<Col, Col, Row, F64, ck::Tuple<>, Scale>>;
+
+TYPED_TEST_SUITE(TestContractionBilinear, BilinearKernelTypes);
+TYPED_TEST_SUITE(TestContractionScale, ScaleKernelTypes);
+
+TYPED_TEST(TestContractionBilinear, bilinear)
+{
+    this->p_cd_element_op = std::make_unique<Bilinear>(1.f, 1.f);
+    this->Run();
+    this->p_cd_element_op = std::make_unique<Bilinear>(-0.5f, 0.5f);
+    this->Run();
+}
+
+TYPED_TEST(TestContractionScale, scale)
+{
+    this->p_cd_element_op = std::make_unique<Scale>(1.f);
+    this->Run();
+    this->p_cd_element_op = std::make_unique<Scale>(0.5f);
+    this->Run();
+}
diff --git a/test/contraction/test_contraction_interface.cpp b/test/contraction/test_contraction_interface.cpp
new file mode 100644
index 0000000000..12a307f5da
--- /dev/null
+++ b/test/contraction/test_contraction_interface.cpp
@@ -0,0 +1,195 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <stdexcept>
+#include <vector>
+
+#include "gtest/gtest.h"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+
+using Pass     = ck::tensor_operation::element_wise::PassThrough;
+using Bilinear = ck::tensor_operation::element_wise::Bilinear;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F32 = float;
+using F64 = double;
+
+template <ck::index_t ABlockTransferSrcVectorDim,
+          ck::index_t BBlockTransferSrcVectorDim,
+          ck::index_t CDEBlockTransferScalarPerVector>
+class ContractionInstanceWrapper
+{
+    public:
+    static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+    static constexpr ck::index_t NumDim = 2;
+    // clang-format off
+    using ContractionDeviceInstance = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|         DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer|             ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer|              BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|                  CBlockTransfer|
+        //#####################################|        |        |        |  Type|  Type|    Type| DataType|           Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|               SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|               SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl|                 ScalarPerVector|
+        //#####################################|        |        |        |      |      |        |         |               |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |                           |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |                           |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|                   _NWaveNPerXdl|
+        //#####################################|        |        |        |      |      |        |         |               |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |                           |               |               |          |                |               |               |                           |               |               |          |            |            |                             |                                |
+        DeviceContractionMultipleD_Xdl_CShuffle<  NumDim,  NumDim,  NumDim,   F32,   F32,     F32,      F32, ck::Tuple<F32>,   F32,         Pass,        Pass,     Bilinear,       GemmSpec,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>, ABlockTransferSrcVectorDim,              4,              4,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>, BBlockTransferSrcVectorDim,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>, CDEBlockTransferScalarPerVector>;
+    // clang-format on
+
+    bool isSupported(std::vector<ck::index_t>& ADims,
+                     std::vector<ck::index_t>& BDims,
+                     std::vector<ck::index_t>& DDims,
+                     std::vector<ck::index_t>& EDims,
+                     std::vector<ck::index_t>& AStrides,
+                     std::vector<ck::index_t>& BStrides,
+                     std::vector<ck::index_t>& DStrides,
+                     std::vector<ck::index_t>& EStrides) const
+    {
+        auto contraction = ContractionDeviceInstance{};
+
+        auto argument = contraction.MakeArgument(nullptr,
+                                                 nullptr,
+                                                 std::array<const void*, 1>{nullptr},
+                                                 nullptr,
+                                                 ADims,
+                                                 AStrides,
+                                                 BDims,
+                                                 BStrides,
+                                                 std::array<std::vector<ck::index_t>, 1>{DDims},
+                                                 std::array<std::vector<ck::index_t>, 1>{DStrides},
+                                                 EDims,
+                                                 EStrides,
+                                                 Pass{},
+                                                 Pass{},
+                                                 Bilinear{1.f, 1.f});
+        return contraction.IsSupportedArgument(argument);
+    }
+};
+
+template <typename DataTypeA,
+          typename DataTypeB,
+          typename DataTypeC,
+          typename DataTypeD,
+          ck::index_t NumDim>
+class ContractionDeviceOpWrapper
+{
+
+    protected:
+    using DeviceOp = ck::tensor_operation::device::DeviceContractionMultipleD<NumDim,
+                                                                              NumDim,
+                                                                              NumDim,
+                                                                              DataTypeA,
+                                                                              DataTypeB,
+                                                                              ck::Tuple<DataTypeC>,
+                                                                              DataTypeD,
+                                                                              Pass,
+                                                                              Pass,
+                                                                              Bilinear>;
+
+    public:
+    bool IsSupportedInstance(std::vector<ck::index_t>& Dims,
+                             std::vector<ck::index_t>& Strides) const
+    {
+
+        bool supported     = false;
+        const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+        for(auto& op_ptr : op_ptrs)
+        {
+            auto argument_ptr =
+                op_ptr->MakeArgumentPointer(nullptr,
+                                            nullptr,
+                                            std::array<const void*, 1>{nullptr},
+                                            nullptr,
+                                            Dims,
+                                            Strides,
+                                            Dims,
+                                            Strides,
+                                            std::array<std::vector<ck::index_t>, 1>{Dims},
+                                            std::array<std::vector<ck::index_t>, 1>{Strides},
+                                            Dims,
+                                            Strides,
+                                            Pass{},
+                                            Pass{},
+                                            Bilinear{1.f, 1.f});
+
+            supported = supported || op_ptr->IsSupportedArgument(argument_ptr.get());
+        }
+        return supported;
+    }
+};
+
+TEST(TestContractionInterface, IncorrectNumDims)
+{
+    std::vector<std::vector<ck::index_t>> Dims    = {{4, 4}, {4, 4, 4, 4}, {4, 4, 4, 4, 4, 4}};
+    std::vector<std::vector<ck::index_t>> Strides = {{1, 1}, {1, 1, 1, 1}, {1, 1, 1, 1, 1, 1}};
+    ContractionDeviceOpWrapper<F32, F32, F32, F32, 1> wrapper_1d;
+    ContractionDeviceOpWrapper<F32, F32, F32, F32, 2> wrapper_2d;
+    ContractionDeviceOpWrapper<F32, F32, F32, F32, 3> wrapper_3d;
+    EXPECT_FALSE(wrapper_1d.IsSupportedInstance(Dims[0], Strides[0]));
+    EXPECT_TRUE(wrapper_2d.IsSupportedInstance(Dims[1], Strides[1]));
+    EXPECT_FALSE(wrapper_3d.IsSupportedInstance(Dims[2], Strides[2]));
+}
+
+TEST(TestContractionInterface, IncorrectDataTypes)
+{
+    std::vector<ck::index_t> Dims    = {4, 4, 4, 4};
+    std::vector<ck::index_t> Strides = {64, 16, 4, 1};
+    ContractionDeviceOpWrapper<F32, F32, F64, F64, 2> wrapper_1;
+    ContractionDeviceOpWrapper<F64, F64, F32, F32, 2> wrapper_2;
+    EXPECT_FALSE(wrapper_1.IsSupportedInstance(Dims, Strides));
+    EXPECT_FALSE(wrapper_2.IsSupportedInstance(Dims, Strides));
+}
+
+TEST(TestContractionSupportedArgs, ABMemoryAccess)
+{
+    std::vector<ck::index_t> Dims           = {4, 4, 4, 4};
+    std::vector<ck::index_t> Strides        = {64, 16, 4, 1};
+    std::vector<ck::index_t> StridesM1      = {4, 1, 64, 16};
+    std::vector<ck::index_t> StridesK1      = {64, 16, 4, 1};
+    std::vector<ck::index_t> InvalidStrides = {4, 4, 4, 4};
+    // Memory access to A
+    ContractionInstanceWrapper<1, 2, 4> wrapperA1;
+    ContractionInstanceWrapper<2, 2, 4> wrapperA2;
+    EXPECT_FALSE(
+        wrapperA1.isSupported(Dims, Dims, Dims, Dims, InvalidStrides, Strides, Strides, Strides));
+    EXPECT_FALSE(
+        wrapperA2.isSupported(Dims, Dims, Dims, Dims, InvalidStrides, Strides, Strides, Strides));
+    EXPECT_TRUE(
+        wrapperA1.isSupported(Dims, Dims, Dims, Dims, StridesM1, Strides, Strides, Strides));
+    EXPECT_TRUE(
+        wrapperA2.isSupported(Dims, Dims, Dims, Dims, StridesK1, Strides, Strides, Strides));
+    // Memory access to B
+    ContractionInstanceWrapper<2, 1, 4> wrapperB1;
+    ContractionInstanceWrapper<2, 2, 4> wrapperB2;
+    EXPECT_FALSE(
+        wrapperB1.isSupported(Dims, Dims, Dims, Dims, Strides, InvalidStrides, Strides, Strides));
+    EXPECT_FALSE(
+        wrapperB2.isSupported(Dims, Dims, Dims, Dims, Strides, InvalidStrides, Strides, Strides));
+    EXPECT_TRUE(
+        wrapperB1.isSupported(Dims, Dims, Dims, Dims, Strides, StridesM1, Strides, Strides));
+    EXPECT_TRUE(
+        wrapperB2.isSupported(Dims, Dims, Dims, Dims, Strides, StridesK1, Strides, Strides));
+}
+
+TEST(TestContractionSupportedArgs, DEMemoryAccess)
+{
+    std::vector<ck::index_t> Dims           = {4, 4, 4, 4};
+    std::vector<ck::index_t> Strides        = {64, 16, 4, 1};
+    std::vector<ck::index_t> InvalidStrides = {64, 16, 1, 4};
+    ContractionInstanceWrapper<2, 2, 4> wrapper;
+    // Memory access to D
+    EXPECT_FALSE(
+        wrapper.isSupported(Dims, Dims, Dims, Dims, Strides, Strides, InvalidStrides, Strides));
+    EXPECT_TRUE(wrapper.isSupported(Dims, Dims, Dims, Dims, Strides, Strides, Strides, Strides));
+    // Memory access to E
+    EXPECT_FALSE(
+        wrapper.isSupported(Dims, Dims, Dims, Dims, Strides, Strides, Strides, InvalidStrides));
+    EXPECT_TRUE(wrapper.isSupported(Dims, Dims, Dims, Dims, Strides, Strides, Strides, Strides));
+}
diff --git a/test/conv_util/CMakeLists.txt b/test/conv_util/CMakeLists.txt
new file mode 100644
index 0000000000..7a46039f15
--- /dev/null
+++ b/test/conv_util/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_gtest_executable(test_conv_util conv_util.cpp)
+target_link_libraries(test_conv_util PRIVATE utility)
diff --git a/test/conv_util/conv_util.cpp b/test/conv_util/conv_util.cpp
new file mode 100644
index 0000000000..6922bbbcc7
--- /dev/null
+++ b/test/conv_util/conv_util.cpp
@@ -0,0 +1,156 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <string>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+
+namespace {
+
+class TestConvUtil : public ::testing::Test
+{
+    public:
+    void SetNDParams(std::size_t ndims, std::size_t s, std::size_t d, std::size_t p)
+    {
+        conv_params = ck::utils::conv::ConvParam(ndims,
+                                                 2,
+                                                 128,
+                                                 192,
+                                                 256,
+                                                 std::vector<ck::index_t>(ndims, 3),
+                                                 std::vector<ck::index_t>(ndims, 71),
+                                                 std::vector<ck::index_t>(ndims, s),
+                                                 std::vector<ck::index_t>(ndims, d),
+                                                 std::vector<ck::index_t>(ndims, p),
+                                                 std::vector<ck::index_t>(ndims, p));
+    }
+
+    protected:
+    // -------  default 2D -------
+    // input GNCHW {2, 128, 192, 71, 71},
+    // weights GKCYX {2, 256, 192, 3, 3},
+    // stride {s, s},
+    // dilations {d, d},
+    // padding {{p, p}, {p, p}
+    ck::utils::conv::ConvParam conv_params;
+};
+
+} // namespace
+
+TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths1D)
+{
+    // stride 2, dilation 1, pad 1
+    SetNDParams(1, 2, 1, 1);
+    std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(
+        out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D."));
+
+    // stride 1, dilation 1, pad 1
+    SetNDParams(1, 1, 1, 1);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(
+        out_spatial_len, std::vector<ck::index_t>{71}, "Error: ConvParams 1D stride {1}."));
+
+    // stride 2, dilation 1, pad 2
+    SetNDParams(1, 2, 1, 2);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
+                                     std::vector<ck::index_t>{37},
+                                     "Error: ConvParams 1D padding left/right {2}."));
+
+    // stride 2, dilation 2, pad 2
+    SetNDParams(1, 2, 2, 2);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(
+        out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D dilation {2}."));
+
+    // stride 3, dilation 2, pad 1
+    SetNDParams(1, 3, 2, 1);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(
+        ck::utils::check_err(out_spatial_len,
+                             std::vector<ck::index_t>{23},
+                             "Error: ConvParams 1D strides{3}, padding {1}, dilations {2}."));
+}
+
+TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths2D)
+{
+    // stride 2, dilation 1, pad 1
+    SetNDParams(2, 2, 1, 1);
+    std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
+                                     std::vector<ck::index_t>{36, 36},
+                                     "Error: ConvParams 2D default constructor."));
+
+    // stride 1, dilation 1, pad 1
+    SetNDParams(2, 1, 1, 1);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(
+        out_spatial_len, std::vector<ck::index_t>{71, 71}, "Error: ConvParams 2D stride {1,1}."));
+
+    // stride 2, dilation 1, pad 2
+    SetNDParams(2, 2, 1, 2);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
+                                     std::vector<ck::index_t>{37, 37},
+                                     "Error: ConvParams 2D padding left/right {2,2}."));
+
+    // stride 2, dilation 2, pad 2
+    SetNDParams(2, 2, 2, 2);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(
+        out_spatial_len, std::vector<ck::index_t>{36, 36}, "Error: ConvParams 2D dilation {2,2}."));
+
+    // stride 3, dilation 2, pad 1
+    SetNDParams(2, 3, 2, 1);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(
+        ck::utils::check_err(out_spatial_len,
+                             std::vector<ck::index_t>{23, 23},
+                             "Error: ConvParams 2D strides{3,3}, padding {1,1}, dilations {2,2}."));
+}
+
+TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths3D)
+{
+    // stride 2, dilation 1, pad 1
+    SetNDParams(3, 2, 1, 1);
+    std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(
+        out_spatial_len, std::vector<ck::index_t>{36, 36, 36}, "Error: ConvParams 3D."));
+
+    // stride 1, dilation 1, pad 1
+    SetNDParams(3, 1, 1, 1);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
+                                     std::vector<ck::index_t>{71, 71, 71},
+                                     "Error: ConvParams 3D stride {1, 1, 1}."));
+
+    // stride 2, dilation 1, pad 2
+    SetNDParams(3, 2, 1, 2);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
+                                     std::vector<ck::index_t>{37, 37, 37},
+                                     "Error: ConvParams 3D padding left/right {2, 2, 2}."));
+
+    // stride 2, dilation 2, pad 2
+    SetNDParams(3, 2, 2, 2);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
+                                     std::vector<ck::index_t>{36, 36, 36},
+                                     "Error: ConvParams 3D dilation {2, 2, 2}."));
+
+    // stride 3, dilation 2, pad 1
+    SetNDParams(3, 3, 2, 1);
+    out_spatial_len = conv_params.GetOutputSpatialLengths();
+    EXPECT_TRUE(ck::utils::check_err(
+        out_spatial_len,
+        std::vector<ck::index_t>{23, 23, 23},
+        "Error: ConvParams 3D strides{3, 3, 3}, padding {1, 1, 1}, dilations {2, 2, 2}."));
+}
diff --git a/test/convnd_bwd_data/CMakeLists.txt b/test/convnd_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000000..f734b46f53
--- /dev/null
+++ b/test/convnd_bwd_data/CMakeLists.txt
@@ -0,0 +1,9 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_gtest_executable(test_convnd_bwd_data convnd_bwd_data.cpp)
+   target_link_libraries(test_convnd_bwd_data PRIVATE utility device_conv1d_bwd_data_instance device_conv2d_bwd_data_instance device_conv3d_bwd_data_instance)
+   set(target 1)
+ endif()
+endforeach()
\ No newline at end of file
diff --git a/test/convnd_bwd_data/convnd_bwd_data.cpp b/test/convnd_bwd_data/convnd_bwd_data.cpp
new file mode 100644
index 0000000000..9d2b6cf577
--- /dev/null
+++ b/test/convnd_bwd_data/convnd_bwd_data.cpp
@@ -0,0 +1,93 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <tuple>
+#include <gtest/gtest.h>
+
+#include "profiler/profile_conv_bwd_data_impl.hpp"
+
+template <typename Tuple>
+class TestConvndBwdData : public ::testing::Test
+{
+    protected:
+    using DataType = std::tuple_element_t<0, Tuple>;
+    std::vector<ck::utils::conv::ConvParam> conv_params;
+
+    template <ck::index_t NDimSpatial>
+    void Run()
+    {
+        for(auto& param : conv_params)
+        {
+            bool pass;
+            EXPECT_FALSE(conv_params.empty());
+            pass = ck::profiler::profile_conv_bwd_data_impl<
+                NDimSpatial,
+                ck::tuple_element_t<NDimSpatial - 1,
+                                    ck::Tuple<ck::tensor_layout::convolution::NWC,
+                                              ck::tensor_layout::convolution::NHWC,
+                                              ck::tensor_layout::convolution::NDHWC>>,
+                ck::tuple_element_t<NDimSpatial - 1,
+                                    ck::Tuple<ck::tensor_layout::convolution::KXC,
+                                              ck::tensor_layout::convolution::KYXC,
+                                              ck::tensor_layout::convolution::KZYXC>>,
+                ck::tuple_element_t<NDimSpatial - 1,
+                                    ck::Tuple<ck::tensor_layout::convolution::NWK,
+                                              ck::tensor_layout::convolution::NHWK,
+                                              ck::tensor_layout::convolution::NDHWK>>,
+                DataType,
+                DataType,
+                DataType>(true,  // do_verification
+                          1,     // init_method integer value
+                          false, // do_log
+                          false, // time_kernel
+                          param);
+            EXPECT_TRUE(pass);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<std::tuple<float>,
+                                     std::tuple<ck::half_t>,
+                                     std::tuple<ck::bhalf_t>,
+                                     std::tuple<std::int8_t>>;
+TYPED_TEST_SUITE(TestConvndBwdData, KernelTypes);
+
+// 1d
+TYPED_TEST(TestConvndBwdData, Conv1dBwdData)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back({1, 1, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
+    this->conv_params.push_back({1, 1, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
+    this->conv_params.push_back({1, 1, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
+    this->template Run<1>();
+}
+
+// 2d
+TYPED_TEST(TestConvndBwdData, Conv2dBwdData)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back(
+        {2, 1, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
+    this->conv_params.push_back(
+        {2, 1, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->conv_params.push_back(
+        {2, 1, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
+    this->template Run<2>();
+}
+
+// 3d
+TYPED_TEST(TestConvndBwdData, Conv3dBwdData)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back(
+        {3, 1, 128, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->conv_params.push_back(
+        {3, 1, 128, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    this->conv_params.push_back(
+        {3, 1, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->template Run<3>();
+}
diff --git a/test/convnd_fwd/CMakeLists.txt b/test/convnd_fwd/CMakeLists.txt
new file mode 100644
index 0000000000..745aceffc9
--- /dev/null
+++ b/test/convnd_fwd/CMakeLists.txt
@@ -0,0 +1,9 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_gtest_executable(test_convnd_fwd convnd_fwd.cpp)
+   target_link_libraries(test_convnd_fwd PRIVATE utility device_conv2d_fwd_instance)
+   set(target 1)
+ endif()
+endforeach()
diff --git a/test/convnd_fwd/convnd_fwd.cpp b/test/convnd_fwd/convnd_fwd.cpp
new file mode 100644
index 0000000000..fe8798ceb8
--- /dev/null
+++ b/test/convnd_fwd/convnd_fwd.cpp
@@ -0,0 +1,92 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <tuple>
+#include <gtest/gtest.h>
+
+#include "profiler/profile_conv_fwd_impl.hpp"
+
+template <typename Tuple>
+class TestConvndFwd : public ::testing::Test
+{
+    protected:
+    using DataType = std::tuple_element_t<0, Tuple>;
+    std::vector<ck::utils::conv::ConvParam> conv_params;
+
+    template <ck::index_t NDimSpatial>
+    void Run()
+    {
+        for(auto& param : conv_params)
+        {
+            bool pass;
+            EXPECT_FALSE(conv_params.empty());
+            pass = ck::profiler::profile_conv_fwd_impl<
+                NDimSpatial,
+                ck::tuple_element_t<NDimSpatial - 1,
+                                    ck::Tuple<ck::tensor_layout::convolution::NWC,
+                                              ck::tensor_layout::convolution::NHWC,
+                                              ck::tensor_layout::convolution::NDHWC>>,
+                ck::tuple_element_t<NDimSpatial - 1,
+                                    ck::Tuple<ck::tensor_layout::convolution::KXC,
+                                              ck::tensor_layout::convolution::KYXC,
+                                              ck::tensor_layout::convolution::KZYXC>>,
+                ck::tuple_element_t<NDimSpatial - 1,
+                                    ck::Tuple<ck::tensor_layout::convolution::NWK,
+                                              ck::tensor_layout::convolution::NHWK,
+                                              ck::tensor_layout::convolution::NDHWK>>,
+                DataType,
+                DataType,
+                DataType>(true,  // do_verification
+                          1,     // init_method integer value
+                          false, // do_log
+                          false, // time_kernel
+                          param);
+            EXPECT_TRUE(pass);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<std::tuple<float>,
+                                     std::tuple<ck::half_t>,
+                                     std::tuple<ck::bhalf_t>,
+                                     std::tuple<std::int8_t>>;
+TYPED_TEST_SUITE(TestConvndFwd, KernelTypes);
+
+// 1d
+TYPED_TEST(TestConvndFwd, Conv1dFwd)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back({1, 1, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
+    this->conv_params.push_back({1, 1, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
+    this->conv_params.push_back({1, 1, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
+    this->template Run<1>();
+}
+
+// 2d
+TYPED_TEST(TestConvndFwd, Conv2dFwd)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back(
+        {2, 1, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
+    this->conv_params.push_back(
+        {2, 1, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->conv_params.push_back(
+        {2, 1, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
+    this->template Run<2>();
+}
+// 3d
+TYPED_TEST(TestConvndFwd, Conv3dFwd)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back(
+        {3, 1, 128, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->conv_params.push_back(
+        {3, 1, 128, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    this->conv_params.push_back(
+        {3, 1, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->template Run<3>();
+}
diff --git a/test/data_type/CMakeLists.txt b/test/data_type/CMakeLists.txt
new file mode 100644
index 0000000000..2b63727f19
--- /dev/null
+++ b/test/data_type/CMakeLists.txt
@@ -0,0 +1,7 @@
+if (USE_BITINT_EXTENSION_INT4)
+  add_gtest_executable(test_int4 int4.cpp)
+  target_link_libraries(test_int4 PRIVATE utility)
+endif()
+
+add_gtest_executable(test_fp8 fp8.cpp)
+target_link_libraries(test_fp8 PRIVATE utility)
diff --git a/test/data_type/fp8.cpp b/test/data_type/fp8.cpp
new file mode 100644
index 0000000000..5004fe9527
--- /dev/null
+++ b/test/data_type/fp8.cpp
@@ -0,0 +1,123 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/type_convert.hpp"
+
+using ck::f8_convert_sr;
+using ck::f8_t;
+using ck::half_t;
+using ck::type_convert;
+
+TEST(FP8, NumericLimits)
+{
+    EXPECT_EQ(ck::NumericLimits<f8_t>::Min(), 0x08);
+    EXPECT_EQ(ck::NumericLimits<f8_t>::Max(), 0x77);
+    EXPECT_EQ(ck::NumericLimits<f8_t>::Lowest(), 0xF7);
+    EXPECT_EQ(ck::NumericLimits<f8_t>::QuietNaN(), 0x80);
+}
+
+TEST(FP8, ConvertFP32Nearest)
+{
+    // fix the tolerance value
+    float abs_tol = 1e-6;
+    // convert 0 float to fp8 and back, check if holds
+    ASSERT_NEAR(0.0f, type_convert<float>(type_convert<f8_t>(0.0f)), abs_tol);
+    // convert minimal float to fp8 and back, check if holds
+    ASSERT_NEAR(std::numeric_limits<float>::min(),
+                type_convert<float>(type_convert<f8_t>(std::numeric_limits<float>::min())),
+                abs_tol);
+    // convert maximal f8_t to float and check if equal to 240.0
+    ASSERT_NEAR(240.0f, type_convert<float>(type_convert<f8_t>(240.0f)), abs_tol);
+    // convert maximal float to fp8 and back, check if clipped to 240.0
+    ASSERT_NEAR(240.0f,
+                type_convert<float>(type_convert<f8_t>(std::numeric_limits<float>::max())),
+                abs_tol);
+    // convert inf float to f8_t and check if it is qNan
+    ASSERT_NEAR(0x80, type_convert<f8_t>(std::numeric_limits<float>::infinity()), abs_tol);
+    // positive float value to fp8 and back, check if holds
+    float pos_float = 0.0078125f;
+    ASSERT_NEAR(pos_float, type_convert<float>(type_convert<f8_t>(pos_float)), abs_tol);
+    // negative float value to fp8 and back, check if holds
+    float neg_float = -0.0156250f;
+    ASSERT_NEAR(neg_float, type_convert<float>(type_convert<f8_t>(neg_float)), abs_tol);
+}
+
+TEST(FP8, ConvertFP32Stochastic)
+{
+    // fix the tolerance value
+    float abs_tol = 1e-6;
+    // convert 0 float to fp8 and back, check if holds
+    ASSERT_NEAR(0.0f, type_convert<float>(f8_convert_sr<f8_t>(0.0f)), abs_tol);
+    // convert minimal float to fp8 and back, check if holds
+    ASSERT_NEAR(std::numeric_limits<float>::min(),
+                type_convert<float>(f8_convert_sr<f8_t>(std::numeric_limits<float>::min())),
+                abs_tol);
+    // convert maximal f8_t to float and check if equal to 240.0
+    ASSERT_NEAR(240.0f, type_convert<float>(f8_convert_sr<f8_t>(240.0f)), abs_tol);
+    // convert maximal float to fp8 and back, check if clipped to 240.0
+    ASSERT_NEAR(240.0f,
+                type_convert<float>(f8_convert_sr<f8_t>(std::numeric_limits<float>::max())),
+                abs_tol);
+    // convert inf float to f8_t and check if it is qNan
+    ASSERT_NEAR(0x80, f8_convert_sr<f8_t>(std::numeric_limits<float>::infinity()), abs_tol);
+    // positive float value to fp8 and back, check if holds
+    float pos_float = 0.0078125f;
+    ASSERT_NEAR(pos_float, type_convert<float>(f8_convert_sr<f8_t>(pos_float)), abs_tol);
+    // negative float value to fp8 and back, check if holds
+    float neg_float = -0.0156250f;
+    ASSERT_NEAR(neg_float, type_convert<float>(f8_convert_sr<f8_t>(neg_float)), abs_tol);
+}
+
+TEST(FP8, ConvertFP16Nearest)
+{
+    // fix the tolerance value
+    float abs_tol = 1e-3;
+    // convert 0 fp16 to fp8 and back, check if holds
+    ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(type_convert<f8_t>(half_t{0.0})), abs_tol);
+    // convert minimal fp16 to fp8 and back, check if holds
+    ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
+                type_convert<half_t>(type_convert<f8_t>(ck::NumericLimits<half_t>::Min())),
+                abs_tol);
+    // convert maximal f8_t to fp16 and check if equal to 240.0
+    ASSERT_NEAR(half_t{240.0}, type_convert<half_t>(type_convert<f8_t>(half_t{240.0})), abs_tol);
+    // convert maximal fp16 to fp8 and back, check if clipped to 240.0
+    ASSERT_NEAR(half_t{240.0},
+                type_convert<half_t>(type_convert<f8_t>(ck::NumericLimits<half_t>::Max())),
+                abs_tol);
+    // convert QuietNaN fp16 to f8_t and check if it is QuietNaN
+    ASSERT_NEAR(0x80, type_convert<f8_t>(ck::NumericLimits<half_t>::QuietNaN()), abs_tol);
+    // positive fp16 value to fp8 and back, check if holds
+    half_t pos_half = half_t{0.0078125};
+    ASSERT_NEAR(pos_half, type_convert<half_t>(type_convert<f8_t>(pos_half)), abs_tol);
+    // negative fp16 value to fp8 and back, check if holds
+    half_t neg_half = half_t{-0.0156250};
+    ASSERT_NEAR(neg_half, type_convert<half_t>(type_convert<f8_t>(neg_half)), abs_tol);
+}
+
+TEST(FP8, ConvertFP16Stochastic)
+{
+    // fix the tolerance value
+    float abs_tol = 1e-3;
+    // convert 0 fp16 to fp8 and back, check if holds
+    ASSERT_NEAR(half_t{0.0}, type_convert<half_t>(f8_convert_sr<f8_t>(half_t{0.0})), abs_tol);
+    // convert minimal fp16 to fp8 and back, check if holds
+    ASSERT_NEAR(ck::NumericLimits<half_t>::Min(),
+                type_convert<half_t>(f8_convert_sr<f8_t>(ck::NumericLimits<half_t>::Min())),
+                abs_tol);
+    // convert maximal f8_t to fp16 and check if equal to 240.0
+    ASSERT_NEAR(half_t{240.0}, type_convert<half_t>(f8_convert_sr<f8_t>(half_t{240.0})), abs_tol);
+    // convert maximal fp16 to fp8 and back, check if clipped to 240.0
+    ASSERT_NEAR(half_t{240.0},
+                type_convert<half_t>(f8_convert_sr<f8_t>(ck::NumericLimits<half_t>::Max())),
+                abs_tol);
+    // convert QuietNaN fp16 to f8_t and check if it is QuietNaN
+    ASSERT_NEAR(0x80, f8_convert_sr<f8_t>(ck::NumericLimits<half_t>::QuietNaN()), abs_tol);
+    // positive fp16 value to fp8 and back, check if holds
+    half_t pos_half = half_t{0.0078125};
+    ASSERT_NEAR(pos_half, type_convert<half_t>(f8_convert_sr<f8_t>(pos_half)), abs_tol);
+    // negative fp16 value to fp8 and back, check if holds
+    half_t neg_half = half_t{-0.0156250};
+    ASSERT_NEAR(neg_half, type_convert<half_t>(f8_convert_sr<f8_t>(neg_half)), abs_tol);
+}
diff --git a/test/data_type/int4.cpp b/test/data_type/int4.cpp
new file mode 100644
index 0000000000..07549c1c48
--- /dev/null
+++ b/test/data_type/int4.cpp
@@ -0,0 +1,211 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <bitset>
+#include <cinttypes>
+#include <cstdint>
+#include <iomanip>
+#include "gtest/gtest.h"
+#include <hip/hip_runtime.h>
+
+#include "ck/host_utility/hip_check_error.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/math_v2.hpp"
+#include "ck/utility/get_id.hpp"
+#include "ck/library/utility/device_memory.hpp"
+
+using ck::int4_t;
+
+TEST(Int4, BaseArithmetic)
+{
+    int4_t a{1};
+    int4_t b{-2};
+    EXPECT_EQ(a + a, int4_t{2});
+    EXPECT_EQ(a - a, int4_t{0});
+    EXPECT_EQ(a + b, int4_t{-1});
+    EXPECT_EQ(a - b, int4_t{3});
+    EXPECT_EQ(a * a, int4_t{1});
+    EXPECT_EQ(a * b, int4_t{-2});
+    EXPECT_EQ(b * b, int4_t{4});
+    EXPECT_EQ(a / b, int4_t{0});
+    a = int4_t{4};
+    EXPECT_EQ(a / b, int4_t{-2});
+    b = int4_t{2};
+    EXPECT_EQ(a % b, int4_t{0});
+}
+
+TEST(Int4, NumericLimits)
+{
+    EXPECT_EQ(ck::NumericLimits<int4_t>::Min(), int4_t{-8});
+    EXPECT_EQ(ck::NumericLimits<int4_t>::Max(), int4_t{7});
+    EXPECT_EQ(ck::NumericLimits<int4_t>::Lowest(), int4_t{-8});
+}
+
+TEST(Int4, MathOpsV2)
+{
+    int4_t a{4};
+    int4_t b{-5};
+
+    EXPECT_EQ(ck::math::abs(a), int4_t{4});
+    EXPECT_EQ(ck::math::abs(b), int4_t{5});
+    EXPECT_FALSE(ck::math::isnan(b));
+}
+
+namespace {
+
+__global__ void copy(const int4_t* src, std::int8_t* dst, ck::index_t N)
+{
+    ck::index_t tid = ck::get_thread_global_1d_id();
+
+    const int8_t* src_i8 = reinterpret_cast<const int8_t*>(src);
+
+    if(tid < N)
+    {
+        for(ck::index_t i = tid; i < N; i += ck::get_grid_size())
+        {
+            dst[i] = src_i8[i];
+        }
+    }
+}
+
+__global__ void copy_with_static_cast(const int4_t* src, std::int8_t* dst, ck::index_t N)
+{
+    ck::index_t tid = ck::get_thread_global_1d_id();
+
+    if(tid < N)
+    {
+        for(ck::index_t i = tid; i < N; i += ck::get_grid_size())
+        {
+            dst[i] = static_cast<std::int8_t>(src[i]);
+        }
+    }
+}
+
+} // anonymous namespace
+
+TEST(Int4, CopyAsI8PositiveValue)
+{
+    constexpr std::size_t SIZE = 100;
+    std::vector<int4_t> h_src_i4(SIZE, 7);
+    std::vector<std::int8_t> h_src_i8(SIZE, 7);
+    std::vector<std::int8_t> h_dst_i8(SIZE, 0);
+
+    DeviceMem d_src_i4(h_src_i4.size() * sizeof(int4_t));
+    DeviceMem d_dst_i8(h_dst_i8.size() * sizeof(std::int8_t));
+
+    d_src_i4.SetZero();
+    d_dst_i8.SetZero();
+
+    d_src_i4.ToDevice(h_src_i4.data());
+
+    copy<<<1, 64>>>(reinterpret_cast<const int4_t*>(d_src_i4.GetDeviceBuffer()),
+                    reinterpret_cast<std::int8_t*>(d_dst_i8.GetDeviceBuffer()),
+                    SIZE);
+    hip_check_error(hipDeviceSynchronize());
+    d_dst_i8.FromDevice(h_dst_i8.data());
+
+    for(std::size_t i = 0; i < SIZE; ++i)
+    {
+        EXPECT_EQ(h_src_i8[i], h_dst_i8[i]);
+    }
+}
+
+TEST(Int4, DISABLED_CopyAsI8NegativeValue)
+{
+    constexpr std::size_t SIZE = 32;
+    std::vector<int4_t> h_src_i4(SIZE, -8);
+    std::vector<std::int8_t> h_src_i8(SIZE, -8);
+    std::vector<std::int8_t> h_dst_i8(SIZE, 0);
+
+    DeviceMem d_src_i4(h_src_i4.size() * sizeof(int4_t));
+    DeviceMem d_dst_i8(h_dst_i8.size() * sizeof(std::int8_t));
+
+    d_src_i4.SetZero();
+    d_dst_i8.SetZero();
+
+    d_src_i4.ToDevice(h_src_i4.data());
+
+    copy<<<1, 64>>>(reinterpret_cast<const int4_t*>(d_src_i4.GetDeviceBuffer()),
+                    reinterpret_cast<std::int8_t*>(d_dst_i8.GetDeviceBuffer()),
+                    SIZE);
+    hip_check_error(hipDeviceSynchronize());
+    d_dst_i8.FromDevice(h_dst_i8.data());
+
+    for(std::size_t i = 0; i < SIZE; ++i)
+    {
+        EXPECT_EQ(h_src_i8[i], h_dst_i8[i]);
+    }
+}
+
+TEST(Int4, CopyAsI8NegativeValueStaticCast)
+{
+    constexpr std::size_t SIZE = 32;
+    std::vector<int4_t> h_src_i4(SIZE, -8);
+    std::vector<std::int8_t> h_src_i8(SIZE, -8);
+    std::vector<std::int8_t> h_dst_i8(SIZE, 0);
+
+    DeviceMem d_src_i4(h_src_i4.size() * sizeof(int4_t));
+    DeviceMem d_dst_i8(h_dst_i8.size() * sizeof(std::int8_t));
+
+    d_src_i4.SetZero();
+    d_dst_i8.SetZero();
+
+    d_src_i4.ToDevice(h_src_i4.data());
+
+    copy_with_static_cast<<<1, 64>>>(reinterpret_cast<const int4_t*>(d_src_i4.GetDeviceBuffer()),
+                                     reinterpret_cast<std::int8_t*>(d_dst_i8.GetDeviceBuffer()),
+                                     SIZE);
+    hip_check_error(hipDeviceSynchronize());
+    d_dst_i8.FromDevice(h_dst_i8.data());
+
+    for(std::size_t i = 0; i < SIZE; ++i)
+    {
+        EXPECT_EQ(h_src_i8[i], h_dst_i8[i]);
+    }
+}
+
+TEST(Int4, DISABLED_BitwiseRepresentation)
+{
+    using bit8_t = std::bitset<8>;
+
+    int4_t a_i4{3};
+    std::int8_t a_i8 = *reinterpret_cast<std::int8_t*>(&a_i4);
+    std::int8_t b_i8{3};
+#if 0
+    std::cout << std::hex << std::showbase << static_cast<int32_t>(a_i8)
+              << ", " << static_cast<int32_t>(b_i8) << std::endl;
+#endif
+    EXPECT_EQ(bit8_t{static_cast<std::uint64_t>(a_i8)}, bit8_t{static_cast<std::uint64_t>(b_i8)});
+
+    a_i4 = int4_t{-3};
+    a_i8 = *reinterpret_cast<std::int8_t*>(&a_i4);
+    b_i8 = std::int8_t{-3};
+#if 0
+    std::cout << std::hex << std::showbase << static_cast<int32_t>(a_i8)
+              << ", " << static_cast<int32_t>(b_i8) << std::endl;
+#endif
+    EXPECT_EQ(bit8_t{static_cast<std::uint64_t>(a_i8)}, bit8_t{static_cast<std::uint64_t>(b_i8)});
+}
+
+TEST(Int4, BitwiseRepresentationStaticCast)
+{
+    using bit8_t = std::bitset<8>;
+
+    int4_t a_i4{3};
+    std::int8_t a_i8 = static_cast<std::int8_t>(a_i4);
+    std::int8_t b_i8{3};
+#if 0
+    std::cout << std::hex << std::showbase << static_cast<int32_t>(a_i8)
+              << ", " << static_cast<int32_t>(b_i8) << std::endl;
+#endif
+    EXPECT_EQ(bit8_t{static_cast<std::uint64_t>(a_i8)}, bit8_t{static_cast<std::uint64_t>(b_i8)});
+
+    a_i4 = int4_t{-3};
+    a_i8 = static_cast<std::int8_t>(a_i4);
+    b_i8 = std::int8_t{-3};
+#if 0
+    std::cout << std::hex << std::showbase << static_cast<int32_t>(a_i8)
+              << ", " << static_cast<int32_t>(b_i8) << std::endl;
+#endif
+    EXPECT_EQ(bit8_t{static_cast<std::uint64_t>(a_i8)}, bit8_t{static_cast<std::uint64_t>(b_i8)});
+}
diff --git a/test/elementwise_normalization/CMakeLists.txt b/test/elementwise_normalization/CMakeLists.txt
new file mode 100644
index 0000000000..74a3e4999e
--- /dev/null
+++ b/test/elementwise_normalization/CMakeLists.txt
@@ -0,0 +1,6 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  add_custom_target(test_elementwise_normalization)
+  add_gtest_executable(test_elementwise_layernorm_fp16 test_elementwise_layernorm_fp16.cpp)
+  target_link_libraries(test_elementwise_layernorm_fp16 PRIVATE utility device_elementwise_normalization_instance)
+  add_dependencies(test_elementwise_normalization test_elementwise_layernorm_fp16)
+endif()
\ No newline at end of file
diff --git a/test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp b/test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp
new file mode 100644
index 0000000000..d5ce77dc2b
--- /dev/null
+++ b/test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp
@@ -0,0 +1,47 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_elementwise_layernorm_impl.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestElementwiseLayernorm : public ::testing::Test
+{
+    protected:
+    using ADataType     = std::tuple_element_t<0, Tuple>;
+    using BDataType     = std::tuple_element_t<1, Tuple>;
+    using GammaDataType = std::tuple_element_t<2, Tuple>;
+    using BetaDataType  = std::tuple_element_t<3, Tuple>;
+    using AccDataType   = std::tuple_element_t<4, Tuple>;
+    using YDataType     = std::tuple_element_t<5, Tuple>;
+
+    void Run()
+    {
+        // M, N
+        std::vector<std::vector<ck::index_t>> lengths = {
+            {1, 1}, {25, 16}, {39, 777}, {100, 200}, {1024, 1024}, {48 * 256, 2048}, {4096, 8192}};
+
+        for(auto length : lengths)
+        {
+            bool success = ck::profiler::profile_elementwise_layernorm_impl<ADataType,
+                                                                            BDataType,
+                                                                            GammaDataType,
+                                                                            BetaDataType,
+                                                                            AccDataType,
+                                                                            YDataType>(
+                true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<
+    // ADataType, BDataType, GammaDataType, BetaDataType, AccDataType, YDataType>
+    std::tuple<F16, F16, F16, F16, F32, F16>>;
+
+TYPED_TEST_SUITE(TestElementwiseLayernorm, KernelTypes);
+TYPED_TEST(TestElementwiseLayernorm, Test_FP16) { this->Run(); }
diff --git a/test/gemm/CMakeLists.txt b/test/gemm/CMakeLists.txt
new file mode 100644
index 0000000000..7a8836bfef
--- /dev/null
+++ b/test/gemm/CMakeLists.txt
@@ -0,0 +1,30 @@
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+add_test_executable(test_gemm_fp32 gemm_fp32.cpp)
+target_link_libraries(test_gemm_fp32 PRIVATE utility)
+target_link_libraries(test_gemm_fp32 PRIVATE device_gemm_instance)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+add_test_executable(test_gemm_fp16 gemm_fp16.cpp)
+target_link_libraries(test_gemm_fp16 PRIVATE utility)
+target_link_libraries(test_gemm_fp16 PRIVATE device_gemm_instance)
+add_library(gemm_standalone_xdl_fp16_instances STATIC
+    instance/gemm_f16_nn_instance.cpp
+    instance/gemm_f16_nt_instance.cpp
+    instance/gemm_f16_tn_instance.cpp
+    instance/gemm_wavelet_f16_tn_instance.cpp
+    instance/gemm_f16_tt_instance.cpp
+)
+add_test_executable(test_gemm_standalone_xdl_fp16 gemm_standalone_xdl_fp16.cpp)
+target_link_libraries(test_gemm_standalone_xdl_fp16 PRIVATE gemm_standalone_xdl_fp16_instances utility)
+target_include_directories(test_gemm_standalone_xdl_fp16 PRIVATE instance/)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+add_test_executable(test_gemm_bf16 gemm_bf16.cpp)
+target_link_libraries(test_gemm_bf16 PRIVATE utility)
+target_link_libraries(test_gemm_bf16 PRIVATE device_gemm_instance)
+endif()
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+add_test_executable(test_gemm_int8 gemm_int8.cpp)
+target_link_libraries(test_gemm_int8 PRIVATE utility)
+target_link_libraries(test_gemm_int8 PRIVATE device_gemm_instance)
+endif()
\ No newline at end of file
diff --git a/test/gemm/gemm_bf16.cpp b/test/gemm/gemm_bf16.cpp
new file mode 100644
index 0000000000..cde5c45aea
--- /dev/null
+++ b/test/gemm/gemm_bf16.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+#include "test/gemm/gemm_util.hpp"
+
+using ADataType   = ck::bhalf_t;
+using BDataType   = ck::bhalf_t;
+using CDataType   = ck::bhalf_t;
+using AccDataType = float;
+
+#include "run_gemm_test.inc"
+
+int main() { return run_gemm_test(); }
diff --git a/test/gemm/gemm_fp16.cpp b/test/gemm/gemm_fp16.cpp
new file mode 100644
index 0000000000..cad250c6fb
--- /dev/null
+++ b/test/gemm/gemm_fp16.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+#include "test/gemm/gemm_util.hpp"
+
+using ADataType   = ck::half_t;
+using BDataType   = ck::half_t;
+using CDataType   = ck::half_t;
+using AccDataType = float;
+
+#include "run_gemm_test.inc"
+
+int main() { return run_gemm_test(); }
diff --git a/test/gemm/gemm_fp32.cpp b/test/gemm/gemm_fp32.cpp
new file mode 100644
index 0000000000..c35aa77ea7
--- /dev/null
+++ b/test/gemm/gemm_fp32.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+#include "test/gemm/gemm_util.hpp"
+
+using ADataType   = float;
+using BDataType   = float;
+using CDataType   = float;
+using AccDataType = float;
+
+#include "run_gemm_test.inc"
+
+int main() { return run_gemm_test(); }
diff --git a/test/gemm/gemm_fp64.cpp b/test/gemm/gemm_fp64.cpp
new file mode 100644
index 0000000000..e67c8ba4f3
--- /dev/null
+++ b/test/gemm/gemm_fp64.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+#include "test/gemm/gemm_util.hpp"
+
+using ADataType   = double;
+using BDataType   = double;
+using CDataType   = double;
+using AccDataType = double;
+
+#include "run_gemm_test.inc"
+
+int main() { return run_gemm_test(); }
diff --git a/test/gemm/gemm_int8.cpp b/test/gemm/gemm_int8.cpp
new file mode 100644
index 0000000000..6ece05e306
--- /dev/null
+++ b/test/gemm/gemm_int8.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+#include "test/gemm/gemm_util.hpp"
+
+using ADataType   = int8_t;
+using BDataType   = int8_t;
+using CDataType   = int8_t;
+using AccDataType = int32_t;
+
+#include "run_gemm_test.inc"
+
+int main() { return run_gemm_test(); }
diff --git a/test/gemm/gemm_standalone_xdl_fp16.cpp b/test/gemm/gemm_standalone_xdl_fp16.cpp
new file mode 100644
index 0000000000..201a49dcd3
--- /dev/null
+++ b/test/gemm/gemm_standalone_xdl_fp16.cpp
@@ -0,0 +1,167 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gemm_util.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+
+#include "gemm_f16_nn_instance.hpp"
+#include "gemm_f16_nt_instance.hpp"
+#include "gemm_f16_tn_instance.hpp"
+#include "gemm_f16_tt_instance.hpp"
+#include "gemm_wavelet_f16_tn_instance.hpp"
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using F16         = ck::half_t;
+using ADataType   = F16;
+using BDataType   = F16;
+using AccDataType = float;
+using CDataType   = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+using ck::gemm_util::GemmParams;
+using ck::tensor_operation::device::BaseOperator;
+using ck::tensor_operation::device::DeviceGemm;
+using namespace ck::tensor_operation::device::instance;
+
+using DeviceGemmNN =
+    DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>;
+using DeviceGemmNT =
+    DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>;
+using DeviceGemmTN =
+    DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>;
+using DeviceGemmTT =
+    DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>;
+
+struct LayoutConfig
+{
+    bool ARowMajor;
+    bool BRowMajor;
+    bool CRowMajor;
+};
+
+int main(int argc, char* argv[])
+{
+    // Class DeviceGemm is templated by layout and precision types so it is not an option to contain
+    // them in a single vector. Instead we use abstract BaseOperator class and dynamic_cast() it
+    // upon invocation.
+    // And since DeviceGemm does not expose template arg information, an extra book keeping class
+    // LayoutConfig is used for determining which type a BaseOperator instance should be cast to.
+    using OpFactoryFn = void (*)(std::vector<std::unique_ptr<BaseOperator>>&);
+
+    std::vector<std::tuple<GemmParams, LayoutConfig, OpFactoryFn>> problems = {
+        // clang-format off
+    // 104 tiles
+    {GemmParams{2048, 3328, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_256x256},
+    {GemmParams{2048, 1664, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_256x128},
+    {GemmParams{1024, 1664, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_128x128},
+    {GemmParams{1024,  832, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_128x64},
+    {GemmParams{2048, 3328, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_256x256},
+    {GemmParams{2048, 1664, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_256x128},
+    {GemmParams{1024, 1664, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_128x128},
+    {GemmParams{1024,  832, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_128x64},
+    {GemmParams{2048, 3328, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_256x256},
+    {GemmParams{2048, 1664, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_256x128},
+    {GemmParams{1024, 1664, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_128x128},
+    {GemmParams{1024,  832, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_128x64},
+    {GemmParams{2048, 3328, 4096}, LayoutConfig{true, false, true}, add_gemm_wavelet_f16_tn_256x256},
+    {GemmParams{2048, 1664, 4096}, LayoutConfig{true, false, true}, add_gemm_wavelet_f16_tn_256x128},
+    {GemmParams{1024, 1664, 4096}, LayoutConfig{true, false, true}, add_gemm_wavelet_f16_tn_128x128},
+    {GemmParams{1024,  832, 4096}, LayoutConfig{true, false, true}, add_gemm_wavelet_f16_tn_128x64},
+    {GemmParams{2048, 3328, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_256x256},
+    {GemmParams{2048, 1664, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_256x128},
+    {GemmParams{1024, 1664, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_128x128},
+    {GemmParams{1024,  832, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_128x64},
+    // 110 tiles
+    {GemmParams{2560, 2816, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_256x256},
+    {GemmParams{2560, 1408, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_256x128},
+    {GemmParams{1280, 1408, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_128x128},
+    {GemmParams{1280,  704, 4096}, LayoutConfig{false, false, true}, add_gemm_f16_nn_128x64},
+    {GemmParams{2560, 2816, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_256x256},
+    {GemmParams{2560, 1408, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_256x128},
+    {GemmParams{1280, 1408, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_128x128},
+    {GemmParams{1280,  704, 4096}, LayoutConfig{false, true, true}, add_gemm_f16_nt_128x64},
+    {GemmParams{2560, 2816, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_256x256},
+    {GemmParams{2560, 1408, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_256x128},
+    {GemmParams{1280, 1408, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_128x128},
+    {GemmParams{1280,  704, 4096}, LayoutConfig{true, false, true}, add_gemm_f16_tn_128x64},
+    {GemmParams{2560, 2816, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_256x256},
+    {GemmParams{2560, 1408, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_256x128},
+    {GemmParams{1280, 1408, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_128x128},
+    {GemmParams{1280,  704, 4096}, LayoutConfig{true, true, true}, add_gemm_f16_tt_128x64},
+        // clang-format on
+    };
+
+    bool do_verification = true;
+    bool time_kernel     = true;
+
+    if(argc == 1)
+    {
+        // use default
+    }
+    else if(argc == 3)
+    {
+        do_verification = std::stoi(argv[1]);
+        time_kernel     = std::stoi(argv[2]);
+    }
+    else
+    {
+        std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
+                  << "arg2: time kernel (0=no, 1=yes)" << std::endl;
+        return 0;
+    }
+
+    bool pass = true;
+    for(auto& p : problems)
+    {
+        GemmParams& problem_size          = std::get<0>(p);
+        const LayoutConfig& layout_config = std::get<1>(p);
+        const auto& factory               = std::get<2>(p);
+        std::vector<std::unique_ptr<BaseOperator>> ops;
+        factory(ops);
+
+        // overwrite strides
+        problem_size.StrideA = layout_config.ARowMajor ? problem_size.K : problem_size.M;
+        problem_size.StrideB = layout_config.BRowMajor ? problem_size.N : problem_size.K;
+        problem_size.StrideC = layout_config.CRowMajor ? problem_size.N : problem_size.M;
+
+        if(!layout_config.ARowMajor && !layout_config.BRowMajor)
+        {
+            auto op_ptr = dynamic_cast<DeviceGemmNN*>(ops[0].get());
+            pass &= ck::gemm_util::TestGemm<AccDataType>{}(
+                op_ptr, problem_size, do_verification, time_kernel);
+        }
+        else if(!layout_config.ARowMajor && layout_config.BRowMajor)
+        {
+            auto op_ptr = dynamic_cast<DeviceGemmNT*>(ops[0].get());
+            pass &= ck::gemm_util::TestGemm<AccDataType>{}(
+                op_ptr, problem_size, do_verification, time_kernel);
+        }
+        else if(layout_config.ARowMajor && !layout_config.BRowMajor)
+        {
+            auto op_ptr = dynamic_cast<DeviceGemmTN*>(ops[0].get());
+            pass &= ck::gemm_util::TestGemm<AccDataType>{}(
+                op_ptr, problem_size, do_verification, time_kernel);
+        }
+        else if(layout_config.ARowMajor && layout_config.BRowMajor)
+        {
+            auto op_ptr = dynamic_cast<DeviceGemmTT*>(ops[0].get());
+            pass &= ck::gemm_util::TestGemm<AccDataType>{}(
+                op_ptr, problem_size, do_verification, time_kernel);
+        }
+    }
+
+    std::cout << (pass ? "ALL TESTS PASSED" : "SOME TESTS FAILED") << std::endl;
+    return pass ? 0 : 1;
+}
diff --git a/test/gemm/gemm_util.hpp b/test/gemm/gemm_util.hpp
new file mode 100644
index 0000000000..6c46f4ee89
--- /dev/null
+++ b/test/gemm/gemm_util.hpp
@@ -0,0 +1,267 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace gemm_util {
+
+struct GemmParams
+{
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA = 1024;
+    ck::index_t StrideB = 1024;
+    ck::index_t StrideC = 1024;
+};
+
+template <typename GemmInstance,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+void RunHostGEMM(const Tensor<ADataType>& A,
+                 const Tensor<BDataType>& B,
+                 Tensor<CDataType>& C,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+{
+    auto ref_gemm    = GemmInstance{};
+    auto ref_invoker = ref_gemm.MakeInvoker();
+
+    auto ref_argument = ref_gemm.MakeArgument(A, B, C, a_element_op, b_element_op, c_element_op);
+
+    ref_invoker.Run(ref_argument);
+}
+
+template <typename DeviceGemmPtr_,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+bool RunDeviceGEMM(DeviceGemmPtr_& gemmPtr,
+                   const ck::gemm_util::GemmParams& params,
+                   const Tensor<ADataType>& A,
+                   const Tensor<BDataType>& B,
+                   Tensor<CDataType>& C,
+                   AElementwiseOperation a_element_op,
+                   BElementwiseOperation b_element_op,
+                   CElementwiseOperation c_element_op,
+                   bool time_kernel)
+{
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * A.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * B.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * C.mDesc.GetElementSpaceSize());
+
+    auto invoker_ptr = gemmPtr->MakeInvokerPointer();
+    auto argument_ptr =
+        gemmPtr->MakeArgumentPointer(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+                                     static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+                                     static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+                                     params.M,
+                                     params.N,
+                                     params.K,
+                                     params.StrideA,
+                                     params.StrideB,
+                                     params.StrideC,
+                                     a_element_op,
+                                     b_element_op,
+                                     c_element_op);
+
+    if(gemmPtr->IsSupportedArgument(argument_ptr.get()))
+    {
+        a_m_k_device_buf.ToDevice(A.mData.data());
+        b_k_n_device_buf.ToDevice(B.mData.data());
+        float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t flop      = std::size_t(2) * params.M * params.N * params.K;
+        std::size_t num_btype = sizeof(ADataType) * params.M * params.K +
+                                sizeof(BDataType) * params.K * params.N +
+                                sizeof(CDataType) * params.M * params.N;
+
+        float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << std::endl;
+
+        c_m_n_device_buf.FromDevice(C.mData.data());
+
+        return true;
+    }
+    else
+    {
+        std::cout << "device_gemm with the specified compilation parameters does "
+                     "not support this GEMM problem"
+                  << std::endl;
+
+        return false;
+    }
+}
+
+template <typename AccDataType>
+struct TestGemm
+{
+    template <typename ADataType,
+              typename BDataType,
+              typename CDataType,
+              typename ALayout,
+              typename BLayout,
+              typename CLayout>
+    auto PrepareGemmTensor(const ck::gemm_util::GemmParams& params)
+    {
+        auto f_host_tensor_descriptor =
+            [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+                using namespace ck::literals;
+
+                if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+                {
+                    return HostTensorDescriptor({row, col}, {stride, 1_uz});
+                }
+                else
+                {
+                    return HostTensorDescriptor({row, col}, {1_uz, stride});
+                }
+            };
+
+        Tensor<ADataType> a_m_k(
+            f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
+        Tensor<BDataType> b_k_n(
+            f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
+        Tensor<CDataType> c_m_n_host_result(
+            f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
+        Tensor<CDataType> c_m_n_device_result(
+            f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
+
+        auto f_generate_tensor_value = [](auto& tensor, auto type) {
+            using dataType = decltype(type);
+
+            tensor.GenerateTensorValue(GeneratorTensor_2<dataType>{-5, 5});
+        };
+
+        f_generate_tensor_value(a_m_k, ADataType{});
+        f_generate_tensor_value(b_k_n, BDataType{});
+
+        std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+        std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+        std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+
+        return std::make_tuple(a_m_k, b_k_n, c_m_n_host_result, c_m_n_device_result);
+    }
+
+    template <template <class...> class DeviceGemmPtr_,
+              typename ALayout,
+              typename BLayout,
+              typename CLayout,
+              typename ADataType,
+              typename BDataType,
+              typename CDataType,
+              typename AElementwiseOperation,
+              typename BElementwiseOperation,
+              typename CElementwiseOperation>
+    auto operator()(DeviceGemmPtr_<ALayout,
+                                   BLayout,
+                                   CLayout,
+                                   ADataType,
+                                   BDataType,
+                                   CDataType,
+                                   AElementwiseOperation,
+                                   BElementwiseOperation,
+                                   CElementwiseOperation>* gemmPtr,
+                    const GemmParams& params = GemmParams{},
+                    bool do_verification     = true,
+                    bool time_kernel         = false)
+    {
+        std::cout << "ALayout = " << ALayout{}.name << ", BLayout = " << BLayout{}.name
+                  << ", CLayout = " << CLayout{}.name << std::endl;
+        std::cout << gemmPtr->GetTypeString() << std::endl;
+
+        auto host_tensors =
+            PrepareGemmTensor<ADataType, BDataType, CDataType, ALayout, BLayout, CLayout>(params);
+
+        const Tensor<ADataType>& a  = std::get<0>(host_tensors);
+        const Tensor<BDataType>& b  = std::get<1>(host_tensors);
+        Tensor<CDataType>& c_host   = std::get<2>(host_tensors);
+        Tensor<CDataType>& c_device = std::get<3>(host_tensors);
+
+        auto a_element_op = AElementwiseOperation{};
+        auto b_element_op = BElementwiseOperation{};
+        auto c_element_op = CElementwiseOperation{};
+
+        using ReferenceGemmInstance =
+            ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                      BDataType,
+                                                      CDataType,
+                                                      AccDataType,
+                                                      AElementwiseOperation,
+                                                      BElementwiseOperation,
+                                                      CElementwiseOperation>;
+
+        if(do_verification)
+        {
+            ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
+                a, b, c_host, a_element_op, b_element_op, c_element_op);
+        }
+
+        // Act
+        bool is_supported = ck::gemm_util::RunDeviceGEMM(
+            gemmPtr, params, a, b, c_device, a_element_op, b_element_op, c_element_op, time_kernel);
+
+        if(is_supported && do_verification)
+        {
+            // Assert
+            bool res = false;
+            if(std::is_same<CDataType, float>::value)
+            {
+                res = ck::utils::check_err(c_device, c_host);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, ck::half_t>::value)
+            {
+                res = ck::utils::check_err(c_device, c_host);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, ck::bhalf_t>::value)
+            {
+                res = ck::utils::check_err(c_device, c_host);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, int8_t>::value)
+            {
+                res = ck::utils::check_err(c_device, c_host);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, double>::value)
+            {
+                res = ck::utils::check_err(c_device, c_host);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+
+            return res;
+        }
+        else
+        {
+            return true;
+        }
+    }
+};
+
+} // namespace gemm_util
+} // namespace ck
diff --git a/test/gemm/instance/gemm_f16_nn_instance.cpp b/test/gemm/instance/gemm_f16_nn_instance.cpp
new file mode 100644
index 0000000000..9016257f13
--- /dev/null
+++ b/test/gemm/instance/gemm_f16_nn_instance.cpp
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "gemm_f16_nn_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using gemm_f16_nn_256x256 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   256,    32,   2,   8,   32,   32,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_nn_256x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_nn_128x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   8,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_nn_128x64 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Col,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   8,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_gemm_f16_nn_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nn_256x256{});
+}
+
+void add_gemm_f16_nn_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nn_256x128{});
+}
+
+void add_gemm_f16_nn_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nn_128x128{});
+}
+
+void add_gemm_f16_nn_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nn_128x64{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/test/gemm/instance/gemm_f16_nn_instance.hpp b/test/gemm/instance/gemm_f16_nn_instance.hpp
new file mode 100644
index 0000000000..e174b99a1d
--- /dev/null
+++ b/test/gemm/instance/gemm_f16_nn_instance.hpp
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+void add_gemm_f16_nn_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_nn_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_nn_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_nn_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/test/gemm/instance/gemm_f16_nt_instance.cpp b/test/gemm/instance/gemm_f16_nt_instance.cpp
new file mode 100644
index 0000000000..27103b88d4
--- /dev/null
+++ b/test/gemm/instance/gemm_f16_nt_instance.cpp
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "gemm_f16_nt_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using gemm_f16_nt_256x256 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   256,    32,   2,   2,   32,   32,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_nt_256x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   2,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_nt_128x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   2,   2,   32,   32,    2,    2,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_nt_128x64 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Col,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   2,   2,   32,   32,    2,    1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_gemm_f16_nt_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nt_256x256{});
+}
+
+void add_gemm_f16_nt_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nt_256x128{});
+}
+
+void add_gemm_f16_nt_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nt_128x128{});
+}
+
+void add_gemm_f16_nt_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_nt_128x64{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/test/gemm/instance/gemm_f16_nt_instance.hpp b/test/gemm/instance/gemm_f16_nt_instance.hpp
new file mode 100644
index 0000000000..c624425e69
--- /dev/null
+++ b/test/gemm/instance/gemm_f16_nt_instance.hpp
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+void add_gemm_f16_nt_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_nt_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_nt_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_nt_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/test/gemm/instance/gemm_f16_tn_instance.cpp b/test/gemm/instance/gemm_f16_tn_instance.cpp
new file mode 100644
index 0000000000..5b11f4dad9
--- /dev/null
+++ b/test/gemm/instance/gemm_f16_tn_instance.cpp
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "gemm_f16_tn_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using gemm_f16_tn_256x256 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   256,    32,   8,   8,   32,   32,    4,   4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_tn_256x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_tn_128x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_tn_128x64 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_gemm_f16_tn_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tn_256x256{});
+}
+
+void add_gemm_f16_tn_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tn_256x128{});
+}
+
+void add_gemm_f16_tn_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tn_128x128{});
+}
+
+void add_gemm_f16_tn_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tn_128x64{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/test/gemm/instance/gemm_f16_tn_instance.hpp b/test/gemm/instance/gemm_f16_tn_instance.hpp
new file mode 100644
index 0000000000..563e10600a
--- /dev/null
+++ b/test/gemm/instance/gemm_f16_tn_instance.hpp
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+void add_gemm_f16_tn_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_tn_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_tn_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_tn_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/test/gemm/instance/gemm_f16_tt_instance.cpp b/test/gemm/instance/gemm_f16_tt_instance.cpp
new file mode 100644
index 0000000000..9032150f0c
--- /dev/null
+++ b/test/gemm/instance/gemm_f16_tt_instance.cpp
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+#include "gemm_f16_tt_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using gemm_f16_tt_256x256 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   256,    32,   8,   2,   32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_tt_256x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   256,   128,    32,   8,   2,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_tt_128x128 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   2,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_tt_128x64 = std::tuple<
+    // clang-format off
+        //#####################| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //#####################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //#####################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //#####################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGemm_Xdl_CShuffle<     Row,      Row,    Row,   F16,   F16,   F16,     F32,      F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_gemm_f16_tt_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tt_256x256{});
+}
+
+void add_gemm_f16_tt_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tt_256x128{});
+}
+
+void add_gemm_f16_tt_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tt_128x128{});
+}
+
+void add_gemm_f16_tt_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tt_128x64{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/test/gemm/instance/gemm_f16_tt_instance.hpp b/test/gemm/instance/gemm_f16_tt_instance.hpp
new file mode 100644
index 0000000000..62914d7ac2
--- /dev/null
+++ b/test/gemm/instance/gemm_f16_tt_instance.hpp
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+void add_gemm_f16_tt_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_tt_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_tt_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_f16_tt_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/test/gemm/instance/gemm_wavelet_f16_tn_instance.cpp b/test/gemm/instance/gemm_wavelet_f16_tn_instance.cpp
new file mode 100644
index 0000000000..23ff89da60
--- /dev/null
+++ b/test/gemm/instance/gemm_wavelet_f16_tn_instance.cpp
@@ -0,0 +1,96 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "gemm_wavelet_f16_tn_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using gemm_f16_tn_256x256 = std::tuple<
+    // clang-format off
+    //#####################             | ALayout| BLayout| CLayout| AData| BData| AccData| CShuffle| CData|           A|           B|           C|           GEMM| NumGemmK| ABBlockTransfer|       BlockGemm|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+    //#####################             |        |        |        |  Type|  Type|    Type| DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| ThreadGroupSize| ThreadGroupSize| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+    //#####################             |        |        |        |      |      |        |         |      |   Operation|   Operation|   Operation|               |    Stage|                |                |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+    //#####################             |        |        |        |      |      |        |         |      |            |            |            |               |         |                |                |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+    DeviceGemm_Xdl_WaveletModel_CShuffle<     Row,     Col,     Row,   F16,   F16,     F32,     F16,    F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,             256,             256,   256,   256,    32,   8,   8,   32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_tn_256x128 = std::tuple<
+    // clang-format off
+    //#####################             | ALayout| BLayout| CLayout| AData| BData| AccData| CShuffle| CData|           A|           B|           C|           GEMM| NumGemmK| ABBlockTransfer|       BlockGemm|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+    //#####################             |        |        |        |  Type|  Type|    Type| DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| ThreadGroupSize| ThreadGroupSize| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+    //#####################             |        |        |        |      |      |        |         |      |   Operation|   Operation|   Operation|               |    Stage|                |                |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+    //#####################             |        |        |        |      |      |        |         |      |            |            |            |               |         |                |                |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+    DeviceGemm_Xdl_WaveletModel_CShuffle<     Row,     Col,     Row,   F16,   F16,     F32,      F16,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,             256,             256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_tn_128x128 = std::tuple<
+    // clang-format off
+    //#####################             | ALayout| BLayout| CLayout| AData| BData| AccData| CShuffle| CData|           A|           B|           C|           GEMM| NumGemmK| ABBlockTransfer|       BlockGemm|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+    //#####################             |        |        |        |  Type|  Type|    Type| DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| ThreadGroupSize| ThreadGroupSize| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+    //#####################             |        |        |        |      |      |        |         |      |   Operation|   Operation|   Operation|               |    Stage|                |                |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+    //#####################             |        |        |        |      |      |        |         |      |            |            |            |               |         |                |                |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+    DeviceGemm_Xdl_WaveletModel_CShuffle<     Row,     Col,     Row,   F16,   F16,     F32,      F16,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,             256,             256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+using gemm_f16_tn_128x64 = std::tuple<
+    // clang-format off
+    //#####################             | ALayout| BLayout| CLayout| AData| BData| AccData| CShuffle| CData|           A|           B|           C|           GEMM| NumGemmK| ABBlockTransfer|       BlockGemm|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+    //#####################             |        |        |        |  Type|  Type|    Type| DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| ThreadGroupSize| ThreadGroupSize| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+    //#####################             |        |        |        |      |      |        |         |      |   Operation|   Operation|   Operation|               |    Stage|                |                |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+    //#####################             |        |        |        |      |      |        |         |      |            |            |            |               |         |                |                |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+    DeviceGemm_Xdl_WaveletModel_CShuffle<     Row,     Col,     Row,   F16,   F16,     F32,      F16,   F16, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,             256,             256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_gemm_wavelet_f16_tn_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tn_256x256{});
+}
+
+void add_gemm_wavelet_f16_tn_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tn_256x128{});
+}
+
+void add_gemm_wavelet_f16_tn_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tn_128x128{});
+}
+
+void add_gemm_wavelet_f16_tn_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances)
+{
+    add_device_operation_instances(instances, gemm_f16_tn_128x64{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/test/gemm/instance/gemm_wavelet_f16_tn_instance.hpp b/test/gemm/instance/gemm_wavelet_f16_tn_instance.hpp
new file mode 100644
index 0000000000..ef269d78ee
--- /dev/null
+++ b/test/gemm/instance/gemm_wavelet_f16_tn_instance.hpp
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <memory>
+#include <vector>
+
+#include "include/ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_gemm_wavelet_f16_tn_256x256(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_wavelet_f16_tn_256x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_wavelet_f16_tn_128x128(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+void add_gemm_wavelet_f16_tn_128x64(std::vector<std::unique_ptr<BaseOperator>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
diff --git a/test/gemm/run_gemm_test.inc b/test/gemm/run_gemm_test.inc
new file mode 100644
index 0000000000..d208bb5a7b
--- /dev/null
+++ b/test/gemm/run_gemm_test.inc
@@ -0,0 +1,41 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+int run_gemm_test()
+{
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    auto test = [&](auto a_layout, auto b_layout, auto c_layout) {
+        bool pass = true;
+
+        using DeviceOp = ck::tensor_operation::device::DeviceGemm<decltype(a_layout),
+                                                                  decltype(b_layout),
+                                                                  decltype(c_layout),
+                                                                  ADataType,
+                                                                  BDataType,
+                                                                  CDataType,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>;
+
+        const auto gemmPtrs =
+            ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+                DeviceOp>::GetInstances();
+
+        for(auto& gemmPtr : gemmPtrs)
+        {
+            pass &= ck::gemm_util::TestGemm<AccDataType>{}(gemmPtr.get());
+        }
+
+        return pass;
+    };
+
+    bool pass = test(Row{}, Row{}, Row{}) && test(Row{}, Col{}, Row{}) &&
+                test(Col{}, Row{}, Row{}) && test(Col{}, Col{}, Row{});
+
+    std::cout << "TestGemm ..... " << (pass ? "SUCCESS" : "FAILURE") << std::endl;
+    return pass ? 0 : 1;
+}
diff --git a/test/gemm_layernorm/CMakeLists.txt b/test/gemm_layernorm/CMakeLists.txt
new file mode 100644
index 0000000000..ba0a99b67f
--- /dev/null
+++ b/test/gemm_layernorm/CMakeLists.txt
@@ -0,0 +1,13 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+  if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+   add_custom_target(test_gemm_layernorm)
+   add_gtest_executable(test_gemm_add_relu_add_layernorm_fp16 test_gemm_add_relu_add_layernorm_fp16.cpp)
+   target_link_libraries(test_gemm_add_relu_add_layernorm_fp16 PRIVATE utility device_gemm_add_relu_add_layernorm_instance)
+   add_dependencies(test_gemm_layernorm test_gemm_add_relu_add_layernorm_fp16)
+   set(target 1)
+  endif()
+ endif()
+endforeach()
diff --git a/test/gemm_layernorm/test_gemm_add_relu_add_layernorm_fp16.cpp b/test/gemm_layernorm/test_gemm_add_relu_add_layernorm_fp16.cpp
new file mode 100644
index 0000000000..3f05996878
--- /dev/null
+++ b/test/gemm_layernorm/test_gemm_add_relu_add_layernorm_fp16.cpp
@@ -0,0 +1,77 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_gemm_add_relu_add_layernorm_impl.hpp"
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestGemmAddReluAddLayernorm : public ::testing::Test
+{
+    protected:
+    using ADataType        = std::tuple_element_t<0, Tuple>;
+    using BDataType        = std::tuple_element_t<1, Tuple>;
+    using AccDataType      = std::tuple_element_t<2, Tuple>;
+    using D0DataType       = std::tuple_element_t<3, Tuple>;
+    using D1DataType       = std::tuple_element_t<4, Tuple>;
+    using EMeanVarDataType = std::tuple_element_t<5, Tuple>;
+    using GammaDataType    = std::tuple_element_t<6, Tuple>;
+    using BetaDataType     = std::tuple_element_t<7, Tuple>;
+    using HDataType        = std::tuple_element_t<8, Tuple>;
+    using ALayout          = std::tuple_element_t<9, Tuple>;
+    using BLayout          = std::tuple_element_t<10, Tuple>;
+    using D0Layout         = std::tuple_element_t<11, Tuple>;
+    using D1Layout         = std::tuple_element_t<12, Tuple>;
+    using HLayout          = std::tuple_element_t<13, Tuple>;
+
+    void Run()
+    {
+        std::vector<std::vector<ck::index_t>> lengths = {
+            {1024, 1024, 1024}, {2048, 640, 640}, {1, 1, 1}};
+
+        for(auto length : lengths)
+        {
+            int M        = length[0];
+            int N        = length[1];
+            int K        = length[2];
+            int StrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+            int StrideB  = ck::is_same_v<BLayout, Row> ? N : K;
+            int StrideD0 = 0;
+            int StrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
+            int StrideH  = ck::is_same_v<HLayout, Row> ? N : M;
+
+            bool success = ck::profiler::profile_gemm_add_relu_add_layernorm_impl<ADataType,
+                                                                                  BDataType,
+                                                                                  AccDataType,
+                                                                                  D0DataType,
+                                                                                  D1DataType,
+                                                                                  EMeanVarDataType,
+                                                                                  GammaDataType,
+                                                                                  BetaDataType,
+                                                                                  HDataType,
+                                                                                  ALayout,
+                                                                                  BLayout,
+                                                                                  D0Layout,
+                                                                                  D1Layout,
+                                                                                  HLayout>(
+                true, 1, false, false, M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideH);
+
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<
+    std::tuple<F16, F16, F32, F16, F16, F16, F16, F16, F16, Row, Row, Row, Row, Row>,
+    std::tuple<F16, F16, F32, F16, F16, F16, F16, F16, F16, Row, Col, Row, Row, Row>,
+    std::tuple<F16, F16, F32, F16, F16, F16, F16, F16, F16, Col, Row, Row, Row, Row>,
+    std::tuple<F16, F16, F32, F16, F16, F16, F16, F16, F16, Col, Col, Row, Row, Row>>;
+
+TYPED_TEST_SUITE(TestGemmAddReluAddLayernorm, KernelTypes);
+TYPED_TEST(TestGemmAddReluAddLayernorm, Test_FP16) { this->Run(); }
diff --git a/test/gemm_reduce/CMakeLists.txt b/test/gemm_reduce/CMakeLists.txt
new file mode 100644
index 0000000000..43c8d60745
--- /dev/null
+++ b/test/gemm_reduce/CMakeLists.txt
@@ -0,0 +1,5 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  add_test_executable(test_gemm_reduce_fp16 gemm_reduce_fp16.cpp)
+  target_link_libraries(test_gemm_reduce_fp16 PRIVATE utility)
+  target_link_libraries(test_gemm_reduce_fp16 PRIVATE device_gemm_reduce_instance)
+endif()
\ No newline at end of file
diff --git a/test/gemm_reduce/gemm_reduce_fp16.cpp b/test/gemm_reduce/gemm_reduce_fp16.cpp
new file mode 100644
index 0000000000..35a149f52c
--- /dev/null
+++ b/test/gemm_reduce/gemm_reduce_fp16.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "profiler/profile_gemm_reduce_impl.hpp"
+
+int main()
+{
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    int M = 512;
+    int N = 256;
+    int K = 128;
+
+    bool pass = true;
+
+    pass = pass &&
+           ck::profiler::
+               profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Row, Row, Row>(
+                   true, 1, false, false, M, N, K, K, N, N);
+
+    pass = pass &&
+           ck::profiler::
+               profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Row, Col, Row>(
+                   true, 1, false, false, M, N, K, K, K, N);
+
+    pass = pass &&
+           ck::profiler::
+               profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Col, Row, Row>(
+                   true, 1, false, false, M, N, K, M, N, N);
+
+    pass = pass &&
+           ck::profiler::
+               profile_gemm_reduce_impl<ck::half_t, ck::half_t, ck::half_t, float, Col, Col, Row>(
+                   true, 1, false, false, M, N, K, M, K, N);
+
+    if(pass)
+    {
+        std::cout << "test GEMM+Reduce fp16: Pass" << std::endl;
+        return 0;
+    }
+    else
+    {
+        std::cout << "test GEMM+Reduce fp16: Fail" << std::endl;
+        return -1;
+    }
+}
diff --git a/test/gemm_split_k/CMakeLists.txt b/test/gemm_split_k/CMakeLists.txt
new file mode 100644
index 0000000000..caf30fca59
--- /dev/null
+++ b/test/gemm_split_k/CMakeLists.txt
@@ -0,0 +1,9 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_gtest_executable(test_gemm_splitk test_gemm_splitk.cpp)
+   target_link_libraries(test_gemm_splitk PRIVATE utility device_gemm_splitk_instance)
+   set(target 1)
+ endif()
+endforeach()
diff --git a/test/gemm_split_k/test_gemm_splitk.cpp b/test/gemm_split_k/test_gemm_splitk.cpp
new file mode 100644
index 0000000000..9eba5bba37
--- /dev/null
+++ b/test/gemm_split_k/test_gemm_splitk.cpp
@@ -0,0 +1,66 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "gtest/gtest.h"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "test_gemm_splitk_util.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+namespace {
+
+template <typename X, typename Y>
+struct tuple_concat;
+
+template <typename... Xs, typename... Ys>
+struct tuple_concat<std::tuple<Xs...>, std::tuple<Ys...>>
+{
+    using type = std::tuple<Xs..., Ys...>;
+};
+
+} // namespace
+
+template <typename Tuple>
+class TestGemmSplitK_MK_KN
+    : public ck::test::TestGemmSplitK<typename tuple_concat<std::tuple<Row, Row>, Tuple>::type>
+{
+};
+
+template <typename Tuple>
+class TestGemmSplitK_MK_NK
+    : public ck::test::TestGemmSplitK<typename tuple_concat<std::tuple<Row, Col>, Tuple>::type>
+{
+};
+
+template <typename Tuple>
+class TestGemmSplitK_KM_KN
+    : public ck::test::TestGemmSplitK<typename tuple_concat<std::tuple<Col, Row>, Tuple>::type>
+{
+};
+
+template <typename Tuple>
+class TestGemmSplitK_KM_NK
+    : public ck::test::TestGemmSplitK<typename tuple_concat<std::tuple<Col, Col>, Tuple>::type>
+{
+};
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    //         ADataType, BDataType, CDataType
+    std::tuple<      F16,       F16,       F16>,
+    std::tuple<      F32,       F32,       F32>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestGemmSplitK_MK_KN, KernelTypes);
+TYPED_TEST_SUITE(TestGemmSplitK_MK_NK, KernelTypes);
+TYPED_TEST_SUITE(TestGemmSplitK_KM_KN, KernelTypes);
+TYPED_TEST_SUITE(TestGemmSplitK_KM_NK, KernelTypes);
+
+#include "test_gemm_splitk_ut_cases.inc"
diff --git a/test/gemm_split_k/test_gemm_splitk_ut_cases.inc b/test/gemm_split_k/test_gemm_splitk_ut_cases.inc
new file mode 100644
index 0000000000..54b9c6c9e3
--- /dev/null
+++ b/test/gemm_split_k/test_gemm_splitk_ut_cases.inc
@@ -0,0 +1,217 @@
+#pragma once
+
+TYPED_TEST(TestGemmSplitK_MK_KN, SmallM)
+{
+    std::vector<int> Ms{0, 1, 2, 3, 4, 5, 6};
+    constexpr int N = 512;
+    constexpr int K = 320;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmSplitK_MK_NK, SmallM)
+{
+    std::vector<int> Ms{0, 1, 2, 3, 4, 5, 6};
+    constexpr int N = 512;
+    constexpr int K = 320;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = K;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmSplitK_KM_KN, SmallM)
+{
+    std::vector<int> Ms{0, 1, 2, 3, 4, 5, 6};
+    constexpr int N = 512;
+    constexpr int K = 320;
+
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, M, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmSplitK_KM_NK, SmallM)
+{
+    std::vector<int> Ms{0, 1, 2, 3, 4, 5, 6};
+    constexpr int N = 512;
+    constexpr int K = 320;
+
+    constexpr int StrideB = K;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, M, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmSplitK_MK_KN, MidLargeM)
+{
+    std::vector<int> Ms{127, 255, 312, 799, 1573};
+    constexpr int N = 512;
+    constexpr int K = 320;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmSplitK_MK_NK, MidLargeM)
+{
+    std::vector<int> Ms{127, 255, 312, 799, 1573};
+    constexpr int N = 512;
+    constexpr int K = 320;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = K;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmSplitK_KM_KN, MidLargeM)
+{
+    std::vector<int> Ms{127, 255, 312, 799, 1573};
+    constexpr int N = 512;
+    constexpr int K = 320;
+
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, M, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmSplitK_KM_NK, MidLargeM)
+{
+    std::vector<int> Ms{127, 255, 312, 799, 1573};
+    constexpr int N = 512;
+    constexpr int K = 320;
+
+    constexpr int StrideB = K;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, M, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmSplitK_MK_KN, PaddK)
+{
+    std::vector<int> Ms{127};
+    constexpr int N = 512;
+    constexpr int K = 437;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmSplitK_MK_NK, PaddK)
+{
+    std::vector<int> Ms{127};
+    constexpr int N = 512;
+    constexpr int K = 437;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = K;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmSplitK_KM_KN, PaddK)
+{
+    std::vector<int> Ms{127};
+    constexpr int N = 512;
+    constexpr int K = 437;
+
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, M, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmSplitK_KM_NK, PaddK)
+{
+    std::vector<int> Ms{127};
+    constexpr int N = 512;
+    constexpr int K = 437;
+
+    constexpr int StrideB = K;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, M, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmSplitK_MK_KN, Regular)
+{
+    std::vector<int> Ms{512};
+    constexpr int N = 512;
+    constexpr int K = 512;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmSplitK_MK_NK, Regular)
+{
+    std::vector<int> Ms{512};
+    constexpr int N = 512;
+    constexpr int K = 512;
+
+    constexpr int StrideA = K;
+    constexpr int StrideB = K;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, StrideA, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmSplitK_KM_KN, Regular)
+{
+    std::vector<int> Ms{512};
+    constexpr int N = 512;
+    constexpr int K = 512;
+
+    constexpr int StrideB = N;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, M, StrideB, StrideC);
+}
+
+TYPED_TEST(TestGemmSplitK_KM_NK, Regular)
+{
+    std::vector<int> Ms{512};
+    constexpr int N = 512;
+    constexpr int K = 512;
+
+    constexpr int StrideB = K;
+    constexpr int StrideC = N;
+
+    for(int M : Ms)
+        this->Run(M, N, K, M, StrideB, StrideC);
+}
diff --git a/test/gemm_split_k/test_gemm_splitk_util.hpp b/test/gemm_split_k/test_gemm_splitk_util.hpp
new file mode 100644
index 0000000000..8243747a69
--- /dev/null
+++ b/test/gemm_split_k/test_gemm_splitk_util.hpp
@@ -0,0 +1,78 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <string>
+#include <sstream>
+#include <tuple>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "include/ck/utility/data_type.hpp"
+#include "profiler/profile_gemm_splitk_impl.hpp"
+
+namespace ck {
+namespace test {
+
+template <typename Tuple>
+class TestGemmSplitK : public testing::Test
+{
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using F32 = float;
+
+    protected:
+    using ALayout   = std::tuple_element_t<0, Tuple>;
+    using BLayout   = std::tuple_element_t<1, Tuple>;
+    using CLayout   = Row;
+    using ADataType = std::tuple_element_t<2, Tuple>;
+    using BDataType = std::tuple_element_t<3, Tuple>;
+    using CDataType = std::tuple_element_t<4, Tuple>;
+
+    public:
+    static constexpr bool verify_     = true;
+    static constexpr int init_method_ = 1; // decimal value initialization
+    static constexpr bool log_        = false;
+    static constexpr bool bench_      = false; // measure kernel performance
+    std::vector<int> k_batches_;
+
+    void SetUp() override { k_batches_ = {1, 2, 3, 5, 8}; }
+
+    void Run(const int M,
+             const int N,
+             const int K,
+             const int StrideA,
+             const int StrideB,
+             const int StrideC)
+    {
+        for(auto kb : k_batches_)
+        {
+            RunSingle(M, N, K, StrideA, StrideB, StrideC, kb);
+        }
+    }
+
+    void RunSingle(const int M,
+                   const int N,
+                   const int K,
+                   const int StrideA,
+                   const int StrideB,
+                   const int StrideC,
+                   int kbatch = 1)
+    {
+        bool pass = ck::profiler::profile_gemm_splitk_impl<ADataType,
+                                                           BDataType,
+                                                           F32,
+                                                           CDataType,
+                                                           ALayout,
+                                                           BLayout,
+                                                           CLayout>(
+            verify_, init_method_, log_, bench_, M, N, K, StrideA, StrideB, StrideC, kbatch);
+        EXPECT_TRUE(pass);
+    }
+};
+
+} // namespace test
+} // namespace ck
diff --git a/test/grouped_convnd_bwd_data/CMakeLists.txt b/test/grouped_convnd_bwd_data/CMakeLists.txt
new file mode 100644
index 0000000000..436ab2af76
--- /dev/null
+++ b/test/grouped_convnd_bwd_data/CMakeLists.txt
@@ -0,0 +1,6 @@
+if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940")
+    add_gtest_executable(test_grouped_convnd_bwd_data test_grouped_convnd_bwd_data.cpp)
+    target_link_libraries(test_grouped_convnd_bwd_data PRIVATE utility device_grouped_conv2d_bwd_data_instance device_grouped_conv3d_bwd_data_instance)
+    add_gtest_executable(test_grouped_convnd_bwd_data_interface test_grouped_convnd_bwd_data_interface.cpp)
+    target_link_libraries(test_grouped_convnd_bwd_data_interface PRIVATE utility device_grouped_conv2d_bwd_data_instance)
+endif()
\ No newline at end of file
diff --git a/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp b/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp
new file mode 100644
index 0000000000..bdad37fe14
--- /dev/null
+++ b/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp
@@ -0,0 +1,112 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <tuple>
+#include <vector>
+
+#include <gtest/gtest.h>
+
+#include "profiler/profile_grouped_conv_bwd_data_impl.hpp"
+
+template <typename Tuple>
+class TestGroupedConvndBwdData : public ::testing::Test
+{
+    protected:
+    using DataType  = std::tuple_element_t<0, Tuple>;
+    using OutLayout = std::tuple_element_t<1, Tuple>;
+    using WeiLayout = std::tuple_element_t<2, Tuple>;
+    using InLayout  = std::tuple_element_t<3, Tuple>;
+
+    std::vector<ck::utils::conv::ConvParam> conv_params;
+
+    template <ck::index_t NDimSpatial>
+    void Run()
+    {
+        EXPECT_FALSE(conv_params.empty());
+        bool pass = true;
+        for(auto& param : conv_params)
+        {
+            pass = pass && ck::profiler::profile_grouped_conv_bwd_data_impl<NDimSpatial,
+                                                                            OutLayout,
+                                                                            WeiLayout,
+                                                                            InLayout,
+                                                                            DataType,
+                                                                            DataType,
+                                                                            DataType>(
+                               true,  // do_verification
+                               1,     // init_method: integer value
+                               false, // do_log
+                               false, // time_kernel
+                               param);
+        }
+        EXPECT_TRUE(pass);
+    }
+};
+
+using namespace ck::tensor_layout::convolution;
+
+using KernelTypes2d = ::testing::Types<std::tuple<float, GNHWK, GKYXC, GNHWC>,
+                                       std::tuple<ck::half_t, GNHWK, GKYXC, GNHWC>,
+                                       std::tuple<ck::bhalf_t, GNHWK, GKYXC, GNHWC>,
+                                       std::tuple<float, NHWGK, GKYXC, NHWGC>,
+                                       std::tuple<ck::half_t, NHWGK, GKYXC, NHWGC>,
+                                       std::tuple<ck::bhalf_t, NHWGK, GKYXC, NHWGC>>;
+
+using KernelTypes3d = ::testing::Types<std::tuple<float, GNDHWK, GKZYXC, GNDHWC>,
+                                       std::tuple<ck::half_t, GNDHWK, GKZYXC, GNDHWC>,
+                                       std::tuple<ck::bhalf_t, GNDHWK, GKZYXC, GNDHWC>,
+                                       std::tuple<float, NDHWGK, GKZYXC, NDHWGC>,
+                                       std::tuple<ck::half_t, NDHWGK, GKZYXC, NDHWGC>,
+                                       std::tuple<ck::bhalf_t, NDHWGK, GKZYXC, NDHWGC>>;
+
+template <typename Tuple>
+class TestGroupedConvndBwdData2d : public TestGroupedConvndBwdData<Tuple>
+{
+};
+
+template <typename Tuple>
+class TestGroupedConvndBwdData3d : public TestGroupedConvndBwdData<Tuple>
+{
+};
+
+TYPED_TEST_SUITE(TestGroupedConvndBwdData2d, KernelTypes2d);
+TYPED_TEST_SUITE(TestGroupedConvndBwdData3d, KernelTypes3d);
+
+TYPED_TEST(TestGroupedConvndBwdData2d, Test2D)
+{
+    this->conv_params.clear();
+
+    this->conv_params.push_back(
+        {2, 2, 4, 192, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->conv_params.push_back(
+        {2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->conv_params.push_back(
+        {2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
+    this->conv_params.push_back(
+        {2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
+    this->conv_params.push_back({2, 1, 1, 1, 32, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->conv_params.push_back({2, 1, 1, 64, 3, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->conv_params.push_back({2, 1, 1, 1, 1, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->template Run<2>();
+}
+
+TYPED_TEST(TestGroupedConvndBwdData3d, Test3D)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back(
+        {3, 2, 16, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->conv_params.push_back(
+        {3, 2, 2, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    this->conv_params.push_back(
+        {3, 2, 32, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->conv_params.push_back(
+        {3, 1, 1, 1, 32, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    this->conv_params.push_back(
+        {3, 1, 1, 64, 3, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    this->conv_params.push_back(
+        {3, 1, 1, 1, 1, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    this->template Run<3>();
+}
diff --git a/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface.cpp b/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface.cpp
new file mode 100644
index 0000000000..bc592ba665
--- /dev/null
+++ b/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface.cpp
@@ -0,0 +1,178 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
+
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+#include <gtest/gtest.h>
+
+using DataType    = ck::half_t;
+using AccDataType = float;
+using Pass        = ck::tensor_operation::element_wise::PassThrough;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+using ConvBackwardDataSpecialization =
+    ck::tensor_operation::device::ConvolutionBackwardDataSpecialization;
+
+static constexpr auto ConvBwdDataDefault   = ConvBackwardDataSpecialization::Default;
+static constexpr auto Filter1x1Stride1Pad0 = ConvBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+template <typename Tuple, ConvBackwardDataSpecialization ConvSpec>
+class TestGroupedConvndBwdData : public ::testing::Test
+{
+    protected:
+    static constexpr ck::index_t NDimSpatial = 2;
+
+    using OutLayout = std::tuple_element_t<0, Tuple>;
+    using WeiLayout = std::tuple_element_t<1, Tuple>;
+    using InLayout  = std::tuple_element_t<2, Tuple>;
+
+    // clang-format off
+    using GroupedConvBwdDataDeviceInstance = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
+    // ######| NDimSpatial|   ALayout|   BLayout|   DsLayout|  ELayout|       AData|       BData|     AccData|         CShuffle|       DsData|      EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
+    // ######|            |          |          |           |         |        Type|        Type|        Type|         DataType|         Type|       Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
+    // ######|            |          |          |           |         |            |            |            |                 |             |           |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
+    // ######|            |          |          |           |         |            |            |            |                 |             |           |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
+    < NDimSpatial, OutLayout, WeiLayout, ck::Tuple<>, InLayout, DataType, DataType, AccDataType, DataType,   ck::Tuple<>, DataType, Pass, Pass,    Pass,  ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>;
+    // clang-format on
+
+    ck::utils::conv::ConvParam conv_param;
+
+    template <ck::index_t NDimSpatial>
+    bool Run()
+    {
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        std::array<ck::index_t, NDimSpatial + 3> out_lengths{};
+        std::array<ck::index_t, NDimSpatial + 3> out_strides{};
+        std::array<ck::index_t, NDimSpatial + 3> wei_lengths{};
+        std::array<ck::index_t, NDimSpatial + 3> wei_strides{};
+        std::array<ck::index_t, NDimSpatial + 3> in_lengths{};
+        std::array<ck::index_t, NDimSpatial + 3> in_strides{};
+        std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+        std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+        std::array<ck::index_t, NDimSpatial> input_left_pads{};
+        std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+        auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
+
+        copy(out_g_n_k_wos_desc.GetLengths(), out_lengths);
+        copy(out_g_n_k_wos_desc.GetStrides(), out_strides);
+        copy(wei_g_k_c_xs_desc.GetLengths(), wei_lengths);
+        copy(wei_g_k_c_xs_desc.GetStrides(), wei_strides);
+        copy(in_g_n_c_wis_desc.GetLengths(), in_lengths);
+        copy(in_g_n_c_wis_desc.GetStrides(), in_strides);
+        copy(conv_param.conv_filter_strides_, conv_filter_strides);
+        copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+        copy(conv_param.input_left_pads_, input_left_pads);
+        copy(conv_param.input_right_pads_, input_right_pads);
+
+        auto conv = GroupedConvBwdDataDeviceInstance{};
+
+        auto argument = conv.MakeArgument(nullptr,
+                                          nullptr,
+                                          std::array<const void*, 0>{},
+                                          nullptr,
+                                          out_lengths,
+                                          out_strides,
+                                          wei_lengths,
+                                          wei_strides,
+                                          {},
+                                          {},
+                                          in_lengths,
+                                          in_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          Pass{},
+                                          Pass{},
+                                          Pass{});
+        return conv.IsSupportedArgument(argument);
+    }
+};
+
+using GNHWC = ck::tensor_layout::convolution::GNHWC;
+using NHWGC = ck::tensor_layout::convolution::NHWGC;
+
+using GKYXC = ck::tensor_layout::convolution::GKYXC;
+
+using GNHWK = ck::tensor_layout::convolution::GNHWK;
+using NHWGK = ck::tensor_layout::convolution::NHWGK;
+
+using KernelTypes =
+    ::testing::Types<std::tuple<GNHWK, GKYXC, GNHWC>, std::tuple<NHWGK, GKYXC, NHWGC>>;
+
+template <typename Tuple>
+class TestGroupedConvndBwdDataDefault : public TestGroupedConvndBwdData<Tuple, ConvBwdDataDefault>
+{
+};
+
+template <typename Tuple>
+class TestGroupedConvndBwdDataFilter1x1
+    : public TestGroupedConvndBwdData<Tuple, Filter1x1Stride1Pad0>
+{
+};
+
+TYPED_TEST_SUITE(TestGroupedConvndBwdDataDefault, KernelTypes);
+TYPED_TEST_SUITE(TestGroupedConvndBwdDataFilter1x1, KernelTypes);
+
+TYPED_TEST(TestGroupedConvndBwdDataFilter1x1, SpecializationCheck)
+{
+    // Check filter 3,3 instead of 1,1
+    this->conv_param  = {2, 2, 4, 192, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
+    bool is_supported = this->template Run<2>();
+    EXPECT_FALSE(is_supported);
+
+    // Check strides 2,2 instead of 1,1
+    this->conv_param = {2, 2, 4, 192, 192, {1, 1}, {28, 28}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
+    is_supported     = this->template Run<2>();
+    EXPECT_FALSE(is_supported);
+
+    // Check with pad
+    this->conv_param = {2, 2, 4, 192, 192, {1, 1}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}};
+    is_supported     = this->template Run<2>();
+    EXPECT_FALSE(is_supported);
+
+    // Supported version
+    this->conv_param = {2, 2, 4, 192, 192, {1, 1}, {28, 28}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
+    is_supported     = this->template Run<2>();
+    EXPECT_TRUE(is_supported);
+}
+
+TYPED_TEST(TestGroupedConvndBwdDataDefault, VectorLoadCheck)
+{
+    // vector load for A
+    this->conv_param  = {2, 2, 128, 129, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
+    bool is_supported = this->template Run<2>();
+    EXPECT_FALSE(is_supported);
+    // vector load for B, E, Ds
+    this->conv_param = {2, 2, 128, 128, 257, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
+    is_supported     = this->template Run<2>();
+    EXPECT_FALSE(is_supported);
+}
diff --git a/test/grouped_convnd_bwd_weight/CMakeLists.txt b/test/grouped_convnd_bwd_weight/CMakeLists.txt
new file mode 100644
index 0000000000..f1a3f04b77
--- /dev/null
+++ b/test/grouped_convnd_bwd_weight/CMakeLists.txt
@@ -0,0 +1,11 @@
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_gtest_executable(test_grouped_convnd_bwd_weight test_grouped_convnd_bwd_weight.cpp)
+   target_link_libraries(test_grouped_convnd_bwd_weight PRIVATE utility device_grouped_conv1d_bwd_weight_instance device_grouped_conv2d_bwd_weight_instance device_grouped_conv3d_bwd_weight_instance)
+   add_gtest_executable(test_grouped_convnd_bwd_weight_interface test_grouped_convnd_bwd_weight_interface.cpp)
+   target_link_libraries(test_grouped_convnd_bwd_weight_interface PRIVATE utility device_grouped_conv1d_bwd_weight_instance device_grouped_conv2d_bwd_weight_instance device_grouped_conv3d_bwd_weight_instance)
+   set(target 1)
+ endif()
+endforeach()
\ No newline at end of file
diff --git a/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp b/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp
new file mode 100644
index 0000000000..ae15f6c0c9
--- /dev/null
+++ b/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp
@@ -0,0 +1,140 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <tuple>
+#include <vector>
+
+#include <gtest/gtest.h>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "profiler/profile_grouped_conv_bwd_weight_impl.hpp"
+
+template <typename Tuple>
+class TestGroupedConvndBwdWeight : public ::testing::Test
+{
+    protected:
+    using InDataType  = std::tuple_element_t<0, Tuple>;
+    using WeiDataType = std::tuple_element_t<1, Tuple>;
+    using OutDataType = std::tuple_element_t<2, Tuple>;
+    using InLayout    = std::tuple_element_t<3, Tuple>;
+    using WeiLayout   = std::tuple_element_t<4, Tuple>;
+    using OutLayout   = std::tuple_element_t<5, Tuple>;
+    using NDimSpatial = std::tuple_element_t<6, Tuple>;
+
+    std::vector<ck::utils::conv::ConvParam> conv_params;
+    ck::index_t split_k{2};
+
+    void Run()
+    {
+        EXPECT_FALSE(conv_params.empty());
+        bool pass = true;
+
+        for(auto& param : conv_params)
+        {
+            pass = pass && ck::profiler::profile_grouped_conv_bwd_weight_impl<NDimSpatial{},
+                                                                              InLayout,
+                                                                              WeiLayout,
+                                                                              OutLayout,
+                                                                              InDataType,
+                                                                              WeiDataType,
+                                                                              OutDataType>(
+                               true,  // do_verification
+                               1,     // init_method: integer value
+                               false, // do_log
+                               false, // time_kernel
+                               param,
+                               split_k);
+        }
+        EXPECT_TRUE(pass);
+    }
+};
+
+template <typename Tuple>
+class TestGroupedConvndBwdWeight1d : public TestGroupedConvndBwdWeight<Tuple>
+{
+};
+
+template <typename Tuple>
+class TestGroupedConvndBwdWeight2d : public TestGroupedConvndBwdWeight<Tuple>
+{
+};
+
+template <typename Tuple>
+class TestGroupedConvndBwdWeight3d : public TestGroupedConvndBwdWeight<Tuple>
+{
+};
+
+using namespace ck::tensor_layout::convolution;
+
+using KernelTypes1d = ::testing::Types<
+    std::tuple<float, float, float, GNWC, GKXC, GNWK, ck::Number<1>>,
+    std::tuple<ck::half_t, ck::half_t, ck::half_t, GNWC, GKXC, GNWK, ck::Number<1>>,
+    std::tuple<ck::bhalf_t, float, ck::bhalf_t, GNWC, GKXC, GNWK, ck::Number<1>>>;
+using KernelTypes2d = ::testing::Types<
+    std::tuple<float, float, float, GNHWC, GKYXC, GNHWK, ck::Number<2>>,
+    std::tuple<ck::half_t, ck::half_t, ck::half_t, GNHWC, GKYXC, GNHWK, ck::Number<2>>,
+    std::tuple<ck::bhalf_t, float, ck::bhalf_t, GNHWC, GKYXC, GNHWK, ck::Number<2>>,
+    std::tuple<float, float, float, NHWGC, GKYXC, NHWGK, ck::Number<2>>,
+    std::tuple<ck::half_t, ck::half_t, ck::half_t, NHWGC, GKYXC, NHWGK, ck::Number<2>>,
+    std::tuple<ck::bhalf_t, float, ck::bhalf_t, NHWGC, GKYXC, NHWGK, ck::Number<2>>>;
+using KernelTypes3d = ::testing::Types<
+    std::tuple<float, float, float, GNDHWC, GKZYXC, GNDHWK, ck::Number<3>>,
+    std::tuple<ck::half_t, ck::half_t, ck::half_t, GNDHWC, GKZYXC, GNDHWK, ck::Number<3>>,
+    std::tuple<ck::bhalf_t, float, ck::bhalf_t, GNDHWC, GKZYXC, GNDHWK, ck::Number<3>>,
+    std::tuple<float, float, float, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>,
+    std::tuple<ck::half_t, ck::half_t, ck::half_t, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>,
+    std::tuple<ck::bhalf_t, float, ck::bhalf_t, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>>;
+
+TYPED_TEST_SUITE(TestGroupedConvndBwdWeight1d, KernelTypes1d);
+TYPED_TEST_SUITE(TestGroupedConvndBwdWeight2d, KernelTypes2d);
+TYPED_TEST_SUITE(TestGroupedConvndBwdWeight3d, KernelTypes3d);
+
+TYPED_TEST(TestGroupedConvndBwdWeight1d, Test1D)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back({1, 2, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
+    this->conv_params.push_back({1, 2, 32, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
+    this->conv_params.push_back({1, 2, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
+    this->conv_params.push_back({1, 1, 1, 1, 32, {3}, {32}, {1}, {1}, {1}, {1}});
+    this->conv_params.push_back({1, 1, 1, 64, 3, {3}, {32}, {1}, {1}, {1}, {1}});
+    this->conv_params.push_back({1, 1, 1, 1, 1, {3}, {32}, {1}, {1}, {1}, {1}});
+    this->Run();
+}
+
+TYPED_TEST(TestGroupedConvndBwdWeight2d, Test2D)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back(
+        {2, 2, 64, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
+    this->conv_params.push_back(
+        {2, 2, 4, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->conv_params.push_back(
+        {2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
+    this->conv_params.push_back({2, 1, 1, 1, 32, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->conv_params.push_back({2, 1, 1, 64, 3, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->conv_params.push_back({2, 1, 1, 1, 1, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->Run();
+}
+
+TYPED_TEST(TestGroupedConvndBwdWeight3d, Test3D)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back(
+        {3, 2, 16, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->conv_params.push_back(
+        {3, 2, 2, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    this->conv_params.push_back(
+        {3, 2, 32, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->conv_params.push_back(
+        {3, 1, 1, 1, 32, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    this->conv_params.push_back(
+        {3, 1, 1, 64, 3, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    this->conv_params.push_back(
+        {3, 1, 1, 1, 1, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    this->Run();
+}
diff --git a/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface.cpp b/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface.cpp
new file mode 100644
index 0000000000..cfbf13f00e
--- /dev/null
+++ b/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface.cpp
@@ -0,0 +1,179 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp"
+
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+#include <gtest/gtest.h>
+
+using F16         = ck::half_t;
+using F32         = float;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+using ConvolutionBackwardWeightSpecialization =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization;
+
+static constexpr auto ConvBwdWeightDefault = ConvolutionBackwardWeightSpecialization::Default;
+static constexpr auto Filter1x1Stride1Pad0 =
+    ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
+
+template <typename Tuple, ConvolutionBackwardWeightSpecialization ConvSpec>
+class TestGroupedConvndBwdWeight : public ::testing::Test
+{
+    protected:
+    static constexpr ck::index_t NDimSpatial = 2;
+
+    using InLayout  = std::tuple_element_t<2, Tuple>;
+    using WeiLayout = std::tuple_element_t<1, Tuple>;
+    using OutLayout = std::tuple_element_t<0, Tuple>;
+
+    // clang-format off
+    using GroupedConvBwdWeightDeviceInstance = ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Xdl_CShuffle
+        //##########|     Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|                      ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer|
+        //##########|     Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                            Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|
+        //##########| Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|                    Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl|
+        //##########|        |         |          |          |       |        |        |        |            |            |            |                                  |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |
+        <         NDimSpatial,  InLayout, WeiLayout,OutLayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,   128,    32,   128,     4,  8,   32,   32,    1,    2,  S<1, 4, 4,  8>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,              2,              8,              1,      true,  S<1, 4, 16, 2>,  S<0, 3, 1, 2>,  S<0, 2, 1, 3>,             2,              8,              4,      true,           1,           1,   S<1, 32, 1, 4>,               8>;
+    // clang-format on
+
+    ck::utils::conv::ConvParam conv_param;
+    ck::index_t split_k{2};
+
+    template <ck::index_t NDimSpatial>
+    bool Run()
+    {
+
+        const auto in_g_n_c_wis_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+
+        const auto wei_g_k_c_xs_desc =
+            ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
+                conv_param);
+
+        const auto out_g_n_k_wos_desc =
+            ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
+                conv_param);
+
+        std::array<ck::index_t, NDimSpatial + 3> input_lengths{};
+        std::array<ck::index_t, NDimSpatial + 3> filter_lengths{};
+        std::array<ck::index_t, NDimSpatial + 3> output_lengths{};
+        std::array<ck::index_t, NDimSpatial + 3> input_strides{};
+        std::array<ck::index_t, NDimSpatial + 3> weights_strides{};
+        std::array<ck::index_t, NDimSpatial + 3> output_strides{};
+        std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+        std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+        std::array<ck::index_t, NDimSpatial> input_left_pads{};
+        std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+        auto range_copy = [](const auto& from, auto to) { std::copy(begin(from), end(from), to); };
+
+        range_copy(in_g_n_c_wis_desc.GetLengths(), begin(input_lengths));
+        range_copy(in_g_n_c_wis_desc.GetStrides(), begin(input_strides));
+        range_copy(wei_g_k_c_xs_desc.GetLengths(), begin(filter_lengths));
+        range_copy(wei_g_k_c_xs_desc.GetStrides(), begin(weights_strides));
+        range_copy(out_g_n_k_wos_desc.GetLengths(), begin(output_lengths));
+        range_copy(out_g_n_k_wos_desc.GetStrides(), begin(output_strides));
+        range_copy(conv_param.conv_filter_strides_, begin(conv_filter_strides));
+        range_copy(conv_param.conv_filter_dilations_, begin(conv_filter_dilations));
+        range_copy(conv_param.input_left_pads_, begin(input_left_pads));
+        range_copy(conv_param.input_right_pads_, begin(input_right_pads));
+
+        auto conv = GroupedConvBwdWeightDeviceInstance{};
+
+        auto argument = conv.MakeArgument(nullptr,
+                                          nullptr,
+                                          nullptr,
+                                          input_lengths,
+                                          input_strides,
+                                          filter_lengths,
+                                          weights_strides,
+                                          output_lengths,
+                                          output_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          PassThrough{},
+                                          PassThrough{},
+                                          PassThrough{},
+                                          split_k);
+        return conv.IsSupportedArgument(argument);
+    }
+};
+
+using GNHWC = ck::tensor_layout::convolution::GNHWC;
+using NHWGC = ck::tensor_layout::convolution::NHWGC;
+
+using GKYXC = ck::tensor_layout::convolution::GKYXC;
+
+using GNHWK = ck::tensor_layout::convolution::GNHWK;
+using NHWGK = ck::tensor_layout::convolution::NHWGK;
+
+using KernelTypes =
+    ::testing::Types<std::tuple<GNHWK, GKYXC, GNHWC>, std::tuple<NHWGK, GKYXC, NHWGC>>;
+
+template <typename Tuple>
+class TestGroupedConvndBwdWeightDefault
+    : public TestGroupedConvndBwdWeight<Tuple, ConvBwdWeightDefault>
+{
+};
+
+template <typename Tuple>
+class TestGroupedConvndBwdWeightFilter1x1
+    : public TestGroupedConvndBwdWeight<Tuple, Filter1x1Stride1Pad0>
+{
+};
+
+TYPED_TEST_SUITE(TestGroupedConvndBwdWeightDefault, KernelTypes);
+TYPED_TEST_SUITE(TestGroupedConvndBwdWeightFilter1x1, KernelTypes);
+
+TYPED_TEST(TestGroupedConvndBwdWeightFilter1x1, SpecializationCheck)
+{
+    // Check filter 3,3 instead of 1,1
+    this->conv_param  = {2, 2, 4, 192, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
+    bool is_supported = this->template Run<2>();
+    EXPECT_FALSE(is_supported);
+
+    // Check strides 2,2 instead of 1,1
+    this->conv_param = {2, 2, 4, 192, 192, {1, 1}, {28, 28}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
+    is_supported     = this->template Run<2>();
+    EXPECT_FALSE(is_supported);
+
+    // Check with pad
+    this->conv_param = {2, 2, 4, 192, 192, {1, 1}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}};
+    is_supported     = this->template Run<2>();
+    EXPECT_FALSE(is_supported);
+
+    // Supported version
+    this->conv_param = {2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}};
+    is_supported     = this->template Run<2>();
+    EXPECT_TRUE(is_supported);
+}
+
+TYPED_TEST(TestGroupedConvndBwdWeightDefault, VectorLoadCheck)
+{
+    // vector load for A
+    this->conv_param  = {2, 2, 128, 129, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
+    bool is_supported = this->template Run<2>();
+    EXPECT_FALSE(is_supported);
+    // vector load for B, E, Ds
+    this->conv_param = {2, 2, 128, 128, 257, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}};
+    is_supported     = this->template Run<2>();
+    EXPECT_FALSE(is_supported);
+}
diff --git a/test/grouped_convnd_fwd/CMakeLists.txt b/test/grouped_convnd_fwd/CMakeLists.txt
new file mode 100644
index 0000000000..38da884734
--- /dev/null
+++ b/test/grouped_convnd_fwd/CMakeLists.txt
@@ -0,0 +1,3 @@
+add_gtest_executable(test_grouped_convnd_fwd grouped_convnd_fwd.cpp)
+target_link_libraries(test_grouped_convnd_fwd PRIVATE utility device_grouped_conv1d_fwd_instance device_grouped_conv2d_fwd_instance device_grouped_conv3d_fwd_instance)
+
diff --git a/test/grouped_convnd_fwd/grouped_convnd_fwd.cpp b/test/grouped_convnd_fwd/grouped_convnd_fwd.cpp
new file mode 100644
index 0000000000..c856255ea3
--- /dev/null
+++ b/test/grouped_convnd_fwd/grouped_convnd_fwd.cpp
@@ -0,0 +1,284 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "profiler/profile_grouped_conv_fwd_impl.hpp"
+
+class TestGroupedConvNdFwd : public ::testing::Test
+{
+    protected:
+    std::vector<ck::utils::conv::ConvParam> conv_params;
+};
+
+// 1d GNWC/GKXC/GNWK
+TEST_F(TestGroupedConvNdFwd, GroupedConv1dFwdGNWC)
+{
+    conv_params.clear();
+    conv_params.push_back({1, 2, 128, 128, 256, {1}, {14}, {2}, {1}, {0}, {0}});
+    conv_params.push_back({1, 2, 128, 128, 256, {3}, {28}, {1}, {1}, {1}, {1}});
+    conv_params.push_back({1, 2, 128, 128, 256, {1}, {3}, {1}, {1}, {0}, {0}});
+    conv_params.push_back({1, 1, 1, 1, 32, {3}, {32}, {1}, {1}, {1}, {1}});
+    conv_params.push_back({1, 1, 1, 64, 3, {3}, {32}, {1}, {1}, {1}, {1}});
+
+    for(auto& param : conv_params)
+    {
+        bool pass;
+
+        // fp32
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<1,
+                                                           ck::tensor_layout::convolution::GNWC,
+                                                           ck::tensor_layout::convolution::GKXC,
+                                                           ck::tensor_layout::convolution::GNWK,
+                                                           float,
+                                                           float,
+                                                           float>(true,  // do_verification
+                                                                  1,     // init_method
+                                                                  false, // do_log
+                                                                  false, // time_kernel
+                                                                  param);
+
+        EXPECT_TRUE(pass);
+
+        // fp16
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<1,
+                                                           ck::tensor_layout::convolution::GNWC,
+                                                           ck::tensor_layout::convolution::GKXC,
+                                                           ck::tensor_layout::convolution::GNWK,
+                                                           ck::half_t,
+                                                           ck::half_t,
+                                                           ck::half_t>(true,  // do_verification
+                                                                       1,     // init_method
+                                                                       false, // do_log
+                                                                       false, // time_kernel
+                                                                       param);
+
+        EXPECT_TRUE(pass);
+
+        // bf16
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<1,
+                                                           ck::tensor_layout::convolution::GNWC,
+                                                           ck::tensor_layout::convolution::GKXC,
+                                                           ck::tensor_layout::convolution::GNWK,
+                                                           ck::bhalf_t,
+                                                           ck::bhalf_t,
+                                                           ck::bhalf_t>(true,  // do_verification
+                                                                        1,     // init_method
+                                                                        false, // do_log
+                                                                        false, // time_kernel
+                                                                        param);
+
+        EXPECT_TRUE(pass);
+
+        // int8
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<1,
+                                                           ck::tensor_layout::convolution::GNWC,
+                                                           ck::tensor_layout::convolution::GKXC,
+                                                           ck::tensor_layout::convolution::GNWK,
+                                                           int8_t,
+                                                           int8_t,
+                                                           int8_t>(true,  // do_verification
+                                                                   1,     // init_method
+                                                                   false, // do_log
+                                                                   false, // time_kernel
+                                                                   param);
+
+        EXPECT_TRUE(pass);
+    }
+}
+
+// 2d GNHWC/GKYXC/GNHWK
+TEST_F(TestGroupedConvNdFwd, GroupedConv2dFwdGNHWC)
+{
+    conv_params.clear();
+    conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
+    conv_params.push_back({2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
+    conv_params.push_back({2, 1, 1, 1, 32, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    conv_params.push_back({2, 1, 1, 64, 3, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    conv_params.push_back({2, 1, 1, 1, 1, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+
+    for(auto& param : conv_params)
+    {
+        bool pass;
+
+        // fp32
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
+                                                           ck::tensor_layout::convolution::GNHWC,
+                                                           ck::tensor_layout::convolution::GKYXC,
+                                                           ck::tensor_layout::convolution::GNHWK,
+                                                           float,
+                                                           float,
+                                                           float>(true,  // do_verification
+                                                                  1,     // init_method
+                                                                  false, // do_log
+                                                                  false, // time_kernel
+                                                                  param);
+
+        EXPECT_TRUE(pass);
+
+        // fp16
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
+                                                           ck::tensor_layout::convolution::GNHWC,
+                                                           ck::tensor_layout::convolution::GKYXC,
+                                                           ck::tensor_layout::convolution::GNHWK,
+                                                           ck::half_t,
+                                                           ck::half_t,
+                                                           ck::half_t>(true,  // do_verification
+                                                                       1,     // init_method
+                                                                       false, // do_log
+                                                                       false, // time_kernel
+                                                                       param);
+
+        EXPECT_TRUE(pass);
+
+        // bf16
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
+                                                           ck::tensor_layout::convolution::GNHWC,
+                                                           ck::tensor_layout::convolution::GKYXC,
+                                                           ck::tensor_layout::convolution::GNHWK,
+                                                           ck::bhalf_t,
+                                                           ck::bhalf_t,
+                                                           ck::bhalf_t>(true,  // do_verification
+                                                                        1,     // init_method
+                                                                        false, // do_log
+                                                                        false, // time_kernel
+                                                                        param);
+
+        EXPECT_TRUE(pass);
+
+        // int8
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
+                                                           ck::tensor_layout::convolution::GNHWC,
+                                                           ck::tensor_layout::convolution::GKYXC,
+                                                           ck::tensor_layout::convolution::GNHWK,
+                                                           int8_t,
+                                                           int8_t,
+                                                           int8_t>(true,  // do_verification
+                                                                   1,     // init_method
+                                                                   false, // do_log
+                                                                   false, // time_kernel
+                                                                   param);
+
+        EXPECT_TRUE(pass);
+    }
+}
+
+// 3d GNDHWC/GKZYXC/GNDHWK
+TEST_F(TestGroupedConvNdFwd, GroupedConv3dFwdGNDHWC)
+{
+    conv_params.clear();
+    conv_params.push_back(
+        {3, 2, 128, 128, 256, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    conv_params.push_back(
+        {3, 2, 128, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    conv_params.push_back(
+        {3, 2, 128, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    conv_params.push_back(
+        {3, 1, 1, 1, 32, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    this->conv_params.push_back(
+        {3, 1, 1, 64, 3, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    conv_params.push_back(
+        {3, 1, 1, 1, 1, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+
+    for(auto& param : conv_params)
+    {
+        bool pass;
+
+        // fp32
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<3,
+                                                           ck::tensor_layout::convolution::GNDHWC,
+                                                           ck::tensor_layout::convolution::GKZYXC,
+                                                           ck::tensor_layout::convolution::GNDHWK,
+                                                           float,
+                                                           float,
+                                                           float>(true,  // do_verification
+                                                                  1,     // init_method
+                                                                  false, // do_log
+                                                                  false, // time_kernel
+                                                                  param);
+
+        EXPECT_TRUE(pass);
+
+        // fp16
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<3,
+                                                           ck::tensor_layout::convolution::GNDHWC,
+                                                           ck::tensor_layout::convolution::GKZYXC,
+                                                           ck::tensor_layout::convolution::GNDHWK,
+                                                           ck::half_t,
+                                                           ck::half_t,
+                                                           ck::half_t>(true,  // do_verification
+                                                                       1,     // init_method
+                                                                       false, // do_log
+                                                                       false, // time_kernel
+                                                                       param);
+
+        EXPECT_TRUE(pass);
+
+        // bf16
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<3,
+                                                           ck::tensor_layout::convolution::GNDHWC,
+                                                           ck::tensor_layout::convolution::GKZYXC,
+                                                           ck::tensor_layout::convolution::GNDHWK,
+                                                           ck::bhalf_t,
+                                                           ck::bhalf_t,
+                                                           ck::bhalf_t>(true,  // do_verification
+                                                                        1,     // init_method
+                                                                        false, // do_log
+                                                                        false, // time_kernel
+                                                                        param);
+
+        EXPECT_TRUE(pass);
+
+        // int8
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<3,
+                                                           ck::tensor_layout::convolution::GNDHWC,
+                                                           ck::tensor_layout::convolution::GKZYXC,
+                                                           ck::tensor_layout::convolution::GNDHWK,
+                                                           int8_t,
+                                                           int8_t,
+                                                           int8_t>(true,  // do_verification
+                                                                   1,     // init_method
+                                                                   false, // do_log
+                                                                   false, // time_kernel
+                                                                   param);
+
+        EXPECT_TRUE(pass);
+    }
+}
+
+// 2d NHWGC/KYXGC/NHWGK
+TEST_F(TestGroupedConvNdFwd, GroupedConv2dFwdNHWGC)
+{
+    conv_params.clear();
+    conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
+    conv_params.push_back({2, 2, 128, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    conv_params.push_back({2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
+    conv_params.push_back({2, 1, 1, 1, 32, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    conv_params.push_back({2, 1, 1, 64, 3, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    conv_params.push_back({2, 1, 1, 1, 1, {3, 3}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+
+    for(auto& param : conv_params)
+    {
+        bool pass;
+
+        // fp16
+        pass = ck::profiler::profile_grouped_conv_fwd_impl<2,
+                                                           ck::tensor_layout::convolution::NHWGC,
+                                                           ck::tensor_layout::convolution::GKYXC,
+                                                           ck::tensor_layout::convolution::NHWGK,
+                                                           ck::half_t,
+                                                           ck::half_t,
+                                                           ck::half_t>(true,  // do_verification
+                                                                       1,     // init_method
+                                                                       false, // do_log
+                                                                       false, // time_kernel
+                                                                       param);
+
+        EXPECT_TRUE(pass);
+    }
+}
diff --git a/test/grouped_gemm/CMakeLists.txt b/test/grouped_gemm/CMakeLists.txt
new file mode 100644
index 0000000000..476d953ed8
--- /dev/null
+++ b/test/grouped_gemm/CMakeLists.txt
@@ -0,0 +1,16 @@
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+   add_custom_target(test_grouped_gemm)
+   add_gtest_executable(test_grouped_gemm_splitk test_grouped_gemm_splitk.cpp)
+   add_gtest_executable(test_grouped_gemm_interface test_grouped_gemm_interface.cpp)
+   target_link_libraries(test_grouped_gemm_splitk PRIVATE utility device_grouped_gemm_instance)
+   target_link_libraries(test_grouped_gemm_interface PRIVATE utility device_grouped_gemm_instance)
+   
+   add_dependencies(test_grouped_gemm test_grouped_gemm_splitk test_grouped_gemm_interface)
+   set(target 1)
+ endif()
+endforeach()
+endif()
diff --git a/test/grouped_gemm/test_grouped_gemm_interface.cpp b/test/grouped_gemm/test_grouped_gemm_interface.cpp
new file mode 100644
index 0000000000..6ff3a787e7
--- /dev/null
+++ b/test/grouped_gemm/test_grouped_gemm_interface.cpp
@@ -0,0 +1,202 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <stdexcept>
+#include <vector>
+#include "gtest/gtest.h"
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "test_grouped_gemm_util.hpp"
+
+class TestGGemmSplitKInterface_MKNKMN : public ::testing::Test
+{
+    protected:
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    using ALayout = Row;
+    using BLayout = Col;
+    using ELayout = Row;
+
+    static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+    template <ck::tensor_operation::device::GemmSpecialization GemmSpec,
+              ck::index_t KPerBlock,
+              ck::index_t K1,
+              ck::index_t ABlockTransferSrcScalarPerVector,
+              ck::index_t BBlockTransferSrcScalarPerVector,
+              ck::index_t CDEBlockTransferScalarPerVector_NPerBlock>
+    using GGemmInstance =
+        ck::test::DeviceGroupedGemmSplitkInstanceWrapper<ALayout,
+                                                         BLayout,
+                                                         ELayout,
+                                                         GemmSpec,
+                                                         KPerBlock,
+                                                         K1,
+                                                         ABlockTransferSrcScalarPerVector,
+                                                         BBlockTransferSrcScalarPerVector,
+                                                         CDEBlockTransferScalarPerVector_NPerBlock>;
+
+    using DefaultGGemmInstance = GGemmInstance<GemmDefault, 32, 8, 4, 8, 8>;
+};
+
+TEST_F(TestGGemmSplitKInterface_MKNKMN, TileSize)
+{
+    std::vector<int> Ms{128, 256, 188, 512};
+    constexpr int N = 256;
+    constexpr int K = 128;
+
+    std::vector<int> Ns(Ms.size(), N);
+    std::vector<int> Ks(Ms.size(), K);
+    std::vector<int> StrideAs(Ms.size(), K);
+    std::vector<int> StrideBs(Ms.size(), K);
+    std::vector<int> StrideCs(Ms.size(), N);
+
+    // M % MPerBlock
+    EXPECT_FALSE(DefaultGGemmInstance{}.IsSupported(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs));
+
+    Ms = std::vector<int>{256, 128, 128, 512};
+    Ns = std::vector<int>{256, 177, 128, 512};
+    // N % NPerBlock
+    EXPECT_FALSE(DefaultGGemmInstance{}.IsSupported(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs));
+}
+
+TEST_F(TestGGemmSplitKInterface_MKNKMN, VectorLoadWidth)
+{
+    static constexpr auto GemmMNKPadding =
+        ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+    using PaddedGGemmInstance = GGemmInstance<GemmMNKPadding, 32, 8, 4, 8, 8>;
+
+    std::vector<int> Ms{128, 256, 256, 512};
+    constexpr int N = 256;
+    constexpr int K = 512;
+
+    std::vector<int> Ns(Ms.size(), N);
+    std::vector<int> Ks(Ms.size(), K);
+    std::vector<int> StrideAs(Ms.size(), K);
+    std::vector<int> StrideBs(Ms.size(), K);
+    std::vector<int> StrideCs(Ms.size(), N);
+
+    // K % ABlockTransferSrcScalarPerVector
+    Ks = std::vector<int>{256, 177, 128, 512};
+    EXPECT_FALSE(PaddedGGemmInstance{}.IsSupported(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs));
+
+    Ks = std::vector<int>{256, 164, 128, 512};
+    // K % BBlockTransferSrcScalarPerVector
+    EXPECT_FALSE(PaddedGGemmInstance{}.IsSupported(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs));
+
+    Ks = std::vector<int>(4, 128);
+    Ns = std::vector<int>{256, 127, 128, 512};
+    // N % CBlockTransferScalarPerVector_NWaveNPerXDL
+    EXPECT_FALSE(PaddedGGemmInstance{}.IsSupported(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs));
+}
+
+TEST_F(TestGGemmSplitKInterface_MKNKMN, KLoops)
+{
+    std::vector<int> Ms{128, 256, 256, 512};
+    constexpr int N      = 256;
+    constexpr int K      = 128;
+    constexpr int kbatch = 4;
+
+    std::vector<int> Ns(Ms.size(), N);
+    std::vector<int> Ks(Ms.size(), K);
+    std::vector<int> StrideAs(Ms.size(), K);
+    std::vector<int> StrideBs(Ms.size(), K);
+    std::vector<int> StrideCs(Ms.size(), N);
+
+    // kloops % 2
+    Ks = std::vector<int>{256, 512, 320, 768};
+    EXPECT_TRUE(
+        DefaultGGemmInstance{}.IsSupported(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, kbatch));
+
+    // Not all gemms have same value for main_k0_block_loop!
+    Ks = std::vector<int>{256, 512, 512, 512};
+    EXPECT_THROW(DefaultGGemmInstance{}.Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, kbatch),
+                 std::runtime_error);
+}
+
+class TestGGemmSplitKInterface_KMKNNM : public ::testing::Test
+{
+    protected:
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    using ALayout = Col;
+    using BLayout = Row;
+    using ELayout = Col;
+
+    static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+    template <ck::tensor_operation::device::GemmSpecialization GemmSpec,
+              ck::index_t KPerBlock,
+              ck::index_t K1,
+              ck::index_t ABlockTransferSrcScalarPerVector,
+              ck::index_t BBlockTransferSrcScalarPerVector,
+              ck::index_t CDEBlockTransferScalarPerVector_NPerBlock>
+    using GGemmInstance =
+        ck::test::DeviceGroupedGemmSplitkInstanceWrapper<ALayout,
+                                                         BLayout,
+                                                         ELayout,
+                                                         GemmSpec,
+                                                         KPerBlock,
+                                                         K1,
+                                                         ABlockTransferSrcScalarPerVector,
+                                                         BBlockTransferSrcScalarPerVector,
+                                                         CDEBlockTransferScalarPerVector_NPerBlock>;
+
+    using DefaultGGemmInstance = GGemmInstance<GemmDefault, 32, 8, 4, 8, 4>;
+};
+
+TEST_F(TestGGemmSplitKInterface_KMKNNM, TileSize)
+{
+    std::vector<int> Ms{128, 256, 188, 512};
+    constexpr int N = 256;
+    constexpr int K = 128;
+
+    std::vector<int> Ns(Ms.size(), N);
+    std::vector<int> Ks(Ms.size(), K);
+    std::vector<int> StrideAs(Ms.size(), K);
+    std::vector<int> StrideBs(Ms.size(), K);
+    std::vector<int> StrideCs(Ms.size(), N);
+
+    // M % MPerBlock
+    EXPECT_FALSE(DefaultGGemmInstance{}.IsSupported(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs));
+
+    Ms = std::vector<int>{128, 256, 256, 512};
+    Ns = std::vector<int>{256, 177, 128, 512};
+    // N % NPerBlock
+    EXPECT_FALSE(DefaultGGemmInstance{}.IsSupported(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs));
+}
+
+TEST_F(TestGGemmSplitKInterface_KMKNNM, VectorLoadWidth)
+{
+    static constexpr auto GemmMNKPadding =
+        ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+    using PaddedGGemmInstance = GGemmInstance<GemmMNKPadding, 32, 8, 2, 8, 4>;
+
+    std::vector<int> Ms{128, 256, 256, 512};
+    constexpr int N = 256;
+    constexpr int K = 512;
+
+    std::vector<int> Ns(Ms.size(), N);
+    std::vector<int> Ks(Ms.size(), K);
+    std::vector<int> StrideAs(Ms.size(), K);
+    std::vector<int> StrideBs(Ms.size(), K);
+    std::vector<int> StrideCs(Ms.size(), N);
+
+    // M % ABlockTransferSrcScalarPerVector
+    Ms = std::vector<int>{256, 177, 128, 512};
+    EXPECT_FALSE(PaddedGGemmInstance{}.IsSupported(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs));
+
+    Ms = std::vector<int>{128, 256, 256, 512};
+    Ns = std::vector<int>{256, 164, 128, 512};
+    // N % BBlockTransferSrcScalarPerVector
+    EXPECT_FALSE(PaddedGGemmInstance{}.IsSupported(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs));
+
+    Ns = std::vector<int>{128, 256, 256, 512};
+    Ms = std::vector<int>{256, 130, 128, 512};
+    // M % CBlockTransferScalarPerVector_NWaveNPerXDL
+    EXPECT_FALSE(PaddedGGemmInstance{}.IsSupported(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs));
+}
diff --git a/test/grouped_gemm/test_grouped_gemm_splitk.cpp b/test/grouped_gemm/test_grouped_gemm_splitk.cpp
new file mode 100644
index 0000000000..d9282fa924
--- /dev/null
+++ b/test/grouped_gemm/test_grouped_gemm_splitk.cpp
@@ -0,0 +1,34 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "gtest/gtest.h"
+#include "test_grouped_gemm_util.hpp"
+
+using F16 = ck::half_t;
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using RRR_F16_F16_F16 = ck::test::TestGroupedGemm<std::tuple<Row, Row, Row, F16, F16, F16>>;
+using RCR_F16_F16_F16 = ck::test::TestGroupedGemm<std::tuple<Row, Col, Row, F16, F16, F16>>;
+
+using RRR_F16_F16_F16_LargeK = ck::test::TestGroupedGemm<std::tuple<Row, Row, Row, F16, F16, F16>>;
+using RCR_F16_F16_F16_LargeK = ck::test::TestGroupedGemm<std::tuple<Row, Col, Row, F16, F16, F16>>;
+
+const std::vector<int> KBATCH{1, 2, 3, 5, 8};
+
+INSTANTIATE_TEST_SUITE_P(TestGroupedGemm_splitk_MK_KN, RRR_F16_F16_F16, testing::ValuesIn(KBATCH));
+INSTANTIATE_TEST_SUITE_P(TestGroupedGemm_splitk_MK_NK, RCR_F16_F16_F16, testing::ValuesIn(KBATCH));
+INSTANTIATE_TEST_SUITE_P(TestGroupedGemm_splitk_LargeK_MK_KN,
+                         RRR_F16_F16_F16_LargeK,
+                         testing::Values(32, 64));
+INSTANTIATE_TEST_SUITE_P(TestGroupedGemm_splitk_LargeK_MK_NK,
+                         RCR_F16_F16_F16_LargeK,
+                         testing::Values(32, 64));
+
+#include "test_grouped_gemm_ut_cases.inc"
diff --git a/test/grouped_gemm/test_grouped_gemm_ut_cases.inc b/test/grouped_gemm/test_grouped_gemm_ut_cases.inc
new file mode 100644
index 0000000000..d94d140d97
--- /dev/null
+++ b/test/grouped_gemm/test_grouped_gemm_ut_cases.inc
@@ -0,0 +1,180 @@
+#pragma once
+
+TEST_P(RRR_F16_F16_F16, TinyCases)
+{
+    const std::vector<int> Ms{0, 1};
+    constexpr int N = 768;
+    constexpr int K = 544;
+
+    const std::vector<int> Ns(Ms.size(), N);
+    const std::vector<int> Ks(Ms.size(), K);
+    const std::vector<int> StrideAs(Ms.size(), K);
+    const std::vector<int> StrideBs(Ms.size(), N);
+    const std::vector<int> StrideCs(Ms.size(), N);
+
+    this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
+}
+
+TEST_P(RRR_F16_F16_F16, SmallCases)
+{
+    const std::vector<int> Ms{2, 1, 3, 4, 5, 0};
+    constexpr int N = 768;
+    constexpr int K = 544;
+
+    const std::vector<int> Ns(Ms.size(), N);
+    const std::vector<int> Ks(Ms.size(), K);
+    const std::vector<int> StrideAs(Ms.size(), K);
+    const std::vector<int> StrideBs(Ms.size(), N);
+    const std::vector<int> StrideCs(Ms.size(), N);
+
+    this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
+}
+
+TEST_P(RRR_F16_F16_F16, MidCases)
+{
+    const std::vector<int> Ms{167, 183, 177, 153, 139, 204};
+    constexpr int N = 768;
+    constexpr int K = 544;
+
+    const std::vector<int> Ns(Ms.size(), N);
+    const std::vector<int> Ks(Ms.size(), K);
+    const std::vector<int> StrideAs(Ms.size(), K);
+    const std::vector<int> StrideBs(Ms.size(), N);
+    const std::vector<int> StrideCs(Ms.size(), N);
+
+    this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
+}
+
+TEST_P(RRR_F16_F16_F16, Regular)
+{
+    const std::vector<int> Ms{64, 128, 256};
+    constexpr int N = 768;
+    constexpr int K = 320;
+
+    const std::vector<int> Ns(Ms.size(), N);
+    const std::vector<int> Ks(Ms.size(), K);
+    const std::vector<int> StrideAs(Ms.size(), K);
+    const std::vector<int> StrideBs(Ms.size(), N);
+    const std::vector<int> StrideCs(Ms.size(), N);
+
+    this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
+}
+
+TEST_P(RRR_F16_F16_F16, MNKPadded)
+{
+    const std::vector<int> Ms{127, 150, 188, 210};
+    constexpr int N = 136;
+    constexpr int K = 280;
+
+    const std::vector<int> Ns(Ms.size(), N);
+    const std::vector<int> Ks(Ms.size(), K);
+    const std::vector<int> StrideAs(Ms.size(), K);
+    const std::vector<int> StrideBs(Ms.size(), N);
+    const std::vector<int> StrideCs(Ms.size(), N);
+
+    this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
+}
+
+TEST_P(RCR_F16_F16_F16, TinyCases)
+{
+    const std::vector<int> Ms{0, 1};
+    constexpr int N = 768;
+    constexpr int K = 544;
+
+    const std::vector<int> Ns(Ms.size(), N);
+    const std::vector<int> Ks(Ms.size(), K);
+    const std::vector<int> StrideAs(Ms.size(), K);
+    const std::vector<int> StrideBs(Ms.size(), K);
+    const std::vector<int> StrideCs(Ms.size(), N);
+    this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
+}
+
+TEST_P(RCR_F16_F16_F16, SmallCases)
+{
+    const std::vector<int> Ms{2, 1, 3, 4, 5, 0};
+    constexpr int N = 768;
+    constexpr int K = 544;
+
+    const std::vector<int> Ns(Ms.size(), N);
+    const std::vector<int> Ks(Ms.size(), K);
+    const std::vector<int> StrideAs(Ms.size(), K);
+    const std::vector<int> StrideBs(Ms.size(), K);
+    const std::vector<int> StrideCs(Ms.size(), N);
+
+    this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
+}
+
+TEST_P(RCR_F16_F16_F16, MidCases)
+{
+    const std::vector<int> Ms{167, 183, 177, 153, 139, 204};
+    constexpr int N = 768;
+    constexpr int K = 544;
+
+    const std::vector<int> Ns(Ms.size(), N);
+    const std::vector<int> Ks(Ms.size(), K);
+    const std::vector<int> StrideAs(Ms.size(), K);
+    const std::vector<int> StrideBs(Ms.size(), K);
+    const std::vector<int> StrideCs(Ms.size(), N);
+
+    this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
+}
+
+TEST_P(RCR_F16_F16_F16, Regular)
+{
+    const std::vector<int> Ms{32, 64, 128, 256};
+    constexpr int N = 768;
+    constexpr int K = 320;
+
+    const std::vector<int> Ns(Ms.size(), N);
+    const std::vector<int> Ks(Ms.size(), K);
+    const std::vector<int> StrideAs(Ms.size(), K);
+    const std::vector<int> StrideBs(Ms.size(), K);
+    const std::vector<int> StrideCs(Ms.size(), N);
+
+    this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
+}
+
+TEST_P(RCR_F16_F16_F16, MNKPadded)
+{
+    const std::vector<int> Ms{127, 150, 188, 210};
+    constexpr int N = 136;
+    constexpr int K = 280;
+
+    const std::vector<int> Ns(Ms.size(), N);
+    const std::vector<int> Ks(Ms.size(), K);
+    const std::vector<int> StrideAs(Ms.size(), K);
+    const std::vector<int> StrideBs(Ms.size(), K);
+    const std::vector<int> StrideCs(Ms.size(), N);
+
+    this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
+}
+
+TEST_P(RRR_F16_F16_F16_LargeK, TestLargeKBatch)
+{
+    const std::vector<int> Ms{188, 210};
+    constexpr int N = 768;
+    constexpr int K = 4096;
+
+    const std::vector<int> Ns(Ms.size(), N);
+    const std::vector<int> Ks(Ms.size(), K);
+    const std::vector<int> StrideAs(Ms.size(), K);
+    const std::vector<int> StrideBs(Ms.size(), N);
+    const std::vector<int> StrideCs(Ms.size(), N);
+
+    this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
+}
+
+TEST_P(RCR_F16_F16_F16_LargeK, TestLargeKBatch)
+{
+    const std::vector<int> Ms{188, 210};
+    constexpr int N = 768;
+    constexpr int K = 4096;
+
+    const std::vector<int> Ns(Ms.size(), N);
+    const std::vector<int> Ks(Ms.size(), K);
+    const std::vector<int> StrideAs(Ms.size(), K);
+    const std::vector<int> StrideBs(Ms.size(), K);
+    const std::vector<int> StrideCs(Ms.size(), N);
+
+    this->Run(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, this->GetParam());
+}
diff --git a/test/grouped_gemm/test_grouped_gemm_util.hpp b/test/grouped_gemm/test_grouped_gemm_util.hpp
new file mode 100644
index 0000000000..04b31dcc91
--- /dev/null
+++ b/test/grouped_gemm/test_grouped_gemm_util.hpp
@@ -0,0 +1,249 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+#include <string>
+#include <sstream>
+#include <tuple>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/ck.hpp"
+#include "ck/stream_config.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/utility/tuple.hpp"
+#include "ck/utility/number.hpp"
+#include "profiler/profile_grouped_gemm_impl.hpp"
+
+namespace ck {
+namespace test {
+
+template <typename Range>
+std::string serialize_range(const Range& range)
+{
+    std::stringstream ss;
+    for(auto& r : range)
+    {
+        ss << r << ", ";
+    }
+    std::string str = ss.str();
+    return std::string(str.begin(), str.end() - 2);
+}
+
+template <typename Tuple>
+class TestGroupedGemm : public testing::TestWithParam<int>
+{
+    protected:
+    using ALayout   = std::tuple_element_t<0, Tuple>;
+    using BLayout   = std::tuple_element_t<1, Tuple>;
+    using ELayout   = std::tuple_element_t<2, Tuple>;
+    using ADataType = std::tuple_element_t<3, Tuple>;
+    using BDataType = std::tuple_element_t<4, Tuple>;
+    using EDataType = std::tuple_element_t<5, Tuple>;
+
+    public:
+    static constexpr bool verify_     = true;
+    static constexpr int init_method_ = 1; // decimal value initialization
+    static constexpr bool log_        = false;
+    static constexpr bool bench_      = false; // measure kernel performance
+
+    void SetUp() override {}
+
+    void Run(const std::vector<int>& Ms,
+             const std::vector<int>& Ns,
+             const std::vector<int>& Ks,
+             const std::vector<int>& StrideAs,
+             const std::vector<int>& StrideBs,
+             const std::vector<int>& StrideCs,
+             int kbatch = 1)
+    {
+        bool pass = ck::profiler::profile_grouped_gemm_impl<ADataType,
+                                                            BDataType,
+                                                            EDataType,
+                                                            float,
+                                                            ALayout,
+                                                            BLayout,
+                                                            ELayout>(
+            verify_, init_method_, log_, bench_, Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, kbatch);
+        EXPECT_TRUE(pass);
+    }
+};
+
+template <typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          tensor_operation::device::GemmSpecialization GemmSpec,
+          ck::index_t KPerBlock,
+          ck::index_t K1,
+          ck::index_t ABlockTransferSrcScalarPerVector,
+          ck::index_t BBlockTransferSrcScalarPerVector,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock>
+struct DeviceGroupedGemmSplitkInstanceWrapper
+{
+    using F16         = half_t;
+    using F32         = float;
+    using Row         = ck::tensor_layout::gemm::RowMajor;
+    using Col         = ck::tensor_layout::gemm::ColumnMajor;
+    using PassThrough = tensor_operation::element_wise::PassThrough;
+
+    using EmptyTuple = ck::Tuple<>;
+
+    template <ck::index_t... Is>
+    using S = ck::Sequence<Is...>;
+
+    template <ck::index_t N>
+    using I = ck::Number<N>;
+
+    using ABlockTransferThreadClusterArrageOrder =
+        std::conditional_t<std::is_same_v<ALayout, Row>, S<0, 2, 1, 3>, S<0, 1, 3, 2>>;
+    using ABlockTransferSrcAccessOrder =
+        std::conditional_t<std::is_same_v<ALayout, Row>, S<0, 2, 1, 3>, S<0, 1, 3, 2>>;
+    using ABlockTransferSrcVectorDim = std::conditional_t<std::is_same_v<ALayout, Row>, I<3>, I<2>>;
+    using ABlockTransferDstScalarPerVector_K1 =
+        std::conditional_t<std::is_same_v<ALayout, Row>, I<8>, I<2>>;
+    using ABlockLdsAddExtraM = std::conditional_t<std::is_same_v<ALayout, Row>, I<1>, I<0>>;
+
+    using BBlockTransferThreadClusterArrageOrder =
+        std::conditional_t<std::is_same_v<BLayout, Row>, S<0, 1, 3, 2>, S<0, 2, 1, 3>>;
+    using BBlockTransferSrcAccessOrder =
+        std::conditional_t<std::is_same_v<BLayout, Row>, S<0, 1, 3, 2>, S<0, 2, 1, 3>>;
+    using BBlockTransferSrcVectorDim = std::conditional_t<std::is_same_v<BLayout, Row>, I<2>, I<3>>;
+    using BBlockTransferDstScalarPerVector_K1 =
+        std::conditional_t<std::is_same_v<ALayout, Row>, I<2>, I<8>>;
+    using BBlockLdsAddExtraM = std::conditional_t<std::is_same_v<ALayout, Row>, I<0>, I<1>>;
+
+    using DeviceGroupedGemmSplitKInstance =
+        tensor_operation::device::DeviceGroupedGemmXdlSplitKCShuffle<
+            ALayout,
+            BLayout,
+            EmptyTuple,
+            ELayout,
+            F16,
+            F16,
+            F32,
+            F16,
+            EmptyTuple,
+            F16,
+            PassThrough,
+            PassThrough,
+            PassThrough,
+            GemmSpec,
+            1,
+            128,
+            128,
+            128,
+            KPerBlock,
+            K1,
+            K1,
+            32,
+            32,
+            4,
+            2,
+            S<1, 4, 16, 1>,
+            ABlockTransferThreadClusterArrageOrder,
+            ABlockTransferSrcAccessOrder,
+            ABlockTransferSrcVectorDim::value,
+            ABlockTransferSrcScalarPerVector,
+            ABlockTransferDstScalarPerVector_K1::value,
+            ABlockLdsAddExtraM::value,
+            S<1, 4, 16, 1>,
+            BBlockTransferThreadClusterArrageOrder,
+            BBlockTransferSrcAccessOrder,
+            BBlockTransferSrcVectorDim::value,
+            BBlockTransferSrcScalarPerVector,
+            BBlockTransferDstScalarPerVector_K1::value,
+            BBlockLdsAddExtraM::value,
+            1,
+            1,
+            S<1, 16, 1, 8>,
+            CDEBlockTransferScalarPerVector_NPerBlock>;
+
+    bool IsSupported(const std::vector<int>& Ms,
+                     const std::vector<int>& Ns,
+                     const std::vector<int>& Ks,
+                     const std::vector<int>& StrideAs,
+                     const std::vector<int>& StrideBs,
+                     const std::vector<int>& StrideCs,
+                     int kbatch = 1) const
+    {
+        std::size_t n_groups = Ms.size();
+        EXPECT_TRUE(Ns.size() == n_groups && Ks.size() == n_groups && StrideAs.size() == n_groups &&
+                    StrideBs.size() == n_groups && StrideCs.size() == n_groups)
+            << "The number of groups is not consistent!";
+
+        std::vector<tensor_operation::device::GemmDesc> gemm_descs;
+
+        for(std::size_t i = 0; i < n_groups; ++i)
+        {
+            gemm_descs.push_back(tensor_operation::device::GemmDesc{
+                Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}});
+        }
+
+        std::vector<const void*> p_As(n_groups, nullptr);
+        std::vector<const void*> p_Bs(n_groups, nullptr);
+        std::vector<void*> p_Cs(n_groups, nullptr);
+        auto p_Ds = std::vector<std::array<const void*, 0>>{};
+
+        auto ggemm_instance = DeviceGroupedGemmSplitKInstance{};
+        auto argument       = ggemm_instance.MakeArgument(
+            p_As, p_Bs, p_Ds, p_Cs, gemm_descs, PassThrough{}, PassThrough{}, PassThrough{});
+        if(kbatch > 1)
+        {
+            ggemm_instance.SetKBatchSize(argument, kbatch);
+        }
+
+        return ggemm_instance.IsSupportedArgument(argument);
+    }
+
+    float Run(const std::vector<int>& Ms,
+              const std::vector<int>& Ns,
+              const std::vector<int>& Ks,
+              const std::vector<int>& StrideAs,
+              const std::vector<int>& StrideBs,
+              const std::vector<int>& StrideCs,
+              int kbatch = 1) const
+    {
+        std::size_t n_groups = Ms.size();
+        EXPECT_TRUE(Ns.size() == n_groups && Ks.size() == n_groups && StrideAs.size() == n_groups &&
+                    StrideBs.size() == n_groups && StrideCs.size() == n_groups)
+            << "The number of groups is not consistent!";
+
+        std::vector<tensor_operation::device::GemmDesc> gemm_descs;
+
+        for(std::size_t i = 0; i < n_groups; ++i)
+        {
+            gemm_descs.push_back(tensor_operation::device::GemmDesc{
+                Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}});
+        }
+
+        std::vector<const void*> p_As(n_groups, nullptr);
+        std::vector<const void*> p_Bs(n_groups, nullptr);
+        std::vector<void*> p_Cs(n_groups, nullptr);
+        auto p_Ds = std::vector<std::array<const void*, 0>>{};
+
+        auto ggemm_instance = DeviceGroupedGemmSplitKInstance{};
+        auto argument       = ggemm_instance.MakeArgument(
+            p_As, p_Bs, p_Ds, p_Cs, gemm_descs, PassThrough{}, PassThrough{}, PassThrough{});
+        if(kbatch > 1)
+        {
+            ggemm_instance.SetKBatchSize(argument, kbatch);
+        }
+
+        EXPECT_TRUE(ggemm_instance.IsSupportedArgument(argument));
+        auto invoker = ggemm_instance.MakeInvoker();
+        DeviceMem gemm_desc_workspace(ggemm_instance.GetWorkSpaceSize(&argument));
+        ggemm_instance.SetWorkSpacePointer(&argument, gemm_desc_workspace.GetDeviceBuffer());
+        return invoker.Run(argument, StreamConfig{nullptr, false});
+    }
+};
+
+} // namespace test
+} // namespace ck
diff --git a/test/image_to_column/CMakeLists.txt b/test/image_to_column/CMakeLists.txt
new file mode 100644
index 0000000000..0feb827b55
--- /dev/null
+++ b/test/image_to_column/CMakeLists.txt
@@ -0,0 +1,4 @@
+add_gtest_executable(test_image_to_column test_image_to_column.cpp)
+target_link_libraries(test_image_to_column PRIVATE utility device_image_to_column_instance)
+add_gtest_executable(test_image_to_column_interface test_image_to_column_interface.cpp)
+target_link_libraries(test_image_to_column_interface PRIVATE utility)
diff --git a/test/image_to_column/test_image_to_column.cpp b/test/image_to_column/test_image_to_column.cpp
new file mode 100644
index 0000000000..0b17cac2d0
--- /dev/null
+++ b/test/image_to_column/test_image_to_column.cpp
@@ -0,0 +1,121 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <tuple>
+#include <vector>
+
+#include <gtest/gtest.h>
+
+#include "profiler/profile_image_to_column_impl.hpp"
+
+template <typename Tuple>
+class TestImageToColumn : public ::testing::Test
+{
+    protected:
+    using InDataType  = std::tuple_element_t<0, Tuple>;
+    using OutDataType = std::tuple_element_t<1, Tuple>;
+    using InLayout    = std::tuple_element_t<2, Tuple>;
+
+    std::vector<ck::utils::conv::ConvParam> conv_params;
+
+    template <ck::index_t NDimSpatial>
+    void Run()
+    {
+        EXPECT_FALSE(conv_params.empty());
+        bool pass = true;
+        for(auto& param : conv_params)
+        {
+            pass = pass && ck::profiler::profile_image_to_column_impl<NDimSpatial,
+                                                                      InLayout,
+                                                                      InDataType,
+                                                                      OutDataType>(
+                               true,  // do_verification
+                               1,     // init_method: integer value
+                               false, // do_log
+                               false, // time_kernel
+                               param);
+        }
+        EXPECT_TRUE(pass);
+    }
+};
+
+using namespace ck::tensor_layout::convolution;
+
+using KernelTypes1d = ::testing::Types<std::tuple<float, float, GNWC>,
+                                       std::tuple<ck::bhalf_t, ck::bhalf_t, GNWC>,
+                                       std::tuple<ck::half_t, ck::half_t, GNWC>,
+                                       std::tuple<int8_t, int8_t, GNWC>>;
+
+using KernelTypes2d = ::testing::Types<std::tuple<float, float, GNHWC>,
+                                       std::tuple<ck::bhalf_t, ck::bhalf_t, GNHWC>,
+                                       std::tuple<ck::half_t, ck::half_t, GNHWC>,
+                                       std::tuple<int8_t, int8_t, GNHWC>>;
+
+using KernelTypes3d = ::testing::Types<std::tuple<float, float, GNDHWC>,
+                                       std::tuple<ck::bhalf_t, ck::bhalf_t, GNDHWC>,
+                                       std::tuple<ck::half_t, ck::half_t, GNDHWC>,
+                                       std::tuple<int8_t, int8_t, GNDHWC>>;
+
+template <typename Tuple>
+class TestImageToColumn1d : public TestImageToColumn<Tuple>
+{
+};
+
+template <typename Tuple>
+class TestImageToColumn2d : public TestImageToColumn<Tuple>
+{
+};
+
+template <typename Tuple>
+class TestImageToColumn3d : public TestImageToColumn<Tuple>
+{
+};
+
+TYPED_TEST_SUITE(TestImageToColumn1d, KernelTypes1d);
+TYPED_TEST_SUITE(TestImageToColumn2d, KernelTypes2d);
+TYPED_TEST_SUITE(TestImageToColumn3d, KernelTypes3d);
+
+TYPED_TEST(TestImageToColumn1d, Test1D)
+{
+    this->conv_params.clear();
+
+    this->conv_params.push_back({1, 1, 4, 1, 192, {3}, {28}, {1}, {1}, {1}, {1}});
+    this->conv_params.push_back({1, 1, 64, 1, 64, {3}, {14}, {1}, {1}, {1}, {1}});
+    this->conv_params.push_back({1, 1, 64, 1, 64, {1}, {7}, {2}, {1}, {0}, {0}});
+    this->conv_params.push_back({1, 1, 64, 1, 64, {1}, {3}, {1}, {1}, {0}, {0}});
+    // ScalarPerVector should be 1
+    this->conv_params.push_back({1, 1, 4, 1, 1, {3}, {28}, {1}, {1}, {1}, {1}});
+    // stride != 1
+    this->conv_params.push_back({1, 1, 1, 1, 4, {3}, {28}, {2}, {1}, {1}, {1}});
+    // dilation != 1
+    this->conv_params.push_back({1, 1, 1, 1, 4, {3}, {28}, {1}, {2}, {1}, {1}});
+    this->template Run<1>();
+}
+
+TYPED_TEST(TestImageToColumn2d, Test2D)
+{
+    this->conv_params.clear();
+
+    this->conv_params.push_back(
+        {2, 1, 4, 1, 192, {3, 3}, {28, 28}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->conv_params.push_back(
+        {2, 1, 64, 1, 64, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
+    this->conv_params.push_back({2, 1, 64, 1, 64, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
+    this->conv_params.push_back({2, 1, 64, 1, 64, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
+    this->template Run<2>();
+}
+
+TYPED_TEST(TestImageToColumn3d, Test3D)
+{
+    this->conv_params.clear();
+    this->conv_params.push_back(
+        {3, 1, 16, 1, 64, {1, 1, 1}, {7, 7, 7}, {2, 2, 2}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->conv_params.push_back(
+        {3, 1, 2, 1, 64, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}});
+    this->conv_params.push_back(
+        {3, 1, 32, 1, 64, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}});
+    this->template Run<3>();
+}
diff --git a/test/image_to_column/test_image_to_column_interface.cpp b/test/image_to_column/test_image_to_column_interface.cpp
new file mode 100644
index 0000000000..ea8b9632e1
--- /dev/null
+++ b/test/image_to_column/test_image_to_column_interface.cpp
@@ -0,0 +1,196 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iostream>
+#include <initializer_list>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp"
+
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+
+#include <gtest/gtest.h>
+
+using DataType = float;
+using InLayout = ck::tensor_layout::convolution::GNWC;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+template <ck::index_t ScalarPerVector, bool IsCPacked>
+class TestImageToColumnInterface : public ::testing::Test
+{
+    protected:
+    static constexpr ck::index_t NDimSpatial = 1;
+
+    // clang-format off
+    using DeviceImgToColInstance = ck::tensor_operation::device::DeviceImageToColumnImpl
+        //#####################|        Num| InLayout| InDataType| OutDataType| Block|  MPer|  KPer|    Thread|         Scalar|
+        //#####################|        Dim|         |           |            |  Size| Block| Block|   Cluster|            Per|
+        //#####################|    Spatial|         |           |            |      |      |      |   Lengths|         Vector|
+        //#####################|           |         |           |            |      |      |      |          |               |
+                              < NDimSpatial, InLayout,   DataType,    DataType,   256,   128,   128, S<16, 16>,ScalarPerVector>;
+    // clang-format on
+
+    ck::utils::conv::ConvParam conv_param;
+
+    bool Run()
+    {
+
+        const auto N = conv_param.N_;
+        const auto C = conv_param.C_;
+        const auto FakeC =
+            conv_param.C_ / 2; // Fake C to simulate the behavior that C is not packed
+
+        const ck::index_t NDoHoWo =
+            N *
+            ck::accumulate_n<ck::index_t>(
+                conv_param.output_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+        const ck::index_t CZYX =
+            C *
+            ck::accumulate_n<ck::index_t>(
+                conv_param.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+
+        const auto in_desc =
+            ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
+                conv_param);
+        const auto out_desc = HostTensorDescriptor({NDoHoWo, CZYX});
+
+        std::array<ck::index_t, NDimSpatial> input_spatial_lengths{};
+        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
+        std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
+        std::array<ck::index_t, NDimSpatial + 3> input_g_n_c_wis_strides{};
+        std::array<ck::index_t, 2> output_m_k_strides{};
+        std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+        std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+        std::array<ck::index_t, NDimSpatial> input_left_pads{};
+        std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+        auto copy = [](const auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
+
+        copy(conv_param.input_spatial_lengths_, input_spatial_lengths);
+        copy(conv_param.filter_spatial_lengths_, filter_spatial_lengths);
+        copy(conv_param.output_spatial_lengths_, output_spatial_lengths);
+        copy(in_desc.GetStrides(), input_g_n_c_wis_strides);
+        copy(out_desc.GetStrides(), output_m_k_strides);
+        copy(conv_param.conv_filter_strides_, conv_filter_strides);
+        copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
+        copy(conv_param.input_left_pads_, input_left_pads);
+        copy(conv_param.input_right_pads_, input_right_pads);
+
+        auto img2col  = DeviceImgToColInstance{};
+        auto argument = img2col.MakeArgument(nullptr,
+                                             nullptr,
+                                             N,
+                                             IsCPacked ? C : FakeC,
+                                             input_spatial_lengths,
+                                             filter_spatial_lengths,
+                                             output_spatial_lengths,
+                                             input_g_n_c_wis_strides,
+                                             output_m_k_strides,
+                                             conv_filter_strides,
+                                             conv_filter_dilations,
+                                             input_left_pads,
+                                             input_right_pads);
+
+        return img2col.IsSupportedArgument(argument);
+    }
+};
+
+class TestImageToColumnInterface1ScalarPerVector : public TestImageToColumnInterface<1, true>
+{
+};
+
+class TestImageToColumnInterface4ScalarPerVector : public TestImageToColumnInterface<4, true>
+{
+};
+
+class TestImageToColumnInterface4ScalarPerVectorFakeC : public TestImageToColumnInterface<4, false>
+{
+};
+
+TEST_F(TestImageToColumnInterface1ScalarPerVector, X1ScalarPerVector)
+{
+    // vector load C * X % ScalarPerVector
+    this->conv_param  = {1, 1, 1, 1, 1, {3}, {3}, {1}, {1}, {0}, {0}};
+    bool is_supported = this->Run();
+    EXPECT_TRUE(is_supported);
+    // vector load C * left_pad_x % ScalarPerVector
+    this->conv_param = {1, 1, 1, 1, 1, {4}, {3}, {1}, {1}, {3}, {0}};
+    is_supported     = this->Run();
+    EXPECT_TRUE(is_supported);
+    // vector load C * right_pad_x % ScalarPerVector
+    this->conv_param = {1, 1, 1, 1, 1, {4}, {3}, {1}, {1}, {0}, {3}};
+    is_supported     = this->Run();
+    EXPECT_TRUE(is_supported);
+
+    // vector load C % ScalarPerVector, right_pad and stride
+    this->conv_param = {1, 1, 1, 1, 1, {4}, {3}, {2}, {1}, {0}, {3}};
+    is_supported     = this->Run();
+    EXPECT_TRUE(is_supported);
+    // vector load C % ScalarPerVector, left_pad and stride
+    this->conv_param = {1, 1, 1, 1, 1, {4}, {3}, {2}, {1}, {3}, {0}};
+    is_supported     = this->Run();
+    EXPECT_TRUE(is_supported);
+    // vector load C % ScalarPerVector, dilation
+    this->conv_param = {1, 1, 1, 1, 1, {4}, {3}, {1}, {2}, {0}, {0}};
+    is_supported     = this->Run();
+    EXPECT_TRUE(is_supported);
+
+    // C = 4
+    this->conv_param = {1, 1, 1, 1, 4, {3}, {3}, {1}, {1}, {3}, {3}};
+    is_supported     = this->Run();
+    EXPECT_TRUE(is_supported);
+}
+
+TEST_F(TestImageToColumnInterface4ScalarPerVector, X4ScalarPerVector)
+{
+    // vector load C * X % ScalarPerVector
+    this->conv_param  = {1, 1, 1, 1, 1, {3}, {3}, {1}, {1}, {0}, {0}};
+    bool is_supported = this->Run();
+    EXPECT_FALSE(is_supported);
+    // vector load C * left_pad_x % ScalarPerVector
+    this->conv_param = {1, 1, 1, 1, 1, {4}, {3}, {1}, {1}, {3}, {0}};
+    is_supported     = this->Run();
+    EXPECT_FALSE(is_supported);
+    // vector load C * right_pad_x % ScalarPerVector
+    this->conv_param = {1, 1, 1, 1, 1, {4}, {3}, {1}, {1}, {0}, {3}};
+    is_supported     = this->Run();
+    EXPECT_FALSE(is_supported);
+
+    // vector load C % ScalarPerVector, right_pad and stride
+    this->conv_param = {1, 1, 1, 1, 1, {4}, {3}, {2}, {1}, {0}, {3}};
+    is_supported     = this->Run();
+    EXPECT_FALSE(is_supported);
+    // vector load C % ScalarPerVector, left_pad and stride
+    this->conv_param = {1, 1, 1, 1, 1, {4}, {3}, {2}, {1}, {3}, {0}};
+    is_supported     = this->Run();
+    EXPECT_FALSE(is_supported);
+    // vector load C % ScalarPerVector, dilation
+    this->conv_param = {1, 1, 1, 1, 1, {4}, {3}, {1}, {2}, {0}, {0}};
+    is_supported     = this->Run();
+    EXPECT_FALSE(is_supported);
+
+    // C = 4
+    this->conv_param = {1, 1, 1, 1, 4, {3}, {3}, {1}, {1}, {3}, {3}};
+    is_supported     = this->Run();
+    EXPECT_TRUE(is_supported);
+}
+
+TEST_F(TestImageToColumnInterface4ScalarPerVectorFakeC, X4ScalarPerVectorFakeC)
+{
+    // C = 3
+    this->conv_param  = {1, 1, 1, 1, 3, {4}, {3}, {1}, {1}, {0}, {0}};
+    bool is_supported = this->Run();
+    EXPECT_FALSE(is_supported);
+    // C = 4
+    this->conv_param = {1, 1, 1, 1, 8, {4}, {3}, {1}, {1}, {0}, {0}};
+    is_supported     = this->Run();
+    EXPECT_TRUE(is_supported);
+}
diff --git a/test/magic_number_division/CMakeLists.txt b/test/magic_number_division/CMakeLists.txt
new file mode 100644
index 0000000000..e7fc6ee5df
--- /dev/null
+++ b/test/magic_number_division/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_test_executable(test_magic_number_division magic_number_division.cpp)
+target_link_libraries(test_magic_number_division PRIVATE utility)
diff --git a/test/magic_number_division/magic_number_division.cpp b/test/magic_number_division/magic_number_division.cpp
new file mode 100644
index 0000000000..ba863c2cf7
--- /dev/null
+++ b/test/magic_number_division/magic_number_division.cpp
@@ -0,0 +1,122 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/utility/magic_division.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+__global__ void gpu_magic_number_division(uint32_t magic_multiplier,
+                                          uint32_t magic_shift,
+                                          const int32_t* p_dividend,
+                                          int32_t* p_result,
+                                          uint64_t num)
+{
+    uint64_t global_thread_num = blockDim.x * gridDim.x;
+
+    uint64_t global_thread_id = blockIdx.x * blockDim.x + threadIdx.x;
+
+    for(uint64_t data_id = global_thread_id; data_id < num; data_id += global_thread_num)
+    {
+        p_result[data_id] =
+            ck::MagicDivision::DoMagicDivision(p_dividend[data_id], magic_multiplier, magic_shift);
+    }
+}
+
+__global__ void
+gpu_naive_division(int32_t divisor, const int32_t* p_dividend, int32_t* p_result, uint64_t num)
+{
+    uint64_t global_thread_num = blockDim.x * gridDim.x;
+
+    uint64_t global_thread_id = blockIdx.x * blockDim.x + threadIdx.x;
+
+    for(uint64_t data_id = global_thread_id; data_id < num; data_id += global_thread_num)
+    {
+        p_result[data_id] = p_dividend[data_id] / divisor;
+    }
+}
+
+int main(int, char*[])
+{
+    uint64_t num_divisor  = 1UL << 12;
+    uint64_t num_dividend = 1UL << 20;
+
+    std::vector<int32_t> divisors_host(num_divisor);
+    std::vector<int32_t> dividends_host(num_dividend);
+
+    // generate divisor
+    for(uint64_t i = 0; i < num_divisor; ++i)
+    {
+        divisors_host[i] = i + 1;
+    }
+
+    // generate dividend
+    for(uint64_t i = 0; i < num_dividend; ++i)
+    {
+        dividends_host[i] = i;
+    }
+
+    DeviceMem dividends_dev_buf(sizeof(int32_t) * num_dividend);
+    DeviceMem naive_result_dev_buf(sizeof(int32_t) * num_dividend);
+    DeviceMem magic_result_dev_buf(sizeof(int32_t) * num_dividend);
+
+    std::vector<int32_t> naive_result_host(num_dividend);
+    std::vector<int32_t> magic_result_host(num_dividend);
+
+    dividends_dev_buf.ToDevice(dividends_host.data());
+
+    bool pass = true;
+
+    for(std::size_t i = 0; i < num_divisor; ++i)
+    {
+        // run naive division on GPU
+        gpu_naive_division<<<1024, 256>>>(
+            divisors_host[i],
+            static_cast<const int32_t*>(dividends_dev_buf.GetDeviceBuffer()),
+            static_cast<int32_t*>(naive_result_dev_buf.GetDeviceBuffer()),
+            num_dividend);
+
+        // calculate magic number
+        uint32_t magic_multiplier, magic_shift;
+
+        ck::tie(magic_multiplier, magic_shift) =
+            ck::MagicDivision::CalculateMagicNumbers(divisors_host[i]);
+
+        // run magic division on GPU
+        gpu_magic_number_division<<<1024, 256>>>(
+            magic_multiplier,
+            magic_shift,
+            static_cast<const int32_t*>(dividends_dev_buf.GetDeviceBuffer()),
+            static_cast<int32_t*>(magic_result_dev_buf.GetDeviceBuffer()),
+            num_dividend);
+
+        naive_result_dev_buf.FromDevice(naive_result_host.data());
+        magic_result_dev_buf.FromDevice(magic_result_host.data());
+
+        bool res = ck::utils::check_err(magic_result_host, naive_result_host);
+
+        if(!res)
+        {
+            pass = false;
+            continue;
+        }
+    }
+
+    if(pass)
+    {
+        std::cout << "test magic number division: Pass" << std::endl;
+        return 0;
+    }
+    else
+    {
+        std::cout << "test magic number division: Fail" << std::endl;
+        return -1;
+    }
+}
diff --git a/test/normalization/CMakeLists.txt b/test/normalization/CMakeLists.txt
new file mode 100644
index 0000000000..2beda4dd74
--- /dev/null
+++ b/test/normalization/CMakeLists.txt
@@ -0,0 +1,19 @@
+if(DTYPES MATCHES "fp16" OR DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+  add_custom_target(test_normalization)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+  add_gtest_executable(test_layernorm2d_fp32 test_layernorm2d_fp32.cpp)
+  add_gtest_executable(test_groupnorm_fp32 test_groupnorm_fp32.cpp)
+  target_link_libraries(test_layernorm2d_fp32 PRIVATE utility device_normalization_instance)
+  target_link_libraries(test_groupnorm_fp32 PRIVATE utility device_normalization_instance)
+  add_dependencies(test_normalization test_layernorm2d_fp32)
+  add_dependencies(test_normalization test_groupnorm_fp32)
+endif()
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  add_gtest_executable(test_layernorm2d_fp16 test_layernorm2d_fp16.cpp)
+  add_gtest_executable(test_groupnorm_fp16 test_groupnorm_fp16.cpp)
+  target_link_libraries(test_layernorm2d_fp16 PRIVATE utility device_normalization_instance)
+  target_link_libraries(test_groupnorm_fp16 PRIVATE utility device_normalization_instance)
+  add_dependencies(test_normalization test_layernorm2d_fp16)
+  add_dependencies(test_normalization test_groupnorm_fp16)
+endif()
diff --git a/test/normalization/test_groupnorm_fp16.cpp b/test/normalization/test_groupnorm_fp16.cpp
new file mode 100644
index 0000000000..325ea75fe5
--- /dev/null
+++ b/test/normalization/test_groupnorm_fp16.cpp
@@ -0,0 +1,51 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_groupnorm_impl.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestGroupnorm : public ::testing::Test
+{
+    protected:
+    using XDataType       = std::tuple_element_t<0, Tuple>;
+    using GammaDataType   = std::tuple_element_t<1, Tuple>;
+    using BetaDataType    = std::tuple_element_t<2, Tuple>;
+    using ComputeDataType = std::tuple_element_t<3, Tuple>;
+    using YDataType       = std::tuple_element_t<4, Tuple>;
+
+    void Run()
+    {
+        // [N, H, W, G, C], reduce H, W, C
+        std::vector<std::vector<ck::index_t>> lengths = {{1, 1, 1, 1, 1},
+                                                         {1, 2, 3, 4, 5},
+                                                         {256, 9, 9, 9, 9},
+                                                         {1, 64, 64, 32, 10},
+                                                         {1, 32, 32, 32, 20},
+                                                         {2, 32, 32, 32, 30},
+                                                         {2, 32, 32, 32, 40},
+                                                         {1, 16, 16, 32, 40}};
+
+        for(auto length : lengths)
+        {
+            bool success =
+                ck::profiler::profile_groupnorm_impl<XDataType,
+                                                     GammaDataType,
+                                                     BetaDataType,
+                                                     ComputeDataType,
+                                                     YDataType>(true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<
+    // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType>
+    std::tuple<F16, F16, F16, F32, F16>>;
+
+TYPED_TEST_SUITE(TestGroupnorm, KernelTypes);
+TYPED_TEST(TestGroupnorm, Test_FP16) { this->Run(); }
diff --git a/test/normalization/test_groupnorm_fp32.cpp b/test/normalization/test_groupnorm_fp32.cpp
new file mode 100644
index 0000000000..ec88442fc0
--- /dev/null
+++ b/test/normalization/test_groupnorm_fp32.cpp
@@ -0,0 +1,49 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_groupnorm_impl.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestGroupnorm : public ::testing::Test
+{
+    protected:
+    using XDataType       = std::tuple_element_t<0, Tuple>;
+    using GammaDataType   = std::tuple_element_t<1, Tuple>;
+    using BetaDataType    = std::tuple_element_t<2, Tuple>;
+    using ComputeDataType = std::tuple_element_t<3, Tuple>;
+    using YDataType       = std::tuple_element_t<4, Tuple>;
+
+    void Run()
+    {
+        // [N, H, W, G, C], reduce H, W, C
+        std::vector<std::vector<ck::index_t>> lengths = {{1, 1, 1, 1, 1},
+                                                         {1, 2, 3, 4, 5},
+                                                         {256, 9, 9, 9, 9},
+                                                         {1, 64, 64, 32, 10},
+                                                         {1, 32, 32, 32, 20},
+                                                         {1, 16, 16, 32, 40}};
+
+        for(auto length : lengths)
+        {
+            bool success =
+                ck::profiler::profile_groupnorm_impl<XDataType,
+                                                     GammaDataType,
+                                                     BetaDataType,
+                                                     ComputeDataType,
+                                                     YDataType>(true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<
+    // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType>
+    std::tuple<F32, F32, F32, F32, F32>>;
+
+TYPED_TEST_SUITE(TestGroupnorm, KernelTypes);
+TYPED_TEST(TestGroupnorm, Test_FP32) { this->Run(); }
diff --git a/test/normalization/test_layernorm2d_fp16.cpp b/test/normalization/test_layernorm2d_fp16.cpp
new file mode 100644
index 0000000000..2222740fcc
--- /dev/null
+++ b/test/normalization/test_layernorm2d_fp16.cpp
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_layernorm_impl.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestLayernorm2d : public ::testing::Test
+{
+    protected:
+    using XDataType       = std::tuple_element_t<0, Tuple>;
+    using GammaDataType   = std::tuple_element_t<1, Tuple>;
+    using BetaDataType    = std::tuple_element_t<2, Tuple>;
+    using ComputeDataType = std::tuple_element_t<3, Tuple>;
+    using YDataType       = std::tuple_element_t<4, Tuple>;
+
+    void Run()
+    {
+        // [N, D], reduce D
+        std::vector<std::vector<ck::index_t>> lengths = {
+            {4, 256}, {8, 511}, {9, 1032}, {4, 2048}, {1, 8192}, {4000, 2000}};
+
+        for(auto length : lengths)
+        {
+            bool success = ck::profiler::profile_layernorm_impl<XDataType,
+                                                                GammaDataType,
+                                                                BetaDataType,
+                                                                ComputeDataType,
+                                                                YDataType,
+                                                                2>(true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<
+    // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType>
+    std::tuple<F16, F16, F16, F32, F16>>;
+
+TYPED_TEST_SUITE(TestLayernorm2d, KernelTypes);
+TYPED_TEST(TestLayernorm2d, Test_FP16) { this->Run(); }
diff --git a/test/normalization/test_layernorm2d_fp32.cpp b/test/normalization/test_layernorm2d_fp32.cpp
new file mode 100644
index 0000000000..30fbe06c60
--- /dev/null
+++ b/test/normalization/test_layernorm2d_fp32.cpp
@@ -0,0 +1,45 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_layernorm_impl.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestLayernorm2d : public ::testing::Test
+{
+    protected:
+    using XDataType       = std::tuple_element_t<0, Tuple>;
+    using GammaDataType   = std::tuple_element_t<1, Tuple>;
+    using BetaDataType    = std::tuple_element_t<2, Tuple>;
+    using ComputeDataType = std::tuple_element_t<3, Tuple>;
+    using YDataType       = std::tuple_element_t<4, Tuple>;
+
+    void Run()
+    {
+        // [N, D], reduce D
+        std::vector<std::vector<ck::index_t>> lengths = {
+            {4, 256}, {8, 511}, {9, 1032}, {4, 2048}, {1, 8192}, {4000, 2000}};
+
+        for(auto length : lengths)
+        {
+            bool success = ck::profiler::profile_layernorm_impl<XDataType,
+                                                                GammaDataType,
+                                                                BetaDataType,
+                                                                ComputeDataType,
+                                                                YDataType,
+                                                                2>(true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<
+    // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType>
+    std::tuple<F32, F32, F32, F32, F32>>;
+
+TYPED_TEST_SUITE(TestLayernorm2d, KernelTypes);
+TYPED_TEST(TestLayernorm2d, Test_FP32) { this->Run(); }
diff --git a/test/pool/CMakeLists.txt b/test/pool/CMakeLists.txt
new file mode 100644
index 0000000000..fac806897a
--- /dev/null
+++ b/test/pool/CMakeLists.txt
@@ -0,0 +1,16 @@
+add_custom_target(test_pool)
+
+add_gtest_executable(test_avg_pool3d_bwd test_avg_pool3d_bwd.cpp)
+add_gtest_executable(test_max_pool3d_bwd test_max_pool3d_bwd.cpp)
+add_gtest_executable(test_avg_pool3d_fwd test_avg_pool3d_fwd.cpp)
+add_gtest_executable(test_max_pool3d_fwd test_max_pool3d_fwd.cpp)
+
+target_link_libraries(test_avg_pool3d_bwd PRIVATE utility device_avg_pool3d_bwd_instance)
+target_link_libraries(test_max_pool3d_bwd PRIVATE utility device_max_pool_bwd_instance)
+target_link_libraries(test_avg_pool3d_fwd PRIVATE utility device_pool3d_fwd_instance)
+target_link_libraries(test_max_pool3d_fwd PRIVATE utility device_pool3d_fwd_instance)
+
+add_dependencies(test_pool test_avg_pool3d_bwd)
+add_dependencies(test_pool test_max_pool3d_bwd)
+add_dependencies(test_pool test_avg_pool3d_fwd)
+add_dependencies(test_pool test_max_pool3d_fwd)
diff --git a/test/pool/test_avg_pool3d_bwd.cpp b/test/pool/test_avg_pool3d_bwd.cpp
new file mode 100644
index 0000000000..fbd03fdf45
--- /dev/null
+++ b/test/pool/test_avg_pool3d_bwd.cpp
@@ -0,0 +1,74 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_avg_pool3d_bwd_impl.hpp"
+#include "test_pool_fwd_common.hpp"
+
+template <typename Tuple>
+class TestAvgPool3dBwd : public ::testing::Test
+{
+    protected:
+    using DOutDataType    = std::tuple_element_t<0, Tuple>;
+    using DInDataType     = std::tuple_element_t<1, Tuple>;
+    using ComputeDataType = std::tuple_element_t<2, Tuple>;
+    using DOutLayout      = std::tuple_element_t<3, Tuple>;
+    using DInLayout       = std::tuple_element_t<4, Tuple>;
+
+    std::vector<PoolingParam> params;
+
+    void Run()
+    {
+        for(auto param : params)
+        {
+            bool success =
+                ck::profiler::profile_avg_pool3d_bwd_impl<DOutDataType,
+                                                          DInDataType,
+                                                          ComputeDataType,
+                                                          DOutLayout,
+                                                          DInLayout>(true,
+                                                                     2,
+                                                                     false,
+                                                                     false,
+                                                                     param.length_,
+                                                                     param.window_spatial_lengths_,
+                                                                     param.window_strides_,
+                                                                     param.window_dilations_,
+                                                                     param.input_left_pads_,
+                                                                     param.input_right_pads_);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+#if defined(CK_ENABLE_FP16) && defined(CK_ENABLE_BF16) && defined(CK_ENABLE_FP32)
+using KernelTypes = ::testing::Types<std::tuple<F16, F16, F32, NDHWC, NDHWC>,
+                                     std::tuple<BF16, BF16, F32, NDHWC, NDHWC>,
+                                     std::tuple<F32, F32, F32, NDHWC, NDHWC>>;
+#elif defined(CK_ENABLE_FP16) && defined(CK_ENABLE_FP32)
+using KernelTypes = ::testing::Types<std::tuple<F16, F16, F32, NDHWC, NDHWC>,
+                                     std::tuple<F32, F32, F32, NDHWC, NDHWC>>;
+#elif defined(CK_ENABLE_BF16) && defined(CK_ENABLE_FP32)
+using KernelTypes = ::testing::Types<std::tuple<BF16, BF16, F32, NDHWC, NDHWC>,
+                                     std::tuple<F32, F32, F32, NDHWC, NDHWC>>;
+#elif defined(CK_ENABLE_FP16) && defined(CK_ENABLE_BF16)
+using KernelTypes = ::testing::Types<std::tuple<F16, F16, F32, NDHWC, NDHWC>,
+                                     std::tuple<BF16, BF16, F32, NDHWC, NDHWC>>;
+#elif defined(CK_ENABLE_FP16)
+using KernelTypes = ::testing::Types<std::tuple<F16, F16, F32, NDHWC, NDHWC>>;
+#elif defined(CK_ENABLE_BF16)
+using KernelTypes = ::testing::Types<std::tuple<BF16, BF16, F32, NDHWC, NDHWC>>;
+#elif defined(CK_ENABLE_FP32)
+using KernelTypes = ::testing::Types<std::tuple<F32, F32, F32, NDHWC, NDHWC>>;
+#endif
+
+TYPED_TEST_SUITE(TestAvgPool3dBwd, KernelTypes);
+TYPED_TEST(TestAvgPool3dBwd, Test_Pool)
+{
+    // length, window_length, window_stride, window_dilation, left_pad, right_pad
+    this->params = {{{1, 1, 1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}},
+                    {{2, 16, 64, 64, 64}, {4, 4, 4}, {4, 4, 4}, {2, 2, 2}, {0, 0, 0}, {0, 0, 0}},
+                    {{2, 32, 30, 30, 30}, {2, 2, 2}, {2, 2, 2}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}}};
+
+    this->Run();
+}
diff --git a/test/pool/test_avg_pool3d_fwd.cpp b/test/pool/test_avg_pool3d_fwd.cpp
new file mode 100644
index 0000000000..fc196a8a07
--- /dev/null
+++ b/test/pool/test_avg_pool3d_fwd.cpp
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_pool3d_fwd_impl.hpp"
+#include "test_pool_fwd_common.hpp"
+
+template <typename Tuple>
+class TestAvgPool3dFwd : public ::testing::Test
+{
+    protected:
+    using InDataType      = std::tuple_element_t<0, Tuple>;
+    using OutDataType     = std::tuple_element_t<1, Tuple>;
+    using ComputeDataType = std::tuple_element_t<2, Tuple>;
+    using IndexDataType   = std::tuple_element_t<3, Tuple>;
+
+    std::vector<PoolingParam> params;
+
+    void Run()
+    {
+        for(auto param : params)
+        {
+            bool success =
+                ck::profiler::profile_pool3d_fwd_impl<InDataType,
+                                                      OutDataType,
+                                                      ComputeDataType,
+                                                      IndexDataType,
+                                                      ck::tensor_layout::convolution::NDHWC,
+                                                      ck::tensor_layout::convolution::NDHWC,
+                                                      ck::ReduceTensorOp::AVG,
+                                                      false,
+                                                      false>(true,
+                                                             2,
+                                                             false,
+                                                             false,
+                                                             param.length_,
+                                                             param.window_spatial_lengths_,
+                                                             param.window_strides_,
+                                                             param.window_dilations_,
+                                                             param.input_left_pads_,
+                                                             param.input_right_pads_);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+#ifdef CK_ENABLE_FP16
+using KernelTypes =
+    ::testing::Types<std::tuple<F16, F16, F32, I32>, std::tuple<F32, F32, F32, I32>>;
+#else
+using KernelTypes = ::testing::Types<std::tuple<F32, F32, F32, I32>>;
+#endif
+TYPED_TEST_SUITE(TestAvgPool3dFwd, KernelTypes);
+TYPED_TEST(TestAvgPool3dFwd, Test_Pool)
+{
+    // length, window_length, window_stride, window_dilation, left_pad, right_pad
+    this->params = {{{1, 1, 1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}},
+                    {{2, 16, 64, 64, 64}, {64, 64, 64}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}},
+                    {{2, 16, 64, 64, 64}, {4, 4, 4}, {4, 4, 4}, {2, 2, 2}, {0, 0, 0}, {0, 0, 0}},
+                    {{2, 32, 30, 30, 30}, {2, 2, 2}, {2, 2, 2}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}}};
+
+    this->Run();
+}
diff --git a/test/pool/test_max_pool3d_bwd.cpp b/test/pool/test_max_pool3d_bwd.cpp
new file mode 100644
index 0000000000..8d52bde4da
--- /dev/null
+++ b/test/pool/test_max_pool3d_bwd.cpp
@@ -0,0 +1,79 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_max_pool3d_bwd_impl.hpp"
+#include "test_pool_fwd_common.hpp"
+
+template <typename Tuple>
+class TestMaxPool3dBwd : public ::testing::Test
+{
+    protected:
+    using DOutDataType  = std::tuple_element_t<0, Tuple>;
+    using DInDataType   = std::tuple_element_t<1, Tuple>;
+    using IndexDataType = std::tuple_element_t<2, Tuple>;
+
+    using InDataType  = DInDataType;
+    using OutDataType = DOutDataType;
+
+    std::vector<PoolingParam> params;
+
+    void Run()
+    {
+        for(auto param : params)
+        {
+            bool success =
+                ck::profiler::profile_max_pool3d_bwd_impl<InDataType,
+                                                          OutDataType,
+                                                          IndexDataType,
+                                                          DOutDataType,
+                                                          DInDataType,
+                                                          false>(true,
+                                                                 2,
+                                                                 false,
+                                                                 false,
+                                                                 param.length_,
+                                                                 param.window_spatial_lengths_,
+                                                                 param.window_strides_,
+                                                                 param.window_dilations_,
+                                                                 param.input_left_pads_,
+                                                                 param.input_right_pads_);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+#if defined(CK_ENABLE_FP16) && defined(CK_ENABLE_BF16) && defined(CK_ENABLE_FP32)
+using KernelTypes = ::testing::Types<std::tuple<F16, F16, I32, NDHWC, NDHWC>,
+                                     std::tuple<BF16, BF16, I32, NDHWC, NDHWC>,
+                                     std::tuple<F32, F32, I32, NDHWC, NDHWC>>;
+#elif defined(CK_ENABLE_FP16) && defined(CK_ENABLE_FP32)
+using KernelTypes = ::testing::Types<std::tuple<F16, F16, I32, NDHWC, NDHWC>,
+                                     std::tuple<F32, F32, I32, NDHWC, NDHWC>>;
+#elif defined(CK_ENABLE_BF16) && defined(CK_ENABLE_FP32)
+using KernelTypes = ::testing::Types<std::tuple<BF16, BF16, I32, NDHWC, NDHWC>,
+                                     std::tuple<F32, F32, I32, NDHWC, NDHWC>>;
+#elif defined(CK_ENABLE_FP16) && defined(CK_ENABLE_BF16)
+using KernelTypes = ::testing::Types<std::tuple<F16, F16, I32, NDHWC, NDHWC>,
+                                     std::tuple<BF16, BF16, I32, NDHWC, NDHWC>>;
+#elif defined(CK_ENABLE_FP16)
+using KernelTypes = ::testing::Types<std::tuple<F16, F16, I32, NDHWC, NDHWC>>;
+#elif defined(CK_ENABLE_BF16)
+using KernelTypes = ::testing::Types<std::tuple<BF16, BF16, I32, NDHWC, NDHWC>>;
+#elif defined(CK_ENABLE_FP32)
+using KernelTypes = ::testing::Types<std::tuple<F32, F32, I32, NDHWC, NDHWC>>;
+#endif
+
+TYPED_TEST_SUITE(TestMaxPool3dBwd, KernelTypes);
+TYPED_TEST(TestMaxPool3dBwd, Test_Pool)
+{
+    // length, window_length, window_stride, window_dilation, left_pad, right_pad
+    this->params = {{{1, 1, 1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}},
+                    {{2, 16, 64, 64, 64}, {4, 4, 4}, {4, 4, 4}, {2, 2, 2}, {0, 0, 0}, {0, 0, 0}},
+                    {{2, 32, 30, 30, 30}, {2, 2, 2}, {2, 2, 2}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}}};
+
+    // this->params = {{{2, 32, 30, 30, 30}, {2, 2, 2}, {2, 2, 2}, {1, 1, 1}, {1, 1, 1}, {1, 1,
+    // 1}}};
+
+    this->Run();
+}
diff --git a/test/pool/test_max_pool3d_fwd.cpp b/test/pool/test_max_pool3d_fwd.cpp
new file mode 100644
index 0000000000..7189f1b104
--- /dev/null
+++ b/test/pool/test_max_pool3d_fwd.cpp
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "profiler/profile_pool3d_fwd_impl.hpp"
+#include "test_pool_fwd_common.hpp"
+
+template <typename Tuple>
+class TestMaxPool3dFwd : public ::testing::Test
+{
+    protected:
+    using InDataType      = std::tuple_element_t<0, Tuple>;
+    using OutDataType     = std::tuple_element_t<1, Tuple>;
+    using ComputeDataType = std::tuple_element_t<2, Tuple>;
+    using IndexDataType   = std::tuple_element_t<3, Tuple>;
+
+    std::vector<PoolingParam> params;
+
+    void Run()
+    {
+        for(auto param : params)
+        {
+            // max pool
+            bool success =
+                ck::profiler::profile_pool3d_fwd_impl<InDataType,
+                                                      OutDataType,
+                                                      ComputeDataType,
+                                                      IndexDataType,
+                                                      ck::tensor_layout::convolution::NDHWC,
+                                                      ck::tensor_layout::convolution::NDHWC,
+                                                      ck::ReduceTensorOp::MAX,
+                                                      false,
+                                                      false>(true,
+                                                             2,
+                                                             false,
+                                                             false,
+                                                             param.length_,
+                                                             param.window_spatial_lengths_,
+                                                             param.window_strides_,
+                                                             param.window_dilations_,
+                                                             param.input_left_pads_,
+                                                             param.input_right_pads_);
+            EXPECT_TRUE(success);
+
+            // max pool + index
+            success = ck::profiler::profile_pool3d_fwd_impl<InDataType,
+                                                            OutDataType,
+                                                            ComputeDataType,
+                                                            IndexDataType,
+                                                            ck::tensor_layout::convolution::NDHWC,
+                                                            ck::tensor_layout::convolution::NDHWC,
+                                                            ck::ReduceTensorOp::MAX,
+                                                            false,
+                                                            true>(true,
+                                                                  2,
+                                                                  false,
+                                                                  false,
+                                                                  param.length_,
+                                                                  param.window_spatial_lengths_,
+                                                                  param.window_strides_,
+                                                                  param.window_dilations_,
+                                                                  param.input_left_pads_,
+                                                                  param.input_right_pads_);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+#ifdef CK_ENABLE_FP16
+using KernelTypes =
+    ::testing::Types<std::tuple<F16, F16, F32, I32>, std::tuple<F32, F32, F32, I32>>;
+#else
+using KernelTypes = ::testing::Types<std::tuple<F32, F32, F32, I32>>;
+#endif
+
+TYPED_TEST_SUITE(TestMaxPool3dFwd, KernelTypes);
+TYPED_TEST(TestMaxPool3dFwd, Test_Pool)
+{
+    // length, window_length, window_stride, window_dilation, left_pad, right_pad
+    this->params = {{{1, 1, 1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}},
+                    {{2, 16, 64, 64, 64}, {64, 64, 64}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}},
+                    {{2, 16, 64, 64, 64}, {4, 4, 4}, {4, 4, 4}, {2, 2, 2}, {0, 0, 0}, {0, 0, 0}},
+                    {{2, 32, 30, 30, 30}, {2, 2, 2}, {2, 2, 2}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}}};
+
+    this->Run();
+}
diff --git a/test/pool/test_pool_fwd_common.hpp b/test/pool/test_pool_fwd_common.hpp
new file mode 100644
index 0000000000..5917a27e56
--- /dev/null
+++ b/test/pool/test_pool_fwd_common.hpp
@@ -0,0 +1,36 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "ck/ck.hpp"
+
+using F16  = ck::half_t;
+using BF16 = ck::bhalf_t;
+using F32  = float;
+using I32  = int32_t;
+using ck::index_t;
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
+
+struct PoolingParam
+{
+    PoolingParam(const std::vector<index_t>& length,
+                 const std::vector<index_t>& window_spatial_lengths,
+                 const std::vector<index_t>& window_strides,
+                 const std::vector<index_t>& window_dilations,
+                 const std::vector<index_t>& input_left_pads,
+                 const std::vector<index_t>& input_right_pads)
+        : length_(length),
+          window_spatial_lengths_(window_spatial_lengths),
+          window_strides_(window_strides),
+          window_dilations_(window_dilations),
+          input_left_pads_(input_left_pads),
+          input_right_pads_(input_right_pads)
+    {
+    }
+    std::vector<index_t> length_;
+    std::vector<index_t> window_spatial_lengths_;
+    std::vector<index_t> window_strides_;
+    std::vector<index_t> window_dilations_;
+    std::vector<index_t> input_left_pads_;
+    std::vector<index_t> input_right_pads_;
+};
diff --git a/test/reduce/CMakeLists.txt b/test/reduce/CMakeLists.txt
new file mode 100644
index 0000000000..fb436165ea
--- /dev/null
+++ b/test/reduce/CMakeLists.txt
@@ -0,0 +1,7 @@
+add_test_executable(test_reduce_no_index reduce_no_index.cpp)
+add_test_executable(test_reduce_with_index reduce_with_index.cpp)
+target_link_libraries(test_reduce_no_index PRIVATE utility)
+target_link_libraries(test_reduce_no_index PRIVATE device_reduce_instance)
+target_link_libraries(test_reduce_with_index PRIVATE utility)
+target_link_libraries(test_reduce_with_index PRIVATE device_reduce_instance)
+
diff --git a/test/reduce/reduce_no_index.cpp b/test/reduce/reduce_no_index.cpp
new file mode 100644
index 0000000000..1ab452442d
--- /dev/null
+++ b/test/reduce/reduce_no_index.cpp
@@ -0,0 +1,248 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <getopt.h>
+
+#include "ck/library/utility/host_common_util.hpp"
+#include "profiler/profile_reduce_impl.hpp"
+
+using namespace ck;
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"reduceDimensions", required_argument, nullptr, 'R'},
+                                       {"scales", required_argument, nullptr, 'S'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths;
+    std::vector<int> reduceDims;
+    std::vector<float> scales;
+
+    int data_type;
+    int init_method = 1;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths "
+                     "(only 4-d tensor supported)"
+                  << std::endl;
+        std::cout << "--reduceDimensions or -R comma seperated list of dimension indexes to reduce "
+                     "(only 1 or 3 or 4 dimensions supported)"
+                  << std::endl;
+        std::cout << "--scales or -S, comma separated two float values for alpha and beta"
+                  << std::endl;
+        std::cout << "Arg1 -- data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2 -- init method(0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:S:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'S':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                scales = getTypeValuesFromString<float>(optarg);
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 2 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type   = std::atoi(argv[optind++]);
+        init_method = std::atoi(argv[optind]);
+
+        if(scales.empty())
+        {
+            scales.push_back(1.0f);
+            scales.push_back(0.0f);
+        };
+
+        if(inLengths.size() != 4 ||
+           (reduceDims.size() != 1 && reduceDims.size() != 3 && reduceDims.size() != 4))
+            return (-1);
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return (-1);
+
+        return (0);
+    };
+};
+
+bool test_reduce_no_index(int data_type,
+                          int init_method,
+                          std::vector<int> reduceDims,
+                          std::vector<size_t> inLengths,
+                          ReduceTensorOp reduceOpId,
+                          bool propagateNan,
+                          float alpha,
+                          float beta)
+{
+    using ck::profiler::profile_reduce_impl;
+
+    bool result = true;
+
+    if(data_type == 0)
+    {
+        result = profile_reduce_impl<float, float, float>(true,
+                                                          init_method,
+                                                          false,
+                                                          false,
+                                                          inLengths,
+                                                          reduceDims,
+                                                          reduceOpId,
+                                                          propagateNan,
+                                                          false,
+                                                          alpha,
+                                                          beta);
+    }
+    else if(data_type == 1)
+    {
+        result = profile_reduce_impl<ck::half_t, float, ck::half_t>(true,
+                                                                    init_method,
+                                                                    false,
+                                                                    false,
+                                                                    inLengths,
+                                                                    reduceDims,
+                                                                    reduceOpId,
+                                                                    propagateNan,
+                                                                    false,
+                                                                    alpha,
+                                                                    beta);
+    }
+    else if(data_type == 3)
+    {
+        result = profile_reduce_impl<int8_t, int32_t, int8_t>(true,
+                                                              init_method,
+                                                              false,
+                                                              false,
+                                                              inLengths,
+                                                              reduceDims,
+                                                              reduceOpId,
+                                                              propagateNan,
+                                                              false,
+                                                              alpha,
+                                                              beta);
+    }
+    else if(data_type == 5)
+    {
+        result = profile_reduce_impl<ck::bhalf_t, float, ck::bhalf_t>(true,
+                                                                      init_method,
+                                                                      false,
+                                                                      false,
+                                                                      inLengths,
+                                                                      reduceDims,
+                                                                      reduceOpId,
+                                                                      propagateNan,
+                                                                      false,
+                                                                      alpha,
+                                                                      beta);
+    }
+    else if(data_type == 6)
+    {
+        result = profile_reduce_impl<double, double, double>(true,
+                                                             init_method,
+                                                             false,
+                                                             false,
+                                                             inLengths,
+                                                             reduceDims,
+                                                             reduceOpId,
+                                                             propagateNan,
+                                                             false,
+                                                             alpha,
+                                                             beta);
+    }
+
+    return (result);
+};
+
+constexpr ReduceTensorOp reduceOpId = ReduceTensorOp::AVG;
+constexpr bool propagateNan         = false;
+
+int main(int argc, char* argv[])
+{
+    SimpleAppArgs args;
+
+    bool result = true;
+
+    if(argc == 1)
+    {
+        int data_type   = 1;
+        int init_method = 2;
+        std::vector<size_t> inLengths{64, 4, 280, 80};
+        std::vector<std::vector<int>> v_reduceDims{
+            {0, 1, 2, 3}, {0, 1, 2}, {1, 2, 3}, {0, 1, 3}, {0, 2, 3}, {0}, {1}, {2}, {3}};
+
+        for(auto& reduceDims : v_reduceDims)
+            result = result && test_reduce_no_index(data_type,
+                                                    init_method,
+                                                    reduceDims,
+                                                    inLengths,
+                                                    reduceOpId,
+                                                    propagateNan,
+                                                    1.0f,
+                                                    0.0f);
+    }
+    else
+    {
+        if(args.processArgs(argc, argv) < 0)
+        {
+            throw std::runtime_error(
+                "Invalid input arguments, test_reduce_no_index could not be executed!");
+        };
+
+        result = test_reduce_no_index(args.data_type,
+                                      args.init_method,
+                                      args.reduceDims,
+                                      args.inLengths,
+                                      reduceOpId,
+                                      propagateNan,
+                                      args.scales[0],
+                                      args.scales[1]);
+    }
+
+    std::cout << "test_reduce_no_index ..... " << (result ? "SUCCESS" : "FAILURE") << std::endl;
+
+    return (result ? 0 : -1);
+}
diff --git a/test/reduce/reduce_with_index.cpp b/test/reduce/reduce_with_index.cpp
new file mode 100644
index 0000000000..0301669c5a
--- /dev/null
+++ b/test/reduce/reduce_with_index.cpp
@@ -0,0 +1,248 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <getopt.h>
+
+#include "ck/library/utility/host_common_util.hpp"
+#include "profiler/profile_reduce_impl.hpp"
+
+using namespace ck;
+
+static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
+                                       {"reduceDimensions", required_argument, nullptr, 'R'},
+                                       {"scales", required_argument, nullptr, 'S'},
+                                       {"help", no_argument, nullptr, '?'},
+                                       {nullptr, 0, nullptr, 0}};
+
+class SimpleAppArgs
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths;
+    std::vector<int> reduceDims;
+    std::vector<float> scales;
+
+    int data_type;
+    int init_method = 1;
+
+    public:
+    void show_usage(const char* cmd)
+    {
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths "
+                     "(only 4-d tensor supported)"
+                  << std::endl;
+        std::cout << "--reduceDimensions or -R comma seperated list of dimension indexes to reduce "
+                     "(only 1 or 3 or 4 dimensions supported)"
+                  << std::endl;
+        std::cout << "--scales or -S, comma separated two float values for alpha and beta"
+                  << std::endl;
+        std::cout << "Arg1 -- data type (1: fp32, 3: int8, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2 -- init method(0=no init, 1=single integer value, 2=scope integer "
+                     "value, 3=decimal value)"
+                  << std::endl;
+    };
+
+    int processArgs(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:S:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'S':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                scales = getTypeValuesFromString<float>(optarg);
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return (-1);
+                };
+                break;
+            default: show_usage(argv[0]); return (-1);
+            };
+        };
+
+        if(optind + 2 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type   = std::atoi(argv[optind++]);
+        init_method = std::atoi(argv[optind]);
+
+        if(scales.empty())
+        {
+            scales.push_back(1.0f);
+            scales.push_back(0.0f);
+        };
+
+        if(inLengths.size() != 4 ||
+           (reduceDims.size() != 1 && reduceDims.size() != 3 && reduceDims.size() != 4))
+            return (-1);
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return (-1);
+
+        return (0);
+    };
+};
+
+bool test_reduce_with_index(int data_type,
+                            int init_method,
+                            std::vector<int> reduceDims,
+                            std::vector<size_t> inLengths,
+                            ReduceTensorOp reduceOpId,
+                            bool propagateNan,
+                            float alpha,
+                            float beta)
+{
+    using ck::profiler::profile_reduce_impl;
+
+    bool result = true;
+
+    if(data_type == 0)
+    {
+        result = profile_reduce_impl<float, float, float>(true,
+                                                          init_method,
+                                                          false,
+                                                          false,
+                                                          inLengths,
+                                                          reduceDims,
+                                                          reduceOpId,
+                                                          propagateNan,
+                                                          true,
+                                                          alpha,
+                                                          beta);
+    }
+    else if(data_type == 1)
+    {
+        result = profile_reduce_impl<ck::half_t, ck::half_t, ck::half_t>(true,
+                                                                         init_method,
+                                                                         false,
+                                                                         false,
+                                                                         inLengths,
+                                                                         reduceDims,
+                                                                         reduceOpId,
+                                                                         propagateNan,
+                                                                         true,
+                                                                         alpha,
+                                                                         beta);
+    }
+    else if(data_type == 3)
+    {
+        result = profile_reduce_impl<int8_t, int8_t, int8_t>(true,
+                                                             init_method,
+                                                             false,
+                                                             false,
+                                                             inLengths,
+                                                             reduceDims,
+                                                             reduceOpId,
+                                                             propagateNan,
+                                                             true,
+                                                             alpha,
+                                                             beta);
+    }
+    else if(data_type == 5)
+    {
+        result = profile_reduce_impl<ck::bhalf_t, float, ck::bhalf_t>(true,
+                                                                      init_method,
+                                                                      false,
+                                                                      false,
+                                                                      inLengths,
+                                                                      reduceDims,
+                                                                      reduceOpId,
+                                                                      propagateNan,
+                                                                      true,
+                                                                      alpha,
+                                                                      beta);
+    }
+    else if(data_type == 6)
+    {
+        result = profile_reduce_impl<double, double, double>(true,
+                                                             init_method,
+                                                             false,
+                                                             false,
+                                                             inLengths,
+                                                             reduceDims,
+                                                             reduceOpId,
+                                                             propagateNan,
+                                                             true,
+                                                             alpha,
+                                                             beta);
+    }
+
+    return (result);
+};
+
+constexpr ReduceTensorOp reduceOpId = ReduceTensorOp::AMAX;
+constexpr bool propagateNan         = false;
+
+int main(int argc, char* argv[])
+{
+    SimpleAppArgs args;
+
+    bool result = true;
+
+    if(argc == 1)
+    {
+        int data_type   = 1;
+        int init_method = 2;
+        std::vector<size_t> inLengths{64, 4, 280, 80};
+        std::vector<std::vector<int>> v_reduceDims{
+            {0, 1, 2, 3}, {0, 1, 2}, {1, 2, 3}, {0, 1, 3}, {0, 2, 3}, {0}, {1}, {2}, {3}};
+
+        for(auto& reduceDims : v_reduceDims)
+            result = result && test_reduce_with_index(data_type,
+                                                      init_method,
+                                                      reduceDims,
+                                                      inLengths,
+                                                      reduceOpId,
+                                                      propagateNan,
+                                                      1.0f,
+                                                      0.0f);
+    }
+    else
+    {
+        if(args.processArgs(argc, argv) < 0)
+        {
+            throw std::runtime_error(
+                "Invalid input arguments, test_reduce_with_index could not be executed!");
+        };
+
+        result = test_reduce_with_index(args.data_type,
+                                        args.init_method,
+                                        args.reduceDims,
+                                        args.inLengths,
+                                        reduceOpId,
+                                        propagateNan,
+                                        args.scales[0],
+                                        args.scales[1]);
+    }
+
+    std::cout << "test_reduce_with_index ..... " << (result ? "SUCCESS" : "FAILURE") << std::endl;
+
+    return (result ? 0 : -1);
+}
diff --git a/test/reference_conv_fwd/CMakeLists.txt b/test/reference_conv_fwd/CMakeLists.txt
new file mode 100644
index 0000000000..b40b9a1ed0
--- /dev/null
+++ b/test/reference_conv_fwd/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_gtest_executable(test_reference_conv_fwd reference_conv_fwd.cpp)
+target_link_libraries(test_reference_conv_fwd PRIVATE utility)
diff --git a/test/reference_conv_fwd/reference_conv_fwd.cpp b/test/reference_conv_fwd/reference_conv_fwd.cpp
new file mode 100644
index 0000000000..b3328e4b36
--- /dev/null
+++ b/test/reference_conv_fwd/reference_conv_fwd.cpp
@@ -0,0 +1,392 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cmath>
+#include <cstdlib>
+#include <numeric>
+#include <type_traits>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
+
+namespace {
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+template <ck::index_t NDimSpatial,
+          typename InDataType    = float,
+          typename WeiDataType   = float,
+          typename OutDataType   = float,
+          typename InLayout      = ck::tensor_layout::convolution::GNHWC,
+          typename WeiLayout     = ck::tensor_layout::convolution::GKYXC,
+          typename OutLayout     = ck::tensor_layout::convolution::GNHWK,
+          typename FillInputOp   = ck::utils::FillMonotonicSeq<InDataType>,
+          typename FillWeightsOp = ck::utils::FillConstant<WeiDataType>>
+Tensor<OutDataType>
+run_reference_convolution_forward(const ck::utils::conv::ConvParam& conv_param,
+                                  const FillInputOp& fill_input_op     = FillInputOp{},
+                                  const FillWeightsOp& fill_weights_op = FillWeightsOp{0.5f})
+{
+    const auto in_g_n_c_wis_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
+
+    const auto wei_g_k_c_xs_desc =
+        ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
+
+    const auto out_g_n_k_wos_desc =
+        ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
+
+    Tensor<InDataType> input(in_g_n_c_wis_desc);
+    Tensor<WeiDataType> weights(wei_g_k_c_xs_desc);
+    Tensor<OutDataType> host_output(out_g_n_k_wos_desc);
+
+    fill_input_op(input.begin(), input.end());
+    fill_weights_op(weights.begin(), weights.end());
+    ck::ranges::fill<OutDataType>(host_output, 0.f);
+
+    auto ref_conv     = ck::tensor_operation::host::ReferenceConvFwd<NDimSpatial,
+                                                                 InDataType,
+                                                                 WeiDataType,
+                                                                 OutDataType,
+                                                                 InElementOp,
+                                                                 WeiElementOp,
+                                                                 OutElementOp>();
+    auto ref_invoker  = ref_conv.MakeInvoker();
+    auto ref_argument = ref_conv.MakeArgument(input,
+                                              weights,
+                                              host_output,
+                                              conv_param.conv_filter_strides_,
+                                              conv_param.conv_filter_dilations_,
+                                              conv_param.input_left_pads_,
+                                              conv_param.input_right_pads_,
+                                              InElementOp{},
+                                              WeiElementOp{},
+                                              OutElementOp{});
+
+    ref_invoker.Run(ref_argument);
+    return host_output;
+}
+
+} // anonymous namespace
+
+// Eeference convolution assume dimensions of tensor descriptors are in GNCDHW/GKCZYX/GNKDHW order,
+// regardless of physical tensor layouts in  memory.
+// Some tests below assume dimensions of tensor descriptors can be in other order, and therefore
+// are disabled
+// TODO: add more tests, which comply with assumption about dimension order of reference convolution
+// and add tests for more physical layout
+#if 0
+TEST(ReferenceConvolutionFWD, Conv2DGNHWC)
+{
+    ck::utils::conv::ConvParam conv_param(2,
+                                          1,
+                                          1,
+                                          1,
+                                          2,
+                                          std::vector<ck::index_t>{3, 3},
+                                          std::vector<ck::index_t>{6, 6},
+                                          std::vector<ck::index_t>{1, 1},
+                                          std::vector<ck::index_t>{1, 1},
+                                          std::vector<ck::index_t>{0, 0},
+                                          std::vector<ck::index_t>{0, 0});
+
+    auto out_tensor = run_reference_convolution_forward<2>(conv_param);
+    std::vector<std::size_t> ref_dims{1, 1, 4, 4, 1};
+    std::vector<float> ref_data{130.5,
+                                148.5,
+                                166.5,
+                                184.5,
+                                238.5,
+                                256.5,
+                                274.5,
+                                292.5,
+                                346.5,
+                                364.5,
+                                382.5,
+                                400.5,
+                                454.5,
+                                472.5,
+                                490.5,
+                                508.5};
+    EXPECT_TRUE(ck::utils::check_err(
+        out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
+    EXPECT_TRUE(ck::utils::check_err(out_tensor, ref_data, "Error: incorrect results!"));
+}
+
+TEST(ReferenceConvolutionFWD, Conv2DGNHWCStridesDilationsPadding)
+{
+    ck::utils::conv::ConvParam conv_param(2,
+                                          1,
+                                          1,
+                                          2,
+                                          2,
+                                          std::vector<ck::index_t>{3, 3},
+                                          std::vector<ck::index_t>{12, 12},
+                                          std::vector<ck::index_t>{2, 2},
+                                          std::vector<ck::index_t>{2, 2},
+                                          std::vector<ck::index_t>{1, 1},
+                                          std::vector<ck::index_t>{1, 1});
+
+    auto out_tensor                   = run_reference_convolution_forward<2>(conv_param);
+    std::vector<std::size_t> ref_dims = std::vector<std::size_t>{1, 5, 5, 2};
+    std::vector<float> ref_data{
+        210.,  210.,  327.,   327.,   351.,   351.,   375.,   375.,   399.,   399.,
+        459.,  459.,  706.5,  706.5,  742.5,  742.5,  778.5,  778.5,  814.5,  814.5,
+        747.,  747.,  1138.5, 1138.5, 1174.5, 1174.5, 1210.5, 1210.5, 1246.5, 1246.5,
+        1035., 1035., 1570.5, 1570.5, 1606.5, 1606.5, 1642.5, 1642.5, 1678.5, 1678.5,
+        1323., 1323., 2002.5, 2002.5, 2038.5, 2038.5, 2074.5, 2074.5, 2110.5, 2110.5};
+    EXPECT_TRUE(ck::utils::check_err(
+        out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
+    EXPECT_TRUE(ck::utils::check_err(out_tensor, ref_data, "Error: incorrect results!"));
+}
+
+TEST(ReferenceConvolutionFWD, Conv1DGNWC)
+{
+    ck::utils::conv::ConvParam conv_param(1,
+                                          1,
+                                          1,
+                                          1,
+                                          2,
+                                          std::vector<ck::index_t>{3},
+                                          std::vector<ck::index_t>{6},
+                                          std::vector<ck::index_t>{1},
+                                          std::vector<ck::index_t>{1},
+                                          std::vector<ck::index_t>{0},
+                                          std::vector<ck::index_t>{0});
+
+    auto out_tensor =
+        run_reference_convolution_forward<1,
+                                          float,
+                                          float,
+                                          float,
+                                          ck::tensor_layout::convolution::GNWC,
+                                          ck::tensor_layout::convolution::GKXC,
+                                          ck::tensor_layout::convolution::GNWK>(conv_param);
+    std::vector<std::size_t> ref_dims{1, 1, 4, 1};
+    std::vector<float> ref_data{7.5, 13.5, 19.5, 25.5};
+    EXPECT_TRUE(ck::utils::check_err(
+        out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
+    EXPECT_TRUE(ck::utils::check_err(out_tensor, ref_data, "Error: incorrect results!"));
+}
+
+TEST(ReferenceConvolutionFWD, Conv1DGNWCStridesDilationsPadding)
+{
+    ck::utils::conv::ConvParam conv_param(1,
+                                          1,
+                                          1,
+                                          2,
+                                          2,
+                                          std::vector<ck::index_t>{3},
+                                          std::vector<ck::index_t>{12},
+                                          std::vector<ck::index_t>{2},
+                                          std::vector<ck::index_t>{2},
+                                          std::vector<ck::index_t>{1},
+                                          std::vector<ck::index_t>{1});
+
+    auto out_tensor =
+        run_reference_convolution_forward<1,
+                                          float,
+                                          float,
+                                          float,
+                                          ck::tensor_layout::convolution::GNWC,
+                                          ck::tensor_layout::convolution::GKXC,
+                                          ck::tensor_layout::convolution::GNWK>(conv_param);
+    std::vector<std::size_t> ref_dims{1, 1, 5, 2};
+    std::vector<float> ref_data{9., 9., 19.5, 19.5, 31.5, 31.5, 43.5, 43.5, 55.5, 55.5};
+    EXPECT_TRUE(ck::utils::check_err(
+        out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
+    EXPECT_TRUE(ck::utils::check_err(out_tensor, ref_data, "Error: incorrect results!"));
+}
+
+TEST(ReferenceConvolutionFWD, Conv1DGNWCSameOutputSize)
+{
+    ck::utils::conv::ConvParam conv_param(1,
+                                          1,
+                                          2,
+                                          16,
+                                          4,
+                                          std::vector<ck::index_t>{3},
+                                          std::vector<ck::index_t>{16},
+                                          std::vector<ck::index_t>{1},
+                                          std::vector<ck::index_t>{1},
+                                          std::vector<ck::index_t>{1},
+                                          std::vector<ck::index_t>{1});
+
+    auto out_tensor2 = run_reference_convolution_forward<1,
+                                                         float,
+                                                         float,
+                                                         float,
+                                                         ck::tensor_layout::convolution::GNWC,
+                                                         ck::tensor_layout::convolution::GKXC,
+                                                         ck::tensor_layout::convolution::GNWK>(
+        conv_param, ck::utils::FillMonotonicSeq<float>{0.f, 0.1f});
+
+    std::vector<std::size_t> ref_dims{1, 2, 16, 16};
+    std::vector<float> ref_data{
+        1.4,       1.4,       1.4,       1.4,       1.4,       1.4,       1.4,       1.4,
+        1.4,       1.4,       1.4,       1.4,       1.4,       1.4,       1.4,       1.4,
+        3.3,       3.3,       3.3,       3.3,       3.3,       3.3,       3.3,       3.3,
+        3.3,       3.3,       3.3,       3.3,       3.3,       3.3,       3.3,       3.3,
+        5.7,       5.7,       5.7,       5.7,       5.7,       5.7,       5.7,       5.7,
+        5.7,       5.7,       5.7,       5.7,       5.7,       5.7,       5.7,       5.7,
+        8.1,       8.1,       8.1,       8.1,       8.1,       8.1,       8.1,       8.1,
+        8.1,       8.1,       8.1,       8.1,       8.1,       8.1,       8.1,       8.1,
+        10.5,      10.5,      10.5,      10.5,      10.5,      10.5,      10.5,      10.5,
+        10.5,      10.5,      10.5,      10.5,      10.5,      10.5,      10.5,      10.5,
+        12.900001, 12.900001, 12.900001, 12.900001, 12.900001, 12.900001, 12.900001, 12.900001,
+        12.900001, 12.900001, 12.900001, 12.900001, 12.900001, 12.900001, 12.900001, 12.900001,
+        15.3,      15.3,      15.3,      15.3,      15.3,      15.3,      15.3,      15.3,
+        15.3,      15.3,      15.3,      15.3,      15.3,      15.3,      15.3,      15.3,
+        17.7,      17.7,      17.7,      17.7,      17.7,      17.7,      17.7,      17.7,
+        17.7,      17.7,      17.7,      17.7,      17.7,      17.7,      17.7,      17.7,
+        20.1,      20.1,      20.1,      20.1,      20.1,      20.1,      20.1,      20.1,
+        20.1,      20.1,      20.1,      20.1,      20.1,      20.1,      20.1,      20.1,
+        22.5,      22.5,      22.5,      22.5,      22.5,      22.5,      22.5,      22.5,
+        22.5,      22.5,      22.5,      22.5,      22.5,      22.5,      22.5,      22.5,
+        24.900002, 24.900002, 24.900002, 24.900002, 24.900002, 24.900002, 24.900002, 24.900002,
+        24.900002, 24.900002, 24.900002, 24.900002, 24.900002, 24.900002, 24.900002, 24.900002,
+        27.300001, 27.300001, 27.300001, 27.300001, 27.300001, 27.300001, 27.300001, 27.300001,
+        27.300001, 27.300001, 27.300001, 27.300001, 27.300001, 27.300001, 27.300001, 27.300001,
+        29.7,      29.7,      29.7,      29.7,      29.7,      29.7,      29.7,      29.7,
+        29.7,      29.7,      29.7,      29.7,      29.7,      29.7,      29.7,      29.7,
+        32.100002, 32.100002, 32.100002, 32.100002, 32.100002, 32.100002, 32.100002, 32.100002,
+        32.100002, 32.100002, 32.100002, 32.100002, 32.100002, 32.100002, 32.100002, 32.100002,
+        34.5,      34.5,      34.5,      34.5,      34.5,      34.5,      34.5,      34.5,
+        34.5,      34.5,      34.5,      34.5,      34.5,      34.5,      34.5,      34.5,
+        23.8,      23.8,      23.8,      23.8,      23.8,      23.8,      23.8,      23.8,
+        23.8,      23.8,      23.8,      23.8,      23.8,      23.8,      23.8,      23.8,
+        27.,       27.,       27.,       27.,       27.,       27.,       27.,       27.,
+        27.,       27.,       27.,       27.,       27.,       27.,       27.,       27.,
+        41.7,      41.7,      41.7,      41.7,      41.7,      41.7,      41.7,      41.7,
+        41.7,      41.7,      41.7,      41.7,      41.7,      41.7,      41.7,      41.7,
+        44.100002, 44.100002, 44.100002, 44.100002, 44.100002, 44.100002, 44.100002, 44.100002,
+        44.100002, 44.100002, 44.100002, 44.100002, 44.100002, 44.100002, 44.100002, 44.100002,
+        46.5,      46.5,      46.5,      46.5,      46.5,      46.5,      46.5,      46.5,
+        46.5,      46.5,      46.5,      46.5,      46.5,      46.5,      46.5,      46.5,
+        48.899998, 48.899998, 48.899998, 48.899998, 48.899998, 48.899998, 48.899998, 48.899998,
+        48.899998, 48.899998, 48.899998, 48.899998, 48.899998, 48.899998, 48.899998, 48.899998,
+        51.3,      51.3,      51.3,      51.3,      51.3,      51.3,      51.3,      51.3,
+        51.3,      51.3,      51.3,      51.3,      51.3,      51.3,      51.3,      51.3,
+        53.7,      53.7,      53.7,      53.7,      53.7,      53.7,      53.7,      53.7,
+        53.7,      53.7,      53.7,      53.7,      53.7,      53.7,      53.7,      53.7,
+        56.100002, 56.100002, 56.100002, 56.100002, 56.100002, 56.100002, 56.100002, 56.100002,
+        56.100002, 56.100002, 56.100002, 56.100002, 56.100002, 56.100002, 56.100002, 56.100002,
+        58.5,      58.5,      58.5,      58.5,      58.5,      58.5,      58.5,      58.5,
+        58.5,      58.5,      58.5,      58.5,      58.5,      58.5,      58.5,      58.5,
+        60.899998, 60.899998, 60.899998, 60.899998, 60.899998, 60.899998, 60.899998, 60.899998,
+        60.899998, 60.899998, 60.899998, 60.899998, 60.899998, 60.899998, 60.899998, 60.899998,
+        63.3,      63.3,      63.3,      63.3,      63.3,      63.3,      63.3,      63.3,
+        63.3,      63.3,      63.3,      63.3,      63.3,      63.3,      63.3,      63.3,
+        65.7,      65.7,      65.7,      65.7,      65.7,      65.7,      65.7,      65.7,
+        65.7,      65.7,      65.7,      65.7,      65.7,      65.7,      65.7,      65.7,
+        68.1,      68.1,      68.1,      68.1,      68.1,      68.1,      68.1,      68.1,
+        68.1,      68.1,      68.1,      68.1,      68.1,      68.1,      68.1,      68.1,
+        70.5,      70.5,      70.5,      70.5,      70.5,      70.5,      70.5,      70.5,
+        70.5,      70.5,      70.5,      70.5,      70.5,      70.5,      70.5,      70.5,
+        72.9,      72.9,      72.9,      72.9,      72.9,      72.9,      72.9,      72.9,
+        72.9,      72.9,      72.9,      72.9,      72.9,      72.9,      72.9,      72.9,
+        49.4,      49.4,      49.4,      49.4,      49.4,      49.4,      49.4,      49.4,
+        49.4,      49.4,      49.4,      49.4,      49.4,      49.4,      49.4,      49.4};
+    EXPECT_TRUE(ck::utils::check_err(
+        out_tensor2.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
+    EXPECT_TRUE(ck::utils::check_err(out_tensor2, ref_data, "Error: incorrect results!"));
+}
+#endif
+
+TEST(ReferenceConvolutionFWD, Conv3DGNCDHW)
+{
+    ck::utils::conv::ConvParam conv_param(3,
+                                          1,
+                                          1,
+                                          1,
+                                          2,
+                                          std::vector<ck::index_t>{3, 3, 3},
+                                          std::vector<ck::index_t>{6, 6, 6},
+                                          std::vector<ck::index_t>{1, 1, 1},
+                                          std::vector<ck::index_t>{1, 1, 1},
+                                          std::vector<ck::index_t>{0, 0, 0},
+                                          std::vector<ck::index_t>{0, 0, 0});
+
+    auto out_tensor = run_reference_convolution_forward<3,
+                                                        float,
+                                                        float,
+                                                        float,
+                                                        ck::tensor_layout::convolution::GNCDHW,
+                                                        ck::tensor_layout::convolution::GKCZYX,
+                                                        ck::tensor_layout::convolution::GNKDHW>(
+        conv_param, ck::utils::FillMonotonicSeq<float>{0.f, 0.1f});
+    std::vector<std::size_t> ref_dims{1, 1, 1, 4, 4, 4};
+    std::vector<float> ref_data{
+        407.7,     410.40002, 413.09998, 415.80002, 423.90002, 426.6,     429.30002, 432.,
+        440.1,     442.80002, 445.5,     448.2,     456.30002, 459.,      461.7,     464.40002,
+        504.90002, 507.6,     510.30002, 513.,      521.1,     523.8,     526.5,     529.2001,
+        537.3,     540.,      542.7001,  545.4,     553.5,     556.2001,  558.9,     561.6,
+        602.10004, 604.8,     607.5,     610.2,     618.3,     621.,      623.7,     626.4,
+        634.5,     637.2,     639.9,     642.60004, 650.7,     653.4,     656.10004, 658.8,
+        699.3,     702.,      704.7,     707.4,     715.5,     718.2,     720.9,     723.60004,
+        731.7,     734.4001,  737.10004, 739.8,     747.9001,  750.60004, 753.3,     756.};
+    EXPECT_TRUE(ck::utils::check_err(out_tensor.mDesc.GetLengths(),
+                                     ref_dims,
+                                     "Error [case 1]: wrong output tensor dimensions!"));
+    EXPECT_TRUE(ck::utils::check_err(out_tensor, ref_data, "Error [case 1]: incorrect results!"));
+}
+
+TEST(ReferenceConvolutionFWD, Conv3DGNCDHWStridesDilations)
+{
+    ck::utils::conv::ConvParam conv_param(3,
+                                          1,
+                                          1,
+                                          2,
+                                          2,
+                                          std::vector<ck::index_t>{3, 3, 3},
+                                          std::vector<ck::index_t>{12, 12, 12},
+                                          std::vector<ck::index_t>{3, 3, 3},
+                                          std::vector<ck::index_t>{1, 1, 1},
+                                          std::vector<ck::index_t>{0, 0, 0},
+                                          std::vector<ck::index_t>{0, 0, 0});
+
+    auto out_tensor = run_reference_convolution_forward<3,
+                                                        float,
+                                                        float,
+                                                        float,
+                                                        ck::tensor_layout::convolution::GNCDHW,
+                                                        ck::tensor_layout::convolution::GKCZYX,
+                                                        ck::tensor_layout::convolution::GNKDHW>(
+        conv_param, ck::utils::FillMonotonicSeq<float>{0.f, 0.1f});
+    std::vector<std::size_t> ref_dims{1, 1, 2, 4, 4, 4};
+    std::vector<float> ref_data{
+        2756.7002, 2764.7998, 2772.9001, 2781.,     2853.9001, 2862.,     2870.1,    2878.2002,
+        2951.1,    2959.2002, 2967.2998, 2975.4001, 3048.2998, 3056.4001, 3064.5,    3072.6,
+        3923.1,    3931.2,    3939.2998, 3947.4,    4020.2998, 4028.4001, 4036.5002, 4044.5999,
+        4117.5,    4125.6,    4133.7,    4141.8,    4214.7,    4222.8,    4230.9004, 4239.,
+        5089.5,    5097.5996, 5105.7,    5113.8,    5186.7,    5194.8,    5202.9,    5211.,
+        5283.9004, 5292.,     5300.0996, 5308.2,    5381.0996, 5389.2,    5397.3,    5405.4004,
+        6255.9004, 6264.0005, 6272.1,    6280.2,    6353.1,    6361.2,    6369.301,  6377.4,
+        6450.301,  6458.4,    6466.5,    6474.6,    6547.5,    6555.6,    6563.699,  6571.801,
+        2756.7002, 2764.7998, 2772.9001, 2781.,     2853.9001, 2862.,     2870.1,    2878.2002,
+        2951.1,    2959.2002, 2967.2998, 2975.4001, 3048.2998, 3056.4001, 3064.5,    3072.6,
+        3923.1,    3931.2,    3939.2998, 3947.4,    4020.2998, 4028.4001, 4036.5002, 4044.5999,
+        4117.5,    4125.6,    4133.7,    4141.8,    4214.7,    4222.8,    4230.9004, 4239.,
+        5089.5,    5097.5996, 5105.7,    5113.8,    5186.7,    5194.8,    5202.9,    5211.,
+        5283.9004, 5292.,     5300.0996, 5308.2,    5381.0996, 5389.2,    5397.3,    5405.4004,
+        6255.9004, 6264.0005, 6272.1,    6280.2,    6353.1,    6361.2,    6369.301,  6377.4,
+        6450.301,  6458.4,    6466.5,    6474.6,    6547.5,    6555.6,    6563.699,  6571.801};
+    EXPECT_TRUE(ck::utils::check_err(out_tensor.mDesc.GetLengths(),
+                                     ref_dims,
+                                     "Error [case 2]: wrong output tensor dimensions!"));
+    EXPECT_TRUE(ck::utils::check_err(
+        out_tensor, ref_data, "Error [case 2]: incorrect results!", 1e-4f, 1e-6f));
+}
diff --git a/test/softmax/CMakeLists.txt b/test/softmax/CMakeLists.txt
new file mode 100644
index 0000000000..4ba4012625
--- /dev/null
+++ b/test/softmax/CMakeLists.txt
@@ -0,0 +1,11 @@
+add_custom_target(test_softmax)
+
+add_gtest_executable(test_softmax_rank3 test_softmax_rank3.cpp)
+add_gtest_executable(test_softmax_rank4 test_softmax_rank4.cpp)
+add_gtest_executable(test_softmax_interface test_softmax_interface.cpp)
+target_link_libraries(test_softmax_rank3 PRIVATE utility device_softmax_instance)
+target_link_libraries(test_softmax_rank4 PRIVATE utility device_softmax_instance)
+target_link_libraries(test_softmax_interface PRIVATE utility device_softmax_instance)
+add_dependencies(test_softmax test_softmax_rank3)
+add_dependencies(test_softmax test_softmax_rank4)
+add_dependencies(test_softmax test_softmax_interface)
diff --git a/test/softmax/test_softmax_interface.cpp b/test/softmax/test_softmax_interface.cpp
new file mode 100644
index 0000000000..25f666f0ea
--- /dev/null
+++ b/test/softmax/test_softmax_interface.cpp
@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <stdexcept>
+#include <vector>
+
+#include "gtest/gtest.h"
+#include "test_softmax_util.hpp"
+
+class TestSoftmaxInterface : public ::testing::Test
+{
+    protected:
+    template <ck::index_t Rank, ck::index_t NumReduceDims>
+    using SoftmaxInstance =
+        ck::DeviceSoftmaxInstanceWrapper<Rank, NumReduceDims, 256, 1, 256, 1, 8, 1, 8, 8>;
+};
+
+TEST_F(TestSoftmaxInterface, IncorrectReduceDims)
+{
+    std::vector<ck::index_t> lengths{2, 128, 1536};
+    std::vector<ck::index_t> strides{128 * 1536, 1536, 1};
+
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported(lengths, strides, {-1})), std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported(lengths, strides, {3})), std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported(lengths, strides, {0, 1})),
+                 std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported(lengths, strides, {})), std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 2>{}.IsSupported(lengths, strides, {2, -1})),
+                 std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 2>{}.IsSupported(lengths, strides, {2, 4})),
+                 std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 2>{}.IsSupported(lengths, strides, {2})), std::runtime_error);
+}
+
+TEST_F(TestSoftmaxInterface, IncorrectLengthsSize)
+{
+    std::vector<ck::index_t> lengths{128, 1536};
+    std::vector<ck::index_t> strides{128 * 1536, 1536, 1};
+    std::vector<ck::index_t> reduce_dims{2};
+
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported({128, 1536}, strides, reduce_dims)),
+                 std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported({}, strides, reduce_dims)),
+                 std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported({1, 8, 128, 1536}, strides, reduce_dims)),
+                 std::runtime_error);
+}
+
+TEST_F(TestSoftmaxInterface, IncorrectStridesSize)
+{
+    std::vector<ck::index_t> lengths{2, 128, 1536};
+    std::vector<ck::index_t> reduce_dims{2};
+
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported(lengths, {1536, 1}, reduce_dims)),
+                 std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported(lengths, {}, reduce_dims)),
+                 std::runtime_error);
+    EXPECT_THROW((SoftmaxInstance<3, 1>{}.IsSupported(lengths, {1, 2, 3, 4}, reduce_dims)),
+                 std::runtime_error);
+}
+
+TEST_F(TestSoftmaxInterface, UnsupportedLengths)
+{
+    using SoftmaxInstance1 = ck::DeviceSoftmaxInstanceWrapper<3, 1, 256, 1, 256, 1, 8, 1, 8, 4>;
+    EXPECT_FALSE(SoftmaxInstance1{}.IsSupported({2, 128, 1500}, {128 * 1500, 1500, 1}, {2}));
+    EXPECT_FALSE(SoftmaxInstance1{}.IsSupported({2, 127, 1536}, {127 * 1536, 1536, 1}, {2}));
+    EXPECT_FALSE(SoftmaxInstance1{}.IsSupported({2, 128, 1537}, {128 * 1537, 1537, 1}, {2}));
+
+    // Reduction of middle dimensions
+    using SoftmaxInstance2 = ck::DeviceSoftmaxInstanceWrapper<3, 3, 256, 8, 32, 8, 8, 0, 8, 4>;
+    EXPECT_FALSE(SoftmaxInstance2{}.IsSupported({2, 128, 1536}, {128 * 1536, 1536, 1}, {0, 1, 2}));
+
+    // Reduction of middle dimensions
+    using SoftmaxInstance3 = ck::DeviceSoftmaxInstanceWrapper<3, 1, 256, 8, 32, 8, 8, 0, 4, 8>;
+    EXPECT_FALSE(SoftmaxInstance3{}.IsSupported({2, 128, 1536}, {128 * 1536, 1536, 1}, {2}));
+    EXPECT_FALSE(SoftmaxInstance3{}.IsSupported({2, 128, 1537}, {128 * 1537, 1537, 1}, {1}));
+    EXPECT_FALSE(SoftmaxInstance3{}.IsSupported({2, 128, 1540}, {128 * 1540, 1540, 1}, {1}));
+    EXPECT_FALSE(SoftmaxInstance3{}.IsSupported({2, 127, 1536}, {127 * 1536, 1536, 1}, {1}));
+}
+
+TEST_F(TestSoftmaxInterface, UnsupportedInstance)
+{
+    // Instance with InSrcVectorDim = 1, can't reduce middle dims if in/out vec size != 1
+    using SoftmaxInstance1 = ck::DeviceSoftmaxInstanceWrapper<3, 1, 256, 8, 32, 1, 8, 1, 8, 8>;
+    EXPECT_FALSE(SoftmaxInstance1{}.IsSupported({2, 128, 1024}, {128 * 1024, 1024, 1}, {0}));
+}
diff --git a/test/softmax/test_softmax_rank3.cpp b/test/softmax/test_softmax_rank3.cpp
new file mode 100644
index 0000000000..a8b950ce6b
--- /dev/null
+++ b/test/softmax/test_softmax_rank3.cpp
@@ -0,0 +1,35 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <stdexcept>
+#include <vector>
+
+#include "gtest/gtest.h"
+#include "test_softmax_util.hpp"
+
+template <ck::index_t N>
+using I = ck::Number<N>;
+#ifdef CK_ENABLE_FP16
+using F16 = ck::half_t;
+#endif
+using F32 = float;
+
+template <typename Tuple>
+class TestSoftmax : public ck::TestSoftmax<Tuple>
+{
+};
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    //         InDataType, AccDataType, OutDataType, Rank
+#ifdef CK_ENABLE_FP16
+    std::tuple<       F16,         F32,         F16,    I<3>>,
+#endif
+    std::tuple<       F32,         F32,         F32,    I<3>>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestSoftmax, KernelTypes);
+
+#include "test_softmax_ut_cases.inc"
diff --git a/test/softmax/test_softmax_rank4.cpp b/test/softmax/test_softmax_rank4.cpp
new file mode 100644
index 0000000000..cbd790ac9b
--- /dev/null
+++ b/test/softmax/test_softmax_rank4.cpp
@@ -0,0 +1,35 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <stdexcept>
+#include <vector>
+
+#include "gtest/gtest.h"
+#include "test_softmax_util.hpp"
+
+template <ck::index_t N>
+using I = ck::Number<N>;
+#ifdef CK_ENABLE_FP16
+using F16 = ck::half_t;
+#endif
+using F32 = float;
+
+template <typename Tuple>
+class TestSoftmax : public ck::TestSoftmax<Tuple>
+{
+};
+
+// clang-format off
+using KernelTypes = ::testing::Types<
+    //         InDataType, AccDataType, OutDataType, Rank
+#ifdef CK_ENABLE_FP16
+    std::tuple<       F16,         F32,         F16,    I<4>>,
+#endif
+    std::tuple<       F32,         F32,         F32,    I<4>>
+    >;
+// clang-format on
+
+TYPED_TEST_SUITE(TestSoftmax, KernelTypes);
+
+#include "test_softmax_ut_cases.inc"
diff --git a/test/softmax/test_softmax_ut_cases.inc b/test/softmax/test_softmax_ut_cases.inc
new file mode 100644
index 0000000000..cf5e4d2d2d
--- /dev/null
+++ b/test/softmax/test_softmax_ut_cases.inc
@@ -0,0 +1,60 @@
+#pragma once
+
+TYPED_TEST(TestSoftmax, ReduceOutermostDim)
+{
+    std::vector<ck::index_t> reduce_dims{this->Rank - 1};
+    this->Run(reduce_dims);
+}
+
+TYPED_TEST(TestSoftmax, ReduceMiddleDim)
+{
+    for(int dim = 0; dim < this->Rank - 1; ++dim)
+    {
+        std::vector<ck::index_t> reduce_dims{dim};
+        this->Run(reduce_dims);
+    }
+}
+
+TYPED_TEST(TestSoftmax, ReduceMultipleDimsWithOutermost)
+{
+    for(int dim = 0; dim < this->Rank - 1; ++dim)
+    {
+        std::vector<ck::index_t> reduce_dims{dim, this->Rank - 1};
+        this->Run(reduce_dims);
+    }
+}
+
+TYPED_TEST(TestSoftmax, ReduceMultipleMiddleDims)
+{
+    std::vector<ck::index_t> reduce_dims{0, 1};
+    if(this->Rank >= 3)
+    {
+        this->Run(reduce_dims);
+    }
+
+    if(this->Rank >= 4)
+    {
+        reduce_dims = std::vector<ck::index_t>{0, 2};
+        this->Run(reduce_dims);
+        reduce_dims = std::vector<ck::index_t>{0, 1, 2};
+        this->Run(reduce_dims);
+    }
+}
+
+TYPED_TEST(TestSoftmax, ReduceAllDims)
+{
+    std::vector<ck::index_t> reduce_dims(this->Rank);
+    std::iota(std::begin(reduce_dims), std::end(reduce_dims), 0);
+    this->Run(reduce_dims);
+}
+
+TYPED_TEST(TestSoftmax, ReduceOddLengths)
+{
+    this->in_lengths_ = {{3, 63, 1032}};
+    if(this->Rank >= 4)
+    {
+        this->in_lengths_ = {{1, 3, 63, 1032}};
+    }
+    this->Run({this->Rank - 1});
+    this->Run({this->Rank - 2});
+}
diff --git a/test/softmax/test_softmax_util.hpp b/test/softmax/test_softmax_util.hpp
new file mode 100644
index 0000000000..1409af8453
--- /dev/null
+++ b/test/softmax/test_softmax_util.hpp
@@ -0,0 +1,226 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <string>
+#include <sstream>
+#include <tuple>
+#include <vector>
+#include <gtest/gtest.h>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "include/ck/utility/data_type.hpp"
+#include "profiler/profile_softmax_impl.hpp"
+
+namespace ck {
+
+template <typename Range>
+std::string serialize_range(const Range& range)
+{
+    std::stringstream ss;
+    for(auto& r : range)
+    {
+        ss << r << ", ";
+    }
+    std::string str = ss.str();
+    return std::string(str.begin(), str.end() - 2);
+}
+
+template <typename Tuple>
+class TestSoftmax : public ::testing::Test
+{
+    protected:
+    using InDataType              = std::tuple_element_t<0, Tuple>;
+    using AccDataType             = std::tuple_element_t<1, Tuple>;
+    using OutDataType             = std::tuple_element_t<2, Tuple>;
+    static constexpr index_t Rank = std::tuple_element_t<3, Tuple>{}.value;
+
+    public:
+    std::vector<std::vector<index_t>> in_lengths_ = {{2, 128, 1024}, {4, 16, 8448}, {128, 128, 64}};
+    std::vector<std::vector<AccDataType>> scales_ = {{2, 0}, {0, 2}, {2, 2}};
+    bool bench_                                   = false; // measure kernel performance
+    bool verify_                                  = true;
+
+    void SetUp() override
+    {
+        if constexpr(Rank == 4)
+        {
+            in_lengths_ = std::vector<std::vector<index_t>>{
+                {1, 2, 128, 1024}, {2, 4, 16, 8448}, {1, 128, 128, 64}};
+        }
+    }
+
+    void RunSingle(std::vector<index_t> in_length,
+                   std::vector<index_t> reduce_dims,
+                   AccDataType alpha,
+                   AccDataType beta)
+    {
+        int init_method = 1; // integer value initialization
+        bool log        = false;
+        std::vector<ck::index_t> strides; // intenionally empty, to get packed layout.
+        bool pass = false;
+
+        if constexpr(Rank == 3)
+        {
+            if(reduce_dims.size() == 1)
+                pass = ck::profiler::
+                    profile_softmax_impl<InDataType, AccDataType, OutDataType, Rank, 1>(verify_,
+                                                                                        init_method,
+                                                                                        log,
+                                                                                        bench_,
+                                                                                        in_length,
+                                                                                        strides,
+                                                                                        reduce_dims,
+                                                                                        alpha,
+                                                                                        beta);
+            else if(reduce_dims.size() == 2)
+                pass = ck::profiler::
+                    profile_softmax_impl<InDataType, AccDataType, OutDataType, Rank, 2>(verify_,
+                                                                                        init_method,
+                                                                                        log,
+                                                                                        bench_,
+                                                                                        in_length,
+                                                                                        strides,
+                                                                                        reduce_dims,
+                                                                                        alpha,
+                                                                                        beta);
+            else if(reduce_dims.size() == 3)
+                pass = ck::profiler::
+                    profile_softmax_impl<InDataType, AccDataType, OutDataType, Rank, 3>(verify_,
+                                                                                        init_method,
+                                                                                        log,
+                                                                                        bench_,
+                                                                                        in_length,
+                                                                                        strides,
+                                                                                        reduce_dims,
+                                                                                        alpha,
+                                                                                        beta);
+        }
+        else if constexpr(Rank == 4)
+        {
+            if(reduce_dims.size() == 1)
+                pass = ck::profiler::
+                    profile_softmax_impl<InDataType, AccDataType, OutDataType, Rank, 1>(verify_,
+                                                                                        init_method,
+                                                                                        log,
+                                                                                        bench_,
+                                                                                        in_length,
+                                                                                        strides,
+                                                                                        reduce_dims,
+                                                                                        alpha,
+                                                                                        beta);
+            else if(reduce_dims.size() == 2)
+                pass = ck::profiler::
+                    profile_softmax_impl<InDataType, AccDataType, OutDataType, Rank, 2>(verify_,
+                                                                                        init_method,
+                                                                                        log,
+                                                                                        bench_,
+                                                                                        in_length,
+                                                                                        strides,
+                                                                                        reduce_dims,
+                                                                                        alpha,
+                                                                                        beta);
+            else if(reduce_dims.size() == 3)
+                pass = ck::profiler::
+                    profile_softmax_impl<InDataType, AccDataType, OutDataType, Rank, 3>(verify_,
+                                                                                        init_method,
+                                                                                        log,
+                                                                                        bench_,
+                                                                                        in_length,
+                                                                                        strides,
+                                                                                        reduce_dims,
+                                                                                        alpha,
+                                                                                        beta);
+            else if(reduce_dims.size() == 4)
+                pass = ck::profiler::
+                    profile_softmax_impl<InDataType, AccDataType, OutDataType, Rank, 4>(verify_,
+                                                                                        init_method,
+                                                                                        log,
+                                                                                        bench_,
+                                                                                        in_length,
+                                                                                        strides,
+                                                                                        reduce_dims,
+                                                                                        alpha,
+                                                                                        beta);
+        };
+
+        EXPECT_TRUE(pass);
+    }
+
+    void Run(std::vector<index_t> reduce_dims = {})
+    {
+        if(reduce_dims.empty())
+        {
+            reduce_dims.push_back(Rank - 1);
+        }
+
+        for(auto in_length : this->in_lengths_)
+        {
+            for(auto scale : this->scales_)
+            {
+                this->RunSingle(in_length, reduce_dims, scale[0], scale[1]);
+            }
+        }
+    }
+};
+
+template <index_t Rank,
+          index_t NumReduceDim,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          index_t InSrcVectorDim,
+          index_t InSrcVectorSize,
+          index_t OutDstVectorSize>
+struct DeviceSoftmaxInstanceWrapper
+{
+    using F16  = half_t;
+    using F32  = float;
+    using Pass = tensor_operation::element_wise::PassThrough;
+
+    using InDataType   = F16;
+    using AccDataType  = F32;
+    using OutDataType  = F16;
+    using InElementOp  = Pass;
+    using AccElementOp = Pass;
+
+    using DeviceSoftmaxInstance = tensor_operation::device::DeviceSoftmaxImpl<InDataType,
+                                                                              AccDataType,
+                                                                              OutDataType,
+                                                                              InElementOp,
+                                                                              AccElementOp,
+                                                                              Rank,
+                                                                              NumReduceDim,
+                                                                              BlockSize,
+                                                                              MThreadClusterSize,
+                                                                              KThreadClusterSize,
+                                                                              MThreadSliceSize,
+                                                                              KThreadSliceSize,
+                                                                              InSrcVectorDim,
+                                                                              InSrcVectorSize,
+                                                                              OutDstVectorSize>;
+
+    bool IsSupported(const std::vector<index_t> in_lengths,
+                     const std::vector<index_t> in_strides,
+                     const std::vector<index_t> reduce_dims) const
+    {
+        auto softmax  = DeviceSoftmaxInstance{};
+        auto argument = softmax.MakeArgument(in_lengths,
+                                             in_strides,
+                                             reduce_dims,
+                                             1,       // alpha
+                                             1,       // beta
+                                             nullptr, // in_dev
+                                             nullptr, // in_out
+                                             Pass{},  // in elementwise op
+                                             Pass{}); // acc elementwise op
+        return softmax.IsSupportedArgument(argument);
+    }
+};
+
+} // namespace ck
diff --git a/test/space_filling_curve/CMakeLists.txt b/test/space_filling_curve/CMakeLists.txt
new file mode 100644
index 0000000000..a527268042
--- /dev/null
+++ b/test/space_filling_curve/CMakeLists.txt
@@ -0,0 +1 @@
+add_test_executable(test_space_filling_curve space_filling_curve.cpp)
diff --git a/test/space_filling_curve/space_filling_curve.cpp b/test/space_filling_curve/space_filling_curve.cpp
new file mode 100644
index 0000000000..a192ecb28f
--- /dev/null
+++ b/test/space_filling_curve/space_filling_curve.cpp
@@ -0,0 +1,195 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+#include <iostream>
+#include <numeric>
+#include <cassert>
+
+#include "ck/ck.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_space_filling_curve.hpp"
+
+using namespace ck;
+
+void traverse_using_space_filling_curve_linear();
+void traverse_using_space_filling_curve_snakecurved();
+
+int main(int argc, char** argv)
+{
+    (void)argc;
+    (void)argv;
+
+    traverse_using_space_filling_curve_linear();
+    traverse_using_space_filling_curve_snakecurved();
+
+    return 0;
+}
+
+void traverse_using_space_filling_curve_linear()
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+
+    using TensorLengths    = Sequence<3, 2, 2>;
+    using DimAccessOrder   = Sequence<2, 0, 1>;
+    using ScalarsPerAccess = Sequence<1, 1, 1>;
+    using SpaceFillingCurve =
+        SpaceFillingCurve<TensorLengths, DimAccessOrder, ScalarsPerAccess, false>;
+
+    constexpr auto expected = make_tuple(make_tuple(0, 0, 0),
+                                         make_tuple(0, 1, 0),
+                                         make_tuple(1, 0, 0),
+                                         make_tuple(1, 1, 0),
+                                         make_tuple(2, 0, 0),
+                                         make_tuple(2, 1, 0),
+                                         make_tuple(0, 0, 1),
+                                         make_tuple(0, 1, 1),
+                                         make_tuple(1, 0, 1),
+                                         make_tuple(1, 1, 1),
+                                         make_tuple(2, 0, 1),
+                                         make_tuple(2, 1, 1));
+
+    constexpr index_t num_access = SpaceFillingCurve::GetNumOfAccess();
+
+    static_assert(num_access == reduce_on_sequence(TensorLengths{} / ScalarsPerAccess{},
+                                                   math::multiplies{},
+                                                   Number<1>{}));
+
+    static_for<1, num_access, 1>{}([&](auto i) {
+        constexpr auto idx_curr = SpaceFillingCurve::GetIndex(i);
+
+        static_assert(idx_curr[I0] == expected[i][I0]);
+        static_assert(idx_curr[I1] == expected[i][I1]);
+        static_assert(idx_curr[I2] == expected[i][I2]);
+
+        constexpr auto backward_step = SpaceFillingCurve::GetBackwardStep(i);
+        constexpr auto expected_step = expected[i - I1] - expected[i];
+        static_assert(backward_step[I0] == expected_step[I0]);
+        static_assert(backward_step[I1] == expected_step[I1]);
+        static_assert(backward_step[I2] == expected_step[I2]);
+    });
+
+    static_for<0, num_access - 1, 1>{}([&](auto i) {
+        constexpr auto idx_curr = SpaceFillingCurve::GetIndex(i);
+
+        static_assert(idx_curr[I0] == expected[i][I0]);
+        static_assert(idx_curr[I1] == expected[i][I1]);
+        static_assert(idx_curr[I2] == expected[i][I2]);
+
+        constexpr auto forward_step  = SpaceFillingCurve::GetForwardStep(i);
+        constexpr auto expected_step = expected[i + I1] - expected[i];
+        static_assert(forward_step[I0] == expected_step[I0]);
+        static_assert(forward_step[I1] == expected_step[I1]);
+        static_assert(forward_step[I2] == expected_step[I2]);
+    });
+}
+
+void traverse_using_space_filling_curve_snakecurved()
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+
+    using TensorLengths    = Sequence<16, 10, 9>;
+    using DimAccessOrder   = Sequence<2, 0, 1>;
+    using ScalarsPerAccess = Sequence<4, 2, 3>;
+    using SpaceFillingCurve =
+        SpaceFillingCurve<TensorLengths, DimAccessOrder, ScalarsPerAccess, true>;
+
+    constexpr auto expected = make_tuple(make_tuple(0, 0, 0),
+                                         make_tuple(0, 2, 0),
+                                         make_tuple(0, 4, 0),
+                                         make_tuple(0, 6, 0),
+                                         make_tuple(0, 8, 0),
+                                         make_tuple(4, 8, 0),
+                                         make_tuple(4, 6, 0),
+                                         make_tuple(4, 4, 0),
+                                         make_tuple(4, 2, 0),
+                                         make_tuple(4, 0, 0),
+                                         make_tuple(8, 0, 0),
+                                         make_tuple(8, 2, 0),
+                                         make_tuple(8, 4, 0),
+                                         make_tuple(8, 6, 0),
+                                         make_tuple(8, 8, 0),
+                                         make_tuple(12, 8, 0),
+                                         make_tuple(12, 6, 0),
+                                         make_tuple(12, 4, 0),
+                                         make_tuple(12, 2, 0),
+                                         make_tuple(12, 0, 0),
+                                         make_tuple(12, 0, 3),
+                                         make_tuple(12, 2, 3),
+                                         make_tuple(12, 4, 3),
+                                         make_tuple(12, 6, 3),
+                                         make_tuple(12, 8, 3),
+                                         make_tuple(8, 8, 3),
+                                         make_tuple(8, 6, 3),
+                                         make_tuple(8, 4, 3),
+                                         make_tuple(8, 2, 3),
+                                         make_tuple(8, 0, 3),
+                                         make_tuple(4, 0, 3),
+                                         make_tuple(4, 2, 3),
+                                         make_tuple(4, 4, 3),
+                                         make_tuple(4, 6, 3),
+                                         make_tuple(4, 8, 3),
+                                         make_tuple(0, 8, 3),
+                                         make_tuple(0, 6, 3),
+                                         make_tuple(0, 4, 3),
+                                         make_tuple(0, 2, 3),
+                                         make_tuple(0, 0, 3),
+                                         make_tuple(0, 0, 6),
+                                         make_tuple(0, 2, 6),
+                                         make_tuple(0, 4, 6),
+                                         make_tuple(0, 6, 6),
+                                         make_tuple(0, 8, 6),
+                                         make_tuple(4, 8, 6),
+                                         make_tuple(4, 6, 6),
+                                         make_tuple(4, 4, 6),
+                                         make_tuple(4, 2, 6),
+                                         make_tuple(4, 0, 6),
+                                         make_tuple(8, 0, 6),
+                                         make_tuple(8, 2, 6),
+                                         make_tuple(8, 4, 6),
+                                         make_tuple(8, 6, 6),
+                                         make_tuple(8, 8, 6),
+                                         make_tuple(12, 8, 6),
+                                         make_tuple(12, 6, 6),
+                                         make_tuple(12, 4, 6),
+                                         make_tuple(12, 2, 6),
+                                         make_tuple(12, 0, 6));
+
+    constexpr index_t num_access = SpaceFillingCurve::GetNumOfAccess();
+
+    static_assert(num_access == reduce_on_sequence(TensorLengths{} / ScalarsPerAccess{},
+                                                   math::multiplies{},
+                                                   Number<1>{}));
+
+    static_for<1, num_access, 1>{}([&](auto i) {
+        constexpr auto idx_curr = SpaceFillingCurve::GetIndex(i);
+
+        static_assert(idx_curr[I0] == expected[i][I0]);
+        static_assert(idx_curr[I1] == expected[i][I1]);
+        static_assert(idx_curr[I2] == expected[i][I2]);
+
+        constexpr auto backward_step = SpaceFillingCurve::GetBackwardStep(i);
+        constexpr auto expected_step = expected[i - I1] - expected[i];
+        static_assert(backward_step[I0] == expected_step[I0]);
+        static_assert(backward_step[I1] == expected_step[I1]);
+        static_assert(backward_step[I2] == expected_step[I2]);
+    });
+
+    static_for<0, num_access - 1, 1>{}([&](auto i) {
+        constexpr auto idx_curr = SpaceFillingCurve::GetIndex(i);
+
+        static_assert(idx_curr[I0] == expected[i][I0]);
+        static_assert(idx_curr[I1] == expected[i][I1]);
+        static_assert(idx_curr[I2] == expected[i][I2]);
+
+        constexpr auto forward_step  = SpaceFillingCurve::GetForwardStep(i);
+        constexpr auto expected_step = expected[i + I1] - expected[i];
+        static_assert(forward_step[I0] == expected_step[I0]);
+        static_assert(forward_step[I1] == expected_step[I1]);
+        static_assert(forward_step[I2] == expected_step[I2]);
+    });
+}
diff --git a/test/wmma_op/CMakeLists.txt b/test/wmma_op/CMakeLists.txt
new file mode 100644
index 0000000000..e553253c62
--- /dev/null
+++ b/test/wmma_op/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_test_executable(test_wmma_op wmma_op.cpp)
+target_link_libraries(test_wmma_op PRIVATE utility)
diff --git a/test/wmma_op/wmma_op.cpp b/test/wmma_op/wmma_op.cpp
new file mode 100644
index 0000000000..47d8c7ed6f
--- /dev/null
+++ b/test/wmma_op/wmma_op.cpp
@@ -0,0 +1,67 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cstdlib>
+#include <iostream>
+#include <numeric>
+#include <tuple>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "test/wmma_op/wmma_op_util.hpp"
+
+template <typename SrcType,
+          typename DstType,
+          typename GPUAccType,
+          typename CPUAccType,
+          ck::index_t AccNum>
+bool run_test()
+{
+    using Row         = ck::tensor_layout::gemm::RowMajor;
+    using Col         = ck::tensor_layout::gemm::ColumnMajor;
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    bool pass         = true;
+
+    const auto matmul_default = ck::wmma_op_util::matmul<SrcType, DstType, GPUAccType, AccNum>;
+    const auto matmul_swizzle_a =
+        ck::wmma_op_util::matmul_swizzle_a<SrcType, DstType, GPUAccType, AccNum>;
+
+    const auto wmma_kernel_container = std::make_tuple(matmul_default, matmul_swizzle_a);
+
+    ck::static_for<0, 2, 1>{}([&](auto i) {
+        pass &=
+            ck::wmma_op_util::TestWmma<decltype(std::get<ck::Number<i>{}>(wmma_kernel_container)),
+                                       SrcType,
+                                       SrcType,
+                                       DstType,
+                                       GPUAccType,
+                                       CPUAccType,
+                                       decltype(Row{}),
+                                       decltype(Col{}),
+                                       decltype(Row{}),
+                                       PassThrough,
+                                       PassThrough,
+                                       PassThrough,
+                                       AccNum>{}(std::get<ck::Number<i>{}>(wmma_kernel_container));
+    });
+
+    return pass ? 1 : 0;
+}
+int main(int, char*[])
+{
+    bool pass = true;
+    // clang-format off
+    //              |SrcType     |DstType     |GPUAccType  |CPUAccType |AccNum
+    pass &= run_test<ck::half_t,  ck::half_t,  float,       float,      8     >();
+    pass &= run_test<ck::bhalf_t, ck::bhalf_t, float,       float,      8     >();
+    pass &= run_test<ck::half_t,  ck::half_t,  ck::half_t,  ck::half_t, 16    >();
+    pass &= run_test<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t, float,      16    >();
+    pass &= run_test<int8_t,      int8_t,      int32_t,     int32_t,    8     >();
+    // clang-format on
+
+    std::cout << "TestGemm ..... " << (pass ? "SUCCESS" : "FAILURE") << std::endl;
+    return pass ? 0 : 1;
+}
diff --git a/test/wmma_op/wmma_op_util.hpp b/test/wmma_op/wmma_op_util.hpp
new file mode 100644
index 0000000000..49782bce6e
--- /dev/null
+++ b/test/wmma_op/wmma_op_util.hpp
@@ -0,0 +1,409 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/utility/amd_wmma.hpp"
+
+namespace ck {
+namespace wmma_op_util {
+
+template <typename src_vec, typename acc_vec>
+__device__ void builtin_wmma_naive_selector(const src_vec&, const src_vec&, acc_vec&)
+{
+}
+
+template <>
+__device__ void
+builtin_wmma_naive_selector<half16_t,
+                            StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, float, 1, 8, true>>(
+    const half16_t& reg_a,
+    const half16_t& reg_b,
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, float, 1, 8, true>& reg_c)
+{
+    intrin_wmma_f32_16x16x16_f16_w32<16, 16>::Run(
+        reg_a, reg_b, reg_c.GetVectorTypeReference(Number<0>{}));
+}
+
+template <>
+__device__ void
+builtin_wmma_naive_selector<bhalf16_t,
+                            StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, float, 1, 8, true>>(
+    const bhalf16_t& reg_a,
+    const bhalf16_t& reg_b,
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, float, 1, 8, true>& reg_c)
+{
+    intrin_wmma_f32_16x16x16_bf16_w32<16, 16>::Run(
+        reg_a, reg_b, reg_c.GetVectorTypeReference(Number<0>{}));
+}
+
+template <>
+__device__ void
+builtin_wmma_naive_selector<half16_t,
+                            StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, half_t, 1, 16, true>>(
+    const half16_t& reg_a,
+    const half16_t& reg_b,
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, half_t, 1, 16, true>& reg_c)
+{
+    intrin_wmma_f16_16x16x16_f16_w32<16, 16, 0>::Run(
+        reg_a, reg_b, reg_c.GetVectorTypeReference(Number<0>{}));
+}
+
+template <>
+__device__ void builtin_wmma_naive_selector<
+    bhalf16_t,
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, bhalf_t, 1, 16, true>>(
+    const bhalf16_t& reg_a,
+    const bhalf16_t& reg_b,
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, bhalf_t, 1, 16, true>& reg_c)
+{
+    intrin_wmma_bf16_16x16x16_bf16_w32<16, 16, 0>::Run(
+        reg_a, reg_b, reg_c.GetVectorTypeReference(Number<0>{}));
+}
+
+template <>
+__device__ void
+builtin_wmma_naive_selector<int8x16_t,
+                            StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, int32_t, 1, 8, true>>(
+    const int8x16_t& reg_a,
+    const int8x16_t& reg_b,
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, int32_t, 1, 8, true>& reg_c)
+{
+    intrin_wmma_i32_16x16x16_iu8_w32<16, 16, true, true, false>::Run(
+        reg_a, reg_b, reg_c.GetVectorTypeReference(Number<0>{}));
+}
+
+#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
+template <>
+__device__ void
+builtin_wmma_naive_selector<int4x16_t,
+                            StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, int32_t, 1, 8, true>>(
+    const int4x16_t& reg_a,
+    const int4x16_t& reg_b,
+    StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, int32_t, 1, 8, true>& reg_c)
+{
+    intrin_wmma_i32_16x16x16_iu4_w32<16, 16, true, true, false>::Run(
+        reg_a, reg_b, reg_c.GetVectorTypeReference(Number<0>{}));
+}
+#endif
+
+template <typename src_t, typename dst_t, typename acc_t, index_t acc_num>
+__global__ void matmul(const src_t* a, const src_t* b, dst_t* c)
+{
+    __shared__ src_t p_shared[16 * 16 * 2];
+    const int lIdx = threadIdx.x;
+    // a and b fragments are stored in 8 VGPRs each, in packed format, so 16 elements each for a and
+    // b a_frag will store one column of the 16x16 matrix tile b_frag will store one row of the
+    // 16x16 matrix tile
+    using src_vec  = typename vector_type<src_t, 16>::type;
+    src_vec a_frag = {};
+    src_vec b_frag = {};
+
+    src_vec a_temp = {};
+    src_vec b_temp = {};
+    // initialize c fragment to 0
+    using acc_vec = StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, acc_t, 1, acc_num, true>;
+    acc_vec c_thread_buf_;
+
+    // lane is (0-31) mod 16 instead of 0-31 due to matrix replication in gfx11
+    // see https://atlvsp3.amd.com/sp3_gfx11_5_instructions.pdf page 482
+    // TODO: remove this dependency in gfx12 https://ontrack-internal.amd.com/browse/DEGFXSP3-101
+    const int lane    = lIdx % 16;
+    const int lane_lo = lIdx / 2;
+    const int lane_hi = lIdx % 2;
+    for(int ele = 0; ele < 8; ++ele)
+    {
+        a_temp[ele] = a[8 * lane_hi + 16 * lane_lo + ele];
+    }
+
+    for(int ele = 0; ele < 8; ++ele)
+    {
+        b_temp[ele] = b[8 * lane_hi + 16 * lane_lo + ele];
+    }
+
+    __syncthreads();
+
+    for(int ele = 0; ele < 8; ++ele)
+    {
+        p_shared[8 * 16 * lane_hi + 8 * lane_lo + ele] = a_temp[ele];
+    }
+
+    for(int ele = 0; ele < 8; ++ele)
+    {
+        p_shared[8 * 16 * lane_hi + 8 * lane_lo + ele + 16 * 16] = b_temp[ele];
+    }
+
+    asm volatile("\
+    s_waitcnt lgkmcnt(0) \n \
+    s_barrier \
+    " ::);
+
+    for(int ele = 0; ele < 16; ++ele)
+    {
+        b_frag[ele] = p_shared[(ele / 8) * 16 * 8 + 8 * lane + ele % 8 + 16 * 16];
+    }
+    // follow origin design
+    for(int ele = 0; ele < 16; ++ele)
+    {
+        a_frag[ele] = p_shared[(ele / 8) * 16 * 8 + 8 * lane + ele % 8];
+    }
+
+    asm volatile("\
+    s_waitcnt lgkmcnt(0) \n \
+    s_barrier \
+    " ::);
+
+    // sync threads, similar to mma_sync
+    // __syncthreads();
+    builtin_wmma_naive_selector<src_vec, acc_vec>(a_frag, b_frag, c_thread_buf_);
+    // since only fp16_fp32 asm wmma implemented for experiment purpose, restrict test case to fp16
+    // when enable this ck::amd_assembly_wmma_f32_16x16x16_f16_w32(a_frag, b_frag,
+    // c_thread_buf_.GetVectorTypeReference(Number<0>{}).template AsType<float8_t>()(Number<0>{}));
+    __syncthreads();
+    // wait for results, similar to mma_sync
+    static_for<0, 8, 1>{}([&](auto ele) {
+        const int r = ele * 2 + (lIdx / 16);
+        // store results from unpacked c_thread_buf_ output
+        c[16 * r + lane] = ck::type_convert<dst_t>(c_thread_buf_[Number<ele * acc_num / 8>{}]);
+    });
+}
+
+template <typename src_t, typename dst_t, typename acc_t, index_t acc_num>
+__global__ void matmul_swizzle_a(const src_t* a, const src_t* b, dst_t* c)
+{
+    const int lIdx = threadIdx.x;
+
+    using src_vec  = typename vector_type<src_t, 16>::type;
+    src_vec a_frag = {};
+    src_vec b_frag = {};
+    using acc_vec  = StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr, acc_t, 1, acc_num, true>;
+    acc_vec c_thread_buf_;
+
+    const int lane = lIdx % 16;
+
+    for(int ele = 0; ele < 16; ++ele)
+    {
+        b_frag[ele] = b[16 * lane + ele];
+    }
+
+    const int offset_m = (((lane & 1) << 3) | (lane >> 1));
+    for(int ele = 0; ele < 16; ++ele)
+    {
+        a_frag[ele] = a[16 * offset_m + ele];
+    }
+
+    __syncthreads();
+    builtin_wmma_naive_selector<src_vec, acc_vec>(a_frag, b_frag, c_thread_buf_);
+    __syncthreads();
+
+    static_for<0, 8, 1>{}([&](auto ele) {
+        const int blk = lIdx / 16;
+        const int r   = ele;
+        c[16 * 8 * blk + 16 * r + lane] =
+            ck::type_convert<dst_t>(c_thread_buf_[Number<ele * acc_num / 8>{}]);
+    });
+}
+
+struct GemmParams
+{
+    GemmParams() : M(16), N(16), K(16), StrideA(16), StrideB(16), StrideC(16), alpha(1), beta(0) {}
+
+    ck::index_t M;
+    ck::index_t N;
+    ck::index_t K;
+
+    ck::index_t StrideA;
+    ck::index_t StrideB;
+    ck::index_t StrideC;
+
+    float alpha;
+    float beta;
+};
+
+template <typename GemmInstance,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+void RunHostGEMM(const Tensor<ADataType>& A,
+                 const Tensor<BDataType>& B,
+                 Tensor<CDataType>& C,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+{
+    auto ref_gemm     = GemmInstance{};
+    auto ref_invoker  = ref_gemm.MakeInvoker();
+    auto ref_argument = ref_gemm.MakeArgument(A, B, C, a_element_op, b_element_op, c_element_op);
+
+    ref_invoker.Run(ref_argument);
+}
+
+template <typename KernelType, typename ADataType, typename BDataType, typename CDataType>
+bool RunDeviceGEMM(KernelType kernel,
+                   const Tensor<ADataType>& A,
+                   const Tensor<BDataType>& B,
+                   Tensor<CDataType>& C)
+{
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * A.mDesc.GetElementSpaceSize());
+    DeviceMem b_n_k_device_buf(sizeof(BDataType) * B.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * C.mDesc.GetElementSpaceSize());
+
+    a_m_k_device_buf.ToDevice(A.mData.data());
+    b_n_k_device_buf.ToDevice(B.mData.data());
+    kernel<<<1, 32>>>(static_cast<const ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+                      static_cast<const BDataType*>(b_n_k_device_buf.GetDeviceBuffer()),
+                      static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()));
+    c_m_n_device_buf.FromDevice(C.mData.data());
+
+    return true;
+}
+
+template <typename DeviceWmma,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename GPUAccDataType,
+          typename CPUAccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t CAccNum>
+struct TestWmma
+{
+    auto PrepareGemmTensor(const ck::wmma_op_util::GemmParams& params)
+    {
+        auto f_host_tensor_descriptor =
+            [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+                if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+                {
+                    return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                                std::vector<std::size_t>({stride, 1}));
+                }
+                else
+                {
+                    return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                                std::vector<std::size_t>({1, stride}));
+                }
+            };
+
+        Tensor<ADataType> a_m_k(
+            f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
+        Tensor<BDataType> b_n_k(
+            f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
+        Tensor<CDataType> c_m_n_host_result(
+            f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
+        Tensor<CDataType> c_m_n_device_result(
+            f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
+
+        auto f_generate_tensor_value = [](auto& tensor, auto type) {
+            using dataType = decltype(type);
+
+            tensor.GenerateTensorValue(GeneratorTensor_2<dataType>{-5, 5});
+        };
+
+        f_generate_tensor_value(a_m_k, ADataType{});
+        f_generate_tensor_value(b_n_k, BDataType{});
+
+        return std::make_tuple(a_m_k, b_n_k, c_m_n_host_result, c_m_n_device_result);
+    }
+
+    auto operator()(const DeviceWmma& wmma_kernel)
+    {
+        std::cout << "ALayout = " << ALayout{}.name << ", BLayout = " << BLayout{}.name
+                  << ", CLayout = " << CLayout{}.name << std::endl;
+
+        // Arrange
+        ck::wmma_op_util::GemmParams params;
+        params.M       = 16;
+        params.N       = 16;
+        params.K       = 16;
+        params.StrideA = 16;
+        params.StrideB = 16;
+        params.StrideC = 16;
+
+        auto host_tensors = PrepareGemmTensor(params);
+
+        const Tensor<ADataType>& a  = std::get<0>(host_tensors);
+        const Tensor<BDataType>& b  = std::get<1>(host_tensors);
+        Tensor<CDataType>& c_host   = std::get<2>(host_tensors);
+        Tensor<CDataType>& c_device = std::get<3>(host_tensors);
+
+        auto a_element_op = AElementwiseOperation{};
+        auto b_element_op = BElementwiseOperation{};
+        auto c_element_op = CElementwiseOperation{};
+
+        using ReferenceGemmInstance =
+            ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                      BDataType,
+                                                      CDataType,
+                                                      CPUAccDataType,
+                                                      AElementwiseOperation,
+                                                      BElementwiseOperation,
+                                                      CElementwiseOperation>;
+        ck::wmma_op_util::RunHostGEMM<ReferenceGemmInstance>(
+            a, b, c_host, a_element_op, b_element_op, c_element_op);
+
+        // Act
+        bool is_supported = ck::wmma_op_util::RunDeviceGEMM(wmma_kernel, a, b, c_device);
+
+        if(is_supported)
+        {
+            // Assert
+            bool res = false;
+            if(std::is_same<CDataType, float>::value)
+            {
+                res = ck::utils::check_err(c_device.mData, c_host.mData);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, ck::half_t>::value)
+            {
+                res = ck::utils::check_err(c_device.mData, c_host.mData);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, ck::bhalf_t>::value)
+            {
+                // 0.5 Pixel Error Tolerance is introduced by Accumulator difference.
+                // BF16 WMMA Accumulator is in BF16 Type while On Host-side Accumulator is Float.
+                res = ck::utils::check_err(
+                    c_device.mData, c_host.mData, "Error: Incorrect results!", 0, 1.0);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, int8_t>::value)
+            {
+                res = ck::utils::check_err(c_device.mData, c_host.mData);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else if(std::is_same<CDataType, double>::value)
+            {
+                res = ck::utils::check_err(c_device.mData, c_host.mData);
+                std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+            }
+            else
+            {
+                std::cout << "UNSUPPORTED CDataType" << std::endl;
+            }
+
+            return res;
+        }
+        else
+        {
+            return true;
+        }
+    }
+};
+
+} // namespace wmma_op_util
+} // namespace ck